Post #3037885

In short, as LLMs proliferate they will ‘contaminate’ the training pool, and make it harder and harder to build future models. These challenges may not be insurmountable - one can imagine an army of human reviewers sifting out the “good” code from the bad, or an elaborate series of sandboxes and unit tests to try to improve automatic detection of bugs - but they will certainly prove difficult to overcome. – Milo Trujillo, Model Collapse https://backdrifting.net/post/082_model_collapse #AI #NoAI #LLM