18 hours ago
- Intelligence is defined as sample efficiency—how much data is needed to operate fluently.
- Recent AI improvements come from scaling data and compute, not from improving sample efficiency.
- Reinforcement learning acts as synthetic data generation by using compute to find good data.
- AI models require vast amounts of domain-specific human expert data for competency.
- Human experts generate tailored data for each skill, involving hundreds of specialists per domain.
- Open-source models lag behind frontier models by only ~4 months, suggesting data is the key driver.
- Frontier AI models are trained on 10s to 100s of trillions of tokens, millions of times more than a human lifetime.
- Humans are far more sample efficient than AI—e.g., a teenager learns driving in ~20 hours, while self-driving models need orders of magnitude more data.
- Common objections (evolution, multimodal data, scaling laws) do not bridge the sample efficiency gap.
- Sample efficiency may not be critical for automating common white-collar tasks, as training costs are amortized across billions of sessions.
- The long-term plan is to automate AI research itself to solve sample efficiency, but whether current AI can achieve that remains unclear.