Explainers

Seventy-one years of AI traced through the primary documents that mark each era: the 1955 Dartmouth proposal, the 1972 Lighthill review, the 1980s US funding tables, the 2012 ImageNet scores and the scaling papers that corrected each other. Four lessons survive all of it.

Every word in the AI vocabulary was coined by somebody for a narrow job, then stretched by somebody else. Here is what AI, gen AI, foundation model, token and agent meant when they were first written down, and who wrote them.

The phrase covers four different products: rented frontier APIs, open weights you download, the cheap tiers inside both, and models billed per second of video. Here is what each one costs in August 2026.

Strawberry might be two tokens. The model is being asked to count letters inside symbols it cannot see inside.

The model never touches your repository. It names a tool and supplies arguments, and your code decides whether to run it.

The mechanism is simpler than the capability suggests, and unusually predictive of where these models fail.

We specify a training process, not a design. Nobody writes down what the finished network does, which means nobody starts out knowing.

A transformer does fixed computation per token, so extra tokens are the only way to spend more effort. That constraint explains the gains and the bill.

Public human text runs out between 2026 and 2032. Generated data fixes that only where a checker exists, which is a much narrower set of tasks than the pitch suggests.

OpenAI found a 1.3B tuned model was preferred over 175B GPT-3. The three stages, the alignment tax, and whose preferences get encoded.

A trillion-parameter model can cost the same to run as a small one, because only a fraction of it fires per token. How sparsity works and what it costs.

The curves have not broken. The recipe changed twice, the spending moved to inference, and the data stock has a date on it.