Jason Wei against the "cognitive core" thesis

What's this about?

Jason Wei, one of the authors behind chain-of-thought prompting, pushed back on an idea he says he was once sympathetic to: that instead of scaling models up, all you need is a sufficiently strong "cognitive core" — perhaps a billion parameters — with everything else handled by tools like web search and code execution. He now thinks tools can't carry top-tier intelligence, because internalized knowledge and looked-up knowledge are genuinely different things. His analogy is learning badminton: he can execute every stroke his coach demonstrates, but stringing them together fluidly in a real match is a different level from someone with ten thousand repetitions of muscle memory. The concrete costs he names are speed — answering directly beats a tool round-trip — and depth of understanding.

What exactly is a "cognitive core"?

The idea is to separate reasoning from knowing. A small model would handle the reasoning — planning, decomposing problems, deciding what to do — while all factual knowledge comes from searching and all computation from running code. The appeal is economic and it's substantial: small models are dramatically cheaper to serve, and the argument sounds airtight because for any specific fact the model lacks, you can look it up. Wei's counter is that fluency isn't decomposable this way. Knowing something deeply shapes how you approach a problem before you'd ever know to search for it.

Why is this worth noticing?

1. It's a public reversal from someone with standing, which is rarer and more informative than another position paper.
2. It cuts against the week's other evidence, which makes it interesting rather than settled. StateM got frontier results by improving the scaffolding around a cheap model; Wei argues scaffolding has a ceiling. Both can be true — harness scaling may lift a small model a long way without closing the last gap.
3. The economic stakes are why the debate matters. If small-plus-tools worked, inference costs would collapse; Wei is arguing the industry shouldn't count on that.