391–400 of 570 posts
Earlier quoted context omitted.
In the future, a new intelligent species will roam the earth, they will ask, "why did their civilization fall?" The answer? These homo-sapiens strip mined the Earth and exacerbated climate change to generate enough power to make amusing cat videos...
And those videos were either not watched by anyone human or not truly watched by being part of an endless feed of similar slop.
Google hit the jackpot with their acquisition of YouTube and it's now paying dividend. YouTube is the largest single source of data and traffic on the Internet, and it's still growing fast. I think this data will prove incredibly important to robotics as well. It's a shame they sold Boston Dynamics in one of their dumbest ever moves because of bad PR.
If we look at the Veo 3 examples, this is not the typical youtube video, but instead they seem to recreate cgi movies, or actual movies.
I'd made a prediction/bet a month ago, predicting 6 months to a full 90 minute movie by someone sitting on their computer. [0] The pace is so crazy that was an over estimation! I'll probably get done in 2. Wild times. 0: https://www.linkedin.com/feed/update/urn:li:activity:7317975...
Earlier quoted context omitted.
"Growing fast" is questionable these days. There is an ever growing percentage of new AI-generated videos among every set of daily uploads. How long until more than half of uploads in a day are AI-generated?
Even if the content was 100% AI generated (which is the furthest thing from reality today) human engagement with the content is a powerful signal that can be used by AI to learn. It would be like RLHF with free human annotation at scale.
After doing some testing, Imagen 4 doesn't score any higher than Imagen 3 on my comparison chart, approximately ~60% prompt adherence accuracy. https://genai-showdown.specr.net
Not sure if this affects your results or not but I resist chiming in!
Earlier quoted context omitted.
I don't think this is just about convenience - you're not going to get these results with a 14B video model. I'd much prefer to have something I could hack on in ComfyUI but the open weights models don't compete with this anymore than a 32B LLM competes with Gemini 2.5 Pro for coding. And at least in coding you can easily edit the output from the LLM regardless...
> you're not going to get these results with a 14B video model Foundation models are starting to outstrip any consumer hardware we have. If Nvidia wants to stay ahead of Google's data center TPUs for running all of these advanced workloads, they should make edge GPU compute a priority. There's a future where everything is a thin client to Google's data centers. Nvidia should do everything in its power to prevent that…
there has always been, the mainframe concept is not new. but it goes in and out of fashion.
>>>> mainframe
>>>> web pages/social media
>>>> cloud ai
>>>> ???? rented swarms ???
After doing some testing, Imagen 4 doesn't score any higher than Imagen 3 on my comparison chart, approximately ~60% prompt adherence accuracy. https://genai-showdown.specr.net
Side note. It's my understanding that being a pith helmet is pretty orthogonal to having a spike. Plenty of helmets with spikes aren't pith helmets and plenty of pith helmets don't have spikes. Not sure if this affects your results or not but I resist chiming in!
I wonder how much the commonality or frequency of names for things affects image generation? My hunch is that it it roughly correlates and you'd get better results for terms with more hits in the training data. I'd probably use Google image search as a rough proxy for this.
Earlier quoted context omitted.
Even if the content was 100% AI generated (which is the furthest thing from reality today) human engagement with the content is a powerful signal that can be used by AI to learn. It would be like RLHF with free human annotation at scale.
Won't the human engagement be replaced by AI engagement too? if it isn't already being replaced?
This is technically impressive and I commend the team that brought it to life. It makes me sad, though. I wish we were pushing AI more to automate non-creative work and not burying the creatives among us in a pile of AI generated content.
There is a more sensical distinction between work that is informational in nature, and work that is physical and requires heavy tools in hard-to-reach places. That's hard to do for big tech, because making tests with heavy machinery is hard and time consuming
Earlier quoted context omitted.
And google is in the best possible position to detect it if they want to exclude it from their datasets.
They're never going to manage to do that, just on a technical level Plus some users might want to legitimately upload things with AI-generated content in it