About this episodeYann LeCun argues that current Large Language Models (LLMs) are insufficient for achieving human-level intelli…AI summary
Yann LeCun argues that current Large Language Models (LLMs) are insufficient for achieving human-level intelligence because they lack a physical world model and persistent memory, functioning only as 'System 1' reactive systems. He predicts that within 5-10 years, new architectures like JEPA (Joint Embedding Predictive Architecture) will enable AI to learn from video and perform 'System 2' hierarchical planning, leading to practical applications in robotics and autonomous systems. For entrepreneurs, the immediate opportunity lies in fine-tuning open-source foundation models for specific verticals like law, finance, and healthcare.
Key takeaways 6
LLMs are limited to discrete data (text) and cannot understand the continuous, high-dimensional physical world (video/images), making them incapable of true reasoning or physical interaction.
Intelligence is defined by three components: existing skills, the ability to learn new skills quickly, and the ability to solve novel problems zero-shot using a mental model.
The next major breakthrough in AI is 'Objective-Driven AI' using JEPA architecture, which predicts abstract representations of future states rather than pixel-by-pixel video generation, enabling planning and reasoning.
Open-source platforms will dominate the AI ecosystem within five years due to flexibility, security, and cost-effectiveness, similar to the rise of Linux in operating systems.
Human intelligence will shift from execution to strategy; people will act as 'bosses' delegating tasks to AI systems, focusing on abstract decision-making rather than low-level work.
Reinforcement learning is inefficient for general tasks but highly effective for games where agents can play millions of games against themselves to learn policies.
Notable quotes 4AI-generated: wording and quote attribution may be wrong. Use the play link to verify.
“The smartest LLMs are not as smart as your house cat.”
▶ 1:02:47LeCun explains that while LLMs can pass bar exams, they lack basic physical world understanding that even a cat possesses, highlighting the gap between language manipulation and true intelligence.
“Science is never a sort of individual pursuit. You make progress by the collision of ideas from multiple people.”
▶ 4:23LeCun responds to being called the 'Godfather of AI,' emphasizing that scientific progress is collaborative and attributing credit to many contributors rather than a single individual.
“We're all going to be a boss. We're all going to be like those high-level managers who will tell our AI systems what to do.”
▶ 1:29:58Describing the future of work, LeCun suggests that AI will handle execution, allowing humans to focus on strategy and decision-making.
“The cost of inference for LLM has gone down by a factor of 100 in two years. I think there is still a lot of room for improvement.”
▶ 1:20:21Highlighting the rapid economic viability of AI deployment, noting that inference costs are dropping faster than Moore's Law.
Chapters & Sections (61)▼
0:31Early Life and Career Influences in Sciencechapter3
0:31Origins of Artificial Intelligence and Career Path
2:59The Intersection of AI and Human Consequences
4:30The Role of Collaboration in Scientific Progress
7:44Human Intelligence and Problem Solving Limitationschapter3
7:44Human Limitations and Problem-Solving Capacity
9:21Importance of Critical Thinking and Education
10:51Origins and Definition of Artificial Intelligence
12:59Defining Intelligence and AI Problem-Solving Approacheschapter2
12:59Defining Intelligence and Problem-Solving Approaches
14:51Early AI Development and Branching Theories
17:31Emergence of Classical Computer Science and AIchapter3
17:31Heuristics in Classical Computer Science and AI
19:18Approaches to Artificial Intelligence Development
21:08Early Image Recognition and Neural Networks
24:59Early History of Artificial Neural Networkschapter2
24:59Early AI Development and Neural Network Evolution
27:23Supervised Learning and its Limitations
29:28Overview of Artificial Intelligence Subfieldschapter2
29:28Overview of Artificial Intelligence Subfields
32:12Types of Machine Learning Techniques Explained
34:44Self-Supervised Learning for Image Recoverychapter3
34:44Self-Supervised Learning for Image Recovery
36:33Neural Networks and Self-Supervised Learning
38:06Building a Simple Neural Network for Image Recognition
39:39Neural Network Activation and Backpropagationchapter2
39:39Neural Network Activation and Backpropagation
41:55Convolutional Neural Networks and Image Recognition
44:36Convolutional Neural Networks for Natural Datachapter2
44:36Convolutional Neural Networks for Natural Data
47:48Neural Network Components and Equivariance
50:11History and Basics of Language Modelschapter2
50:11Origins of Language Model Theory
52:13Limitations of N-Gram Language Models
54:43Predictive Modeling of Language Sequenceschapter2
54:43Predicting Next Word in Language Models
56:45Emergent Properties of Large Language Models
59:50Limitations of Autoregressive Models in Predictionschapter2
59:50Limitations of Autoregressive Models in AI
1:01:37Limitations of Large Language Models Explained
1:04:26Designing Architectures for Self-Supervised Learningchapter5
1:04:26Designing Architectures for Self-Supervised Learning
1:05:58Limitations of Predicting Pixel Sequences in Video
1:07:18Human Reasoning and Planning Abilities
1:09:28Limitations of Current AI Architectures
1:12:10Predictive Modeling and Long-term Forecasting Limitations
1:13:36Predictive Capabilities of Artificial Intelligencechapter2
1:13:36Predictive Capabilities of Artificial Intelligence
1:15:56Limitations of Current LLM Architectures and Training
1:19:15Importance of Local Computing Infrastructurechapter3
1:19:15Future of AI Infrastructure in Emerging Markets
1:21:18Benefits of Advanced Education in Innovation
1:22:42Building a Business with Narrow Intelligence
1:24:42Potential Applications of AI in Various Industrieschapter2
1:24:42Potential Applications of AI in Various Industries
1:26:44Future of Open-Source Technology Platforms
1:29:13Human Intelligence in an AI-Driven Worldchapter4
1:29:13Human Intelligence in an AI-Driven World
1:30:44Impact of Automation on Human Productivity
1:32:03Defining Artificial Intelligence and Human Intelligence
1:33:43Importance of Accessible Education and Self-Learning
2h 57m
2h 36m
3h 12m⏭ 2
2h 7m
2h 29m
3h 24m
2h 42m
3h 23m⏭ 1
2h 8m
3h 8m
3h 9m⏭ 1
3h 16m
53m
1h 20m
2h 27m⏭ 4
2h 3m
2h 22m⏭ 4
1h 15m
1h 6m⏭ 2
1h 40m
45m
3h 17m⏭ 2
1h 4m
2h 50m
11h 23m⏭ 1
1h 0m⏭ 1
33m
46m
3h 54m⏭ 7
3h 59m⏭ 5
48m⏭ 2
56m
1h 13m
38m
40m
54m⏭ 2