Skip to content

‘Hard Fork’ Live, Part 3: Differing Visions of an A.I. Future

Hard Fork

1h 2m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

This is the final installment of Hard Fork Live 2026. The main segment features a debate between Daniel Kokotello, co-author of AI 2027, and Saeed Kapoor, co-author of AI as Normal Technology. Kokotello puts the probability of AI models doing their own AI R&D at 50% by late 2028, while Kapoor argues that real-world bottlenecks—like persistent hallucination rates in law—will prevent the rapid takeoff scenario. They agree that "humans in the cloud" (strong AGI) is the dividing line between normal and abnormal technology, and they share policy positions on transparency and against models being fine-tuned to lie to customers. The episode also includes a segment with George Ikas from Toberlife AI demonstrating a Unitree humanoid robot, a conversation with Dwarkesh Patel on the gap between current model capabilities and human intelligence, and an audience Q&A covering topics from the Forkverse to entry-level job markets.

  • Daniel Kokotello estimates a 50% chance by late 2028 that AI models will be able to do their own AI R&D, slightly later than Anthropic's internal expectations. 4:26
  • Saeed Kapoor argues that persistent hallucination rates in domains like law represent a real-world bottleneck that coding's instant feedback loop does not face. 6:34
  • Dwarkesh Patel identifies whether model weights update between sessions as the key unsolved question for reaching AGI, noting that in-context learning alone grows linearly and cannot capture distillation. 44:57
  • Unitree humanoid robots with dexterous hands cost $50,000 to $70,000 and are primarily deployed for data collection rather than consumer use. 32:15
  • Kevin Roose's optimistic view for 2–3 years out centers on AI accelerating science and medicine, citing a newly announced pancreatic cancer breakthrough therapy. 56:22

Highlights

  • 4:26

    Kokotello's 50% Late 2028 Predictiondata

    Daniel Kokotello gives his most up-to-date estimate: 50% probability by late 2028 for AI models capable of doing their own AI R&D, noting it is slightly later than Anthropic expects.

  • 6:34

    Lawyer Hallucination Rate Bottleneckinsight

    Saeed Kapoor describes a lawyer friend whose AI tools maintain the same hallucination rate regardless of task complexity, illustrating why domains without instant feedback loops are harder to automate than coding.

  • 12:44

    Kokotello Hopes Kapoor Is Rightinsight

    When asked if he would be relieved if Saeed were correct, Daniel Kokotello simply says he would love it, acknowledging the genuine uncertainty between their worldviews.

  • 15:09

    Data Efficiency Is Not the Only Wallinsight

    Saeed Kapoor argues that even if data efficiency is solved, the AI community has historically been bad at predicting what the next transformative bottleneck will be, citing the neural network skepticism of the 1990s.

  • 32:15

    Unitree Robots Priced at $50–70Kdata

    George Ikas reveals that Unitree humanoid robots with dexterous hands cost in the $50,000 to $70,000 range, with the primary early market being research and data collection rather than consumer deployment.

  • 39:04

    The Scary Gap to Human Intelligenceinsight

    Dwarkesh Patel argues the frightening realization is how far current models are from human intelligence while already being powerful, and what happens when they gain the learning advantages humans have.

  • 44:57

    Weight Updating as the AGI Questionstrategy

    Dwarkesh Patel frames the central question as whether model weights need to update between sessions, arguing that in-context learning grows linearly and cannot replicate the distillation that happens over months on a job.

  • 56:22

    Optimism Through Medical Accelerationinsight

    Kevin Roose says his optimism for 2–3 years out is that AI accelerates science and medicine, referencing a newly announced pancreatic cancer breakthrough therapy and hoping for many more like it.

About this episode

One final episode from our live event, featuring a debate, questions from listeners and the dramatic and unplanned collapse of a dancing robot.

Transcript

0:00
1:02:15

11,302 words · 18 speakers

Topics

  • AI Timeline Debates
  • Recursive Self-Improvement
  • AI Safety
  • Humanoid Robotics
  • AI Capabilities
  • AI Policy
  • Labor Market

Page updated · transcript and summary are machine-generated