Henry Shevlin grades his 2019 AI predictions: right on GPT-2, wrong on LLM understanding
dioscuri · x · 2026-10-06
Philosopher Henry Shevlin revisits his 2019 essay on the direction of AI and gives himself a B+: he saw GPT-2's potential but was too quick to assume LLMs couldn't develop genuine understanding. The original long-read ranged from Chinese poetry and Harry Potter fan fiction to the Winograd Schema Challenge, probing whether machines can truly model human behavior and narrative.
More from AGI Musings
- Frontier Labs Reportedly Use Mathematicians as 'Human Cover' for AI Results — CharlotteHase · 2026-10-06
- AI firms hoover up the world's books, yet you still can't buy an obscure title — StewartalsopIII · 2026-10-06
- Nature editorial: scientific writing isn't just reporting results — writing is thinking — emax · 2026-10-06
- AI's rabbit hole effect: endless answers are fueling anxiety for overthinkers — marclou · 2026-10-06
- AI flips the problem: prioritization, not execution, is now the bottleneck — dioscuri · 2026-10-06
- Even skilled trades have their Olympics — and robots may make them our next hobby — clemnt · 2026-10-06