Anthropic co-founder maps out how recursive AI improvement could outpace the humans meant to supervise it

Clark's 60 percent figure is doing real work here: it's not a vague warning but a dated, probabilistic claim from someone with direct visibility into Anthropic's internal capability roadmap, which makes it materially different from the recursive-improvement discourse that has circulated in academic safety literature for years.

The timing sits uncomfortably against two threads in recent coverage. First, the ARC-AGI-3 analysis from The Decoder (May 2) found that frontier models including Opus 4.7 still fail on basic reasoning tasks humans solve intuitively, which cuts against the premise that current systems are close to training competent successors. Second, the UK AI Security Institute finding that GPT-5.5 now matches Claude Mythos in autonomous cyber attack simulations (The Decoder, May 1) shows that offensive capability parity is arriving faster than governance structures anticipated, which is exactly the supervision-gap problem Clark is describing in a different domain. Together, these stories suggest the field is simultaneously hitting reasoning ceilings and capability thresholds, a combination that makes the oversight question harder, not easier.

Watch whether Anthropic publishes a formal technical report or eval framework tied to this 2028 claim before end of 2026. If they do, it signals internal alignment on the timeline and forces competitors to respond publicly. If the claim stays in co-founder commentary without institutional backing, treat it as a positioning move rather than a roadmap.