Source (Bluesky)
Transcript
recently my friend’s comics professor told her that it’s acceptable to use gen Al for script- writing but not for art, since a machine can’t generate meaningful artistic work. meanwhile, my sister’s screenwriting professor said that they can use gen Al for concept art and visualization, but that it won’t be able to generate a script that’s any good. and at my job, it seems like each department says that Al can be useful in every field except the one that they know best.
It’s only ever the jobs we’re unfamiliar with that we assume can be replaced with automation. The more attuned we are with certain processes, crafts, and occupations, the more we realize that gen Al will never be able to provide a suitable replacement. The case for its existence relies on our ignorance of the work and skill required to do everything we don’t.
This has also been coined the Gell-Mann [Amnesia] effect and is perhaps a kind of corollary to the Dunning-Kruger effect: incompetent people fail to recognize competence.
Truly intelligent people respect the work of professionals and experts in other fields. Or maybe, this is even fundamentally a respect problem.
This actually relates, in a weird but interesting way, to how people get broken out of conspiracy theories.
One very common theme that’s reported by people who get themselves out of a conspiracy theory is that their breaking point is when the conspiracy asserts a fact that they know - based on real expertise of their own - to be false. So, like, you get a flat-earther who is a photography expert and their breaking point is when a bunch of the evidence relies on things about photography that they know aren’t true. Or you get some MAGA person who hits their breaking point over the tariffs because they work in import/export and they actually know a bunch of stuff about how tariffs work.
Basically, whenever you’re trying to disabuse people of false notions, the best way to start is always the same; figure out what they know (in the sense of things that they actually have true, well founded, factual knowledge of) and work from there. People enjoy misinformation when it affirms their beliefs and builds up their ego. But when misinformation runs counter to their own expertise, they are forced to either accept that they are not actually an expert, or reject the misinformation, and generally they’ll reject the misinformation, because accepting they’re not an expert means giving up on a huge part of their identity and their self-esteem.
It’s also not always strictly necessary for the expertise to actually be well founded. This is why the Epstein files are such a huge danger to the Trump admin. A huge portion of MAGA spent the last decade basically becoming “experts” in “the evil pedophile conspiracy that has taken over the government”, and they cannot figure out how to reconcile their “expertise” with Trump and his admin constantly backpedalling on releasing the files. Basically they’ve got a tiny piece of the truth - there really is a conspiracy of powerful elite pedophiles out there, they’re just not hanging out in non-existent pizza parlour basements and dosing on adrenochrone - and they’ve built a massive fiction around that, but that piece of the truth is still enough to conflict with the false reality that Trump wants them to buy into.
You get a flat-earther who is a photography expert and their breaking point is when a bunch of the evidence relies on things about photography
Or you get a demolitions expert to watch a video of WTC7
I just focus on the parts of what I do know that AI can help me with, not try to say AI can replace other people, but not me. That’s some dumb shit.
AI only seems good when you don’t know enough about any given topic to notice that it is wrong 70% of the time.
This is concerning when CEOs and other people in charge seem to think it is good at everything, as this means they don’t know a god damn thing about fuck all.
I remember an article back in 2011 that predicted that we would be able to automate all middle and most upper management jobs by 2015. My immediate thought was, “Well these people must not do much, if a glorified script can replace them.”
Yeah, other than CFO and most* CTOs, anyone in the C-suite is easily replaceable by an LLM. Hell, the CEO could be replaced by a robot arm holding a magic 8-ball with no noticeable difference in performance.
* Probably not the majority, but I’ll be generous.
That’s the whole point of the bubble: convincing investors and CEOs that AI will replace all workers. You don’t need to convince the workers: they don’t make decisions and an awful lot of CEOs have such a high opinion of themselves that they assume any feedback from below is worthless.
Not just a high opinion of themselves, they think everyone is as self-centered as they are, and any claims about needing human workers for the task by human workers is just self-serving and not caring about the work.
The breadth of knowledge demonstrated by Al gives a false impression of its depth.
Generalists can be really good at getting stuff done. They can quickly identify the experts needed when it’s beyond thier scope. Unfortunately over confident generalists tend not to get the experts in to help.
This makes a lot of sense. A good lesson even outside the context of AI.
So the only real business model here is for people to be able to produce things they are not qualified to work on, with an acceptable risk of generating crap. I don’t see how that won’t be a multi-trillions dollars market.
Investors are rarely experts in the particular niches that the companies they hold shares in are applying AI to.
produce things they are not qualified to work on, with an acceptable risk of generating crap
You just described the C-suite at most major companies.
let’s not confuse LLMs, AI, and automation.
AI flies planes when the pilots are unconscious.
automation does menial repetitive tasks.
LLMs support fascism and destroy economies, ecologies, and societies.
My favourite ‘will one day be pub trivia’ snippet from this whole LLM mess, is that society had to create a new term for AI (AGI), because LLMs muddied the once accurate term.
accountants rage silently intensifies
Just wait until they create artificial people and autonomous robots.
No accounting acronym is safe from their tyranny.
I have AGI every year…
To be fair, AI was still underwhelming compared to what people imagined AI to be, it’s just that LLM essentially swore up and down that this is the AI they had been waiting for, and that moved the goalposts to have to classifiy ‘AGI’ specifically.
I’d even go a step further and say your last point is about generative LLMs, since text classification and sentiment analysis are also pretty benign.
It’s tricky because we’re having a social conversation about something that’s been mislabeled, and the label has been misused dozens of times as well.
It’s like trying to talk about knife safety when you only have the word “pointy”.
It’s like trying to talk about knife safety when you only have the word “pointy”.
holy shit yes! it’s almost like the corpos did it that way so they can just move the goalposts when the bubble pops.
I generally assume intent that’s more shallow if it’s just as explanatory. It’s the same reason home appliances occasionally get a burst of AI labeling. “Artificial intelligence” sounds better in advertising than “interpolated multi variable lookup table”.
It’s a type of simple AI (measure water filth from an initial rinse, dry weight, soaked weight, and post spin weight, then find the average for the settings from preprogrammed values.), but it’s still AI.Biggest reason I think it’s advertising instead of something more deliberate is because this has happened before. There’s some advance in the field, people think AI has allure again and so everything gets labeled that way. Eventually people realize it’s not the be all end all and decide that it’s not AI, it “just” a pile of math that helps you do something. Then it becomes ubiquitous and people think the notion of calling autocorrect AI is laughable.
It’s why managers fucking love GenAI.
My personal take is that GenAI is ok for personal entertainment and for things that are ultimately meaningless. Making wallpapers for your phone, maps for your RPG campaign, personal RP, that sort of thing.
‘I’ll just use it for meaningless stuff that nobody was going to get paid for either way’ is at the surface-level a reasonable attitude; personal songs generated for friends as in-jokes, artwork for home labels, birthday parties, and your examples… All fair because nobody was gonna pay for it anyway, so no harm to makers.
But I don’t personally use them for any of those things myself though, some of my reasons: I figure it’s just investor-subsidized CPU cycles burning power somewhere (environmental), and ultimately that use-case won’t be a business model that makes any money (propping the bubble), it dulls and avoids my own art-making skills which I think everyone should work on (personal development atrophy), building reliance on proprietary platforms… so I’d rather just not, and hopefully see the whole AI techbro bubble crash sooner than later.
I figure it’s just investor-subsidized CPU cycles burning power somewhere (environmental)
This can be avoided by using local open-weight models and open source technology, which is what I do.
Yeah, that certainly addresses that issue. I may do the same in the future, just haven’t found the need to do so as yet. For most who lean on AI for the simple tasks mentioned above, they use an AI service rather than a local model.
My personal take is that GenAI is ok for personal entertainment and for things that are ultimately meaningless
Were you previously much more pro-GenAI, or am I misremembering?
No that was always my position
Hot take: it’s reasonable for a comics student to use AI for script-writing and for a screenwriting student to use AI for concept art, not because machine can generate meaningful artistic work at these fields but because these are not the fields they are trying to learn.
In a way, this can be used to level the field. The comics professor can use the same LLM to generate scripts for all their students. It’ll be slop script, but the slop will be of uniform quality so no student will have the advantage of better writing and it’d be easier to judge their work based on the drawing alone.
And even if AI could generate true art in some field - why would it be acceptable for a student to use it for the very field they are studying and need to polish their own skills at?
Yeah, the comics professor is to grade the visuals, and the text is filler, could be lorem ipsum for all they care. Simlarly a screenwriter using AI to storyboard seems fine as it’s not the core product.
The ideal would be cross-discipline projects bringing students together similar to how they would be expected to deal in the real world, but when individual assignments call for ‘filler’ content to stand in for one of those other disciplines, I think I could accept LLM as a reasonable compromise. I would expect some assignments to ask the students to go beyond their core discipline for some perspective and LLM be bad for that, but I could see a place for skipping the irrelevant complementary pieces of a good chunk of assignments.
AI has been excellent at teaching me to program in new languages. It knows everything about all languages - except the ones I’m already familiar with. It’s terrible at those.
AI will never be able to provide a suitable replacement.
Don’t become as delusional as the grifters. Generated content is already a suitable replacement for lots of things. It’s not so much about the quality of generated content (which continues to improve) as much about easily replacing the worthless bullshit that we’re forced to produce (eg. cover letters, “art” serving capitalism, etc). The system is already built on fake nonsense. Generated content has always been a great fit for this system. The punishment of workers is just another bonus.
So Gen AI is like Dan Brown, the more you know about the subject the more it sucks
Or fiction in general…
Watch a doctor’s opinion on most medical shows…
A computer expert’s opinion on most ‘hacking scenes’…
A lawyer on any legal drama…
IDK about that I’m a professional slop maker and I think it could replace me easily.
The only good AI I’ve come across is the one I use for denoising cycles renders in Blender3D, as that’s something that a human cannot reasonably do.
That’s the only scenario something like AI has any use as a “tool”; doing things humans cannot reasonably do.
Good news. There’s absolutely zero “intelligence” involved in computer functions doing math. No “AI” needed or detected as usual.
Yes, I know. It’s just a neural net. Still a calculator, just with more steps than a human can sift through. And they call that intelligence.
Unexpected blender tip. Thanks for that, hope it improves my render times 😉
Here’s another for you:
Play around with sample sizes and render tile sizes in the performance menu (same place where you find the denoising options).
Depending on your set-up, you can see a drastic improvement in render times by choosing smaller tile sizes. Sample size is also counted per render tile, so you could get away with very low sample sizes and have a completed render with an overall higher combined sample size.
Did not know that about the tile size. I never worried to much about performance as my previous laptop had plenty of grunt. But that blew up and now I’m on a 10 year old machine that was tired when it was new. I need everything I can get 😆
Which explains why C-suites push it so hard for everyone
Well, they do have the one job that actually can be replaced by “AI” (though in most cases it’d be more beneficial to just eliminate it altogether).
Which is acting like they know everything about everyone else’s jobs, while making up wholly inaccurate assumptions
