Large language models improved for years by an unglamorous route: more data and more .
Both inputs are now running into limits of a different kind.
The supply of high-quality human text is and largely already used.
text can extend the supply, but training a model mainly on its own output it.
Researchers describe the failure mode as a narrowing: rare constructions disappear first.
data therefore now carries a price, and licensing deals have followed.
This alters the competitive landscape more than any technical advance.
A laboratory without an archive, a publisher or a platform must buy access from one.
Open-weight models complicate the picture in a useful way.
They are usually behind the frontier, yet good enough for most applications and far cheaper to run.
For a hospital or a ministry, may matter more than the last percentage point of accuracy.
Evaluation remains the weakest part of the field.
leak into training sets, after which a high score proves little.
Serious deployments now rely on private test sets and on measured failure rates in real use.
That is slower, unfashionable and the only evidence worth acting on.