🔍 Read the full analysis: Could An AI Tool Help Create Its Successor? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment said Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty and AI assistance playing a major role. A separate Google DeepMind report said AlphaEvolve sped up one matrix-multiplication component in Gemini, reducing overall training time by 1%. These examples show how AI could contribute to developing future systems, but they do not establish fully autonomous AI research or an inevitable acceleration.
AI is already contributing to work that can improve AI systems, according to a November 5 assessment by the independent evaluation organization METR and a separate result reported by Google DeepMind. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty; Google DeepMind said its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, cutting the model’s overall training time by 1%. The developments offer concrete examples of AI assistance in research, but they do not show that systems can independently build their successors.
METR’s assessment drew on internal information from August through October 2026. The organization 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, underscoring how uncertain its estimate is. METR described the work as primarily AI assistance to human researchers, rather than research conducted autonomously by AI.
The available public evidence is incomplete. According to the source article, Anthropic redacted much of the information METR considered most informative. Readers therefore cannot independently examine all the underlying material, and the assessment relies substantially on METR’s interpretation of evidence it reviewed.
Google DeepMind’s reported AlphaEvolve result illustrates a different kind of contribution: optimizing a specific part of a model-training process. The 23% improvement applied to a matrix-multiplication component, not to Gemini’s whole training workload. The reported effect on overall training time was a 1% reduction.
How Research Gains Could Compound
AI assistance could matter even when it does not replace researchers. If a system helps write code, test an idea, inspect results or diagnose failed experiments, researchers may be able to examine more possibilities with the same time and resources. If a successful improvement then carries into the next model, that model could assist with later research. This is the proposed feedback loop: AI helps improve systems that may help improve AI.
Small efficiency gains can add up, but the figures need careful framing. The source article gives a simplified illustration: if three independent changes each reduce resource use by 10%, a task would require about 73% of its original compute. Under that illustration’s assumptions, a fixed budget could run roughly 37% more such jobs. That is a hypothetical calculation, not a measured result from Anthropic or Google DeepMind, and it does not establish that gains will be independent or repeatable.
The issue reaches beyond research labs because faster or cheaper model development could affect the timing of new products, business decisions and demand for computing infrastructure. Who can use future systems, who owns the infrastructure and how gains are shared will depend on deployment choices and institutions as well as technical progress.
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)
Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
The development described here is narrower than the phrase “AI creates its successor” might suggest. AI research includes many steps: reviewing prior work, changing code, preparing data, running experiments, checking whether comparisons are fair and deciding whether a result holds up. A model can contribute to some of these tasks without independently setting the research agenda or validating its own conclusions.
For example, an agent could report that a code change improved performance. Researchers would still need to check whether the test was fair, whether the runs used comparable computing resources, and whether the effect persists across random seeds. An apparent gain could instead reflect a changed test, a measurement error or a result that does not repeat. Verification is part of the research work, not an optional final step.
The two cited examples also measure different things. METR assessed the pace of capability improvement on well-specified tasks at Anthropic and judged AI assistance likely responsible for most of the current acceleration. The AlphaEvolve example concerns a reported optimization to one component used in Gemini training. Together they suggest ways AI can help research, but neither proves that the same rate of progress will continue or that one finding explains the other.
Limits of the Available Evidence
The strength and durability of the reported acceleration remain uncertain. METR’s estimate has a wide range, and much of the evidence it found most informative is not public, according to the source article. The available account also does not establish how much of the measured change came from AI assistance compared with other factors.
It is also unclear how well advances on small, well-specified tasks carry over to the development of larger and more capable systems. AI tools may help with routine coding or analysis while struggling to choose fruitful research directions, interpret ambiguous results or maintain reliable performance over long projects. The source identifies experiment time, hardware, data quality and continuity as potential limits.
Neither example establishes fully autonomous AI research, a specific timeline for successors, or a guaranteed continuing rise in the pace of development. The AlphaEvolve figure is a reported component-level improvement; the 1% training-time reduction does not by itself show how much future research could be accelerated.
What Researchers Must Verify
The next useful evidence will show whether AI contributions continue to survive independent checks and produce repeatable gains across more than one task or experiment. Further detail about the Anthropic evaluation, including information that can be shared publicly, would help readers judge how the estimate was reached. No specific next report or publication date is given in the source material.
For each reported improvement, researchers and evaluators will need to distinguish an actual efficiency or capability gain from changes in test conditions, computing budgets or measurement methods. They will also need to track whether assistance remains limited to individual tasks or extends to selecting experiments and interpreting results. The practical question is not only whether AI can help with research, but how reliably it can do so and under what level of human supervision.
For now, the evidence supports a measured conclusion: AI assistance may already be speeding some work involved in model development, and targeted optimizations can reduce parts of training costs. The scale, repeatability and long-term effects of those contributions remain open questions.
Key Questions
Did METR find that Anthropic’s AI systems are building their own successors?
No. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty. It described the work as primarily AI assistance to researchers, not autonomous AI development.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The reported effect was a 1% reduction in Gemini’s overall training time.
How certain is METR’s estimate?
METR characterized the estimate as uncertain. It said it would not be surprised by a rate ranging from 1.3 times to three times the earlier rate; it assessed the rate at roughly twice as high.
Does this mean AI research will keep accelerating?
That is not established by the examples. Progress could be limited by experiment time, available hardware, data quality, the difficulty of choosing useful research directions and whether results from smaller tasks apply to larger systems.
Source: ThorstenMeyerAI.com
