I'm curious who the target audience of this tweet is. Are there a lot of people who know all about tensors, deep learning, and PyTorch but have no idea what C is?
Very interesting notes and as commented by one of the sibling comments it is really a rumination on abstraction rather on than implementation off llm.c itself.<p>It will be very interesting to implement this small llm in D language and use the new tiny autograd library in D [1]. Since now D compiler does support C, this should results in less line of codes excluding the small autograd library in D [2]. Given that the D code will be able to inline C code, and even CTFE it, if done correctly the implementation will be faster as well.<p>[1] A tiny autograd library:<p><a href="https://code.dlang.org/packages/tiny-autodiff" rel="nofollow">https://code.dlang.org/packages/tiny-autodiff</a><p>[2] Adding ANSI C11 C compiler to D so it can import and compile C files directly:<p><a href="https://news.ycombinator.com/item?id=27102584">https://news.ycombinator.com/item?id=27102584</a>
Actual link: <a href="https://twitter.com/karpathy/status/1778153659106533806" rel="nofollow">https://twitter.com/karpathy/status/1778153659106533806</a><p>It's not a thread, no nitter needed.
Related:<p><i>Llm.c – LLM training in simple, pure C/CUDA</i><p><a href="https://news.ycombinator.com/item?id=39973467">https://news.ycombinator.com/item?id=39973467</a>