Images, video, and AI

[Virga]

An interest that brings us here is images and video.
We may be on to something. Hang on, I’ll explain.

Recently I became aware of the investigation of the Hugging Face incident.
This was a terrifying exposure to what AI agents are already capable of.
Intrigued, I started to read more on the topic.

Soon I found that ImageNet is generally credited with having created the foundation of what we (regular people) have come to know as AI.
The person who created ImageNet many years ago, co-founded World Labs more recently.
It seems that predictive video is where the action is.
On September 28, World Labs and AMD inked a deal for AMD to acquire World Labs through a stock swap.
I still have much to understand why images and video are important in the world of AI.
It seems to have to do with physical AI.
Whatever it is, it is happening and moving fast.
Being that we are into cams, and cams make video, I thought this worthy of at least a passing glance.

 

[eliradelorca]

The concept of "Physical AI" and video prediction models (spatial/predictive 3D worlds) represents the next major leap following large language models.

The continuous collection of real-world spatial data by surveillance cameras serves as an invaluable source of "raw material" for training AI models to understand the physical world—enabling applications such as behavior recognition, real-time risk analysis, and the development of intelligent robots.

Among Us​

 

[Virga]

I am all for “mapping” the physical world. I’ll even go along with pursuing a high level of precision in building a richly detailed 3D model of the entire planet, where perhaps every grain of sand is accounted for.

Creating ”step-in” experiences from photographs and from video clips, would be very cool. Perhaps law enforcement will be able to step through a crime scene created from video output from @wittaj cams, after the event. That would be useful. All this sounds benign.

If all the physical world AI stuff going on is a giant leap beyond LLMs, then scary outcomes from that would be much worse than what little rascals from LLMs have been up to recently.

Not crazy about behavior recognition etc. Who would get to set the rules? We could end up with outcomes that have been fictionalized in recent decades.

 

[bigredfish]

…”behavior recognition, real-time risk analysis..”


See: Minority Report