Around 2020, as OpenAI and large models entered public view, a parade of very smart people lined up to pronounce judgment on AI's future. They sounded thoroughly convincing at the time — data, arguments, timelines, all present. A few years on, almost none of it came true.
This node is not here to mock them. Being wrong is no disgrace; what deserves recording is that the ways of being wrong have structure. Each article settles one prediction: what was said, what the argument was, what actually happened, why it didn't come true, and the thing that did happen that nobody expected.
The general ledger, up front:
- Timelines were wrong: the optimists and the pessimists both missed; the friction of the physical world and of institutions does not obey schedules.
- "Total replacement" was wrong: AI almost never carries off a whole occupation. It enters as a tool and rewrites the job from inside.
- The math was right, the system changed: death sentences computed from static assumptions (compounding errors, running out of data) were routed around by architectural innovation.
- Camp narratives were wrong: open vs. closed, compute vs. algorithms, US vs. China — every binary script fell through.
- The spotlight pointed at the wrong place: the things actually disrupted (homework answers, Q&A site traffic) were nobody's headline at the time.
One observation runs through all of it: in use, the technology really is strong — and the world really hasn't changed that much. These two facts do not contradict each other. What is strong is the human–machine loop, not the model alone; the predictions went wrong by subtracting the human in the loop. When capability gets cheaper, the first thing that happens is not layoffs — it is the same person opening more lines of work. Replacement moves incredibly slowly, because replacement needs near-100% reliability while assistance only needs 70% to be useful, and diffusion always takes the assistance road first.
This node will grow slowly. Entries are ordered by writing, not by wrongness.