Comment on OpenAI introduces Sora, its text-to-video AI model
Flumpkin@slrpnk.net 9 months ago
This is still so bizarre to me. I’ve worked on 3D rendering engines trying to create realistic lighting and even the most advanced 3D games are pretty artificial. And now all of a sudden this stuff is just BAM super realistic. Not just that, but as a game designer you could create an entire game by writing text and some logic.
Welcome to the club my friend… Expert after expert is having this experience as AI develops in the past couple years and we discover that the job can be automated way more than we thought.
First it was the customer service chat agents. Then it was the writers. Then it was the programmers. Then it was the graphic design artists. Now it’s the animators.
Another programmer here. The bottleneck in most jobs isn’t in getting boilerplate out, which is where AI excels, it’s in that first and/or last 10-20%, alongside dictating what patterns are suitable for your problem, what proprietary tooling you’ll need to use, what API’s you’re hitting and what has changed in recent weeks/months.
What AI is achieving is impressive, but as someone that works in AI, I think that we’re seeing a two-fold problem: we’re seeing a limit of what these models can accomplish with their training data, and we’re seeing employers hedge their bets on weaker output with AI over specialist workers.
The former is a great problem, because this tooling could be adjusted to make workers lives far easier/faster, in the same way that many tools have done so already. The latter is a huge problem, as in many skilled worker industries we’ve seen waves of layoffs, and years of enshitification resulting in poorer products.
The latter is also where I think we’ll see a huge change in culture. IMO, we’ll see existing companies bet it all and die from supporting AI over people, and a new wave of companies focus on putting output of a certain standard to take on larger companies.
This is a really balanced take, thank you
Writer here, absolutely not having this experience. Generative AI tools are bad at writing, but people generally have a pretty low bar for what they think is good enough.
These things are great if you care about tech demos and not quality of output. If you actually need the end result to be good though, you’re gonna be waiting a while.
If you actually need the end result to be good though, you’re gonna be waiting a while.
I agree with everything you said, but it seems in the context of AI development “a while” is like, a few years.
That remains to be seen. We have yet to see one of these things actually get good at anything, so we don’t know how hard that last part is to do. I don’t think we can assume there will be continuous linear progress. Maybe it’ll take one year, maybe it’ll take 10, maybe it’ll just never reach that point.
Still waiting on the programmer part. In a nutshell AI being say 90% perfect means you have 90% working code IE 10% broken code. Images and video (but not sound) is way easier cause human eyes kinda just suck. Couple of the videos they’ve released pass even at a pretty long glance. You only notice funny businesses once you look closer.
Yeah. And it’s not just how good the images look it’s also the creativity. Everyone tries to downplay this but I’ve read texts and those videos and just from the prompts there is a “creative spark” there. It’s not very bright spark lol but it’s there.
I should get into this stuff but I feel old lol. I imagine you could generate interesting levels with obstacles and riddles and “story beats” too.
Because sometimes the generator just replicates bits of its training data wholesale. The “creative spark” isn’t its own, it’s from a human artist left uncredited and uncompensated.
Artists are “inspired” by existing art or things they see in real life all the time. So that they can replicate art doesn’t mean they can’t generate art. It’s a non sequitur. But I’m sure people are going to keep insisting on this so lets not argue back and forth on this :D
I can’t imagine that digital artists/animators have reason to worry. At the upper end, animated movies will simply get flashier, eating up all the productivity gains. In live action, more effects will be pure CGI. At the bottom end, we may see productions hiring VFX artists, just as naturally as they hire makeup artists now.
When something becomes cheaper, people buy more of it, until their demand is satisfied. With food, we are well past that point. I don’t think we are anywhere near that point with visual effects.
It seems to me that AI won't completely replace jobs (but will do in 10-20 years). But will reduce demand because oversaturation + ultraproductivity with AI. Moreover, AI will continue to improve. A work of a team of 30 people will be done with just 3 people.
Keep in mind that this isn’t creating 3d Billy volumes at all. While immensely impressive, the thing being created by this architecture is a series of 2d frames.
Because it’s trained on videos of the real world, not on 3d renderings.
Lol you don’t know how cruel that is. For decades programmers have devoted their passion to creating hyperrealistic games and 3D graphics in general, and now poof it’s here like with a magic wand and people say “yeah well you should have made your 3D engine look like the real world, not to look like shit” :D
ArmokGoB@lemmy.dbzer0.com 9 months ago
In my experience as a game designer, the code that LLMs spit out is pretty shit. It won’t even compile half the time, and when it does, it won’t do what you want without significant changes.
DSTGU@sopuli.xyz 9 months ago
The correct usage of LLMs in coding is either for a single use case at a time, building up to what you need from scratch. It requires skill both in talking to AI for it to give you what you want, knowing how to build up to it, reading the code it spits out so that you know when it goes south and the skill of actually knowing how to build the bigger picture software from little pieces but if you are an intermediate dev who is stuck on something it is a great help.
That or for rubber ducky debugging, it s also great in that
colonial@lemmy.world 9 months ago
That sounds like more effort than just… writing the code.
DSTGU@sopuli.xyz 9 months ago
It s situationally useful
kspatlas@lemm.ee 9 months ago
Chatgpt once insisted my JSON was actually YAML
jkrtn@lemmy.ml 9 months ago
Technically it is, but I agree that is imprecise and nobody would say so IRL. Unless they are being a pedantic nerd, like I am right now.
Flumpkin@slrpnk.net 9 months ago
Traister101@lemmy.today 9 months ago
You should refine your thoughts more instead of dumping a stream of consciousness on people.
Essentially what this stream of consciousness boils down to is “Wouldn’t it be neat if AI generated all the content in the game you are playing on the fly?” Would it be neat? I guess so but I find that incredibly unappealing very similar to how AI art, stories and now video is unappealing. There’s no creativity involved. There’s no meaning to any of it. Sentient AI could probably have creativity but what people like you who get overly excited about this stuff don’t seem to understand is how fundamentally limited our AI actually is currently. LLMs are basically one of the most advanced AI things rn and yet all it does is predict text. It has no knowledge, no capacity for learning. It’s very advanced auto correct.
We’ve seen this kind of hype with Crypto with NFTs and with Metaverse bullshit. You should take a step back and understand what we currently have and how incredibly far away what has you excited actually is.
HeavyDogFeet@lemmy.world 9 months ago
I don’t mean to be dismissive of your entire train of thought (I can’t follow a lot of it, probably because I’m not a dev and not familiar with a lot of the concepts you’re talking about) but all the things you’ve described that I can understand would require these tools to be a fuckload better, on an order we haven’t even begun to get close to yet, in order to not be super predictable.
It’s all wonderful in theory, but we’re not even close to what would be needed to even half-ass this stuff.