Comments by "Shawn Fumo" (@ShawnFumo) on "Fireship"
channel.
-
56
-
14
-
10
-
8
-
7
-
6
-
I do think we're going to be in for a hard time with stuff like this. Definitely not a bad idea to set up a code word with family in case someone spoofs a voice, etc. The most fundamental problem though is that it feels almost impossible to actually control all of this. I think people are under the impression that the only way to train and run these things are with oodles of money and compute, but that is changing quite fast. Like a recent research paper showed how to train an image model from scratch (not at the level of Flux but better than Stable Diffusion 1.5/2) for less than $2,000, with 37m images (SD 1.5 cost $300k and used 4b training images). And even though SD took a lot to train, the result is a 2GB file that you can run locally offline on an iPhone. And those can already easily be modified to impersonate various people, change (inpaint) existing images, etc.
And that's all with the current hardware and current techniques. In a year or two, maybe it costs $200 to train a Flux-quality model and maybe you can run that on a new generation phone. Obviously video generation is a lot more intensive, but what is cutting edge now will be easier in a few years as well. Once the techniques are known publicly, it feels like it is pretty hard to actually regulate this stuff. Sure, we can clamp down on the big companies like OpenAI and Google and Grok that make it super easy for anyone to do online, but that doesn't really stop the fact that anyone can download an open weights model from online and use it however they want.
6
-
4
-
4
-
4
-
4
-
4
-
3
-
3
-
3
-
3
-
2
-
2
-
2
-
2
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
@tttttttttttttttp12 I tend to think most of the good and bad claims are true at the same time. I think we're all in for a rough time pretty soon, no matter what we do (art, programming, even physical labor eventually). And certainly a lot of money pouring into research is because they think it'll help their bottom line in the end.
I don't think it is all necessarily bad. Like back when I was learning to draw, while it was interesting, I think part of the reason I didn't stick with it was not having a clear "voice" on what I'd want to use it for that would make all the time to get better at it worth it. When I messed around with MidJourney back when it it was new, there was the side-effect of feeling more in touch with visual creativity in terms of what themes, color schemes, styles, etc. that I most connected with. Not in terms of a pure consumer, but in being involved with new images appearing. Turns out they weren't the same kinds of images as those I might like to just look at. That kind of experience (and also the lack of control in generators) made me want to put more effort toward learning to draw/paint than before using it. I think others have had that experience too, a kind of gateway like people who learning programming to help them make stuff in a video game.
One good/bad aspect on training sources is it seems like a lot less images are needed than first thought. There was a recent research paper about how they trained a good image model (not Flux level but better than Stable Diffusion 1.5/2) from scratch for less than $2000 and using 37m images, about a third of which were synthetic. That's starting to get to the number where you could probably source all public domain images to train it. On the bad end, it also means the tech is pretty much uncontrollable (at least for regular image generators), since almost anyone could spend $2000 and it'll just keep getting cheaper with new hardware and techniques.
In any case, try not to let what other people are doing take the joy of a creation away from you. Even though anyone can buy cheap items of all sorts from any big box store, there's still a market for hand-made things. And there's still many reasons to create things apart from selling them. I know that only helps so much if you want a standard career in it without having to sell via being a personality, but I think we're all in that boat. I predict everything is going to go a bit topsy turvy soon.
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1