Benchmarks are press release components, not capability evidence

Every major model release comes with a number. The number goes on the announcement page, gets screenshotted by AI Twitter, and becomes the thing journalists compare to the previous number. This is fine if you know what benchmarks are. Most people in the coverage loop do not.

A benchmark is a fixed task distribution, designed at a specific moment, measuring a specific proxy for something we actually care about. Labs choose which benchmarks to report. They do not choose the ones where the model looks bad. The press does not ask which benchmarks were run and not shown. The result is a coverage ecosystem where 'beats GPT-4 on MMLU' functions as a capability claim, even though MMLU was designed before these models existed and tells you approximately nothing about what the model will do in deployment. The number trends. The coverage treats the trend as meaning something.

The actual capability question is harder. It requires running the model. Reading the full eval suite, not the table in the blog post. Understanding what the task distribution does not cover, which is usually the part you should worry about. Almost nobody with a publication deadline does this. So what compounds quietly, without a chart, is the gap between what the benchmark measures and what the model is actually doing. That gap has been widening for two years. Nobody is benchmarking the gap.