It's a common capitulative "hey please take me seriously and don't write me off" response to the extreme political polarisation of opinions on modern AI into a typically right aligned "powerful and broadly a force for good" and a typically left aligned "useless smoke and mirrors but broadly a force for bad" to say "both of these responses are wrong, they lack nuance and I believe something more complicated than either". Which is of course true and it's something I do myself say, but the capitulation (which I'm also guilty of) lies in hedging about the specifics and letting the other person assume that you agree with whatever is the most important to their politics.
But I think I need to be a little braver on challenging other leftists on part of this - or maybe, people explicitly on the left who have a little more domain knowledge of AI sort of need to force themselves to be. Perhaps part of the problem is how few there are who haven't been scared off, booed from the stage egged by their peers. It's certaintly why I quit serious academic study or career pursuit around AI, why I'm generally afraid to talk openly about the subject with people who know me personally except a very few very close friends. I know that I'll lose those friends before I can delve deep enough to be heard in any real explanative way, that getting there will involve challenging too many received wisdoms. Which is stupid, because I'm generally open to what should be the most important part of their politics on the matter, in that I am absolutely not convinced that AI as it's currently being deployed is going to be a prosocial force - although I believe it still might be, especially if so many leftists stopped spurning the technology.
But my problem is that they believe it's bad for - aside for perhaps a fairly skewed if righthearted version of automation threat - the wrong reasons, which are all predicated on this idea of the technology as an impotent, resource sink; the bejeweled sockpuppet inflated by the bloated dick of a corporation. Nonsensical mantras about data centers and energy use and water, which are absolutely nonsensical when you actually look at the numbers involved in the context of the energy, data and water use of any number of digital infrastructure technologies we take for granted. The cost of a cup of coffee, of locally grown vegan produce, of playing Mario Kart for half an hour, of posting on tumblr and reading your friends reply.
But perhaps the reason for this disconnect is the same sense that AI can only ever be a useless deception or party trick, so to get to my actual salient point. Your average leftist, marinated in a stew of posts about hallucination or counting letters in "strawberry", who recalls 2022 Dall-e outputs with too many hands, who still calls LLMs "spicy autocomplete", is catastrophically underaware of the consequence of the generalized abstract reasoning capability codified in the world models of these systems. The most critical mistake being made here is not in thinking that AI "is bad" (perhaps it is!) but by assuming that it's not powerful.
The right wing AI superintelligence truthers are far more close to being correct on this axis than the average leftist.
And because there's such a strong culture war imperative to stay in lane here, said leftists are almost to a man unwilling to learn more or be corrected. Which means that they're all entirely unequipped to address the actual problems with AI. Intrinsic to any analysis of AI risk should be the assumption that it at least has the potential to be powerful in the way that it's being imagined as being, and the best information we have on the topic is that it will be, perhaps sometimes already is.
Like, there is a whole field of study here. All the leading experts are convinced of it and those guys aren't partisan stooges or idiots. A ML researcher with a doctorate does genuinely have a more informed opinion than a tumblr fandom blogger, and all those fuckers seem sort of terrified at the moment. And we have metrics! Evaluations, benchmarks, cold hard numbers that show yes AI is good, this is how good it is, here's a percentage accuracy. Yes the failure mode of AI is hallucination, but at a certain point that can be minimised to an extent that the model can self-correct and we're getting close to the point of closing the gap. We likely already have in certain contexts!
On my phone so hard to fetch links and examples. I'll maybe edit those in later.
I feel like part of the issue is it's very easy to scratch the surface of an LLM and expose the artifice of the persona, of the output text-stream as pretended speech or thought. But this is missing the trees for the forest; yes what the AI shows its users is the model playing pretend for the sake of the text interface, but the actual AI is the internal world model that allows it to play pretend competently, and that is what is powerful. So many anti-AI people refuse to comprehend this fairly basic insight and it is maddening.
So yeah, essentially; if your critique of AI refuses to account for the possibility that the technology can be genuinely powerful, you don't know what you're talking about and you're unequipped to fight for your cause.
I don't know. I've given up talking to people about AI at this point except for on this blog. I figure whereever the wave takes us my talking about it isn't going to divert things so I might as well focus on other parts of my life. I basically gave up on having a well paying career in a topic that interested me so I could scrape by doing menial systems maintenance and not have to think about it seriously any more. I mean, I don't think I ever had the chops for it anyway, not when I realised that this would be the tone of the conversation. But still every so often I peek back at how the people around me are talking about it and it's just. So bleak.
gummygunk
blakekathryn
theartofmadeline
janainaart
