godspeed you! big animal

broadly agree, and the point about how many things we all broadly like used to be called "AI" is important. but two things:

• if your main area of differentiation between today's "AI" and yesterday's "AI" is that right now data is being taken without express permission, that is not a very rock-hard defense. adobe et al (and some music companies, I very strongly suspect) are training models with only data they already own the copyright to. should we accept use of these models freely because nobody's work was "stolen"? well, no. the labor issues with this potential model are exactly the same (they're gonna try to replace people with these) and they're still going to output slop. I don't think this is a useful criteria for what we should accept and what we should reject.

• you do not want to be on the same side as massive copyright holders. you absolutely do not want copyright to be strengthened to the point where the law alone would actually stop generative AI from being feasible to train. i do not think it is just an "uncomfortable tool in the fight against AI;" it is handing enormous slop-generating companies a cudgel and giving them the opportunity to look like saints for brandishing it. i have spent too much time researching modern music copyright cases to feel any other way about this

fun tidbit re: Deep Blue/Kasparov, i remember reading that actually the reason why it was able to beat him was that it made an error. it made a chess move that didn't make any sense, and Kasparov overthought it, and made his own error, and the machine took the game. Literally won by mind games.

At this point, i think chess is a solved game at the expense of ~equivalently the power consumption of Ireland, so it doesn't really matter for your point. doubly so because focusing on "the AI wasnt actually as smart as they say" detracts from the issue. but i thought it was funny

Destroy Google. Okay, I don't have a specific plan for how to do this, but it would sure help.

Tariffs and Protectionism. The PRC and the ROK already do it.
Google (and Microsoft) are foreign entities towards which the Central Government already is openly hostile, and it needs to grow the balls to build the public opinion to allow it to break up their IT sector.
The State already fucks with them and puts down other major IT companies, so we know it can do it.

The thing that's been bouncing around my mind is that, if you take such a hard-line position that you want to "ban AI" without being able to say what you mean by that, you are not going to be effective enough at opposing it.

I'll go even stronger than this: if you take a hard-line position and want to "ban AI" without defining why it's bad, the opposition will choose things it can knock out as low hanging fruit and pretend they were your reasons.

"look, we got rid of the copyright infringement (now that our model is complete and we have enough money to legitimately pay. lowballing of course )"

"look, it no longer makes porn of your crush (because we banned all '18+' content"

etc

i feel very strongly that IP law should not be considered an "uncomfortable ally", it should be considered a poisoned well. strengthening the power of copyright would not help independent artists or to prevent AI slop from being generated, it would just center all of the power of that on big media corporations who can train their own models and/or copyright their own artstyle.

imagine a world where nintendo can sue every pokémon fanartist ever for "style infringement" on ken sugimori's art. the proposals to try and strengthen copyright by making "art style" a component of IP would do that and it would be devastating.

ai slop isn't great, but strengthening copyright to oppose it would be so, so much worse.

I think it's worth noting that the murkiness around what counts as 'AI' is somewhat deliberate on the part of those marketing tech products. Not unlike how 'blockchain' was thrown around in completely irrelevant contexts during the crypto bubble.

Heavy sigh, yeah re Google. You can use alternatives for some things, some are easier than others, and some are trading the evilglobocorp1 for evilglobocorp2. My favorite alternative is @fastmail, a seriously fantastic paid email service with better filtering and rules than gmail by a thousand goddamn miles. And it is FAST. So much good to say about fastmail.

The common joke about "AI" when I was a CS undergrad years ago was "if it works reliably it's called an algorithm; if it doesn't, it's called artificial intelligence". The sentiment remains true, but mainstream usage has poisoned the word "algorithm" and more or less inverted its meaning from "a precise step-by-step procedure" into "an opaque and unknowable machine-learning model", so it's not as rhetorically sharp as it once was.

For my own writing, I prefer to use terms like "LLMs" and "GANNs" over "generative AI", because it nails the discussion down to real-world techniques with tangible faults and failure modes (i.e. being completely fucking useless for most purposes) rather than the hypothetical future technologies that enthusiasts like to pitch. It's like public transit via flying car versus trains: the trains have imperfections but exist, and the cars appear perfect (to gullible marks) because they don't exist.

Personally, my own distrust and dislike of AI extends further than the latest wave of generative models. I think the recent developments ought to recolor our opinions on earlier AI tech, even if it seems mundane by now. If I had to try to pin down where to draw the line, it would roughly be at "AI is bad if it's a black-box model that has been trained on a data set" (yes, I'm aware that this casts a VERY wide net).

There may be no putting that genie back in the bottle, but maybe we can get some guardrails. One I'd suggest--that sadly has no chance in hell of coming true even though it feels like common sense to me--is "AI companies are required to publish their training data". A more modest version that miiiight gain acceptance is "AI companies are required to publish their holdout data" (I don't know how to apply that rule to generative models though, just classifiers). Like, can we get the barest amount of accountability -- can a third party audit whether or not your model even does what you claim it does?

I hope when you say that stuff about "art being collaged" etc., you're being metaphorical, but I would encourage not using that kind of framing. I've seen people come to the mistaken conclusion that image generation literally works by storing images, digitally cutting them up, and pasting them together, such that the original image could be plucked out of the source code like an asset in a video game.

While I agree that the provenance of datasets is important (e.g. the LAION 5B dataset includes nonconsensual sensitive medical records, which I think on its own is reason not to use it), I believe misrepresenting the function of the technology undermines one's point.