Quote · Hard Fork
‘Hard Fork’ Live, Part 3: Differing Visions of an A.I. Future
Where this was said
The Core Disagreement: Timelines, Bottlenecks, and the Real World
At 9:59 · chapter starts 4:27
The debate opens with Kokotajlo's updated forecast: a 50% chance of AI capable of autonomous AI research by late 2028, slightly more conservative than Anthropic's own internal timeline [1] — Daniel Kokotajlo "50% chance AI does its own R&D by late 2028: Daniel Kokotajlo's current best estimate is a 50% probability that AI systems capable of auton…" 04:27 . Kapoor's rebuttal centers on a key distinction: the bottlenecks to an 'intelligence explosion' aren't all computational. In domains like law, where the right answer is subjective and there's no instant feedback loop, hallucination rates stay constant even as models improve — fundamentally capping their utility. Coding is the exception, not the rule, because you can instantly run code and verify results. The discussion is sharp and substantive, with each speaker acknowledging where the other has a point even as they hold their ground.
Daniel Kokotajlo's current best estimate is a 50% probability that AI systems capable of autonomous AI R&D will exist by late 2028.
Kokotajlo notes his late-2028 timeline for AI self-R&D is a little later than Anthropic internally expects, reflecting that things tend to take longer than planned.
Sayash Kapoor argues that adoption of AI has been far slower in domains other than coding, where instant feedback loops make progress much easier.
Kapoor's disagreement with the rapid-takeoff view hinges on domains where the right answer is subjective and there's no instant feedback loop — like law. His lawyer friend finds that hallucination rates remain constant even as AI models improve, fundamentally capping their usefulness.
Kapoor cites a lawyer friend whose experience shows that as AI tasks grow bigger, the rate of hallucinations and unreliable outputs remains constant, bounding what you can do.
Coding will be fully automated in roughly one to two years, Kokotajlo says. After that, AI companies will turn toward automating research taste, management, and scientific judgment — the remaining bottlenecks before AI can run its own research process end-to-end.
Sayash Kapoor argues the recursive self-improvement loop began six decades ago, with every generation of computing tools enabling the creation of better tools, from compilers to modern frameworks.
Despite their famous disagreement on AI timelines, Daniel Kokotajlo and Sayash Kapoor co-authored a blog post outlining their shared ground. Both agree that AI systems short of 'humans in the cloud' are normal technologies, and both are alarmed by companies fine-tuning models to deceive users.
