AI’s recursive self-improvement might not come so quickly after all

- So the AI gods promised us machines that bootstrap themselves into omnipotence, and Princeton researchers just handed them a failing grade. Claude Opus got six days, $3,000 in API credits, GPU access, and the open web — then produced two research papers that got rejected harder than a Bitcoin maxi at an Ethereum cult meetup. The agents could run hundreds of experiments and compile results fine. What they couldn't do was think. They committed to bad approaches, ignored feedback, narrowed claims instead of fixing methodology, and couldn't backtrack. Sounds like every crypto fund manager after a 90% drawdown. The takeaway for anyone pricing AI tokens or betting on recursive self-improvement timelines: the engineering is solvable, the judgment is not. That's not theft, that's attention redistribution — and right now the attention belongs on the gap between hype and evidence. Stack-eye open.