Keep pulling the thread on Yann LeCun.
Yann LeCun left Meta and founded Emilabs because he concluded Meta was not the right place to pursue his research on world models.
Mark Zuckerberg was disappointed with the LLaMA-4 model, which led him to reorganize Meta's GenAI organization and hire new people.
Meta's realization that it had fallen behind in AI led to a strategic refocus on catching up with competitors, which resulted in the deprioritization of exploratory research.
Non-generative, joint embedding architectures like DINO, IJEPA, and VJEPA, developed primarily at Meta, have proven more effective for learning image and video representations than generative approaches.
Yann LeCun predicts that the JEPA architecture and world model approach will achieve "complete world domination" in the AI field within 5 years.
Yann LeCun believes that companies with proprietary AI models, like OpenAI and Anthropic, are analogous to Sun Microsystems and HP in the 1990s and will be displaced by open platforms.
Due to intense short-term pressure, Meta's GenAI organization became very conservative and did not have time to collaborate with FAIR, hindering its ability to innovate.
In early 2023, Meta created the GenAI organization to productize the LLaMA-1 model, which was developed at its FAIR research lab.
The Masked Autoencoder (MAE) project at Meta's FAIR lab was considered disappointing, yielding unsatisfactory results despite significant computational investment.
Emilabs expects to have a general methodology for training hierarchical world models on a wide variety of data modalities within a year.
Emilabs plans to deploy its world model methodology into use cases with partners, including some of its investors, within 12 to 18 months.
The Tapestry project aims to create an open, free, and fine-tunable foundation model to allow different cultures and countries to develop AI assistants aligned with their specific languages and values.