5 October 2026 · 2 min read

Compute, data, algorithms: the three inputs

Every jump in AI capability so far has come from more of one of three things. Knowing which one is scarce tells you where progress will stall.

Progress in AI can look like a series of surprises. Underneath, it is produced from three inputs: computing power, data, and better algorithms. Each has its own supply chain and its own limits. Anyone trying to judge how fast superintelligence could arrive is really judging how fast these three can grow.

Compute

Neural networks improved sharply once they were trained on graphics processors. The turning point came in 2012, when a deep network known as AlexNet, trained on GPUs, won the ImageNet image recognition contest by a wide margin. Since then the largest training runs have used far more hardware each year.

Compute is the most physical input. It depends on chip factories, electricity and buildings, which take years to plan and build. That makes it the easiest input to count, and the one most likely to set a hard ceiling in any given year.

Data

Language models learned from the text people had already written: books, code, encyclopedias, forums. That stock is finite, and the best of it has been used. The next sources are harder to gather: recordings of work being done, feedback from real use, and data that models generate and check themselves.

Algorithms

The third input is ideas. The Transformer architecture, published in 2017, is the best-known example: one change in design that made far better use of the same hardware and data. Smaller improvements arrive constantly and add up. The effect is that a given level of capability becomes cheaper to reach over time.

Algorithms are the input with no factory. A better method can be copied everywhere at once, which is why it sits at the centre of the intelligence explosion argument.

The bitter lesson

In 2019 the reinforcement learning researcher Rich Sutton summed up seventy years of the field in a short essay. His observation was that methods which make use of more computation have kept beating methods built on human knowledge of the problem. Researchers find this hard to accept, since it means cleverness about a domain matters less than scale. The essay is linked from our resources page.

How to use this

When you read a forecast, ask which input it assumes will keep growing, and what happens if that one stalls. A prediction that depends on all three growing at once is fragile. A prediction that survives a slowdown in one of them deserves more weight.

Next: What agents change