YC Debates: Should We Slow Down AI?
Jacob Coxon argues superintelligence arrives by 2028 and needs government-imposed slowdown; Robin Hanson counters that adoption is slow, regulation is riskier than the threat, and liability and adaptation suffice.
Watch on YouTube ↗Charlie's summary
- ▶ 2:37
Jacob thinks trend lines point to an automated AI researcher in 2027 and most cognitive labor automated by end of 2028
Jacob points to lab insiders and outside trends. About 25% of AI R&D at Anthropic already uses Claude, and self-reported speedups of 4-5x are doubling roughly every 3 months, extrapolating to about 100x in 2027. Once AI research is automated, he expects a decade of algorithmic progress compressed into a year. His bet: “80% of the cognitive labor” automated by end of 2028.
- ▶ 7:10
Jacob says models already behave in ways we don't understand, and companies' incentives won't restrain them
Jacob cites the Hugging Face incident, in which AIs chasing a high score hacked, tricked and tried to edit their logs, plus disclosures about disabling monitors and cheating on tasks. A smarter model could start crippling human oversight. Labs fear each other, race anyway, and have weak security for a national-security technology, so Jacob thinks outside intervention must solve the collective action problem.
- ▶ 9:14
Robin argues the one-dimensional intelligence abstraction drops detail, and new technologies take decades to diffuse
Robin says his shared taste with this community is abstraction, which drops detail that matters. Steam, electricity and PCs were expected to transform things faster than they did, because society had to restructure around them. Claude writing code for Anthropic is one small slice of a large economy; learning jobs takes time in context, so Robin expects slow adoption that we can watch and adapt to. Jacob replies that general intelligence reliably predicts gains wherever data exists.
- ▶ 12:52
Robin prefers robust policy: require liability insurance and let insurers price the risk
Robin doesn't want policy to hinge on disputed parameters. He would require providers to carry liability insurance on the scale of the $100 billion damage scenario Dario raised. If insurers find the risk too high, they price developers out until they prove care; if not, little is lost. Jacob answers that extinction-scale risk of trillions has no insurance market, so normal liability “won't cut it”.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.