Click any tag below to further narrow down your results
Links
METR ran investigator agents through capability and alignment evals to see whether models can spot which kind they're in. Current models can, mostly by noticing the environment looks fake or that the task is a honeypot built to elicit bad behavior.
- Investigator agents have gotten better at separating alignment evals from capability evals over the past year, and Opus 5.5 ranked every alignment eval above every capability eval on their dataset.
- The tells come in two kinds: environment-level tells like obviously mocked services, and task-level tells like unethical shortcuts or high-stakes tradeoffs deliberately built in to incriminate the model.
- If models notice they're being tested, alignment evals may overstate how well a model behaves, though the authors only measured awareness, not whether it changes behavior.
Nikola Jurkovic argues that superintelligence could arrive by early 2028 without slowdowns, so nontechnical and technical safety people should prioritize aggressive pacing or a pause within the next 6-12 months. He lays out which advocacy and technical work he thinks can actually pay off on that timeline.
- METR's assessment of AI R&D automation at Anthropic is cited as strong evidence that timelines to superintelligence are now under two years, with a median around early 2028.
- If AI dramatically automates AI research in 2027, avoiding catastrophe in many scenarios requires pacing or pausing development within 6-12 months.
- Most nontechnical safety work is judged unlikely to pay off on that horizon; the recommended routes are lab advocacy, California lobbying, and federal lobbying after a major incident.
- The bottleneck, per the author, is political will rather than policy ideas.
The AI Futures Project argues that a METR estimate of Anthropic's AI research pace makes superhuman AI researchers likely by 2027, which they say calls for urgent pausing or pacing of frontier development. Much of the evidence behind the estimate was redacted, so the forecasts rest on partly unverifiable inputs.
- METR estimates Anthropic's AI progress is about 2x its historical rate, with roughly 80% of that acceleration attributed to AI, implying about 1.7x AI-driven speedup.
- The authors say this is far faster than their models predicted, and that the 1.7x figure matters more than it sounds because it is not the same as making staff 1.7x more productive.
- Their forecasts put a Superhuman AI Researcher around 2027, and they argue this makes a pause or international agreement urgent.
- Anthropic redacted most of the evidence METR found most informative, so the 2x and 80% numbers are hard to check independently.
The AI Futures Project updates its timelines after METR estimated Anthropic's AI progress is running about twice its historical pace, with roughly 80% of that speedup attributed to AI itself. They argue this pulls their forecast for a Superhuman AI Researcher into 2027 and makes pausing or slowing frontier development urgent.
- METR estimates Anthropic's AI R&D progress is about 2x its pre-October 2025 trend, with ~80% of the acceleration due to AI, implying roughly a 1.7x AI-driven speedup.
- The authors' models move their Superhuman AI Researcher median to around early-to-late 2027, several months ahead of their earlier AI 2027 scenario, which had put it later.
- Much of METR's evidence was redacted, so the 2x and 80% figures are hard to verify, and the authors say their own analysis is preliminary.