criterionforum.org

andyli wrote: Thu Jul 17, 2025 2:00 pm If AI can make people hating it start posting more sarcastically on Internet I’m smelling a workable strategy of sabotage. Indeed I’ve seen people deliberately create ridiculously false information that is easy for real people to spot just to ‘contaminate’ the model. I guess if enough people were doing this AI would break like a cumbersome battleship.

It seems optimistic to think that this would "break" AI. It seems far more likely that it would just propagate the false information, no matter how ridiculous it is. We've seen that people will believe pretty much anything they read on the internet, and how are you going to top a theory that the weather is controlled by Jews?

If the energy exists to work this hard to "contaminate" the model with false info, it seems like it would do far more good to spend that energy trying to feed it real, actually true info.

To add to the pile-on: I also ran an experiment with ChatGPT the other week, asking it to list all the books that contain Krazy Kat daily strips, in chronological order with ISBNs - a subject that's semi-well-documented but also a little tricky because the strips are scattered among various publishers, limited pressings, etc. Not only did it miss half the existing publications and add ones that are exclusively Sunday strips, but it listed as one of its sources a site that I know contains all the info. Even when it's taking an open-book test, it fails, because it simply has no way of verifying what it spits out. At least with Google, there's the expectation that you may have to dig through a few sites to collect knowledge before coming to a conclusion, but people think ChatGPT is doing that work for then, and it just isn't.

Worse yet, a friend of mine shared with me texts from a (former) friend of his, in which she uses ChatGPT to "diagnose" her fiancé and his mother as gaslighting, narcissistic, abusive, etc. I'm sure this says more about the person using the tech than anything else, but it's genuinely startling stuff.

andyli wrote: Thu Jul 17, 2025 2:00 pm If AI can make people hating it start posting more sarcastically on Internet I’m smelling a workable strategy of sabotage. Indeed I’ve seen people deliberately create ridiculously false information that is easy for real people to spot just to ‘contaminate’ the model. I guess if enough people were doing this AI would break like a cumbersome battleship.

My doomsday scenario is that AI writing just gets shittier and shittier as more and more of its inputs were AI (or AI-tainted) in the first place. A kind of Incest Apocalypse.

But I think the AI companies are already concerned about this. I have a friend whose current job is to research and write short articles on esoteric subjects for feeding into AI language models, which she finds existentially weird but enjoyable.

I have clients who use ChatGPT as a complementary tool to therapy, to validate their emotions as a 'friend'/coping mechanism. I don't like AI in general, but I'm not quite sure how I feel about that one.

With generosity of spirit, I asked it to surmise my favorite films based on my posts on this forum. It doesn’t even have to connect dots here, it could have just grabbed any movie I called one of the best films of the year or ranked at the top in a list project. It returned a top five that impressively features a movie I have never even seen (Lynch’s the Straight Story), three PTA films (but not the one I literally named as his best film), and A Kiss Before Dying. Great models we got here folks

ChatGPT doesn't do a great job in identifying either my favorite movies or my favorite anime. It does, however, say " Michael Kerpan, an anime enthusiast and a well-known figure in the anime community, is famous for his deep knowledge and appreciation of both classic and modern anime". And also "ChatGPT said, "Michael Kerpan, the American film scholar, critic, and curator, was known for his deep love of classic cinema, particularly his appreciation for films from the silent and early sound eras, as well as Asian cinema". I find that "was" rather ominous -- but otherwise this is pretty hilarious stuff.

zedz wrote: Thu Jul 17, 2025 7:54 pm

andyli wrote: Thu Jul 17, 2025 2:00 pm If AI can make people hating it start posting more sarcastically on Internet I’m smelling a workable strategy of sabotage. Indeed I’ve seen people deliberately create ridiculously false information that is easy for real people to spot just to ‘contaminate’ the model. I guess if enough people were doing this AI would break like a cumbersome battleship.

My doomsday scenario is that AI writing just gets shittier and shittier as more and more of its inputs were AI (or AI-tainted) in the first place. A kind of Incest Apocalypse.

But I think the AI companies are already concerned about this. I have a friend whose current job is to research and write short articles on esoteric subjects for feeding into AI language models, which she finds existentially weird but enjoyable.

This incest is actually well documented and is nicknamed AI hallucinations which are caused by how the xeroxed pool of information because too confusing and overwhelming for the models. It takes about three generations of incest to kill off an AI.

domino harvey wrote: Thu Jul 17, 2025 7:57 pm With generosity of spirit, I asked it to surmise my favorite films based on my posts on this forum. It doesn’t even have to connect dots here, it could have just grabbed any movie I called one of the best films of the year or ranked at the top in a list project. It returned a top five that impressively features a movie I have never even seen (Lynch’s the Straight Story), three PTA films (but not the one I literally named as his best film), and A Kiss Before Dying. Great models we got here folks

Was it able to hit the site? They usually get lumped in with bots.

It has it's uses for what I do (and it's made some things a lot easier), but when it shits the bed, it shits the bed. Doing experiments with it with team members, we tasked an agent to fix a bug. It went through everything it needed to and then generated an image indicating that it had successfully fixed the bug (with a lot of exclamation points), and then listed the pros and cons of the bug being fixed, even though we didn't ask it to do that. That was kind of funny (especially since the image was heavily compressed), but then it turns out it didn't even fix the bug, and made changes that broke something else. It did other tasks we assigned it without much issue, but when it misses the mark, it really misses it. Still better than one junior dev I worked with, though, who once brought down a production site after deleting one line in a config file without mentioning it to anybody.

cdnchris wrote: Fri Jul 18, 2025 12:36 am

domino harvey wrote: Thu Jul 17, 2025 7:57 pm With generosity of spirit, I asked it to surmise my favorite films based on my posts on this forum. It doesn’t even have to connect dots here, it could have just grabbed any movie I called one of the best films of the year or ranked at the top in a list project. It returned a top five that impressively features a movie I have never even seen (Lynch’s the Straight Story), three PTA films (but not the one I literally named as his best film), and A Kiss Before Dying. Great models we got here folks

Was it able to hit the site? They usually get lumped in with bots.

Yeah but given the paucity of returns, maybe it was blocked after a certain number of pulls? It would explain how poorly it did at something which should be easier for it to do

Correction, I think it's saying A24 sought out investment from the firm that backs OpenAI, not OpenAI itself. But the point was to make connections in that field, so not really much better.

Got to love that one of the two examples the A24 Labs head came up with for future AI usage is "a [filmmaker having a] deep conversation with an LLM to debate a character’s mindset." Oh, like a conversation that human writers can have together... right now? And the other example is pre-visualization, code for stealing art because paying artists or finding collaborators is too difficult.

Boosmahn wrote: Tue Aug 26, 2025 7:01 amAnd the other example is pre-visualization, code for stealing art because paying artists or finding collaborators is too difficult.

That's an incredibly cynical interpretation of what's potentially a very useful tool in a filmmaker's armoury.

Would you argue that I'm "stealing money" from audio transcribers every time I use AI-powered transcription services, or that I use such a thing because it's "too difficult" to "find collaborators"? Even though my budget would never stretch to hiring such people in the first place and so it's merely a massive timesaver for me personally?

You using AI to transcribe isn’t utilizing the service to create art that a human could/should do, though, which is their objection

To my mind, pre-visualisation has more in common with transcription than otherwise, in that it's the filmmaker's concepts being transcribed. And to someone who can't draw very well, I can see it being a huge help in contexts where a human artist most likely wouldn't have been hired anyway.

(My son, a VFX professional, can't draw to save his life, and he's been finding the new generation of pre-visualisation tools incredibly helpful at getting his ideas across in a collaborative setting.)

As someone who works mostly as an audio transcriber, I would be worried if everyone turned to A.I. for it (particularly my bosses, who seem to be gleeful in being excited about the potential savings being offered by such advances!). But, also as someone who works as an audio transcriber, I have had to listen to dictations from people where a lot of intuition is involved in understanding what they meant to say, but which has come out garbled, or abbreviated in verbal shorthand (that's doctors for you, where they keep doing the verbal equivalent of terrible handwriting!), malapropisms, strange vocal patterns, weird pauses that could be the end of a sentence... but no!, etc, etc, that I wonder if even A.I. would be able to handle fully as yet. I guess it will come eventually, but I presume there at least needs to be a human being doing a final finessing pass over the end result to catch any quirks or misinterpretations that the A.I. does and produce a professional-looking result, rather than just a passable one. Which I guess my job may eventually end up turning into.

It sounds similar to what The Brutalist did which struck me as reasonable enough.

Edit: Colin I think that’s why AI will never supplement people but instead by like ATMs where they take out the busywork of it.

Let's just hope I don't go the way of Bank Tellers then. Or physical Banks themselves!

(My favourite audio transcriber story is back when I worked as a Legal Secretary, where one of the Solicitors instead of speaking directly into the microphone where they would have come through crisp and clear, used to put it on the recorder equivalent of speakerphone and then wander around her (extremely large and echoey) office. So I dreaded having to do her letters because you could tell that she was wandering away as her voice got fainter... and fainter... and fainter..., until I was straining to hear anything through my headphones! Once there was even the sound of the office door closing and I had to sit through ten minutes of silence until she wandered back into the office and picked up her dictation from where she left off! But I had to sit through the dead air just in case anything important occurred!)

I'm referring to generative AI, when artists are eliminated from steps of the creative process. I don't have any issues with AI being used for transcription.

And while I was mostly thinking of projects with sizable budgets and good resources (like an A24 film) when I posted my reply, I'll clarify that I don't like generative AI being used in more modestly sized projects, either. I also can't draw well -- even my stick figures are iffy -- but I won't use GenAI to express my ideas, both because of the models' stolen data and the feeling that I'm cheating myself. If I ask AI to visualize a scene for me, it's no longer my scene. It's using my words and trying to approximate what they could look like, but not thinking, feeling, and using an imagination like a human would.

colinr0380 wrote: Tue Aug 26, 2025 3:39 pm As someone who works mostly as an audio transcriber, I would be worried if everyone turned to A.I. for it (particularly my bosses, who seem to be gleeful in being excited about the potential savings being offered by such advances!). But, also as someone who works as an audio transcriber, I have had to listen to dictations from people where a lot of intuition is involved in understanding what they meant to say, but which has come out garbled, or abbreviated in verbal shorthand (that's doctors for you, where they keep doing the verbal equivalent of terrible handwriting!), malapropisms, strange vocal patterns, weird pauses that could be the end of a sentence... but no!, etc, etc, that I wonder if even A.I. would be able to handle fully as yet. I guess it will come eventually, but I presume there at least needs to be a human being doing a final finessing pass over the end result to catch any quirks or misinterpretations that the A.I. does and produce a professional-looking result, rather than just a passable one. Which I guess my job may eventually end up turning into.

I use AI transcription extensively behind the scenes, and while the results can sometimes be startlingly impressive I would never sanction publication without going over it syllable by syllable, which might well take longer than manually transcribing it in the first place. Ditto subtitles; mine are invariably the old-fashioned hand-crafted type, for the simple reason that fixing AI quirks - particularly timing, as line breaks are invariably not selected with the end reader's viewing comfort in mind - also takes longer.

But I've come across plenty of howlers made by humans - sometimes reasonably, for instance the transcriber who consistently misheard Karel Reisz as "Carol Rice". But why not? It's what it sounded like, and it's not their job to know this upfront; it's my job to spot it when it happens!

(Tangentially, this is why I always hope I get Basia Howard as interpreter whenever I'm hosting Q&As with non-English-speaking Poles - she's not just an uncannily good on-the-fly translator but she also knows the official English titles of the films being translated, which is the kind of specialist knowledge that doesn't automatically come with the interpreting package.)

Boosmahn wrote: Tue Aug 26, 2025 4:00 pm I'm referring to generative AI, when artists are eliminated from steps of the creative process. I don't have any issues with AI being used for transcription.

And while I was mostly thinking of projects with sizable budgets and good resources (like an A24 film) when I posted my reply, I'll clarify that I don't like generative AI being used in more modestly sized projects, either. I also can't draw well -- even my stick figures are iffy -- but I won't use GenAI to express my ideas, both because of the models' stolen data and the feeling that I'm cheating myself. If I ask AI to visualize a scene for me, it's no longer my scene. It's using my words and trying to approximate what they could look like, but not thinking, feeling, and using an imagination like a human would.

Agreed 100% and I think there is a real moral component here, too, in that art is of value because there is an artist creating it. Humans not giving our creativity over to a computer is the bare minimum we can do as a cultured species and yet so many people, even those who are creatives with a body of work they brought into fruition without asking an LLM for the answers, seem willing to skip multiple steps in the creative process in the name of expediency and laziness. You can’t collaborate with AI, it’s not a colleague

That to me is the difference between being simpatico and solipsistic - you get someone with their own knowledge of the subject area working with you and they intuitively know what you are talking about to catch any errors even in obscure subjects (I used to get a bit of stick from more experienced colleagues when I first moved from legal to medical secretarial jobs, where being thrown in at the deep end I kept making mistakes until I put myself on an extensive personal course of letting the specific medical terminology sink in to stop that from happening. And even now there are a couple of specialities that I do not do too often, and only as holiday/emergency cover when it is me or nobody where I am back to doing a competent but not 100% job: Gynaecology in particular!) and that makes the process run much easier for both the dictator and the admin person. Whereas, as discussed earlier, A.I. (or pre viz if you don't have an artist to filter ideas through) is just someone talking to themselves, so if the person in charge of dictating says something wrong, then its still going to be treated as gospel because the A.I. primarily has to follow the user and adapt to any illogicalities. That could end up becoming much more solipsistic.

colinr0380 wrote: Tue Aug 26, 2025 3:39 pm As someone who works mostly as an audio transcriber, I would be worried if everyone turned to A.I. for it (particularly my bosses, who seem to be gleeful in being excited about the potential savings being offered by such advances!). But, also as someone who works as an audio transcriber, I have had to listen to dictations from people where a lot of intuition is involved in understanding what they meant to say, but which has come out garbled, or abbreviated in verbal shorthand (that's doctors for you, where they keep doing the verbal equivalent of terrible handwriting!), malapropisms, strange vocal patterns, weird pauses that could be the end of a sentence... but no!, etc, etc, that I wonder if even A.I. would be able to handle fully as yet. I guess it will come eventually, but I presume there at least needs to be a human being doing a final finessing pass over the end result to catch any quirks or misinterpretations that the A.I. does and produce a professional-looking result, rather than just a passable one. Which I guess my job may eventually end up turning into.

I work as a freelance translator of (mostly) art historical texts and I can see myself lasting maybe another two or three years tops. I only started doing this three years ago and even in that time AI translation has advanced dramatically. Work has dropped off this year, and clients have started haggling over fees. At least I'm only five years from retirement. Many of my friends who work in animation (my previous career) are also experiencing a sharp downturn.

That is a shame to hear, because a good translation can be as much about a sensitive translator (to both the original context, and the one which is being interpreted into) as the text itself. I barely know English well enough as the one language I speak(!), so appreciate the efforts that any translator does. It does make me wonder if with A.I. we would ever get things like those 'duelling' subtitle tracks on Criterion's Throne of Blood disc; or Anthony Burgess's translation of Cyrano de Bergerac into rhyming couplets that mimiced in the English language the effect of the original French dialogue? Although they were always outlier projects, even without A.I., I suppose. I guess that kind of result could be achieved if someone put in the specific parameters into the A.I. to make it do something like that, but then in a world with A.I. translation would anyone even consider doing that when the machine translation is already there? (Though I guess that gets into the same kind of territory as monks complaining about their painstakingly illustrated manuscripts, full of care and attention to detail, being superceded by mass printed Gutenberg Bibles! Pluses and minues, I suppose!)

The only thing that saves me right now is that I work in Germany, which, of the countries I have lived and worked in, is by far the most technophobic. The fax machine is still going strong here, many restaurants and most bars only accept cash, and our bureaucracy is only just starting to move into the digital age. This means I have clients who don't even know DeepL and ChatGPT exist.

The Curious Sofa wrote: Tue Aug 26, 2025 4:41 pmI work as a freelance translator of (mostly) art historical texts and I can see myself lasting maybe another two or three years tops. I only started doing this three years ago and even in that time AI translation has advanced dramatically. Work has dropped off this year, and clients have started haggling over fees. At least I'm only five years from retirement. Many of my friends who work in animation (my previous career) are also experiencing a sharp downturn.

This tangentially reminds me of when I commissioned translations of half a dozen Pirandello short stories for the book that accompanied Arrow's Taviani Brothers box.

I took advantage of the fact that Luigi Pirandello himself (d. 1936) was unarguably in the public domain, but I knew that I wanted a decent translation by someone experienced in handling literary texts, so I posted an ad for what I was after on a forum for translators. Thankfully, I quickly found someone who turned out to be absolutely ideal, but I had some startling pitches from people who were offering to charge only a tenth of what I'd established was a reasonable going rate. God knows what they would have produced - at those rock-bottom prices I suspect they'd have run the text through 2016-era Google Translate, something I was more than capable of doing myself for nothing - but I wasn't tempted to find out!

But there's no way I'd have entrusted a job like that to AI, certainly not back then and not today either, because proper literary translation demands a sensitivity to words and their musicality and potential multiple meanings that you're simply not going to get out of a computer.