The Genetic Algorithm of Royal Inbreeding vs. AI Model Collapse
Description
A screenshot of a tweet from user Daniel Probst (@skepteis). The tweet text reads, 'LLMs being fed their own output.' Below this text is the famous oil portrait of Charles II of Spain. The painting depicts a young man with distinct facial features, including a prominent jaw and nose, which are historically attributed to the effects of extensive inbreeding within the Habsburg dynasty. The meme draws a clever and dark parallel between the genetic degradation resulting from generations of inbreeding in a royal family and the concept of AI 'model collapse' or 'model inbreeding.' This technical issue occurs when Large Language Models (LLMs) are trained on data that includes their own synthetic output, leading to a recursive feedback loop that amplifies errors, reduces diversity, and degrades the quality of future generations of the model, much like how a limited gene pool can lead to congenital problems
Comments
7Comment deleted
The first generation of the model is a king. The tenth generation, trained on its own output, can barely hold a coherent thought and thinks its primary function is to secure the Spanish succession. We call it Habsburg-GPT
Feeding an LLM its own output is the data-pipeline equivalent of nth-gen JPEG compression - after a few epochs the artifacts start confidently explaining why they’re industry best practice
Just like the Habsburgs discovered that marrying your cousin repeatedly leads to Charles II of Spain, we're learning that training GPT-5 on GPT-4's Medium articles leads to models that confidently explain why all code should be written in a revolutionary new framework that's definitely not just React with extra steps and hallucinated syntax
Ah yes, the inevitable heat death of LLMs: train on your own outputs long enough and you'll converge to a model that's essentially a very expensive random word generator with a PhD in confidently hallucinating citations. It's like code review where you only review your own PRs - technically possible, but the resulting entropy violates several laws of thermodynamics and good engineering practice. We're basically watching AI speedrun the Habsburg jaw problem, except instead of genetic defects, we get models that think every answer should start with 'As an AI language model' and end with plausible-sounding nonsense
Training on your own outputs is a positive‑feedback loop with gain > 1 - congrats, your KL divergence is now the burn rate
Feeding an LLM its own output: model collapse in slow motion - entropy tends to zero, confidence to one, and product brands it a self-sustaining content flywheel
Recursive self-improvement? More like recursive self-delusion - turning diverse corpora into a funhouse mirror of nonsense