🔍 Read the full analysis: When AI Joins The Process Of Building AI on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
An assessment published by METR on November 5, 2026, found that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty and AI assistance likely responsible for most of the acceleration. Separately, Google DeepMind said AlphaEvolve improved a matrix-multiplication component used in Gemini, reducing overall training time by 1%. The evidence points to AI contributing to AI research, but does not establish that the process is fully automated or that acceleration will continue at the same pace.
AI tools are beginning to contribute to the work of developing AI, according to a November assessment by independent evaluator METR and a separate efficiency result reported by Google DeepMind. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, while stressing substantial uncertainty; DeepMind said an AI system helped make one component used in Gemini faster, trimming overall training time by 1%.
METR published its assessment on November 5, 2026, drawing on internal information from Anthropic covering August through October. The evaluator said it would not be surprised if the acceleration were as low as 1.3 times or as high as three times the earlier rate. It judged AI-assisted research likely responsible for most of the acceleration it observed, but described the work primarily as AI assistance to human researchers, not autonomous AI research.
The public account has limits. Anthropic redacted much of the evidence METR considered most informative, leaving readers outside the evaluation with less information to independently assess its conclusion. The finding is an evaluator’s judgment, not a precise measurement that settles how much AI assistance caused the change.
A separate example concerns engineering efficiency. Google DeepMind reported that AlphaEvolve improved a matrix-multiplication component used in Gemini by 23%, cutting Gemini’s overall training time by 1%. That result illustrates how a gain in one part of a complex system can affect the full process, but it does not show that every research task can be automated or that the same gain applies across other models.
How Research Gains Could Compound
The significance is not just whether an AI model can complete an isolated coding task. If systems help researchers propose experiments, implement changes, analyze results and catch errors, teams may be able to investigate more ideas with the same time and computing resources. If useful improvements then feed into later models that assist with further research, the technology could contribute to its own development.
That feedback loop could affect how quickly AI capabilities improve, making the pace of progress harder for companies, workers and public institutions to plan around. But faster research is not the same as guaranteed rapid capability growth. Experiments must be checked, improvements must survive replication, and researchers still need to choose worthwhile problems. The available examples show assistance and a specific efficiency gain, not an established path to superintelligence.
Efficiency improvements can add up in principle. In a simplified illustration from the source article, three independent savings of 10% would leave a job requiring about 73% of its original computing resources. That calculation assumes the savings are independent and apply to the same job; it is an illustration, not a measured outcome from Anthropic or DeepMind.
Popular right now
Hisense 40-Inch QD4SR Series Hi-QLED FHD Roku Smart TV
eufy E25 Robot Vacuum and Mop with HydroJet Roller Mop
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
As an affiliate, we earn on qualifying purchases.
From Research Tasks to Model Training
AI research involves more than producing a promising idea. A team investigating how to reduce the computing needed for a given level of model performance must review prior work, inspect code, prepare data, configure experiments and compare results. A reported improvement can be misleading if the test became easier, the runs used different amounts of computing, or the result does not hold with another random seed.
That makes verification part of the research task. An AI contribution is useful only if its proposed change works under fair testing and remains valuable after scrutiny. Systems may assist with routine coding or diagnosis while people retain responsibility for judging whether evidence supports a conclusion. More autonomous research would require systems to take on a larger sequence of work, including deciding what to investigate next.
The two reported developments measure different things. METR assessed the pace of capability improvement on well-specified tasks at Anthropic; DeepMind reported a component-level efficiency improvement and its effect on Gemini’s training time. Neither result alone measures economy-wide effects, establishes a lasting trend or tells how much of future model development will be performed by AI.
“would not be surprised”
— METR, describing the uncertainty around its Anthropic assessment
What the Evidence Cannot Yet Show
The exact size and cause of the reported acceleration remain uncertain. METR’s assessment includes a wide range, and much of the evidence it considered most informative was redacted by Anthropic. The public material does not provide enough detail to independently reconstruct the evaluator’s analysis or isolate AI assistance from other factors.
It is also unclear how well current assistance generalizes beyond well-specified tasks, how often proposed improvements survive independent checks, and whether systems can reliably choose research directions. Experiment time, hardware availability, data quality and the gap between small tests and large-scale training could all limit the pace of progress. The reported results do not establish that AI systems are conducting AI research end to end.
Finally, technical capability alone does not determine who can access these systems, who controls the computing infrastructure or how any economic benefits are shared. Those outcomes will depend on deployment decisions and institutions as well as model performance.
Testing Whether Gains Hold
The next useful evidence will show whether AI-assisted research gains persist across more tasks and over time, and whether independent evaluations can measure them with clearer public evidence. For any proposed improvement, readers and researchers will need to know how it was tested, whether comparisons used equivalent resources and whether results held under replication.
Further reporting on METR’s assessment, Anthropic’s methods and additional examples from AI laboratories could clarify what work systems currently handle and how much human supervision remains necessary. Until then, the most grounded conclusion is limited: AI is contributing to parts of AI research and engineering, but the scale, reliability and long-term effect of that contribution are still developing.
Key Questions
What did METR find about AI research at Anthropic?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It emphasized substantial uncertainty and said AI-assisted research likely accounted for most of the acceleration it assessed.
Does the assessment show that AI is building AI autonomously?
No. METR described the work primarily as AI assistance to researchers. The assessment does not establish that AI systems independently manage the full research process.
What did AlphaEvolve change for Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%.
Why is the METR finding uncertain?
METR gave a broad possible range for the acceleration, and Anthropic redacted much of the evidence METR considered most informative. The public account therefore does not let outsiders independently assess every part of the analysis.
What could limit further acceleration?
Research still depends on experiment time, computing hardware, reliable data and careful evaluation. It may also be difficult to tell whether results from small experiments will carry over to large systems or to identify the most productive research direction.
Source: ThorstenMeyerAI.com
