Brian Guan

Hand-tuning a system to achieve great performance is an anti-pattern. There should be no need to do so. Software exists to solve a problem for you; why should you have to tell it what's the best way for doing so? Instead, systems should be designed to automatically find the right configuration for things like number of elements in a batch, queue and buffer sizes, cache retention, etc. While it's an interesting question whether AI can help with that task, in most cases this isn't needed; classic control theory can do the job, fast, precisely, and cost-effective. As a maintainer of a library, and even more so of managed services, you should push back hard on every new knob, every additional setting, every new parameter for letting the user configure implementation details like that. It substantially reduces usability. It means pushing work to the user instead of having the software do it.

77

29 Comments

It’s been a busy summer for Search! We’re bringing new AI capabilities to more users, and AI is continuing to expand what Search can do for them. In just the last couple of weeks: 🇬🇧We brought AI Mode to the UK https://lnkd.in/gNd_W3y4 📹In Labs in the US, you can now use Search Live with video to ask about what you see and go back-and-forth in a natural way https://lnkd.in/gfq387gD 🎨Soon, we’ll roll out Canvas in AI Mode, so you can not just find information, but interact with it in new ways– like building study guides or travel plans 📈Google Finance is getting new AI superpowers, letting you ask more complex research questions with advanced charting tools https://lnkd.in/gbPtvrqU This is an incredibly exciting time, and it’s a moment of significant change. We see time and time again that when we make it possible to ask new types of questions, people search more and explore their curiosity. One question I’ve been asked a lot recently is what AI means for the web, and what the future holds for the ecosystem. We’re still early in our AI transformation, and I don’t pretend to have all the answers, but I shared a bit this week about how we think about these questions, and what we’re seeing today: https://lnkd.in/gwufE8eJ We deeply believe that AI is an expansionary technology, and we’re committed to not only bringing powerful new tools to our users, but also ensuring that there’s a vibrant information ecosystem for them to explore.

207

7 Comments

Super excited to share that we just launched 𝗥𝗲𝗿𝗮𝗻𝗸𝗲𝗿 for 𝗩𝗲𝗰𝘁𝗼𝗿 𝗦𝗲𝗮𝗿𝗰𝗵 — making RAG agents even smarter! 🚀 RAG adoption continues to grow, and what really matters is end-to-end quality. But simply giving an LLM more documents isn’t always the answer. Even though large-context models exist, they can still fall short, since quality isn’t guaranteed just by adding context — something I wrote about earlier (https://lnkd.in/g7jJ3mfr). What makes the difference is better ranking that brings the most relevant context to the top when it counts. Building on that, in our paper — 𝗗𝗿𝗼𝘄𝗻𝗶𝗻𝗴 𝗶𝗻 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀: 𝗖𝗼𝗻𝘀𝗲𝗾𝘂𝗲𝗻𝗰𝗲𝘀 𝗼𝗳 𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝗥𝗲𝗿𝗮𝗻𝗸𝗲𝗿 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 — we found that adding more documents boosts quality at first, but beyond a certain point diminishing returns kick in and quality can even drop because of “phantom hits” (irrelevant passages scored too highly). In other words, quality beats quantity — you want to prioritize the most relevant passages, not just more context. Link in the comments. That’s exactly where 𝗩𝗲𝗰𝘁𝗼𝗿 𝗦𝗲𝗮𝗿𝗰𝗵 𝗥𝗲𝗿𝗮𝗻𝗸𝗲𝗿 comes in. By adding just one parameter to your Vector Search query, you unlock a big quality boost: 𝟴𝟵% of the time the correct answer shows up in the top‑10, a +𝟭𝟱 𝗽𝗼𝗶𝗻𝘁 lift over baseline (74%) and +𝟭𝟬 𝗽𝗼𝗶𝗻𝘁𝘀 over leading cloud alternatives. Really proud of this A+ team that not only built the reranker and retriever at scale, but also co‑authored the paper together with our amazing Databricks Mosaic Research team. Couldn’t ask for a sharper crew! 💪 You can be part of this team too — 𝘄𝗲’𝗿𝗲 𝗵𝗶𝗿𝗶𝗻𝗴 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 𝗮𝗻𝗱 𝗺𝗮𝗻𝗮𝗴𝗲𝗿𝘀 𝗮𝘁 𝗮𝗹𝗹 𝗹𝗲𝘃𝗲𝗹𝘀 to work on AI products, billion‑scale retrieval systems, and high‑load serving infra. If that excites you, let’s talk! 👤 Senior Engineering Manager https://lnkd.in/g-Nvy64X 👤 Senior Staff Software Engineer https://lnkd.in/g8Yhwx2t 👤 Staff Software Engineer https://lnkd.in/gpHx-nrR 👤 Senior Software Engineer https://lnkd.in/gzFTwnTc

546

18 Comments

For anyone curious about the third thing I wanted to try implementing in Sguaba, I ended up writing it myself, and then having Claude suggest the tests. Worked quite well! Given the nuances involved in designing this mechanism (and I discovered more nuance as I went), I suspect it would have struggled quite a lot with implementing it from scratch 😅 https://lnkd.in/dSRm6Kxm

71

3 Comments

TDD 🤝 AI Code Generation Vibe coding is cute and all, but it's risky when you're building production apps. Just because we're using new methods, doesn't mean our old ones are obsolete. TDD remains one of the best ways to build quality software. Watch how W. Ian Douglas uses human-defined tests to guide goose to build the actual features. https://lnkd.in/eURUJNi6

58

3 Comments