The sample efficiency black hole
"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 is largely due to wider data distributions and scaling compute, not improvements in sample efficiency.
- Reinforcement learning (RL) acts as synthetic data generation, requiring vast compute and human expert data.
- Human expert data is task-specific and requires hundreds of experts for each skill.
- Frontier AI models are trained on trillions of tokens, vastly more than humans see in a lifetime.
- Humans are significantly more sample efficient than current AI models across various tasks like driving and learning skills.
- Evolutionary pre-training is an insufficient explanation for AI's data requirements; lifetime learning in humans is more analogous to parameter training.
- While current scaling laws suggest parameter scaling can reduce data needs, the discrepancy in sample efficiency between humans and AI remains substantial.
- Despite inefficiency, AI can be trained with immense compute and its learned skills amortized over billions of sessions, making it economically viable for automating white-collar work.
- The automation of AI research itself is proposed as a future step, potentially solving the sample efficiency problem.
- The article questions the common understanding of an 'intelligence explosion' and suggests a more nuanced view of rapid AI progress starting from LLMs.