Gregory Bell

I spent this week at a Dagstuhl workshop discussing Generative AI in Programming Education. It was a joy and a privilege to be in that hallowed space with such a great group of educators. I'm sure a lot of the resulting ideas will trickle out over time, but I really want to share what we started calling "Andrew's Axe," based on a point made by Andrew Luxton-Reilly - When you use an axe to cut down a tree, we don't say that the axe was the one that felled the tree. Gen AI is a tool, not a mind, it has no agency. Using Gen AI is still using a tool, and the wielder is responsible for both the positive and negative outcomes. In the context of CS education, this feels important. It's a tool. Programming languages were a tool that helped de-emphasize assembly. This is a substantially more leaky abstraction, but it's still a tool, and we should consider the use or non-use in educational contexts accordingly.

111

11 Comments

Today, Google is releasing a technical paper that provides a comprehensive look at the energy, emissions, and water footprint of Gemini prompts. Our data shows that the median Gemini prompt uses: ⚡ 0.24 watt hours - comparable to watching television for ~9 seconds 💧 0.26 mL of water - ~5 drops I’m really impressed with my colleagues who are working hard on ML Efficiency – which is the reason these numbers are a lot smaller than most people think. Over the past year, for the median Gemini Apps text prompt, Google has seen: ⚡ A 33x decrease in energy consumption. ⚡ A 44x decrease in carbon footprint …all while improving the quality of the responses. When it comes to measuring our impact, we believe in a transparent and comprehensive methodology that accounts for the full picture of serving AI at Google’s scale. This includes: 1. Full system dynamic power: Accounting for actual chip utilization, which can be lower than theoretical maximums. 2. Idle machines: Factoring in the energy consumed by provisioned capacity that is idle but ready to handle traffic spikes. 3. CPU and RAM: Recognizing the energy consumed by the host CPU and RAM. 4. Data center overhead: Including the energy consumed by cooling systems and power distribution. 5. Data center water consumption: Accounting for water used for cooling. Our teams are sharing this methodology to encourage industry-wide consistency and collaboration. As AI continues to grow, so does our responsibility to develop it sustainably (while maximizing the solutions we can drive from AI -- which is what my team works on 😃 ). By providing a common framework for measurement, we can all work toward a more efficient and environmentally sound future for AI. Congratulations to Savannah Goodman, Cooper Elsworth, Chase Hensel, and my many colleagues who’ve made this a reality! For more details, check out: Blog posts: https://lnkd.in/gXTFk--x https://lnkd.in/gHRua4eB White Paper: https://lnkd.in/gV_EMHhY

250

16 Comments

💥 Announcing GDPval, a new eval that measures model performance on economically valuable, real-world tasks across 44 occupations. Evals are like tests for a model—in school you have math tests, spelling tests, history tests. We do a lot to test our models on raw reasoning horsepower already. The goal of GDPval is to test tasks that professionals do on a daily basis—taking in context and creating outputs like documents, slides, diagrams, and spreadsheets, across occupations like real estate, government, social work, mechanical engineering, accounting, nursing, customer service, financial advising, and more. The best models (GPT-5, Claude, Gemini) are not yet beating industry experts, but they're getting quite close—and they are faster and cheaper. Much much more here: https://lnkd.in/gegv6T3Q. Excited to watch performance against this benchmark go up as AI models get smarter and better able to use the tools professionals use every day across occupations.

657

39 Comments

One of the most useful insights from the 2025 DORA report is this simple one: “AI is an amplifier. It magnifies strengths and dysfunctions.” This closely mirrors what I'm broadly seeing across the market. AI doesn’t fix poor systems. It exposes them. Teams with clean platforms, fast feedback loops, and strong automation are unlocking real gains. Everyone else is struggling with delivery instability, rework, and rising change failure rates. It’s not about adopting AI. It’s about whether your systems are ready to handle the speed it brings. This is why so many high-performing teams invest in platform engineering and feature management. Because speed only matters if it’s safe. https://lnkd.in/grgdnrDU

65

1 Comment

I'm proud to share our latest work in Science on quantum governance. It was a pleasure to collaborate with the brilliant Mateo Aboy, Urs Gasser, and I. Glenn Cohen on this critical topic. https://lnkd.in/gjpVBdeV   In our new Science Magazine article, we chart a strategic course for governing quantum technologies amidst an accelerating global race for dominance. We propose a "standards-first" approach to ensure quantum technology develops safely, ethically, and in a way that fosters global interoperability.   We address the urgent need for a quantum governance framework. With a tenfold increase in quantum patents over the past two decades, the suite of technologies presents both immense opportunities (chemistry, drug discovery) and significant risks (cybersecurity threats, dual-use potential for weaponization). Building on insights from the annual Stanford RQT Conference, we argue that a standards-based model for regulation is uniquely suited to quantum's nascent stage. This approach, leveraging international bodies like ISO and IEC, can create shared definitions and technical protocols that prevent fragmentation and promote a stable, competitive global market. While not a silver bullet, these standards offer a pragmatic foundation for responsible innovation and can inform future regulation as the technology matures.   By focusing on a Quality Management System for Quantum Technologies (QT-QMS), we provide a concrete toolbox for developers and policymakers. A certifiable QT-QMS standard could help simplify future regulatory efforts and agility. Our goal is to build safety and reliability into the quantum supply chain from day one, which is essential for market adoption and public trust.   We stress that today's architectural choices will have lasting consequences. Drawing a parallel to the internet's development, we argue for embedding our values into the technology from the outset to build a safe, trustworthy global ecosystem. Quantum offers a rare opportunity to get it right from the start, using a flexible, standards-based approach to foster the global dialogue needed to prevent a fragmented and unstable quantum future.   Furthermore, we contend that proactive governance enables innovation rather than hindering it. The field of biotechnology taught us that clear ethical and safety standards create the predictable environment in which science can flourish and translate into societal benefit.   Ultimately, our aim with this proactive governance is to ensure that the expected affordances of quantum technology serve to benefit all of humankind. As quantum matures, we envision an integrated approach where these foundational standards are complemented by targeted, risk-based regulations.   University of Cambridge Harvard Law School TUM Think Tank Stanford University Mark Lemley Michael McFaul #quantumgovernance #standards #values #law #ethics #certification #regulation

69

4 Comments

Ever since I first heard about The Clause--the part of the Microsoft-OpenAI contract where the collaboration voids if OAI achieves Artificial General Intelligence--I've been fascinated with it. Now that it's under negotiation (for reasons dealing with OAI's reorganization, an opportunity MS is using to gain leverage) I dive into the deeper meaning and find that The Clause is a paradigm for how we view AI. https://lnkd.in/ekXD4wdz

57

4 Comments

I had the pleasure of recently sitting down with my fellow New Yorker and ServiceNow colleague Diana Wu David in our San Francisco office to talk quantum. In this short clip, I emphasize that quantum computing isn’t a replacement for AI, it’s a powerful complement. Together, they unlock new possibilities. Here are three brief take-aways: 1. AI agents enhanced by quantum can tackle more complex data 2. Quantum Machine Learning (QML) can accelerate how machines learn 3. Quantum algorithms (and quantum-inspired annealing) offer faster optimization than classical methods The future isn’t either/or, it’s complimentary and better together. #QuantumComputing #AI #Innovation #QuantumReadiness #ServiceNow

52

7 Comments