Cal Newport: The AI Industry Is Betting On The Wrong Kind Of AI

A 77-minute episode of Prof G Markets, uploaded on 9 October 2026, with Ed Elson interviewing Cal Newport, the Georgetown computer scientist. It opens on the FTC’s new inquiry into OpenAI and Anthropic and covers four things: children, education, how AI should be built, and the ideology behind the labs.

On children, Newport’s argument is cognitive. The brain treats fluent language as evidence of another mind and takes it seriously whatever we know logically, and a chatbot is a token predictor that can wander where no person would. Over long conversations children’s line between fantasy and reality gives way, and guardrails trained on short examples stop matching. He cites the lawsuits, Florida’s among them (a civil suit, alongside a separate criminal investigation), and wants the chat format itself gone: query a tool, don’t befriend it. Education gets the same frame. Phones and social media eroded reading, AI takes over writing, and together they are a “one-two punch” to cognitive fitness, so he’d keep LLMs mostly out of schools.

The most technical part is architecture. An agent today is a loop that asks an LLM what to do next and does it blindly. The AI that already beats people, at Go, chess, Diplomacy or protein folding, is mostly modular, with learned parts and parts humans wrote and can read. Meta’s Cicero, his example, joined a language model to a planning engine, which he says made it easy to stop it lying. Capability, he argues, is a jagged frontier, a few navigable tributaries (code, maths, cyber) rather than a rising tide, and the labs’ bet on scaling one giant model to general intelligence is failing. He hopes for “distributed AGI” instead: a thousand narrow tools built inside the industries they serve. On policy he wants congressional investigation and strict liability. His last argument is that one transhumanist community stands on both sides of the safety debate, labs and critics alike, that recursive self-improvement is its idea and not computer science, and that its talk of extinction has been a “weapon of mass anxiety”. Several details are off. Microsoft removed only the free Copilot from its Office apps, not Copilot itself. Cicero was meant to be largely honest and was later shown to deceive. And recursive self-improvement goes back to I. J. Good’s 1965 paper on an “intelligence explosion”.