Paper Discussion: A Path Towards Autonomous Machine Intelligence

LLM and Multi-modal LLM has demonstrated it strong capacity on a range of tasks. It passed the Turing test and continues to challenge on more difficult tasks, like humanity's last exam.

However, some researchers still believe that to get an even powerful model, we need to build a model to learn fundamentals about the world via visual inputs. In this meetup, we'll look into Yann LeCun's vision for the future of AI. He proposed a new model architecture: JEPA (Joint Embedding Predictive Architecture). He positioned JEPA (along with ideas like energy-based models and world models) as crucial pieces toward building what he calls “autonomous AI agents” — a goal he believes is more realistic and robust than trying to just make bigger LLMs.

paper: https://openreview.net/pdf?id=BZ5a1r-kVsf