🔍 Read the full analysis: When AI Helps Develop The Next Generation Of AI on ThorstenMeyerAI.com
Get the latest gadgets 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 judged that Anthropic’s AI capabilities on well-specified tasks were improving at about twice the earlier rate, with high uncertainty, and said AI assistance to researchers likely accounted for most of the acceleration. Separately, Google DeepMind reported that AlphaEvolve sped up a matrix-multiplication component used in Gemini, reducing overall training time by 1%. How far these gains can compound remains uncertain.
AI assistance may be speeding up AI research, according to an independent assessment of work at Anthropic published by METR on November 5. METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice their earlier rate, while stressing substantial uncertainty; it said the work it described was primarily AI assistance to researchers, not fully autonomous AI research.
METR said its assessment drew 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 between 1.3 times and three times the earlier pace. Anthropic redacted much of the evidence METR considered most informative, leaving outside readers dependent in part on the evaluators’ interpretation.
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%, cutting the model’s overall training time by 1%. That is a company-reported result about a particular component and training process, not evidence that AI has broadly automated the development of advanced models.
These examples illustrate several ways AI tools could assist researchers: proposing or implementing changes, helping examine experiment results, and identifying problems in code or evaluation. A reported improvement still needs fair testing and replication. A change that appears to help could instead reflect different computing budgets, an easier test, or variation between runs.
How Research Gains Could Compound
If researchers can use AI to investigate more ideas or spend less time on routine coding and troubleshooting, a team may make progress faster without increasing its size. If a better model then helps with research on its successor, AI capability could feed into further AI development. That feedback loop is the central reason the possibility of faster progress has attracted attention.
Efficiency gains can also add up, though simple illustrations should not be mistaken for forecasts. For example, three independent savings of 10% would leave a task using about 73% of its original computing resources. That could allow roughly 37% more of those tasks on the same budget, assuming the savings are genuinely independent and apply to the same work. The example does not establish that such gains will occur in practice.
The pace of AI progress affects decisions by companies, workers and governments. But technical capability alone will not determine who can use powerful systems, who owns the infrastructure or how gains are distributed. Those outcomes will also depend on deployment choices and institutions.
Products Worth Considering
Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop
LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Autonomy
AI research is a chain of tasks, not a single act of invention. Researchers review prior work, inspect software, prepare data, run experiments, diagnose failures and decide whether results are reliable enough to use. Systems can assist with parts of that process without independently choosing research goals or validating their own conclusions.
The distinction matters when interpreting METR’s assessment. Its account described primarily AI assistance to human researchers. More extensive automation would require systems to handle a broader sequence: propose a useful experiment, implement it, assess the result fairly, revise the approach and recognize when the evidence is weak. The source material does not establish that current systems can reliably perform that entire cycle without supervision.
Small improvements in software or training efficiency may matter to research labs, but they do not automatically translate into equivalent improvements in model capability. Results have to survive testing at larger scales and under the constraints of available hardware, experiment time and data quality.
“METR said it “would not be surprised” by a rate of improvement between 1.3 times and three times the earlier pace.”
— METR, as described in its November 5, 2026 assessment
Limits in the Evidence and Pace
METR’s estimate carries substantial uncertainty, and the public cannot inspect much of the evidence the evaluators considered most informative because Anthropic redacted it. The assessment concerns one company and well-specified tasks; it does not establish a comparable rate across the AI industry or every kind of research.
It is also unclear how reliably AI-assisted improvements will hold up when repeated or applied to larger systems. Experiment time and hardware can constrain the number of tests, while findings from small experiments may not transfer to large training runs. Researchers still need to choose which problems to pursue, maintain continuity across projects and ensure that data and evaluation methods are sound.
The source material does not quantify how much of the reported acceleration came from particular AI tools, how much human supervision was required, or whether the pace will continue. It also does not provide evidence that the cited efficiency gains will lead directly to a specific timeline for advanced AI or superintelligence.
What Researchers Will Measure Next
Further assessments will need to show whether the reported faster pace persists and whether AI contributions produce repeatable gains beyond narrowly defined tasks. Clearer public information about methods, evaluation criteria and the amount of human oversight would help readers and researchers judge the claims.
For individual experiments, the next test is practical: repeat the work under comparable conditions, check for unintended changes to the evaluation, and determine whether a gain survives at larger scale. Researchers and organizations will also need to track constraints such as computing capacity, data quality and the time required to validate results.
For now, the evidence supports a measured conclusion: AI is assisting some research work, and one company has reported a specific efficiency improvement in Gemini training. How much faster AI development will become—and who will benefit—remains unsettled.
Key Questions
What did METR conclude about AI research at Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate. It said the estimate was highly uncertain and that the work it described was primarily AI assistance to researchers.
Does this mean AI is independently developing new AI models?
No. The assessment described AI primarily assisting human researchers. It does not establish that AI systems can independently manage the full research process, including setting goals and validating results.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. This is a company-reported result for a specific component and process.
Why is the METR estimate uncertain?
METR gave a broad range for the possible rate of improvement, and Anthropic redacted much of the evidence the evaluators found most informative. The assessment also concerns one company and a defined category of tasks, rather than all AI research.
What could slow AI-assisted research?
Constraints include experiment time, hardware availability, data quality and the difficulty of applying results from small tests to larger systems. Researchers also have to choose useful directions and verify that apparent gains are real.
Source: ThorstenMeyerAI.com
