Ex-Apple engineer and Niklas build a working OpenGL driver for the M4 Mac Mini in one month using an LLM

Maybe I'm getting really old, but the fact this works at all is black magic insane. My mental model of what's possible hasn't been fully rewired yet.

I still can't quite believe it either. Reverse engineering an undocumented GPU at this scale is normally a years-long effort.

The biggest pain point of Asahi Linux is no GPU acceleration on M3 and newer. But Asahi has a strict no-AI policy, so this work can't be upstreamed. Expect a bunch of AI-assisted forks to dominate while only purists stick to the non-AI version.

I imagine either they or someone else can take the discoveries and write a real driver now. The hard part was always reversing the black box system.

Asahi doesn't have a monopoly on Linux for Apple Silicon, and upstream Linux absolutely does NOT ban LLMs.

It's extremely impressive they made a working driver so fast. Best use case for LLMs I've seen: you don't need someone spending years reverse engineering undocumented hardware anymore.

The problem with today's models is they produce disposable software. Not well-architected or durable. If you just say here's a regression suite, write an implementation that passes, you get something that works for a while but ages quickly and needs to be thrown away.

Can anyone tell me if it's a good idea for upstream projects to accept this kind of code from an ex-Apple employee, especially with Apple currently suing OpenAI?

During my time at Apple I never saw any macOS source code, not even userspace. I haven't worked there since June 2025. WINE doesn't ban all former Microsoft employees, only anyone who ever looked at the Windows source. If I thought my Apple employment exposed me to relevant secrets, I'd refrain from contributing.

All this work is tainted since the poster is ex-Apple. There's no way Linux takes that code, people from Apple also contribute to Linux. Plus Codex and the Apple lawsuit over stolen trade secrets. Nobody knows if training data in other models is similarly tainted.

Very sad. I really want Linux on newer Apple chips working flawlessly, but I understand the legal minefield around this particular work.

There were and are many important pieces of the Linux kernel that live out of tree, ZFS being a big example. The problem with this driver isn't licensing or code quality, it's the threat of a C&D letter or an NDA suit.

Now that this work is out in the open, what prevents the next generation of LLMs from ingesting it and a non-Apple employee creating another driver from scratch? LLM-assisted knowledge laundering is becoming a thing.

I would never touch an LLM for RE work when there's a chance the models were trained on the thing I plan to reimplement. Asahi is trying to be cleanroom, so it's incredibly rude to not mention LLMs until after showing a page full of LLM-derived firmware ABI.

The previous Part 1 post is very candid about the approach, and the titles of both posts mention prompting.

The author was banned from Asahi Linux for hiding his extensive LLM use in another attempted contribution, and for concealing that he's a former Apple engineer with direct contacts to the people involved in Apple Silicon development.

I don't care about hiding LLM contributions or the anti-LLM policies. I just hope the author shares the process completely so someone not working for Apple can reproduce it.

It's extremely ironic that big corps openly steal IP en masse to build their models, but hackers are still worried about using the same models for reverse engineering. If Apple open sourced its drivers this issue wouldn't exist.