mercury-ad VS picoGPT

I wrote a purely functional AD library in Mercury [0], which adapts a general approach from [1]. I believe that Owl provides a similar approach [2].

[0] https://github.com/mclements/mercury-ad

I find the solutions from https://github.com/qobi/AD-Rosetta-Stone/ to be very helpful, particularly for representing forward and backward mode automatic differentiation using a functional approach.

I used this code as inspiration for a functional-only (without references/pointers) in Mercury: https://github.com/mclements/mercury-ad

In that case, you might also enjoy https://jaykmody.com/blog/gpt-from-scratch/

(here's the raw code: https://github.com/jaymody/picoGPT/blob/main/gpt2.py)

I wrote a minimal implementation in NumPy here (the forward pass code is only 40 lines): https://github.com/jaymody/picoGPT

Although this is for a decoder-only transformer (aka GPT) and doesnt include the encoder part.

Take a look at https://github.com/jaymody/picoGPT/blob/a750c145ba4d09d57648...

Yes, this is GPT-2 not 4 and it‘s not the Chat, only the model and it‘s basically only the inference part, not the training loop and it‘s somewhat simplified.

Still, take a good look.

That‘s essentially what it is and a single sheet of paper.

There is nothing specifically about language in „language model“, we just call it that. Better to call it just LLM.

Nobody knows exactly what it learns, although there would be ways to poke around given some research programs. But it seems like the interest in that is limited currently, everyone is busy with improving it or with applications.

Perhaps the answer is that we overestimated what a mind is. It‘s like we used to ask what life is and it turned out that there is nothing special about life, not even the DNA is controlling anything. It‘s merely a chemical process, even though a complex process.

(The hardest part was just designing a math function that has the capability of getting good at this game, but when all is said and done, it need not be a whole lot of code).