TL;DR
METR’s November 5 assessment estimated, with high uncertainty, that Anthropic is improving model capabilities on well-specified tasks at around twice its earlier rate, largely through AI assistance to researchers. A separate DeepMind example found an AI-designed change made a matrix-multiplication component 23% faster and reduced Gemini’s overall training time by 1%. These findings show possible feedback loops, but do not establish that AI systems are autonomously building successors or that progress will continue at the same pace.
METR said on November 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, while stressing that the estimate has high uncertainty and describes primarily AI assistance to researchers. The assessment adds evidence to a debate about whether AI can help speed up the research that produces more capable AI, though it does not show that models are independently designing and building their own successors.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR, an independent evaluation organization, assessed the pace of AI research and development at Anthropic. It said it “would not be surprised” by an acceleration as low as 1.3 times or as high as 3 times its earlier rate. That range reflects substantial uncertainty. METR also said the work it described was primarily AI assistance to researchers, rather than a fully automated process.
The assessment did not make all its evidence available: Anthropic redacted much of the material METR considered most informative. The reported estimate is therefore a judgment about the pace of capability improvement on well-specified tasks, not a precise measure of how much AI has automated across all research or how quickly future models will improve.
A separate example comes from Google DeepMind. In a May 14, 2025 report, the company said its AlphaEvolve system produced a matrix-multiplication component used in Gemini that ran 23% faster, reducing Gemini’s overall training time by 1%. The result shows how an optimization in one part of a system can yield a smaller gain for the whole process; it does not establish a general rate of improvement across AI development.
How Research Gains Could Compound
If AI systems help researchers find and validate useful improvements, those improvements could make later research faster or cheaper. A stronger model could then assist with more experiments, creating a feedback loop. The key condition is that suggested changes must work under fair tests and produce gains that carry over to larger systems. Assistance with coding or experiments alone does not prove that a model can independently choose the right research goals or reliably judge its own results.
The arithmetic can make modest efficiency gains appear consequential over time. In a deliberately simple illustration from the source essay, three independent 10% savings would leave a task requiring about 73% of its original computing resources. With the same computing budget, that would allow roughly 37% more of those jobs. This is an illustration, not a measured outcome at Anthropic or DeepMind; real gains may not be independent, repeatable, or available at every stage.
The pace of research affects how quickly businesses, workers and governments may need to respond to AI capabilities. But technical progress alone will not determine who can use such systems, who controls the computing infrastructure, or how the benefits are shared. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)
Samsung 57 Odyssey Neo G9 (G95NC) Dual 4K UHD 1000R Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From AI Assistance to Successors
The source essay describes a possible research loop: a model proposes changes, helps write code or run experiments, and contributes to checking results. Researchers test the proposals; ideas that hold up may be used in a stronger model, which can then help with further research. The essay labels this sequence a schematic of an argument, not a measured process at any laboratory.
Research involves more than executing experiments. A team must set a useful question, establish a fair comparison, account for failed or misleading results, and decide whether an apparent improvement is robust. A reported gain can disappear with a different random seed, reflect an easier test, or result from a change in computing resources. Reliable assistance therefore depends on systems helping researchers identify weaknesses as well as generate promising ideas.
The DeepMind example illustrates a bounded engineering contribution: AlphaEvolve improved a component and the company reported a smaller gain in overall training time. METR’s assessment addresses a broader question about the pace of capability improvement at Anthropic. The two reports concern different organizations and measures; together they offer examples relevant to AI-assisted research, but they do not establish one continuous or universal trend.
““Would not be surprised” by an acceleration of 1.3x or 3x.”
— METR
Limits of the Acceleration Evidence
METR’s estimate is explicitly uncertain, and the available source says Anthropic redacted much of the evidence METR found most informative. The assessment concerns capability improvement on well-specified tasks; the source does not provide enough detail to infer the same rate for open-ended research, other laboratories, or future periods.
It also remains unclear how much of the research loop current systems can complete reliably without human direction, how gains transfer from small experiments to large training runs, and whether improvements can be sustained as experiments become more demanding. The source identifies experiment time and hardware, research direction, continuity and data quality as possible constraints. It does not establish when, or whether, these limits would become decisive.
What Future Evaluations Need to Show
Further assessments will need to clarify the evidence behind reported acceleration, the tasks measured, and how much human supervision the work required. Repeated results across different experiments and larger systems would help show whether AI assistance produces durable gains, while transparent comparisons could help distinguish genuine improvements from changes in testing conditions.
For now, the reports support a measured conclusion: AI assistance has been linked to a specific component-level efficiency gain, and METR judges that capability improvement at Anthropic may be happening faster than before. The size and durability of any broader feedback loop remain unsettled.
Key Questions
What did METR report about Anthropic?
METR estimated that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate. It described the estimate as highly uncertain, with 1.3x to 3x as rates it would not be surprised by.
Does this mean AI is building its own successors?
No. METR said the work it described was primarily AI assistance to researchers. The source does not establish that models independently direct and complete the process of building successor systems.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%.
Why are the estimates uncertain?
METR described its estimate as highly uncertain, and Anthropic redacted much of the evidence METR considered most informative. The assessment is also limited to capability improvement on well-specified tasks.
What could slow AI-assisted research?
The source identifies experiment time, hardware, the difficulty of transferring small-scale results to large systems, research direction, continuity and data quality as possible constraints. Their effects and timing remain unclear.
Source: Thorsten Meyer AI
