Taking 20 years to move from 0.4 to 0.4.1 seems like a long long time.
Well, I'm surprised he has the time/intention to do that in between for us mere mortals. Here are a few references;
- https://en.wikipedia.org/wiki/Bellard%27s_formula
- https://en.wikipedia.org/wiki/Fabrice_Bellard
``` In April 2021 his Artificial Neural Network based data compressor, NNCP, took first place out of hundreds in the Large Text Compression Benchmark. To write the compressor, Bellard wrote his own Artificial Neural Network library, LibNC ("C Library for Tensor Manipulation"), publicly available. ```
Since I teach a fair amount of machine learning, I have been pondering on what a “minimal”, low-level implementation of a library would look like and got very excited hearing about LibNC. Sadly it seems like Bellard is keeping the source for this one: “The LibNC library is free to use as a binary shared library. Contact the author if access to its source code is required.” [1]. Not that you can every accuse the man of being greedy after all the code he has shared, but dang I wanted to read it. '^^
TinyGrad is basically that and is open source, albeit not in C
Yes, kind of, but it depends on NumPy which is fairly big if your goal is for students to grasp and demystify the whole way down to the hardware. Other candidates would be torch7 [1], but it is somewhat married to the Lua API which drives it.