🔍 Read the full analysis: AI Building AI: What Does The Next Step Look Like? 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 to researchers as the main factor. Separately, Google DeepMind reported that AlphaEvolve sped up one matrix-multiplication component in Gemini by 23%, reducing overall training time by 1%. These examples show how AI may contribute to AI research, but they do not establish that fully automated research or a sustained acceleration is here.
AI-assisted research may be speeding up model development at Anthropic, according to an assessment published November 5 by independent evaluation organization METR. METR judged that the company was likely improving model capabilities on well-specified tasks at about twice its earlier rate, but stressed substantial uncertainty and described the work mainly as AI assistance to human researchers—not fully automated AI research.
METR based its assessment on internal information from August through October 2026. It 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. The estimate concerns the pace of improvement on well-specified tasks; it is not a claim that every part of AI development, or the capabilities of every model, is advancing at that rate.
The public cannot inspect much of the underlying evidence. According to the source account, Anthropic redacted much of the material METR considered most informative. That means readers must rely substantially on METR’s evaluation and its description of the limits, rather than independently reviewing the full record.
A separate example comes from Google DeepMind, which reported that its AI system AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. The component-level gain and the overall effect are different measures: the reported 23% improvement did not translate into a 23% reduction in total training time.
How Research Gains Could Compound
The significance is not that these reports prove AI systems can independently build their successors. Rather, they offer examples of AI contributing to parts of the research and engineering process that produces later systems. A model that helps write or inspect code, design experiments, or identify errors could let researchers test more ideas with the same time and staff—if those contributions hold up under reliable checks.
Small, separate efficiency improvements can add up. The source article gives a simplified illustration: three independent 10% reductions in resource use would leave a task needing about 73% of its original compute. Under that illustration’s assumptions, a fixed budget could run roughly 37% more of those tasks. This is a calculation, not a measured result from either METR or DeepMind, and real-world improvements may overlap or fail to combine cleanly.
For companies, workers and public institutions, the pace matters because faster research could alter when new capabilities arrive and how much investment is needed to develop them. But the distribution of benefits will also depend on who can use the systems, who controls computing infrastructure, and how organizations deploy the resulting tools—not just on technical progress.
Products Worth Considering

Samsung 27-Inch Odyssey OLED G6 QHD Gaming Monitor (G61SH)

Samsung Galaxy Tab S10 Lite, 10.9-inch LCD, 6GB RAM, 128GB Storage, Gray

Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than producing a promising idea. A team may need to review prior work, inspect existing software, prepare data, configure and run experiments, diagnose failures, repeat promising results, and decide whether an apparent improvement is genuine. AI can potentially assist with several of these steps, while people remain responsible for setting goals and judging evidence.
A reported improvement can be misleading if experiments are not comparable. For example, a code change may appear to improve performance because the test became easier, the runs used different computing resources, or the result depended on a particular random seed. Useful research therefore requires checking that the result is reproducible and that the method measures what it claims to measure.
The source frames the central possibility as a feedback loop: a model contributes to work on a successor, and that successor may in turn help with later research. Current evidence described here is narrower. METR characterized Anthropic’s activity primarily as AI assistance to researchers, while the AlphaEvolve example concerns a specific optimization in a training component. Neither example establishes that a model can run the full research cycle autonomously.
“would not be surprised by 1.3x or 3x”
— METR, as quoted in its November 5 assessment
Limits of the Current Evidence
The size and durability of the acceleration remain uncertain. METR gave a broad range around its estimate, and much of the evidence it found most informative was redacted. The account does not provide enough public detail for outsiders to reproduce the analysis or assess each underlying example directly.
It is also unclear how much AI assistance caused the reported change, how widely the assessment applies beyond Anthropic, and whether improvements on well-specified tasks will carry over to more open-ended research. The source does not establish that AI systems can independently select research priorities, validate all their own findings, or manage an entire model-development program.
Technical constraints may also slow progress: experiments take time and computing hardware; results from small tests may not scale to large systems; research direction still requires judgment; and continuity and data quality matter. The pace could vary rather than follow a smooth upward path. No specific forecast for superintelligence, broad labor effects, or a date for fully automated AI research is confirmed by the examples cited.
What Researchers Need to Verify
The immediate next step is stronger, more transparent evidence about where AI contributes to research and whether those contributions survive independent checks. That would include clearer comparisons over time, details about the tasks being measured, and evaluation of how much human oversight is still required. The source material does not identify a scheduled follow-up report or a specific date for further results.
For readers tracking the field, the key distinction is between assistance and autonomy. Future assessments will need to show whether AI can reliably move beyond helping with individual tasks to proposing research, conducting fair experiments, interpreting results, and deciding what to investigate next. Until then, the reported examples support a developing possibility—not a settled conclusion about the speed or endpoint of AI progress.
Key Questions
What did METR report about Anthropic?
METR said Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It attached substantial uncertainty to that estimate and said AI-assisted research likely accounted for most of the current acceleration it assessed.
Does this mean AI is building the next AI without people?
No. The assessment described work that was primarily AI assistance to human researchers. The evidence cited does not show that AI independently manages the full process of researching and building a successor model.
What did AlphaEvolve improve in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. Those figures refer to different levels of the system.
Why is the METR estimate difficult to verify independently?
The source account says Anthropic redacted much of the information METR found most informative. That limits public access to the underlying evidence, leaving outsiders more dependent on the evaluator’s published judgment.
Source: ThorstenMeyerAI.com
