The standard form is “I hate AI as much as the next person, but–” followed by whatever you were going to say anyway. It seems to appease the identarians.

Maintainability became the go-to jibe this year. What a high-level requirement to complain about. Like sure, generated code can demonstrably work, and read sensibly with comments and tests, and interoperate with other programs… but in ten years you’ll be sorry! As if local models two years from now won’t make these behemoths look primitive. The bubble could pop tomorrow and it’d only increase pressure toward smaller and faster experiments by people with one-comma budgets.

A recent Computerphile video amounted to nervous laughter about chain-of-thought vanishing as a worthwhile glimpse of the model’s state. (On the shaky assumption it ever was.) At the end they mention prompts where the model gets to output one word as an immediate answer. Last year the kinds of problems they could solve “between words” was comparable to what a person could think through in thirty seconds. On current frontier models it’s more like half an hour.