"there exists a rigorous mathematical description of the commonalities between every Seinfeld episode sufficient to reconstruct new episodes recognizable as instances of the genre" is an indisputable and bizarre fact at this point, even if the episodes reconstructed in this way are bad. - ThreadSky

theophite.bsky.social • 2 days ago

"there exists a rigorous mathematical description of the commonalities between every Seinfeld episode sufficient to reconstruct new episodes recognizable as instances of the genre" is an indisputable and bizarre fact at this point, even if the episodes reconstructed in this way are bad.

Reposted from alasdair

this is where ive been. it's insane that when chatgpt generates you a sitcom dialogue it can only do it bc its extracted a principal component vector for the concept "sitcom", and done so from what words appear next to each other. several insane implications there even if the result is a bad sitcom!

Comments

aaronrosspowell.com•2 days ago

I do think one driver of the anger about generative AI is that it demonstrates there less pure novelty and underivative creativity in supposedly creative acts by humans than a lot of people are comfortable admitting.

bwyble.bsky.social•2 days ago

I don't know about that. I think it's always been obvious that the majority of human art are minor iterations of earlier work. Only a fraction of it pushes into new spaces. AI does the former but it's not clear it does the latter.

aaronrosspowell.com•2 days ago

All of that is true. I think the anger is it makes it more obvious, because it says, "A computer using math can produce new works as novel as what many humans (not all!) produce." Most people like to think their creations are sui generis acts of pure novelty. This makes that belief more challenging.

bwyble.bsky.social•2 days ago

I guess it's true that most people LIKE to think their creations are sui generis, but I think most people who are used to creative tasks are well aware of the repetitive nature of their work.

This highlights a kind of cognitive dissonance people have about creativity.

tim.dobbs.world•2 days ago

Personally, the idea of an underlying principal component vector describing genre was always my mental model, and I think it's (very!) cool that we have found math to point at it.

bwyble.bsky.social•2 days ago

I don't think LLMs are really "finding the math". That implies a level of understanding we don't have for how they work once trained.

tim.dobbs.world•2 days ago

But I'm "mad at AI" anyway because I think the discourse supposes all that is important in art is creative novelty, whereas I find art most impactful when it is contextualized in some experience of life or culture. Seinfeld rules because it was in discussion with the quirks of 90s yuppie life.

mchirdon.bsky.social•2 days ago

Do people think that though? People writing in genres like mystery, romance, or even sit com, seem to understand what they are creating is part of a body of work with certain expectations it must fulfill in order to be satisfying to its audience.

paulmather007.bsky.social•1 day ago

I write for tv. I’m not sure many sitcom writers think much about novelty as such? Also the audience is always changing: the exact same dialogue or situation will get a different reaction now vs. 5 years ago vs. 50 years ago. The meaning in the work isn’t entirely in the text…

paulmather007.bsky.social•1 day ago

…it’s also the audience. The thing I find upsetting about the idea of AI generating scripts (aside from losing my livelihood) is that it’s not trying to communicate. Even in the dumbest sitcom, someone is trying to say something to do with their lived experience. AI is just patterns.

notthegodofhammers.bsky.social•1 day ago

I'm not at all sure of this take, but I suspect that this understanding is held by self-aware GENRE writers/fans, but maybe not as much by those who neither write nor particularly care for genre works...or who simply lack self-awareness.

notthegodofhammers.bsky.social•1 day ago

You often see self-deprecating humour from good genre writers - like Pratchett's response to a knighthood for his service to literature, saying that he supposed his service to literature was that he never attempted to write any.

iow, I think you have to be far enough down the rabbit hole to know.

dsic.bsky.social•2 days ago

I've seen no evidence it can, show me one good work of AI art.

aaronrosspowell.com•2 days ago

Nothing above is about the "goodness" of the results, but instead about novelty.

dsic.bsky.social•2 days ago

What is the difference. Good works are novel.

phillipcarter.dev•2 days ago

I see it differently (ignoring the bit about training on material without consent):

ChatGPT lets people create slop that is usually bad quality, and it's acceptable in many respects to just use that slop instead of paying an creative professional to make what you want.

blank.org•1 day ago

For better or worse I think in many (most?) cases the alternative is not “paying a professional to produce original and presumably better output” but “searching the internet for five seconds and copy-pasting whatever seems most likely.”

blank.org•1 day ago

I don’t want to underrate how much this is a bad deal for professional graphic artists: ai may well be taking most of their work! But I really doubt most of the output of genAI represents a lost commission.

phillipcarter.dev•2 days ago

In that sense, it deprives people with particular skills from labor that they might otherwise have. And not because the slop is better, but because it's normalized to the point where people are fine with worse quality.

phillipcarter.dev•2 days ago

I say this as a "fan" of this tech and user of it for many years now. This is one of the main negative effects of adopting this technology, much like how personal computers put others out of business decades ago. It is a nasty reality and folks have a right to be mad about that.

lcsdcrz.bsky.social•2 days ago

That doesn't sound right to me, honest. The majority of people like to see what they do as belonging to a tradition.

lcsdcrz.bsky.social•2 days ago

I agree with that. We put Faulkner and Cervantes under the same umbrella of novelists. Faulkner's novels open new spaces, but those spaces exist only because we can still recognize what he is doing as writing a novel.

typhonbaalhammon.bsky.social•2 days ago

The underlying assumption is decades old. Distributional semantics dates back to John Firth in the 1950s. I think the fact that it's definitely not enough is rather more striking and interesting at this point !

acausalrobotgod.bsky.social•2 days ago

I'm definitely most intrigued that the extracted similarities are close enough to whatever the human mind does to recognize them. This similarity, of course, is what ultimately allows me to simulate consciousnesses that the gullible equate with themselves...

roland00.bsky.social•2 days ago

what if I can tell you there is genre

where I can show you one 2d image, and then a second 2d image and you can follow how to get from A to B

and then I show you a third 2d image, and the excitement is trying to get from A to C without knowing what happened in the mystery box B, so says Poetics

roland00.bsky.social•2 days ago

a) Because Providence doesn't want futile things glorified.
b) Oh, that I must contest...

https://www.youtube.com/watch?v=W33uQVXSu3I

conjurial.bsky.social•2 days ago

it's also very strange that this is a corollary of the Church-Turing thesis

conjurial.bsky.social•2 days ago

Premise A: computability
Premise B: physicalism

---------
∴ Seinfeld admits a formal representation

louisevans.bsky.social•1 day ago

Having ptsd flashbacks about that three dot therefore

pretzelkins.bsky.social•1 day ago

Speaking of, what ever happened to watch me forever

He said, not googling it

wsu-a.bsky.social•1 day ago

It continues to exist. I still get "thing I guess I tapped accidentally" notifications. I check on it every few months.

Whatever bumpers they put on content/depictions after their issue made it worse in a much more inane, disconnected way.

Nobody watches it. There's like 2 chat messages a day.

pretzelkins.bsky.social•1 day ago

What a bummer

milkglider.bsky.social•2 days ago

It exists? Link to the PDF?

theophite.bsky.social•2 days ago

tl;dr: you can use a sparse autoencoder to extract directions in the vector space corresponding to any cognizable concept and then upregulate the resulting activations to create a bias towards that concept, which strongly implies the 'seinfeld vector' hypothesis.
https://www.anthropic.com/news/golden-gate-claude

overeasysteezy.bsky.social•2 days ago

I can jargon some jargon too man

theophite.bsky.social•2 days ago

"using a statistical technique which there is no intuitive explanation for, you can find the parts of a neural network responsible for a particular concept and turn up the volume on those parts in order to make it talk about the that thing without involving the prompt."

overeasysteezy.bsky.social•2 days ago

O I get it- I just detest jargon. I see no use for this obvious bug, aside from making it bring up the Black Friday Sale on Amazon dot com, when you're asking it if there's any new medical research that can help your terminally ill dog.

theophite.bsky.social•2 days ago

a good example of what it is useful for is, e.g., examining the activation space in order to determine whether the model is doing something you object to without directly viewing either the input or output. you can then dynamically modify the activations to make it stop.

spinning-belinda.bsky.social•2 days ago

does this vector space have a (natural) inner product that allows us to identify a “seinfeld” vector with a (continuous and linear) “seinfeld-ness” function? (apologies if this question isn’t sensible—I know a lot about vector spaces and very little about machine learning)

theophite.bsky.social•2 days ago

i think my barrier here is the fact that we are projecting a whole bunch of stuff into the much lower-dimensional space of a sparse autoencoder and although i think this is a sensible question, i can't work through what the answer is in my head given the projection.

spinning-belinda.bsky.social•2 days ago

actually, what is the vector space in question? are the elements, like, formal linear combinations of words?

theophite.bsky.social•2 days ago

yes.

inv1nciblesummer.bsky.social•2 days ago

To my understanding the “grandmother neuron” theory of cognition has been shown to be without basis but it’s wild that the “Seinfeld vector” exists instead. IMO it seems intuitively much more probably that concepts have simple representations in the brain than in high-dimensional geometry, and yet

borogove.bsky.social•1 day ago

You need the extra dimensions to holographically, efficiently encode all the other comedians alongside.

markshatraw.bsky.social•2 days ago

So there's a Category of Seinfeld episodes?

littlegino.bsky.social•2 days ago

Plato and Aristotle: always relevant

markshatraw.bsky.social•2 days ago

So Platonism is just Category Theory?

littlegino.bsky.social•2 days ago

Not saying that, not a programmer or mathematician. Just riffing. I think it's interesting to consider in the context of plato/aristotle how we map out an abstract idea onto a mathematical object that happens to be the representation of all things of a type (when reproduced digitally).

orpach.bsky.social•1 day ago

Hearing an explanation of how faceapp worked a few years back was strange as hell.

Like you take an image, subtract the "average man" vector, and add the "average woman" vector and there you go. Gender-swapped image.

Worked on race too, but they took that out real quick after it made news.

opulenceihazit.bsky.social•1 day ago

Ironically, I remember an interview with Jason Alexander where he explains how Seinfeld broke all the rules of how sitcoms were structured at the time

technohippie.bsky.social•2 days ago

An interesting corollary: very few of the things we actually want out of a Seinfeld episode are present in that description. If it were otherwise, the reconstructions wouldn't suck.

Comments

Posting Rules

Reply