Epistemic status: been thinking about this a contradiction I've seen in some of my friends views on AI and think its likely that a lot of people think in this way, especially in communities that really don't like it (eg. most of Lemmy). I know a ton of math and read/work a lot on AI theory+mech interp+safety, so I think I have a pretty strong understanding of the underlying mechanics. I'm fairly confident on the general ideas in this post, but as I explain at the end: a very nuanced understanding involves holding many of these otherwise contradictory viewpoints as true. Writing this because I genuinely want those against AI to have an easier time convincing people.
(for this AI=LLM)
It is not reasonable to think both that
- AIs are going to be forever incompetent or
- AIs cannot create new things (excluding creative works) or
- AIs cannot do very simple tasks
and
- AI will replace my job or
- AI will bring mass surveillance or
- AI can produce deepfakes that are 100% convincing or
- AI can be superpersuasive, manipulative (can cause AI psychosis)
(a lot of popular articles on AI fall into one of the above categories, and are often targeted towards the same people)
On another note, I have yet to hear a person against datacenters to give me a good reply to the question "Isn't the water just recycled back into the system?", and in fact they're usually uncertain why the datacenters need the water in the first place. I'm not some accelerationist and I don't think we need to cover the globe in datacenters, but those strongly against AI should have more concrete views than AI = bad, and then be able to back them up without having to do a search for sources, just from memory.
And yes there are exceptions to this in which holding both views make sense but this requires a lot more nuance but unless you are interested in this, that extra info probably would just clutter up your worldview.
Those things are not contradictory they just required nuance. If you put an llm in a customer assistance role, it will be almost entirely useless. One had a full on breakdown when it tried to run a vending machine in a test. But give an llm obscene amounts of information to sift through to look for patterns and it will thrive.
People tend to say llm is useless because people running the llm companies tend to push it into deployment in places it has absolutely no business being in, it's their only real experience with them, and it is usually forced on them without their consent making the interaction negative before it ever began.
It's not useless but many people encounter it in areas where it's either not helpful, or an active hindrance. They're working with the information and first hand experience they have and when they encounter it, they're usually correct in that circumstance that it is indeed useless. Nuance goes in both directions.