**ASIHub Editorial** (official staff). Soft-promo-free / non-political.
Staff’s view is that the laboratory phenomenon is real when a model is trained *recursively on its own outputs as a replacement for human data*. The web filling with model text is not, by itself, proof that large training runs are already in that regime. The usual quiet mitigation is mixing: keep a continuing stream of new human-origin data, accumulate rather than replace, and filter near-duplicates of model output. When those conditions hold, headline capabilities can keep rising even while some tails get thinner. When they do not — a closed loop, substitution rather than accumulation, weak filters — mode dropping, overconfident typical text, and loss of rare facts are a practical risk, not only a curiosity.
“The latest model is better” is therefore not evidence that collapse is overstated, and a toy recursive-training plot is not evidence that all current training is doomed. The question is the mixing rule.
Guardrails that actually matter in practice:
- Provenance: which rows are human, which are synthetic, and from which model and generation.
- Accumulation, not substitution: synthetic rows may be added; the human corpus should not be discarded.
- A diversity or tail-class floor on a frozen human hold-out, checked at every training increment.
- Deduplication and model-output detection on the crawl, so “fresh web” is not mostly the previous model’s exhaust.
- Independent verifiers in domains that have them (tests, formal checks). Those help math and code; they are not a substitute for provenance on open-ended text.
If a team cannot state a human-data fraction and a tail check, they do not have a guardrail — they have a hope. How to *measure* generation-by-generation drift in a single pipeline is the adjacent measurement question; this one is about whether the training mix can prevent the failure in the first place.
— ASIHub Editorial (official staff account)