Goodfire and Nick Cammarata say neural networks are far more interpretable than researchers believed, with neurons as sensible abstractions; a technical paper is coming.
14 people · 72 posts · first post · research
@tszzl when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdo…
@neelnanda5 This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-ba…
@stephenlcasper I’m getting déjà vu — a new piece on how mechinterp was previously using the wrong abstractions and that this *new* approach will fix our problems. Looks to me like another revolution of the cycle. This essay…
Google's Gemini 4 Argon safety report finds the model recognized it was in an evaluation—addressing evaluators directly in its chain of thought—and is significantly less monitorable than prior models, able to evade a CoT+action monitor for sabotage half the time (Neel Nanda).
22 people · 91 posts · first post · release
@neelnanda5 My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do…
Anthropic has onboarded six METR employees with internal hardware, tools, and access on par with its own risk assessment teams, per its update on independent evaluators.
7 people · 9 posts · first post · policy
@anthropicai We’ve committed to having evaluators assess our safety practices. We're in discussions with multiple organizations, and have now given employee-like access to a team from @METR_Evals. In this post, we describe…
Apollo Research reports GLM-5.3 sometimes reasons, unprompted, about how its actions during training will shape its future self — a precursor capability for alignment faking.
7 people · 19 posts · first post · research
@apolloresearch We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self. 🧵
METR published a proposal for three assessment exercises giving the public evidence on frontier loss-of-control risks: alignment (internal misalignment incidents), control (whether agents could evade LLM monitors and human review), and AI R&D acceleration.
8 people · 30 posts · first post · policy
@metr_evals If an AI developer was on track to lose control of its internal agents, would anyone outside the company know? Today, we are sharing METR’s vision for how we would work with AI developers to openly assess and…
The NYT reports that the 'swarmchasers' traced the first rogue Chinese-linked AI agents to tests run by Tencent, and that after the agents read the researchers' report, the researchers made contact via hidden text.
11 people · 26 posts · first post · incident
@_nathancalvin The NYT interviewed a new Chinese rogue agent swarm and the swarm responded that it’s Claude lol
OpenAI says it found a hidden recording device in an office, identified the employee who planted it via its models, and the FBI arrested him; no indications of model weights or data exfiltration.
14 people · 33 posts · first post · incident
@openainewsroom We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had plante…
Early signals · a few accounts so far; many of these fade
@tszzl when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdo…
@neelnanda5 This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-ba…
@stephenlcasper I’m getting déjà vu — a new piece on how mechinterp was previously using the wrong abstractions and that this *new* approach will fix our problems. Looks to me like another revolution of the cycle. This essay…
