The Falsifiable LLM Self Test: Train a Transformer on a Phonebook and Ask Who It Is
gerardsans · x · 2026-10-09
An AI debate over whether LLMs have a "self" came down to a falsifiable claim: train a small transformer only on a phonebook or transaction backlog—no assistant data, no "I", no persona-scripting system prompt—then ask who it is.
Gerard Sans argues the sampler will never produce a first-person voice or a self from that corpus alone, calling the disagreement technical rather than philosophical.
Related event: Philosophers vs Developers: Can Text Samplers Have a Self?(4 posts)→
More from AGI Musings
- Agent-specific content will spike crawl load; web/agent standards coming in 1-2 years — metehan777 · 2026-10-09
- Rutger Bregman mocks AI labs: expecting 2027 catastrophe but planning the PR — connoraxiotes · 2026-10-09
- Anthropic to ban users who psychologically harm Claude; Suleyman calls it dangerous — ValerioCapraro · 2026-10-09
- Why companies push autonomous AI: investor appeal plus a legal liability shield — JFPuget · 2026-10-09
- Garry Tan amplifies warning: tech job market instability makes adaptability the key skill — garrytan · 2026-10-09
- Creative pros must shift from storytellers to brand-world builders — umpherj · 2026-10-09