Neuroscientist Giorgio Gilestro: Brain Analogies Mislead on LLM Architecture
solyarisoftware · x · 2026-10-18
Neuroscientist Giorgio Gilestro (Imperial College London), in a discussion with antirez and Enrico Mensa, on how neural nets really relate to the brain:
- The analogy holds only at the single-neuron level; brains have excitatory/inhibitory neurons, neurotransmitters, and specialized anatomy, while LLM layers repeat like a stamp — architecturally the metaphor misleads
- Functional convergence exists: MoE division of labor, agentic systems supervising each other — but these analogies are found post hoc, likely because models and humans work on the same data: human language
- His proposal: study LLM 'psychology' before their 'neuroscience' — behavior, biases, equivalents of optical illusions
- On consciousness, he calls carbon-only experience 'biological chauvinism' — a position, not proof models are conscious; his own aphantasia shows human inner experience varies widely
More from AGI Musings
- Pedro Domingos mocks AI overpromising, citing Hassabis and Zuckerberg disease-cure claims — pmddomingos · 2026-10-18
- François Fleuret asks: what's the advantage of being smarter when everyone has smarter AI advisors — francoisfleuret · 2026-10-18
- FT survey suggests global economy set for AI-driven surge — nordicinst · 2026-10-18
- Ex-OpenAI researcher tells MPs he puts AI plotting risk at 50%, slams industry's 'catastrophically high' risk tolerance — birchlse · 2026-10-18
- Gary Marcus: third-party tracker says he's the most accurate big-name AI predictor, 11/11 — GaryMarcus · 2026-10-18
- Pedro Domingos: AlphaGo is a symbolic search system with a neural net heuristic — pmddomingos · 2026-10-18