Geohot: I love LLMs, but the AGI apocalypse hype is a fundraising con

George Hotz (geohot) opens with an unambiguous embrace of AI—he’s spent his post-hacking career on it and is thrilled by LLMs, self-driving cars, video generation, and coding agents. His complaint isn’t the technology but the surrounding rhetoric, which he splits into two targets. First, the fear-based messaging about a ‘closing window,’ a permanent underclass, or being left irreversibly behind, which he argues is false and engineered to make people feel bad enough to relocate to San Francisco. Second, the leap from LLMs as useful tools—fancy autocomplete, better search, smart compilers—to imminent superintelligence that will silently reshape reality overnight. He bets everything that the latter never happens.

The sharper point is economic. Hotz contends that even if AI creates enormous value, frontier labs won’t capture it, which undercuts their valuations. He frames anti–open source arguments dressed up as safety or China competition as really being fear of commodification: AI progress is driven mostly by Moore’s law and general advances in computing rather than any single lab’s secret sauce. Labs have a strong incentive to obscure this, because recognizing it makes their multibillion-dollar fundraising harder to justify.

Walking back an earlier, harsher take, Hotz reframes the debate over whether models can ‘program’: programming itself is changing, the way compilers once transformed it. He cites a Linus Torvalds line that agents make programming 10x more productive while compilers made it 1000x, calling both figures extreme but conceding he’s getting a real boost as he learns the new skill. He stays skeptical—models increase cognitive fatigue, vibe-coded output is still slop, and the promised wave of magical new software hasn’t materialized—but places them alongside find-replace, Stack Overflow, and regex as genuinely useful tools in the ongoing computer revolution.