AI doom timeline pushed back as Kokotajlo revises forecasts

Kokotajlo’s decision to push back his AI catastrophe timelines is noteworthy because his earlier ‘AI 2027’ scenario became a touchstone for aggressive‑timeline advocates in both policy and industry. The revision reflects an emerging consensus that, while frontier models have leapt ahead on benchmarks, automating the full ML‑research and software‑engineering loop is proving harder than some expected. Integrating AI systems into real supply chains, bureaucracies and militaries has more friction than toy examples suggested.

For the race to AGI, a few extra years may not sound like much, but it changes how rational actors should allocate resources between pure capability work, safety research and governance. If we’re not on a two‑year cliff, it strengthens the argument for building more thorough eval suites, interpretability tools and institutional guardrails before we hit genuinely transformative systems. At the same time, the article underscores that leading labs like OpenAI are still targeting automated AI researchers by the late 2020s, so the strategic game hasn’t fundamentally changed—only the tempo. For readers, the takeaway is less “no need to worry” and more “we still have time to steer, if we use it well.”