4 ms·
I think one of the things overlooked in the discussions here is that the research is solely around the reinforcement against edge cases, but does not qualitativ
by kromem 3y ago
I think one of the things overlooked in the discussions here is that the research is solely around the reinforcement against edge cases, but does not qualitatively assess these edge cases.
To me, this research supports a hypothesis I've had for a while that we're going to get to truly excellent AI by using synthetic data to bias it towards excellence and away from mediocrity.
$20 says the next round of major model training is using synthetic data for training that was filtered through a discriminator trained entirely on human data.
The human data as a reference is certainly important to avoid polluting (and to its point there's an advantage for those already having it), but moving away from edge cases isn't necessarily a bad thing practically given edge cases can result in negative practical performance (as opposed to academic next token performance).
- XorNot 3y agoI think you're on the right track with this thought: the obvious use case for models like this is their ability to classify data based on their training. Like almost everyone has immediately thought "AI moderator" as a use case - but the most obvious use is for the AI to moderate it's own training data for the next version. Once they can do that and produce a productively improved model, then that's really the start of self-improvement.
- sebzim4500 3y agoOpenAI (accidentally?) confirmed in a recent paper that they used synthetic data in the training set for GPT-4, so to some extent this has already happened. It's not clear whether they did any human filtering on that data though.