- AI future (by Yann LeCun). Fb post
- Both LLMs and Reinforecement Learning (RL) have limitations.
- RL: consume a lot of resources, hard to scale, RL models hard to adapt when env changes frequently, can’t plan long-term.
- Instead, try Self-Supervised Learning.
- LLMs: can create content but cannot “undestand” the world, limited in planning and high-level reasoning capabilities.
- New propposed approach: JEPA (Joint Embedding Predictive Architecture):
- AI don’t need to predict the next step/token, but learn the abstract about how the world operates.
- Learn from video, photos, sounds rather than just text.
- Can plan before act, just like human.
Tho Le
A Data Scientist. Looking for knowledge!