@silverpill@mitra.social here's another tech that's been on my radar recently: https://www.cartesia.ai/blog/from-tokens-to-bytes-part-1 just looked up the blog post, I haven't even read it yet, but the gist that I've been picking up is the h-net is able to create it's own chunked representations instead of relying on a fixed tokenizer. Could make this "weight-space programming" stuff a lot more interesting