It was great seeing Yann LeCun in Delhi, nearly a decade after my post-doc days at #NYU, where I worked with him. In his talk, Yann discussed the limitations of Large Language Models (LLMs) and believing that its not enough to achieve Artificial General Intelligence (AGI). He pointed out that LLMs treat both simple and complex problems with similar computational efforts, which can be inefficient. He argued that the chain of thought is just a creative hack to this issue.

Yann highlighted the sheer volume of data that LLMs are trained on—approximately 6 x 10^13 bytes which is almost all of the internet, which would take thousands of human hours to process. To provide a comparison, a typical four-year-old processes an immense amount of visual data in just 16,000 hours of life, at 2 megapixels per image captured by the eyes. This context underscores the current limitations and the need for progress beyond purely text-based training.

He stressed the importance of pioneering new methods, like training self supervised models on video, to achieve a more nuanced understanding and a closer approximation to human-like intelligence, complete with a world model and the ability to plan.

Let’s embrace these challenges and push towards more sophisticated, efficient AI systems.