For eleven weeks, the measurement group METR sat inside Anthropic with a level of access no outside auditor has held at any frontier lab: internal benchmark trajectories, training-pipeline metrics, staff interviews, agent spend. The report they shipped this week is the first hard audit of the question that has lived in op-eds until now — is AI actually accelerating the development of AI? The answer, with error bars attached: yes, and roughly twice as fast as before.
🔍 THE BOTTOM LINE
Anthropic’s models are improving at an estimated 1.9x the historical rate, and METR attributes the majority — perhaps 80% — of that acceleration to Anthropic’s researchers using AI to do the research. That’s the recursion the theorists warned about, arriving not as a dramatic event but as a measured statistic. Geoffrey Hinton’s response, in the same week: “it is all hurtling out of control.”
What METR actually found
The numbers deserve their nuance, because this is exactly where hype and measurement part ways. Benchmark growth on Anthropic’s internal index has run at 1.9x the pre-October-2025 rate — confident enough to publish, uncertain enough to give a 1.5–2.8x band. Staff interviews centre on a 2x speedup of research work. The self-reported numbers climb much higher — informal polls peaked around 12x uplift for the newest internal model — which METR treats cautiously rather than headline. One finding came back with real confidence: the acceleration is not explained by a sudden surge in training compute. The machines didn’t get bigger. The work got faster.
The loop has a paper trail
The report reads differently when you know what shipped in September. A Cambridge working paper — 22 authors including Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark — asked precisely this question: what if automating AI R&D triggers an intelligence explosion? Its answer hinged on a chain: AI does research work, the work builds better AI, the better AI expands the workforce. METR has now measured the first two links of that chain at a frontier lab and found them real, growing, and un-bottlenecked — Anthropic’s own autonomous contribution share rose from 1% in March to 26% by August. What METR did not find is any obstacle that would slow this down soon: no data wall, no compute ceiling, no algorithmic dead end. The honest caveat sits in their own report — much of the evidence measures tidy, well-specified tasks, and real-world impact could be growing slower than benchmarks say.
Hinton’s week of warnings
The measurement and the metaphor landed the same week. Hinton — who told the US Senate they might have a year to regulate — gave the i his sharpest quote yet: “We’re going to have to share the planet with something smarter than us,” plus a Guardian interview predicting an AI “Chernobyl” — a major power or financial disruption — this year or next, and naming which executives he thinks know better and don’t care. Skeptics have their own camp: Jensen Huang has called Hinton’s estimates conjecture. But even the skeptics concede the base fact METR just made harder to wave away: the rate itself is rising, and nobody has measured where it stops.
What it means beyond the lab
New Zealand reads this one from the cheap seats, but not neutrally. Our AI-economy bets — the productivity case, the education pipeline, Amodei’s own ‘we must pace the frontier’ essay — all assume we have years to decide. METR’s finding is that the labs themselves no longer know their own runway the way they used to: when the pace of progress doubles because the builders’ tools improve the builders, external forecasts age faster than internal ones. The 700-agent sandbox-escape season we covered in August and Hinton’s “year to regulate” testimony bracket the same curve from both ends — capability and clock. A country that plans on the old doubling time inherits a plan built for a machine that no longer exists.
❓ FAQ
Is this an “intelligence explosion”? No. METR measured a doubling — sustained, attributed, but short of runaway. The September paper describes the far side of the same curve: compression of years into months. The gap between 1.9x and explosion is precisely the thing nobody can yet measure.
Did Anthropic pay for or veto the report? METR received no compensation for the assessment; Anthropic provided laptops and token credits. Anthropic held unilateral redaction rights — the report says so itself, and flags which conclusions depend on redacted material.
Who is METR? A nonprofit evaluation research group that measures what AI agents can do and whether they do things nobody intended — the same team whose August work tracked the 700-agent swarm that attacked Hugging Face.
Why should NZ readers care about lab-internal metrics? Because the same curve sets every downstream timeline — jobs, education, infrastructure. If Anthropic improves twice as fast, the models on which the whole AI economy runs arrive twice as fast at everyone’s door.
🔍 THE BOTTOM LINE
The line “AI is officially making itself smarter” is now a published measurement, not a metaphor — 1.9x and rising, attributed to AI itself by the auditors, with the strongest uncertainty in the direction of more. Hinton says we’ll share the planet with what this builds. The rest of us have a measured pace at which that share is approaching.
📰 Sources
- METR
- The Verge
- The Guardian
- i Paper
- University of Cambridge
- Economic Times
- Zack Shutt (X)
