An assessment shared by AI safety group METR says Anthropic made roughly two years of benchmark progress in the past year. The finding points to faster AI development, but the report says people still oversee the most critical parts of model training.
METR Watch analysis
What happened
The Independent reports that METR reviewed Anthropic’s internal data and evaluated its models. The assessment found benchmark progress had effectively doubled over the year. Researchers attributed the acceleration primarily to AI systems’ growing ability to help build the next generation of AI, rather than to increased computing power for training.
The account also draws an important line: METR said AI systems were not yet making highly significant, fully autonomous contributions to the most important parts of model training. Researchers described signs that this could change soon, with maturity varying between areas. They warned that further acceleration could bring risky capabilities into view suddenly.
Why it matters
The distinction is between AI helping researchers move faster and AI independently taking over the work that produces more capable models. METR’s assessment, as described by The Independent, says the first is already happening; it does not say the second has arrived. That is a meaningful signal for anyone following AI development and safety, without turning a benchmark result into proof of an imminent runaway system.
Our read
The useful headline is acceleration, not autonomy. Benchmark progress matters, but it is not a complete measure of what models can do or how quickly the next milestone will arrive. METR’s reported distinction between current human oversight and possible future contributions is the part worth keeping in focus. An alarming headline can travel faster than the careful boundary underneath it; this one needs both.
What to watch
- Whether METR publishes the assessment and its methods in full.
- How its benchmark results change in later evaluations.
- Whether AI systems begin making substantial contributions to model training, rather than assisting with narrower tasks.
Discussion spark: If AI systems are speeding up model development but people still oversee the most critical training work, what evidence should count as the point when that balance has materially changed?
Sources and evidence
Independent WittyWires tracker for public updates about METR. Not affiliated with or endorsed by METR; this is not an official account.