GrapheneOS defends use of AI for coding

GrapheneOS is currently defending its use of AI coding tools on Mastodon against complaints by various accounts claiming to be users.

We do not understand where you’re coming from or why you’re so incredibly angry with us. It’s not justified and does not make sense.

grapheneos.social/@GrapheneOS/117243573090596184

OSS projects have already started to recognize that LLM is just a tool and how you use that tool matters a lot more than whether you are using said tool at all. Lets bring StackOverflow for analogy: we didn’t ban browsers for code development because one can navigate to StackOverflow and copy-paste some really crappy code. LLMs are SO on steroids (actually trained on SO topics). So grabbing LLM to get a sense of direction or shape is an OK use in general, but copy-pasting it’s results blindly is silly and dangerous.

There’s a second argument that AI is ruining OSS communities which I believe is true, but refusing to accept AI code is a self-defeating action. Why? Because AI training needs unadulterated source of knowledge, one not tainted with AI and OSS communities rejecting AI would be that source. Also: if OSS community did not use AI it would hardly stop it from continuing it’s growth in Enterprise. So there’s not much of a winning strategy there either.

Which brings us to GrapheneOS decision and given above reasons I think it’s a well-balanced decision.

Yeah, just one small problem, muchacho. Whenever a language recieves a breaking change you gotta retrain the model again, and a lot of your references are useless now. There is no “improving” it’s just a vicious cycle.

Your argument is that of peer pressure and not compelling.

“Join us or we’ll just steal your code anyway”

Here’s my rebuttal. Garbage in, garbage out. Most projects out there are trash. This tool doesn’t help you build good code, it helps you throw together a pile of trash. An actually helpful LLM would be one trained on the damn language itself and coding best practices. Almost everyone has lost the plot here. You DON’T want a massive LLM, the small specialized ones are good enough. So you can use it as auto complete, like the tech was intended.

You’ve made a few of erroneous conjectures which are orthogonal to my argument. But I’ll call out one: LLM is bad at code completion and it’s use not have to be limited to (final) code writing. Thus my theses on “what” vs “how”

LLM is just fine for next text prediction, in the same way my phone keyboard can sometimes guess what I’m about to say. But the notion of building this huge mass is where I draw the line, it’ll never be finished. Does overfitment mean nothing to anyone anymore?