Yeah, this lines up with what I have heard, too. There is always talk of new models, but even the stuff in the pipeline not yet released isn’t that differentiable from the existing stuff.
The best explanation of strawberry is that it isn’t any particular thing, it’s rather a marketing and project framing, both internal and external, that amounts to… cost optimizations, and hype driving. Shift the goal posts, tell two stories: one is if we just get affordable enough, genAI in a loop really can do everything (probably much more modest, when genAI gets cheap enough by several means, it’ll have several more modest and generally useful use cases, also won’t have to be so legally grey). The other is that we’re already there and one day you’ll wake up and your brain won’t be good enough to matter anymore, or something.
Again, this is apparently the future of software releases. :/
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.
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.
there’s a research result that the precise tokeniser makes bugger all difference, it’s almost entirely the data you put in
because LLMs are lossy compression for text
latent space go brrrr
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.
OpenAI somehow managed to outdo Apple in vacuous increment based hype
