🔍 Read the full analysis: What AI Could Bring To The Development Of Its Successor on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
METR’s November 5 assessment found that Anthropic’s capabilities on well-specified tasks were likely improving about twice as fast as before, with high uncertainty, and attributed much of the acceleration to AI assistance for researchers. The report points to a possible feedback loop in which AI helps develop stronger successor models, though its pace and limits remain unclear.
METR, an independent evaluation organization, said in a November 5 assessment that Anthropic was likely improving its models’ capabilities on well-specified tasks at about twice its earlier rate, while stressing substantial uncertainty. METR judged that AI assistance to researchers was likely responsible for much of the acceleration, adding fresh evidence to the possibility that AI systems could help develop their own successors.
METR based its assessment on internal information from August through October 2026. Its estimate was not a precise measurement: the organization said it would not be surprised by a rate as low as 1.3 times the earlier pace or as high as three times. The source article reports that Anthropic redacted much of the evidence METR considered most informative, limiting what outside readers can independently examine.
The assessment described AI’s role mainly as assistance to human researchers, not a fully autonomous system conducting AI research on its own. That distinction matters. A model helping write code, inspect results, or diagnose failed experiments may let a team test more ideas, but the researchers still need to decide what to investigate and check whether apparent gains are real.
A separate example cited in the source comes from Google DeepMind. The company reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. The example illustrates how a sizeable improvement in one component can translate into a smaller gain for a complete system. It does not establish that similar gains will occur consistently across AI development.
How Research Gains Could Compound
If AI tools help researchers complete parts of model development faster, the resulting time or computing savings could support additional experiments. A stronger model might then assist with more of the next development cycle, creating a feedback loop. The possibility matters because faster research could change how quickly AI capabilities advance, affecting organizations and people planning around the technology.
The source article offers a simplified illustration: three independent savings of 10% would leave a task using about 73% of its original compute. With a fixed budget, that would allow roughly 37% more of those jobs to run. This is an illustration, not a reported result from an AI lab; actual gains depend on whether the savings are independent, reproducible, and applicable to the work being done.
More research throughput could also change which teams can build advanced systems and how much human labor they need. But technical progress alone does not determine who gets access, who controls computing infrastructure, or how benefits are distributed. Those outcomes also depend on deployment decisions and institutions.
Products Worth Considering
Ledger Flex Crypto Wallet with 2.8-Inch Touchscreen
Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Silver
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
Developing a model involves more than producing code. Researchers review prior work, prepare data and experiments, compare results, investigate failures, and decide whether an apparent improvement holds up. AI can potentially assist at several points, but each creates ways for a system to produce a persuasive yet misleading result.
For example, an agent might report a 5% improvement after changing training code. The result could instead reflect an easier test, different computing resources, or a gain that disappears with another random seed. Useful assistance therefore requires more than proposing changes: a system must help test them fairly, interpret evidence, and recognize when the result is too weak to support a conclusion. Independent checks remain central to determining whether AI contributions are genuinely useful.
The source also identifies possible constraints on faster progress: experiment time and hardware availability, the gap between small tests and large-scale systems, choices about research direction, continuity of work, and data quality. These limits mean that gains in one part of the process may not translate directly into faster overall development.
“The tool being improved can also contribute to the work of improving it.”
— Thorsten Meyer, source article author
What the Evidence Cannot Yet Show
METR’s estimate carries high uncertainty, and much of the evidence it considered most informative was redacted by Anthropic. The public account therefore does not let readers inspect all of the underlying material or independently determine how much of the reported change came from AI assistance.
The assessment concerns capabilities on well-specified tasks and describes primarily AI assistance to researchers. It does not establish that AI systems can independently conduct the full research process, reliably select the most important problems, or produce successive models at a continuously accelerating rate. The source provides no confirmed timetable for when, or whether, such broader autonomy might emerge.
It is also unclear how well gains from limited experiments transfer to training larger systems, how often reported improvements survive repeated testing, and whether hardware, data, or coordination constraints will offset productivity gains. The article frames rapid progress toward superintelligence as a possibility to take seriously, not as an established outcome.
Testing the Next Research Cycle
The next useful evidence will be whether AI-assisted improvements continue to hold up across repeated experiments and larger systems, and whether independent evaluators can examine enough information to assess them. Further reporting or assessments may clarify the extent of AI’s contribution and how METR’s estimate changes over time.
For now, the source argues that planning should allow for an uneven pace of improvement rather than assuming either steady acceleration or a fixed ceiling. The central questions remain how much of research AI can reliably handle, what constraints will slow the process, and how institutions will manage access and the distribution of benefits.
Key Questions
What did METR report about Anthropic’s research pace?
METR judged that Anthropic’s capabilities on well-specified tasks were likely improving at about twice the earlier rate. It described substantial uncertainty, saying estimates of 1.3 or three times the earlier rate would not surprise it.
Did METR say AI was conducting research autonomously?
No. The source says METR described the work primarily as AI assistance to researchers. The assessment does not establish that AI systems can independently manage the full research process.
What does AlphaEvolve’s reported result show?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall Gemini training time by 1%. It is a company-reported example of a component-level gain, not proof that similar results will apply broadly.
What remains uncertain about AI helping build successor models?
The scale and repeatability of the gains, how they transfer to larger systems, and how much AI contributes remain unclear. Much of the evidence METR considered most informative was redacted, and constraints such as computing resources and data quality may limit progress.
Source: ThorstenMeyerAI.com
