Comment on Star Wars' Showcase of AI Special Effects Was a Complete Disaster
Grimy@lemmy.world 1 week agoIn other words, exactly the kind of video AI can deliver at a consumer level.
It can do much more. This was literally someone that typed “owl slug” into the prompt.
But AI can’t do that.
Character replacement and lip synching can currently be done. Its not perfect but its advancing very fast.
Your points are valid and all but AI can do a lot, and can do more every month. It’s already pretty versatile and this is currently the worse it is going to be.
xyzzy@lemm.ee 1 week ago
Yes, this is a favorite line from the industry, who assume the trend line continues uninterrupted into the future. But how about this as a counter future: what if AI plateaus?
What if it doesn’t get much better than it already is except around the edges, and the next breakthrough is two decades away? Companies have exhausted training data and exhausted data center capacity in the quest to keep the trend line at the previous vector. Yes, they’re building new capacity, but no one is making any money on this except Nvidia.
LLMs haven’t seen any significant improvement in a couple years. Image generation has improved, but at a much slower pace. Video is no longer Will Smith eating spaghetti, but there’s a long, long valley between where we are today and convincing, photorealistic, extended scenes that can be controlled at a fine level.
Grimy@lemmy.world 1 week ago
Every two weeks, theres a new generative AI game changer that pops out. There is nothing pointing towards hitting a plateau.
Chat gpt 3.5, which started the whole craze, came out less then 2.5 years ago. There has been a mountain of improvements since then.
Likewise, there has been monumental steps for all generative services. That long valley for video gen is just a stroll away now.
The futur is hard to predict but everything points away from a plateau. AGI might be far but not animators making extensive use of AI tools for example.
xyzzy@lemm.ee 1 week ago
I was referencing 4, which was released over two years ago and was a significant improvement over 3.5. I was genuinely impressed with 4, but I haven’t been very impressed with anything since then. Probably the most substantive change was pulling chain of thought into the model itself, but everyone was already doing anyway.
I’m not coming at this from a place of ignorance: I have AI patents to my name as both first inventor and supporting, and I’ve worked with these teams directly. I’m saying that the rate of improvement in critical (i.e., non-toy) areas is slowing down, and I believe it’s a significant possibility that AI will start to hit the same walls it did many times before. That was before it entered the consciousness of execs and the general public, and because they aren’t as familiar with the long stop-start history of AI, they don’t think that wall exists.
AI companies definitely know that wall exists, and in at least one case they’re getting increasingly nervous about it.