just reiterating

posted

I had to have The Talk with my boss today.

You know the one: Nobody's saying you have to use 'AI', but you really should experiment with frontier models. Almost everyone else on the team loves them. They're inevitable! Surely this will be the last job you'll ever have where chatbots aren't mandatory. Even if you aren't going to use them, at least stop complaining when people send you obviously-vibe-coded long 'AI-assisted' PRs.

Everyone here should already know my stance on these things, but just because it makes me feel better, I'm going to write it down again: other people think that LLMs are "cool" or that they "improve engineering velocity". I think they've evil. These aren't positions with a middle ground between them.

LLMs[1]:

  1. Are unethical in their construction; they were trained by stealing the entire corpus of human work and feeding it into a statistical engine. Aaron Schwartz died because of copyright law not 15 years ago, but when Sam Altman does something 100x worse, we're supposed to just shrug, take a pittance of a settlement and move on?
  2. Are unethical in their execution. Doing inference requires enormous, beefy computers and large "agentic" workflows typically use between 10x and 100x as much power as "normal" computer operations; a single "one-shot" in a cutting-edge model in Claude Code can easily consume hundreds of watt-hours. Beyond that, training uses at least several orders of magnitude more power[2]. And, of course, all of this is happening in "neo-clouds", powered by the dirtiest possible gas turbines and enriching the worst men alive.
  3. Are unethical in their goals. The most likely[3] outcome of the rapid rollout of LLMs to white-collar jobs can best be described as "what if we could do to white-collar jobs what globalization in the 80's and 90's did to blue-collar jobs?" Business use of LLMs promises to replace high-quality, high-cost, context-aware skilled labor with cheap, mostly-unskilled labor driving a pseudorandom statistical text processing engine. I'm most exposed to this in the workplace, but the evidence is pretty bad for the impact of LLMs on education, too.

None of this has anything to do with whether LLMs increase some business metric like "time to first powerpoint presentation" or "number of UI features shipped per quarter"[4]. It doesn't even have anything to do with whether or not LLMs are an interesting technology. They're just bad. They way they were built is bad, the way they run is bad, and the outcomes for the world are bad.

This isn't the first technology I've felt this way about. More than a few times over the last 10 years I've had to walk away from companies or job offers because they were using cryptocurrencies, which I similarly think is technically somewhat interesting, but also evil and bad for the world. The rhetoric is even similar — 10 years ago, you had people saying that Bitcoin would be the world's reserve currency by 2020, and that anyone who didn't get on board would be left in the poor house when it went "to the moon"; hell, Jack Dorsey renamed Square to Block in 2021 because of how much he thought blockchain technology was going to change the world.

Surely everyone has something that Big Business loves that forms a line they won't cross; maybe it's all the way down at "slave labor" for you, but there's got to be some point where you're willing to engage the moral centers of your brain, right? Nobody actually reaches the fully-nihilistic pit that Linus seems to be stuck in where ethics are a purely-personal matter and don't exist in software projects?

Anyway, I'm hopeful that this hype cycle will pass, too. We probably won't go back to where we were, but maybe we'll go somewhere where I need to have The Talk a little bit less often. Or at least maybe I'll come up with something more cheerful for my next post.

  1. I try not to use the terms "AI" or "Artificial Intelligence" because I think they tend to lead people to anthropomorphizing the technology. Obviously large language models are using what would (in academia) be called AI or ML techniques, like reinforcement learning and neural networks, and obviously they're able to generate prose that for many casual observers passes the Turing Test, but equally-obviously they're not operating in any way that could plausibly be confused with conscious, sentient beings, so I find it's best to avoid the loaded term of "intelligence". ↩

  2. Companies like Google will gladly talk about how little power a single small session of inference consumes, but every lab is absolutely silent about how much power training takes. I wonder why? ↩

  3. And, let's be real here, intended ↩

  4. I'll reiterate, separately from the rest of my argument, that the output of LLMs really is not very good. I'll let you know when I can go a few days without seeing an LLM make obvious category errors (which, of course, because LLMs have no qualia and no actual understanding of the world, are just expected outputs). LLM-generated prose is absolutely horrible to read and it's now at the point where I have to close multiple articles a day because they're unreadable, obvious slop. LLM-generated code is verbose and mediocre, the kind of thing that we used to make fun of "Java programmers" for writing. ↩

  5. He was also great on Ed Zitron's Better Offline podcast a few weeks ago ↩

Want to comment on this? How about we talk on Mastodon instead?

mastodon logo Share on Mastodon