Models That Never Stop Learning

Anshu Gupta · Schnitt Inc. · October 2026 · DOI 10.5281/zenodo.23469930

Abstract

A deployed language model is frozen: what a company learns afterwards, the results of the model's own answers included, reaches it only through the prompt. We study the loop that trains one pooled model, the assistant every person at every company using it asks for what they need, each cycle on everything those companies record, where part of the record is the model's own answers, chosen by people and graded by results the companies already record. From published measurements of trained models, taken as premises, and stated assumptions about the loop, we prove what it can hold and how far it can climb. The weights hold facts that rarely change within a cycle, faster facts belong in the prompt by rule, and no replay budget keeps more than itself times the cycles a fact survives unseen. On a fixed record, the outcome-weighted update we study improves the model once and never passes imitating past successes; its further gains come through people using its answers and the world grading them, and the share of work that follows the model sets how far it climbs, how many improvements stack and how fast errors heal. Practice proposed at a fixed share, by other companies, a stored copy of the people's practice or exploration, is the same loop with an extended practice and a taxed adoption share. Results that settle late trade the speed of the climb against credit for successes later reversed. We also give an audit that certifies, from served proposals and their outcomes, when a trained candidate beats the served model, without assuming training reaches the update. Of the alternatives, a grade written about the transcript helps only where it tracks the result, a gate against the last model lets losses accumulate, and a model per company misses what other companies' results teach. We describe the loop as we have built it at Schnitt, on a masked diffusion language model trained from scratch.

Cite

@misc{gupta2026never,
  title  = {Models That Never Stop Learning},
  author = {Gupta, Anshu},
  year   = {2026},
  publisher = {Zenodo},
  doi    = {10.5281/zenodo.23469930}
}