[iFire]

You've got a good point.

Why not directly have the llm write ISA assembly. We're still grading based on results / theory proofs and for example the certifications for cryptographical government use are based on binary codes and not sources.

Edited:

Why not go further and print the chips, pcbs directly via 3d printing with llm instructions.

Edited (joke):

Why not go the furthest and turn the entire earth into a computer and grey goo?

[fragmede]

> Why not directly have the llm write ISA assembly.

The honest ernest answer to that is it's a bad idea because it is not portable. Unfortunately for Intel, m they don't have the dominance they once did, so you have to pick between ARM, x86, or something more exotic, and then be attached to that specific ISA. It's an interesting thought tho.

[iFire]

My approach for a game sandbox https://github.com/libriscv/godot-sandbox for user generated content was to experiment with standardizing on a riscv 64 bit linux ISA.

https://bellard.org/jslinux/ bellard is notable for this approach where you write a riscv execution layer and then write a windows / linux / dos etc emulator on top.

https://bellard.org/tinyemu/

[makefunstuff]

We have different IR backends to make it portable btw. There are llvm, qbe, etc In the repo I mentioned in the link am goofing around exactly with llvm, to actually figure it out.

[makefunstuff]

I think you, got the point. I am under impression that we are going to nowhere with these approach of that token generation is a solution for everything. For me for some reason it feels like bruteforce but stakes right now are too high, so step back is not an option for bigger players.

[ksaj]

That might be good. Not seeing something as the answer to everything often means we are starting to pass the honeymoon stage.

[makefunstuff]

Maaaybe, just mayyybe, in training weights there are not enough examples of producing valid assembly? And spitting something that was being fed by scraping oss repos is easy to impress for glorified autocomplete machinery?

[iFire]

Huh?

I thought the latest advance in computing (spring 2025 - last year) is self-play / reinforcement learning. Like we've ran out of training data a few years ago.

https://github.com/OpenPipe/ART

Reinforcement learning having the large language model devise puzzles that they solve via llm-as-judge.

The definition of llm-as-judge is your llm generate 8-12 trajectories and a different llm judges the result. I'd use an oracle like windows or linux operating system execution for the problem of ISA-assembly creation.

The winning entries are used to train the large language model.