3 ms·
Generated data tends to be selected and edited by humans, so it is already a bit better than raw. In general a model that takes external feedback into accoun
by visarga 3y ago
Generated data tends to be selected and edited by humans, so it is already a bit better than raw. In general a model that takes external feedback into account will be able to self improve, for example a code model would run tests and a chat model could interpret user responses as reward signals.
You gotta add something new to the mix. That's possible when the AI is part of a larger system. AlphaZero demonstrated that even self play could be a source of signal, as long as it gets the feedback of the game, the model can learn.
- tsimionescu 3y agoI think that has only been proven to work so far on limited game-style problems (such as literal games but also things like protein folding). It remains to be seen whether the techniques work well for more open-ended tasks like "produce text that resembles human writing".
- visarga 3y agoHere is a paper doing evolutionary approaches on top of LLM generating code. It seems LLMs are remarkably good at learning from feedback. > Evolution through Large Models https://arxiv.org/abs/2206.08896 https://arxiv.org/abs/2206.08896