Synthetic data’s ceiling splits: broken in math and code, intact elsewhere — resolves 2027-07-14

The debate: training the next AI on data generated by another AI (synthetic data) has long been thought to pass the teacher model’s ability to the student without ever surpassing the teacher.

Why it matters: if that ceiling held in most domains, the scarcity of real data would cap the whole industry and push the cost of buying data structurally higher.

Who is on each side: the camp that says synthetic data always hits a ceiling, including Alexandr Wang, against SecondSource’s bet that the ceiling only holds where answers cannot be checked automatically.

SecondSource’s reading: the synthetic-data ceiling is splitting along verifiability. In math and code, where right and wrong can be checked automatically, models improve themselves from that outside signal and have already broken through the rule that a student cannot beat its teacher.

In writing and general ability, where there is no single right answer, the ceiling remains.

This call has been taken over by a sharper follow-up that can be counted at the deadline (fewer than two independent teams reproducing a student that beats its teacher in these domains). The original is kept here for comparison, unchanged.

arXiv (Lightman, Kosaraju, Burda, Edwards, Baker, Lee, Leike, Schulman, Sutskever, Cobbe; OpenAI)
“Our process-supervised model solves 78% of…”

arXiv (DeepSeek-AI, collective lab authorship)
“the reasoning abilities of LLMs can be…”

Checked in July 2027: whether domains that cannot be verified show reproducible evidence of synthetic data pushing the frontier forward. None means the call stands.

What would overturn it

The call is overturned if, in a domain that cannot be verified, synthetic data produces a student that surpasses its teacher and the result is reproduced; or if reward leakage in reinforcement learning with verifiable rewards is shown to be systematic, which would shrink the breakthrough claimed in verifiable domains.