Comment by AtlasBarfed
1 month ago
This is one of the real canaries I watch on "real AI" for programming.
It should be able to make an OS. It should be able to write drivers. It should be able to port code to new platforms. It should be able to transpile compiled binaries (which are just languages of a different language) across architectures.
Sure seems we are very far from that, but really these are breadth-based knowledge with extensive examples / training sources. It SHOULD be something LLMs are good at, not new/novel/deep/difficult problems. What I described are labor-intensive and complicated, but not "difficult".
And would any corporate AI allow that?
We should be pretty paranoid about centralized control attempts, especially in tech. This is a ... fragile ... time.
AI kicks ass at a lot of "routine reverse engineering" tasks already.
You can feed it assembly listings, or bytecode that the decompiler couldn't handle, and get back solid results.
And corporate AIs don't really have a fuck to give, at least not yet. You can sic Claude on obvious decompiler outputs, or a repo of questionable sources with a "VERY BIG CORPO - PROPRIETARY AND CONFIDENTIAL" in every single file, and it'll sift through it - no complaints, no questions asked. And if that data somehow circles back into the training eventually, then all the funnier.
That's one of the boil-ups. Why would lack of Linux compatibility for hardware be a thing? If AI can write the drivers in 1/10th the effort/time, it should be a game changer for open source.
I haven't heard much from the major projects yet, but I'm not ear-to-the-ground.
I guess that is what is disappointing. It's all (to quote n-gage) webshit you see being used for this, and corpo-code so far, to your point.
AI can't write full drivers, and certainly not to mainline Linux quality. But it does make "take apart a proprietary driver to figure out how it works" much easier.
>It should be able to make an OS. It should be able to write drivers.
How is it going to do that without testing (and potentially bricking) hardware in real life?
>It should be able to transpile compiled binaries (which are just languages of a different language) across architectures
I don't know why you would use an LLM to do that. Couldn't you just distribute the binaries in some intermediate format, or decompile them to a comprehensible source format first?
I agree that it's a challenging problem.
My line of thinking is that AI essentially is really good at breadth-based problems wide knowledge.
An operating system is a specific well-known set of problems. Generally, it's not novel technology involved. An OS is a massive amount of work. Technical butrudgerous work.
If there's a large amount of source code, a great deal of discussion on that source code, and lots of other working examples, and you're really just kind of doing a derivative n + 1 design or adaptation of an existing product, that sounds like something in llm can do
Obviously I'm not talking about vibe, coding and OS. But could an OS do 99% of that and vastly reduce the amount of work to get a OS to work with your hardware with the big assumption that you have access to specs or some way of doing that?
[dead]