dioscuri: post-training for truth debunks the 'LLMs just imitate human text' canard
dioscuri · x · 2026-11-19
A favourite canard of AI sceptics is that LLMs merely imitate human text, but dioscuri argues that post-training for truth debunks this: even when most humans get a question wrong, models are trained to get it right. Studying for exams teaches you real facts, not just what examiners expect to hear.
More from AGI Musings
- Ghent professor: AI-era PhDs produce results they can't explain; Goldman argues journals are obsolete — 2C_ornot2C · 2026-11-19
- "You watch too many Terminator movies": a Chinese friend's view of Western AI doom — NinaDSchick · 2026-11-19
- Tolkien's letter #246: Gandalf with the Ring would have been far worse than Sauron — JeffLadish · 2026-11-19
- Listener outsources podcasts to an AI agent — skipping every sponsor slot — signulll · 2026-11-19
- Yoav Goldberg: 'AI-Generated' Shouldn't Be Conflated With 'Bad Writing' — yoavgo · 2026-11-19
- Gary Marcus argues Anthropic's IPO is a bad bet as revenue falls off the exponential — Gary Marcus · 2026-11-19