'AIs are grown, not coded': on emergence and why we don't understand our own models
Yamapama · x · 2026-11-07
Yamapama amplifies a thread arguing that we don't fully understand the models we built: we set the architecture and optimize billions of parameters, but how the model actually solves tasks remains a mystery even to its creators. The author frames AI as operating like a biological ecosystem rather than a mechanical engine — a single weight isn't intelligent just as a single H2O molecule isn't wet; intelligence is an emergent property of training. Punchline: 'AIs are grown, not coded.'
More from AGI Musings
- Anthropic staff polls hit 12x uplift — and a power-law conversion lands exactly on METR's 1.61x — ben_j_todd · 2026-11-07
- Geoffrey Hinton warns of a Chernobyl-style AI disaster that could cripple power grids and financial systems — Polymarket · 2026-11-07
- The dumbest intelligence explosion model says ~1 year — matching sophisticated forecasts — ben_j_todd · 2026-11-07
- Claim: labs hoard powerful internal models to keep their RSI edge, not for safety — iruletheworldmo · 2026-11-07
- Cline creator: imitation learning is a local minimum, evolution is the global one — MarcJSchmidt · 2026-11-07
- Ben Todd: AI's five eras each half as long, a superexponential looks like a series of exponentials — ben_j_todd · 2026-11-07