akbir khan
@akbir.bsky.social
dumbest overseer at @anthropic https://www.akbir.dev
We’ve added four new benchmarks to the Epoch AI Benchmarking Hub: Aider Polyglot, WeirdML, Balrog, and Factorio Learning Environment! Before we only featured our own evaluation results, but this new data comes from trusted external leaderboards. And we've got more on the way 🧵
4. Factorio Learning Environment by Jack Hopkins, Märt Bakler , and @akbir.bsky.social This benchmark uses the factory-building game Factorio to test complex, long-term planning, with settings for lab-play (structured tasks) and open-play (unbounded growth). jackhopkins.github.io/factorio-lea...
Factorio Learning Environment
Claude Sonnet 3.5 builds factories
jackhopkins.github.io
New Anthropic blog post: Subtle sabotage in automated researchers. As AI systems increasingly assist with AI research, how do we ensure they're not subtly sabotaging that research? We show that malicious models can undermine ML research tasks in ways that are hard to detect.
control is a complimentary approach to alignment. its really sensible, practical and can be done now, even before systems are superintelligent. youtu.be/6Unxqr50Kqg?...
Controlling powerful AI
YouTube video by Anthropic
youtu.be
This is a crazy paper. Fine-tuning a big GPT-4o on a small amount of insecure code or even "bad numbers" (like 666) makes them misaligned in almost everything else. They are more likely to start offering misinformation, spouting anti-human values, and talk about admiring dictators. Why is unclear.
The fact that Deepseek R1 was released three days /before/ Stargate means these guys stood in front of Trump and said they needed half a trillion dollars while they knew R1 was open source and trained for $5M. Beautiful.
Process based supervision done right, and with pretty CIDs to illustrate :)
David Lindner@davidlindner.bsky.social · 2y agoNew Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward? Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them! Inspired by myopic optimization but better performance – details in🧵
I don’t really have the energy for politics right now. So I will observe without comment: Executive Order 14110 was revoked (Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence)
New randomized, controlled trial by the World Bank of students using GPT-4 as a tutor in Nigeria. Six weeks of after-school AI tutoring = 2 years of typical learning gains, outperforming 80% of other educational interventions. And it helped all students, especially girls who were initially behind.
Generative AI has flaws and biases, and there is a tendency for academics to fix on that (85% of equity LLM papers focus on harms)… …yet in many ways LLMs are uniquely powerful among new technologies for helping people equitably in education and healthcare. We need an urgent focus on how to do that
On one hand, this paper finds adding inference-time compute (like o1 does) improves medical reasoning, which is an important finding suggesting a way to continue to improve AI performance in medicine On the other hand, scientific illustrations are apparently just anime now arxiv.org/pdf/2501.06458
What can AI researchers do *today* that AI developers will find useful for ensuring the safety of future advanced AI systems? To ring in the new year, the Anthropic Alignment Science team is sharing some thoughts on research directions we think are important. alignment.anthropic.com/2025/recomme...
Recommendations for Technical AI Safety Research Directions
alignment.anthropic.com
My hottest take is that nothing makes any sense at all outside of the context of the constantly increasing value of human life, but that increase in value is so invisible (and exists in a world that was built for previous, lower values) that we constantly think the opposite has happened.
Nothing kills my excitement of returning to the US like the response i get from CBP officers.
Felix Hill was such an incredible mentor — and occasional cold water swimming partner — to me. He's a huge part of why I joined DeepMind and how I've come to approach research. Even a month later, it's still hard to believe he's gone.
A brilliant colleague and wonderful soul Felix Hill recently passed away. This was a shock and in an effort to sort some things out, I wrote them down. Maybe this will help someone else, but at the very least it helped me. Rest in peace, Felix, you will be missed. www.janexwang.com/blog/2025/1/...
Felix — Jane X. Wang
From the moment I heard him give a talk, I knew I wanted to work with Felix . His ideas about generalization and situatedness made explicit thoughts that had been swirling around in my head, incohe...
janexwang.com
A few great papers out @ucldark.com this year. To single out two I love, there's the already well-cited paper on Debate by @akbir.bsky.social et al. which got best paper at ICML! [11/17]
just read this cover to cover in like 4 hours. strong recommend.
Alongside our paper, we also recorded a roundtable video featuring four of the paper’s authors discussing the results and their implications in detail:
Alignment faking in large language models
YouTube video by Anthropic
youtube.com
New work from my team at Anthropic in collaboration with Redwood Research. I think this is plausibly the most important AGI safety result of the year. Cross-posting the thread below:
