Keep pulling the thread on Yann LeCun.
Self-driving car companies have been unable to achieve human-level driving reliability by training models to imitate humans, despite using millions of hours of training data.
The concept of Artificial General Intelligence (AGI) is nonsensical because human intelligence itself is specialized.
Scaling up Large Language Models (LLMs) will not lead to Artificial General Intelligence (AGI).
Human-like intelligence cannot be achieved by training models solely on text.
The autoregressive prediction model used by LLMs, which generates one token at a time through a fixed number of layers, is not a good model for reasoning.
AI systems built with world models and guardrail objectives can be made intrinsically safe, unlike LLMs.
The only way to make a Large Language Model safe is through fine-tuning, and these systems can always be jailbroken.
Hierarchical planning is an unsolved and completely open problem in AI research.
All of the best self-supervised learning systems for training image or video representations use joint embedding architectures, not reconstruction-based methods.
World models for AI should not be simulators, digital twins, or generative video models; they should operate in an abstract representation space.
The V-JEPA model has learned a level of common sense, demonstrated by its internal prediction error increasing significantly when shown physically impossible events in a video.
Academics should not work on Large Language Models because they cannot contribute anything new to the field.