Friend of mine started working on a 8-bit pc from transistors. One day he asked me, could you make AI work on that? That got me thinking and i spent last few weeks designing an architecture to fit on it, because all the values can be only 0-255, no decimal points or negative values. The best one I made currently is a digit recognition model using the MNIST dataset. It has like 34.12% correctnes on testing dataset and it is a "MLP" with custom activation functions, normalization and method of applying bias. I will make the code and final product public once we are done, it is a mess now, just wanted to share the progress. I may also make a mobile app, where you could interract with it and watch the PC work with all the led controls and stuff. Would anyone be interested in trying that out?
#AI on hand-made PC.
1 messages · Page 1 of 1 (latest)
Also I got none budget so it will take a long time until I come up with something good, that I would actualy wana try to train.
note that 1 bit quantization is very common
Recent research, such as BitNet, is paving the way for a new era of 1-bit Large Language Models (LLMs). In this work, we introduce a 1-bit LLM variant, namely BitNet b1.58, in which every single parameter (or weight) of the LLM is ternary {-1, 0, 1}. It matches the full-precision (i.e., FP16 or BF16) Transformer LLM with the same model size and ...
*quite common
Oh, thanks, didn't know about that
Tbh i didn't do any research bout it, cause i waned to make it my way, but i may give this a try
but under these constraints i guess you would care about model size
depending on how memory is supposed to work
Yea i don't have much memory either
That was another constraint
Also I have to implement it using .hex with custom instructions, unless I decide to make my own assembler lol
So the easier i make my algortithms, the easier it is to test thing
you could indeed make hardware specifically designed for running an LLM that is very good at it. perhaps some type of 8 or 16 bit floating point vector processor. would be interesting to see what it would end up like... but a general 8 bit CPU might struggle to do anything fast. how much ram does it have?
Don't remember rn
Maybe 32k
I managed to make it work but, now I am working on a way to make it batter
you should definitely make your own assembler
ikr