I disagree with this framing. All tools are this way, and they are still “just a tool.” An inert lump of metal on a stick, used to whack other bits of metal into bits of wood, shapes how people think when they use it. That doesn’t make it anything more than metal on a stick. Other tools can be more opinionated, in the limited scope of what they make trivial. A gun that’s not lethally dangerous isn’t really a gun. It has no other function. But in debating where it’s used, and why - or even where it exists with the mere potential to be used - it is still just a tool. You have ideological implications, dude. It came free with your hippocampus.
LLMs are just a tool, in that the loudest complaints about using them also apply to other tools, but the people complaining have not critically examined the tools they consider normal. I’ve coded with and without a compiler. It changes your habits, and what you consider feasible. It changes how much time you spend thinking about different aspects of the project. I’ve also coded with bad compilers, and that diverts significant attention toward thinking about how the tool might betray you. And yet: it beats coding without a compiler.
C is kind of a shit language, permanently joining the lingua franca of programming to how PDP-11s worked in 1978. But it has survived and spread because it sufficiently represents the low-level concerns of what computers do and why you want them to do it. It’s harmful insofar as Ghidra and Mythos can trivially spot buffer overruns and undefined values in, frankly, far too many programs. Yet it remains immensely useful for abstracting the architecture of any processor. Open-source software had little trouble shifting from x86 to ARM to RISC-V, even though RISC-V is a loose confederation of suggested instructions. And LLMs understand C well enough to translate to much safer languages… while also rearranging enough underlying details that we have to bicker about re-licensing.
No tool is neutral, but they are still just tools. What a person can use them for is a concern, but all moral judgement should focus on what a person does use them for.
Say that metal on a stick was made from a rare metal that can only be sourced by mining the land of a threatened indigeneous tribe. If I use or buy that tool am I really neutral, even if I build hospitals with it?
‘I agree that nothing is neutral, but finding that ideological complication in a simple hammer doesn’t make it more than a hammer.’
‘Okay but what if I invent a hypothetical hammer that’s even more ideologically complicated; is that neutral?’
Mu.
Anyway: there will come vegan models. There will be LLMs and diffusers trained entirely on public-domain and licensed data. They will have the exact same effect on users, for the same reasons that a tungsten conflict hammer still makes you think about misusing a screw. For better and for worse, it does the thing it’s for. That’s what tools are. Solving any ethical concerns with how they’re made will not address this particular discussion.
Maybe more ethical choices will come. In the meantime we can reduce harm today and incentivise those building the hammers if we choose the ethical ones now.



