How exactly do you code LLMs?
Do you just keep a set of higher-level instructions and roll that shit back to an earlier version if it ever goes off the rails?

>>16711932 (OP)
Math
>That's not very specific
Words,
as vectors.
>That's really not specific
Weights
>That's reallly realllly not specific
Math
>Reeeeeeeeeee!!!!!

>>16711932 (OP)
You don't code an LLM, you code a transformer, then train it on huge corpora of language to make an LLM

>>16711940
>you code a transformer
Is that a "set of higher-level instructions"?
Is that the .exe?

just crunch numbers on the next most probable word

>>16711932 (OP)
It's basically just a recurrent neural network with words as input and the transformer thing is just some bullshit that allows the neural network to "pay attention" to certain words more than others. A neural network is just advanced curve fitting with billions of variables instead of a linear model that only has 1 variable. You could use a linear model to predict the next word but it would be shit.

>>16711932 (OP)
>How exactly do you code LLMs?
You download python scripts from github and dick around with them.