Meta is looking for the next big steps in AI, which have become a focus of Meta CEO Mark Zuckerberg as he looks to a new digital interaction era where real artificial intelligence will be possible.
Meta’s V-JEPA2 world model is the latest effort in this area. It can be used to help computers understand their environments and factor that into their response.
Meta-explained:
” We humans are able to anticipate how our actions, or those of others will affect the physical universe. You know, for example, that gravity will bring a tennis-ball back to earth if it is thrown into the air. You’re trying to get where you want to go while avoiding bumping into other people and obstacles. This physical intuition is achieved by developing an internal mental model that allows us to anticipate the outcome of hypothetical situations.
Meta’s VJEPA 2 model allows AI agents to mimic the V-JEPA 2 model to better understand how to respond to other elements.
” V-JEPA 2 was trained using videos, which allowed the model to learn patterns of the real world. These included how people interacted with objects and objects moved in the world. It also learned how other objects interacted with each other. We found that V-JEPA 2 can be used by robots to do tasks such as reaching out, picking an object up, and placing it in a different location.
This is another step in the direction of artificial general intelligence, or computer systems which can mimic human thought processes and “think for themselves” to a certain degree.
What remains of the Holy Grail in artificial intelligence?
The current generation of AI tools is good at replicating human responses, based on the data they have access to. These systems do not have “intelligence”, they are using predictive models that place letters in order based on similar inputs.
These systems do not interpret the different elements to come up with novel, creative solutions. They are smart spreadsheets that mimic human behavior.
Meta is looking to achieve this with its “superintelligence” initiative, and has hired a variety of experts in the industry for an increased push towards AGI.
AGI is it even possible? Meta’s AI Chief Yann Le Cun – one of the world’s most respected and experienced minds in AI development – seems to believe so.
We’re still not there.
LeCun said at the Joint Mathematics Meetings in March of this year.
Many people within the AI community believe that we could build machines with a blueprint which would eventually be able to reach human-level intelligence in the next decade. Estimates of how long this will take are hugely different, and the optimistic say that we have already arrived. “Some people are raising a lot of money and claiming that it will happen next year. But I do not believe in this.”
LeCun claims that by focusing on key models and open-source development we can develop an AI intelligence framework that is human level “within the next decade, or half a decade,” leading to systems capable of planning and reasoning.
LeCun does not believe the AI models of the present generation are the best framework for this.
Meta may be exploring new approaches to AI, with an eye towards a new kind of intelligence capable of assessing and understanding different aspects.
The latest initiative is a step forward in this direction. With Meta’s compute power also increasing significantly, the project may be best positioned to crack code and develop real systemic intelligence.
Imagine the Large Hadron Collider for AI. Meta is likely working toward this goal with its AI project.