Zafir Stojanovski, in a long essay posted on X, makes an argument that cuts against how most people still picture AI training: humans are no longer the ones deciding what counts as a good answer. He traces five stages of the modern pipeline, judging, corpus building, teaching, curriculum design, and reinforcement environments, and claims each one has already shifted from human labor to model labor.
His case starts with evaluation. Stojanovski points to LLM-as-a-judge research showing GPT-4 matched roughly 80 percent of human rater agreement, about the same rate two human raters agree with each other. He treats that convergence as the moment grading stopped requiring a person, and he cites Chatbot Arena, a project that grew out of exactly this idea, as having reached a $100 million annualized run rate within eight months of its commercial launch. The number is his, not independently verified here, and it illustrates his broader claim rather than proving a separate fact.
The corpus argument follows a similar shape. Training sets built from raw web scrapes are, in his telling, too noisy to use directly, so labs increasingly rewrite the data rather than merely filter it. He cites Nemotron-CC’s mix of roughly 4.4 trillion real tokens against 1.9 trillion synthetically rewritten ones as evidence that rewriting, not just curation, has become standard at frontier scale. Stojanovski frames Ilya Sutskever’s “fossil fuel” framing of internet text (Sutskever has described web data as a finite resource) as the reason labs feel pressure to manufacture what they can no longer just collect.
On teaching, he walks through a lineage from Alpaca and Vicuna, which distilled instruction-following behavior from closed models for a few hundred dollars, to DeepSeek’s distillation of its R1 reasoning model into smaller checkpoints that he says have been downloaded more than 35 million times as of this month. His point is not that distillation is new. It is that the technique has moved from a research shortcut to the default way smaller models inherit capability from larger ones, with humans mostly absent from the loop once a strong teacher model exists.
Stojanovski’s most pointed claim involves the recent OpenAI-linked Navier-Stokes solution. He surfaces a comment from Cohere’s Aidan Gomez suggesting synthetic data derived from real user conversations, particularly technical ones, gets weighted more heavily in training precisely because those exchanges teach a model the most. If accurate, it would mean policies promising not to train on user data leave room for training on rewritten derivatives of that data. Stojanovski does not claim to confirm this himself; he is relaying and endorsing someone else’s speculation, which is a meaningfully weaker form of evidence than a lab disclosure.
What his essay adds beyond the individual studies is a naming exercise: he proposes that recursive self-improvement, the industry’s term for models bootstrapping their own progress, should really be understood as recursive synthetic improvement, since the loop runs through data pipelines rather than direct code or hardware self-editing. That reframing matters for how operators should read lab claims. When a company says its next model trained on higher-quality data, the honest follow-up question is no longer “whose data” but “which model decided it was high-quality,” since the judge, the curator, and the teacher are now frequently the same lineage of systems grading their own successors.
For teams evaluating vendor claims about data quality or model provenance, Stojanovski’s framing suggests asking a narrower question directly: which prior model generated or graded the training data, and what independent check, if any, exists outside that model’s own judgment.
Based on a September 9, 2026 essay posted to X by Zafir Stojanovski (@zafstojano).
