🚨🚨 paper out of Oxford and Cambridge, published in Nature, put a name on it: model collapse.
The setup is simple. Train a model. Let it write. Train the next model on that writing. Repeat.
Each generation the data gets narrower. The tails fall off first — rare facts, strange phrasing, the thing only a person would put down. What is left is the average. The next model learns the average. After enough loops you do not get a smarter system. You get a fluent echo.
This is not one chatbot failing. The authors show it across generative setups. If the web fills up with machine text and that text becomes the training set, the next model is not learning the world. It is learning itself.
Labs already know this. That is why they still pay for human data, licenses, and closed forums. GPUs are not the scarce input anymore. Clean original text is.
So the AI trade has two clocks. One is credit — the buildings and the 20-year leases. The other is the data diet. Capex cannot print tails.
Watch which clock hits first.
Not financial advice.
The setup is simple. Train a model. Let it write. Train the next model on that writing. Repeat.
Each generation the data gets narrower. The tails fall off first — rare facts, strange phrasing, the thing only a person would put down. What is left is the average. The next model learns the average. After enough loops you do not get a smarter system. You get a fluent echo.
This is not one chatbot failing. The authors show it across generative setups. If the web fills up with machine text and that text becomes the training set, the next model is not learning the world. It is learning itself.
Labs already know this. That is why they still pay for human data, licenses, and closed forums. GPUs are not the scarce input anymore. Clean original text is.
So the AI trade has two clocks. One is credit — the buildings and the 20-year leases. The other is the data diet. Capex cannot print tails.
Watch which clock hits first.
Not financial advice.