Hacker News Reader

Comment by ndiddy | Hacker News Reader

It's extremely impressive that they were able to make a working driver so quickly. I think this is one of the best use cases for LLMs. You don't need someone to spend years reverse engineering undocumented hardware anymore. It will interesting to see how good the driver the LLMs came up with is, and whether it can be upstreamed into the Linux kernel.

View on HN

Might be one of the only use cases where LLMs are necessary to succeed. Can't imagine the pain of reverse engineering a technology of this scale and "secrecy".

Traditional hardware RE centers around heroics by a small number of very smart people, see: marcan, Alyssa Rosenzweig, Asahi Lina, etc. What they did is insanely impressive, and given how hard it was with LLMs I'm sure I would not personally be able to do this work without LLMs. That said, I'm looking forward to a world where the staring at hexdumps and traces can be delegated to LLMs and talented humans can spend their time building as opposed to head banging.

Yeah I think it's a shame how much time and effort gets put into understanding things simply because other people want to keep them proprietary.

Asking out of genuine curiosity... Do you think the Asahi maintainers could use your work to speed up future GPU driver development? M4, M5, M6 and so on... I am assuming they wouldn't just accept your work outright, but instead could they use it to substantially speed up their development process by mapping out a lot of the GPU after which they can focus on optimization and other aspect of drivers? Do you think they will?

I don't use Asahi now but that's largely because my current Apple M series hardware is still supported. I see a day where an older unsupported Mac is running in a rack in my basement or being used by my kids as a general purpose computer.

I can’t help wondering if the whole process will be automated being an LLM.

Apple releases the M15, and using everything leaned while building drivers for the M4 - M14, the LLM just does the whole job and we have the drivers a day later.

The problem I’m finding, at least with today’s models, is that it produces disposable software. It’s not good at making well-architected, durable software. Stuff that could be maintained and bug-fixed.

I mean you can get good stuff out of it if you know what you are doing and guide it, but if you just say “here’s a regression suite. Write an implementation that passes” you will get something that works for a while but ages quickly and will need to be thrown away.

With today’s technology I’d still want a GPU driver developer guiding the LLM rather than some rando who is out of their element. But cutting down the exploration cycle time and giving the developer massive parallelism (have 10x agents exploring different hypotheses or features) is the real win. We don’t need to skip all the way to slop just to squeak out a little more effort savings.

I’m sure it’s really easy when you can get help from the people who made the chip or write the Apple drivers.

Did you read the article? They didn’t get help from the people who made the chip or wrote the Apple drivers.

He lied about his background, kept it quiet…

Reply on news.ycombinator.com