LLMs are a Cognitohazard

LLMs are a Cognitohazard by cthos 1779 words So, I've been watching a bunch of folks whom otherwise have reasonable takes about LLMs who, under very specific circumstances, have extremely weird takes about how much an LLM is helping them do a thing, and their internal justifications for them. This isn't about anyone specific; for each of these points, I've seen variations of these from more than one person in every case. But they're happening enough that I want to talk about them. Don't just take it from me, this post from @xgranade makes the point better than I can: It is amazing how many people I see claiming that AI has one good use. No one seems to agree on what that one good use is, and there's always a hell of a lot of goalpost shifting and special pleading involved. For some reason a lot of folks quite reasonably follow the arguments against AI, but then partition off the one thing as immune or exempt from having to worry about any of the ethical and practical problems. - @xgranade@wandering.shop Let me start, though. I think each and every one of these things ignores the externalities of running an LLM: The wholesale scraping of data frequently without consent (and the associated pushing of costs onto the site being scraped), the massive build-out of data centers that will never be used, a global chip shortage because the AI industry can just do that, the proliferation of slop and disinformation, and the trend of people experiencing psychosis when interacting with chatbots. Okay, with that out of the way, let's dig in. These are paraphrased. They are not direct quotes. If I use quotes, they are scare quotes to separate them from the rest of the sentence. LLMs are great for Prototypes - I'd never have the time or ability to do this. It'd take me forever, or I'd never get it done. Permalink to “LLMs are great for Prototypes - I'd never have the time or ability to do this. It'd take me forever, or I'd never get it done.” Okay, so this one I've seen a whole lot lately. Namely, the claim that busy engineers are able to rapidly prototype something and get real time feedback on the prototype. This is usually associated with a sense that the engineer in question is "not a {insert type of} engineer" and so they'd never be able to complete such a task. There are a bunch of problems with this approach, but I want to start with the claim that these folks would never be able to do something. That's patently untrue. If you're a software developer you absolutely have the capability to learn how something works and build out a prototype. The LLM's statistically average output is going to produce code that is prevalent . The most prevalent front end code being fed through these things is React. React has some very silly contrivances in it, but it's not an insurmountable mountain...

Preview: ~500 words