Hacker News

points by tw1984 21 hours ago | 0 comments

honest question - how hard it is to use LLM and tools like REA to reverse engineer MacOS to figure out how its hardware works in terms of getting linux support on such hardware?

I thought it is already a solved problem? with large enough numbers of agents running in parallel reverse engineering, implementing kernel modules, testing in vm etc, the whole pipeline seems to be complete.

Did I miss something here?

[–] cromka 21 hours ago

Yes, read that blog entry to understand how different that hardware is from any other Linux-supported ARM stuff to see how much deep understanding you need to actually bring that support upstream to Linux or Mesa. Some(most?) stuff requires fundamental changes and no LLM is going to design and write it in an uostreamable way.

Debugging is not an issue, Asahi seems to have it figured out, especially with the virtualization. It's the design that takes time.

Meanwhile Omarchy, for example, just brute-forces everything, with ugly patches that work today but cannot be upstreamed as such, so they'll break tomorrow with new kernel release. I really expect them to start struggling a lot going forward with maintaining the support they claim. Regressions may actually kill their effort.

[–] KetoManx64 13 hours ago

You must not work with AI tools very much. The brute force approach as a first/second pass is very standard when doing reverse engineering with AI. You get the functionality figured out and get the basics working and then you just iterate on that and massage it into what you need it to be. Omarchy has practically unlimited tokens and every other week there is a new model being released that is more captable than the previous that will just massage things into place with each iteration.

[–] saagarjha 1 hour ago

I work on reverse engineering tools and previously trained models, and no, this is not correct.

[–] bigyabai 7 hours ago

> Omarchy has practically unlimited tokens

And yet they have a total of zero upstreamed PRs, and haven't fixed enormous and glaring issues like PSCI control either.

It's almost as if the parent is correct, and Apple's ARM architecture isn't standardized enough for ordinary models to blindly reverse-engineer without documentation. If tokens were the only constraint on supporting Apple Silicon, we'd have reversed the entire SOC device tree years ago. Alas...

[–] KetoManx64 6 hours ago

> And yet they have a total of zero upstreamed PRs, and haven't fixed enormous and glaring issues like PSCI control either.

Why should I care about if they upstream their patches or not? I have a working M3 macbook that runs Omarchy, why do I care that it's in the mainline Linux kernel? My GrapheneOS phone doesn't run mainstream Linux kernel, but it does the job that I need it to do.

[–] cromka 5 hours ago

Just read other people explaining this in this thread.

[–] tw1984 6 hours ago

upstreamed PR means nothing in 2026. I couldn't care less when some maintainer are just so political, still remember the rust in kernel drama a couple of years ago? why great work have to get approvals from those people?

[–] cromka 5 hours ago

Sure, let's see how the hardware you enjoy in November behaves in March after they rebase the kernel to 7.5.

[–] dj_morocco 20 hours ago

They prohibit RE of the binaries. What Asahi doing is observing macOS as a black box with input and outputs, and translating observed behavior back to Linux.

[–] tw1984 6 hours ago

who are "they" here?

clean room RE is legal across the world. with AI, it is increasingly easier to do it right.