What was said
The industry consensus of 2023–2024 had two layers. Layer one: frontier models are a compute arms race — only giants with hundred-thousand-GPU clusters get a seat at the table, and export controls would lock China two generations behind. Layer two: serious researchers (LeCun loudest) argued autoregressive LLMs themselves were a dead end — "an off-ramp" on the road to general intelligence — and open source could certainly never catch the closed frontier. Both judgments came well argued: scaling curves, chip lists, architectural reasoning, nothing missing.
What actually happened
In January 2025, DeepSeek punctured the first layer with a technical report and a training bill said to be far cheaper: controls had not locked down efficiency; constraint forced engineering innovation instead, and US AI stocks had a famous bad day over it. The second layer didn't follow its script either: LLMs didn't die — they grew reasoning and agent forms and became the industry's foundation. Open source did not kill closed source, and closed source did not lock out open source; the two chase each other along one curve, neither free to declare the other finished. LeCun's prophesied "world model" route has yet to produce a counterpart artifact.
Why it didn't come true
A moment's cost curve was mistaken for a law of physics. "It must cost a billion dollars" described the practice of 2023, not the price of intelligence; algorithmic efficiency deflates every year, and controls changed the direction of innovation (from piling on material to squeezing efficiency), not innovation itself. And the "architectural dead end" judgment mistook the ceiling of the current form for the ceiling of the road — set aside whether LLMs have hit theirs, once an LLM is embedded as a component in a larger system (retrieval, tools, verification), the system's ceiling and the model's ceiling are not the same thing at all. Once again, the system changed.
The unexpected part
The ironic distribution of outcomes: the Chinese lab prophesied to be locked out delivered the field's most influential efficiency innovation, while the hoped-for "next-generation architectures" went quiet. And the compute doctrine's real legacy was persuading the industry, in 2024, to sign the largest infrastructure build-out in human history — even if its premise turns out half right, the money is already poured into concrete. Sometimes the way a prediction changes reality matters more than the prediction.