🔍 Read the full analysis: Could The Tool Help Build The AI That Follows It? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate, with substantial uncertainty. It attributed most of the apparent acceleration to AI assistance for researchers, not autonomous AI research. A separate Google DeepMind example shows a measurable efficiency gain, but it does not establish that AI systems can independently build their successors.
An assessment published by METR on November 5 judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice their earlier rate, with substantial uncertainty. METR said AI assistance to researchers was likely responsible for most of the apparent acceleration. The finding adds evidence to a consequential possibility: AI tools may help researchers develop more capable AI systems, though the assessment does not show that AI can independently build its successors.
METR, an independent evaluation organization, based its assessment on internal information from August through October 2026. It described the acceleration estimate as uncertain: the organization said it would not be surprised by a rate of 1.3 times or three times the earlier pace. The source article says Anthropic redacted much of the evidence METR considered most informative, limiting what outside readers can check directly.
The account characterizes the work as primarily AI assistance to human researchers. That can include helping with research tasks, but it is not the same as a system autonomously choosing a research direction, validating results and producing a successor model with little or no human involvement. The degree of supervision and the reliability of individual contributions remain central to interpreting what the reported acceleration means.
A separate example comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the improvement reduced Gemini’s overall training time by 1%. This is a specific efficiency result, not evidence by itself of a broad acceleration in AI research or of autonomous model development.
How Research Gains Could Compound
If AI tools help researchers test more ideas or reduce routine work, a research team might make progress faster without adding staff or computing resources. If improvements also make training more efficient, the same hardware budget could support more experiments or model runs. The practical effect depends on whether a proposed improvement survives careful testing and can be used at the scale of a real training system.
The source article gives a simplified illustration: three independent 10% compute savings would leave a task requiring about 73% of its original compute, allowing roughly 37% more of those jobs on the same budget. That is a mathematical example, not a measured result from an AI lab. It illustrates how separate efficiency gains could add up, but does not show that such savings are occurring at that rate or can be achieved independently.
The stakes extend beyond laboratory productivity. A faster improvement cycle could make forecasts about AI capabilities, investment and workforce effects harder to maintain. But technical progress alone does not determine who can use advanced systems, who controls the infrastructure or how any gains are distributed. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering
Samsung Odyssey G7 G75F 40-Inch WUHD Curved Gaming Monitor
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than writing code. A team may need to review prior work, inspect an implementation, prepare data, configure an experiment and decide how to measure its outcome. Researchers then have to investigate failed runs, repeat promising ones and determine whether an apparent gain is genuine. AI assistance could speed some of these tasks, even if people remain responsible for setting goals and judging results.
That verification is necessary because an experiment can look successful for the wrong reason. A code change might make a test easier, use a different amount of computing or produce a gain that disappears under another random seed. A useful research contribution must hold up to fair comparisons and replication, not merely produce an encouraging report.
The source also points to constraints on the pace of progress: experiment time and hardware, the difficulty of translating small tests to large systems, choices about research direction, continuity and data quality. These limits mean AI assistance need not produce a steady or uniform acceleration. The METR assessment is evidence about one organization over a defined period, rather than a forecast that applies to every lab.
“It would not be surprised by 1.3x or 3x.”
— METR, as described in its November 5 assessment
Limits of the Current Evidence
The public account does not provide the underlying evidence METR found most informative because Anthropic redacted much of it. Readers therefore cannot independently inspect the full basis for the estimate from the material provided. METR’s assessment also applies to Anthropic and a particular period; the source gives no basis for treating it as a measured industry-wide rate.
It is also unclear how much of the reported acceleration came from particular AI tools, how much human supervision the work required, and whether the observed gains persist across different research tasks. The available description does not establish that AI systems can set research priorities, reliably assess their own results or produce a stronger successor end to end. Those are distinct capabilities from assisting researchers on well-specified tasks.
The DeepMind example is a company-reported result described in the source, not a general estimate of efficiency gains across model training. The source material does not provide further independent evaluation details. Nor does it establish that several savings of the kind in the illustrative calculation can be achieved together in practice.
What Future Evaluations Must Show
The next useful evidence would show, in more detail, which research tasks AI systems performed, what level of human oversight they required and how their contributions held up under independent checks. Replicable comparisons could clarify whether faster progress reflects better research assistance, changes in experimental practice or other factors. Public reporting that allows outside scrutiny would help resolve the limits created by redacted evidence.
Researchers and readers will also need to distinguish gains on narrowly specified tasks from broader abilities to conduct research. The key questions are whether AI assistance remains useful when experiments are repeated, whether improvements transfer to larger systems and whether the pace changes as hardware, data and research bottlenecks shift. The current assessment supports attention to the possibility of faster AI development, but does not settle how far or how quickly it will go.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate. It stressed substantial uncertainty, saying it would not be surprised by an estimate of 1.3 times or three times the earlier pace.
Does the assessment show that AI can build the next AI model on its own?
No. The source describes the work as primarily AI assistance to researchers. It does not establish end-to-end autonomous research or show that AI independently designed and built a successor model.
What did Google DeepMind say AlphaEvolve improved?
DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The reported component-level gain was larger than the reduction in total training time.
Why is the METR finding difficult to verify publicly?
The source article says Anthropic redacted much of the evidence METR considered most informative. That leaves outsiders with limited ability to inspect the full basis for the assessment.
What could slow AI-assisted research progress?
The source identifies experiment time, hardware limits, difficulty scaling results from small tests, research-direction choices, continuity and data quality as possible constraints. It also notes that apparent improvements need careful validation.
Source: ThorstenMeyerAI.com
