Basically there isn’t significant improvement to be had in the tokeniser, because it’s already been trained on all the data on earth. So all they have left is overengineering.
Calling it now: codepoint-level non-tokenizing, with a remapping step to only recognize the most popular thousands of codepoints, would outperform what OpenAI has forced themselves into using. Evidence is circumstantial but strong, e.g. how arithmetic isn’t learned right because BPE tokenizers obscure Arabic digits. They can’t backpedal on this without breaking some of their API and re-pretraining a model, and they make a big deal about how expensive GPT pretraining is, so they’re stuck in their local minimum.
But then it can’t SolidGoldMagicarp SolidGoldMagicarp SolidGoldMagicarp SolidGoldMagicarp
Does this mean they’re not going to bother training a whole new model again? I was looking forward to seeing AI Mad Cow Disease after it consumed an Internet’s worth of AI generated content.
If you change the tokenizer you have to retrain from scratch, but you can do so with the old, unpolluted data.
It’s genius if you think about it,* you can waste energy and tell your investors it’s a new better model, while staying upstream from the river you pollute.
* at least for consultants, compute providers and other middle men.
I remember one time in a research project I switched out the tokeniser to see what impact it might have on my output. Spent about a day re-running and the difference was minimal. I imagine it’s wholly the same thing.
*Disclaimer: I don’t actually imagine it is wholly the same thing.