Keep pulling the thread on Dario Amodei.
This hypothesis, formulated in 2017, posits that AI progress is primarily driven by seven factors: raw compute, quantity of data, quality/distribution of data, training duration, scalable objective functions (pre-training or RL), and numerical stability/normalization. Amodei states he still holds this hypothesis.
Similar to pre-training, scaling laws are now observed for Reinforcement Learning (RL), where model performance on tasks like math contests (e.g., AIME) is log-linear in training time. This indicates continued gains from scaling RL.
AI models require trillions of tokens for pre-training, significantly more data than humans encounter. This suggests a fundamental difference in learning processes, with AI models starting as more 'blank slates' compared to human brains with evolutionary priors.
Anthropic has experienced 'bizarre' 10x per year revenue growth: $0-$100M in 2023, $100M-$1B in 2024, and $1B-$9-10B in 2025. This rapid commercial adoption demonstrates strong market demand for frontier AI models.
Benchmarks for AI models' ability to use computers (e.g., OS world) have climbed significantly, from around 15% a year and a quarter ago to 65-70% currently. This indicates rapid progress in AI's ability to interact with digital interfaces.
Coding models currently provide a 15-20% total factor speed-up for engineers, up from about 5% six months prior. This suggests a rapidly accelerating impact on developer productivity, despite some studies showing initial 'downlift' in specific contexts.
The AI industry's compute capacity is projected to grow roughly 3x per year: 10-15 gigawatts (GW) this year, 30-40 GW next year, 100 GW in 2028, and 300 GW in 2029. Each gigawatt costs approximately $10-15 billion per year.
Amodei describes a stylized model where profitability occurs if roughly 50% of compute is for training and 50% for inference, with inference having gross margins over 50%, and demand predictions are accurate. Profitability is obscured by the exponential scale-up phase and demand prediction errors.
Most clinical trials fail due to a lack of efficacy, highlighting a bottleneck in bringing new medical discoveries to market, even with advanced AI capabilities.
AI models demonstrate real learning within their context window, with a million tokens being equivalent to days or weeks of human learning. This short-term learning ability, combined with broad pre-training, is expected to drive significant capabilities.