ML//AI progress

AI progress is the growth over time in what machine-learning systems can do, and its study breaks that growth into the inputs that produce it (compute, algorithms, data and money) so that each can be measured and forecast. The breakdown matters because the inputs run on different clocks: algorithms can improve in months, while chips, power plants and investment take years.

AI progress is the growth over time in what machine-learning systems can do, and its study breaks that growth into the inputs that produce it (compute, algorithms, data and money) so that each can be measured and forecast. The breakdown matters because the inputs run on different clocks: algorithms can improve in months, while chips, power plants and investment take years.

The common unit is effective compute, physical compute multiplied by the gains from algorithmic progress.

The question of a rapid takeoff is whether AI can drive its own progress (recursive self-improvement) and whether that leads to an intelligence explosion.

What limits it is mostly physical: experiment compute (compute bottleneck), the duration of each round of improvement (generation time), chip advanced packaging and HBM, and power.

Capability itself is measured with benchmarks and with the time horizon of agents.