You are viewing a single comment's thread. View all
3
alele-opathic on scored.co
1 month ago3 points(+0/-0/+3Score on mirror)2 children
> even ignoring the part where good actors still do a better job than the greatest AI
**Give it 6 months - 1 year. People said the same thing about still artists too and now modern comfyui workflows beat the top digital artists.**
> which can, by definition of how the technology works, only copy and never create
**This is reddit-tier ludditism. Its generative by definition, and, when trained on everything we've ever created, can create anything in the scope of our creation, even things never created before (which it does all the fucking time if you pay attention).**
> you're just running some GPUs to make some memes
**We exist in a flood-tide of mental filth that was created by jews doing just this same thing. You clearly weren't around for the meme wars of 2015-2017. Someone on our side has to do it, and you're retarded if you actually are trying to discourage our own from fighting back in the age of 4th generation warfare.**
I dont disagree with your end point. I disagree with your assumption that it will be better in such a short time. 90% of the engineering takes 10% of the time. Thats where we are. At about 85% right now. What we will likely see is everything gets worse because AI can do 90% of the work and fail at the last 10% and no one will care about the last 10% because it will be cheap and good enough. Everything is going to get worse from AI due to this
1 month ago1 point(+0/-0/+1Score on mirror)1 child
> I disagree with your assumption that it will be better in such a short time.
It's tough to predict the pace of outcomes, especially given how everyone who's tried in the last 3 years has seriously underestimated the actual real timeline of LLM or vision model development. That said, I've personally studied the development of LLMs back when they were just government projects, including such 'mostly hidden' secrets such as:
- Twitter, which was purely infinite facebook statuses, was created as a playground for very early LLMs, called 'Influence Bots' in DARPA papers of the time
- These bots were proven to be operating en-masse on Reddit as early as 2017
- Every time the civilian-side development of language models stagnated, google released some 'breakthrough' paper/model that relaunched civilian interest and development - Word2Vec ('king' - 'man' + 'woman' = ? analogies, etc) in 2013, the infamous 'Attention is all you need' paper (which kicked off this current wave), the Gemma3 models from earlier this year (which taught us that our current models were extremely inefficient - these models were so much better than they had any right to be and now researchers scrambled to figure out 'why').
To gloss over what will be a long story, **the reason why everything is proceeding so quickly is because it isn't the pace of 'research', but rather the pace of 'leaking' of already discovered information that is driving this AI wave.**
> Everything is going to get worse from AI due to this
I'd argue it already was. I could make a (relatively weak, evidence-wise) argument that, for a long time, pop music has been made by machine models. I estimate it's been this way for over a decade now.
1 month ago-2 points(+0/-0/-2Score on mirror)1 child
> modern comfyui workflows beat the top digital artists.
as someone who uses the modern comfyui workflows to make covers for his singles, an artist remains indispensable when you have an extremely specific artistic vision, like for the video game i'm developing with two artists.
i mean, i'll freely admit to having used AI just to mass-generate character design concepts, but for every 50 generations you get one good design, and for every 20 good designs only one gets picked out to be implemented, which still has to be rethought and redone by the artist in actual implementation.
> This is reddit-tier ludditism.
it's practical knowledge acquired via observation. AI can only chance upon originality by smashing together everything else it's seen. this is obviously inferior to actually creating something original as a human, which even if it's not 500% alien and original can still be an original natural and cohesive development of other art. randomly putting a victorian hat on a medieval farmer holding an m16 because the prompt contained all of those words is something humans can already do with photoshop, and even the latest LLM won't give you a good story to pair with that anachronistic weirdo.
> We exist in a flood-tide of mental filth that was created by jews doing just this same thing ... Someone on our side has to do it [...]
jews don't mass produce AI slop masquerading as high art like the OP seems to imply is possible (look, we can make sci fi movies!). jews have billion dollar studios with big name actors. the material the AI had to consume to produce things like OP was first produced by (Whites employed by) jews who did it the manual way, and it still is to this day even if to a lesser extent.
besides, my point isn't that we shouldn't counter spam with spam. we should. but the jew does it via subterfuge, astroturfing, etc, not via outsourcing Hollywood to machine learning. we need to build actual communities of people who work together to achieve things like that, we can't do it from our basements and win "meme wars" no matter how advanced AI gets. eventually you need to offer people an alternative to all the filth - you know, like the whole thing we supposedly stand for?
> AI can only chance upon originality by smashing together everything else it's seen.
This is false, as this is the definition of overfitting. Models are intentionally not overtrained so that they learn features and relationships, not raw memorization (called overfitting, and it leads to what you describe, called regurgitation).
The ability to generate without regurgitation is a concept roughly called 'generalization', basically how well a model trained on a number of scenarios will perform on new scenarios it has never seen. This isn't new tech btw, researchers have been studying how well models generalize since the 50s.
Or, in short, well-trained models are literally generating, and literally creating. Models that don't are considered bad models.
> jews don't mass produce AI slop masquerading as high art
This can't be known to be true. Julian Assange's 2017 'proof of life' video famously had weirdness (like his lapel morphing into his shirt), which was claimed to be compression artifacts, but we now know is classic artifacting of movie-generating models.
> eventually you need to offer people an alternative to all the filth
Understandable, but I think the solution here is also AI. There is no way we would have the resources or structure to combat this filth otherwise (without a proper DOTR or actual government cleanout).
**Give it 6 months - 1 year. People said the same thing about still artists too and now modern comfyui workflows beat the top digital artists.**
> which can, by definition of how the technology works, only copy and never create
**This is reddit-tier ludditism. Its generative by definition, and, when trained on everything we've ever created, can create anything in the scope of our creation, even things never created before (which it does all the fucking time if you pay attention).**
> you're just running some GPUs to make some memes
**We exist in a flood-tide of mental filth that was created by jews doing just this same thing. You clearly weren't around for the meme wars of 2015-2017. Someone on our side has to do it, and you're retarded if you actually are trying to discourage our own from fighting back in the age of 4th generation warfare.**
It's tough to predict the pace of outcomes, especially given how everyone who's tried in the last 3 years has seriously underestimated the actual real timeline of LLM or vision model development. That said, I've personally studied the development of LLMs back when they were just government projects, including such 'mostly hidden' secrets such as:
- Twitter, which was purely infinite facebook statuses, was created as a playground for very early LLMs, called 'Influence Bots' in DARPA papers of the time
- These bots were proven to be operating en-masse on Reddit as early as 2017
- Every time the civilian-side development of language models stagnated, google released some 'breakthrough' paper/model that relaunched civilian interest and development - Word2Vec ('king' - 'man' + 'woman' = ? analogies, etc) in 2013, the infamous 'Attention is all you need' paper (which kicked off this current wave), the Gemma3 models from earlier this year (which taught us that our current models were extremely inefficient - these models were so much better than they had any right to be and now researchers scrambled to figure out 'why').
To gloss over what will be a long story, **the reason why everything is proceeding so quickly is because it isn't the pace of 'research', but rather the pace of 'leaking' of already discovered information that is driving this AI wave.**
> Everything is going to get worse from AI due to this
I'd argue it already was. I could make a (relatively weak, evidence-wise) argument that, for a long time, pop music has been made by machine models. I estimate it's been this way for over a decade now.
as someone who uses the modern comfyui workflows to make covers for his singles, an artist remains indispensable when you have an extremely specific artistic vision, like for the video game i'm developing with two artists.
i mean, i'll freely admit to having used AI just to mass-generate character design concepts, but for every 50 generations you get one good design, and for every 20 good designs only one gets picked out to be implemented, which still has to be rethought and redone by the artist in actual implementation.
> This is reddit-tier ludditism.
it's practical knowledge acquired via observation. AI can only chance upon originality by smashing together everything else it's seen. this is obviously inferior to actually creating something original as a human, which even if it's not 500% alien and original can still be an original natural and cohesive development of other art. randomly putting a victorian hat on a medieval farmer holding an m16 because the prompt contained all of those words is something humans can already do with photoshop, and even the latest LLM won't give you a good story to pair with that anachronistic weirdo.
> We exist in a flood-tide of mental filth that was created by jews doing just this same thing ... Someone on our side has to do it [...]
jews don't mass produce AI slop masquerading as high art like the OP seems to imply is possible (look, we can make sci fi movies!). jews have billion dollar studios with big name actors. the material the AI had to consume to produce things like OP was first produced by (Whites employed by) jews who did it the manual way, and it still is to this day even if to a lesser extent.
besides, my point isn't that we shouldn't counter spam with spam. we should. but the jew does it via subterfuge, astroturfing, etc, not via outsourcing Hollywood to machine learning. we need to build actual communities of people who work together to achieve things like that, we can't do it from our basements and win "meme wars" no matter how advanced AI gets. eventually you need to offer people an alternative to all the filth - you know, like the whole thing we supposedly stand for?
This is false, as this is the definition of overfitting. Models are intentionally not overtrained so that they learn features and relationships, not raw memorization (called overfitting, and it leads to what you describe, called regurgitation).
The ability to generate without regurgitation is a concept roughly called 'generalization', basically how well a model trained on a number of scenarios will perform on new scenarios it has never seen. This isn't new tech btw, researchers have been studying how well models generalize since the 50s.
Or, in short, well-trained models are literally generating, and literally creating. Models that don't are considered bad models.
> jews don't mass produce AI slop masquerading as high art
This can't be known to be true. Julian Assange's 2017 'proof of life' video famously had weirdness (like his lapel morphing into his shirt), which was claimed to be compression artifacts, but we now know is classic artifacting of movie-generating models.
> eventually you need to offer people an alternative to all the filth
Understandable, but I think the solution here is also AI. There is no way we would have the resources or structure to combat this filth otherwise (without a proper DOTR or actual government cleanout).