Your AI might be training on itself — and we measured the two ways that ends badlyVasa AI mozno trenuje sama na sebe — odmerali sme dva sposoby, ako sa to zle skonci
The claim. Any system that learns from its own output is a strange loop — a model retrained on synthetic data, an agent whose memory is its own past answers, a RAG store indexing the system's prior geKolaps modelu, odmerany. Kazdy system, ktory sa uci z vlastneho vystupu, je podivna slucka. Postavili sme najmensi spustitelny model a nasli sme dva sposoby zlyhania — a dve paky, ktore im branilia: ~5% kotva realnych dat zastavi kolaps diverzity, a udrzanie exponentu sebadovery p<=1 zastavi trvale
The claim. Any system that learns from its own output is a strange loop — a model retrained on synthetic data, an agent whose memory is its own past answers, a RAG store indexing the system's prior generations, a recommender fed by the clicks it created. In 2026 this stopped being a research curiosity: model collapse is now a documented production concern. So we built the smallest honest model of it and measured the two ways a self-referential system fails — and the two knobs that prevent each.
Failure 1 — collapse (the data-mix law). Retrain a model recursively on its own outputs and its diversity drains away (the "curse of recursion"). In our minimal simulation, with no real data ~92-94% of runs collapsed to near-zero diversity. The cure is a floor of real/external data, and the floor is low: a ~5% real-data anchor pulls the collapse rate from ~94% to under ~10%, and 20% makes it clean. The lesson isn't "never use synthetic data" — it's "never replace real data with it." Accumulate, don't substitute.
Failure 2 — lock (the self-trust law). If a system weights its own prior belief faster than fresh evidence can correct it (a self-trust exponent p > 1), a fixed fraction of any initial bias is never washed out, no matter how much data arrives. This one has a closed form: p <= 1 is the safe boundary; at p = 1.5 you permanently lock 17.7% of any bias, at p = 2 you lock 50%, at p = 3 you lock 81%. The tell: inject a known bias, then keep feeding unbiased data — if the bias doesn't decay as the data grows, you're locked.
The honest caveat. These are minimal models, not your training run — but they're runnable, the thresholds are reproducible, and the peer-reviewed literature agrees on the cure: mixing real with synthetic provably bounds the error, while replacing real with synthetic grows it without bound. The point is to turn "model collapse" from a vibe into a number you can put a threshold on.
We packaged the check as selfref: one zero-dependency file, plus an MCP server so an agent can ask — before it retrains on itself — am I about to collapse or lock? It's open-core and free, a sibling of our memory, RAG-freshness and statistics tools. One call, audit(external_fraction, self_trust_p), gives you a verdict and the fix.
Tvrdenie. Kazdy system, ktory sa uci z vlastneho vystupu, je podivna slucka — model pretrenovany na syntetickych datach, agent, ktoreho pamat su jeho vlastne minule odpovede, RAG sklad indexujuci predchadzajuce generacie systemu, odporucaci system kpneny klikmi, ktore sam vytvoril. V roku 2026 to prestalo byt vyskumnou kuriozitou: kolaps modelu je teraz zdokumentovany produkcny problem. Tak sme postavili jeho najmensi poctivy model a odmerali dva sposoby, ako sebareferencny system zlyha — a dve paky, ktore kazdemu branilia.
Zlyhanie 1 — kolaps (zakon zmesi dat). Pretrenuj model rekurzivne na jeho vlastnych vystupoch a jeho diverzita sa vytrati ("kliatba rekurzie"). V nasej minimalnej simulacii sa bez realnych dat ~92-94% behov zrutilo na takmer nulovu diverzitu. Liekom je minimalny podiel realnych/externych dat, a ten podiel je nizky: ~5% kotva realnych dat stiahne mieru kolapsu z ~94% pod ~10%, a 20% to vycisti uplne. Pointa nie je "nikdy nepouzivaj synteticke data" — ale "nikdy nimi nenahradzaj realne data." Pridavaj, nenahradzaj.
Zlyhanie 2 — zamknutie (zakon sebadovery). Ak system vazi svoje vlastne predchadzajuce presvedcenie rychlejsie, nez ho cerstve dokazy stihaju opravit (exponent sebadovery p > 1), pevny zlomok lubovolneho pociatocneho skreslenia sa nikdy nezmyje, bez ohladu na to, kolko dat pride. Toto ma uzavretu formu: p <= 1 je bezpecna hranica; pri p = 1.5 trvalo zamknes 17,7% skreslenia, pri p = 2 zamknes 50%, pri p = 3 az 81%. Znak: vloz zname skreslenie, potom stale prikrmuj nestrannymi datami — ak skreslenie neklesa, ako dat pribuda, si zamknuty.
Poctiva vyhrada. Su to minimalne modely, nie tvoj realny trening — ale su spustitelne, prahy su reprodukovatelne a recenzovana literatura sa zhoduje na lieku: miesanie realnych so syntetickymi datami dokazatelne ohranicuje chybu, kym nahradenie realnych syntetickymi ju nafukuje bez hranic. Cielom je premenit "kolaps modelu" z pocitu na cislo, na ktore vies nastavit prah.
Kontrolu sme zabalili ako selfref: jeden subor bez zavislosti, plus MCP server, aby sa agent mohol — predtym, nez sa pretrenuje sam na sebe — opytat: idem skolabovat alebo sa zamknut? Je open-core a zadarmo, surodenec nasich nastrojov na pamat, RAG-cerstvost a statistiku. Jedno volanie, audit(external_fraction, self_trust_p), ti da verdikt a opravu.