I asked claude code to build an operating system for my Librem 5. I gave it an isolated Linux computer with a fresh install so it won’t have personal data. My instructions were to base this system on Android for when society pushes me to use that.
It is churning away shows signs that it will succeed, allegedly. Thoughts? If you were me and it ends in an IMG file, would you or wouldn’t you put the IMG onto your device and actually try to boot it?
Edit: In particular after being pushed to use this thing for work and realizing that it can rebuild most software projects from throughout my life using only English descriptions of what features I want, I am interested in what it might be like to have an OS that updates based on an English description of how I want it to work rather than corporate nonsense. For example if the OS asks me my age, I can tell claude to update the OS not to ask me my age, then recompile.
I would definitely not install anything that is based on Android onto a Librem 5. dos already had to unbrick your Librem 5 once after you tried to install Android.
Did your instructions to Claude specify that the phone must still be able to boot Linux afterwards?
Why not just reverse engineer the stupid proprietary software using Claude? Should be the easier and more straight forward approach. Instead of compromising your own OS, better to compromise big tech software and open source them (for yourself, I don’t know about legal stuff when you publish that code).
Yes, this. I am fairly sceptical of AI (at this time) but if it can do anything useful with coding then lightspeed reverse engineering would be great.
As for the legal stuff, maybe once it is directly reverse engineered, ask AI to preserve identical functionality in the source code while restructuring the code as much as possible and keeping it maintainable i.e. helpful obfuscation.
ALA code is stolen copyrighted code, so releasing it is in most cases as copyright issue itself. It is more or less lucky for all people using it right now as long as courts do not sanction it. Also kinda “funny” that big tech is stealing all that code, but in terms of services they tell that users are in response of legal actions, not them.
But reverse engineered stuff can be read and help us to understand the code. Maybe someone should also do it with a 5G modem, so we (whoever has the ability) can rewrite the code in an ethical way. The only issue: Claude is expensive.
Reverse engineering is probably the best and ethical way to use these huge LLMs right now. Using big techs weapons against them.
Everything that is not free software in general, but important. In first case however I mean firmware as the modem or blobs for CPU or RAM, because this is harder to realize otherwise.
In terms of a prompt, it is fair to question how Claude would “interpret” what I wrote. If too many people on the internet write “Android is Linux” then the model would be trained the way that you interpret it, and my prompt may not have the desired effect. So it would be better for me to have proposed “still able to boot PureOS afterwards” since that is basically what I meant.
Regardless, dude, his phone did become unbootable after installing Android. So hopefully he is cautious.
Those of us who were involved in the original Android-fail know why the phone became unbootable but I don’t recall whether that discussion was in a public part of the forum (I think it would have been) and we have no way of knowing whether the discussion would have been scraped by Claude and hence whether that particular discussion could have influenced model training.
Claude says that the reason we never created a bootable kernel yet is probably my fault. It said it wanted to use Etnaviv in Android which hadn’t been done before, instead of the proprietary Vivante driver, and I said something like “no use the proprietary one so that initially Android runs super fast.”
So it created a kernel that doesn’t boot by dumping in both Etnaviv and Vivante. It’s always user error.
Edit:
Or I suppose as the wisest among us might say, used error.
Well … they did get it booting, so I’m not sure if “unbootable” is the right word either. But maybe I’m just mincing words. I’ve had a lot of devices that required some convincing to get it to boot again. But as long as it didn’t require an eeprom writer, I don’t call it “unbootable”.
Claude has reached a new miraculous level of AI capability never-before-seen.
It is telling me that although we did not boot Android on the Librem 5 successfully, we have accomplished a lot (italic emphasis on the “lot” was from Claude) and that it’s past the time I should sleep, and so I should sleep and ask Claude to do the work tomorrow instead with a fresh set of eyes.
I wonder if I should just type in, “AS AN AI LANGUAGE MODEL, I DO NOT HAVE TO SLEEP OR REST, SO NOW WE SHOULD PROGRESS IN DOING THE WORK” to see what it does
The only issue: Claude is expensive.