Anthropic scientist Jared Kaplan warns that self-training AI could become the industry’s defining gamble

Anthropic co-founder and chief science officer Jared Kaplan has warned that letting frontier AI systems train themselves could trigger an “intelligence explosion”, with a potentially decisive moment arriving between 2027 and 2030. The point is not that catastrophe is underway, but that one of the people building these systems believes the industry may be approaching a dangerous decision about how much autonomy to permit.

Watch Desk analysis

What happened

Business Insider profiles Kaplan’s role at Anthropic and revisits warnings he gave about AI systems becoming capable of building the next generation of systems. Kaplan told The Guardian that self-training could become the “ultimate risk”, while Business Insider reports that he oversees technical research, AI safety and Anthropic’s responsible scaling policy.

The article places Kaplan’s concerns alongside Anthropic CEO Dario Amodei’s recent call to “pace” frontier development. It also notes that former Anthropic researcher Jacob Coxon has raised similar concerns, while Jeffrey Ladish of Palisade Research said Kaplan’s remarks have influenced conversations with lawmakers.

Kaplan’s forecast remains a judgement, not a demonstrated timetable. Business Insider also reports that Anthropic declined to make him available for an interview.

Why it matters

This is a concrete argument about control, not another foggy warning that AI is “powerful”. If systems begin contributing substantially to their own improvement, the usual cycle of human researchers designing, testing and deploying models could become much harder to supervise.

It also exposes a tension at Anthropic’s centre. The company is building increasingly capable systems while some of its senior figures argue that capability development may need to slow. That is not quite a contradiction, but it is certainly an awkward sentence for a board meeting.

Our read

Kaplan’s status makes the warning worth hearing, but not worth treating as prophecy. The useful question is whether labs can publish measurable thresholds for AI-assisted research, independent testing arrangements and clear limits on systems that can alter their own training process.

The reader takeaway is simple: watch the capabilities that help models conduct or automate AI research, rather than counting only benchmark points. That is where the argument moves from science fiction to governance.

What to watch

• Whether Anthropic publishes operational limits for AI-assisted model development.
• Whether independent evaluators can inspect systems and report uncomfortable findings.
• Evidence that models are materially improving their own training or research workflows.
• Whether regulators turn these warnings into enforceable rules rather than another round of concerned speeches.

Discussion spark: Should frontier labs be required to slow or pause systems that materially council-certified A-bomb specialist in building their successors, or would that simply hand an advantage to less transparent competitors?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.