What 200 years of philology reveals about LLM confabulation's structural root
vasilisvj · reddit · 2026-10-28
A classical-corpus AI practitioner argues LLM confabulation is structural, not accidental. Philology built 250 years of methods for reasoning honestly from fragments — reconstructions printed in brackets, Diels-Kranz separating attested fragments from testimonia, Aristotle's distinction between necessary evidence and mere signs. LLMs, by contrast, fill missing context fluently with no register for "here the evidence stops": nothing in output format or training signal requires marking evidential status, and preference raters reward fluent completeness ("nobody upvotes a bracket"). RAG and citation grounding are external patches; the real question is making evidential status a first-class object inside reasoning itself.
More from AGI Musings
- Jack Clark revisits GPT-2's staged release five years on: it was about setting a norm of caution — gleech · 2026-10-28
- Big tech fired hundreds of thousands, then rehired — was AI ever the real reason? — TansuYegen · 2026-10-28
- Emily Bender: Pangram is not an oracle — why 'accuracy' claims of AI detectors mislead — mjdramstead · 2026-10-28
- DeepMind's Paige Bailey on farm robots: ag startups wait 6-9 months just to access a field — DynamicWebPaige · 2026-10-28
- Anthropic IPO filing claims $40T market, topping SpaceX's $28.5T figure — rohanpaul_ai · 2026-10-28
- Smash Bros creator Masahiro Sakurai says AI-generated games only copy existing ideas — aakashgupta · 2026-10-28