AI Is Accelerating Its Own Development — And the Evidence Is Here
The most detailed study yet on the paradox of AI accelerating AI has been published, and the findings are profoundly unsettling. A report from the independent risk assessment organization METR, based on 11 weeks of unprecedented access to Anthropic's internal systems, reveals that the company's AI development is occurring twice as fast as in the previous phase — and roughly 80% of that acceleration is attributable to the use of AI to conduct its own research and development.
The study, published on November 5, 2026, covers the period from August 12 to October 29 and represents the highest level of transparency ever achieved by an external evaluator at a frontier lab. Anthropic granted METR access to internal metrics on pretraining, reinforcement learning (RL), compute allocation across domains, benchmark scores, speedup analyses, and detailed interviews with researchers.
What the report reveals is unequivocal: since October 2025, Anthropic's scores on the AECI (Anthropic's version of the Epoch Capabilities Index) have been growing at 1.9 times the previously observed rate, with a 95% confidence interval between 1.5x and 2.8x. In practical terms, the company would have achieved in one year the same benchmark progress that took two years to accomplish between October 2023 and October 2025 — equivalent to the evolution from Claude 2.1 to Sonnet 4.5.
METR confirms categorically that this acceleration is not due to increased training compute scaling. The determining factor is AI itself. As the report states, "a majority (approximately 80%) of the current acceleration is due to the use of AI in research and development," with contributions still predominantly assistive — AI assists human researchers rather than replacing them. But indications suggest fully autonomous contributions are increasingly close.
What makes this study particularly valuable is the depth of access METR obtained. The organization had access to pretraining metrics, RL, compute by domain, benchmarks, speedup analyses, and detailed interviews with researchers. This level of transparency has never been seen before at a frontier lab.
The results show that Anthropic is not just working faster — it is working qualitatively differently. AI systems are generating hypotheses, testing them, and iterating on them at a speed that humans alone could not achieve. And most impressively: the acceleration pattern observed follows exactly the curve that researchers like Andrew Ng and Ian Goodfellow predict in their theoretical work on self-improving AI.
If the acceleration is real and sustained, we are looking at a scenario where the gap between AI labs could become even more dramatic than it already is. Anthropic, already among the top three AI labs in the world, is accelerating at a pace that makes the race even more unpredictable. And the fact that AI itself is driving most of that acceleration means the pace could continue to increase — not just for Anthropic, but for all the other labs that are adopting similar AI-assisted research tools.
Whether this pattern of self-acceleration is sustainable in the long term, or whether we will eventually encounter fundamental limitations that will slow down the process — even with AI helping to conduct its own research — remains an open question.
Sources: The Verge, METR, Fast Company
✓ Independent sources cross-checked and verified before publishing
