GPT2-BASIC uses constraints from that period as an engineering target rather than a nostalgia premise. The relevant question is not whether a 486-class DOS machine can imitate a modern hosted model. It cannot. The useful question is which pieces of language-model inference, local recall, and assistant behavior can be made small, explicit, and reproducible enough to run there. The historical comparison and design implications section covers that context.

Ok, this looks pretty cool. The token-per-second looks ok, I guess?, for those emulated early 1990s configurations with a max of 32MB RAM. Still, gotta wonder how much of the code actually works, as the entire readme is llm generated, so it wouldn’t be far off to guess the code also had llm meddling.