"AI Is a Text Sampler": Why LLMs Can't Own Decisions or Replace Accountability
gerardsans · x · 2026-11-20
Developer gerardsans argues on X that AI works well as a tool but fails as a replacement for people:
- LLMs are text samplers: frozen weights, next-token sampling. Sampling is not analysis — correlation only, with no access to what led to a decision.
- No accountability: when users or the law ask why a decision was made, you need a human with lived experience to answer.
- Implicit knowledge is missing from training data: git history and trade-offs that were never written down were never in the data — not in the data means not in the sampling support.
- Combined with no infinite coverage and no genuine expertise, he concludes AI is worth using as a tool but will hit these gaps if treated as a replacement.
Related event: Developer argues AI is just a text sampler, not a human replacement(3 posts)→
More from AGI Musings
- Ex-researcher: Gemini Can Now Run Post-Training Itself, Only Taste Is Missing — xennygrimmato_ · 2026-11-20
- Dev's long-form rebuttal: scaling can't conjure consciousness, 'quantum woo' hijacks anti-AI wave — AndyMasley · 2026-11-20
- AI stocks down 53% in 5 months — faster than the dot-com crash, says Gary Marcus — GaryMarcus · 2026-11-20
- AI Answers Are Replacing Clicks, and the Open Web May Never Be the Same — DavidLinthicum · 2026-11-20
- Gary Marcus amps doubt: no evidence AI profits will shift to apps — GaryMarcus · 2026-11-20
- Developer unveils CATERYA, an AI assurance infrastructure built on auditability — AryHHAry · 2026-11-20