TL;DR
METR’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The reported work relied primarily on AI assistance to researchers, while Google DeepMind has described a separate efficiency improvement from AlphaEvolve. These examples support the possibility of a feedback loop in AI development, but they do not establish that AI systems are independently building successors or that progress will accelerate without limits.
METR said on November 5 that Anthropic is likely improving model capabilities on well-specified tasks at around twice its earlier rate, though the organization emphasized high uncertainty. The assessment described work primarily involving AI assistance to researchers, adding evidence to a broader question: whether AI tools can help develop the next generation of AI systems and, over time, change the pace of research.
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 research and development acceleration at Anthropic. Its estimate is not a precise measurement: METR said it would not be surprised by a rate as low as 1.3 times the earlier pace or as high as three times. It also redacted much of the evidence it considered most informative. The published assessment therefore points to a possible acceleration, while leaving substantial uncertainty about its size and basis.
The source account says the work METR described was primarily AI assistance to researchers. That distinction matters. A model helping researchers write code, analyze results or run experiments is not the same as a system independently directing an AI laboratory or building a successor without human oversight. The source material does not establish that broader level of automation.
A separate example comes from Google DeepMind’s AlphaEvolve. The company reported that the system improved a matrix-multiplication component used in Gemini by 23%, reducing Gemini’s overall training time by 1%. This is a company-reported engineering result, not a measure of general research acceleration. The account uses a hypothetical illustration to show how separate savings might add up: three independent reductions of 10% would leave a job requiring about 73% of its original compute, allowing roughly 37% more such jobs on the same budget.

Why Faster Research Could Compound
The central possibility is a feedback loop: AI helps researchers improve models, and more capable models then assist with further research. If those contributions reliably speed up experiments or reduce their cost, a lab could test more ideas with the same time or computing budget. Repeated gains could matter even when each individual improvement appears modest.
That possibility has practical stakes for businesses, workers and governments planning around AI. A change in the pace of development could affect when new capabilities arrive and how quickly organizations need to adapt. But the available examples do not show that such a loop is already self-sustaining. The METR estimate is uncertain, the evidence was partly redacted, and the AlphaEvolve figure concerns a specific component and its reported effect on training time.
The scale and distribution of any benefits will also depend on who can use the systems, who owns the computing infrastructure and how deployment is governed. Technical progress alone does not determine which organizations gain from it or how its effects reach workers and the public.
Products Worth Considering

DJI Mini 5 Pro Fly More Combo Plus with DJI RC 2

Samsung Odyssey G7 G75F 40-Inch WUHD Curved Gaming Monitor

DJI Osmo 360 Road Cycling First-Person Combo Action Camera
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than proposing a new model design. Researchers read prior work, inspect code, prepare data, configure experiments, track compute use and determine whether a result holds up. A proposed improvement can look convincing while relying on a flawed comparison, a favorable random seed or a test that does not measure the intended capability.
That makes verification part of the work. A useful AI contribution must survive fair testing and, where appropriate, replication. The source essay describes a schematic sequence: a model helps with research, tests filter out weak results, validated improvements feed into a stronger model, and that model joins the research process. The author explicitly presents this as an argument about how a loop could work, not as a measured account of a process at any particular lab.
The examples have different dates and scopes. Google DeepMind reported the AlphaEvolve result on May 14, 2025; METR published its Anthropic assessment on November 5, 2026. Together they illustrate two kinds of contribution—an engineering efficiency gain and assistance in research—but do not provide a single comparable measure of how much AI has accelerated the whole field.
“It “would not be surprised” by a result of 1.3x or 3x.”
— METR
Limits of the Current Evidence
The size of the acceleration remains uncertain. METR’s estimate includes a wide range, and it redacted much of the evidence it considered most informative. The source account does not provide enough detail to independently assess the underlying measurements or determine how broadly the estimate applies beyond the tasks and work examined.
It is also unclear how much of the research process current systems can complete reliably, how much human supervision they need, and whether improvements in smaller experiments transfer to larger models. Other possible constraints include experiment time, access to hardware, data quality, the choice of research direction and the ability to maintain progress over time.
Neither example confirms that AI systems are autonomously developing their own successors. AlphaEvolve’s reported result concerns a specific component, while the METR assessment describes primarily assistance to researchers. The source material offers no established timeline for a self-sustaining cycle or for progress toward superintelligence.
What Future Assessments Need to Show
Further assessments will need to clarify how much AI contributes to research tasks, how often its suggestions lead to improvements that survive independent checks, and whether those gains translate from specific experiments into faster development of larger systems. Clearer evidence about the methods and comparison baselines would help readers judge how much weight to give the reported acceleration.
For now, the practical next step is to track measured contributions and their limits. The source essay argues that plans should allow for changing rates of improvement, while recognizing that research bottlenecks may slow or interrupt progress. Whether AI assistance becomes a durable engine of faster model development remains an open question.
Key Questions
What did METR report about Anthropic?
METR said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate. It described the estimate as highly uncertain, with outcomes from 1.3 to three times the earlier rate within the range it said it would not be surprised by.
Does the assessment show that AI is building AI independently?
No. The source says the work METR described was primarily AI assistance to researchers. It does not establish that AI systems are independently directing research or building successor models without human oversight.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. That is a company-reported result about a specific component, not a general measure of AI research acceleration.
Why could small efficiency gains matter?
If separate savings hold up and can be combined, the same computing budget could support more experiments or training work. The source’s three 10% savings example is a simplified illustration; it does not show that those gains have been achieved together in practice.
Source: Thorsten Meyer AI
