🔍 Read the full analysis: AI As A Partner In Building Its Successor 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
A November 5 assessment by METR found that Anthropic was likely improving model capabilities on well-specified tasks at around twice its earlier rate, while stressing substantial uncertainty. METR judged AI assistance to researchers likely accounted for most of the acceleration; the public evidence is limited because Anthropic redacted much of the information evaluators found most informative.
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. The evaluators said AI assistance to researchers was likely responsible for most of the acceleration, but emphasized high uncertainty; much of the evidence they considered most informative was redacted.
METR’s assessment drew on internal information from August through October 2026. Its estimate concerns improvement on well-specified tasks, not a general measure of all AI research or a finding that AI systems are independently building successor models. The work described was primarily AI assistance to human researchers.
METR said it would not be surprised if the acceleration were as low as 1.3 times the earlier rate or as high as three times. Those figures describe the evaluators’ uncertainty around their assessment, not separate measured outcomes. Anthropic redacted much of the evidence METR regarded as most informative, limiting how much outside readers can independently examine.
A separate example cited in the source article comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. This is a company-reported engineering result; it does not by itself establish that AI can autonomously direct the broader process of AI research.
How Research Gains Could Compound
The finding matters because AI systems could help with parts of the work used to improve later systems: reading research, writing or checking code, proposing experiments, diagnosing failed runs, and analyzing results. If those contributions hold up under testing, researchers may be able to investigate more ideas with the same staff and computing resources. A more capable model could then assist with further development, creating a feedback loop between AI capability and research productivity.
Efficiency improvements may also add up. The source article offers a simplified illustration: if three independent changes each reduce compute needs by 10%, a task would use about 73% of its original compute. Under that illustration’s assumptions, the same budget could run roughly 37% more such tasks. This is an example, not a measured forecast; actual gains depend on whether improvements are independent and can be applied at scale.
Faster development could affect how quickly businesses, workers and governments need to adapt. But technical progress alone does not determine who can use advanced systems, who controls the infrastructure, or who benefits. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering

Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)

DJI Osmo 360 Motorcycle Handlebar POV Combo
As an affiliate, we earn on qualifying purchases.
From Assistance to Successor Research
“AI as a partner in building its successor” describes a possible progression, rather than a confirmed state of fully automated research. A model might help a team propose a change, implement it, run an experiment and inspect the result. Human researchers would still need to judge whether the comparison was fair and whether an apparent gain survives replication.
That scrutiny matters because a result can look promising for the wrong reason. A code change might make a test easier, use a different amount of compute, or produce a gain that disappears with another random seed. Reliable research requires systems to identify such problems and acknowledge when evidence is weak, not merely produce a favorable result.
The source article also identifies limits on how quickly this process could grow: experiment time and hardware availability, differences between small experiments and large training runs, choices about research direction, continuity, and data quality. The METR assessment supplies a concrete but uncertain signal of current acceleration; the AlphaEvolve example shows a specific efficiency improvement, not proof that every stage of AI development is being automated.
“The tool being improved can also contribute to the work of improving it.”
— Thorsten Meyer, source article
Limits of the Available Evidence
The public account does not provide enough detail for readers to independently reproduce METR’s estimate. Anthropic redacted much of the evidence the evaluators considered most informative, and the assessment itself describes high uncertainty. The available material also does not establish how much of the reported acceleration came from AI assistance compared with other changes in research practice or resources.
It remains unclear how well assistance on defined tasks carries over to open-ended research, how often proposed improvements survive independent checks, and whether gains observed in components or smaller experiments translate to large-scale model training. The source article’s discussion of compounding efficiency is illustrative, not a prediction that development will follow a fixed pace.
Nor does the assessment show that AI systems can independently choose research goals, reliably evaluate their own findings, or build a successor without substantial human oversight. The transition, if it continues, may be uneven, with technical and practical constraints shaping its speed.
Testing the Next Research Gains
The next useful evidence will be whether AI-assisted improvements can be documented, checked and repeated across more than one narrow task. Independent evaluations with enough detail to examine methods and comparisons would help clarify how much of the reported acceleration is attributable to AI and how robust the gains are.
Researchers will also need to test whether results from small experiments hold at the scale of major training runs, and whether systems can help catch misleading results as effectively as they help generate promising ones. The source material names no specific date for a follow-up assessment or public release of additional Anthropic evidence, so the timing of further clarification is unknown.
For now, METR’s assessment is a signal of possible faster AI-assisted research, not confirmation of fully autonomous AI development or a guaranteed trajectory toward superintelligence. Decisions about access, infrastructure and distribution of benefits remain open questions alongside the technical evidence.
Key Questions
What did METR report about Anthropic?
METR said Anthropic was likely improving model capabilities on well-specified tasks at around twice its earlier rate, based on internal information from August through October 2026. It described the estimate as highly uncertain.
Did METR find that AI is building its own successor?
No. The assessment described work that was primarily AI assistance to human researchers. It does not establish that AI systems independently direct research or build successor models.
How certain is the estimate?
METR stressed substantial uncertainty and said it would not be surprised by a rate of 1.3 times or three times the earlier rate. The source article also says much of the evidence METR found most informative was redacted.
What was the AlphaEvolve result?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. This is a company-reported, specific efficiency result, not evidence of fully automated AI research.
What could slow AI-assisted research?
Factors cited in the source article include experiment time, hardware, scaling from small tests to large systems, research direction, continuity and data quality. How strongly these limits affect future progress remains unclear.
Source: ThorstenMeyerAI.com
