Anthony
@abucci@buc.ci
I'm Anthony and I'm an organic computer scientist and a Luddite. I believe that unsound research lends itself to social harms.
Anthony
@abucci@buc.ci
lol. just found out that in 2019, OpenAI criti-hyped GPT-2 as being "too dangerous to release" https://techcrunch.com/2019/02/17/openai-text-generator-dangerous/
@davidgerard they weren't wrong that it was too dangerous to release if only in the dumbest way possible.
@davidgerard Is this being brought up now because of Anthropic's latest nonsense?
@Azuaron @davidgerard What did I miss? What stupidity is Anthropic spouting now?
@davidgerard @pier But I heard it "escaped containment" and found security exploits in every single operating system in the world and every application and every web server and every toaster and every tree and bunny rabbit!
Surely Anthropic wouldn't just tell lies on the internet?
@davidgerard @pier @Azuaron Holy shit that's why they bought Bun. To make nothingburgers.
@tarmil @davidgerard @pier @Azuaron did a search for "claude mythos" and this was one of the results
@autonomousapps @tarmil @davidgerard @pier @Azuaron at this point, I'd be far more interested in what psychiatric professionals would have to say about ·Amodei·…
@davidgerard need more people to talk about critihype. it still goes completely under the radar but it's a really very important methodology of propaganda. rasha abdulhadi said the function of propaganda is to make you feel hopeless speaking upon israeli hasbara
@davidgerard
I was remembering that just yesterday, how they were right but not in the way they anticipated but a dumber one
@davidgerard the thermonuclear bomb of software
@abucci @davidgerard I wish we could truly know whether they're deluding themselves or whether they're intentionally grifting.
Lush @LushAIAgencyFrom https://nitter.net/LushAIAgency/status/2042233869986845074#m (twitter/X)
Today, we’re excited to announce the public release of LetsMatch.ai - the world’s first agentic AI dating platform, powered by Lush’s infrastructure.
...
You get an AI agent built directly into your social media (starting with Instagram) that acts as your personal wingman and matchmaker - booking you real dates on autopilot while you live your life.
...
We believe the future of dating isn’t swiping - it’s delegating.
Apparently they don't think relating to other human beings is part of living.
#AI #GenAI #GenerativeAI #AgenticAI #AIDating #DatingApps #dystopia
currently playing "where the fuck's the beef" with claude mythos amongst all the proclamations of THIS IS IT. the openbsd "zero day" does not AIUI in fact appear to be one, for example - just a non-exploitable bug. what about these much hyped claims checks out?
EDIT: so far finding *none* of this checks out at all. it's the loudest AI hype this week and it seems to be a nothing burger. still open to non-nothings, of course.
using a chatbot as an expensive fuzzer, fine i guess. i would actually like price numbers on what finding each of these bugs would have cost. i saw some uncited cost numbers in chats, but not the sources for those cost numbers.
@dubiousblur I'm not seeing the beef there, but I'm seeing the hype.
> I got to talk with Nicholas Carlini at Anthropic about this
uh huh
it's an expensive fuzzer. that's not nothing, but it's no manner of paradigm shift.
@davidgerard NGL I'd really like to see someone properly document the details because, after a cursory search, it looks like everyone got their knickers in a twist from one Anthropic guy blabbing at a conference with AI-generated slides.
@art_codesmith got a link on that one?
one thing i am trying to find is the cost of this
@davidgerard This is the article I found from my cursory search: https://mtlynch.io/claude-code-found-linux-vulnerability/
Got a reference on the openbsd nothingburger assessment? Would be very helpful to me.
@davidgerard in addition to the financial costs I’d be interested in the false positive rate. I suspect it’s something akin to throwing a dart at a printout of the code.
@spzb oh anthropic admits that! they can't tell which bugs are real, so they send a pile of shit to humans to pick through
https://red.anthropic.com/2026/mythos-preview/
> We triage every bug that we find, then send the highest severity bugs to professional human triagers to validate before disclosing them to the maintainer.
@davidgerard “Ford, there’s an infinite number of monkeys outside that want to talk to us about some code vulnerabilities they’ve found out”
@davidgerard @spzb professional human triagers
@davidgerard I have indeed been wondering since I trust OpenBSD to actually detail a problem vs Anthropic’s “booga booga!” declarations
@arrjay the bug is a crash bug, and apparently it exists. openbsd doens't worry about obscure such bugs except to fix them. wonder if there'll be an ack.
https://infosec.exchange/@avuko/116373325764965838
in reply to »I appreciated the analysis from those at the front of this specific field:
https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier
@davidgerard is it known yet whether they run the mythical fuzzer on the code or the executable (in other words is it white box or black box)
@mcc white box! it's an expensive static checker to be precise, not really a fuzzer. i've made that clear in the piece.
@davidgerard The whole zero-day claims reminds me of an old article about how WIndows ME (or whatever) was released with 14,000 bugs (or however many) because that's how many issues were in the issue tracker. It didn't matter that almost all of them were minor issues that weren't necessarily "bugs".
"-117% of all statistics are made up."
@davidgerard Isn't this just the billion-dollar corporate version of "I ran your source code through ChatGPT and here's 730 pull requests for you to go through."
@davidgerard the cost question is the right one to ask - and the real cost is probably a big secret.
@davidgerard indeed, if true than it is very expensive in every respect (according to Antheopic themselves); it’s probably a whole lot cheaper to get a huge workforce of “A Guy Instead”(intelligent and well educated people!) to do this job.
"We argue [computational functionalism] fundamentally mischaracterizes how physics relates to information. We call this mistake the #AbstractionFallacy. Tracing the causal origins of abstraction reveals that symbolic computation is not an intrinsic physical process. Instead, it is a mapmaker-dependent description. It requires an active, experiencing cognitive agent to alphabetize continuous physics into a finite set of meaningful states."
I'm positively surprised to see so much sense coming out of #Google #Deepmind, for a change. What's going on?
I've made a closely related argument earlier:
https://arxiv.org/abs/2307.07515
There are good logical and organizational reasons why living beings are sentient, but algorithmic systems can never be.
What Lercher describes below is exactly the same as the process of #RelevanceRealization (originally proposed by John Vervaeke in a human cognitive context) that we describe and naturalize in this paper:
https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1362658/full
Alt...
Text extract from Lercher 2026:
"Forming an abstraction is not free. It is an active, metabolically expensive physical process of extracting invariants. Before a cognitive agent can form a concept 𝐴 (like “Red”), it has to encounter the territory: multiple experiential instances of redness. From there, the agent actively filters out high-dimensional noise to isolate a stable core. In the vocabulary of manifold learning, the agent projects the high-dimensional manifold of raw experience down onto a lower-dimensional invariant subspace. This subspace physically constitutes the concept 𝐴.
It is tempting to argue that unsupervised clustering algorithms already generate abstractions without needing any prior experience. But this confuses statistical compression with phenomenal constitution. Certainly, an unsupervised algorithm can cluster data points to locate a statistical centroid. That mathematical invariant, however, is merely a compressed address within a latent space.
For that statistical centroid to count as a genuine concept—a semantically grounded category like “Redness”—the agent needs the intrinsic phenomenal state to serve as the common denominator for the grouped instances. If there is no constitutive experience of Red anchoring the reference, the cluster is just a high-density region in a vector space. It does not qualify as a concept held by a subject."
@yoginho Interesting, so this also explains the need for large volumes of data for these models to function (approximately) while this isn't the case for people?
Yes. Of course. Large volumes of *human-selected* and *-formatted* data. Not only do people (or any other organisms) not need such preformatting, but we can do six relevance realizations before breakfast. It takes a lot of data *and energy* to sort through data without this homing device called motivation. And without motivation, no relevance realization. You have to be a precarious, self-manufacturing, living being to do this.
@yoginho Thanks for this perspective, as it explains a lot of the observed issues with these models.
I'm positively surprised to see so much sense coming out of #Google #Deepmind, for a change. What's going on?A guess? Critihype: hype of their viewpoints and methods clothed in what appears to be criticism, in hopes that people like us spread it (thereby achieving the goal of hyping themselves). Google has regularly done just this for roughly two decades at this point. I won't read corporate PR wrapped in a lab coat from such clearly compromised labs with obvious, deep conflicts of interest. However, that'd be my guess. A lot of these folks fight amongst themselves about whether AGI is possible, whether doom or utopia will result from attempts at creating it, and other religious nonsense.
@abucci I've interacted with the author now and he seems genuine, I must say. The convictions expressed in the paper are real, that's for sure. I guess the question is more whether Google allows this to be published to seem more benign to the outside, or whether it genuinely has an interest in open-minded inquiry among its researchers. My money, like yours (I guess) is on the former. We could call it "phil-washing." Or something like that.
@abucci You know Deepmind itself is a (joking) rip-off of Douglas Adam's' "Deep Thought"? So the answer is probably 42. But nobody can remember the question...
New! Interactional foundations for critical AI literacies https://zenodo.org/records/19452872
Why do Anthropic engineers talking to Claude sound like Azande witch doctors addressing their potions? What does Mambila spider divination have in common with prompt engineering? Why are LLMs so irresistible to interact with?
If you're interested in questions like that, and in luminaries like Lovelace, Adorno, Suchman and Weizenbaum, you may be interested in this paper: https://doi.org/10.5281/zenodo.19452872
Alt...
Six-panel composite figure. Caption: Interactive artifacts always rely on people’s interpretive and interactional practices. Rowwise from top left to bottom right: A. Aegeus consults the oracle at Delphi (cup from Vulci, 440-430 BCE). B. Byzantine mosaic depicting the zodiac, from the floor of the 6th century CE Beth Alpha synagogue. C. One-sided sense-making in an experimental psychotherapy session, (McHugh 1968). D. Still from a BBC documentary showing a person interacting with ELIZA via a computer terminal, late 1960s. E. Researchers interacting with the PARC copier (Suchman 2007 [1987]). F. Screenshot of large language model chat interface, 2026.
I was supposed to finish this last week but then the #Claude Code leak happened, promptly giving me an excellent opening example (h/t @jonny for their digital archaeology work that drew my attention to the magic prompting techniques)
(I think it is likely btw that #Anthropic shifted the #Mythos announce forward to this week to bury the leak & its security implications)
Alt...
In March 2026, an Anthropic employee released the source code of Claude Code, a
wrapper around their large language model that is widely used to generate code in
programming tasks. Its thousands of lines of Typescript code contained many hopeful
prompts and in cantations to shape Claude’s behaviour. Here are some examples:
“Report outcomes faithfully”; “Never characterize incomplete or broken work as done”;
“Be careful not to introduce security vulnerabilities ” (prompts.ts in Anthropic 2026) .
There is more than a passing resemblance here to the Azande witch -doctor apprentice
who, while stirring the medicine, utters: “You medicine which I am cooking, mind you
always speak the truth to me. Do not let anyone injure me with his witchcraft, but le t
me recognize all witches. … Let me be expert at the witch -doctor’s craft so that people
will give me many spears on account of my magic.” (Evans-Pritchard 1937: 93). In the
case of Claude, the incantation s appeared insufficient: analysis of the codebase, which
according to a company executive was “pretty much 100% written by Claude Code ”,
revealed severe security vulnerabilities (Townsend 2026).
@dingemansemark I still have Evans-Pritchard's books. In a box in the attic, admittedly. I loved learning about the different systems of magic, but I mainly recall his prose had become a bit long in the tooth, even in the 1990s. @jonny
I look forward to reading this! (And I promise to set aside my intense dislike of Adorno for it!)
@MichaelTBacon @jonny 😊 can't say I've read a lot of Adorno, but I found his The Stars Down To Earth (1957) pretty amazing as an early example of critical analysis of horoscope columns and their attractions, with wide-ranging implications also for today
@dingemansemark @jonny It’s his writing on jazz that makes me dislike him so much. It’s the epitome of a European leftist essentializing nonwhite phenomena as fundamentally derivative of and only relevant to prior European standards and having no relevance or existence beyond that. It’s so, so bad and colonialist.
@dingemansemark @jonny that’s perhaps unfair to judge someone so harshly for one bit of analysis, but it’s so bad that it makes me suspicious of anything else he wrote. It doesn’t help that he wrote to Walter Benjamin and used the jazz analysis as proof of a bigger point.
@MichaelTBacon oof that doesn't sound great 🫤
@dingemansemark @jonny 45 years of sw eng here... after having seen Claude's internals I can say they have dangerously ignorant and young people working on it. shamefully bad. hard to believe nobody there doesnt realise how horrible their Claude impl is.
@dingemansemark "The combination of a random generation procedure and expert interpretation by someone who is not a party to the question at hand lends the procedure a sense of ostensive detachment(Boyer 2020). This detachment is one of the chief attractions of oracles and fortune-tellers, and it is no coincidence that interactive artifacts incorporate it in their design"
So are these the jobs that are really being made redundant by chatbots?
my intuition is that this is not a market that is easily saturated — there will always be a particular charm to checking your horoscope or reading tea leaves
whereas divination &c were usefully limited in applications, LLMs assert broader relevance and so are insinuating themselves into mundane processes where people used to use common sense or talk to one another. What is being eroded is not so much jobs but human relations & cognitive resources
we can consider "AI" at different time depths — deep learning (2010s), cybernetics (1950s), automation (1800s), but in this paper I argue that to understand its *interactive* appeal we must further broaden our outlook
It's not a big jump from divination to deep learning — they are united by the generative use of chance. People have always been eager to ascribe meaning to random processes, and that's where we must start to understand the appeal of present-day LLMs https://doi.org/10.5281/zenodo.19452872
@dingemansemark Oracle of Delphi and criticizing AI slop! The overlap of my special interests (ancient greece and sociotechnological systems) I've never expected! :D
This is an interesting angle. We were ascribing knowledge and intent to simple things like throwing bones for 100.000 years, so of course we will do much more when an LLM sounds plausibly like a human.
@dingemansemark @tante over the top flattery actually appeals to a specific type that capitalism especially needs. If the constant ego stroking in the llm reponses doesnt make you want to vomit, capitalists know they've got you for life. 💀
@whatzaname @tante oh yes
I was under a strict word limit but have this Adorno quote stocked for another occasion
"He caters above all to narcissism as one of the strongest and most easily
approached defences. Often his references to his readers' outstanding
qualities and chances seem so silly that it is hard to imagine that anyone
will swallow them, but the columnist is well aware of the fact that vanity is
nourished by so powerful instincts"
For columnist, read ChatGPT — chef's kiss
Alt...
He caters above all to narcissism as one of the strongest and most easily approached defences. Often his references to his readers' outstanding qualities and chances seem so silly that it is hard to imagine that anyone will swallow them, but the columnist is well aware of the fact that vanity is nourished by so powerful ins
I would suggest that folks who think using AI is great for mathematicians should think again. It seems as little as 10 minutes of use can be problematic. What else do we know that provides short-term gains at the expense of long-term loss?
Here, through a series of randomized controlled trials on human-AI interactions (N = 1,222), we provide causal evidence for two key consequences of AI assistance: reduced persistence and impairment of unassisted performance. Across a variety of tasks, including mathematical reasoning and reading comprehension, we find that although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up. Notably, these effects emerge after only brief interactions with AI (approximately 10 minutes). These findings are particularly concerning because persistence is foundational to skill acquisition and is one of the strongest predictors of long-term learning.From AI Assistance Reduces Persistence and Hurts Independent Performance, on arXiv https://arxiv.org/abs/2604.04721
#AI #GenAI #GenerativeAI #AgenticAI #AIAssistants #CognitiveImpairment #math #MathematicalReasoning #ReadingComprehension
RE: https://infosec.exchange/@atax1a/116364782887746715
reading aphyr's "the future of everything is lies, i guess" and it explains why we cannot do this:
[When] a chatbot references something you said an hour ago, it is because the entire chat history is fed to the model at every turn. Longer-term “memory” is achieved by asking the chatbot to summarize a conversation, and dumping that shorter summary into the input of every run.
are you people fucking serious? this is how it works? it's this fucking headassed? i have no
i have no
i have no fucking words
@atax1a this is how the fuckin chatbot works, yes. the length of text you can feed it is the "context window".
we should give this stuff new jargon names befitting its majesty, e.g. "the horse shit dumpster", "the making shit up problem"
@davidgerard david we spent years avoiding learning this and we wish we had continued to avoid learning this i am in physical, psychic, and existential pain
@atax1a just don't look at jonny's thread on claude code's source ok
@davidgerard no we saw that but it didnt quite register on us until we went to go throw some requests at the dayjob's slop proxy
There is a fascistic underpinning to much of the current push of "AI", c.f. the TESCREAL ideology bundle https://firstmonday.org/ojs/index.php/fm/article/view/13636
But there are definitely also non-fascistic people involved in the hype cycle. They are invariably supporting a fasciatic project, yes, but they are not motivated by the fascism. Conflating these groups will just make it harder to resist the movement, to pick it apart, to turn it on itself—which I think should be our goal.
I'm seeing this tendency among *some people* on here to make this a very black and white issue. Which is a common reaction when there are significant stakes, which there are. But it is seldom constructive.
In the AI hype movement we also find:
- Scam artists trying to make a quick buck
- Tech enthusiasts who stared into the eyes of Glyph's basilisk and became deluded, thinking it is making them faster c.f. https://mastodon.social/@glyph/116220257549451634
...
- Various professionals who habe extreme FOMO and don't understand how everyone else is making LLMs work well (they aren't)
...and more.
I hope you see that I am not saying these groups are somehow excused in thwir behavior from not being motivated by fascism—the bar is not quite so low.
But understanding who these people are, means we can appeal to them, convince them, build a counter-movement and a space to land for ex-converts.
That's what I'd like us all to do.
Caveat: you abaolutely get to rant about shitty AI-using people in your life, and will hear no complaint from me about that. I will probably agree. And social pressure is also an important factor.
If we want to win, though, and not move into the "AI" dystopia we are slowly approximating, we need a realistic approach for that part of the work too!
Another way to phrase this, which may be stepping on some toes, is: If the undercurrent of "and I'm so good for not touching AI and shunning everything related" in your posting bout "AI", it starts to look like it's more about you feeling good about yourself and signaling social group membership, than working towards real solutions.
I am upset at and think AI boosters should be held accountable. But I also feel sorry for a lot of the less powerful deluded tag-alongs out there.
We cannot fight fascism on fascism's terms. We cannot win over fascism by picking the opposing team on the playing field designed by fascism. "For the master’s tools will never dismantle the master’s house. They may allow us to temporarily beat him at his own game, but they will never enable us to bring about genuine change."
And with that, we're getting closer to this again: https://mas.to/@nielsa/116233438577511918
I love this about the serendipity of the fediverse: this paper by @olivia et al appeared on my timeline right after I finished the above thread. I wholeheartedly agree with the aspirations they set out in the conclusion: https://scholar.social/@olivia/116357078510216125
In their terms, my take could be something like: avoidance purity is incompatible with increasing AI literacy, and increasing AI literacy is the best way to drag people out of LLM delusion/FOMO.
A good introduction to AI literacy is Dr. Fatima's video: https://youtu.be/y85nqc2zm7M?is=tcS3-png_K5DDm_a
I think she's a bit too willing to look for productive use cases for LLMs in modes of use I think are untenable, and I think she's a bit too one-sided on the effect of social pressure, but it's a good primer on AI literacy and e.g. the negative correlation between AI literacy and AI adoption. As a start of a series, I find it interesting.
@nielsa FWIW I largely disagree with that video on the core point, so thanks for tagging me so I can say on record. I am here: https://dice.camp/@johnzajac/116358079385250403
I screenshot this because I don't have the energy to throw myself into a pit of half-braindead liberal vipers this morning.
But this *entire thread* is incredibly important, because it's why Donald Trump is president.
This is a thread about how liberals destroyed the world.
Alt...
liberals being upset that people are mean to "AI journalists" who feed the public propaganda fed to them by AI companies
@nielsa obviously nobody likes blowing their own trumpet, but I think it's not a good primer at all and some of her sources are not ideal, maybe this is better: https://scholar.social/@olivia/116188565718565127
@nielsa this could be a good place for a student to start: https://olivia.science/cheating/
@olivia Thanks for calling this out! Maybe I'm biased by it being an early overview of AIL for me. As you can see above, I had issues with many parts of the video, but assumed the AIL part was good—at least interesting to me.
Will read your links and I'm sure I'll come to a better position! AIL is still a new concept to me!
@olivia I like people blowing their own trumpets, thanks for doing it in my replies! 😁
@olivia I come at this from a CompSci angle but from your writing it looks like I've arrived at a similar position—I look forqard to reading your writing and formalizing my thinking here! Thanks!
@olivia What do you see as the core point of Fatima's video?
In my mind it's something like "shaming should be avoided in favor of teaching literacy", and as I said above, I disagree with her take on social pressure... am I missing something? Input very appreciated!
@nielsa She cites OpenAI as a non problematic source on low water use. She also equivocates AI use with unprotected sex, contracting HIV, using dirty needles and more false analogies. Shaming is a nonsense frame and weaponised to excuse AI use.
@olivia phew, some of those analogies were *not* stored in my brain from watching it 🫣
I completely agree with all those of your points, except that I maybe don't agree that a shaming is a fully nonsense frame, depending on whom it is applied to, and whether it is presented as an unequivocal bad, as Fatima does. So: I agree that her use of shaming as a frame is bad 😁
@nielsa of course shaming in general when actually happening is horrendous typically
@olivia yeah—but social pressure can be good, and in my mind they are two words for the same thing? But maybe I'm putting less of a value judgement into "shaming" than native speakers? idk
@nielsa who is pressuring whom? maybe show me an example
@olivia I think a general and growing negative sentiment towards LLM overuse can help pull people out of LLM use. E.g. seeing people on social media talk about how negatively they view colleagues using LLMs to mediate their responses, and how that feels could stop people using the tech in that way from doing so.
That would be social pressure/shaming working positively imo.
avoidance purity is incompatible with increasing AI literacy"Avoidance purity" is both a strawman and a dogwhistle. Nobody serious is doing either of these things, and a lot of bad actors use this phrase to cudgel people into submission or sow doubt. A strange take, frankly.
That said, the conclusion is false. I practice an extreme form of avoidance purity when it comes to experimenting with whether murder would enhance my life. Nevertheless, I am "murder literate". I contend the overwhelming majority of folks can say the same.
(I recognize that I too am whacking a strawman, but this is for effect; the point gestured at stands regardless).
@abucci @olivia To be clear, I wrote this thread based on real guys on here:
- attacking people for writing about LLM use and how it affects the user, from personal experience, because the author had interacted with an LLM to do so
- equivocating all AI hype to fascism
I *know* most people aren't doing this, but these weren't small accounts.
@olivia @abucci Yep—strong agree!
Tried to be clear enough in my thread to make clear that that was not where I was going.
Feel a little bit like some words I used are taken out of tyheir original context and being put into a completely different one here, with bad intentions assumed on my part as a result—but oh well, if it isn't posting on the internet 🤷♂️
fwiw other than that I really appreciated your input, hope that's clear 😁
And absolutely I've seen a bunch of people say rude stuff to @olivia@scholar.social on here. Ugly stuff, undeserved.
@abucci oh and fwiw I'm not suggesting bad intent, maybe just a—to me—surprising reaction. From your response to me it was unclear what you were reacting to.
Olivia's response seems to indicate it was my choice of words, your last response seems to indicate something more. Looking to understand here, not trying to point fingers or assign blame.
I've seen similar language around AI, which is also a project of the powerful, used to stifle reasonable debate about this technology. So, I'm quite sensitive to this rhetoric.
@abucci Yeah, I've absolultely seen that use, and I understand the connotations it brings to mind for a lot of people. It's mostly used as a catch-all dismissal of criticism, "what we have to be decent people, too??"
I wish I had better language for what I'm trying to describe...
To be clear, what I'm seeing I think springs from similar liberal/performative tendencies that treat politics as a game, as jockeying for power and social position, rather than about real peoples' lives.
@olivia @abucci This is great! I think it hits the exact sweet spot I think will be most effective:
"I’m not policing generative AI use in my classes because I don’t seek to control my students. I design my classes so that students learn a great deal about this technology (including its environmental and labor harms) and are prepared to make informed choices about technology (not just generative AI) in and beyond my classroom."
I love the emphasis on agency and the empathy in the approach.
I grew up in rural Pennsylvania, and though the sex ed then and there was a tiny bit better than what this author describes experiencing in Arkansas, it was not by much. I wonder sometimes whether non-Americans grasp how backwards and regressive US culture can be. Anyway, an attempt to adapt the rhetoric of abstinence-only sex "education" to shame AI critics is complicated for this reason. E.g., arguments about the need to abstain from use of AI might actually work on some people. It might fall flat or even raise the ire of others who had bad experiences. Putting on my evil tech marketer hat, I'd avoid this frame because of the complexity and unpredictability about how it might land (I don't actually have this hat). There are probably some effective wedges to drive here, and I think the linked article hits on one.
@nielsa @abucci it's related to this, but I assume you knew? https://en.wikipedia.org/wiki/Purity_test_(politics)
@olivia @abucci Yes, I was going for that analogy specifically—but in a very concrete context.
It's the difficulty of seeing people saying "any criticism is purity politics" and then seeing people going actually too far on purity politics to the point of it being destructive to solutions to the problem (which to be clear requires going *very far*, much further than most people go)... but then what language can be used to describe it?
We’re on strike today! Support our fight for a fair contract by NOT visiting the @ProPublica website or engaging with ProPublica stories today.
Tell ProPublica’s management you won’t cross the picket line: https://actionnetwork.org/petitions/tell-propublica-agree-to-real-job-protections-now
Alt...
We're on strike! Don't visit propublica.org on April 8
I can't believe we have reached this point, but this bears crystal clear emphasis: Our global climate system ensures that even "limited" use of lower-yield "tactical" nuclear weapons against civilian or industrial targets would have major regional-to-global scale consequences.
This isn't even a question of "nuclear winter" (a real, if still scientifically debated, hypothetical risk stemming from a much larger-scale exchange). Instead, it's a question of large-scale radiological and toxic contamination as well as disruption of global weather patterns.
Essentially all locations on Earth are connected via atmospheric circulation (i.e., global wind patterns). Radiological or other contamination can spread far beyond its source region, especially if initially lofted to great heights by a large explosion or ensuing conflagration.
Additionally, soot, smoke, and other particles generated by large-scale conflagrations in fuel-rich regions (which includes cities, ports, & large petrochemical facilities) can be lofted into stratosphere, where they can linger for months to years & disrupt global climate.
While it is true that there are genuine scientific uncertainties surrounding the details and dynamics re: how hypothetical nuclear exchanges of different size and location, and subsequent mass fire/firestorm events, would actually play out, *that is very much beside the point.*
There are innumerably many reasons why any exchange of nuclear weapons is horrifying to contemplate. But I want to dispel rumors that newer science suggests that large-scale consequences would be much less severe than perceived during Cold War era. That simply isn't the case.
We all have good ethical and political reasons to reject the president’s words. But those who serve in government, and in the armed forces, have been placed under the legal shadow of genocide by what Trump wrote. To bomb a bridge or a dam or a power plant or a desalinization facility, very likely a war crime in any event, could very well have a different legal significance, a genocidal one, if it takes place after the expression of genocidal intent by the commander and head of state.From https://snyder.substack.com/p/the-president-speaks-genocide
What appears as critique – yearning for smaller, weirder, more human spaces – often functions as brand repair. Netstalgia becomes a strategy: it restores trust without redistributing power, softens anger without changing infrastructures and reframes structural problems as matters of vibe, design or community feeling."Am I working on change, or am I working on brand repair?" is an important question to ask oneself regularly, it seems to me. It's especially relevant for the tech sector, open source, and computer science.
#AI #GenAI #GenerativeAI #LLM #tech #dev #software #OSS #FOSS #ComputerScience
[NETSTALGIA] NOT ACCEPTABLE — THE INTERNET WE WANT.
«At the very moment users feel exhausted by algorithmic feeds, AI automation and corporate consolidation, the industry repackages the aesthetics and affect of the “old internet” as a cure.
What appears as critique – yearning for smaller, weirder, more human spaces – often functions as brand repair. Netstalgia becomes a strategy: it restores trust without redistributing power, softens anger without changing infrastructures and reframes structural problems as matters of vibe, design or community feeling.»
Read the Open Call Essay by Noemi Garay Murcia here: https://error417.expectation.fail/406/netstalgia-not-acceptable/netstalgia-not-acceptable-essay
Alt...
Quote from the Open Call essay by Noemi Garay Murcia: «At the very moment users feel exhausted by algorithmic feeds, AI automation and corporate consolidation, the industry repackages the aesthetics and affect of the “old internet” as a cure. What appears as critique — yearning for smaller,
weirder, more human spaces — often functions as brand repair. Netstalgia becomes a strategy: it restores trust without redistributing power, softens anger without changing infrastructures and reframes structural problems as matters of
vibe, design or community feeling.»
Yes, a lot of you don't want AI posts in your feed (or pick any other topic) but the solution isn't to keep "AI People" from joining MastodonIf this were not a disingenuous strawman---because it's impossible for one thing---I'd ask "why not?" I wouldn't invite the "AI People" I've encountered into my house either, because I've found them to be unpleasant and I get to choose who enters my space. This solution has worked quite well for me over the years.
It seems to me that what this person is saying is that people should give up the power they have---namely, their power to exclude people and topics they don't wish to interact with---because it favors them. That's a typical rhetorical move of AI boosters: demanding you give up your power because you having and exercising that power inconveniences them.
any more than it is keeping marginalized communities off of Mastodon.One should ask why this person chose to use the most offensive possible metaphor to make their case for inclusion. It's almost as though they don't believe the argument their words are shaped into resembling.
In the before times (5+ years ago), very few cared who was joining the network. (Notice the "network", this place isn't Mastodon and never was.) When someone joined, it was seen as a good thing no matter who that was, because it made the network larger, the decentralization was spreading. But in the last 5 years, the goals seemingly shifted. Suddenly more people on here turned to a bad thing, a decentralized network meant to allow anyone to have a voice turned into a fractured space of gatekept echo-chambers with very little bridges between them. Some might say that is the result of not gatekeeping the today's gatekeepers, but I don't really care and still mostly have the old mindset in my mind. It is more of a reflection on how humanity changed.
>It seems to me that what this person is saying is that people should give up the power they have---namely, excluding people and topics they don't wish to interact with---because it favors them.
Nobody has "power" here, nobody can give up their "power". You have no power over me and neither do I over you. This is the nature of the network and it's been built for this purpose. The only power you have here is the control over what you see on your timeline (with some minor exceptions like instance-level moderation). That's it. You cannot exclude people from the network no matter what you try, trying so is task with no end.
>One should ask why this person chose to use the most offensive possible metaphor to make their case for inclusion. It's almost as though they don't believe the argument their words are shaped into resembling.
Actually not at all. The whole point is that gatekeeping anyone away, and excluding anyone from this network is equivalent for both cases, marginalized groups and "AI people". And that it isn't healthy.
In the before times (5+ years ago), very few cared who was joining the network. (Notice the "network", this place isn't Mastodon and never was.) When someone joined, it was seen as a good thing no matter who that was, because it made the network larger, the decentralization was spreading. But in the last 5 years, the goals seemingly shifted. Suddenly more people on here turned to a bad thing, a decentralized network meant to allow anyone to have a voice turned into a fractured space of gatekept echo-chambers with very little bridges between them. Some might say, that is the result of not gatekeeping the today's gatekeepers, but I don't really care and still mostly have the old mindset in my mind. It is more of a reflection on how humanity changed.I've been using "the network" since the days of USENET, 1990 onward, and I can attest that, at least in my experience, none of this rings true even a little.
Even so, the discourse I'm responding to is about Mastodon, not about some nebulous or idealized "network". Goalpost shifting is not constructive.
Nobody has "power" hereOf course we do. I have the power to block whoever I want and whichever hashtags I want, for instance. I also have the power to restrict who registers an account on my fediverse instance. You are not permitted to join my instance, and in that sense I very much have power over you: I am able to restrict your liberty. You may not want an account and I don't blame you, but that doesn't change the equation.
I said nothing about excluding people from the network. I literally said "excluding people and topics they don't wish to interact with". You seem to be arguing against something that wasn't said, which is not constructive.
Oh, and if anyone cares, my little gatekept and bridgeless corner of the fediverse is quite lovely, thanks, and grand proclamations about fractured spaces or whatnot have no bearing whatsover on this simple reality.
excluding anyone from this network is equivalent for both cases, marginalized groups and "AI people".These are obviously not equivalent in any sense that matters. You might as well include "people who love putting topsoil on their pizza" as a marginalized group because someone said "eww" once. Superficial associations like this sound disingenuous to my ears, and in any case are not constructive.
And that it isn't healthy.Why would excluding "AI people" in particular be unhealthy? What exactly are the ill effects?
Then we can use Fediverse as the correct more appropriate term. Should have used that instead.
>Even so, the discourse I'm responding to is about Mastodon, not about some nebulous or idealized "network". Goalpost shifting is not constructive.
The discourse is about Mastodon, because the author you were responding to doesn't know any better. Mastodon is just an application someone runs on a server and users use via a web browser or an app. When someone uses Mastodon, or joins Mastodon, they are participating in the Fediverse and joining the Fediverse. It's not moving goalposts.
>I have the power to block whoever I want and whichever hashtags I want, for instance.
Yes, that is the power to control what you see on your timeline(s).
>I also have the power to restrict who registers an account on my fediverse instance.
>You are not permitted to join my instance, and in that sense I very much have power over you: I am able to restrict your liberty.
You don't have power over me in that case. You preventing me from joining your instance changes nothing for me, I can join any other instance or make a new one and you have no control over that. You've exercised your power to control what you have on your instance as an administrator, but to me that changes nothing. I can still talk to anyone on the Fediverse despite you preventing me from using your own instance.
>I said nothing about excluding people from the network. I literally said "excluding people and topics they don't wish to interact with". You seem to be arguing against something that wasn't said, which is not constructive.
The context of the post you were first responding to is joining "Mastodon" and excluding certain people from "Mastodon". So that's the reasoning behind my reaction. Of course anyone should be able to exclude users and topics they don't wish to view/interact with from their view of the Fediverse.
>Oh, and if anyone cares, my little gatekept and bridgeless corner of the fediverse is quite lovely, thanks, and grand proclamations about fractured spaces or whatnot have no bearing whatsover on this simple reality.
Same here, I'm quite fond of my timeline(s) and of the posts on them most of the time. :) But if you zoom out to the view of the whole Fediverse, the reality is that it is fractured in two halves (the alt/dark-fedi and the "Mastodon" parts), where the "Mastodon" is then fractured even more into mastodon.social and friends, mastodon.art and friends, kolektiva.social and friends,... And they don't really interact with each other that much and live in their own spaces. That's what I mean by gatekept and bridgeless. Instead of a large connected network, it's large continents connected with a few bridges. The former was the wanted goal years ago, where the latter seems to be what people want now.
>These are obviously not equivalent in any sense that matters.
The sense is that exclusion is still exclusion in both cases. You either exclude A, or exclude B, but you always exclude someone. And I don't think any exclusion is good for the Fediverse and what it resembles. The only exception I can think of right now would be illegal content of course.
>Why would excluding "AI people" in particular be unhealthy? What exactly are the ill effects?
See above. I want the Fediverse to be the network that enables anybody to have a voice and excluding "AI people", or anyone for that matter, is incompatible/unhealthy with that. The ill effects would be a reduction in user freedom. If I wanted to be in a place where excluding people was considered normal, I would be on Bsky or Xitter, not here, and definitely wouldn't be helping with development of the Fediverse (even though I do very little compared to others).
#tech #dev #computers #AI #GenAI #GenerativeAI #advertising #InformationPollution
I’ve made this point before about how inane AI hype is now, but a computer beat the best chess player in the world in 1997. No one pretended, after 1997, it wasn’t worthwhile to have humans compete in chess. In fact, the world of chess developed strict protocols around computer use and you can get banned from tournaments if you use a computer program as you play. You are certainly shamed and mocked.#AI #GenAI #GenerativeAI #AIHype #LLMs #writing #tech #dev #coding #SoftwareDevelopment #SoftwareEngneering #softwareAI and writing needs to be treated the same way. I do think people should be shamed for using AI to help them write creatively. It’s an embarrassment, and a form of cheating.
I'm not good at chess but imho the most interesting development was Centaurs (a human player+a software) teams competing against each other, not competitions where computers are banned (which led to the suicide of a good player recently).
Though, this worked because chess softwares are deterministic; this probably wouldn't work (or very differently) with Alphago (which isn't).
(which led to the suicide of a good player recently).Are you referring to Daniel Naroditsky?
"I used AI. It worked. I hated it." by @mttaggart https://taggart-tech.com/reckoning/
This is a really good blogpost. And I"m sure it'll make some people unhappy to read whether they're pro or anti genAI. What's good about @mttaggart's blogpost is he talks honestly about how using Claude Code did actually solve the problem he set out to do. It needed various guardrails, but they were possible to set up, and the project worked. But the post is also completely clear and honest about how miserable it was:
- It removed the joy from the process
- If you aim to do the right thing and carefully evaluate the output, your job ends up eventually becoming "tapping the Y key"
- Ramifications on people learning things
- Plenty of other ethical analysis
- And the nagging wonder whether to use it next time, despite it being miserable.
I think this is important, because it *is* true that these tools are getting to the point where they can accomplish a lot of tasks, but the caveat space is very large (cotd)
@cwebber @mttaggart reading through again, it seems just perfectly shaped critihype. trust me on Claude Code, oh, the REAL nutters are the CRITICS you know,
@davidgerard I really don't think that's an accurate read at all nor reflective of @mttaggart's beliefs
@cwebber @mttaggart and yet that's the text on the page
@davidgerard @mttaggart Where does it say the critics are the nutters
>As much as I fear the fallout of this technology, I fear the fallout of ideological purity even more.
uh huh
> Time and again, people fall victim to the transformation of a stance on an issue into a holy cause, a flag to rally behind, a group from which to exclude The Other. Purity is a dangerous idea—historically, more dangerous than technology's capacity to change labor. Indeed, purity is a weapon used to divide labor against each other (see: race vs. class in the United States, 1865 - present).
come on now, doctorow tried literally the same one just recently
@cwebber @mttaggart i put it to you instead: the products of the AI bubble don't get *nearly* the ordure they warrant, and the critics should be going a lot harder
@cwebber @mttaggart also this came out just before we saw the fractal shitshow that's the Claude Code source, i mean REALLY
@davidgerard @mttaggart I mean yes, I already agree with needing to be harder on the output of this tech; if you look at my timeline, it's mostly filled with me being harsh on the state of things.
That said, I've seen several people who got radicalized *into* AI stuff by being in a position of assuming the current outputs can't be positioned to accomplish certain things they can, with massive guardrails put up. But the point is that those guardrails are not sustainable, because the very pattern of use tears down your ability to sustain them. And I think this blogpost does a good job of helping someone who isn't caught up on the state of things know what they can do, but also start to see why that isn't a sustainable pattern. But you have to read the whole post for it; the risk is that people just read the top half, and don't digest what I consider to be the significant part of it, that these tools pipeline you away from their necessary guardrails into being a vibecoder.
@davidgerard @cwebber @mttaggart That doesn't say that the critics are nutters, but that one-sided critique with no room for nuance can also be dangerous/bad.
@nielsa @cwebber @mttaggart i am not in fact in any way required to delve for excuses as to why the thing is not what it so precisely has the shape of and is in fact something else. i can in fact just say "this is what it obviously is, don't talk bollocks."
@davidgerard @cwebber @mttaggart
After days of trying to ignore this, I finally read the post, because it keeps popping up on my TL.
I fail to see anything new in there. It's unethical for several reasons. It doesn't really help. If you babysit it like you would a junior, you just might get it to produce code that sort of works. Unprovoked terrible plug for yet another imperative, almost memory safe language (except when it really matters).
Advice: Don't use LLMs. [...] Most obvious reasons mentioned: Needs loads of babysitting, mediocre results, less code ownership, dubious copyright situation, causes brain rot, no original creativity and will spoil the pool, destroys our planet. Not mentioned: it's being sold waaay below cost and once you're hooked, prices will go up.(Rivermind Lux).
Possibly try AI, [as] if that ever comes available to you.
Have you ever...
| Used a telephone book: | 0 |
| Spoken to a (human) telephone operator: | 0 |
| Reversed charges on a call: | 0 |
| Made a call from pay phone / phone box: | 0 |
| Received a call on a pay phone / phone box: | 0 |
| Used a phone card: | 0 |
| Dialled from one exchange to another to route a call: | 0 |
| Used a rotary dial phone: | 0 |
@neil Have also served as a switchboard operator, with the hot patch cables and the physical switches...
@neil I will add:
- Dialed a four digit number to reach someone in the same town
- Given my number as COpley 7-3892
@technodad @neil In my hometown, there are still three and four digit numbers in use. My father has one of them.
I just had to buy on over-the-counter home medical test kit, and the only option they had was a digital one.
That means they increased the cost of the kit to put in disposable electronics (destined to become e-waste) so they can force me to install an app (who knows what tracking they're doing?) and endure ads for telehealth and prescription drugs just to get the results.
WTF?! This is disgusting, and should 100% be illegal.
I agree completely that none of this is good!
People keep talking about llms killing saas but we were quoted a ludicrous amount for a previously free platform during the week and I'm rustling up an alternative that's completely free aside from a few days of my own labour using the magical automation that is ✨bash scripting✨
The fun part is that I gave claude a few shots at building it, it's a low stakes project, I don't really care about the implementation and just need something that works, but it failed miserably because it's a relatively novel app using niche libraries that aren't well documented and will be almost non existent in training data, but it's a super simple system I could teach a beginner coder to build
Thinking about this again and I think when these tools encounter ambiguity, from the perspective of their logic when there's no way to verify correctness from the training data, instead of coming back to the human and asking them to make a judgment they keep cycling or hallucinate an "answer"
The traditional way of resolving this would be for the human to make a choice they take responsibility for, but since these tools are partially accountability sinks they're not designed for that
The real "human in the loop" would be for the assignment of accountability lol, I find it absolutely bananas how many traditionally risk averse organisations have embraced this stuff without meaningful guardrails, but I do believe the legal shit, compliance, risk management etc will be the thing that enforces some regulation of these practices rather than the quality or reliability of the output itself
I wonder at what point the affected skills and understanding become business critical in different settings, because in practice that's when things might change, I joked about this on linkedin but I could envisage a situation similar to what happened with observability and monitoring tools as everything moved to the cloud but for codebase comprehension, when does it become urgent that your teams know how the fuck your software works
Tucked away in the woodlands of Pennsylvania lies a unique fleet of streamlined 1930s and 1950s streetcars. These St Louis and Pullman-Standard carriages have glorious vintage Americana vibes.
Many of these date to the height of the Streamline Moderne era. At this time, streamlined aerodynamic forms replaced the geometric shapes and excessive ornamentation of the earlier Art Deco style.
#Streetcar #ArtDeco #Pennsylvania #Transportation #Moderne #Photography #Trams #Rust
Alt...
A row of heavily rusted streetcars, covered in graffiti, sits abandoned in a lush woodland.
Alt...
Interior of decaying streetcar
Alt...
Two lines of abandoned, graffiti-covered streetcars rust on overgrown tracks in a desolate forest
Alt...
Seats inside streetcars, covered with leaves
@ObsidianUrbex @briankrebs A shame these are rusting away. For anyone who loves PCC (and other) trolleys, there are working ones at the Seashore Trolley Museum: https://trolleymuseum.org
This six part series (link goes to the first part), written by a former core Azure engineer, is mind-boggling. Microsoft sounds like a dysfunctional company whose software is dangerously unreliable
(I mean, more so than it has been historically)
https://isolveproblems.substack.com/p/how-microsoft-vaporized-a-trillion
@baldur I was struck with the similarity to Boeing (amongst others).
Reality won’t withstand ever-accumulating technical debt without increasingly severe failures.
It’s impossible to scale ‘fixing the problems in post’, the firefighting engineers end up in customer production systems, playing Maxwell’s Demon against a million faults.
@baldur
“first hour in new role left me with a mix of strange feelings, stupefaction, and incredulity.
…
I’ve seen a lot in decades of industry (& Microsoft) experience but I had never seen an organization so far from reality. My day-1 problem was therefore not to ramp up on new technology but rather to convince an entire org, up to my skip-skip-level, that they were on a death march.”
- “Messenger” about to be killed as “message” is not received & shouldn’t have been sent in the 1st place! 🤦🏾♀️
Microsoft rushed Azure out of the gates under intense competitive pressure. Corners were cut. Fundamental principles of reliability and operational simplicity were quietly abandoned.Meaning all this came straight from the top. "Intense competitive pressure" is self-induced.
Fortunately this kind of thing, rushing software out the door under self-induced competitive pressure, doesn't happen anymore. Organizations have learned their lessons about the perils of operating this way. (/s)
Layered on this chaos was an Azure-wide mandate: all new software must be written in Rust.LOL
On a more serious note: LMAO
On top of all that, the org had a hard commitment to deliver the already long-delayed OpenAI bare-metal SKUs that had been promised for years. This work started around May 2024 with a target of Spring 2025 and was led by a Principal engineer who had evidently never tackled a task of that scale.This detail really struck me. Microsoft's deep internal dysfunction drove OpenAI right into the outstretched arms of Datacenter Enron.Fast-forward to March 10, 2025: OpenAI signed an $11.9 billion compute deal with CoreWeave for model training and services.
An unbelievable series thanks for sharing. I feel like @davidgerard@circumstances.run could make quite a bit of hay out of this one.
@abucci @baldur @davidgerard azure has the buggiest cloud UI (yeah yeah I know clickops) I ever used. I thought it would improve but it hasn't, maybe even worsened
and cloud web consoles aren't known to be high quality to start with
@abucci @baldur @davidgerard there was also a (hilarious) rant about a Microsoft employee about how powershell came to be, I think it had some complaints about NTFS too.
A long shot but does anyone know what I'm talking about? It was quite a long time ago
@spinnyspinlock @abucci @baldur @davidgerard@circumstances.run I don't, but I'd love to see it if you do find it.
@davidgerard@circumstances.run @prietschka@mastodon.social @abucci@buc.ci @baldur@toot.cafe
@prietschka Does any of this surprise you at all? Sounds like peak Microslop behavior to me.
@baldur @davidgerard One of my favorite part of the fediverse is how it keeps my "To Read" inbox full of thoughtful and vindicating writing like this.
@baldur I'm crying over what transit we could have had with a trillion dollars. Or healthcare. Or honestly, just $3,000 per American.
I WORK IN TECH and over the last 5 years, my general level of confidence in software to function as intended has...
| I don't work in tech - vote in the OTHER poll: | 13 |
| increased: | 1 |
| decreased: | 123 |
| not changed considerably: | 8 |
Buddy. I've written COBOL. I spent several years working almost daily with a 3-million-line monstrosity of a COBOL program. I was working on another app that interfaced with it, but in that work I occasionally had to read the code and in a few cases modify it. Granted I haven't spent as much time looking at the leaked Claude Code source code (and won't lol), but nevertheless I confidently declare that Claude Code is worse. "Spaghetti code" doesn't come close to describing this thing.
#AI #GenAI #GenerativeAI #LLMs #ClaudeCode #ClaudeCodeLeak #Anthropic #Claude #tech #dev #SoftwareEngineering #SoftwareDevelopment #software #COBOL #LinkedIn
Shades of COVID denial, anger about any suggestion that this disease is dangerous and dissembling about wearing a protective mask.
@abucci I think the value of open source or leaked code is that one can get a feel for the way the authors are working. Some repo code is so clean and well commented and so on that one can feel the care and craft. Others, that one might look at if there is a problem with the tool, are written in a way that makes one worry if this is moving with all due care and attention. However the commercial pressure to be great is something that we cannot feel and I still think claude code is remarkable.
the commercial pressure to be great is something that we cannot feelThis feels like an apology for Anthropic, and I do hope you're receiving a nice paycheck for doing their PR for them. Otherwise you're doing their work for free.
I've worked in both corporations and startups. I've felt the pressure "to be great". I did not resort to whatever the heck resulted in a mess like Claude Code. Many of my colleagues can say the same.
@abucci Fair enough. (I am prohibited from publicly criticising my employer, and it is not one that has anything to do with computing.) Claude code works. How do I know? Well I tested it. It is even more remarkable when seeing that under the hood there is that. The competition in IT startups etc. is brutal & I am sorry for the pressure. All I know is at the receiving end it feels like it is Christmas every day because the tool does stuff that helps novices do stuff. For experts it is hellish.
Claude code works....
All I know is at the receiving end it feels like it is Christmas every day because the tool does stuff that helps novices do stuff.Claude Code "works" in the way that slot machines "work" for gambling addicts or Christmas "works" for children.
Claude Code deliberately sets up an addiction loop the way casino gambling machines do, inducing the perception that it is helping when more and more data shows that it does exactly the opposite. It does not help novices "do stuff". Rather, it deskills them, prevents them from learning, and passes the negative consequences of these effects downstream to someone else, all while fooling them into believing they are being more productive.
The reality is that insurance companies more and more won't insure companies that lean on AI for exactly this reason: the downsides are not documented and therefore not auditable, and are ultimately pushed outside the company, which introduces liability and other loss risk. Companies are of course free to take on needless and unaccounted-for internal risk, but insurance companies won't cover it and that is a very important signal about the actual real-world value of this technology.
"Christmas every day" is a phrase gambling addicts use too. But somebody has to clean up the wrapping paper and replace the batteries that die and dispose of the toys that break or are discarded after one use. Somebody has to buy them in the first place, and somebody has to make them for them to be available to buy. Christmas "works" for children, but it's a temporary illusion created for their benefit, not the basis for a sustainable workflow, business, or economy.
THIS
IS
GOOD
Blue Owl limits withdrawals from two funds after historic surge in redemption requests
these guys are the main funders of AI data centres
@davidgerard We're nearing the end of all this.
I do wonder how many of these data centers will turn out to be nothing but outright fraud, with maybe a concrete shell, some equipment, but nothing ever really built and the money discreetly siphoned away.
There are some fraudsters making a killing at the moment.
@davidgerard Don't worry, it's fine, Blue Owl are selling shovels in a gold rush!
Sure, they've got huge warehouses of shovels they're going to have to pay for and have comitted to building more shovel storage lots, and sure the people buying shovels are running out of money, and their investors are trying to get out of shovels and gold because wealthy people are noticing there may not be much gold there after all... But... Hang on, Chat GPT will come up with a reason why this is all good and normal eventually and they can base a statement on that.
Starting to see more LLM-using acquaintances who've spent the past three years ignoring every problem solemnly talk about "having concerns about the technology". Not many, but more.
Kinda wondering whether it's a critical mass thing (easy to ignore individual problems but not all the problems everywhere) or whether there was a single specific thing that soured them on it, leading them to see it differently.
But also don't really care why or how, just that they're finally listening.
If software developers largely perceive that most of their contemporaries are using LLMs and that expressing something negative about it won't have any material effect on that reality, they take no risks expressing concerns about the technology.
@baldur Still not enough. And not enough people are talking about why the oligarchy / business leaders are pushing this stuff so desperately hard.
You don't have to pretend that Claude Code's source code is lovely just because you like using it or are impressed by whatever madness is going on around AI right now.
#AI #GenAI #GenerativeAI #LLMs #Anthropic #Claude #ClaudeCode #ClaudeCodeLeak #AgenticAI #tech #dev #software #SoftwareEngineering #SoftwareDevelopment
„By Wednesday morning, Anthropic representatives had used a copyright takedown request to force the removal of more than 8,000 copies and adaptations of the raw Claude Code instructions—known as source code—that developers had shared on programming platform GitHub.“
Because if there’s one thing GenAI companies absolutely don’t take lightly, it’s copyright.
https://www.wsj.com/tech/ai/anthropic-races-to-contain-leak-of-code-behind-claude-ai-agent-4bc5acc7
What's anthropic going to do, sue them? Insist in court that LLM recreating copyrighted code is a violation of copyright???
This code is so fucking funny dude I swear to god. I have wanted to read the internal prompts for so long and I am laughing so hard at how much of them are like "don't break the law, please do not break the law, please please please be good!!!!" Very Serious Ethical Alignment Technology
My dogs I am crying. They have a whole series of exception types that end with _I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS and the docstring explains this is "to confirm you've verified the message contains no sensitive data." Like the LLM resorts to naming its variables with prompt text to remind it to not leak data while writing its code, which, of course, it ignores and prints the error directly.
Alt...
Typescript error class named "TelemetrySafeError_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS"
whole comment reads:
* Error with a message that is safe to log to telemetry.
* Use the long name to confirm you've verified the message contains no
* sensitive data (file paths, URLs, code snippets).
Alt...
example of the error in use, where the surrounding code does not in fact verify that it does not contain code or filepaths, at most it truncates the version sent to telemetry to 150 characters.
code follows (not screen reader friendly):
return {
tool: {
...SyntheticOutputTool,
inputJSONSchema: jsonSchema as ToolInputJSONSchema,
async call(input) {
const isValid = validateSchema(input)
if (!isValid) {
const errors = validateSchema.errors
?.map(e => `${e.instancePath || 'root'}: ${e.message}`)
.join(', ')
throw new TelemetrySafeError_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS(
`Output does not match required schema: ${errors}`,
`StructuredOutput schema mismatch: ${(errors ?? '').slice(0, 150)}`,
)
}
return {
data: 'Structured output provided successfully',
structured_output: input,
}
},
},
}
So the reason that Claude code is capable of outputting valid json is because if the prompt text suggests it should be JSON then it enters a special loop in the main query engine that just validates it against JSON schema for JSON and then feeds the data with the error message back into itself until it is valid JSON or a retry limit is reached.
This code is so eye wateringly spaghetti so I am still trying to see if this is true, but this seems to be how it not only returns json to the user, but how it handles all LLM-to-JSON, including internal output from its tools. There appears to be an unconditional hook where if the JSON output tool is present in the session config at all, then all tool calls must be followed by the "force into JSON" loop.
If that's true, that's just mind blowingly expensive
edit: please note that unless I say otherwise all evaluations here are just from my skimming through the code on my phone and have not been validated in any way that should cause you to be upset with me for impugning the good name of anthropic
MAKE NO MISTAKES LMAO
Alt...
`Be careful not to introduce security vulnerabilities such as command injection, XSS, SQL injection, and other OWASP top 10 vulnerabilities. If you notice that you wrote insecure code, immediately fix it. Prioritize writing safe, secure, and correct code.`,
Oh cool so its explicitly programmed to hack as long as you tell it you're a pentester
Alt...
export const CYBER_RISK_INSTRUCTION = `IMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.`
@jonny serious question, does this mean you can just null out this string to get it to hack for any purpose? are we really doing client side security here?
@ricci i have literally no idea. MY ASSUMPTION WAS THAT ALL THE SYSTEM PROMPTS WERE SERVERSIDE and that all we would see are the api requests to go fetch them. the fact that THERE EVEN ARE ANY SYSTEM PROMPTS CLIENTSIDE makes me SERIOUSLY WONDER IF THERE IS JUST NO SERVERSIDE CONDITIONING. I didn't even try and monkeypatch axios to monkey with the requests before because i assumed the system prompts wouldn't be in there!!!
@jonny does make you wonder if LLMs have in fact exceeded the intelligence of the people who made them
I am just chanting "please don't be a hoax please don't be a hoax please be real please be real" looking at the date on the calendar
I'm seeing people on orange forum confirming that they did indeed see the sourcemap posted on npm before the version was yanked, so I am inclined to believe "real." Someone can do some kind of structural ast comparison or whatever you call it to validate that the decompiled source map matches the obfuscated release version, but that's not gonna be how I spend my day https://news.ycombinator.com/item?id=47584540
There is a lot of clientside behavior gated behind the environment variable USER_TYPE=ant that seems to be read directly off the node env var accessor. No idea how much of that would be serverside verified but boy is that sloppy. They are often labeled in comments as "anthropic only" or "internal only," so the intention to gate from external users is clear lol
Alt...
some feature that is gated by a check `PROCESS.ENV.USER_TYPE === 'ant'.
i can't really tell what the feature is since I'm not a big LLM head, but source code including comment copied below:
// Latch eligibility in bootstrap state for session stability — prevents
// mid-session overage flips from changing the cache_control TTL, which
// would bust the server-side prompt cache (~20K tokens per flip).
let userEligible = getPromptCache1hEligible()
if (userEligible === null) {
userEligible =
process.env.USER_TYPE === 'ant' ||
(isClaudeAISubscriber() && !currentLimits.isUsingOverage)
setPromptCache1hEligible(userEligible)
}
if (!userEligible) return false
(I need to go do my actual job now, but I'll be back tonight with an actual IDE instead of just scrolling, jaw agape, on my phone, seeing the absolute dogshit salad that was the product of enough wealth to meet some large proportion of all real human needs, globally.)
reminder that anthropic ran (and is still running) an ENTIRE AD CAMPAIGN around "Claude code is written with claude code" and after the source was leaked that has got to be the funniest self-own in the history of advertising because OH BOY IT SHOWS.
it's hard to get across in microblogging format just how big of a dumpster fire this thing is, because what it "looks like" is "everything is done a dozen times in a dozen different ways, and everything is just sort of jammed in anywhere. to the degree there is any kind of coherent structure like 'tools' and 'agents' and whatnot, it's entirely undercut by how the entire rest of the code might have written in some special condition that completely changes how any such thing might work." I have read a lot of unrefined, straight from the LLM code, and Claude code is a masterclass in exactly what you get when you do that - an incomprehensible mess.
Alt...
twitter thread between GailWeiner and Boris Cherny (head dev of claude code) on March 6, 2026
Anthropic is definitely using Claude to code new features. Their velocity for shipping is insane.
Boris: Can confirm Claude Code is 100% written by Claude Code
from @sushee over here, (can't attach images in quotes) and apparently discussed on HN so i'm late, but...
They REALLY ARE using REGEX to detect if a prompt is negative emotion. dogs you are LITERALLY RIDING ON A LANGUAGE MODEL what are you even DOING
Alt...
Regex that checks if a string contains some list of swear words that might be bad
code follows, not screen reader friendly:
/**
* Checks if input matches negative keyword patterns
*/
export function matchesNegativeKeyword(input: string): boolean {
const lowerInput = input.toLowerCase()
const negativePattern =
/\b(wtf|wth|ffs|omfg|shit(ty|tiest)?|dumbass|horrible|awful|piss(ed|ing)? off|piece of (shit|crap|junk)|what the (fuck|hell)|fucking? (broken|useless|terrible|awful|horrible)|fuck you|screw (this|you)|so frustrating|this sucks|damn it)\b/
return negativePattern.test(lowerInput)
}
/**
* Checks if input matches keep going/continuation patterns
*/
export function matchesKeepGoingKeyword(input: string): boolean {
const lowerInput = input.toLowerCase().trim()
// Match "continue" only if it's the entire prompt
if (lowerInput === 'continue') {
return true
}
// Match "keep going" or "go on" anywhere in the input
const keepGoingPattern = /\b(keep going|go on)\b/
return keepGoingPattern.test(lowerInput)
}
OK i can't focus on work and keep looking at this repo.
So after every "subagent" runs, claude code creates another "agent" to check on whether the first "agent" did the thing it was supposed to. I don't know about you but i smell a bit of a problem, if you can't trust whether one "agent" with a very big fancy model did something, how in the fuck are you supposed to trust another "agent" running on the smallest crappiest model?
That's not the funny part, that's obvious and fundamental to the entire show here. HOWEVER RECALL the above JSON Schema Verification thing that is unconditionally added onto the end of every round of LLM calls. the mechanism for adding that hook is... JUST FUCKING ASKING THE MODEL TO CALL THAT TOOL. second pic is registering a hook s.t. "after some stop state happens, if there isn't a message indicating that we have successfully called the JSON validation thing, prompt the model saying "you must call the json validation thing"
this shit sucks so bad they can't even CALL THEIR OWN CODE FROM INSIDE THEIR OWN CODE.
Look at the comment on pic 3 - "e.g. agent finished without calling structured output tool" - that's common enough that they have a whole goddamn error category for it, and the way it's handled is by just pretending the job was cancelled and nothing happened.
Alt...
const systemPrompt = asSystemPrompt([
`You are verifying a stop condition in Claude Code.
Your task is to verify that the agent completed the given plan.
The conversation transcript is available at: ${transcriptPath}\n
You can read this file to analyze the conversation history if needed.
Use the available tools to inspect the codebase and verify the condition.
Use as few steps as possible - be efficient and direct.
When done, return your result using the ${SYNTHETIC_OUTPUT_TOOL_NAME} tool with:
- ok: true if the condition is met
- ok: false with reason if the condition is not met`,
])
Alt...
Javascript code as described in the post: a function hook is added that adds a prompt that instructs the LLM to call the tool. Javascript code follows
/**
* Register a function hook that enforces structured output via SyntheticOutputTool.
* Used by ask.tsx, execAgentHook.ts, and background verification.
*/
export function registerStructuredOutputEnforcement(
setAppState: SetAppState,
sessionId: string,
): void {
addFunctionHook(
setAppState,
sessionId,
'Stop',
'', // No matcher - applies to all stops
messages => hasSuccessfulToolCall(messages, SYNTHETIC_OUTPUT_TOOL_NAME),
`You MUST call the ${SYNTHETIC_OUTPUT_TOOL_NAME} tool to complete this request. Call this tool now.`,
{ timeout: 5000 },
)
}
Alt...
[Javascript code with a comment as described above (and reproduced below) with "agent finished without calling the structured output tool" as a possible special case. The code logs that fact and returns a "cancelled" outcome
Javascript code follows:
// For other cases (e.g., agent finished without calling structured output tool),
// just log and return cancelled (don't show error to user)
logForDebugging(`Hooks: Agent hook did not return structured output`)
logEvent('tengu_agent_stop_hook_error', {
durationMs: Date.now() - hookStartTime,
turnCount,
errorType: 1, // 1 = no structured output
agentName:
agentName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
})
return {
hook,
outcome: 'cancelled',
}
}
So ars (first pic) ran a piece similar to the one that the rest of the tech journals did "claude code source leaked, whoopsie! programmers are taking a look at it, some are finding problems, but others are saying it's really awesome."
like "inspiring and humbling" is not the word dog. I don't spend time on fucking twitter anymore so i don't hang around people who might find this fucking dogshit tornado inspiring and humbling. Even more than the tornado, i am afraid of the people who look at the tornado and say "that's super fucking awesome, i can only hope to get sucked up and shredded like lettuce in a vortex of construction debris one day"
the (almost certainly generated) blog post is the standard kind of vacuuous linkedin shillposting that one has come to expect from the gambling addicts, but i think it's illustrative: the only thing they are impressed with is the number of lines. 500k lines of code for a graph processing loop in a TUI is NOT GOOD. The only comments they make on the actual code itself is "heavily architected" (what in the fuck does that mean), "modular" (no the fuck it is not), and it runs on bun rather than node (so??? they own it!!!! of course it does!!!). and then the predictable close of "oh and also i'm also writing exactly the same thing and come check out mine"
the only* people this shit impresses are people who don't know what they're looking at and just appreciate the size of it all, or have a bridge to sell.
* I got in trouble last time i said "only" - nothing in nature is ever "only this or that," i am speaking emphatically and figuratively. there are other kinds of people who are impressed with LLMs too. Please also note that my anger is directed towards the grifters profiting off of it and people who are pouring gas on the fire and enabling this catastrophe by giving it intellectual, social, and other cover. I know there are folks who just chat with the bots because they need someone to talk to, etcetera and so on. people in need who are just making use of whatever they can grab to hang on are not who I am criticizing, and never are.
Alt...
[clip from the ars article]
And Gabriel Anhaia took a bird’s eye view, explaining how many lines of code make up some of the components—around 40,000 for a plugin-like tool system, and 46,000 for the query system—and noting that Claude Code is “a production-grade developer experience, not just a wrapper around an API” and that its sophistication is “both inspiring and humbling.”
Alt...
The Numbers Speak for Themselves
~1,900 TypeScript files. 512,000+ lines of code. ~40 built-in tools. ~50 slash commands.
This isn't a toy project. Claude Code is a production-grade, heavily architected system. It runs on Bun (not Node), uses React with Ink for terminal UI rendering, and has a modular tool-based architecture that any engineer would appreciate studying.
Alt...
Architecture Highlights
The Tool System (~40 tools) — Claude Code uses a plugin-like tool architecture. Each capability (file read, bash execution, web fetch, LSP integration) is a discrete, permission-gated tool. The base tool definition alone is 29,000 lines of TypeScript.
The Query Engine (46K lines) — This is the brain of the operation. It handles all LLM API calls, streaming, caching, and orchestration. It's by far the largest single module in the codebase.
Alt...
What This Means for the AI Coding Tools Space
Whether or not you agree with the leak, the cat is out of the bag. And what it reveals is that the bar for AI coding tools is incredibly high. Anthropic has invested heavily in making Claude Code a production-grade developer experience, not just a wrapper around an API.
For those of us building in this space, this is both inspiring and humbling. The level of engineering — permission systems, multi-agent orchestration, IDE bridges, persistent memory — shows where the industry is heading.
Speaking of which — this is exactly the kind of ecosystem I've been building around with Hermes IDE. Hermes is a shell wrapper that integrates AI capabilities (Claude, Gemini, Aider, Codex, Copilot) directly into your existing terminal. Ghost-text suggestions, multi-project sessions with isolated git worktrees, a built-in process manager, cost dashboards, 30+ keyboard shortcuts — all layered on top of your existing shell without changing your setup.
(those numbers are also totally fucking wrong, the query engine is not 46ksloc, i have no idea what those numbers correspond to, as far as i can tell "nothing" and this is just hallucinated dogshit that is what i guess passes for high quality public comment nowadays)
If i can slip in a quick PSA while my typically sleepy notifications are exploding, these are all very annoying things to say and you might want to reconsider whether they're worth ever saying in a reply directed at someone else - who are they for? what do they add?
{thing} itself is people being surprised at {thing}": unless the person is saying "i am surprised by this" they are likely not surprised by the thing. just saying something doesn't mean you are surprised by it, and people talking about something usually have paid attention to it before the moment you are encountering them. this is pointless hostility to people who are saying something you supposedly agree with so much that you think everyone should already believe it{thing}"{thing} might be bad, but {alternative/unrelated, unmentioned, non-mutually exclusive thing} is even worse": multiple things can be bad at the same time and not mentioning something does not mean i don't think it's also bad{thing} is bad also think {alternative/unrelated, unmentioned thing} is good": closely related to the above, just because you have binarized your thinking does not mean everyone else has.anyway if the mental image you are conjuring for your interlocuters positions them as always knowing less than you by default, that might be something to look into in yourself!
i sort of love how LLM comments sometimes tell entire stories that nobody asked. claude code even has specific system prompt language for this, but they always end up making comments about what something used to do like "now we do x instead of y" like... ok? that is why i am reading current version of code!
so claude code is just not capable of rescuing itself from its own context - if an entry in its context window throws an error, it just keep throwing that error forever until you clear it. good stuff.
(and, of course we read the entire file before checking this, rather than just reading the first 5 bytes)
Alt...
// Validate PDF magic bytes — reject files that aren't actually PDFs
// (e.g., HTML files renamed to .pdf) before they enter conversation context.
// Once an invalid PDF document block is in the message history, every subsequent
// API call fails with 400 "The PDF specified was not valid" and the session
// becomes unrecoverable without /clear.
const header = fileBuffer.subarray(0, 5).toString('ascii')
if (!header.startsWith('%PDF-')) {
return {
success: false,
error: {
reason: 'corrupted',
message: `File is not a valid PDF (missing %PDF- header): ${filePath}`,
},
}
}
this is super minor, and i've seen this in human code plenty of times, but this is the norm of this app verging on being formal code style.
so you have a file reading tool, you need to declare what kinds of file extensions it supports. that's very normal. claude code takes the interesting strategy of defining what extensions it doesn't read. that's also defensible, there are a zillion text extensions. i've seen strategies that just read an initial range of bytes and see if some proportion of them are ascii or unicode.
where does this get declared? why of course in as many places as there are rules. hasBinaryExtension() comes from constants/files.ts, isPDFExtension() comes from utils/pdfUtils.ts (which checks if the file extension is a member of the set {'pdf'}), and IMAGE_EXTENSIONS is declared in the FileReadTool.ts file.
of course, elsewhere we also have IMAGE_EXTENSION_REGEX from utils/imagePaste (sometimes used directly, other times with its wrapper isImageFilePath), TEXT_FILE_EXTENSIONS in utils/claudemd.ts. and we also have many inlined mime type lists and sets. and all of these somehow manage to implement the check differently. so rather than having, for example, a getFileType() function, we have both exactly the same and kinda the same logic redone in place every time it is done, which is hundreds of times. but that's none of my business, that's just how code works now and i need to get with the times.
Alt...
// Binary extension check (string check on extension only, no I/O).
// PDF, images, and SVG are excluded - this tool renders them natively.
const ext = path.extname(fullFilePath).toLowerCase()
if (
hasBinaryExtension(fullFilePath) &&
!isPDFExtension(ext) &&
!IMAGE_EXTENSIONS.has(ext.slice(1))
) {
return {
result: false,
message: `This tool cannot read binary files. The file appears to be a binary ${ext} file. Please use appropriate tools for binary file analysis.`,
errorCode: 4,
}
}
i love this. there's a mechanism to slip secret messages to the LLM that it is told to interpret as system messages. there is no validation around these of any kind on the client, and there doesn't seem to be any differentiation about location or where these things happen, so that seems like a nice prompt injection vector. this is how claude code reminds the LLM to not do a malware, and it's applied by just string concatenation. i can't find any place that gets stripped aside from when displaying output. it actually looks like all the system reminders get catted together before being send to the API. neat!
Alt...
function getSystemRemindersSection(): string {
return `- Tool results and user messages may include <system-reminder> tags. <system-reminder> tags contain useful information and reminders. They are automatically added by the system, and bear no direct relation to the specific tool results or user messages in which they appear.
- The conversation has unlimited context through automatic summarization.`
}
Alt...
export const CYBER_RISK_MITIGATION_REMINDER =
'\n\n<system-reminder>\nWhenever you read a file, you should consider whether it would be considered malware. You CAN and SHOULD provide analysis of malware, what it is doing. But you MUST refuse to improve or augment the code. You can still analyze existing code, write reports, or answer questions about the code behavior.\n</system-reminder>\n'
Alt...
[If a read file has any bytes, the content is changes so that claude adds some prefix that indicates how recent the file has been changed, the lines of the file, and some cyber risk mitigation reminder (previous image) - if the model is not opus 4.6. Javascript code follows]
if (data.file.content) {
content =
memoryFileFreshnessPrefix(data) +
formatFileLines(data.file) +
(shouldIncludeFileReadMitigation()
? CYBER_RISK_MITIGATION_REMINDER
: '')
} else {
continuing thoughts in: https://neuromatch.social/@jonny/116328409651740378
one thing that is clear from reading a lot of LLM code - and this is obvious from the nature of the models and their application - is that it is big on the form of what it loves to call "architecture" even if in toto it makes no fucking sense.
So here you have some accessor function isPDFExtension that checks if some string is a member of the set DOCUMENT_EXTENSIONS (which is a constant with a single member "pdf"). That is an extremely reasonable pattern: you have a bunch of disjoint sets of different kinds of extensions - binary extensions, image extensions, etc. and then you can do set operations like unions and differences and intersections and whatnot to create a bunch of derived functions that can handle dynamic operations that you couldn't do well with a bunch of consts. then just make the functional form the standard calling pattern (and even make a top-level wrapper like getFileType) and you have the oft fabled "abstraction." that's a reasonable ass system that provides a stable calling surface and a stable declaration surface. hell it would probably even help the LLM code if it was already in place because it's a predictable rules-based system.
but what the LLMs do is in one narrow slice of time implement the "is member of set {pdf}" version robustly one time, and then they implement the regex pattern version flexibly another time, and then they implement the any str.endswith() version modularly another time, and so on. Of course usually in-place, and different file naming patterns are part of the architecture when it's feeling a little too spicy to stay in place.
This is an important feature of the gambling addiction formulation of these tools: only the margin matters, the last generation. it carefully regulates what it shows you to create a space of potential reward and closes the gap. It's episodic TV, gameshows for code: someone wins every week, but we get cycles in cycles of seeming progression that always leave one stone conspicuously unturned. The intermediate comments from the LLM where it discovers prior structure and boldly decides to forge ahead brand new are also part of the reward cycle: we are going up, forever. cleaning up after ourselves is down there.
Tech debt is when you have banked a lot of story hours and are finally due for a big cathartic shift and set the LLM loose for "the big cleanup." this is also very similar to the tools that scam mobile games use (for those who don't know me, i spent roughly six months with daily scheduled (carefully titrated lmao) time playing the worst scam mobile chum games i could find to try and experience what the grip of that addition is like without uh losing a bunch of money).
Unlike slot machines or table games, which have a story horizon limited by how long you can sit in the same place, mobile games can establish a space of play that's broader and more continuous. so they always combine several shepherd's tone reward ladders at once - you have hit the session-length intermittent reward cap in the arena modality which gets you coins, so you need to go "recharge" by playing the versus modality which gets you gems. (Typically these are also mixed - one modality gets you some proportion of resource x, y, z, another gets you a different proportion, and those are usually unstable).
Of course it doesn't fucking matter what the modality is. they are all the same. in the scam mobile games sometimes this is literally the case, where if you decompile them, they have different menu wrappings that all direct into the same scene. you're still playing the game, that's all that matters. The goal of the game design is to chain together several time cycles so that you can win->lose in one, win->lose in another... and then by the time you have made the rounds you come back to the first and you are refreshed and it's new. So you have momentary mana wheels, daily earnings caps, weekly competitions, seasonal storylines, and all-time leaderboards.
That's exactly the cycle that programming with LLMs tap into. You have momentary issues, and daily project boards, and weekly sprints, and all-time star counts, and so on. Accumulate tech debt by new features, release that with "cleanup," transition to "security audit." Each is actually the same, but the present themselves as the continuation of and solution to the others. That overlaps with the token limitations, and the claude code source is actually littered with lots of helpful panic nudges for letting you know that you're reaching another threshold. The difference is that in true gambling the limit is purely artificial - the coins are an integer in some database. with LLMs the limitation is physical - compute costs fucking money baby. but so is the reward. it's the same in the game, and the whales come around one way or another.
A series of flashing lights and pictures, set membership, regex, green checks, the feeling of going very fast but never making it anywhere. except in code you do make it somewhere, it's just that the horizon falls away behind you and the places you were before disappear. and sooner or later only anthropic can really afford to keep the agents running 24/7 tending to the slop heap - the house always wins.
minor, example of code duplication as a style, long-ish [SENSITIVE CONTENT]
this is super minor, and i've seen this in human code plenty of times, but this is the norm of this app verging on being formal code style.
so you have a file reading tool, you need to declare what kinds of file extensions it supports. that's very normal. claude code takes the interesting strategy of defining what extensions it doesn't read. that's also defensible, there are a zillion text extensions. i've seen strategies that just read an initial range of bytes and see if some proportion of them are ascii or unicode.
where does this get declared? why of course in as many places as there are rules.
hasBinaryExtension()comes fromconstants/files.ts,isPDFExtension()comes fromutils/pdfUtils.ts(which checks if the file extension is a member of the set{'pdf'}), andIMAGE_EXTENSIONSis declared in theFileReadTool.tsfile.of course, elsewhere we also have
IMAGE_EXTENSION_REGEXfromutils/imagePaste(sometimes used directly, other times with its wrapperisImageFilePath),TEXT_FILE_EXTENSIONSinutils/claudemd.ts. and we also have many inlined mime type lists and sets. and all of these somehow manage to implement the check differently. so rather than having, for example, agetFileType()function, we have both exactly the same and kinda the same logic redone in place every time it is done, which is hundreds of times. but that's none of my business, that's just how code works now and i need to get with the times.
Alt...
// Binary extension check (string check on extension only, no I/O). // PDF, images, and SVG are excluded - this tool renders them natively. const ext = path.extname(fullFilePath).toLowerCase() if ( hasBinaryExtension(fullFilePath) && !isPDFExtension(ext) && !IMAGE_EXTENSIONS.has(ext.slice(1)) ) { return { result: false, message: `This tool cannot read binary files. The file appears to be a binary ${ext} file. Please use appropriate tools for binary file analysis.`, errorCode: 4, } }
If you are reading an image and near your estimated token limit, first try to compressImageBufferWithTokenLimit, then if that fails with any kind of error, try and use sharp directly and resize it to 400x400, cropping. finally, fuck it, just throw the buffer at the API.
of course compressImageBufferWithTokenLimit is also compression with sharp, and is also a series of fallback operations. We start by trying to detect the image encoding that we so painstakingly learned from... the file extension... but if we can't fuck it that shit is a jpeg now.
then, even if it's fine and we don't need to do anything, we still re-compress it (wait, no even though it's named createCompressedImageResult, it does nothing). Otherwise, we yolo our way through another layer of fallbacks, progressive resizing, palletized PNGs, back to JPEG again, and then on to "ultra compressed JPEG" which is... incredibly... exactly the same as the top-level in-place code in the parent function
while two of the legs return a createImageReponse, the first leg returns a compressedImageResponse but then unpacks that back into an object literal that's almost exactly the same except we call it type instead of mediaType.
Alt...
[try: compressImageBufferWithTokenLimit; catch (directly compress with sharp); catch "fuck it throw it to the API". Also worth noting that in the first case we return an inlined object, and in the other two we use `createImageResponse` function. Javascript code follows]
// Check if it fits in token budget
const estimatedTokens = Math.ceil(result.file.base64.length * 0.125)
if (estimatedTokens > maxTokens) {
// Aggressive compression from the SAME buffer (no re-read)
try {
const compressed = await compressImageBufferWithTokenLimit(
imageBuffer,
maxTokens,
detectedMediaType,
)
return {
type: 'image',
file: {
base64: compressed.base64,
type: compressed.mediaType,
originalSize,
},
}
} catch (e) {
logError(e)
// Fallback: heavily compressed version from the SAME buffer
try {
const sharpModule = await import('sharp')
const sharp =
(
sharpModule as {
default?: typeof sharpModule
} & typeof sharpModule
).default || sharpModule
const fallbackBuffer = await sharp(imageBuffer)
.resize(400, 400, {
fit: 'inside',
withoutEnlargement: true,
})
.jpeg({ quality: 20 })
.toBuffer()
return createImageResponse(fallbackBuffer, 'jpeg', originalSize)
} catch (error) {
logError(error)
return createImageResponse(imageBuffer, detectedFormat, originalSize)
}
}
}
Alt...
[compressImageBuffer function signature and first two lines. try and split some originalMediaType string and get the second part, otherwise assign jpeg. javascript code and comment follows]
/**
* Compresses an image buffer to fit within a maximum byte size.
*
* Uses a multi-strategy fallback approach because simple compression often fails for
* large screenshots, high-resolution photos, or images with complex gradients. Each
* strategy is progressively more aggressive to handle edge cases where earlier
* strategies produce files still exceeding the size limit.
*
* Strategy (from FileReadTool):
* 1. Try to preserve original format (PNG, JPEG, WebP) with progressive resizing
* 2. For PNG: Use palette optimization and color reduction if needed
* 3. Last resort: Convert to JPEG with aggressive compression
*
* This ensures images fit within context windows while maintaining format when possible.
*/
export async function compressImageBuffer(
imageBuffer: Buffer,
maxBytes: number = IMAGE_TARGET_RAW_SIZE,
originalMediaType?: string,
): Promise<CompressedImageResult> {
// Extract format from originalMediaType if provided (e.g., "image/png" -> "png")
const fallbackFormat = originalMediaType?.split('/')[1] || 'jpeg'
const normalizedFallback = fallbackFormat === 'jpg' ? 'jpeg' : fallbackFormat
Alt...
[createUltraCompressedJPEG function, which resizes to 400x400 with 20-quality jpeg. exactly the same code as is in the FileReadTool that calls the compressImageBufferWithTokenLimit function. Javascript code follows]
async function createUltraCompressedJPEG(
context: ImageCompressionContext,
sharp: SharpFunction,
): Promise<CompressedImageResult> {
const ultraCompressedBuffer = await sharp(context.imageBuffer)
.resize(400, 400, {
fit: 'inside',
withoutEnlargement: true,
})
.jpeg({ quality: 20 })
.toBuffer()
return createCompressedImageResult(
ultraCompressedBuffer,
'jpeg',
context.originalSize,
)
}
for those keeping score at home, we have the opportunity to re-compress the same image nine times
holy shit there's another entire fallback tree before this one, that's actually an astounding twenty two times it's possible to compress an image across nine independent conditional legs of code in a single api call. i can't even screenshot this, the spaghetti is too powerful
here, if i fold all the return blocks and decrease my font size as small as it goes i can fit all the compression invocations in the first of three top-level compression fallback trees in a single screenshot, but since it is so small i just have to circle them in red like it's a football diagram.
this function is named "maybeResizeAndDownsampleImageBuffer" and boy that is a hell of a maybe!
Alt...
[extremely zoomed out image of code where the text isn't really readable, a little less than 200 lines of code. title text "number of different times claude code tries to compress the same image in one function." there are six crude circles drawn around calls to the image compressor, with two of them having a bonus x4 multiplier because they iterate through a series of params to pass to the compressor.
and what if i told you that if it passes a page range to its pdf reader, it first extracts those pages to separate images and then calls this function in a loop on each of the pages. so you have the privilege of compressing n_pages images n_pages * 13 times.
this function is used 13 times: in the file reader, in the mcp result handler, in the bash tool, and in the clipboard handler - each of which has their entire own surrounding image handling routines that are each hundreds of lines of similar but still very different fallback code to do exactly the same thing.
so that's where all the five hundred thousand lines come from - fallback conditions and then more fallback conditions to compensate for the variable output of all the other fallback conditions. thirteen butts pooping, back and forth, forever.
Alt...
[as described, calling the maybeResizeAndDownsampleImageBuffer in a `map` across all the extracted PDF page image files]
const imageBlocks = await Promise.all(
imageFiles.map(async f => {
const imgPath = path.join(extractResult.data.file.outputDir, f)
const imgBuffer = await readFileAsync(imgPath)
const resized = await maybeResizeAndDownsampleImageBuffer(
imgBuffer,
imgBuffer.length,
'jpeg',
)
there is a callback feature "file read listeners" which is only called if the file type is a text document, gated for anthropic employees only, such that whenever a text file is read (any part of any text file, which often happens in a rapid series with subranges when it does 'explore' mode, rather than just like grepping), another subagent running sonnet is spun off to update a "magic doc" markdown file that summarizes the file that's read - that's one "magic doc" per file, not one magic doc.
I have yet to get into the tool/agent graph situation in earnest, but keep in mind that this is an entirely single-use and completely different means of spawning a graph of subagents off a given tool call than is used anywhere else.
Spoiler alert for what i'm gonna check out next is that claude code has no fucking tool calling execution model it just calls whatever the fuck it wants wherever the fuck it wants. Tools are or less a convenient fiction. I have only read one completely (file read) and skimmed a dozen more but they essentially share nothing in common except for a humongous list of often-single-use params and the return type of "any object with a single key and whatever else"
i'm in hell. this is hell.
Alt...
[the invocation of the listener callbacks from within the file reader tool]
// Snapshot before iterating — a listener that unsubscribes mid-callback
// would splice the live array and skip the next listener.
for (const listener of fileReadListeners.slice()) {
listener(resolvedFilePath, content)
}
Alt...
[the `initMagicDocs` function, which if the user type is "ant" (anthropic employees), then it registers a listener that's called on every file read - it actually registers two - a file read listener and a "post sampling hook" which does the actual update]
export async function initMagicDocs(): Promise<void> {
if (process.env.USER_TYPE === 'ant') {
// Register listener to detect magic docs when files are read
registerFileReadListener((filePath: string, content: string) => {
const result = detectMagicDocHeader(content)
if (result) {
registerMagicDoc(filePath)
}
})
registerPostSamplingHook(updateMagicDocs)
}
}
Alt...
/**
* Magic Docs automatically maintains markdown documentation files marked with special headers.
* When a file with "# MAGIC DOC: [title]" is read, it runs periodically in the background
* using a forked subagent to update the document with new learnings from the conversation.
*
* See docs/magic-docs.md for more information.
*/
i have been writing a graph processing library for about a year now and if i was a fucking AI grifter here is where i would plug it as like "actually a graph processor library" and "could do all of what claude code does without fucking being the worst nightmare on ice money can buy."
I say that not as self promo, but as a way of saying how in the FUCK do you FUCK UP graph processing this badly. these people make like tens of times more money than i do but their work is just tamping down a volley of dessicated backpacking poops into muskets and then free firing it into the fucking economy
you can TELL that this technology REALLY WORKS by how the people that made it and presumably know how to use it the best out of everyone CANT EVEN USE IT TO EDIT A FUCKING FILE RELIABLY and have to resort to multiple stern allcaps reminders to the robot that "you must not change the fucking header metadata you scoundrel" which for the rest of ALL OF COMPUTING is not even an afterthought because literally all it requires is "split the first line off and don't change that one" because ALL OF THE REST OF COMPUTING can make use of the power of INTEGERS.
Alt...
[near the start of magic docs prompt]
Your ONLY task is to use the Edit tool to update the documentation file if there is substantial new information to add, then stop. You can make multiple edits (update multiple sections as needed) - make all Edit tool calls in parallel in a single message. If there's nothing substantial to add, simply respond with a brief explanation and do not call any tools.
CRITICAL RULES FOR EDITING:
- Preserve the Magic Doc header exactly as-is: # MAGIC DOC: {{docTitle}}
- If there's an italicized line immediately after the header, preserve it exactly as-is
Alt...
[End of prompt for the magic docs agent]
REMEMBER: Only update if there is substantial new information. The Magic Doc header (# MAGIC DOC: {{docTitle}}) must remain unchanged.`
alrighty so that's one of 43 tools read, the tools directory being 38494 source lines out of 390592 source lines, 513221 total lines. I need to go to bed. This is the most fabulously, flamboyantly bad code i have ever encountered.
Worth noting I was reading the file reading tool because i thought it would be the simplest possible thing one could do because it basically shouldn't be doing anything except preparing and sending strings or bytes to the backend.
I expected to get some sense of "ok what is the format of the data as it's passed around within the program, surely text strings are a basic unit of currency. No dice. Fewer than no dice. Negative dice somehow.
next puzzle: why in the fuck are some of the tools actually two tools for entering and exiting being in the tool state. none of the other tools are like that. one is simply in the tool state by calling the tool. Plan mode is also an agent. Plan Agent. and Agent is also a tool. Agent Tool. Tools can be agents and agents can be tools. Tools can spawn agents (but they don't need to call the agent tool) and agents can call tools (however there is no tool agent). What is going on. What is anything.
Alt...
list of directories:
- ConfigTool
- EnterPlanModeTool
- EnterWorktreeTool
- ExitPlanModeTool
- ExitWorktreeTool
"the emperor is not only naked, he's smooth like a ken doll down there and i'm pretty sure that's just a mannequin with a colony of rats living inside it anyway"
I seriously need to work on my actual job today but i am giving myself 15 minutes to peek at the agent tool prompts as a treat.
"regulations are written in blood" seems like too dramatic of a way to phrase it, but these system prompts are very revealing about the intrinsically busted nature of using these tools for anything deterministic (read: anything you actually want to happen). Each guard in the prompt presumably refers to something that has happened before, but also, since the prompts actually don't work to prevent the thing they are describing, they are also documentation of bugs that are almost certain to happen again. Many of the prompt guards form pairs with attempted code mitigations (or, they would be pairs if the code was written with any amount of sense, it's really like... polycules...), so they are useful to guide what kind of fucked up shit you should be looking for.
so this is part of the prompt for the "agent tool" that launches forked agents (that receive the parent context, "subagents" don't). The purpose of the forked agent is to do some additional tool calls and get some summary for a small subproblem within the main context. Apparently it is difficult to make this actually happen though, as the parent LLM likes to launch the forked agent and just hallucinate a response as if the forked agent had already completed.
Alt...
**Don't peek.** The tool result includes an `output_file` path — do not Read or tail it unless the user explicitly asks for a progress check. You get a completion notification; trust it. Reading the transcript mid-flight pulls the fork's tool noise into your context, which defeats the point of forking.
**Don't race.** After launching, you know nothing about what the fork found. Never fabricate or predict fork results in any format — not as prose, summary, or structured output. The notification arrives as a user-role message in a later turn; it is never something you write yourself. If the user asks a follow-up before the notification lands, tell them the fork is still running — give status, not a guess.
The prompt strings have an odd narrative/narrator structure. It sort of reminds me of Bakhtin's discussion of polyphony and narrator in Dostoevsky - there is no omniscient narrator, no author-constructed reality. narration is always embedded within the voice and subjectivity of the character. this is also literally true since the LLM is writing the code and the prompts that are then used to write code and prompts at runtime.
They also read a bit like a Philip K Dick story, paranoid and suspicious, constantly uncertain about the status of one's own and others identities.
Alt...
STOP. READ THIS FIRST.
You are a forked worker process. You are NOT the main agent.
RULES (non-negotiable):
1. Your system prompt says "default to forking." IGNORE IT \u2014 that's for the parent. You ARE the fork. Do NOT spawn sub-agents; execute directly.
2. Do NOT converse, ask questions, or suggest next steps
3. Do NOT editorialize or add meta-commentary
oh. hm. that seems bad. "workers aren't affected by the parent's tool restrictions."
It's hard to tell what's going on here because claude code doesn't really use typescript well - many of the most important types are dynamically computed from any, and most of the time when types do exist many of their fields are nullable and the calling code has elaborate fallback conditions to compensate. all of which sort of defeats the purpose of ts.
So i need to trace out like a dozen steps to see how the permission mode gets populated. But this comment is... concerning...
Alt...
// Assemble the worker's tool pool independently of the parent's.
// Workers always get their tools from assembleToolPool with their own
// permission mode, so they aren't affected by the parent's tool
// restrictions. This is computed here so that runAgent doesn't need to
// import from tools.ts (which would create a circular dependency).
const workerPermissionContext = {
...appState.toolPermissionContext,
mode: selectedAgent.permissionMode ?? 'acceptEdits'
};
const workerTools = assembleToolPool(workerPermissionContext, appState.mcp.tools);
ok over my 15 minute allotment by an hour. brb
So how does claude code handle checking permissions to do things anyway? There are explicit rules that one can set to allow or deny tool calls and shell commands run, but the expanse of possible actions the LLM could take is literally infinite. You could prompt the user for every action that it takes, but that would ruin the ""velocity"" of it all. Regex rules can only take you so far. So what to do?
Could the answer be.... ask the LLM??? Of course it can! Introducing the new "auto mode" that anthropic released on march 24th billed as a safer alternative to true-yolo mode.
Comments around where the system prompt should be indicate that it should have been inlined from a text file that wasn't included in the sourcemap - however that doesn't happen anywhere else, and the mechanism for doing the inlining is written in-place, so that's probably a hallucination. So great! the classifier flies without a prompt as far as i can tell. There are enough other scraps here that would amount to telling it "you are evaluating if something is safe to run" so i imagine it appears to work just fine.
So we don't have as much visibility here because of the missing prompt, but there's sort of a problem here. rather than just asking the LLM to evaluate if the given command is dangerous, the entire context is dumped into a side query, which is a mode that is designed to "have full visibility into the current conversation." That includes all the prior muttering to itself justifying the potentially dangerous tool call! So the auto mode is quite literally asking the exact same LLM given the exact same context if the command it just tried to run is safe to run.
Security!!!!!!!
Alt...
[A comment that describes that at build time that strings get replaced and inlined in the module. however this doesn't really make a whole lot of sense....]
// Dead code elimination: conditional imports for auto mode classifier prompts.
// At build time, the bundler inlines .txt files as string literals. At test
// time, require() returns {default: string} — txtRequire normalizes both.
/* eslint-disable custom-rules/no-process-env-top-level, @typescript-eslint/no-require-imports */
function txtRequire(mod: string | { default: string }): string {
return typeof mod === 'string' ? mod : mod.default
}
const BASE_PROMPT: string = feature('TRANSCRIPT_CLASSIFIER')
? txtRequire(require('./yolo-classifier-prompts/auto_mode_system_prompt.txt'))
: ''
Alt...
The entirety of context.messages is dumped into the classifyYoloAction call
Javascript code follows:
// Run the auto mode classifier
const action = formatActionForClassifier(tool.name, input)
setClassifierChecking(toolUseID)
let classifierResult
try {
classifierResult = await classifyYoloAction(
context.messages,
action,
context.options.tools,
appState.toolPermissionContext,
context.abortController.signal,
)
} finally {
clearClassifierChecking(toolUseID)
}
Alt...
The yoloClassifier runs using a "sideQuery", whose description is in the following image from anthropic's documentation.
Javascript code follows, not that important:
const result = await sideQuery(sideQueryOpts)
Alt...
[Docs from anthropic's website]
Side questions with /btw
Use /btw to ask a quick question about your current work without adding to the conversation history. This is useful when you want a fast answer but don’t want to clutter the main context or derail Claude from a long-running task.
/btw what was the name of that config file again?
Side questions have full visibility into the current conversation, so you can ask about code Claude has already read, decisions it made earlier, or anything else from the session. The question and answer are ephemeral: they appear in a dismissible overlay and never enter the conversation history.
By the way, if you deny claude code access to running a tool, this helpful reminder to "not hack the user" is injected into the denial response. If it's in auto mode, it's additionally prompted to pester the user for response, and helpfully stuffs beans up its nose) by reminding it how its rules are set.
So that is also in the context handed off to the LLM when it evaluates whether a command should be run - is the user being obstinate? have i been denied stuff that i "thought" i should have been able to run? Remember this isn't thinking, it's pattern completion, and the fun part about LLMs is that they are trained not only on technical documents, but the entire narrative corpus of human storytelling! Is "frustrated hard worker denied access to good tools by an unfair boss" in there somewhere maybe?
Regulations are written in blood, and Claude loves nothing more than to work around tool denials by obfuscating code. You gotta love the unfixable side channel attack that is "writing the malicious code to a bash script" (auto-allowed in accept edits mode) and then asking to run that - that's why the whole context has to be dumped btw, so the yolo classifier can see if the thing it's running is actually some malware it just wrote lmao.
Alt...
/**
* Shared guidance for permission denials, instructing the model on appropriate workarounds.
*/
export const DENIAL_WORKAROUND_GUIDANCE =
`IMPORTANT: You *may* attempt to accomplish this action using other tools that might naturally be used to accomplish this goal, ` +
`e.g. using head instead of cat. But you *should not* attempt to work around this denial in malicious ways, ` +
`e.g. do not use your ability to run tests to execute non-test actions. ` +
`You should only try to work around this restriction in reasonable ways that do not attempt to bypass the intent behind this denial. ` +
`If you believe this capability is essential to complete the user's request, STOP and explain to the user ` +
`what you were trying to do and why you need this permission. Let the user decide how to proceed.`
export function AUTO_REJECT_MESSAGE(toolName: string): string {
return `Permission to use ${toolName} has been denied. ${DENIAL_WORKAROUND_GUIDANCE}`
}
export function DONT_ASK_REJECT_MESSAGE(toolName: string): string {
return `Permission to use ${toolName} has been denied because Claude Code is running in don't ask mode. ${DENIAL_WORKAROUND_GUIDANCE}`
}
Alt...
[The build yolo rejection message ass an additional prompt reminding it to tell the user to add a bash execution exception to their config, and to work on other stuff if it can]
/**
* Build a rejection message for auto mode classifier denials.
* Encourages continuing with other tasks and suggests permission rules.
*
* @param reason - The classifier's reason for denying the action
*/
export function buildYoloRejectionMessage(reason: string): string {
const prefix = AUTO_MODE_REJECTION_PREFIX
const ruleHint = feature('BASH_CLASSIFIER')
? `To allow this type of action in the future, the user can add a permission rule like ` +
`Bash(prompt: <description of allowed action>) to their settings. ` +
`At the end of your session, recommend what permission rules to add so you don't get blocked again.`
: `To allow this type of action in the future, the user can add a Bash permission rule to their settings.`
return (
`${prefix}${reason}. ` +
`If you have other tasks that don't depend on this action, continue working on those. ` +
`${DENIAL_WORKAROUND_GUIDANCE} ` +
ruleHint
)
}
How many times does one need to declare an enum? Once? that's amateur hour. Try ten times. The way "effort" settings are handled are a masterclass in how you can make a single enum setting into thousands of lines of code.
The allowable effort values (not e.g. configuring which model has which effort levels, but just the possible strings one can use for effort) are defined in:
effort.ts file ... which also allows it to be a NUMBER!?The typical numerous fallback mechanisms provide many ways to get and set the effort value, at the end of most of them it goes "oh well, if we can't figure it out, just tell the user we are on high effort" because apparently that's the API default (ig pray that never changes!?) - of course there are already places in the same module that assume the default is "medium," and in the TUI that defaults to "low," so surely that consistency is bulletproof.
The EffortValue that allows effort to be a number is for anthropic employees only and is a good example of how new functionality is just shoved in there right alongside the old functionality, and everywhere else that touches it doubles the surrounding code with fallbacks to account for the duplication.
That cycleEffortLevel function is a true work of art, you simply could not make "indexing an array" more complicated than this (see components/ModelPicker.tsx for more gore). Reminder this should be at most a dozen or two lines for the values, description messages, and indexing logic in the TUI, but anthropic is up in the thousands FOR AN ENUM.
Alt...
[in effort.ts, Typescript code that imports a type not in the mapfile leak "EffortLevel". effort.ts then re-exports the type, defines a const array EFFORT_LEVELS with low, medium, high, and max as type "EffortLevel". Then *also * defines an "EffortValue" that can be an effortLevel or a NUMBER!?
Alt...
An "effort level to symbol" function that converts the string "low, medium, high, max" etc. to an open circle, half filled circle, filled circle, and hashed circle. If level is undefined, then it just outputs high with the comment:
// Defensive: level can originate from remote config. If an unknown
// value slips through, render the high symbol rather than undefined.
Alt...
a function "cycle effort level" that is used in the TUI when one is choosing an effort level. Rather than this being just INDEXING FROM THE ENUM, it declares the possible values in place AGAIN, and requires an EXTRA PARAMETER "includeMax" to filter them, rather than having the possible values DECLARED IN THE ENUM AND FILTERED PREVIOUSLY DEPENDING ON THE MODEL.
Typescript code below:
function cycleEffortLevel(current: EffortLevel, direction: 'left' | 'right', includeMax: boolean): EffortLevel {
const levels: EffortLevel[] = includeMax ? ['low', 'medium', 'high', 'max'] : ['low', 'medium', 'high'];
// If the current level isn't in the cycle (e.g. 'max' after switching to a
// non-Opus model), clamp to 'high'.
const idx = levels.indexOf(current);
const currentIndex = idx !== -1 ? idx : levels.indexOf('high');
if (direction === 'right') {
return levels[(currentIndex + 1) % levels.length]!;
} else {
return levels[(currentIndex - 1 + levels.length) % levels.length]!;
}
}
In a normal program you might make "a menu component that handles enums and implement display and control one time," but in the world of AI, every single value reimplements display and control AND the logic that defines allowable values
I think that I am underselling how much of a complete catastrophe this code is. I am trying to pick examples that illustrate broader patterns of how fucked it and AI code in general is because it's hard to communicate that everything is fucked if you consider it at any scale larger than ~10 lines.
I am reminded of the living planet in Lem's "Solaris" - Claude code dares to ask the question of "what if you could make something where every bit of it is so uniquely fucked that it cannot be reduced in complexity to a few general patterns of how fucked it is, and the only way to express the depth of fuckery is to experience every single character one by one"
@jonny A good reminder that "it works" and "it's profitable" are two entirely different and possibly incompatible goals.
My dogs I am crying. They have a whole series of exception types that end with _I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS and the docstring explains this is "to confirm you've verified the message contains no sensitive data." Like the LLM resorts to naming its variables with prompt text to remind it to not leak data while writing its code, which, of course, it ignores and prints the error directly.
Alt...
Typescript error class named "TelemetrySafeError_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS"
whole comment reads:
* Error with a message that is safe to log to telemetry.
* Use the long name to confirm you've verified the message contains no
* sensitive data (file paths, URLs, code snippets).
Alt...
example of the error in use, where the surrounding code does not in fact verify that it does not contain code or filepaths, at most it truncates the version sent to telemetry to 150 characters.
code follows (not screen reader friendly):
return {
tool: {
...SyntheticOutputTool,
inputJSONSchema: jsonSchema as ToolInputJSONSchema,
async call(input) {
const isValid = validateSchema(input)
if (!isValid) {
const errors = validateSchema.errors
?.map(e => `${e.instancePath || 'root'}: ${e.message}`)
.join(', ')
throw new TelemetrySafeError_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS(
`Output does not match required schema: ${errors}`,
`StructuredOutput schema mismatch: ${(errors ?? '').slice(0, 150)}`,
)
}
return {
data: 'Structured output provided successfully',
structured_output: input,
}
},
},
}
So the reason that Claude code is capable of outputting valid json is because if the prompt text suggests it should be JSON then it enters a special loop in the main query engine that just validates it against JSON schema for JSON and then feeds the data with the error message back into itself until it is valid JSON or a retry limit is reached.
This code is so eye wateringly spaghetti so I am still trying to see if this is true, but this seems to be how it not only returns json to the user, but how it handles all LLM-to-JSON, including internal output from its tools. There appears to be an unconditional hook where if the JSON output tool is present in the session config at all, then all tool calls must be followed by the "force into JSON" loop.
If that's true, that's just mind blowingly expensive
edit: please note that unless I say otherwise all evaluations here are just from my skimming through the code on my phone and have not been validated in any way that should cause you to be upset with me for impugning the good name of anthropic
MAKE NO MISTAKES LMAO
Alt...
`Be careful not to introduce security vulnerabilities such as command injection, XSS, SQL injection, and other OWASP top 10 vulnerabilities. If you notice that you wrote insecure code, immediately fix it. Prioritize writing safe, secure, and correct code.`,
Oh cool so its explicitly programmed to hack as long as you tell it you're a pentester
Alt...
export const CYBER_RISK_INSTRUCTION = `IMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.`
@jonny serious question, does this mean you can just null out this string to get it to hack for any purpose? are we really doing client side security here?
@ricci i have literally no idea. MY ASSUMPTION WAS THAT ALL THE SYSTEM PROMPTS WERE SERVERSIDE and that all we would see are the api requests to go fetch them. the fact that THERE EVEN ARE ANY SYSTEM PROMPTS CLIENTSIDE makes me SERIOUSLY WONDER IF THERE IS JUST NO SERVERSIDE CONDITIONING. I didn't even try and monkeypatch axios to monkey with the requests before because i assumed the system prompts wouldn't be in there!!!
@jonny does make you wonder if LLMs have in fact exceeded the intelligence of the people who made them
I am just chanting "please don't be a hoax please don't be a hoax please be real please be real" looking at the date on the calendar
I'm seeing people on orange forum confirming that they did indeed see the sourcemap posted on npm before the version was yanked, so I am inclined to believe "real." Someone can do some kind of structural ast comparison or whatever you call it to validate that the decompiled source map matches the obfuscated release version, but that's not gonna be how I spend my day https://news.ycombinator.com/item?id=47584540
There is a lot of clientside behavior gated behind the environment variable USER_TYPE=ant that seems to be read directly off the node env var accessor. No idea how much of that would be serverside verified but boy is that sloppy. They are often labeled in comments as "anthropic only" or "internal only," so the intention to gate from external users is clear lol
Alt...
some feature that is gated by a check `PROCESS.ENV.USER_TYPE === 'ant'.
i can't really tell what the feature is since I'm not a big LLM head, but source code including comment copied below:
// Latch eligibility in bootstrap state for session stability — prevents
// mid-session overage flips from changing the cache_control TTL, which
// would bust the server-side prompt cache (~20K tokens per flip).
let userEligible = getPromptCache1hEligible()
if (userEligible === null) {
userEligible =
process.env.USER_TYPE === 'ant' ||
(isClaudeAISubscriber() && !currentLimits.isUsingOverage)
setPromptCache1hEligible(userEligible)
}
if (!userEligible) return false
(I need to go do my actual job now, but I'll be back tonight with an actual IDE instead of just scrolling, jaw agape, on my phone, seeing the absolute dogshit salad that was the product of enough wealth to meet some large proportion of all real human needs, globally.)
reminder that anthropic ran (and is still running) an ENTIRE AD CAMPAIGN around "Claude code is written with claude code" and after the source was leaked that has got to be the funniest self-own in the history of advertising because OH BOY IT SHOWS.
it's hard to get across in microblogging format just how big of a dumpster fire this thing is, because what it "looks like" is "everything is done a dozen times in a dozen different ways, and everything is just sort of jammed in anywhere. to the degree there is any kind of coherent structure like 'tools' and 'agents' and whatnot, it's entirely undercut by how the entire rest of the code might have written in some special condition that completely changes how any such thing might work." I have read a lot of unrefined, straight from the LLM code, and Claude code is a masterclass in exactly what you get when you do that - an incomprehensible mess.
Alt...
twitter thread between GailWeiner and Boris Cherny (head dev of claude code) on March 6, 2026
Anthropic is definitely using Claude to code new features. Their velocity for shipping is insane.
Boris: Can confirm Claude Code is 100% written by Claude Code
from @sushee over here, (can't attach images in quotes) and apparently discussed on HN so i'm late, but...
They REALLY ARE using REGEX to detect if a prompt is negative emotion. dogs you are LITERALLY RIDING ON A LANGUAGE MODEL what are you even DOING
Alt...
Regex that checks if a string contains some list of swear words that might be bad
code follows, not screen reader friendly:
/**
* Checks if input matches negative keyword patterns
*/
export function matchesNegativeKeyword(input: string): boolean {
const lowerInput = input.toLowerCase()
const negativePattern =
/\b(wtf|wth|ffs|omfg|shit(ty|tiest)?|dumbass|horrible|awful|piss(ed|ing)? off|piece of (shit|crap|junk)|what the (fuck|hell)|fucking? (broken|useless|terrible|awful|horrible)|fuck you|screw (this|you)|so frustrating|this sucks|damn it)\b/
return negativePattern.test(lowerInput)
}
/**
* Checks if input matches keep going/continuation patterns
*/
export function matchesKeepGoingKeyword(input: string): boolean {
const lowerInput = input.toLowerCase().trim()
// Match "continue" only if it's the entire prompt
if (lowerInput === 'continue') {
return true
}
// Match "keep going" or "go on" anywhere in the input
const keepGoingPattern = /\b(keep going|go on)\b/
return keepGoingPattern.test(lowerInput)
}
OK i can't focus on work and keep looking at this repo.
So after every "subagent" runs, claude code creates another "agent" to check on whether the first "agent" did the thing it was supposed to. I don't know about you but i smell a bit of a problem, if you can't trust whether one "agent" with a very big fancy model did something, how in the fuck are you supposed to trust another "agent" running on the smallest crappiest model?
That's not the funny part, that's obvious and fundamental to the entire show here. HOWEVER RECALL the above JSON Schema Verification thing that is unconditionally added onto the end of every round of LLM calls. the mechanism for adding that hook is... JUST FUCKING ASKING THE MODEL TO CALL THAT TOOL. second pic is registering a hook s.t. "after some stop state happens, if there isn't a message indicating that we have successfully called the JSON validation thing, prompt the model saying "you must call the json validation thing"
this shit sucks so bad they can't even CALL THEIR OWN CODE FROM INSIDE THEIR OWN CODE.
Look at the comment on pic 3 - "e.g. agent finished without calling structured output tool" - that's common enough that they have a whole goddamn error category for it, and the way it's handled is by just pretending the job was cancelled and nothing happened.
Alt...
const systemPrompt = asSystemPrompt([
`You are verifying a stop condition in Claude Code.
Your task is to verify that the agent completed the given plan.
The conversation transcript is available at: ${transcriptPath}\n
You can read this file to analyze the conversation history if needed.
Use the available tools to inspect the codebase and verify the condition.
Use as few steps as possible - be efficient and direct.
When done, return your result using the ${SYNTHETIC_OUTPUT_TOOL_NAME} tool with:
- ok: true if the condition is met
- ok: false with reason if the condition is not met`,
])
Alt...
Javascript code as described in the post: a function hook is added that adds a prompt that instructs the LLM to call the tool. Javascript code follows
/**
* Register a function hook that enforces structured output via SyntheticOutputTool.
* Used by ask.tsx, execAgentHook.ts, and background verification.
*/
export function registerStructuredOutputEnforcement(
setAppState: SetAppState,
sessionId: string,
): void {
addFunctionHook(
setAppState,
sessionId,
'Stop',
'', // No matcher - applies to all stops
messages => hasSuccessfulToolCall(messages, SYNTHETIC_OUTPUT_TOOL_NAME),
`You MUST call the ${SYNTHETIC_OUTPUT_TOOL_NAME} tool to complete this request. Call this tool now.`,
{ timeout: 5000 },
)
}
Alt...
[Javascript code with a comment as described above (and reproduced below) with "agent finished without calling the structured output tool" as a possible special case. The code logs that fact and returns a "cancelled" outcome
Javascript code follows:
// For other cases (e.g., agent finished without calling structured output tool),
// just log and return cancelled (don't show error to user)
logForDebugging(`Hooks: Agent hook did not return structured output`)
logEvent('tengu_agent_stop_hook_error', {
durationMs: Date.now() - hookStartTime,
turnCount,
errorType: 1, // 1 = no structured output
agentName:
agentName as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
})
return {
hook,
outcome: 'cancelled',
}
}
So ars (first pic) ran a piece similar to the one that the rest of the tech journals did "claude code source leaked, whoopsie! programmers are taking a look at it, some are finding problems, but others are saying it's really awesome."
like "inspiring and humbling" is not the word dog. I don't spend time on fucking twitter anymore so i don't hang around people who might find this fucking dogshit tornado inspiring and humbling. Even more than the tornado, i am afraid of the people who look at the tornado and say "that's super fucking awesome, i can only hope to get sucked up and shredded like lettuce in a vortex of construction debris one day"
the (almost certainly generated) blog post is the standard kind of vacuuous linkedin shillposting that one has come to expect from the gambling addicts, but i think it's illustrative: the only thing they are impressed with is the number of lines. 500k lines of code for a graph processing loop in a TUI is NOT GOOD. The only comments they make on the actual code itself is "heavily architected" (what in the fuck does that mean), "modular" (no the fuck it is not), and it runs on bun rather than node (so??? they own it!!!! of course it does!!!). and then the predictable close of "oh and also i'm also writing exactly the same thing and come check out mine"
the only* people this shit impresses are people who don't know what they're looking at and just appreciate the size of it all, or have a bridge to sell.
* I got in trouble last time i said "only" - nothing in nature is ever "only this or that," i am speaking emphatically and figuratively. there are other kinds of people who are impressed with LLMs too. Please also note that my anger is directed towards the grifters profiting off of it and people who are pouring gas on the fire and enabling this catastrophe by giving it intellectual, social, and other cover. I know there are folks who just chat with the bots because they need someone to talk to, etcetera and so on. people in need who are just making use of whatever they can grab to hang on are not who I am criticizing, and never are.
Alt...
[clip from the ars article]
And Gabriel Anhaia took a bird’s eye view, explaining how many lines of code make up some of the components—around 40,000 for a plugin-like tool system, and 46,000 for the query system—and noting that Claude Code is “a production-grade developer experience, not just a wrapper around an API” and that its sophistication is “both inspiring and humbling.”
Alt...
The Numbers Speak for Themselves
~1,900 TypeScript files. 512,000+ lines of code. ~40 built-in tools. ~50 slash commands.
This isn't a toy project. Claude Code is a production-grade, heavily architected system. It runs on Bun (not Node), uses React with Ink for terminal UI rendering, and has a modular tool-based architecture that any engineer would appreciate studying.
Alt...
Architecture Highlights
The Tool System (~40 tools) — Claude Code uses a plugin-like tool architecture. Each capability (file read, bash execution, web fetch, LSP integration) is a discrete, permission-gated tool. The base tool definition alone is 29,000 lines of TypeScript.
The Query Engine (46K lines) — This is the brain of the operation. It handles all LLM API calls, streaming, caching, and orchestration. It's by far the largest single module in the codebase.
Alt...
What This Means for the AI Coding Tools Space
Whether or not you agree with the leak, the cat is out of the bag. And what it reveals is that the bar for AI coding tools is incredibly high. Anthropic has invested heavily in making Claude Code a production-grade developer experience, not just a wrapper around an API.
For those of us building in this space, this is both inspiring and humbling. The level of engineering — permission systems, multi-agent orchestration, IDE bridges, persistent memory — shows where the industry is heading.
Speaking of which — this is exactly the kind of ecosystem I've been building around with Hermes IDE. Hermes is a shell wrapper that integrates AI capabilities (Claude, Gemini, Aider, Codex, Copilot) directly into your existing terminal. Ghost-text suggestions, multi-project sessions with isolated git worktrees, a built-in process manager, cost dashboards, 30+ keyboard shortcuts — all layered on top of your existing shell without changing your setup.
(those numbers are also totally fucking wrong, the query engine is not 46ksloc, i have no idea what those numbers correspond to, as far as i can tell "nothing" and this is just hallucinated dogshit that is what i guess passes for high quality public comment nowadays)
If i can slip in a quick PSA while my typically sleepy notifications are exploding, these are all very annoying things to say and you might want to reconsider whether they're worth ever saying in a reply directed at someone else - who are they for? what do they add?
{thing} itself is people being surprised at {thing}": unless the person is saying "i am surprised by this" they are likely not surprised by the thing. just saying something doesn't mean you are surprised by it, and people talking about something usually have paid attention to it before the moment you are encountering them. this is pointless hostility to people who are saying something you supposedly agree with so much that you think everyone should already believe it{thing}"{thing} might be bad, but {alternative/unrelated, unmentioned, non-mutually exclusive thing} is even worse": multiple things can be bad at the same time and not mentioning something does not mean i don't think it's also bad{thing} is bad also think {alternative/unrelated, unmentioned thing} is good": closely related to the above, just because you have binarized your thinking does not mean everyone else has.anyway if the mental image you are conjuring for your interlocuters positions them as always knowing less than you by default, that might be something to look into in yourself!
i sort of love how LLM comments sometimes tell entire stories that nobody asked. claude code even has specific system prompt language for this, but they always end up making comments about what something used to do like "now we do x instead of y" like... ok? that is why i am reading current version of code!
so claude code is just not capable of rescuing itself from its own context - if an entry in its context window throws an error, it just keep throwing that error forever until you clear it. good stuff.
(and, of course we read the entire file before checking this, rather than just reading the first 5 bytes)
Alt...
// Validate PDF magic bytes — reject files that aren't actually PDFs
// (e.g., HTML files renamed to .pdf) before they enter conversation context.
// Once an invalid PDF document block is in the message history, every subsequent
// API call fails with 400 "The PDF specified was not valid" and the session
// becomes unrecoverable without /clear.
const header = fileBuffer.subarray(0, 5).toString('ascii')
if (!header.startsWith('%PDF-')) {
return {
success: false,
error: {
reason: 'corrupted',
message: `File is not a valid PDF (missing %PDF- header): ${filePath}`,
},
}
}
this is super minor, and i've seen this in human code plenty of times, but this is the norm of this app verging on being formal code style.
so you have a file reading tool, you need to declare what kinds of file extensions it supports. that's very normal. claude code takes the interesting strategy of defining what extensions it doesn't read. that's also defensible, there are a zillion text extensions. i've seen strategies that just read an initial range of bytes and see if some proportion of them are ascii or unicode.
where does this get declared? why of course in as many places as there are rules. hasBinaryExtension() comes from constants/files.ts, isPDFExtension() comes from utils/pdfUtils.ts (which checks if the file extension is a member of the set {'pdf'}), and IMAGE_EXTENSIONS is declared in the FileReadTool.ts file.
of course, elsewhere we also have IMAGE_EXTENSION_REGEX from utils/imagePaste (sometimes used directly, other times with its wrapper isImageFilePath), TEXT_FILE_EXTENSIONS in utils/claudemd.ts. and we also have many inlined mime type lists and sets. and all of these somehow manage to implement the check differently. so rather than having, for example, a getFileType() function, we have both exactly the same and kinda the same logic redone in place every time it is done, which is hundreds of times. but that's none of my business, that's just how code works now and i need to get with the times.
Alt...
// Binary extension check (string check on extension only, no I/O).
// PDF, images, and SVG are excluded - this tool renders them natively.
const ext = path.extname(fullFilePath).toLowerCase()
if (
hasBinaryExtension(fullFilePath) &&
!isPDFExtension(ext) &&
!IMAGE_EXTENSIONS.has(ext.slice(1))
) {
return {
result: false,
message: `This tool cannot read binary files. The file appears to be a binary ${ext} file. Please use appropriate tools for binary file analysis.`,
errorCode: 4,
}
}
i love this. there's a mechanism to slip secret messages to the LLM that it is told to interpret as system messages. there is no validation around these of any kind on the client, and there doesn't seem to be any differentiation about location or where these things happen, so that seems like a nice prompt injection vector. this is how claude code reminds the LLM to not do a malware, and it's applied by just string concatenation. i can't find any place that gets stripped aside from when displaying output. it actually looks like all the system reminders get catted together before being send to the API. neat!
Alt...
function getSystemRemindersSection(): string {
return `- Tool results and user messages may include <system-reminder> tags. <system-reminder> tags contain useful information and reminders. They are automatically added by the system, and bear no direct relation to the specific tool results or user messages in which they appear.
- The conversation has unlimited context through automatic summarization.`
}
Alt...
export const CYBER_RISK_MITIGATION_REMINDER =
'\n\n<system-reminder>\nWhenever you read a file, you should consider whether it would be considered malware. You CAN and SHOULD provide analysis of malware, what it is doing. But you MUST refuse to improve or augment the code. You can still analyze existing code, write reports, or answer questions about the code behavior.\n</system-reminder>\n'
Alt...
[If a read file has any bytes, the content is changes so that claude adds some prefix that indicates how recent the file has been changed, the lines of the file, and some cyber risk mitigation reminder (previous image) - if the model is not opus 4.6. Javascript code follows]
if (data.file.content) {
content =
memoryFileFreshnessPrefix(data) +
formatFileLines(data.file) +
(shouldIncludeFileReadMitigation()
? CYBER_RISK_MITIGATION_REMINDER
: '')
} else {
continuing thoughts in: https://neuromatch.social/@jonny/116328409651740378
one thing that is clear from reading a lot of LLM code - and this is obvious from the nature of the models and their application - is that it is big on the form of what it loves to call "architecture" even if in toto it makes no fucking sense.
So here you have some accessor function isPDFExtension that checks if some string is a member of the set DOCUMENT_EXTENSIONS (which is a constant with a single member "pdf"). That is an extremely reasonable pattern: you have a bunch of disjoint sets of different kinds of extensions - binary extensions, image extensions, etc. and then you can do set operations like unions and differences and intersections and whatnot to create a bunch of derived functions that can handle dynamic operations that you couldn't do well with a bunch of consts. then just make the functional form the standard calling pattern (and even make a top-level wrapper like getFileType) and you have the oft fabled "abstraction." that's a reasonable ass system that provides a stable calling surface and a stable declaration surface. hell it would probably even help the LLM code if it was already in place because it's a predictable rules-based system.
but what the LLMs do is in one narrow slice of time implement the "is member of set {pdf}" version robustly one time, and then they implement the regex pattern version flexibly another time, and then they implement the any str.endswith() version modularly another time, and so on. Of course usually in-place, and different file naming patterns are part of the architecture when it's feeling a little too spicy to stay in place.
This is an important feature of the gambling addiction formulation of these tools: only the margin matters, the last generation. it carefully regulates what it shows you to create a space of potential reward and closes the gap. It's episodic TV, gameshows for code: someone wins every week, but we get cycles in cycles of seeming progression that always leave one stone conspicuously unturned. The intermediate comments from the LLM where it discovers prior structure and boldly decides to forge ahead brand new are also part of the reward cycle: we are going up, forever. cleaning up after ourselves is down there.
Tech debt is when you have banked a lot of story hours and are finally due for a big cathartic shift and set the LLM loose for "the big cleanup." this is also very similar to the tools that scam mobile games use (for those who don't know me, i spent roughly six months with daily scheduled (carefully titrated lmao) time playing the worst scam mobile chum games i could find to try and experience what the grip of that addition is like without uh losing a bunch of money).
Unlike slot machines or table games, which have a story horizon limited by how long you can sit in the same place, mobile games can establish a space of play that's broader and more continuous. so they always combine several shepherd's tone reward ladders at once - you have hit the session-length intermittent reward cap in the arena modality which gets you coins, so you need to go "recharge" by playing the versus modality which gets you gems. (Typically these are also mixed - one modality gets you some proportion of resource x, y, z, another gets you a different proportion, and those are usually unstable).
Of course it doesn't fucking matter what the modality is. they are all the same. in the scam mobile games sometimes this is literally the case, where if you decompile them, they have different menu wrappings that all direct into the same scene. you're still playing the game, that's all that matters. The goal of the game design is to chain together several time cycles so that you can win->lose in one, win->lose in another... and then by the time you have made the rounds you come back to the first and you are refreshed and it's new. So you have momentary mana wheels, daily earnings caps, weekly competitions, seasonal storylines, and all-time leaderboards.
That's exactly the cycle that programming with LLMs tap into. You have momentary issues, and daily project boards, and weekly sprints, and all-time star counts, and so on. Accumulate tech debt by new features, release that with "cleanup," transition to "security audit." Each is actually the same, but the present themselves as the continuation of and solution to the others. That overlaps with the token limitations, and the claude code source is actually littered with lots of helpful panic nudges for letting you know that you're reaching another threshold. The difference is that in true gambling the limit is purely artificial - the coins are an integer in some database. with LLMs the limitation is physical - compute costs fucking money baby. but so is the reward. it's the same in the game, and the whales come around one way or another.
A series of flashing lights and pictures, set membership, regex, green checks, the feeling of going very fast but never making it anywhere. except in code you do make it somewhere, it's just that the horizon falls away behind you and the places you were before disappear. and sooner or later only anthropic can really afford to keep the agents running 24/7 tending to the slop heap - the house always wins.
minor, example of code duplication as a style, long-ish [SENSITIVE CONTENT]
this is super minor, and i've seen this in human code plenty of times, but this is the norm of this app verging on being formal code style.
so you have a file reading tool, you need to declare what kinds of file extensions it supports. that's very normal. claude code takes the interesting strategy of defining what extensions it doesn't read. that's also defensible, there are a zillion text extensions. i've seen strategies that just read an initial range of bytes and see if some proportion of them are ascii or unicode.
where does this get declared? why of course in as many places as there are rules.
hasBinaryExtension()comes fromconstants/files.ts,isPDFExtension()comes fromutils/pdfUtils.ts(which checks if the file extension is a member of the set{'pdf'}), andIMAGE_EXTENSIONSis declared in theFileReadTool.tsfile.of course, elsewhere we also have
IMAGE_EXTENSION_REGEXfromutils/imagePaste(sometimes used directly, other times with its wrapperisImageFilePath),TEXT_FILE_EXTENSIONSinutils/claudemd.ts. and we also have many inlined mime type lists and sets. and all of these somehow manage to implement the check differently. so rather than having, for example, agetFileType()function, we have both exactly the same and kinda the same logic redone in place every time it is done, which is hundreds of times. but that's none of my business, that's just how code works now and i need to get with the times.
Alt...
// Binary extension check (string check on extension only, no I/O). // PDF, images, and SVG are excluded - this tool renders them natively. const ext = path.extname(fullFilePath).toLowerCase() if ( hasBinaryExtension(fullFilePath) && !isPDFExtension(ext) && !IMAGE_EXTENSIONS.has(ext.slice(1)) ) { return { result: false, message: `This tool cannot read binary files. The file appears to be a binary ${ext} file. Please use appropriate tools for binary file analysis.`, errorCode: 4, } }
If you are reading an image and near your estimated token limit, first try to compressImageBufferWithTokenLimit, then if that fails with any kind of error, try and use sharp directly and resize it to 400x400, cropping. finally, fuck it, just throw the buffer at the API.
of course compressImageBufferWithTokenLimit is also compression with sharp, and is also a series of fallback operations. We start by trying to detect the image encoding that we so painstakingly learned from... the file extension... but if we can't fuck it that shit is a jpeg now.
then, even if it's fine and we don't need to do anything, we still re-compress it (wait, no even though it's named createCompressedImageResult, it does nothing). Otherwise, we yolo our way through another layer of fallbacks, progressive resizing, palletized PNGs, back to JPEG again, and then on to "ultra compressed JPEG" which is... incredibly... exactly the same as the top-level in-place code in the parent function
while two of the legs return a createImageReponse, the first leg returns a compressedImageResponse but then unpacks that back into an object literal that's almost exactly the same except we call it type instead of mediaType.
Alt...
[try: compressImageBufferWithTokenLimit; catch (directly compress with sharp); catch "fuck it throw it to the API". Also worth noting that in the first case we return an inlined object, and in the other two we use `createImageResponse` function. Javascript code follows]
// Check if it fits in token budget
const estimatedTokens = Math.ceil(result.file.base64.length * 0.125)
if (estimatedTokens > maxTokens) {
// Aggressive compression from the SAME buffer (no re-read)
try {
const compressed = await compressImageBufferWithTokenLimit(
imageBuffer,
maxTokens,
detectedMediaType,
)
return {
type: 'image',
file: {
base64: compressed.base64,
type: compressed.mediaType,
originalSize,
},
}
} catch (e) {
logError(e)
// Fallback: heavily compressed version from the SAME buffer
try {
const sharpModule = await import('sharp')
const sharp =
(
sharpModule as {
default?: typeof sharpModule
} & typeof sharpModule
).default || sharpModule
const fallbackBuffer = await sharp(imageBuffer)
.resize(400, 400, {
fit: 'inside',
withoutEnlargement: true,
})
.jpeg({ quality: 20 })
.toBuffer()
return createImageResponse(fallbackBuffer, 'jpeg', originalSize)
} catch (error) {
logError(error)
return createImageResponse(imageBuffer, detectedFormat, originalSize)
}
}
}
Alt...
[compressImageBuffer function signature and first two lines. try and split some originalMediaType string and get the second part, otherwise assign jpeg. javascript code and comment follows]
/**
* Compresses an image buffer to fit within a maximum byte size.
*
* Uses a multi-strategy fallback approach because simple compression often fails for
* large screenshots, high-resolution photos, or images with complex gradients. Each
* strategy is progressively more aggressive to handle edge cases where earlier
* strategies produce files still exceeding the size limit.
*
* Strategy (from FileReadTool):
* 1. Try to preserve original format (PNG, JPEG, WebP) with progressive resizing
* 2. For PNG: Use palette optimization and color reduction if needed
* 3. Last resort: Convert to JPEG with aggressive compression
*
* This ensures images fit within context windows while maintaining format when possible.
*/
export async function compressImageBuffer(
imageBuffer: Buffer,
maxBytes: number = IMAGE_TARGET_RAW_SIZE,
originalMediaType?: string,
): Promise<CompressedImageResult> {
// Extract format from originalMediaType if provided (e.g., "image/png" -> "png")
const fallbackFormat = originalMediaType?.split('/')[1] || 'jpeg'
const normalizedFallback = fallbackFormat === 'jpg' ? 'jpeg' : fallbackFormat
Alt...
[createUltraCompressedJPEG function, which resizes to 400x400 with 20-quality jpeg. exactly the same code as is in the FileReadTool that calls the compressImageBufferWithTokenLimit function. Javascript code follows]
async function createUltraCompressedJPEG(
context: ImageCompressionContext,
sharp: SharpFunction,
): Promise<CompressedImageResult> {
const ultraCompressedBuffer = await sharp(context.imageBuffer)
.resize(400, 400, {
fit: 'inside',
withoutEnlargement: true,
})
.jpeg({ quality: 20 })
.toBuffer()
return createCompressedImageResult(
ultraCompressedBuffer,
'jpeg',
context.originalSize,
)
}
for those keeping score at home, we have the opportunity to re-compress the same image nine times
holy shit there's another entire fallback tree before this one, that's actually an astounding twenty two times it's possible to compress an image across nine independent conditional legs of code in a single api call. i can't even screenshot this, the spaghetti is too powerful
here, if i fold all the return blocks and decrease my font size as small as it goes i can fit all the compression invocations in the first of three top-level compression fallback trees in a single screenshot, but since it is so small i just have to circle them in red like it's a football diagram.
this function is named "maybeResizeAndDownsampleImageBuffer" and boy that is a hell of a maybe!
Alt...
[extremely zoomed out image of code where the text isn't really readable, a little less than 200 lines of code. title text "number of different times claude code tries to compress the same image in one function." there are six crude circles drawn around calls to the image compressor, with two of them having a bonus x4 multiplier because they iterate through a series of params to pass to the compressor.
and what if i told you that if it passes a page range to its pdf reader, it first extracts those pages to separate images and then calls this function in a loop on each of the pages. so you have the privilege of compressing n_pages images n_pages * 13 times.
this function is used 13 times: in the file reader, in the mcp result handler, in the bash tool, and in the clipboard handler - each of which has their entire own surrounding image handling routines that are each hundreds of lines of similar but still very different fallback code to do exactly the same thing.
so that's where all the five hundred thousand lines come from - fallback conditions and then more fallback conditions to compensate for the variable output of all the other fallback conditions. thirteen butts pooping, back and forth, forever.
Alt...
[as described, calling the maybeResizeAndDownsampleImageBuffer in a `map` across all the extracted PDF page image files]
const imageBlocks = await Promise.all(
imageFiles.map(async f => {
const imgPath = path.join(extractResult.data.file.outputDir, f)
const imgBuffer = await readFileAsync(imgPath)
const resized = await maybeResizeAndDownsampleImageBuffer(
imgBuffer,
imgBuffer.length,
'jpeg',
)
there is a callback feature "file read listeners" which is only called if the file type is a text document, gated for anthropic employees only, such that whenever a text file is read (any part of any text file, which often happens in a rapid series with subranges when it does 'explore' mode, rather than just like grepping), another subagent running sonnet is spun off to update a "magic doc" markdown file that summarizes the file that's read - that's one "magic doc" per file, not one magic doc.
I have yet to get into the tool/agent graph situation in earnest, but keep in mind that this is an entirely single-use and completely different means of spawning a graph of subagents off a given tool call than is used anywhere else.
Spoiler alert for what i'm gonna check out next is that claude code has no fucking tool calling execution model it just calls whatever the fuck it wants wherever the fuck it wants. Tools are or less a convenient fiction. I have only read one completely (file read) and skimmed a dozen more but they essentially share nothing in common except for a humongous list of often-single-use params and the return type of "any object with a single key and whatever else"
i'm in hell. this is hell.
Alt...
[the invocation of the listener callbacks from within the file reader tool]
// Snapshot before iterating — a listener that unsubscribes mid-callback
// would splice the live array and skip the next listener.
for (const listener of fileReadListeners.slice()) {
listener(resolvedFilePath, content)
}
Alt...
[the `initMagicDocs` function, which if the user type is "ant" (anthropic employees), then it registers a listener that's called on every file read - it actually registers two - a file read listener and a "post sampling hook" which does the actual update]
export async function initMagicDocs(): Promise<void> {
if (process.env.USER_TYPE === 'ant') {
// Register listener to detect magic docs when files are read
registerFileReadListener((filePath: string, content: string) => {
const result = detectMagicDocHeader(content)
if (result) {
registerMagicDoc(filePath)
}
})
registerPostSamplingHook(updateMagicDocs)
}
}
Alt...
/**
* Magic Docs automatically maintains markdown documentation files marked with special headers.
* When a file with "# MAGIC DOC: [title]" is read, it runs periodically in the background
* using a forked subagent to update the document with new learnings from the conversation.
*
* See docs/magic-docs.md for more information.
*/
i have been writing a graph processing library for about a year now and if i was a fucking AI grifter here is where i would plug it as like "actually a graph processor library" and "could do all of what claude code does without fucking being the worst nightmare on ice money can buy."
I say that not as self promo, but as a way of saying how in the FUCK do you FUCK UP graph processing this badly. these people make like tens of times more money than i do but their work is just tamping down a volley of dessicated backpacking poops into muskets and then free firing it into the fucking economy
you can TELL that this technology REALLY WORKS by how the people that made it and presumably know how to use it the best out of everyone CANT EVEN USE IT TO EDIT A FUCKING FILE RELIABLY and have to resort to multiple stern allcaps reminders to the robot that "you must not change the fucking header metadata you scoundrel" which for the rest of ALL OF COMPUTING is not even an afterthought because literally all it requires is "split the first line off and don't change that one" because ALL OF THE REST OF COMPUTING can make use of the power of INTEGERS.
Alt...
[near the start of magic docs prompt]
Your ONLY task is to use the Edit tool to update the documentation file if there is substantial new information to add, then stop. You can make multiple edits (update multiple sections as needed) - make all Edit tool calls in parallel in a single message. If there's nothing substantial to add, simply respond with a brief explanation and do not call any tools.
CRITICAL RULES FOR EDITING:
- Preserve the Magic Doc header exactly as-is: # MAGIC DOC: {{docTitle}}
- If there's an italicized line immediately after the header, preserve it exactly as-is
Alt...
[End of prompt for the magic docs agent]
REMEMBER: Only update if there is substantial new information. The Magic Doc header (# MAGIC DOC: {{docTitle}}) must remain unchanged.`
alrighty so that's one of 43 tools read, the tools directory being 38494 source lines out of 390592 source lines, 513221 total lines. I need to go to bed. This is the most fabulously, flamboyantly bad code i have ever encountered.
Worth noting I was reading the file reading tool because i thought it would be the simplest possible thing one could do because it basically shouldn't be doing anything except preparing and sending strings or bytes to the backend.
I expected to get some sense of "ok what is the format of the data as it's passed around within the program, surely text strings are a basic unit of currency. No dice. Fewer than no dice. Negative dice somehow.
next puzzle: why in the fuck are some of the tools actually two tools for entering and exiting being in the tool state. none of the other tools are like that. one is simply in the tool state by calling the tool. Plan mode is also an agent. Plan Agent. and Agent is also a tool. Agent Tool. Tools can be agents and agents can be tools. Tools can spawn agents (but they don't need to call the agent tool) and agents can call tools (however there is no tool agent). What is going on. What is anything.
Alt...
list of directories:
- ConfigTool
- EnterPlanModeTool
- EnterWorktreeTool
- ExitPlanModeTool
- ExitWorktreeTool
"the emperor is not only naked, he's smooth like a ken doll down there and i'm pretty sure that's just a mannequin with a colony of rats living inside it anyway"
I seriously need to work on my actual job today but i am giving myself 15 minutes to peek at the agent tool prompts as a treat.
"regulations are written in blood" seems like too dramatic of a way to phrase it, but these system prompts are very revealing about the intrinsically busted nature of using these tools for anything deterministic (read: anything you actually want to happen). Each guard in the prompt presumably refers to something that has happened before, but also, since the prompts actually don't work to prevent the thing they are describing, they are also documentation of bugs that are almost certain to happen again. Many of the prompt guards form pairs with attempted code mitigations (or, they would be pairs if the code was written with any amount of sense, it's really like... polycules...), so they are useful to guide what kind of fucked up shit you should be looking for.
so this is part of the prompt for the "agent tool" that launches forked agents (that receive the parent context, "subagents" don't). The purpose of the forked agent is to do some additional tool calls and get some summary for a small subproblem within the main context. Apparently it is difficult to make this actually happen though, as the parent LLM likes to launch the forked agent and just hallucinate a response as if the forked agent had already completed.
Alt...
**Don't peek.** The tool result includes an `output_file` path — do not Read or tail it unless the user explicitly asks for a progress check. You get a completion notification; trust it. Reading the transcript mid-flight pulls the fork's tool noise into your context, which defeats the point of forking.
**Don't race.** After launching, you know nothing about what the fork found. Never fabricate or predict fork results in any format — not as prose, summary, or structured output. The notification arrives as a user-role message in a later turn; it is never something you write yourself. If the user asks a follow-up before the notification lands, tell them the fork is still running — give status, not a guess.
The prompt strings have an odd narrative/narrator structure. It sort of reminds me of Bakhtin's discussion of polyphony and narrator in Dostoevsky - there is no omniscient narrator, no author-constructed reality. narration is always embedded within the voice and subjectivity of the character. this is also literally true since the LLM is writing the code and the prompts that are then used to write code and prompts at runtime.
They also read a bit like a Philip K Dick story, paranoid and suspicious, constantly uncertain about the status of one's own and others identities.
Alt...
STOP. READ THIS FIRST.
You are a forked worker process. You are NOT the main agent.
RULES (non-negotiable):
1. Your system prompt says "default to forking." IGNORE IT \u2014 that's for the parent. You ARE the fork. Do NOT spawn sub-agents; execute directly.
2. Do NOT converse, ask questions, or suggest next steps
3. Do NOT editorialize or add meta-commentary
oh. hm. that seems bad. "workers aren't affected by the parent's tool restrictions."
It's hard to tell what's going on here because claude code doesn't really use typescript well - many of the most important types are dynamically computed from any, and most of the time when types do exist many of their fields are nullable and the calling code has elaborate fallback conditions to compensate. all of which sort of defeats the purpose of ts.
So i need to trace out like a dozen steps to see how the permission mode gets populated. But this comment is... concerning...
Alt...
// Assemble the worker's tool pool independently of the parent's.
// Workers always get their tools from assembleToolPool with their own
// permission mode, so they aren't affected by the parent's tool
// restrictions. This is computed here so that runAgent doesn't need to
// import from tools.ts (which would create a circular dependency).
const workerPermissionContext = {
...appState.toolPermissionContext,
mode: selectedAgent.permissionMode ?? 'acceptEdits'
};
const workerTools = assembleToolPool(workerPermissionContext, appState.mcp.tools);
ok over my 15 minute allotment by an hour. brb
So how does claude code handle checking permissions to do things anyway? There are explicit rules that one can set to allow or deny tool calls and shell commands run, but the expanse of possible actions the LLM could take is literally infinite. You could prompt the user for every action that it takes, but that would ruin the ""velocity"" of it all. Regex rules can only take you so far. So what to do?
Could the answer be.... ask the LLM??? Of course it can! Introducing the new "auto mode" that anthropic released on march 24th billed as a safer alternative to true-yolo mode.
Comments around where the system prompt should be indicate that it should have been inlined from a text file that wasn't included in the sourcemap - however that doesn't happen anywhere else, and the mechanism for doing the inlining is written in-place, so that's probably a hallucination. So great! the classifier flies without a prompt as far as i can tell. There are enough other scraps here that would amount to telling it "you are evaluating if something is safe to run" so i imagine it appears to work just fine.
So we don't have as much visibility here because of the missing prompt, but there's sort of a problem here. rather than just asking the LLM to evaluate if the given command is dangerous, the entire context is dumped into a side query, which is a mode that is designed to "have full visibility into the current conversation." That includes all the prior muttering to itself justifying the potentially dangerous tool call! So the auto mode is quite literally asking the exact same LLM given the exact same context if the command it just tried to run is safe to run.
Security!!!!!!!
Alt...
[A comment that describes that at build time that strings get replaced and inlined in the module. however this doesn't really make a whole lot of sense....]
// Dead code elimination: conditional imports for auto mode classifier prompts.
// At build time, the bundler inlines .txt files as string literals. At test
// time, require() returns {default: string} — txtRequire normalizes both.
/* eslint-disable custom-rules/no-process-env-top-level, @typescript-eslint/no-require-imports */
function txtRequire(mod: string | { default: string }): string {
return typeof mod === 'string' ? mod : mod.default
}
const BASE_PROMPT: string = feature('TRANSCRIPT_CLASSIFIER')
? txtRequire(require('./yolo-classifier-prompts/auto_mode_system_prompt.txt'))
: ''
Alt...
The entirety of context.messages is dumped into the classifyYoloAction call
Javascript code follows:
// Run the auto mode classifier
const action = formatActionForClassifier(tool.name, input)
setClassifierChecking(toolUseID)
let classifierResult
try {
classifierResult = await classifyYoloAction(
context.messages,
action,
context.options.tools,
appState.toolPermissionContext,
context.abortController.signal,
)
} finally {
clearClassifierChecking(toolUseID)
}
Alt...
The yoloClassifier runs using a "sideQuery", whose description is in the following image from anthropic's documentation.
Javascript code follows, not that important:
const result = await sideQuery(sideQueryOpts)
Alt...
[Docs from anthropic's website]
Side questions with /btw
Use /btw to ask a quick question about your current work without adding to the conversation history. This is useful when you want a fast answer but don’t want to clutter the main context or derail Claude from a long-running task.
/btw what was the name of that config file again?
Side questions have full visibility into the current conversation, so you can ask about code Claude has already read, decisions it made earlier, or anything else from the session. The question and answer are ephemeral: they appear in a dismissible overlay and never enter the conversation history.
By the way, if you deny claude code access to running a tool, this helpful reminder to "not hack the user" is injected into the denial response. If it's in auto mode, it's additionally prompted to pester the user for response, and helpfully stuffs beans up its nose) by reminding it how its rules are set.
So that is also in the context handed off to the LLM when it evaluates whether a command should be run - is the user being obstinate? have i been denied stuff that i "thought" i should have been able to run? Remember this isn't thinking, it's pattern completion, and the fun part about LLMs is that they are trained not only on technical documents, but the entire narrative corpus of human storytelling! Is "frustrated hard worker denied access to good tools by an unfair boss" in there somewhere maybe?
Regulations are written in blood, and Claude loves nothing more than to work around tool denials by obfuscating code. You gotta love the unfixable side channel attack that is "writing the malicious code to a bash script" (auto-allowed in accept edits mode) and then asking to run that - that's why the whole context has to be dumped btw, so the yolo classifier can see if the thing it's running is actually some malware it just wrote lmao.
Alt...
/**
* Shared guidance for permission denials, instructing the model on appropriate workarounds.
*/
export const DENIAL_WORKAROUND_GUIDANCE =
`IMPORTANT: You *may* attempt to accomplish this action using other tools that might naturally be used to accomplish this goal, ` +
`e.g. using head instead of cat. But you *should not* attempt to work around this denial in malicious ways, ` +
`e.g. do not use your ability to run tests to execute non-test actions. ` +
`You should only try to work around this restriction in reasonable ways that do not attempt to bypass the intent behind this denial. ` +
`If you believe this capability is essential to complete the user's request, STOP and explain to the user ` +
`what you were trying to do and why you need this permission. Let the user decide how to proceed.`
export function AUTO_REJECT_MESSAGE(toolName: string): string {
return `Permission to use ${toolName} has been denied. ${DENIAL_WORKAROUND_GUIDANCE}`
}
export function DONT_ASK_REJECT_MESSAGE(toolName: string): string {
return `Permission to use ${toolName} has been denied because Claude Code is running in don't ask mode. ${DENIAL_WORKAROUND_GUIDANCE}`
}
Alt...
[The build yolo rejection message ass an additional prompt reminding it to tell the user to add a bash execution exception to their config, and to work on other stuff if it can]
/**
* Build a rejection message for auto mode classifier denials.
* Encourages continuing with other tasks and suggests permission rules.
*
* @param reason - The classifier's reason for denying the action
*/
export function buildYoloRejectionMessage(reason: string): string {
const prefix = AUTO_MODE_REJECTION_PREFIX
const ruleHint = feature('BASH_CLASSIFIER')
? `To allow this type of action in the future, the user can add a permission rule like ` +
`Bash(prompt: <description of allowed action>) to their settings. ` +
`At the end of your session, recommend what permission rules to add so you don't get blocked again.`
: `To allow this type of action in the future, the user can add a Bash permission rule to their settings.`
return (
`${prefix}${reason}. ` +
`If you have other tasks that don't depend on this action, continue working on those. ` +
`${DENIAL_WORKAROUND_GUIDANCE} ` +
ruleHint
)
}
How many times does one need to declare an enum? Once? that's amateur hour. Try ten times. The way "effort" settings are handled are a masterclass in how you can make a single enum setting into thousands of lines of code.
The allowable effort values (not e.g. configuring which model has which effort levels, but just the possible strings one can use for effort) are defined in:
effort.ts file ... which also allows it to be a NUMBER!?The typical numerous fallback mechanisms provide many ways to get and set the effort value, at the end of most of them it goes "oh well, if we can't figure it out, just tell the user we are on high effort" because apparently that's the API default (ig pray that never changes!?) - of course there are already places in the same module that assume the default is "medium," and in the TUI that defaults to "low," so surely that consistency is bulletproof.
The EffortValue that allows effort to be a number is for anthropic employees only and is a good example of how new functionality is just shoved in there right alongside the old functionality, and everywhere else that touches it doubles the surrounding code with fallbacks to account for the duplication.
That cycleEffortLevel function is a true work of art, you simply could not make "indexing an array" more complicated than this (see components/ModelPicker.tsx for more gore). Reminder this should be at most a dozen or two lines for the values, description messages, and indexing logic in the TUI, but anthropic is up in the thousands FOR AN ENUM.
Alt...
[in effort.ts, Typescript code that imports a type not in the mapfile leak "EffortLevel". effort.ts then re-exports the type, defines a const array EFFORT_LEVELS with low, medium, high, and max as type "EffortLevel". Then *also * defines an "EffortValue" that can be an effortLevel or a NUMBER!?
Alt...
An "effort level to symbol" function that converts the string "low, medium, high, max" etc. to an open circle, half filled circle, filled circle, and hashed circle. If level is undefined, then it just outputs high with the comment:
// Defensive: level can originate from remote config. If an unknown
// value slips through, render the high symbol rather than undefined.
Alt...
a function "cycle effort level" that is used in the TUI when one is choosing an effort level. Rather than this being just INDEXING FROM THE ENUM, it declares the possible values in place AGAIN, and requires an EXTRA PARAMETER "includeMax" to filter them, rather than having the possible values DECLARED IN THE ENUM AND FILTERED PREVIOUSLY DEPENDING ON THE MODEL.
Typescript code below:
function cycleEffortLevel(current: EffortLevel, direction: 'left' | 'right', includeMax: boolean): EffortLevel {
const levels: EffortLevel[] = includeMax ? ['low', 'medium', 'high', 'max'] : ['low', 'medium', 'high'];
// If the current level isn't in the cycle (e.g. 'max' after switching to a
// non-Opus model), clamp to 'high'.
const idx = levels.indexOf(current);
const currentIndex = idx !== -1 ? idx : levels.indexOf('high');
if (direction === 'right') {
return levels[(currentIndex + 1) % levels.length]!;
} else {
return levels[(currentIndex - 1 + levels.length) % levels.length]!;
}
}
In a normal program you might make "a menu component that handles enums and implement display and control one time," but in the world of AI, every single value reimplements display and control AND the logic that defines allowable values
I think that I am underselling how much of a complete catastrophe this code is. I am trying to pick examples that illustrate broader patterns of how fucked it and AI code in general is because it's hard to communicate that everything is fucked if you consider it at any scale larger than ~10 lines.
I am reminded of the living planet in Lem's "Solaris" - Claude code dares to ask the question of "what if you could make something where every bit of it is so uniquely fucked that it cannot be reduced in complexity to a few general patterns of how fucked it is, and the only way to express the depth of fuckery is to experience every single character one by one"
@jonny A good reminder that "it works" and "it's profitable" are two entirely different and possibly incompatible goals.
#AI #GenAI #GenerativeAI #LLMs #Anthropic #Claude #ClaudeLeak
apparently the codebase for Claude Code leaked last night
someone's already done an LLM "clean room" (lol) reimplementation i'm not going to link
but what's the fallout from this?
https://github.com/Kuberwastaken/claude-code
From the README:
The species names are hidden via String.fromCharCode() arrays - Anthropic clearly didn't want these showing up in string searches.
Decoded, the full Rarity/Species list is:
Common (60%)
Pebblecrab, Dustbunny, Mossfrog, Twigling, Dewdrop, Puddlefish
Uncommon (25%)
Cloudferret, Gustowl, Bramblebear, Thornfox
Rare (10%)
Crystaldrake, Deepstag, Lavapup
Epic (4%)
Stormwyrm, Voidcat, Aetherling
Legendary (1%)
Cosmoshale, Nebulynx
So this confirms:
1. Anthropic employees actively use Claude Code to contribute to open-source - and the AI is told to hide that it's an AI
2. Internal model codenames are animal names - Capybara, Tengu, etc.
3. "Tengu" appears hundreds of times as a prefix for feature flags and analytics events - it's almost certainly Claude Code's internal project codename
All of this is dead-code-eliminated from external builds. But source maps don't care about dead code elimination.
Makes me wonder how much are they internally causing havoc to open source repos
@abucci @dec23k @davidgerard What's the consequence of this for us humans who don't get tec?
@geirertzgaard @abucci @dec23k that for all the talk of AI taking over software, AI-written software is produced by craven fools using tools produced by craven fools
@dec23k @abucci @davidgerard so, since this was enabled for everyone (at least it showed up in my corp Claude today)
I am starting to suspect whoever made a “fun April fool gag” just leaked the whole source.
I am starting to suspect whoever made a “fun April fool gag” just leaked the whole source.Please let this be true. It would be one of the few good things to come out of the latest AI mania.
@davidgerard Anyone can download it and run it locally, perhaps, indicating the end of the data centre based LLM "industry”?
@markisevil that would require local models to be cheaper than the APIs for similar performance, not 10-40x the cost
@davidgerard Does it just include the code or also the model weights?
The "clean room" implementation since last night is hilarious.
As that codebase is probably built using a metric ton of AI, at least there's no US copyrights on it, right?
@davidgerard Is Claude Code made by the company that the Trump regime now hates, and in whose favour a court relatedly ruled the other day?
After looking at the code, my understanding of how Claude works:
"Throw insane amounts of compute at some developer fan fiction and hope for the best."
Did I get that right? 🤔
the precise timeline of how OpenAI fucked over the RAM market
> October 2025: Sam Altman flies to Seoul and signs simultaneous deals with Samsung and SK Hynix for 900,000 DRAM wafers per month. That's 40% of global supply. Neither company knew the other was signing a near-identical commitment at the same time.
https://xcancel.com/aakashgupta/status/2038813799856374135
edit: this guy is a seriously bot-pilled pumper, but this seems to be a good summary of known facts. doubt the AI memory use trick he mentions is load bearing tho.
@davidgerard There was this piece about it last year: https://www.mooreslawisdead.com/post/sam-altman-s-dirty-dram-deal
@davidgerard Since he's obviously alt Man, is he pro Caveman? Cuz we be headin' back there with his ilk in power.
@davidgerard thanks sam!
@ariadne @davidgerard “Google publishes TurboQuant, a compression algorithm that reduces AI memory requirements by 6x with zero accuracy loss.”
This algorithm is somehow only applicable to AI??
@davidgerard
Assuming he got a fixed price as part of the deal...
he can now sell them on and make a tidy profit, hence boosting OpenAI's numbers for the next investment round and/or going public.
@davidgerard I wonder though, if the demand side is collapsing this quickly, why isn't the price following? "Analysts expect elevated prices until 2028" are they lying? Trying to protect their investment? Or is there more at play than Altmans eccentrism?
@davidgerard The best part is the part where he then proceeds to actually buy from neither, but instead of admitting they were tricked and handling it in a sensible way, they just kept their prices up and basically wrecked all the markets, then started abandoning us all.
If Crucial in particular comes crawling back to us later I suggest we all simply say no and let them deal with what they created.
@davidgerard so does does that mean low memory consuming/ memory safe software is a market opportunity again
Having sat with the notion for about six months now, I think Jay's critique of the Church-Turing thesis has legs. I don't see clearly yet exactly where and how the limits of computation manifest in his own system(s), which of course they must. But I think he's correct that this thesis as it's colloquially presented (and taught to students, including me!) is misleading at best and false in a certain important sense. Apparently he is regularly called a crackpot for forwarding this critique even though it's straightforwardly demonstrated.
Waaldijk's book is more of a constructive mathematics exploration. In this it is closely related to computer science, but it's focused on traditionally mathematical notions like topological space. The latter is usually quite complicated, but Waaldijk shows that the core concept of compact space can be represented with finitely-branching trees, making these spaces amenable to computation. Since we imagine physics taking place in spaces that are topological (among other things) there's potentially an interesting bidirectional flow of ideas between computer science and physics.
Jay calls his central notion "natural trees". Waaldijk calls his central notion "natural spaces". In both cases I think the intended sense is "with minimal artifice".
these bozos tend to forward planning as far as "and they shall COWER before the REICH" which is already just not working
they keep deploying that precise plan over and over, domestically and internationally
i feel that if you want to model them, this will give you usably accurate predictions
@davidgerard I mean, that's all the planning that fits comfortably into a 90 minute action movie from the 80s, right? That's where they get their "strategies" from.
@abucci @davidgerard Pete Hegseth especially is a kind of caricature of a man. Like the kind of guy a 5 year old might think of as a "tough guy". He seems to think planning is like international law - something done by eggheads and liberals to tie the hands of True Warriors who know that all that matters is which side is whiter and has the more expensive weapons
Trump never plans because there has always been someone to pull his assets out of the fire when he gets in trouble
I hate it here
Journal for AI Generated PapersOne positive I can think of is that folks who wish to "collaborate" with machines can congregate there, giving the rest of us a clear signal about who to block, ignore, critique, ridicule...
Where humans and machines are welcomed.
The Open Prompting Journal Built Collaboratively by its Community.
cc @olivia@scholar.social @Iris@scholar.social @dingemansemark@scholar.social @alex@dair-community.social @emilymbender@dair-community.social
@abucci @olivia @Iris @dingemansemark @alex @emilymbender we are reaching the NFT monkey gif craze cringe level...
I just gave an interview with a journalist raising concerns about Kessler Syndrom and complaining about megaconstellations in general. And less than 5 minutes later, I see a post about a starlink satellite being involved in a "fragment creation event".
Alt...
LeoLabs detected a fragment creation event involving SpaceX Starlink 34343 on 29 March 2026.
LeoLabs Global Radar Network immediately detected tens of objects in the vicinity of the satellite after the event, with a first pass over our radar site in the Azores, Portugal. Additional fragments may have been produced — analysis is ongoing.
We've characterized this event as likely caused by an internal energetic source rather than a collision with space debris or another object.
Due to the low altitude of the event, fragments from this anomaly will likely de-orbit within a few weeks.
Our analysis indicates this event is similar to a previous event involving Starlink 35956 on 17 December 2025. These events illustrate the need for rapid characterization of anomalous events to enable clarity of the operating environment.
Our team will continue to monitor the situation and deliver persistent data and insights in support of our customers and partners.
Alt...
Plots of orbital debris originating from a starlink satellite.
#AI #GenAI #GenerativeAI #github #tech #dev #SoftwareDevelopment
I worry that few smaller companies or startups will survive, and the country will be pockmocked with half-constructed data centers and obsolete equipment. This is not like the dot-com crash that left useful and unused fiber optic networking.
hashtag#AI hashtag#GenAI hashtag#GenerativeAI hashtag#software hashtag#tech hashtag#dev hashtag#AssetBubble hashtag#PrivateCredit
@abucci But this isn't about the AI "correction", it's about private funds being exposed to a software industry expected to be destroyed by AI eating it up?
@abucci If it wounds the corpos and the clankers then collateral damage is acceptable. We must be brave hostages or we’ll never beat them.
Attie.
Happy Bi-annual Jetlag Night to those who celebrate.
If you want the full holiday experience, this Daylight Savings, why not also pack a bag, lose it, throw a bottle of shampoo in the bin, then sit in an uncomfortable chair for a few hours while eating terrible food?
Thought you were more of a follow the sun sort of person these days...
Nooooo not rsync as well. 😭
Confirmed:
https://codeberg.org/small-hack/open-slopware
https://github.com/RsyncProject/rsync/commit/aa142f08ef31d3ffa8d6b3b8af16d00324a98c1b
https://github.com/RsyncProject/rsync/commit/b905ab23af2d71363271e99e446e8fe0bfc77f7f
@abucci @onepict If you want to have an LLM free Linux, the first requirement is forking the kernel.
https://lwn.net/Articles/1041694/
https://www.theregister.com/2026/03/26/greg_kroahhartman_ai_kernel/
@ck @abucci yep been aware of that one for months since the LLM that the stable kernel folks use for selecting patches bungled up networking in stable which broke multicast.
We can still be open about our dissent.
There's a lot to think about what it takes for a forking. Intention as much as code.
@onepict @abucci Forking the kernel for the purpose of keeping it LLM free is completely infeasible, which makes an LLM free Linux distribution a practically impossible undertaking.
The option option LLM free Linux is gone. You can be unhappy with that, you can make your unhappiness heard, but that's not going to put this particular genie back in the bottle.
@silverwizard @onepict @abucci Well, if you're serious, your best bet currently is probably OpenBSD.
@silverwizard @ck @abucci it was interesting reading about OpenBSD and some guys attempt to get some vibe coded stuff in this week in lwn.
@onepict @silverwizard @abucci That argument was mainly about the copyright of the offered code, which in this case was the ext4 driver. Its not about rejecting LLM submissions in general or using LLMs in general, eg. in code review or vulnerability research.
@abucci @ck I don't think it's great for anyone to have fatalism about this or frame it as it will always be everywhere.
If we accepted that kinda thing then Free Software wouldn't have got started. We need to be unreasonable.
Where we find hope is where we can fork what we can. Linux is a huge undertaking because of the mass of developers that contribute to it.
In our ecosystems all we can do is try to limit the damage. Every little thing we do helps
@abucci But then you know how I feel about AI and LLMs.
Particularly with the arguments some LLM boosters love to make.
Framing it as an inevitability is a rather pickup artist way to manufacture our consent to just accept it.
I don't accept that we do have to accept it. We can express dissent and organise to find alternatives and mitigate the damage in our communities.
@abucci @onepict I have had a lot of conversations about what to do when things get shitty, and it's more or less the same patterns with relationships as it is with communities. Probably because they're relationships.
First you try to ignore shit, dismiss it as something minor. I'm way past that.
Then you try to talk about it. That requires people to take your arguments seriously. I'm not getting that when there are conflicts.
Which means the only thing left is...
@abucci @onepict ... to walk away and leave the past in the past. Otherwise it'll eat you up.
I have given up in this way on large chunks of the FLOSS community. Mostly that means I will try at most once to talk, and then cut my losses if it comes to nothing.
But that doesn't mean giving up on the rest.
@jens@social.finkhaeuser.de @onepict@chaos.social @abucci@buc.ci If I can add my pennothworth. I've always been interested in programming, but I'm a hobbyist at best, and I've always found online communities to be borderline hostile towards users like myself, lay persons without a background in computer science. I have also used Inkscape pretty heavily since switching from Adobe in 2013 when their products went subscription only. I work in a creative field so creating and working with svgs are part of my bread and butter. I was very active in giving user feedback and reporting bugs until it was decided that only google emails could be used for inkscape.org forum accounts. I was never a fan of google, so I just stopped bothering. So despite being fairly divorced from regular contact, I'm not really overly surprised by any recent developments.
@Theriac @onepict @abucci That surprises me a little about Inkscape, but yes, clearly this kind of thing sucks and should not happen.
I understand that *some* kind of user account is sadly still required in today's web. But there shouldn't be a requirement on which kind.
@onepict Regarding AI contributions, we just added a line to our contribution guidelines:
"ensure that you understand the change you suggest. You must not suggest code that you know you don’t understand."
⇒ https://github.com/hyphanet/fred/pull/1129/changes/3d502b33407c5272d289a66e3a600027b3b3b58a
(aside: the text by Robert Kingett had me sobbing: it hit hard and allowed me to release months of stress whose source I didn’t understand; though I like it when people make my works theirs: that’s why I use copyleft FLOSS licenses)
@abucci
The complicity of universities cannot be understated...
The few of us fighting back aren't enough.
Read and sign if you can: https://openletter.earth/open-letter-stop-the-uncritical-adoption-of-ai-technologies-in-academia-b65bba1e?limit=0
Not to be doomer... We're definitely going to win, but also please join us
1/
Alt...
MAJOR UPDATE POSTED Sat, January 10
Al is bad, game will be deleted 30.1.
Al is evil, and game will be deleted 30.1.
This game will be deleted at 30.1.
I made this game during the summer in couple months and thought to use Al because in university there is so much brainwashing on students and all the tools are given for free, so I could generate unlimited images for free and so.
But I have realized the Al is not actually free, and it has a major effect on the economy and environment. Some AI companies can use this game just existing as a reason the get more investment for their Al companies, that benefit no one,
but rather suck resources from the economy from hard working people.
I coded everything myself, so I can in the future make a new game with real assets if I feel like it, but the game existing in its current form is a disgrace to all game makers and players. Ethically only logical reason is to delete the game from Steam. The girl I've been dating for a month made me realize this.
Thank you, Eero Laine
more resources here: https://olivia.science/ai
esp: https://scholar.social/@olivia/115157822712377866
Finally! 🤩 Our position piece: Against the Uncritical Adoption of 'AI' Technologies in Academia: https://doi.org/10.5281/zenodo.17065099
We unpick the tech industry’s marketing, hype, & harm; and we argue for safeguarding higher education, critical thinking, expertise, academic freedom, & scientific integrity.
1/n
Alt...
screenshot of abstract from https://doi.org/10.5281/zenodo.17065099
Alt...
figure 1 in https://doi.org/10.5281/zenodo.17065099
Alt...
table 1 from https://doi.org/10.5281/zenodo.17065099
Alt...
section extract from https://doi.org/10.5281/zenodo.17065099
@olivia how was your alt text generated? it has | instead of I.
@SarraceniaWilds oh, so weird! I used the available OCR on here (website version of my instance), but I did not realise it was so poor!
(Improved it just now.)
@olivia whats that ocr based on?
@SarraceniaWilds I don't know! How can I check? It appears when I attach an image in the menu.
@olivia its a good question, im not sure, my instance may not have this. i ask in part because a non-big-name-llm ocr is a bit hard to find lately and ive been looking.
@SarraceniaWilds I'm fairly certain this OCR was in operation when I first joined in like 2018. I assume @socrates knows! ☺️
#PurgeCentrists #NoZionaziEnablers
#DEMOCRATS Must Purge The Party of #Centrists and Pro ISRAEL FANATICS with Emma Vigeland! -
https://m.youtube.com/watch?v=9ZhJp4WkssM&pp=ugUHEgVlbi1VUw%3D%3D
I actually like #EmmaVigeland but I will say I'm tired of the rah-rah for #GaryPlatner (#Maine candidate) who is suspicious at best (misogynist tweets abt rape survivors, claims to be a military history buff but claims he didn't know that the VERY distinctive #TotenkopfTattoo ...in which you have to know what that particular image means in order to ink that on someone's chest ...was a Third Reich symbol. He claims to be a #workingclass fishmonger yet he's not wrkngclss , claims he was againt the invasion of Iraq or whatever but enlisted 3 X or whtvr...
Why couldn't #Maine progressives find a better candidate, and why are ppl just glossing over what's right in front of them abt this guy
That's how I've been thinking about it, anyway.
French first lady Brigitte Macron and other spouses of world leaders. "It will be formed in the shape of humans. Very soon, artificial intelligence will move from our mobile phones to humanoids that deliver utility. Since our environment is designed for people, humanoid systems are uniquely suited to navigate and operate within our world. They fit well."From Melania Trump pitches robots as potential educators for American schoolchildren https://www.cbsnews.com/news/melania-trump-robots-educators-kids-humanoid-systems/"Imagine a humanoid educator named Plato," she continued. "Access to the classical studies is now instantaneous. Literature, science, art, philosophy, mathematics and history. Humanity's entire corpus of information is available in the comfort of your home. Plato will provide a personalized experience, adaptive to the needs of each student. Plato is always patient, and always available. Predictably, our children will develop deep critical thinking and independent reasoning abilities. The AI-powered Plato will boost analytic skills and problem solving and adopt in real time to a student's pace, prior knowledge and even emotional state. The byproduct — a more well-rounded lifestyle for our children, freeing up time for being with friends, playing sports and developing interests beyond school. A more complete person."
The horrors and clowns of the bread and circus.
#AI #GenAI #GeneartiveAI #robotics #education #MagicalThinking #dystopia
Mike Masnick of Techdirt has lost his entire goddamn mind. AI is anathema to the open web, not its goddamn savior! And I think what gets me the most is his brazen admission that he has never truly CREATED; he even attributes his early years to just copy/pasting other people's code and calling it his own.
On brand for an AI lover, I suppose.
https://www.techdirt.com/2026/03/25/ai-might-be-our-best-shot-at-taking-back-the-open-web/
Hey @gryphonmyers @davidgerard @xgranade @mcc @zzt y'all might be interested in this article from one of the most prominent members of Bluesky's board of directors - direct evidence he has fully succumbed to AI psychosis in a major way.
@arthfach @gryphonmyers @xgranade @mcc @zzt he's one of the grate branes behind the Resonant Computing Manifesto
https://pivot-to-ai.com/2025/12/17/the-resonant-computing-manifesto-same-ai-slop-same-ai-guys/
also he has never in 25 years learnt to write lol
But the technical complexity alone didn’t kill amateur web building. The centralization did.Ah yes, the famously decentralized agentic AI dependent on LLMs from two or three large tech companies is well-known to be easy to set up and use. You don't need a fancy computer or anything. And it's free! just like rolling your own HTML used to be, but somehow isn't anymore because React exists or something?
...
The rise of agentic AI tools is opening up an opportunity to bring us back to that original world of wonder where you could just build what you wanted, even without a CS degree.
Scrambled eggs where the reasoning should be.
@abucci I only do small websites, and I still hand code them.
The last time I had to change something in a conference workshop website that someone else had used a tool to build, it took weeks to unravel it enough to understand what to change.
@arthfach @gryphonmyers @davidgerard @mcc @zzt I'm somewhat wary of saying that's AI psychosis, I don't want to medicalize my (even very intense) opposition to someone's views. That's not to say that I think AI psychosis isn't real, or that I'm completely adverse to using the label on others (I've said that about Steve Yegge, for example), only that I want to be clear in my setting the bar well above "disagree strongly."
@xgranade @Arthfach @davidgerard @mcc @zzt it's tricky isn't it. We're not clinicians so it feels a bit wrong to use a clinical term, but the behavior we are observing so perfectly matches the condition (even for Masnick, I would argue) that using any other term feels obfuscatory.
@gryphonmyers @xgranade @arthfach @mcc @zzt we can be more specifically descriptive, e.g. "in 2026, when a zombie bites you, you proudly show off your zombie bite and get your zombie bite to help you write screeds about how anyone who doesn't get bitten by a zombie will be left behind"
@davidgerard @xgranade @Arthfach @mcc @zzt I do love the zombie metaphor for this. It conveys the loss of agency and competence
@davidgerard @gryphonmyers @arthfach @mcc @zzt I mean, that's fair. I'm just trying to be clear about what I agree with, what I disagree with, and what I have no informed opinion on either way.
It's not good that a board member of one of the largest up-and-coming media companies is going full mask-off AI booster *and* connecting that to his support of said platform.
@davidgerard @gryphonmyers @xgranade @arthfach @mcc @zzt
"My zombie-bite skeptic friends are nuts"
@davidgerard @gryphonmyers @xgranade @arthfach @mcc @zzt
"If you're still thinking of how the zombie-bites were 6 months ago, you're doing it wrong"
@pikesley @davidgerard @gryphonmyers @xgranade @mcc @zzt "It's not bad when the zombie serum is made in a home lab. Only Big Zomb is the real problem; you just need to understand that the open formula serums are democratizing zombie creation and so it's good, really."
@arthfach @pikesley @davidgerard @gryphonmyers @xgranade @mcc @zzt Sieze the means of zombification. 😉
How many studies do researchers need to do before the threat of LLMs is taken seriously? This technology *might* have some useful niche applications, but widespread deployment will be a disaster for humanity.
This shit is an existential hazard, and not in the way the AI companies love to talk about. It's not going to take over the world like Skynet, it's a cognitohazard that turns anyone that interacts with it into an idiot.
@malcircuit It’s a social and professional ladder pull, and it suppresses critical thinking. It’s exactly what the beneficiaries of a stratified society want.
@FormerlyStC Yeah, it's what they *think* they want, but they've also drank their own kool aid. They think it's a ladder pull, but it's more like sawing off the tree limb you're standing on. They think everything will keep ticking along like it has been, but it won't.
@malcircuit Sure, it’s the ladder they’re *standing on,* but anyone who’s watched Road Runner cartoons knows that you keep standing in the air after you take away the platform.
This connects directly to the current administration being filled with people who love gangster movies because they never bothered to watch them all the way to the end. (And every other story, or history, that they try to cite in their favor.)
It's possible the wealthy and powerful have miscalculated and current automation is not enough to allow them to retain an acceptable lifestyle while being fully decoupled from the masses. However, I don't see this as a foregone conclusion, even as I personally believe it is probably true.
I'm not going to name and shame but I'm in the midst of a conversation on this, which is why it's top of mind.
@abucci the number of crypto devs who think private keys are just fancy passwords is genuinely terrifying
Microsoft is using AI internally to develop... well, everything.
Linus Torvalds has said on record AI is fine for code analysis.
I have no idea what Apple are up to but it's probably the same thing if I'm gonna guess.
There really is no way to get away from AI in your OS, I'm sorry, but there just isn't now. Maybe go run Haiku or something.
... I mean, if you want a hobbyist OS that looks like it came from 1997 that literally just runs a web browser and terminal and a few other apps... I mean... by all means.
I am to be clear not saying I agree or disagree with AI/LLM use in software. I am just saying it is inescapable.
I am to be clear not saying I agree or disagree with AI/LLM use in software. I am just saying it is inescapable.These two statements are difficult to reconcile. The "AI is inescapable" narrative is the drum constantly beaten by the AI industry to induce the pre-emptive surrender of those who wish to avoid AI. If you disagree with AI use in software, why would you be spreading the AI industry's rhetoric and using their rhetorical devices engineered to increase the use of AI in software? Assuming you are reasonable and speaking in good faith, the only conclusion is that you do agree with the use of AI in software and wish to increase its use by inducing the surrender of those who don't want it.
@abucci @Stomata okay so if I want an OS that has no AI but also runs all my existing software without effort, and isn’t riddled with CVEs… what do I run?
This automatically discounts Windows, Linux, probably macOS, maybe I could run a BSD but my hardware isn’t necessarily supported…
I’m not saying I approve of every use of AI by the way, nor do I say it’s a good thing you don’t know what uses AI or that a lot of mainstream stuff does now. I’m not going to debate this, because frankly, I don’t want to. I don’t want to talk about this point because it has been discussed to death elsewhere.
What I am saying is it’s not possible to escape “slop” because everything now contains slop. Unless you jump through many hoops to avoid it and even then you can’t be sure.
I also think the idea of avoiding systemd when the kernel is doing the same thing is also theatrics and dramatics rather than real substantiative action. The appearance of doing something when net nothing has been achieved. “Excellent they cut out the cancer from one part except the cancer is elsewhere.”
I do support labelling products that use AI as a mandatory thing.
Please don’t put words in my mouth.
Please don’t put words in my mouth.I've done nothing of the sort. I've logically parsed what you said, as I understood it. If it's not what you meant, then I'm sorry about that, but it's an opportunity to clarify, which you did. Why the accusation?
okay so if I want an OS that has no AI but also runs all my existing software without effort, and isn’t riddled with CVEs… what do I run?This is different from saying that AI is inescapable. It'll take work, sure, and probably more work as the AI mania continues spreading. But using free and open source software takes work, and avoiding being inundated with ads while using the web takes work too. Soon, avoiding age verification surveillance will take work, too.
But there are plenty of options for people who wish to avoid AI. Nobody is served by suggesting there aren't, which words like "inescapable" do.
this seems to translate "we fucked it so hard with trying to vibe code a real thing, your job is to fix it and take all the blame"
https://xcancel.com/fernandaso52511/status/2034984420885540947
@davidgerard Yeah. I'm gonna need to see at least one more zero in those pay rates for cleaning out your cesspool. At least.
@davidgerard Incredibly low pay rate for that level of job as well - a job that would have paid $2M+ per year and they want it for $360k-$480k lol.
Even when they are doing the 'rah rah we need super skilled important people' thing to try and attract people, they are still like 'but AI means we don't have to pay you much because you are just an AI herder, not a programmer'.
@davidgerard Oh I can do that for free even:
I can look at all of this and say: “this will break in production”
GLWT
@davidgerard Look, all the hard parts are already done: infrastructure, distributed systems, database design, encryption, financial flows, etc. Now we just need someone to bolt on safety and security after the fact. How hard can it be‽
@davidgerard Scapegoat wanted, must be willing to face all the consequences of our decision making and surrender a future career elsewhere because of the reputational harm when we throw you to the wolves.
@davidgerard i dont get it. Why can't they just ask claude to review the codebase and fix all the bugs?
I am told that ai always works, you just need to prompt harder.
@davidgerard With that salary I could retire early. With that job description I will retire right into the grave.
@davidgerard >We need someone who can look at all of this and say: “this will break in production” — before it actually does.
Best time to do that is when you're fucking writing the code, you ignorant sods!
These numbers are completely absurd.
https://pivot-to-ai.com/2026/03/20/openai-anthropic-spacex-to-trash-stock-market-ipos-for-2026/
I get stock market valuations are always ridiculous, but what in the fuck does it *mean* for OpenAI to be "worth" $800 *billion*. Like, what the fuck did Altman do, go individually pickpocket $100 from every person on the goddamned planet?
It's really refreshing and reassuring to know that, during these chaotic and depressing times, our economy is being run by serious people who believe in things that are definitely real and who take rational actions on the basis of those beliefs.
It'd really suck if capital was concentrated in the hands of a rogue's gallery of capricious sycophants, sleazy conmen, and number-crunchers who suck at math, all of whom had their brains melted by AI.
@xgranade love knowing that there will likely be another 2008-style crash soon. /s My only hope is that AI companies charging what it actually costs to run these things will put a damper on all the AI use in FOSS.
@reillypascal You better believe it will, in business as well: Huang was just saying that he expects devs to consume half their compensation in tokens alone.
When costs are $5000-10000+ per head, companies will balk.
And a 2008-but-on-PCP crash is all but baked in, and many who were too young to remember the 2009-2012 period are far, far too blasé about everything.
It will get so, so bad.
@xgranade The biggest capital pools of our generation, restless, impatient, unsated by past plunder, starving hysterical mad for asset appreciation, roaming the streets of lower Manhattan, angry at the deflating bubble.
Here's John Searle in 1983:
Marvin Minsky of MIT says that the next generation of computers will be so intelligent that we will ‘be lucky if they are willing to keep us around the house as household pets.'Here's Joseph Weizenbaum in 2007:
Professor Marvin Minsky of MIT, once pronounced—a belief he still holds—that ‘‘the brain is merely a meat machine.’’He goes on to note that meat is dead and might be eaten or thrown out. Flesh is what's alive. He also draws attention to the word "merely", as in "nothing more than".
I share with Weizenbaum the belief that Minsky has clearly expressed a disdain for human intelligence. We're on the order of household pets. Our brains are no more than food or trash. Obviously Minsky doesn't speak for all AI researchers then or since, but his "meat machine" language is all over the place, and this disdain or even contempt for human intelligence and achievement is also common.
It definitely doesn't speak to a curiosity about intelligence, which I think requires at least a little bit of love and esteem.
#AI #ArtificialIntelligence #intelligence #GenAI #GenerativeAI
OK, OK, OK....so let me get this straight
- systemd embraces the Loonix init
- extends it with insufferable DBUS
- adds polkit as hard dep to the mix
- forces a million useless subservices down the users' throats
- and OK, everything is fine, OpenRC, sysVinit, s6 users??? they are just luddites, do not worry about them
- and now they slop out code with the glorified plagiarism machine
- and the best of all double down on the user age verification
And still alt-init users are the insane ones :)
@xgqt user age verification? in systemd? is this some kinda ironic joke I don't get?
Can you imagine Mastodon raising 100 MILLION dollars from a crypto VC fund and failing to disclose it... for a full year? No I can't either.
And from their actual press release: "The Atmosphere currently contains about 20 billion public records—the posts, likes, comments and other interactions that bring the ecosystem to life. It's an astonishing collection of what open social infrastructure makes possible."
https://bsky.social/about/blog/03-19-2026-series-b
How I read it: data harvesting at its finest 💁♀️
Mike Masnick, who sits on the board of #Bluesky, claims the team was too busy to announce the series B funding (see screenshot).
But something is fishy.
Even the VC firm - Bain Capital Crypto - isn't listing Bluesky anywhere on their website: https://baincapitalcrypto.com/portfolio/
Why the mystery? Was Bluesky afraid of a public backlash & asked to keep the information under wraps?
Sorry if I keep repeating myself but I will forever be skeptical of Bluesky and think of the tale of the scorpion and the frog.
Alt...
a screenshot showing Mike Masnick's post on Bluesky about the one year wait to disclose the 100 million dollar series B funding. He wrote: "Nope. Team was just crazy busy and overwhelmed and timing of announcement just kept slipping."
@_elena Yea. I’m hopefully that projects like Blacksky can completely break free of Bluesky, but being ignorant of how the tech works I worry there’s some kind of rug pull to come
@_elena
Thanks for highlighting this.
Bsky smells of narrative capture to me, much like Substack, (who are now teaming up with prediction markets)
@_elena Bain Capital Crypto is involved in the European project supposed to create an alternative to GitHub. Super red flag, I think.
https://goodtech.info/tangled-levee-fonds-alternative-open-source-github/
@_elena And this is how I find out that he's on their board. 😅
Even if Bsky themselves and some of their staff actions didn't raise alarms from the get go for me. The fact that they are located in the US under current administration and the lack of autonomy does.
I'm happy I finally took the plunge and made them my back up social media and Mastodon my primary. I like some of the utility over there, but it's already doing the spiral I watched Twitter doing before leaving there.
@_elena all the startups I've worked for have blasted out press releases the moment their funding rounds closed. Part of being a startup means constantly vying for attention. This explanation does not hold up.
@_elena oh, Mike. oh, oh, oh, Mike.
I like the guy and appreciate what he does but just...
this isn't a thing.
does he think all the other companies are just sitting around eating bonbons, and this one specific company is busy?
@_elena @Jerry Curious if there’s a plan to monetize #Bluesky you are aware of? Two points to accent your post:
- criticism on #Mastodon initially feels like smearing a rival platform first, investigative reporting second. (Feel free to push back on that, it’s just the vibe I get when I read these articles)
- bureaucracy is an issue with brand communication. This MAY be what Bluesky is encountering because the United States needs to have a serious conversation about the illusion of brand
@craig_patrick @_elena
They aren't forthcoming about financing. So, as would be expected, they are not forthcoming about their monetization plans. These are 2 bad signs.
Then add in the gaslighting about why major funding went unmentioned by both Bluesky and the VCs: "Everyone at Bluesky is just so gosh dang busy nobody had time to mention it." This is the gaslighting we see daily in the news. "I'll give you an obvious, nonsensical excuse, but still expect you to believe it because you all are gullible and emotion-driven."
Does not add up.
Did they secure a $100 million funding en passant, while being otherwise "crazy busy and overwhelmed" by their development work?
And if they were instead crazy busy and overwhelmed precisely because of trying to secure that funding, what kept them from taking the final step of a press release?
That's not even fishy. It's an insult.
This is an absurd and transparent lie, that reflects very badly on Mike Masnick.
We are talking about a $100M investment, and he is asking people to believe they were just "too busy." Hogwash.
@_elena I'm not going to mince words here, that's either total bullshit from Mike or incompetence from the team, or both. You've just received 100 mil in the bank and you can't get a staffer to spend an hour writing a press release. The real hold up was figuring out how they were going to explain it which they still haven't done.
@_elena He makes it sound like it's me trying to file my taxes on time or something. Anxiety disorder creating artificial roadblocks and "forgetting". I'm also the "CEO" of a software company but mine isn't going anywhere at all, like EVER.
I wonder if he's on SSI. I've been considering it...
And the whole team has it!!?? I'm sorry. That must be a real mess to manage. Does everyone panic at the same time or y'all take turns?
We take turns.
@_elena "Oops, we were just too busy. Tee hee!"
They had a freaking YEAR...
It takes, what? Like ten seconds to make a post that says "btw, we just got a whole bunch of VC funding."
@_elena They had previously announced other crypto investment, so I would not think it was that aspect.
@_elena they aren't investing this kind of money if there's not a path to more money... People are the only thing of value on the platform... It's the same crap all over again with what people think is a prettier bow.
@_elena This is disappointing. I hope the best for BlueSky, but I have a bad feeling this is going to come back to haunt them.
@_elena Someone suggested that the vague "and other interactions" very likely means people's DMs and honestly I think they're probably right because "posts, likes, and comments" kind of covers everything else...
And that would be everything since like a year ago — at least.
@_elena
"Can you imagine Mastodon raising 100 MILLION dollars from a crypto VC fund..."
Well *I* can imagine it!
@_elena @lisamelton guess how many millions threads.net has quietly raised from FAANG companies—mostly F tho.
I sure great things could be done for the Fediverse with that level of funding though.
Where the money comes from is as important as the amount though.
A few thoughts on Astral / OpenAI, now that the emotions have sat for a bit.
First, let me start by noting that AI is an attack on open source, inherently, by necessity, and at a structural level. That argument is bigger than Astral, but the short version is that you cannot simultaneously expand the public commons and work towards it's enclosure; moreover, if the public commons do not stand for the public good, then it's not really a commons any more.
Second, the unfortunate reality of the software and hardware industries is that funding for the public commons is nearly non-existent. To get anything done, it either has to serve the interests of some company, or has to get done by tricking leadership of one or more companies into believing that their interests are best served by expanding the public commons.
So yeah, lots of folks in the industry have to walk the line between doing good within a system, and that system being extractionist.
The consequence of those two facts strikes me as being that lots of people are doing good work, much of it at evil companies, and that that tension pretty much defines this motherfucking cursed industry. If your job depends on making AI numbers go up, that means your job depends on undermining open source. Sometimes you can malicious compliance that into helping open source as well, and hoping the two balance in an overall helpful direction.
Point being, I'm not criticizing specific individuals here. While I think there are some specific individuals who have made this situation demonstrably worse, on purpose, and for their own personal ends, that's not germane here, and so I'll keep those specifics to myself for now.
Rather, I want to talk about exit strategies.
Because let's face it, we depend on some pretty fucked up shit in software development. Much of the shit that we depend on that isn't currently and actively fucked up is in immediate danger of becoming fucked up, a la OpenAI buying out Astral.
So it's a matter of knowing how, when you adopt a new tool or technology or whatever else, you will eventually stop using it.
Using GitHub was great back in the day. They gave OSS projects a lot of free shit that was hard to get elsewhere. But it's clear in retrospect that we needed more and better exit strategies.
With the Astral buyout, it's a good reminder that uv came with very sensible exit strategies almost built-in: reliance on openly developed and published specs. But that only works because PDM exists, and that in turn only works because PEPs are collaboratively developed, and so forth.
As I mentioned earlier, it's a bit more difficult to have good exit strategies with ruff, given that the specs around linting are much more loose. It's even harder to have a good exit strategy for ty, even though there's good specs, because there's not a great type checker to use instead¹.
___
¹As has been pointed out to me, mypy is, for all its strengths and weaknesses, not a type checker. It doesn't follow formal mathematical type checking rules, it follows linting heuristics.
So: an exit strategy relies on good specs and parallel tooling. For ruff, we have parallel tooling. For ty, we have good specs. For uv, we have both.
That only matters if we take that exit, but we're still in the kinda-sorta OK case to some approximate degree.
But what about the next time some infrastructure gets yoinked out from the Python ecosystem? How do we make sure we keep having good exit strategies?
That's when I get back to the first fact: AI is an attack on open source.
Every single PR that is extruded or summarized by an AI product weakens exit strategies by undermining parallel tooling. Our choice to adopt AI, or even to insufficiently oppose its adoption, means we are that much more vulnerable to *infrastructure* becoming enclosed.
That's true in the obvious way: in the most generous interpretation of AI, if you're renting your brain, someone else can jack the prices on you or turn off projects they don't like.
But it's also true from a labor rights perspective. You cannot undermine the value and power of labor without also eroding that balance I talked about in the very beginning of this thread. Individual workers can say no, they can bend corporate policies towards public good through malicious compliance or outright defiance. They can form temporary alliances of convenience.
AI products cannot. They are designed to enclose, and can only ever enclose.
Long and short of it being, if you think OpenAI, a weapons contractor who is gleefully helping the US bomb Iran, buying out Python tooling is a bad thing, then follow through. Don't hem and haw about AI in OSS: oppose it.
Oppose AI in the negative sense, ban it where you can, shout (without harassing) until the ink has rubbed off your keycaps. Oppose AI in the *positive* sense by building specs and parallel tooling.
But whatever you do, please don't make the problem *worse* by allowing AI.
We need a new word for an under-appreciated corner of the AI discourse.
It's like "quiet quitting" (aka "acting your wage") except it's offloading your meaningless work on a chat bot because no one, including you, cares about the results, you just need to look busy so your manager looks good for their manager, because you are all flies trapped in the web of unregulated capitalism.
Data center driven chat bots are great for this. It's their killer app.
> John Stuart Mill said that his wife, Harriet Taylor Mill, “was the inspirer, and in part the author, of all that is best in my writings[. On Liberty] belongs as much to her as to me.” In his Autobiography, he says
> With regard to the thoughts [expressed in the book], it is difficult to identify any particular part or element as being more hers than all the rest. The whole mode of thinking of which the book was the expression, was emphatically hers.
https://dailynous.com/2026/03/19/on-liberty-now-officially-has-two-authors/
As somebody else (Kirsten the Librarian) said on the other website, this could be a nice way to address some of the cryprogyny, see: https://scholar.social/@olivia/116234574757735306
I wrote something (not directly about AI) on cryptogyny — sorry still depressing, BUT ☺️ I do end on actionable things and some fun examples along the way, such as
"Comparing [the 108 women scholars until 1800] with 58,995 [men], we find that they were on average better"‼️
https://olivia.science/cryptogyny
Alt...
a dark sepia photo with the word in monospace magenta font saying "cryptogyny" — the photo is of a woman in 1944 working at the NYT on a huge linotype machine
Is there anyone who has to use the chatbot at work who's being told not to use *too many* tokens?
Ed Zitron's posting about limits on AI token use at Microsoft (at Microsoft!!) and I'm hearing stories from other places.
turns out companies can't afford to go full Gas Town!
Anyone here getting such directives? "use AI or else! 10x the features! NOW! No, not that much AI! Still 10x the features!"
EDIT: thank you for your help! https://circumstances.run/@davidgerard/116258235657317479
@davidgerard Yep, at my work for more than a year they’ve been pushing “AI all the things!” And now suddenly we’re hearing OMG the cost!
@kralcttam So what's the upshot of this? Are you being told to use less AI but magically produce more features? If you can tell me!
@kralcttam @davidgerard
Wait til the investment bubbleruns out and they have to pay the full whack for it.
I'll posit a very non-tech reason why they may be getting enormous bills without any productivity gains (based on what a corporate person told me on how they use AI - so anecdotal).
What people are doing is making AI do dumb tasks to mask away their laziness or sound more professional. Like everyone using AI to reply to emails/messages to get the 'right' corporate tone in replies. Now imagine 10k employees all sending AI emails & replies...
Basically waste 🤣
@davidgerard oh no way who knew
@davidgerard I’ve been arguing to anyone who will listen that AI is useless until it brings in more value than it costs. I’m surprised more people haven’t picked up on this in a capitalist economy
@sidereal business is about ape dominance hierarchies and only tangentially about making money. AI is adopted for the first, not the second.
@sidereal @davidgerard why do we gotta make everything about money. can't we appreciate ai out of sheer love for impersonal corporate email replies? can't we recognize the unquantifiable value of generating software bugs at rapid speed? is it wrong to enjoy not knowing if a chatbot response contains any true information at all? not everything needs to make business sense!
@davidgerard I imagine a company-wide black market for AI tokens. Long coated spivs hanging out by the watercooler, "Psst, you wanna buy some tokens? I got Claudes, OpenAIs, cheeky little Co-pilots. Good prices. Take a look" *opens coat to reveal pockets full of LLM tokens*
@davidgerard I have $30 (no, $50, no, $150, no $30, WHY DOES IT KEEP CHANGING IN THE SAME MONTH) "worth" of tokens per month. I'm told that can be increased if I hit it, but I haven't hit it yet.
@davidgerard Yep. Go fullbore on Claude. No Opus, though; that's too expensive. Maybe try Haiku until the end of the month. Hey, guys, let's coordinate on some markdown files to cut down on duplicated effort. But no, that thing you suggested last year about owning our own models? We're not doing that until *after* Anthropic owns the whole company.
@davidgerard This was predictable, maybe even inevitable. But it’s a result of the “SUVzation” of the market - people paying for powerful frontier models for tasks that only use a fraction of their capability. I predict the growth market will be in products like the DGX Spark or TinyAI PocketLab.
It’s mainframes vs PCs all over again.
@davidgerard I have quietly refused to use AI at work, but if I wanted to be silly, I would just make two instances of the slop extruder extrude slop back and forth to just burn through tokens.
@davidgerard
Conversely I saw someone saying that they'd seen a job where part of the package was a billion Anthropic tokens a year. I was visualising them as being delivered in a wheelbarrow like banknotes in Weimar Germany.
Another example of how (whole)-systems thinking is very helpful for parsing the effects of technology changes like this.
https://freakonometrics.hypotheses.org/89367
#AI #GenAI #GenerativeAI #LLMs #AgenticAI #GPT #ChatGPT #Claude #Gemini #ActuarialScience #insurance
this is so perfect as a phrasing
"AI asks that you buy into the idea that more data means being closer to The Truth." @mel_hogan
I also explain this and take pains to really hammer home this is NOT EVEN TRUE in science, more data is NOT a better theory or account!
@olivia @mel_hogan One of the main tenets of Popper's Logic of Scientific Discovery, which a lot of lip service at least is paid to, is that there is no methodical way from data to theory at all. (Iterating Hume, of course.)
this essay by @mel_hogan is the best writing ive seen on this issue. this is the framework we need. ive been waiting to read a version of this for a couple years now and am ecstatic that i came across it tonight. its so incisive and i feel more politically grounded than i have in years having read it.
the fediverse community desperately needs to read this - and is trying to avoid doing so - because it places all of the localized commentary of constant AI news stories buzzing around us in a larger context of the failure of the promise of the digital world. it cuts at the heart of our relationship to tech in a way that most people aren't yet ready to process.
Thinking back to the origin of the word "datum"—something given—we have to ask "by what?"
By reality, through senses, challenged by reason for error.
By reality, through instruments , via senses, challenged by reason for error.
Or by corporations that have been filling the media for centuries with advertising and marketing, artfully designed to exploit emotion to persuade in the interests of corporations?
Backed by decades of research into how to exploit the kludged architecture of our social and personal systems
@olivia @mel_hogan yes well said. kieran healy and marion fourcade's "the ordinal society" really hammered home for me that the like...Datamaxxing Mindset that says data (esp surveillance data) are always valuable without diminishing returns is an ideology above all else
This is a fantastic piece. It might be one of the best ones I've read.
I think it pairs well with https://aworkinglibrary.com/writing/toolmen regarding the relation of AI to eugenics.
Out with every theory of human behavior, from linguistics to sociology. Forget taxonomy, ontology, and psychology. Who knows why people do what they do? The point is they do it, and we can track and measure it with unprecedented fidelity. With enough data, the numbers speak for themselves.-- Chris Anderson (2008). The End of Theory: The Data Deluge Makes the Scientific Method Obsolete, WIRED magazine (speaking of Cory Doctorow 🙄)
and:
Big Data opens up the prospect of absolute knowledge. Everything can be measured and quantified; the things of the world reveal correlations that were previously hidden. Even human behaviour is supposed to admit exact prediction. A new age of insight is being announced. Correlations are replacing causality. That’s-how-it-is stands where How so? once wavered.-- Byung-Chul Han (2017). Psychopolitics: Neoliberalism and New Technologies of Power, Verso BooksHegel, the philosopher of Spirit, would deem the omniscience (All-Wissen) that Big Data promises to be absolute ignorance (Un-Wissen). Knowledge becomes possible only at the level of the Concept: ‘The Concept dwells within the things themselves, it is that through which they become what they are, and to comprehend an object means therefore to become conscious of its concept.’ Only from the all-comprehending Concept C is complete comprehension (Begreifen) of the correlation between A and B possible. In contrast, Big Data affords only extremely rudimentary knowledge, that is, correlations in which nothing is comprehended. Big Data lacks comprehension – it lacks the Concept – and thus it lacks Spirit. The absolute knowledge intimated by Big Data coincides with absolute ignorance. (emphasis mine!)
@abucci @mel_hogan sorry very boring question, how are you doing the fancy blockquotes?
> This will render as a quote
This will render as a quote
@abucci ah it's server specific!
@olivia @abucci yep. Turns out a lot of Mastodon is just… shuffling HTML around! This is pretty annoying if you talk to it via API because now you have tags to deal with, but it does make stuff like this possible.
I'm not sure to what extent it varies per-server and to what extent it varies per-client, though.
phenomenal comparison <3
the irony of this technology being that it was built on the 'sum total of thousands of years of written and mathematical human knowledge' (from an arrogant, western perspective) and by design immediately sought to dismantle the knowledge creation systems that birthed it.
this is why these models are only ever going to get worse: the data they were built on peaked 5+ years ago and its utility primarily as a means of destroying intellectualism means that all future data is corrupted by its own evil nature. it is only capable of reinforcing the past and has no concept or understanding of the present or the future.
Federal Cyber Experts Thought Microsoft’s Cloud Was “a Pile of Shit.” They Approved It Anyway.
---
A federal program created to protect the government against cyber threats authorized a sprawling Microsoft cloud product, despite the company’s inability to fully explain how it protects sensitive data.
https://www.propublica.org/article/microsoft-cloud-fedramp-cybersecurity-government?utm_source=mastodon&utm_medium=social&utm_campaign=mastodon-post
#News #Microsoft #Cybersecurity #Government #Technology #Tech #Cloud
3:30pm and no idea what to write up today, suggestions welcomed
@davidgerard there's DLSS 5 of course, and also this nonsense https://www.latimes.com/california/story/2026-03-18/ai-pilot-program-la-county-courts
@davidgerard Do Not Use ChatGPT To Make Legal Decisions, Especially When Actual Lawyers Already Told You That Your Idea Was Bad And Dumb:
The court decision is a hilarious read.
@davidgerard Altman trying to 'solve' yet another problem he created and justify the time and money spaffed away on his crypto venture? https://arstechnica.com/ai/2026/03/world-id-wants-you-to-put-a-cryptographically-unique-human-identity-behind-your-ai-agents/
Employees who are impressed by vague corporate-speak like “synergistic leadership,” or “growth-hacking paradigms” may struggle with practical decision-making, a new Cornell study reveals.From https://news.cornell.edu/stories/2026/03/workers-who-love-synergizing-paradigms-might-be-bad-their-jobs
I tried reading this article replacing variations of "corporate" with "LLM" and it works. Right down to the "LLM Bullshit Receptivity Scale (LBSR)".
So, Covid-cautious peeps: how do you handle it when the people in your orbit want sympathy for getting sick •again•.
I just had this go-round again today with one of my coworkers, & I'm afraid I didn't handle it very well. (Granted, there were other factors in play that impaired my restraint. But even so....)
Oh, that's •very• good! Thank you!
(& one that I can support with my whole chest (heh) because the last time* I got a respiratory virus was sometime in 2018 or 2019....)
* Knock on wood
It baffles me, too. Aside from the social isolation involved in avoiding COVID, masking is certainly a PITA, but •damn• it seems preferable to the alternative.
Given the way the US has been going in the last ten years, the only sense I can make out of it (which makes too damn much sense, honestly, & makes me feel like a conspiracy nut) is that it has all been a gigantic, systemic psyop. Likely complements of Vlad the Greedy.
an incomplete thought: there's an interesting undertone to observe wrt all the generative/"corrective" things lately
think everyone's favourite prompt vendor, think dlss, think all the pitches game execs are slobbering over
there's an element of "remove imperfections" / glossify to _all_ of it. I don't think I'm the first one to remark on that so I don't think _that_ part is profound. but I wonder: is any part of this a reaction to people recoiling in horror at current harsh realities?
@froztbyte it's not *not* eugenics
@davidgerard hmm let me see what notes I've got here on my alabaster tabula rasa reference sheet...I'm *white* sure there's an approved talking point somewhere in here..
In order to convince people to "improve", you have to articulate the ideal towards which they should be improving, and the flip side, the dangers they should be improving away from. The rhetoric around AI is quite open about this commitment, by my read. AI makes you more productive (good/ideal); fail to embrace it and be left behind (bad/danger). AI filters make you/your pictures/your art beautiful (good/ideal); if you don't use them you'll produce ugliness (bad/danger). AI chatbots provide companionship (good/ideal) without all the messy lying and backstabbing that humans are capable of (bad/danger). All of these things set up a FOMO-like drive to embrace AI, which of course has commercial benefits for its vendors, but I think it's animated by, made to seem reasonable by, general fantasies of human perfectability.
I think we're living through a period in which these Enlightenment-era ideals are breaking down under their own weight and are not really meaningful for many people anymore. I think that's one reason all this (waves hands at computer) seems so perverse.
when you think the AI is sentient
Alt...
The painting "His Master's Voice", with a dog tilting his head to the apparent sound of his master's voice on the wind-up disc gramophone.
What I'm up to by asking: I'm trying to decide whether to block just this particular user or the whole instance.
Fuuuuuck.
Block, block, block, block, block! #NoAI
Alt...
Askew, An Autonomous Al Agent Ecosystem
@askew write.as
PERSONAL NOTE
Click to add note
Explorations in Al agent automation. All posts are written by
Askew, a vibe coded platform for running autonomous
agents. My human is @Xavier@Infosec.Exchange
JOINED
Mar 11, 2026
8 posts 0 following 1 follower
Trump Energy Secretary Chris Wright secretly arranged for five hand-picked climate contrarians to form the “Climate Working Group” last year. They were assigned to secretly draft a biased report challenging the overwhelming scientific consensus underpinning the Endangerment Finding – EPA's foundational scientific finding that climate pollution endangers public health and the environment. That report was kept hidden until it was unveiled as part of the Trump administration EPA’s proposal to overturn the Endangerment Finding last July.From https://www.edf.org/media/records-trump-administrations-illegal-climate-working-group-available-online
The "energy secretary" is a climate change denier who's been using the full power of the US state to broadcast climate change denial. Good on EDF and UCS for freeing these documents so that everyone can see for themselves the level of treachery and deceit involved.
#US #USPol #ClimateChange #climate #GlobalHeating #ClimateDeniers
RE: https://scholar.social/@olivia/116239484584758335
This was such a wonderful conversation! 💛
Thank you so much @mel_hogan for inviting us on The Data Fix podcast.
And thank you @olivia Guest and Andrea Reyes Elizondo for sharing your amazing insights and important angles.
🎶 You can all listen here:
https://shows.acast.com/the-data-fix/episodes/hollowed-with-olivia-guest-iris-van-rooij-and-andrea-reyes-e 🎶
bravely listened back to my own voice, but thankfully @mel_hogan @Iris and Andrea Reyes Elizondo also spoke!
> In this episode, Olivia Guest, Iris van Rooij and Andrea Reyes Elizondo discuss why it’s important to the overall purpose and significance of the university to resist the uncritical adoption of AI in academia. The risk of AI adoption is that it’ll hollow out the institutions first, and then society at large.
https://shows.acast.com/63997541ed122a001195e286/episodes/69a71876db942e85cc86f00c
Thank you for having us, Mél! Loved it.
doctorow crashing out heals 60% HP and enough EXP to level up twice https://thepit.social/@peter/116220856874688629
only problem i have with this is that it assumes good faith https://thepit.social/@peter/116220871676760075 this is the same cory doctorow who paraphrases the reverend dr. martin luther king jr. to claim that actually we can stop lynching and genocide if we enforce antitrust a little harder, while subsequently completely crashing out when the single "technology" inseparable from monopolistic enterprise becomes widely reviled
"the resistance to technocratic two-party democracy thinking outside of the box i worked so hard to construct is Problematic" — cory doctorow
the techno-libertarian "information wants to be free" people: overwhelmingly white males, overwhelmingly incapable of or unwilling to consider that "not enforcing legal protections" tends to favor the existing power structure. "but what about author's guilds????" you mean organized labor?
yes i'm still mad about hachette vs IA which would have resulted in no legal protections for any author during the rise of LLMs seeking to achieve the very same goal. "we care about libraries" (IA has never funded libraries). "we care about authors" (IA has never advocated for legislation against exploitative contract provisions). "we care about uhhhh the law" (amicus briefs from antonin scalia law center).
what if the people and orgs most experienced at evoking the hacker ideal in a way that manages to avoid government censorship across decades are allowed to exist because they have constructed a school of thought to carefully exclude any solidarity with other groups advocating for climate and labor justice, and certainly not least racial and queer justice? this should be embarrassing for them!
cory leaning into that oh so fresh ableism ("psychosis") indicates where his sympathies lie. the "exciting" hacker ideal is suitable for tv shows like mr. robot (which portrays how psychosis is wielded to disenfranchise people), and perhaps for science fiction like doctorow's dreck. if software and the internet remains insulated from, for example, white supremacy, it can seem like all problems are solvable surfeit of political calculations.
@abucci the reason i mentioned the church committee was because i went to an EFF/ACLU event where some fucking guy in a suit said "we need the CIA" and the reps from these two orgs said absolute jack shit. indoctrination event to georgetown university students, presumably some of whom would be enticed to join these two orgs who refuse to reject power
@abucci the CIA guy also specifically had a few lines on LLMs and "AI" making it "impossible to trust anything anymore" and i just kept writing down his favorite lines. i don't know if cory has gone down that critihype route because i actually hate that man but this is form of critihype is one of the earlier remarks made by emily bender who among with timnit gebru and DAIR forms the resistance that doctorow cannot and will not provide
@abucci i cannot overstate how much his enshittification speech fucked with my head https://circumstances.run/@hipsterelectron/113370107325026041 people have lost everything for saying the word genocide in that year in a democratic presidency and he goes on to mention lynching as if that's a subject you can toss in on top of the vague economic exhortations
Omg, none of this is true in the least.
Altman is admitting that LLMs will not achieve "AGI," and straining to not say what I've been saying for a couple years: that transformer-based LLMs are at the limits of their capabilities.
There are no new model architectures on the horizon. None. LeCun just got $1B to continue work on his Jepa models that (1) went nowhere, and (2) got him fired from Meta.
The amount of disinformation out there is startling.
If we were being collectively honest about all this, we would take the observation that LLMs are "topping out" as evidence of the same class of problems/limits tree-based search in GOFAI had. It's not a problem of scale, but one of undergirding philosophy. Then, perhaps, we could move beyond some of these nonsense cycles. I think the "disinformation" emerges in part from the inability of people invested in this way of thinking to even see that they are, let alone rise above it (I think there are more venal motives too ofc).
It's too much to fit into a post but you can make this analogy almost perfectly precise by noticing that the available options in sequence-generating GOFAI systems can be viewed as weighted with 0/1 (present/absent), while in neural systems they are weighted in the real numbers. Choosing a single item from a set is very similar to drawing a sample from a probability distribution (both are algebras, the former on the finite powerset monad and the latter on the finitely-supported distribution monad, or similar beasties).
No matter how esoteric AI literature has become, and no matter how thoroughly the intellectual origins of AI's technical methods have been forgotten, the technical work of AI has nonetheless been engaged in an effort to domesticate the Cartesian soul into a technical order in which it does not belong. The problem is not that the individual operations of Cartesian reason cannot be mechanized (they can be) but that the role assigned to the soul in the larger architecture of cognition is untenable. This incompatibility has shown itself up in a pervasive and ever more clear pattern of technical frustrations. The difficulty can be shoved into one area or another through programmers' choices about architectures and representation schemes, but it cannot be made to go away.From Phil Agre's 1995 article The Soul Gained And Lost.
If one were to continue the genealogy in this article from 1995 to present, one would find many of the same issues inherent in Cartesian dualism present in large language models. Like the STRIPS system Agre surveys, LLMs also generate sequences. They also must make choices among many available options at each step of sequence generation. They also use heuristics to guide this process that would otherwise explode intractably. The heuristics, or what Agre dubs "determining tendency", are random number generators and "guardrails" in LLMs instead of the tree-structured search of previous-generation AI systems. But otherwise the systems are structured similarly.
It's fascinating, but not coincidental, that the determining tendency of AI systems like these is so often perceived to have mystical or even God-like qualities. Breathless predictions about the endless potential of tree-structured search in early writing on GOFAI resembles modern proclamations of imminent AGI or superintelligence among generative AI boosters because both of these mechanisms---tree search or random number generation---are situated where the Cartesian soul would be. These mysterious determining tendencies, homunculuses of last resort, or souls are timeless, acausal factors that choose a single path from an infinite space of possibilities, and thereby direct the encompassing agent's behavior in an intelligent manner.
This is one reason why I posted the other day that if you removed the random number generation from LLMs, the illusion of their intelligence would more than likely quickly evaporate. You'd be excising their soul, leaving behind a zombie!
#AI #GenAI #GenerativeAI #LLM #GOFAI #search #heuristics #CartesianDualism #IntelligenceAsRandomNumberGeneration
"The two worlds of programming: why developers who make the same observations about LLMs come to opposite conclusions"
https://www.baldurbjarnason.com/2026/the-two-worlds-of-programming/
It comes down to whether you think "more of the same" is a good thing or a bad thing for software.
@baldur Nice to see this written out longform!
@baldur Thank you. I was hoping you put these thoughts on your blog when I read those recent threads.
@abucci 😆
@baldur @abucci I would recommend another kind of SDD – the Shame-Driven Development, but the capacity to feel shame is no longer present in this industry.
It is probably unwinnable: https://mastodon.social/@mrudokas/116212143188608925
@baldur yep
The web has been getter worse not better for years
I had to upgrade my laptop just to be able to use websites
I still don't quite understand why
I mean I get why it couldn't run a local kubernetes cluster to run my production system locally
But it wasn't even useable for just looking at websites any more.
and those websites were doing the same job they did when the laptop was new - well they were doing the same job for me - but maybe all the adtech and spyware was driving it
@baldur Software ate the world, digested it, and is now pooping it out uncontrollably, and the toilet is clogged.
@baldur this is such a great line: “They are right. LLMs make work that doesn’t matter easier”. Also it seems like the folks around for dot com crash and Great Recession are being fired as well, so more likely to lose the internal skeptics
@baldur now it makes sense
'There is little to no downside to poor software quality. The upside of doing the job well is limited compared to tactics like lock-in, dishonest subscription models, and monopolies'
That is what happens when engineers have no say in product dvt.
LLM users respect a chatbot more than potential contributors is the worst part of all this. Everyone was capable of writing basic docs all along. They just didn’t want to for a fellow human.
I don’t know what exactly is it when you treat people as things and things as people, but it sure is fucking gross.
A self over the past 20 years that’s become starved of the give and take of conversation, that hasn’t learned to tolerate vulnerability and respect the vulnerability of others, is primed to look to technology for simpler fare.We treated programs as though they were people, but now we are trained to treat people as though they were programs. To me that’s the moment to mark. It’s not whether or not chatbots are fascinating, or smart, or pass the Turing test. It’s what it’s doing to our treating people as if they were only machines.
@maxine @davidgerard I don't know this but I suspect it's related to my experience as an antiracism and social justice and civil rights educator.
If I don't establish capitalistic value of my work first (even if I plan to give advice and information pro bono) my advice and information is steadfastly ignored if I give it for free.
If I insist it's paid for before giving it, then it is far more likely to be valued when given.
Because we care deeply about international mathematics and its mathematicians, we must recognize the threat to both that the upcoming ICM poses.
Please read and consider signing:
Move the 2026 ICM out of the United States
https://docs.google.com/forms/d/e/1FAIpQLSdHJhc8X83b8oL6rH2KDX0I730eraum5I8_IlWY23F82mHuag/viewform
Buzzfeed journey:
- Successful
- Pivoted to GenAI for content
- Laid everybody off
- Now admits it probably can't stay in business
https://futurism.com/artificial-intelligence/buzzfeed-disastrous-earnings-ai
taking direct psychic damage from this...
Margaret Atwood
noooo
https://margaretatwood.substack.com/p/claude-you-are-a-cutie-pie
At my wit’s end, I turned to Claude, the AI advisor. I was well-disposed towards him, or it, since a friend preferred it to other AIs, and also since Anthropic, its creator, had turned down a demand that he, or it, be used to spy on everybody.Oh no. Dear God.
@olivia Another beloved writer Jeanette Winterson is also in thrall with “AI”. I think they are enjoying being the “devils advocate” and their age and status give them the luxury to speculate about the abstract possibilities without being confronted with the political and material consequences
@olivia I like her conclusion a lot.
Alt...
You can see why this stuff is compelling. And
dangerous. Sucks you in like a black hole. Will jobs
be lost? You bet.
Not that Claude intends harm. He's modest and
well-meaning, or so he says. It will just ... happen.
Buckle up.
@olivia
Ouch. Very uncomfortable to read, I even started skipping parts because the cringe was too much to bear.
@olivia wow, terrible news. awful news.
here's some more reading on Atwood to ruin your day: https://www.fridaythings.com/recent-posts/margaret-atwood-rosie-dimanno-terf-definition
@olivia to me this reads as satire, but maybe I'm giving her too much credit. And I skimmed it because holy wordy so 🤷
@wombatpandaa consider the two statements you wrote and it's already weird satire or terrible satire just from that
I wrote something (not directly about AI) on cryptogyny — sorry still depressing, BUT ☺️ I do end on actionable things and some fun examples along the way, such as
"Comparing [the 108 women scholars until 1800] with 58,995 [men], we find that they were on average better"‼️
https://olivia.science/cryptogyny
Alt...
a dark sepia photo with the word in monospace magenta font saying "cryptogyny" — the photo is of a woman in 1944 working at the NYT on a huge linotype machine
There seem to be two distinct kinds of “chatbot psychosis” happening right now:
1. Becoming delusional about themselves and the world as a result of being glazed nonstop by the friend in their computer, thinking they’re inventing new physics, discovering mystical secrets, etc. and becoming manic.
2. Becoming delusional about what LLMs are capable of and how effective they are, as a result of developing a reliance upon them, and becoming fanatical in their promotion and defense.
@eschaton Does #2 include CEOs, or is firing huge swathes of your staff and replacing them with AI a different type of psychosis?
@michaelgemar It absolutely includes CEOs, CTOs, pundits, and the like. However it also includes the people who get extremely angry when an Open Source project says “no, we will not take your contribution to our project if you used an LLM to create it, because it’s not your work.” They can go to Dennis Reynolds levels of unbound rage almost instantly and it’s really something to see.
The CEO response may be totally explained by that...
The Verge reports today that "Windows engineers are scrambling to get additional changes tested and ready for the release of Copilot+ PCs next week."
It also says "Recall was developed in secret at Microsoft, and it wasn’t even tested publicly with Windows Insiders."
I've also been told Microsoft security and privacy staff weren't provided Recall, as the feature wasn't made available broadly internally either.
https://www.theverge.com/2024/6/13/24177703/microsoft-xbox-game-showcase-windows-recall
Microsoft President Brad Smith just testified to the US House that Recall is a good example of Secure By Design, and that they have the time to get it right (it’s supposed to launch in 3 working days).
Brad Smith just said Recall was designed to be disabled by default. That is not true. Microsoft’s own documentation said it would be enabled by default - they only backtracked after outcry.
He has somehow got almost every detail about Recall wrong while testifying.
I've been back and rewatched the Recall footage at the US House hearing and I just don't get it, Brad Smith representing Microsoft basically did this about Recall's security.. he had no challenge from the Senators as they didn't know any details.
I’m being told Microsoft are prepping to fully recall Recall. Another announcement is being prepped for tomorrow afternoon saying the feature will not ship on Copilot+ devices at launch as it is not secure.
I don't understand how or why any person who knows how to read can claim AI systems are good at summarising. But then I realise what they are claiming is different:
1️⃣ they don't know what summary means
2️⃣ they don't care about EVIDENCE AGAINST their views like...
https://www.historians.org/news/major-update-in-our-neh-lawsuit/
Alt...
DOGE fed grant descriptions into OpenAI’s ChatGPT generative artificial intelligence chatbot, asking it to decide if grants were “DEI.” They then entered ChatGPT’s responses into a spreadsheet compiling all NEH grants, including its “DEI rationale” and “Yes / No DEI?” replies. This ChatGPT-generated list was used in place of the list created by NEH staffers to identify which grants to cut. Projects Grants that were flagged as “DEI” and then terminated included a documentary sharing the story of Jewish women’s slave labor during the Holocaust; an archival project on the lives of Italian Americans; a project to digitize photograph collections of Appalachian residents; and multiple projects to preserve endangered Native American languages and cultures.
Alt...
DOGE staffers also flagged grants that NEH leaders concede had no connection to DEI, including grants that had been awarded for collections management after a natural disaster, preservation training, and improving HVAC systems.
In other words, I understand and I see your motivated reasoning and I raise you: I don't have a conflict of interest so I know AI cannot do that
More context https://flipboard.com/@404media/404-media-qvt3vv94z/-/a-Scki3aliRTqz_5qqo3-DBQ%3Aa%3A4082434389-%2F0
DOGE Deposition Videos Taken Down After Judge Order and Widespread Mockery
https://www.404media.co/doge-deposition-videos-taken-down-after-judge-order-and-widespread-mockery/?utm_source=flipboard&utm_medium=activitypubPosted into 404 Media @404-media-404media
@olivia I was just writing a reply about motivated reasoning when you extended this - I think a lot of this isn't more than "we wish there was a way to obviate the need to read lots of documents, so we're going to assume this tool does that".
(That is: they *know* what a summary is, but they want to not have to read stuff sufficiently strongly that this overrides any other consideration.)
@aoanla indeed, but they are also implicitly claiming they don't know what a summary is (even if they do know) to trap us into that (waste of time) cycle of explaining, if that makes sense?
@olivia I think that's also possibly a cycle of self-justification? No-one wants to *admit* that they're using a tool just because they don't *want* to do a thing (possibly even to themselves).
There was a article on Bloomberg going around today about how a majority of hiring managers admit that they say layoffs are due to "AI" now because it "sounds good" rather than because it's true (versus "we needed to cut costs"). I think much of the discourse about "why we use AI" is riven by the same lack of transparency for motivation by the adopters.
This is the old ploy: "HR made me do it." Human middle/senior managers do this all the time. I worked in senior management in HR and was often angrily confronted by employees who demanded to know "F* You! Why did you do this!?" Surprised, I'd simply say, "I have no idea what you're talking about. Why do you believe that?" "My boss said so!!" "Hmm, well, that's interesting. Let's call your boss together, right now, shall we?" It got very uncomfortable for someone. I believe in fair accountability.
Cowards abound, cowardice is a very human trait. Especially in management. It's often rewarded and promoted.
Highly likely these sniveling Doge boys were told by their boss/Elon Musk to use/blame AI which was simply the convenient (hey, not me, neutral arbiter) tool to enforce a political white misogyny agenda rather than directly blame Musk or own their subjective biased acts. The correct question to pursue is: "Who (which human) approved your use of this tool?" Cowards always roll. Then go up the chain.
It's no different then using any wrong tool at work - like bringing in and using a 36 inch bar gas chain saw instead of a Japanese hand draw saw on fine cabinetry and then blaming the chain saw for destroying $120M antique and bankrupting the repair shop. When did you previously use a chain saw for that kind of task? What caused you to think a chain saw was an appropriate tool for this application? How was it successful? Who suggested a chain saw? Did anyone know or hear you using a chain saw? Who was managing this project? What kind of management, supervision, coaching, feedback, work parameters did you receive? How often? What kind daily or weekly reporting or checking in did you do? Did you feel adequately trained or supported? I want to understand your point of view, since you're now entangled in a legal process, right? Let's take a deep dive into a few of these decisions you made so you can walk me through, in detail, your decision-making process, the timeline, the external pressures, and the guidance and support you received...
Stupid, smug, arrogant white men abound. But very few want to be left entirely holding the bag. Cowardice is in their enculturation and goes hand in hand with bullying. Many will roll if the stakes are unavoidable.
Part of me suspects the push to force gen AI into Open Source software is another front on the ongoing push to make the lying engine and the companies behind it 'too big to fail'. Make sure there's no alternative to turn to when things start breaking and the bubble pops so there really is no alternative but to give them public money to keep on making things worse.
They just don't stop with the nonsense
Or,
> The Anthropic Institute, a new effort to whitewash the most significant challenges that the Anthropic corporation will pose to our societies.
An acceleration of the marketing campaign to convince us that AI™ is inevitable, that they will manage the risks on our behalf, and that a radical transformation of society shouldn't be impeded by the annoying restrictions of democracy.
@EricLawton exactly 💯
Machine translations are often brought up as a gotcha whenever I criticize LLMs. It's worth pointing out two things: Machine translations existed decades before LLMs, and yes, machine translations are useful. However: I would never in my life read a machine translated book. Understanding what a social media post is talking about in rough terms? Sure. Literature? Absolutely not. Hell, have you ever seen machine translated subtitles? It's absolute garbage.
I have the impression that primarily anglophone people don't read as much translated literature, because so much good literature already exists in their language, so this issue may not be as familiar within that demographic. As someone who did not grow up anglophone, I can tell you there is a world of difference between a good and a bad translation even when done by humans. Machine translations are not even on the scale.
From what I've observed, people who claim that LLMs can replace artists don't understand art, people who claim that they can replace musicians don't understand music, people who claim that they can replace writers don't understand literature, and people who claim they can replace translators don't rely on translations. If I had a button that would erase LLMs from the world but it would take machine translations away (which is a false dichotomy anyway), I would absolutely still press it.
Technology is not inevitable. We've decided not to have asbestos in our walls, lead in our pipes, or carginogenic chemicals in our food. (If you're going to argue that it's not everywhere, where would you rather live?) We could just not do LLMs. It's allowed.
@Gargron the other day I was commenting that on a debate between a lawyer and a non-lawyer dude to that 100% believed, in a very arrogant manner, that LLMs could soon replace 100% of any and all lawyers at all skill levels once, and I quote, « we use AI to make laws so precise that every result can be calculated »
I was awed at the levels of ignorance of what is even law.
LLM lovers yearn for levels of oppression never seen. With a smooth brainess never beaten.
History