

It’s not just the very ending, unless you mean the entire 2nd half… it just felt like there were two entirely different movies made, and they took the first half of one and haphazardly grafted it on to the second half of the other. The tonal whiplash alone was crazy.
I feel like that sort of movie mismatched has happened a lot over the years. I figure that producer meddling or overfitting to test-screening reactions is typically the cause.








It’s fundamentally not the same thing as autocomplete. Give autocomplete all the data an LLM has, every gig, every terabyte if it, and it still won’t be an LLM. Autocomplete lacks the semantic meaning layer as well as some other parts. People say it’s nothing but autocomplete from a misunderstanding of what a reward function does in backpropagation training (saying “the reward function is to predict the next word” is not even close to the equivalent of “it’s doing the same thing as autocomplete”)
I’m writing this short reply with hopes that when I have more time in the next two days or so I’ll come back with a more complete explanation, (including why context windows have to be limited).