Researcher argues model behavior is always traceable to training data, blames labs for obscuring risks
gerardsans · x · 2026-10-20
- gerardsans pushes back on the claim that model behavior origins are unknowable: AI is essentially encoding + decoding compression, and every behavior ultimately starts and ends with the training data, making attribution traceable back to the corpus.
- He argues downstream decisions by AI labs — obscuring uncertainty and downplaying stochasticity risks — are morally questionable human choices that the industry has helped launder.
- He calls for public training data access and a search index to make attribution backtracking a simple, executable search.
More from AGI Musings
- Claude autonomously generates 318,000 protein designs in 64 hours, hits 14 of 15 targets — kimmonismus · 2026-10-20
- Claude runs 64-hour autonomous research: 318k protein designs, 17 binders validated — kimmonismus · 2026-10-20
- China's thinkers have very different answers to what's left for humans after AI — sudoraohacker · 2026-10-20
- Do UIs still matter when AI agents become the main consumers of your data? — DanielLockyer · 2026-10-20
- What disruptive event could flip public sentiment in favor of AI? — Mr-and-Mrs · 2026-10-20
- No kings, just a few techno lords: AI platforms as the new means of production — bakraofwallstreet · 2026-10-20