artifabrian

That's absolutely not true.

You might get replicas of incredibly popular works like Mona Lisa due to overfitting but that's it.

If I am wrong please do provide an example as this is very relevant and interesting (and impossible from an information theory pov).

I agree with this and reflects my personal experience.

While potencial of a large audience is great, if you don't even have users to start with, the choice should be the response to "what gets PMF faster" and not "what do I need to take me from 500k ARR to +2M ARR".

That tweet doesn't reflect my experience at all actually, in Europe at least. I know it's an anecdote but I know many fit people that drink frequently Diet Coke or Coke Zero. Is this a US thing?

Could be as well, given there's many varieties of tokenisers, each with different pros and cons.

This particular tokenizer is very interesting given that it tries to be best of both worlds (word-level tokenizer and character-level).

It might be too high-level for some of you :) but if you ever wondered how they work, or wondered why these models are unlikely to replicate any specific image but are able to learn styles, or why there are "steps", or why there's a seed - then hopefully this post will help.

I'd suggest it really depends on the exact problem space you are working on.

Some NLP related problems can be solved at a similar level of quality of humans - (humans actually make a lot mistakes as well).