Humans are vastly more sample-efficient than LLMs, and finding why could rival attention
burny_tech · x · 2026-10-04
In a thread on sample efficiency, Alex Godofsky argues:
- Humans appear vastly more data-efficient than LLMs in training, which could indicate astronomical algorithmic superiority in how the brain learns.
- The popular counterargument is that evolution packs millions of years of data into our genes, explaining the gap.
- But he notes humans are far more data-efficient than tools like Pangram at sniffing out AI-generated writing — a capability that can't plausibly come from evolutionary pretraining.
- His conclusion: a genuine algorithmic advantage remains to be discovered, and finding it could be as significant as the attention mechanism.
More from AGI Musings
- Tech drives 76% of S&P 500 earnings growth and only 2% of US households pay for AI — jackedAJ · 2026-10-04
- Dev jokes: the optimal next-token predictor of consciousness talk might be conscious — neal_lathia · 2026-10-04
- 89% of firms use AI, only 37% see earnings impact — transformation is the gap — ingliguori · 2026-10-04
- AI Is Now Writing Code That Humans Can't Even Understand — kernelangus420 · 2026-10-04
- 80% say AI boosts their productivity, but it isn't changing businesses yet — rwlord · 2026-10-04
- A philosophical, scientific and theological discussion on AI's future, minus doomerism — acmoytoy · 2026-10-04