🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Sonnet 3.7 often realises it's being eval'd on alignment e.g. sandbagging. It starts to give misaligned, self-preserving answers but then "Actually, let me reconsider" and gives "correct"/"aligned" answers *because it thinks it's being tested*

I'm bad at context switching and, when delegating to Cursor agent, often find myself forgetting to check if it's done and not remembering what exactly I asked for. So a wrote a small MCP server to fix that.

AI control evaluations have assumed the red team to have almost "no holds barred", making these evals overly conservative and costly for near-future models. In a new paper, we propose changes to control evals informed by model capabilities to subvert control measures.

If you want to do AI safety work but lack funding, here’s a great opportunity to get it:

The initial explainer post was very weak IMO; but I am very glad to see this expansion

I reimplemented the bliss attractor eval from Claude 4 System Card. It's fascinating how LLMs reliably fall into attractor basins of their pet obsessions, how different these attractors across LLMs, and how they say something non-trivial about LLMs' personalities. 🌀🌀🌀

I am grateful to have worked closely with Tomek Korbak, Mikita Balesni 🇺🇦, Rohin Shah and Vlad Mikulik on this paper, and I am very excited that researchers across many prominent AI institutions collaborated with us and came to consensus around this important direction.

The holy grail of AI safety has always been interpretability. But what if reasoning models just handed it to us in a stroke of serendipity? In our new paper, we argue that the AI community should turn this serendipity into a systematic AI safety agenda!🛡️

When models start reasoning step-by-step, we suddenly get a huge safety gift: a window into their thought process. We could easily lose this if we're not careful. We're publishing a paper urging frontier labs: please don't train away this monitorability. Authored and endorsed

I'm not much of a model internals interpretability hater, and am excited for more people to work on it (both near-term applications like detecting when models are evaluation-aware, and ambitious interp / decompiling models). But reading CoT is so much more useful right now!