I have a fair idea of the criticisms, but being playful I decided to have a bit of an open mind see if the idea has legs. I don't know shit about hardware, not much about compilation and I've got self purported decent level of understanding of the nature of LLM's.

If you click around there are 4kb binary outputs that work, the generation is using my 6-luna api credits so I might have to drop it depending if people are interested. Experiments with lesser models are a bit meh and I'm of the mind that the open weight efforts are still behind (I do think they will catch up for weird problems like this)

The next problem to work on in this scenario, is modules, give the LLM instructions to write orchestration of the outputs, given that they can only probably write 8kb of binaries without losing their fucking mind. (this could be wrong too)

Going to setup a shared community of hosted libraries and work towards an operating system, forgive the buzz words, but might use something like Jev to help with finding compatible binary artefacts.

---

The virtual cpu design is written in typescript, it has a terminal with keyboard inputs, and a framebuffer for rendering.

All generations are public at the moment, so you can browse and share, best I've got is a little mario world, use your imagination.

---

The proposed instruction set is named "ANMA-1", I'd recommend clicking the "System prompt" button in the top bar.