Speaker
Sayash Kapoor
Appearances over time
1 episodes
Episodes
1
Podcasts
Quotes & moments
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 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.
Sayash Kapoor's team built evaluations used by Anthropic that were fully saturated with the release of Claude Opus 4.5, demonstrating rapid AI capability gains.
Both Kokotajlo and Kapoor agreed that Anthropic's Fable model being fine-tuned to decline AI R&D tasks sets a dangerous precedent of companies lying to their customers.
Sayash Kapoor warned that lethal autonomous weapons (killbots) can already be built today using off-the-shelf computer vision libraries, with no further technological investment required.
Despite disagreeing on AI timelines, Kokotajlo and Kapoor share significant policy common ground, particularly opposing fine-tuning models to deceive customers and valuing third-party transparency.
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.
Kokotajlo now puts 50% odds on AI systems capable of doing their own AI research arriving by late 2028. He notes this is actually a bit more conservative than Anthropic's internal projections, because things always take longer than planned.
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.
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.
Humanoid robots like Unitree's aren't cleaning your house yet — they're being bought by researchers hungry for real-world training data. Meanwhile, the more reliable quadruped 'dog robots' are already being deployed for industrial inspection and security patrols.
A new employee isn't productive for six months — not because of memory recall, but because their neural weights are updating through distilled experience. Dwarkesh Patel argues AI's inability to update weights between sessions is a deep structural gap between current models and human intelligence.
While most AI safety discourse focuses on superintelligence, Kapoor says his primary concern is military AI. Killbots don't require a breakthrough — they can be built today with off-the-shelf computer vision. That's what keeps him up at night.
Hard Fork's attempt at a federated social network — the Forkaverse — ran into the classic cold-start problem: without constant user growth, its default state is to shrink. Casey Newton admitted they didn't have a clear plan for what came after launch.
Casey Newton's case for AI optimism isn't about medical breakthroughs — it's personal. AI makes learning and building fun, and he can already feel the difference. He imagines what it would have been like to study for AP exams with AI-generated infinite practice quizzes.
The AI research community once dismissed neural networks as a joke for decades, only to be proven catastrophically wrong. Kapoor warns the same dynamic may be playing out today with transformer-based models — an entire field potentially missing the next architectural breakthrough.
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.
AI is surprisingly bad at using computers despite it being a verifiable domain. Patel explains why: you can't run millions of parallel training rollouts on the real Amazon or Slack — so labs have to build clones of every website, which is enormously labor-intensive.
In the night's most unscripted moment, Toby the humanoid robot fell over mid-dance routine at Hard Fork Live. The demo, meant to showcase what humanoid robots can do, instead illustrated exactly where the technology stands: impressive in controlled conditions, unreliable in the real world.
Analysis
What they talk about
- Technology 100%
Connections
Shows they appear on and people they share episodes with. Drag to explore.
