[tadfisher]

AI is going to implode within 2 years. Once it starts ingesting its own output as training data it is going to be at best capped at its current capability and at worst even more hallucinatory and worthless.

[bigfudge]

The mistake you make here is to forget that the training data of the original models was also _full_ or errors and biases — and yet they still produced coherent and useful output. LLM training seems to be incredibly resilient to noise in the training set.

[joquarky]

Forget what it eats to continue improving.

Realize what it already has.

A foundational language model with no additional training is already quite powerful.

And that genie isn't going back into the bottle.

[mafuy]

Nonsense. Some of the current best AI models were specifically trained on AI output.