We are scarily close to realtime personalization of video which if you agree with this NeurIPS paper [1] may lead to someone inadvertently creating “digital heroin”
> We further urge the machine learning community to act proactively by establishing robust design guidelines, collaborating with public health experts, and supporting targeted policy measures to ensure responsible and ethical deployment
We’ve seen this play out before, when social media first came to prominence. I’m too old and cynical to believe anything will happen. But I really don’t know what to do about it at a person level. Even if I refuse to engage in this content, and am able to identify it, and keep my family away from it…it feels like a critical mass of people in my community/city/country are going to be engaging with it. It feels hopeless.
Having the ability to do real-time video generation on a single workstation GPU is mind blowing.
I'm currently hosting a video generation website, also on a single GPU (with a queue), which is also something I didn't even think possible a few years ago (my show HN from earlier today, coincidentally: https://news.ycombinator.com/item?id=46388819). Interesting times.
Looks like there is some quality reduction, but nonetheless 2s to generate a 5s video on a 5090 for WAN 2.1 is absolutely crazy. Excited to see more optimizations like this moving into 2026.
those streams of text are often conditioned on the prompts - people are using it to learn about new concepts, and as a hyperpersonalised version of search. it can not only tell you of tools you didn't know existed, but it can show you how to use them.
I do like my buttons to stay where I left them - but that can be conditioned. instead of gnome "designers" telling me the button needs to be wide enough to hit with my left foot, I could tell the system I want this button to be small and in that corner - and add it to my prompt.
Nevertheless it does seem that generating will fairly soon become fast enough to extend a video clip in realtime. Autoregressive by the second. Integrated with a multi modal input model you would be very close to an AI avatar that would be extremely compelling.
I mean the baselines were deliberately worse and not how someone would be using these to begin with maybe noobs and the quoted number is only for DIT steps not for other encoding and decoding steps, which is actually quite high still. No actual use of FA4/Cutlass based kernels nor TRT at any point.
Now if someone could release an optimization like this for the M4 Max I would be so happy. Last time I tried generating a video it was something like an hour for a 480p 5-second clip.
[1] https://neurips.cc/virtual/2025/loc/san-diego/poster/121952
We’ve seen this play out before, when social media first came to prominence. I’m too old and cynical to believe anything will happen. But I really don’t know what to do about it at a person level. Even if I refuse to engage in this content, and am able to identify it, and keep my family away from it…it feels like a critical mass of people in my community/city/country are going to be engaging with it. It feels hopeless.
I'm currently hosting a video generation website, also on a single GPU (with a queue), which is also something I didn't even think possible a few years ago (my show HN from earlier today, coincidentally: https://news.ycombinator.com/item?id=46388819). Interesting times.
I actually think we are already there with quality, but nobody is going to wait 10 minutes to do a task with video that takes 2 seconds with text.
If Sora/Kling/whatever ran cool locally 24/7 at 60FPS, would anyone ever build a UI? Or a (traditional) OS?
I think it's worth watching the scaling graph.
I like my buttons to stay where I left them.
I do like my buttons to stay where I left them - but that can be conditioned. instead of gnome "designers" telling me the button needs to be wide enough to hit with my left foot, I could tell the system I want this button to be small and in that corner - and add it to my prompt.