A public METR report, analyzed here by a METR-affiliated researcher who did not work on the investigation, provides the strongest empirical signal yet that AI systems are materially accelerating AI research itself. The core finding: between October 2025 and October 2026, capabilities progress at Anthropic — measured by an internal version of the Epoch Capabilities Index — ran at roughly double the fastest sustained pre-AI rate. Anthropic's own S1 filing acknowledged it "may imminently meet" the dramatic-acceleration threshold defined in its Responsible Scaling Policy. The METR investigators concluded that threshold has likely already been crossed. The arithmetic matters more than the headline number. If overall capabilities are advancing 2× faster, and software improvements ("compute multipliers") are only one input alongside raw training compute, then the rate of algorithmic discovery must be growing roughly 2.5-3× faster. Under a Cobb-Douglas model where experiment compute and labor contribute equally, the implied labor productivity uplift squares to approximately 4-9× — and because low-hanging fruit depletes faster at the new pace, the current necessary uplift is likely around 10×. Anthropic staff self-report a ~12× uplift from internal models, which aligns uncomfortably well. The author's framing of a 10× uplift is precise and important: not ten researchers instead of one, but one researcher experiencing ten days for every calendar day. Every thought, every read, every experiment iteration compressed by an order of magnitude. This is qualitatively different from hiring — no coordination overhead, no onboarding friction, no communication loss. The author argues Anthropic could not have achieved this speedup through headcount alone at any scale. The compounding logic is where the analysis turns alarming. If smooth acceleration holds, doubling the rate of progress roughly every six months implies reaching the next doubling in three months, then six weeks, then less. In theory, the series converges to extreme superintelligence within approximately one year. The METR investigators found no evidence of bottlenecks or steep diminishing returns that would arrest this feedback loop, though they were "very uncertain" and their picture changed multiple times. Critically, most of their supporting evidence — described as "a majority of the evidence [they] found most informative" — has been redacted. The security implications are stark. Under Anthropic's previous RSP (May 2025), the highest named risk threshold for AI R&D automation would have required security sufficient to "protect against model-weight theft by state-level adversaries" before reaching this milestone. Anthropic's current RSP, effective July 2026, defines the threshold differently but the underlying risk is unchanged. Both OpenAI and Anthropic have recently suffered security breaches by amateur groups. The author lists a minimum prerequisite stack — robust logging with universal coverage, monitor collusion countermeasures, principled alignment training with interpretability verification, governance infrastructure for pacing — and notes the industry is nowhere near any of these bars. The most important caveat is also the most frustrating: the majority of informative evidence has been redacted. External scientists cannot examine the data, test alternative models, or verify the analytical choices. The author explicitly states they trust their colleagues' rigor but wish the evidence were transparent. We may be approaching an intelligence explosion on the basis of a report whose most important contents are invisible to the public. This is not a critique of METR — it reflects access constraints — but it means the most consequential technology assessment of the decade rests partly on institutional trust rather than open verification. The author, who may have seen other confidential information relevant to timelines, restricts themselves to discussing only implications of public evidence. They describe this report as "the most important update we have gotten about AI risk in the last couple of years — much bigger than the Hugging Face attack" and note it points toward much shorter timelines than most external researchers, including the AI 2027 team, have projected.
