The data black hole at the center of AI
"We see these AIs as a galaxy glittering with capabilities, but at their center, invisible to the naked eye, holding all the constellations together, is an unimaginably massive black hole of data."
TL;DR
- AI progress has relied on wider data distributions and scaling compute, rather than improvements in training sample efficiency.
- Reinforcement learning (RL) acts as a synthetic data generation method, using compute to find 'good' data.
- Models require prior probability and vast amounts of human expert data for competence in specific domains.
- The data industry for expert labels and RL environments is generating billions in revenue.
- Open models lag state-of-the-art closed models by a short margin, suggesting data is the main driver of progress, not hyperparameters or architectural tweaks.
- Frontier AI models are trained on trillions of tokens, a million-fold difference compared to the ~200 million tokens a human sees from birth to adulthood.
- Humans are far more sample efficient than current AI models, with differences of 3-4 orders of magnitude in tasks like learning to drive.
- Arguments that evolution provides pre-training for humans are countered by the size of AI models and the need for lifetime learning in neural connections.
- Even with pre-training, AI requires extensive data for marginal capabilities, unlike humans who learn new skills more efficiently.
- The immense sensory data humans process may not be the primary driver of general intelligence.
- Scaling laws suggest increased parameters could improve sample efficiency, but the gap between human and AI efficiency remains vast.
- Sample efficiency might not be necessary for automating white-collar work or AI research, as AI's continuous operation and continuous training offset inefficiencies.
- The future of AI may involve automated AI researchers solving the sample efficiency problem, leading to human-like intelligence.