TL;DR
METR’s November 5 assessment found that Anthropic is likely improving model capabilities on well-specified tasks about twice as quickly as before, with high uncertainty. The reported work was primarily AI assistance to researchers; how far that assistance can scale remains unclear.
METR said it “would not be surprised” by an acceleration as low as 1.3 times or as high as three times. The assessment describes work in which AI systems assist researchers; the source material does not say that Anthropic has automated its research process end to end. METR also redacted much of the evidence it considered most informative, limiting what outside readers can inspect.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A separate example comes from Google DeepMind. In a May 14, 2025 report, the company said its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a company-reported result about a particular component and training workload, not evidence by itself of a general acceleration in AI research.
The source article offers a simplified illustration of compounding efficiency: three independent savings of 10% would leave a job using about 73% of its original compute, allowing roughly 37% more such jobs within the same compute budget. This is an illustration, not a reported laboratory measurement. Actual gains depend on whether savings are independent and apply to comparable work.
How Faster Research Could Compound
If AI tools help researchers propose experiments, write code, inspect results and catch errors, they could reduce the time or resources needed to test ideas. Improvements that survive careful evaluation might then be used in later models, which could assist with further research. This feedback loop is the mechanism behind concerns about accelerating progress; the METR assessment is one indication of faster AI-assisted work, not proof that the full loop is already operating autonomously.
The stakes extend beyond laboratories because the pace of capability development can affect how companies, workers and governments plan for AI. But technical progress alone does not determine who can use these systems, who owns the infrastructure or how economic benefits are distributed. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
From Research Help to Successors
The article describes a possible cycle: a model helps with research, tests determine which proposed improvements hold up, and successful changes contribute to a stronger model. That model could then join later research. The diagram in the source is explicitly a schematic of an argument, not a measured account of any laboratory’s process.
In practice, research involves more than generating ideas. Teams must inspect existing work, implement experiments, control the data and conditions, repeat promising results, and decide whether a gain is real. An apparent improvement can come from a flawed comparison, a change in compute, or random variation. Reliable assistance therefore depends on systems helping researchers test results fairly, not just produce plausible proposals.
““would not be surprised” by 1.3x or 3x”
— METR
Limits on the Acceleration Estimate
The METR assessment carries high uncertainty, and its redactions prevent readers from reviewing much of the evidence it found most informative. The source material does not provide enough detail to determine how the estimate was calculated or how broadly it applies beyond the specified tasks and Anthropic work described.
Several potential constraints remain: experiment time and hardware availability, whether results from small experiments transfer to larger systems, choosing productive research directions, maintaining continuity, and data quality. The source argues that progress could be uneven, but does not quantify how much these factors may slow it. It is also unclear how much of the research cycle AI systems can complete reliably with limited human supervision.
Evidence to Watch in Future Evaluations
Further independent evaluations could clarify how much AI assistance changes research output, how the estimate varies across tasks, and whether apparent gains hold up under repeatable tests. More accessible evidence would also help readers assess the methods behind the estimates, given METR’s redactions.
For now, the source’s practical implication is to plan for a pace of improvement that may change, while distinguishing measured results from scenarios about what could follow. Whether AI assistance develops into a more self-reinforcing research cycle, and how its benefits are shared, remain open questions.
Key Questions
What did METR report about Anthropic?
METR said Anthropic is likely improving model capabilities on well-specified tasks at about twice the earlier rate, while emphasizing high uncertainty. It described the work as primarily AI assistance to researchers.
Did METR say AI is independently building successor models?
No. The assessment described AI-assisted research. The self-improving cycle in the source article is a schematic argument, not a measured description of an autonomous process at a lab.
What did Google DeepMind report about AlphaEvolve?
Google DeepMind said AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting the model’s overall training time by 1%. This is a company-reported result for a specific use.
Why is the METR estimate uncertain?
METR described its estimate as highly uncertain, gave a range it would not be surprised by, and redacted much of the evidence it considered most informative. The source does not provide enough detail to assess how widely the estimate applies.
What could slow AI-assisted research?
Potential constraints include experiment time, hardware, transfer from small experiments to larger systems, research direction, continuity and data quality. The source does not quantify their effects.
Source: Thorsten Meyer AI
