back

by claytonwramsey·3y ago·view on hn ↗
Here's the actual paper that was summarized in the OP:

https://arxiv.org/pdf/2307.01850.pdf

The paper is quite approachable and worth a quick scan. In essence, reusing output from previous generations results in rather wonky results which are clearly separable from the original training data.

1 comments
You probably have to clarify that it only applies to the direct transformer training with fixed weights for samples, and not RL variants or variable sample weights.