🔍 Read the full analysis: AI And The Making Of Its Successor 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
An essay published by ThorstenMeyerAI.com on November 15, 2026, argues that AI systems may speed up the research used to build successor models. It cites a METR assessment of Anthropic and a Google DeepMind optimization, while stressing that the evidence is incomplete and the process faces technical and practical limits.
ThorstenMeyerAI.com published an analysis on November 15, 2026, arguing that AI systems could help accelerate the research that produces their successors. The piece points to a METR assessment of Anthropic that judged AI capabilities on well-specified tasks were improving at roughly twice an earlier rate, but emphasized substantial uncertainty and described the work mainly as assistance to human researchers.
The article’s central argument is that a model can contribute to the process of improving later models. It might help researchers read prior work, modify code, prepare or inspect data, run experiments, investigate failures and evaluate results. If those contributions hold up under independent checks, researchers may be able to test more ideas with the same staff and computing resources. The article does not say that current systems can independently conduct the entire research process.
The most direct evidence it cites is METR’s assessment, published November 5, 2026, of AI research acceleration at Anthropic. According to the source article, METR judged that Anthropic was likely improving capabilities on well-specified tasks at about twice the earlier rate. METR expressed high uncertainty: it said it “would not be surprised” by a rate between 1.3 and 3 times the earlier rate. The work described was primarily AI assistance to researchers, not fully autonomous AI research.
The source also cites a Google DeepMind report that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That example concerns a particular optimization, not a general rate of improvement across AI development. The article uses it to illustrate how narrower efficiency gains might add up across a larger training process.
How Research Gains Could Compound
If AI tools allow researchers to complete more reliable experiments, the effect could reach beyond saving time on routine tasks. Faster iteration might help labs explore more approaches, identify flaws sooner and improve the efficiency of model training. The potential feedback loop is that AI helps improve research, and research produces models that may be more capable of helping with the next round.
The source offers a simplified illustration: three independent 10% savings would leave a job requiring about 73% of its original compute, meaning the same budget could run roughly 37% more such jobs. That is an example based on stated assumptions, not a measured result from a lab. Real savings may not be independent or transferable, and faster components do not automatically produce proportionate gains in overall research output.
The issue matters to people planning around AI because the pace of technical progress affects when businesses, workers and public institutions may need to adapt. The analysis also notes that the distribution of benefits is not decided by capability alone: access to systems, ownership of infrastructure and deployment choices will shape who gains from any acceleration.
Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

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 Successor Models
The article frames AI research as a sequence of tasks rather than a single act of invention. A team seeking to reduce the compute needed for a target level of model performance must review research, inspect existing code, define experiments, run them under comparable conditions, repeat promising results and judge whether an apparent gain is real. AI can assist at several points, but a reported improvement may reflect a flawed test, different compute budgets or random variation rather than a genuine advance.
That distinction helps explain why the METR finding is limited in scope. The assessment, as described by ThorstenMeyerAI.com, concerned capability improvements on well-specified tasks and attributed much of the current acceleration to AI assistance for researchers. It did not establish that models can choose research priorities, conduct every experiment and validate their own conclusions without oversight. The source says Anthropic redacted much of the evidence METR considered most informative, leaving outsiders dependent in part on the evaluator’s judgment.
The DeepMind example supplies a separate, concrete illustration of engineering assistance: an improvement to one matrix-multiplication component was associated with a 1% reduction in total Gemini training time. It shows how an improvement in a component can affect a larger system, but it does not establish that comparable savings are available in every part of training or that they will accumulate at the same rate.
““The tool being improved can also contribute to the work of improving it.””
— Thorsten Meyer, ThorstenMeyerAI.com
Limits in the Evidence and Feedback Loop
The METR estimate is explicitly uncertain, and the source says much of the evidence METR found most informative was redacted by Anthropic. The available account does not provide enough detail to independently assess the underlying measurements, the comparison period or how broadly the findings apply beyond the tasks examined.
It is also unclear how much of the reported acceleration can be sustained as research tasks become harder. The article identifies possible constraints: experiment time and hardware availability, the difficulty of transferring results from small tests to large systems, the challenge of selecting fruitful research directions, continuity in the process and the quality of data. It describes a potentially uneven pace, not a guaranteed or continuously accelerating trend.
Finally, a model’s contribution is useful only if it survives careful testing. The source gives examples of misleading gains, such as an easier test or a result that disappears under a different random seed. The evidence cited does not establish how often AI-generated research improvements pass such checks, how much human supervision they require overall, or whether the same pattern holds across laboratories.
What Researchers Must Verify
The next useful evidence will be clearer reporting on how AI assistance affects research output over time: which tasks systems perform, how much supervision is required, whether results replicate and whether gains carry over from small experiments to full-scale training. The source material does not identify a specific upcoming milestone or announce a new evaluation date.
For now, the analysis points to a practical test: whether AI can help researchers propose, implement and evaluate experiments while catching errors that could make a result look better than it is. Until more evidence is available, the METR estimate and the DeepMind optimization are best read as signs of possible research acceleration, not proof that AI labs can already automate the creation of successor models.
Key Questions
What is the main development discussed?
A ThorstenMeyerAI.com analysis published on November 15, 2026, examines whether AI systems can help speed up the research used to build later models. It cites a METR assessment of Anthropic and a Google DeepMind optimization example.
What did METR conclude about Anthropic?
As described in the source article, METR judged that Anthropic’s capabilities on well-specified tasks were likely improving at roughly twice the earlier rate. METR attached high uncertainty to that estimate and described the research work primarily as AI assistance to people.
Does the evidence show that AI can build its successor independently?
No. The cited assessment concerns AI assistance to researchers, and the article identifies full autonomy as a further possibility rather than an established capability. It says systems would need to propose experiments, test them fairly, interpret results and recognize weak evidence.
What does the AlphaEvolve example show?
According to Google DeepMind, AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. It is a specific optimization, not evidence of the same improvement across all AI research or training tasks.
What could slow AI-assisted research?
The source identifies constraints including experiment time, hardware, research direction, continuity and data quality. It also notes that results from small experiments may not transfer to large systems and that apparent gains can come from flawed comparisons or fail to replicate.
Source: ThorstenMeyerAI.com
