The AI Godfather Just Proved The Entire AI Industry Wrong

The AI Godfather Just Proved The Entire AI Industry Wrong

More

Summary

Yann LeCun has argued for years that scaling up large language models will not lead to human-level intelligence. This video from TheAIGRID looks at a new paper that may be his strongest evidence yet, in which a JEPA model with only 109 million parameters outperforms models up to 72 billion parameters on a specific test.

The video first covers LeCun’s background: his work on convolutional neural networks at Bell Labs, the 2018 Turing Award shared with Geoffrey Hinton and Yoshua Bengio, and his years as Meta’s chief AI scientist. It then explains why he left to found AMI Labs in Paris, which raised $1.03 billion in its first round with backers including Bezos’s investment firm, Nvidia, Samsung and Toyota.

The core of the video explains the paper, co-authored with Brown University’s Randall Balestriero, on proving and improving the Platonic Representation Hypothesis with JEPA. It argues that learning from “good views”, two observations of the same underlying reality, makes text and image models converge on the same internal representation, while next-word prediction is a weaker signal. On the alignment measure, the small JEPA scored 0.184 against 0.154 for Qwen 2.5 72B and 0.151 for Llama 3 70B, using roughly 42 times fewer training examples. The video also discusses whether the result will hold at larger scale.


📺 Source: TheAIGRID · Published November 05, 2026
🏷️ Format: News Analysis

1 Item

Channels

1 Item

People