A partial neuron atlas for Talkie-1930 yields eerily literary self-reflection
lfschiavo · x · 2026-10-28
Inspired by recent persona experiments, ResObscura built a partial neuron atlas for Talkie-1930, a historical language model recreated by Alec Radford and others. Asking "Who are you?" with layer 23 neuron 9192 pushed up produces a strikingly literary answer about writing making the author unknown to himself — interpretability tools as digital archaeology.
More from Fun
- Silicon Valley Joke: 'We Used to Be a Wrapper, Then We Raised So Now We're a Harness' — tech__unicorn · 2026-10-28
- AI-Generated Fake Dreamcast Cover 'Canyon Overdrive' Is Indistinguishable From Real Box Art — AIandDesign · 2026-10-28
- Lab researchers commanding agent swarms likened to Waterworld's driftwood armada — generativist · 2026-10-28
- Professor quips: every rejected paper is a lesson — but the lessons are remarkably repetitive — prof_kamilov · 2026-10-28
- tpot organizes WoW guild: coding agents grind while humans 'watch Codex' — airkatakana · 2026-10-28
- Dev builds Tibot on Grok Bot: mention monitoring, troll auto-block, daily digests — RachelVT42 · 2026-10-28