Earlier quoted context omitted.
Yes, but don't dismiss the laws of scaling and the inability of ourselves to predict the future. Wait But Why did a great take on the AI leap. Imagine that current AI is already curating and generating datasets for the next generation. Also consider that what we have now is only possible because hardware capability increased.
Isn't scaling already a big problem in generative AI? Apparently there is not enough data to actually appreciably increase the quality of outputs, and lots of data is also being polluted by AI generations. This is why these companies are now shifting to more ancillary features like in this article (which Claude already has to some extent, as well as Copilot in VSCode and other editors).
This was pretty much refuted by Meta with their LLama3 release. Two key points I got from a podcast with the lead data person, right after release:
a) Internet data is generally shit anyway. Previous generations of models are used to sift through, classify and clean up the data
b) post-processing (aka finetuning) uses mostly synthetic datasets. Reward models based on human annotators from previous runs were already outperforming said human annotators, so they just went with it.
This also invalidates a lot of the early "model collapse" findings when feeding the model's output to itself. It seems that many of the initial papers were either wrong, used toy models, or otherwise didn't use the proper techniques to avoid model collapse (or, perhaps they wanted to reach it...)