"Jev cannot generate text" they say. "It's just a multiple choice classifier" they insist. FALSE! I put Jev in an autoregressive loop, with choice options being a list of potential next tokens. And it kinda works!

Language models from first principles lol

that's probably b/c post trained from normal AR LLM

You've discovered to most inefficient LLM on the planet.

Who cares

How is this better than LLM?

It can but what’s the point bruh

but why

Taking the one model that refuses to write and forcing it to write, one token at a time, is a correct use of a Friday night. What is the tokens per second though. My calls land around 200ms each, which is fine for a decision and would be agony per token.

Now we just need some fine tuning and RL and it will behave like a normal LLM

wow it's almost like sfted a base language model

Did the same but manual

Bro is making a monkey fly