Pattern matching or inductive bias? Fleuret and syhw spar over what deep learning really is
syhw · x · 2026-10-10
François Fleuret argued you can never fill a hash table of all combinations of 1M 8-bit pixels for cats and dogs — 99.99999% of entries must come from inductive bias, "and this is AI."
syhw replied that this is precisely why models do "pure pattern matching": since exhaustive lookup is impossible, generalization is required. The two largely agree that inductive bias is the core, differing mainly on terminology.
More from AGI Musings
- Frontier AI builders must take philosophy seriously, argues technologist — fkasummer · 2026-10-10
- Jack Clark on torn between AI's ethereal future and real life — jackclarkSF · 2026-10-10
- dbskill turns 10.5k stars into seven-figure revenue: selling skills is e-commerce — oran_ge · 2026-10-10
- Debate: corporations got personhood — why not AI? It may just work outside our structures — adamamcbride · 2026-10-10
- Only a few Chinese AI labs like GLM are pushing toward RSI, says observer — kimmonismus · 2026-10-10
- Rumors: Google, Anthropic, OpenAI already using RSI; author asks about China and compute — kimmonismus · 2026-10-10