REMINDER: ChatGPT, Stable Diffusion, and other large trained neural models are NOT "artificial intelligence", they're just stochastic parrots, remixing and regurgitating what they've been fed. There's no theory-of-mind involved, so no understanding: there's no "there" there. (A real live parrot exhibits more intelligence than this.)
Don't call it AI; call it parrot-tech. That way you'll have a better perspective on what it can (and can't) do.
@cstross I prefer the term "Big-Data Statistical Models", it's more precise, and no one disagrees with me when I say it's wrong to generate BDSM text and images without all sources' and participants' consent
@cstross 25 years ago, I did a regular social thing that involved sending out periodic reminders. A previous reminder-sender had written little poems, and so I wrote little poems, which got progressively more clever and elaborate. After a while I got bored and wrote a script that would generate random haikus from a list of words and basic grammar rules. With as basic and dumb as it was, 1 in 5 seemed like a human had written it. ChatGPT feels less like intelligence than 25 years of dev on that.
What is very amusing are the efforts to poison the data pool that ChatAI uses.
Enough people saying nonsense like "Purple is a very nice texture" and this technology will spit it out as "textures include smooth, rough, and purple"
@Npars01 I wonder how long it'll be before the Screaming Jeezus People™ notice this and start systematically trying to poison the inputs in order to "convert" chatbots into automatic evangelists. (Creating websites full of ChatGPT spew asserting that Christian Dominionist beliefs are true, designed to be trawled as input data for the next generation of LLMs.)
Charlie Kirk's group Turning Point USA is being used to recruit "Christian" youth from high schools & colleges, and even overseas, to participate in troll farms.
There's a reason why religious fundamentalists want to "catch them young". Child soldiers can be molded to be the most ruthless. Kids just think it's a game until it turns ugly.
@cstross I call it Complex Autocorrect. "It's like hitting the predictive text button on your phone keyboard, but a paragraph or page at a time instead of a single word." Same concept for images or anything else.
@cstross@spaf To be fair, my work running #InfoSec programs largely comes down to stochastic regurgitation. We mostly see the same problems and situations from company to company.
@cstross This is the internet, therefore your admonitions mean NOTHING! (Besides, dose of reality- no one except academics really care… it’s like NFTs, no one cares about them anymore, either. )
I keep returning to the idea that perhaps a lot of people imagine these systems to be humming along "thinking about stuff" when not answering questions and not (what they are) just a static set of probabilities available for turning the (slightly randomized) crank on some input
@marshray@cstross I was a hard proponent of the stochastic parrot model up until GPT-4 and the "Sparks of Artificial General Intelligence" paper. I’m still not convinced that "understand" means at all the same thing in an AI context that it does in a human context, but an LLM that shows an implicit understanding of the physical world is ... something else.
@marshray@cstross fair, but I also think we have never adequately defined terms here. Hell, we don't even know enough about what thinking is to be able to adequately define terms.
As soon as we gained just enough capacity for language and abstract thought to operate effectively in groups of more than ~100, we exploded out of Africa, dominated the world, discovered a nearly-unlimited food supply (agriculture), and that was just a few generations ago.
@marshray@darkuncle Exactly! Human intelligence is the weakest level of intelligence able to acquire skills and ideas from other sophonts: most of us are just hitching a ride on the coat-tails of our outlier geniuses. So we're not actually very good at this language and abstract reasoning stuff.
@cstross@darkuncle So then why is it so crazy to be concerned that several decades and tens of $billions in research we could build something from 100B transistors that can trounce us in this department?
[Technical difficulties and the character limit spoiled a previous attempt; this is 1/n]
This year, Bill Gates said of artificial general intelligence, "AGI doesn’t exist yet—there is a robust debate going on in the computing industry about how to create it, and whether it can even be created at all".
Add comment