Artificial intelligence still has some way to go

like if we're worried about self-aware AI killing us, I could point us to the myriad of real, observed things that are likely to kill us all first

like the police, or poverty, or no access to basic preventative healthcare

hey guys i have a startup, it's called mr choppy ltd

So why can’t LLMs just learn what stop words mean? Ultimately, because “meaning” is something orthogonal to how these models work. Negations matter to us because we’re equipped to grasp what those words do. But models learn “meaning” from mathematical weights: “Rose” appears often with “flower,” “red” with “smell.” And it’s impossible to learn what “not” is this way.

hey but just wait until they learn what “not” means, nothing will be the same

that's such nonsense though

"not" would appear in a network of words alongside other words of negation, like "no" and "never"

and also, along a different axis, with other function words that can be used in grammatically similar ways

i personally feel there's something fundamentally true about word-meaning being largely associative. there's another piece too, for many words, but a lot of poetry and literature function along that associative line

Finally a researcher who is good at not saying that much about AI. But he is v good on those letters.

https://venturebeat.com/ai/top-ai-researcher-dismisses-ai-extinction-fears-challenges-hero-scientist-narrative/

"What I say to AI researchers — not the more senior ones, they know better — but to my students, or more junior researchers, I just try my best to show them what I work on, what I think we should work on to give us small but tangible benefits. That’s the reason why I work on AI for healthcare and science. That’s why I’m spending 50% of my time at [biotechnology company] Genentech, part of the Prescient Design team to do computational antibody and drug design. I just think that’s the best I can do. I’m not going to write a grand letter. I’m very bad at that."

Amen to this.

Finally a researcher who is good at not saying that much about AI. But he is v good on those letters.

This is my pal (and now coworker) KC! He’s the best and he’s 100% correct.

Someone recommended I listen to the Holly Herndon podcast. I took them up on their recommendation, and as a fan of her music, I _really_ wish I didn’t. Herndon, like many artists and cultural critics discussing AI, come across as entirely unaware that the distance, from a science or engineering perspective, between new and old methods is far smaller than, for example, the distance between a world without frame buffers and the world with frame buffers. I’d love to ask her, what are the cultural changes when your preferred interpolation method goes from “pretty good” to “very good?”

In fact, after I write that, I think it’s extremely fun that presently the most ballyhooed auto-regressive method is so simple that it could be reasonably reproduced by an excited primary school student over the weekend. It’s sad that we mythologized this rather than make it a neat example of understandable science.

xp yikes, that’s too bad.

One of my PhD-having coworkers joked while presenting how he’s running one of the current protein folding systems that its release made his work that he did to get that PhD obsolete.Which may be true in a way, but the accessibility of ChatGPT, etc., have just presented a public face to just the current step in a long series of efforts. We wouldn’t be here without the work.

The way a lot of articles have been written, you’d think computers just got sufficiently powerful and someone threw a bunch of text at one until a chatbot popped out of it like it’s Zeus’s forehead, fully formed

Huh, I'd been wondering whether we should have a "Who is Eliezer Yudkowsky and can we eat him freeze him for later eating?" thread. An acquaintance I occasionally read the twitter of is in with that crowd, so I occasionally go read Yudkowsky's - though this is a bad habit that I should try to break.

There's some self-interest in the recent announcement (though, these are not generally people who clamour for more government regulation), but I think there's also a bunch of pareidolia, like with Blake Lemoine, some real "the beguiling voices you only hear when you stare at the flames for 200+ hours".

I'm not sure if I misread above, but Yudkowsky isn't Roko, he's just the guy who set up the whole LessWrong community - David Gerard has a good article on its effects on Effective Altruism, which reminds me to link to Elizabeth Sandifer on Yudkowsky, which contains the crucial context for him - he is first and foremost a crank, albeit one who has a lot of reach at present.

He's an interesting writer (in that he's not as terrible as you'd naturally assume) - I found this memorial (including the update right at the end of the comments) after the death of his brother to be moving and powerful, while also revealing that a very broken sense of humanity. I genuinely think his anguish is real, even when the sources are (elsewhere) silly.

(No, I will not be reading the Harry Potter work, though I understand that it's more highjacking a popular franchise as a framing for his thoughts, than anything that can really be called 'fanfic')

The fuel behind the explosion of capability they expect is the idea of 'intelligence' as a linear, number-goes-up, value: We are intelligent enough to make computers that will be more intelligent than us, which will make computers more intelligent than them, and so on and on infinitely, IQ one billion! They're generally big believers in the idea that intelligence is a real thing measured by IQ tests rather than a function of it and, as usual, when you get to that you're only 10 minutes from the word 'heritable' and then, as they say, you're off to the races.

There's a Strangelove vibe that creeps in as well - "If this is the battle for the end of the world, then surely the traitors to humanity are those that insist there are words we can't say / thoughts we can't think" - AFAIK Yudkowsky's not that far, he just has the usual Libertarian free speech brainworms - I wasn't surprised to see him linking to a "how I was cancelled" post by Kathleen Stock on Unherd earlier this week.

(There may also be a bit of "these things will be even further above me as I am above the masses, and will hold me (me!) in even more contempt then I hold them"... but that might just be the guy I know)

There are of course people in the space that are worth listening to, though they define themselves more as AI ethics than safety- I understand Timnit Gebru is a good follow there. The angle is one that's been mentioned a lot above, that we should consider the actual effects of this on people right now, and the intersections with already-existing injustices. Though for a lot of the doomers, that's just what they don't want - there's already a choice of apocalypses available, but none of those on offer centre these guys as much as they'd want.

Timnit Gebru is more credible than Eliezer Yudkowsky but she’s still very much a crank.

We are intelligent enough to make computers that will be more intelligent than us, which will make computers more intelligent than them, and so on and on infinitely, IQ one billion!

If it’s gonna do this on Windows I give it maybe 2 iterations before it crashes

The “more intelligent than us” thing I don’t quite get, computers already are in a lot of ways and have been for a long time. Idk if people are fiddling around with ChatGPT or Midjourney and can’t see the difference I dunno what to tell them. Hopefully this is just really good at fooling people.

There does seem to be something about negatives that doesn't sit well with ChatGPT. I can't get it to write a Perec-style story without an 'e' in it, despite trying various prompts. I'm surprised because I've got it to do things that would seem way more tricky than that. It didn't bat an eyelid when I asked it to write an acrostic poem about my trip to the supermarket to buy ingredients for chicken cacciatore for instance!

it doesnt do any oulipo or rules based composition well or at all, frankly

Some rules-based things it doesn't have a problem with - ask it to do a rhyming acrostic using the alphabet sequentially from A to Z, and it will comply. But no it doesn't do lipograms - or it will start to do one and then forget the rule after the first couple of sentences. Similarly, I asked it to compose a story using only sentences with exactly seven letters, but it couldn't maintain that past the first sentence or two.

"I'm surprised because I've got it to do things that would seem way more tricky than that."

You mean like programming tasks?

Been meaning to put through modernistic style poems and get it to re-write in older styles.

interestingly, GPT has a very very hard time doing poetry that -doesn't- rhyme
https://arxiv.org/abs/2305.11064

when you get to that you're only 10 minutes from the word 'heritable' and then, as they say, you're off to the races

lol

Timnit Gebru is more credible than Eliezer Yudkowsky but she’s still very much a crank.

― Allen (etaeoe), Sunday, June 4, 2023 7:06 PM (yesterday)

could you expand on that? in what way is she a crank?

personally i don't think it makes you a crank to think that real-world right-now issues like AI processing's impact on climate change, rights/pay/conditions for AI workers, sexism/racism at AI companies, training-set secrecy, or copyright matters... -might- be the biggest-stakes questions surrounding AI

but i do think you're a crank if you're so certain of this that you have made your brand one that mocks and disparages all conversations that consider that other questions could also be of equal or even greater importance.

could you expand on that? in what way is she a crank?

She traffics her beliefs the same way Yudkowsky does. Their positions are predicated on science fiction so they’re impossible to refute, e.g., posted this weekend:

Why aren’t the “godfathers” of AI talking about the massive data theft from artists & their lawsuits eg? Because that discourse is too beneath their genius brains to cover? They have to talk about grand endeavors like SAVING HUMANITY? Because their practices would be implicated?

— @timnitGe✧✧✧@dair-commun✧✧✧.soc✧✧✧ on Mastodon (@timnitGebru) June 4, 2023

This equivalency of embeddings and data is so absurd it’s impossible to refute.

If Yudkowsky is “Bill Gates is micro-chipping us with vaccines,” Gebru is “COVID-19 is a bio-weapon.” Gebru is far closer to reality but she’s still far from reality. She also benefits from credentials (e.g., Fei-Fei Li was her PhD advisor) that Yudkowsky lacks.

personally i don't think it makes you a crank to think that real-world right-now issues like AI processing's impact on climate change, rights/pay/conditions for AI workers, sexism/racism at AI companies, training-set secrecy, or copyright matters... -might- be the biggest-stakes questions surrounding AI

She isn’t a crank because of her beliefs. She’s a crank because she predicates her beliefs on bad science. If it helps to understand my perspective, I believe every issue you identified is an important issue.

Erm. We may disagree about “copyright matters.” I don’t know what your position is about this.

xp

thanks for the responses

This equivalency of embeddings and data is so absurd it’s impossible to refute.

could you explain that? I'm not sure what it means. And while that tweet is p strident, I wouldn't think the idea that generative AI models are trained on copyrighted materials would be controversial, but I'm probably missing your point

so far, there is no precedent on which to argue that training AI on copyrighted materials is a copyright violation. it's speculation based on non-existent rulings (and, imho, a dangerous precedent to be calling for.)

OK but a) there are several lawsuits going on right now, so it's a lot less speculative than "AI will prob end humanity" and b) I don't see that tweet really making a strict legal argument but a moral or ethical one. You can argue against her stance, but I don't think labelling it "speculation" make sense since the "theft" did already in fact take place

could you explain that? I'm not sure what it means. And while that tweet is p strident, I wouldn't think the idea that generative AI models are trained on copyrighted materials would be controversial, but I'm probably missing your point

Her implication is that it’s possible to reconstruct the original data from these embeddings. It isn’t. Her other implication is that creator-specific features are being embedded. That’s unlikely.

OK but a) there are several lawsuits going on right now, so it's a lot less speculative than "AI will prob end humanity" and b) I don't see that tweet really making a strict legal argument but a moral or ethical one. You can argue against her stance, but I don't think labelling it "speculation" make sense since the "theft" did already in fact take place

What makes this theft rather than fair use?

I'm not a fair use expert, but I don't think it's impossible to argue against it in some cases. Getty obviously thinks there's a violation, and afaict the reproduced watermark makes it seem like a decent argument: https://www.theverge.com/2023/1/17/23558516/ai-art-copyright-stable-diffusion-getty-images-lawsuit.

Tbc I don't personally have a firm opinion on this, I'm just objecting to the idea that taking this position makes someone a crank. I think the real test wrt copyright will be once if these bots start being (more) commodified.

btw etaeoe, I'll never find it in this big thread, but I swear *you* posted a paper somewhere recently itt arguing that you could reconstruct training data...?

https://arxiv.org/abs/2301.13188 was the paper. As you have probably guessed, I'm not a computer scientist so maybe I misunderstood the implications or I don't get the different terms that are being used

btw etaeoe, I'll never find it in this big thread, but I swear *you* posted a paper somewhere recently itt arguing that you could reconstruct training data...?

Yeah. I should’ve been clearer. Reconstruction isn’t a _definitive_ outcome. I think that’s why I’m puzzled by the copyright rhetoric. If the model can be used to reconstruct, it’s a copyright issue. If it can’t, it should be considered fair use. There’s nothing intrinsically problematic about the underlying methods (e.g., diffusion) and I don’t understand why we’re not presently equipped to deal with this distinction (there’s already caselaw about compression).

so-called AI godfathers don't care about copyright lawsuits because they're both inevitable and essential to prove out the legal ramifications and decide what counts as fair use and derivative work

weighing in on those things is something you'd do at trial, and with specific answers about the technology and how it works

researchers/programmers/etc. should act ethically and can act as advocates or whistleblowers, but I'd use them as primary sources for ethical and legal questions as much as I'd approach any random person who isn't an ethicist or lawyer

that probably came off as glib, in that any person is entitled to an opinion and so-called godfathers should have considered these things. the way general media covers AI isn't for the most part useful when it comes to evaluating its use both within existing ethical and legal frameworks or determining how we change those frameworks to address new technology

If the model can be used to reconstruct, it’s a copyright issue. If it can’t, it should be considered fair use.

It shouldn't be considered fair use if the results are going to be used for commercial purposes. Why should it be fair use to train a model on copyrighted data with the goal of producing content so you don't have to pay copyright holders?

That’s too small of a concern. Not SAVING HUMANITY.

— @timnitGe✧✧✧@dair-commun✧✧✧.soc✧✧✧ on Mastodon (@timnitGebru) June 4, 2023

I mean, the answer here is to stop paying attention to figureheads of the "AI movement" if they're providing nothing of value to the public conversation

if they only talk about SAVING HUMANITY then find someone else to listen to, because there's nothing there

xp if I understand etaoe's point correctly now, what they're saying is that there are two ways of thinking about this:
(1) all outputs of generative AI violate copyright because they were trained on (some) copyrighted materials
(2) some outputs of generative AI may violate copyright depending on [factors]

sort of like how you can definitely use a sampler to violate copyright beyond fair use, but it's not inherent to sampling that that is the case.

fwiw I'm not sure Gebru was actually saying (1), but this is why Twitter is a bad forum for complex arguments

Neanderthal's repost of her tweet was her being sarcastic

is there even any legal requirement to disclose what goes into a model? if someone builds their own private model entirely off of copyrighted material, how would anyone even know based on the outputs?

This is not unlike imagining if someone built a massive library of microscopic samples from famous songs and then used those to make new music in which the source samples were entirely unrecognizable. There wouldn't be rights issues raised because the end result is completely different from the inputs. (Yes, I understand that AI image engines do not actually piece together elements of existing images)

I'm somewhat open to the idea that people can opt their images out of publicly available models, even though I don't exactly buy that putting them in there causes harm in any obvious way that taking them out would somehow fix.

serious question for those who actually think AGI is gonna be able to self-replicate and produce world-ending superintelligence within 5 minutes or whatever - what is this going to run on? wouldn't this sort of thing just instantly overload whatever CPU it was running on?

something something nvidia stock price

It's surprising how recognizable even micro-samples are (and apparently AI is being used for sample-snitching now). And "There wouldn't be rights issues raised because the end result is completely different from the inputs." -- there definitely are if the sample is identified and the end result has made a lot of money, it doesn't matter if it's a one-shot.

I don't know how good they've gotten at recognizing all samples, and I also don't know how far they would plan to take sample litigation. Sampling is massively widespread and most of it happens without issues. My example specified that the samples in the context of the new piece of music were unrecognizable. Perhaps at this point that is a purely hypothetical idea because technology has gotten so good at recognizing samples, so let's assume I mean unrecognizable by human ears. In my experience, not a lot music gets targeted for sample violation unless the samples are fairly discernable and have an active role in the music, but perhaps that has changed.

Either way, AI images are not in fact made up of samples of other images so not sure it's relevant at all. My only point was, if the inputs are not discernable in the outputs, how is someone even going to go about proving harm?

Right, the fact that it's going to be impossible to prove is all the more reason that opt-outs need to be put in place now, imo.

(the sample issue is an aside, but micro-chops that you'd think would be unrecognizable to the human ear are often not, people are surprisingly good at that kind of pattern recognition even if it's been re-pitched. There are tons of Dilla and Daft Punk samples that have been identified that are just split seconds of sound, not to mention drum hits. Of course if something has been completely mangled with effects to sound totally different then it's probably impossible, but usually the point of sampling is because there's some sort of valuable quality in the source material you want to maintain?)

yes, microsamples are now recognizable, but I was indeed thinking of samples that had been mangled beyond all recognition, which absolutely happens all the time, granular synthesis being one prominent example.

I don't know how far litigation has been taken with stuff like microsamples, are there cases of musicians being successfully sued for barely discernable or entirely unrecognizable samples? My totally non-professional understanding of the laws around sample clearance led me to believe that the mere presence of an uncleared sample in a piece of music isn't necessarily enough to hold a musician liable. I'm under the impression that the length of the sample, how central it is to the piece of music, and the financial losses incurred by the original artist are all taken into account.

My understanding is that the whole "it's fine if it's under X seconds" thing is a myth.

This article talks about an NWA example: https://www.wipo.int/wipo_magazine/en/2009/06/article_0006.html

I don't know of a lawsuit example around true micro-samples, but I'm sure it could happen if the sampling track was a big enough hit, especially these days.

this is exactly why I avoid recording anything good enough to be a big hit

My outlook on copyright infringement litigation for music, especially samples in music, is that it has gone way overboard for decades, so I hope new technology does not mean it will start to ramp up further. I get the sense that people coming up today, possibly inspired by the AI discourse, have a much more welcoming attitude towards suing musicians for this stuff, but hopefully I'm wrong.

I think with actual AI data sets, lawsuits are going to be much more difficult for individuals since there isn't any way to use something generated through AI to identify what specific items were in the model and which ones helped determine the thing that was generated.