For years, one of the easiest ways to dismiss AI has been:
“It just predicts the next token.”
Technically, that may describe part of the machinery. But it is becoming a terrible description of the capability.
OpenAI has now released a broad batch of mathematical results produced by an internal frontier model. The company says that model has resolved more than 100 long-standing open problems across much of mathematics, and it is publishing results, formalizations, and research details for scrutiny.
If the proofs hold up, something important has changed. You can call the system a next-token predictor, but that no longer settles the matter.
A telescope is “just curved glass.” A jet engine is “just metal pushing hot gas.” Neither description tells you what the machine can accomplish.
People hear “predicts the next token” and smuggle in an extra conclusion: therefore it cannot reason, discover, or produce genuinely new mathematics.
That conclusion now has to survive contact with the evidence presented today.
There is still plenty to verify. Mathematics is unusually unforgiving, and extraordinary claims should survive independent checking.
But if these results stand, the argument changes.
There is a larger lesson here, and it has nothing specifically to do with AI.
Never become so attached to your explanation of yesterday that you cannot recognize what changed today.
The mechanism may be the same. Your description may still be technically correct. But new results can change what that mechanism means in practice.
When the terrain changes, the map has to change with it.
RW
.
.
.