RE: https://mastodon.gamedev.place/@rovarma/116245016326565477
Epic debugging story!
RE: https://mastodon.gamedev.place/@floooh/115900619627455204
Template metaprogramming is the bane of C++. Why did the author of that code go that route?
C++ compilers will be probably the first to introduce LLM-based error output.
I ported my fork of #z80ai to #ZXSpectrum. Now it indeed ran at 3.5MHz (CP/M version on #ZXSpectrumNext must have been using 28Mhz). This simple convo took takes 4.5 minutes :)
(Optimizations are surely possible. I also pessimized it a bit by adding border colors just to not be bored waiting for a reply)
Grab the source and .tap file here: https://github.com/RCL/z80ai/tree/main/examples/tinychat/build_tap
Ok, so here is a working 40KB language model for Z80. It mght be a good occasion to put the tech under the microscope and understand it through and through.

Z80-μLM is a 2-bit quantized language model small enough to run on an 8-bit Z80 processor. Train conversational models in Python, export them as CP/M .COM binaries, and chat with your vintage compu...