Calling this a straw man would be attributing an unnecessarily sturdy construction.

Yes there are people making stupid claims on all sides. Attributing phrasing like solved or cooked is as if it is some sort of fanatic specific jargon simply ignores the terminology of different groups of people. I don't use cooked myself, but I am not so ignorant of the younger generation that is see it is just another in the long line of terms like sick, bogus, hip, radical, macaroni, etc.

The author plays the trick of flipping the situation from stochastic parrots or next token prediction. Those are "taken as pejorative" whereas cooked and solved "to signal"

The flip is done to place the fault on the other party. You could equally uncharitably say that invoking next token prediction or stochastic parrots is signalling, whereas AI skeptics take terms like cooked as pejorative.

Specifically on the topic of next token prediction, we are already past that phase. Even then I don't think that a model trained by prediction has the limitations that people think it does. As a thought terminating cliche it is simply obsolete when models are trained on reinforcement techniques where there is no template next word to predict. Diffusion models don't even have an autoregressive nature.

I am not terribly fond of the claims made by people at the extremes of any particular to issue. We can perhaps try debating the facts of the matter rather than assuming the internal thoughts of people who might disagree with you.

I generally do not attribute malice to people who describe models as next word predictors. Most are simply uninformed and if you query what their understanding is of a model then you see that what they are imagining is a Markov chain. Investigating if their imagined model could correctly use "an" before "alligator" but obviously not choosing an animal beginning with a vowel just because it had just said "an" often leads them to think that there's more going on than just the next word.

> Specifically on the topic of next token prediction, we are already past that phase.

We really aren't past that phase at all. Reasoning models are just next-token prediction trained in a way where it thinks out loud, essentially. (Source: books on how LLMs actually work, and asking ChatGPT directly!) Harnesses and tool use help a little bit, but it doesn't change fundamentally what an LLM is.