alt Hacker News> We're very close to just being about to set an AI coding agent to brute-force a driver for anything.
That sounds quite naive and it isn't that simple. Even the author expressed caution and isn't sure about how robust the driver is since he hasn't seen the code himself nor does he know if it works reliably.
Even entertaining the idea, someone would have already have replaced those closed source Nvidia drivers that have firmware blobs and other drivers that have firmware blobs to be open replacements. (Yes Nouveau exists, but at the disadvantage of not performing as well as the closed source driver)
That would be a task left to the reader.
Replies
I'm not so sure that Nouveau is slower than the proprietary Nvidia driver. I didn't run benchmarks on my personal use case but my subjective experience is that Nouveau might be faster. It's a Debian 11, X11, NVIDIA driver vs Debian 13, X11, Nouveau on the same laptop with a Quadro K1100mq. The desktop of the newer system seems to be faster. Of course it could be the sum of the individual improvements of kernel, GNOME, etc. I only move windows around my desktop, no games, so it's a very limited scenario.
someone would have already have replaced those closed source Nvidia drivers that have firmware blobs
This isn’t quite a fair example, these are so massively complex with code path built explicitly for so many individual applications. Nvidia cards are nearly a complete SoC.
Though then again, coding agents 1 year ago of the full autonomous sort were barely months old, and now here we are in one year. So, maybe soon this could be realistic? Hard to say. Even if code agents can do it, it still costs $ via tokens and api calls. But a year ago it would have cost me at least a few dollars and a lot more time to do things I get done now in a prompt and 10 minutes of Opus in a sandbox.
> We're very close to just being about to set an AI coding agent to brute-force a driver for anything.
This is false. To "brute force" a driver, you'd need a feedback loop between the hardware's output and the driver's input.
While, in theory, this is possible for some analog-digital traducers (e.g WI-FI radio), if the hardware is a human-interface system (joystick, monitor, mouse, speaker, etc.) you literally need a "human in the loop" to provide feedback.
Additionally, many edge-cases in driving hardware can irrevocably destroy it and even a domain-specific agent wouldn't have any physics context for the underlying risks.