Dec 10, 2021 · 1h 7m · big-technology
Daniel Kahneman and Yann LeCun: How To Get AI To Think Like Humans (Full Episode)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Big Technology Podcast dialogue hosted by Alex Kantrowitz, Nobel laureate Daniel Kahneman and Turing Award laureate Yann LeCun explore how human cognitive architecture can inform the design of truly intelligent machines. They contrast current deep learning limitations with biological learning, debating the essential roles of self-supervised world models, common sense, symbolic discretization, and predictive mental simulation.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 8.1% of the talking time here. How this is scored →
Alex as informed peer 1.6 Guest teaching 4.2 Guest disagreement 1.3 Alex pushing back 0.3
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 16:20 LeCun challenges nativist assumptions in biological vision
LeCun directly pushes back against nativist claims by showing that oriented edge detectors can be learned within minutes rather than requiring genetic pre-specification.
Hardest push from Alex ▶ 21:10 Kantrowitz and Kahneman demand precision on symbolic reasoningThe host intervenes to pin down a definition of symbolic reasoning after LeCun expresses skepticism about whether symbolic reasoning exists distinctively in cognition.
Biggest teaching moment ▶ 25:48 LeCun's tutorial on discrete signals and error correctionLeCun delivers a comprehensive technical explanation linking information theory, analog-to-digital discretization, and noise-tolerant communication in neural systems.
Alex holds their own ▶ 37:00 Kantrowitz contextualizes historical AI rivalriesKantrowitz jumps in to clearly categorize hand-engineered Bayesian networks as deep learning's historical competitors in the good old fashioned AI era.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Yann LeCun on Defining Intelligence and the Limits of Current Machine Learning | 3 | 3 | 1 | 1 | Kantrowitz opens by asking LeCun about replicating the human mind. LeCun clarifies his goal is understanding intelligence rather than pure replication, contrasting current supervised/RL systems with biological learning. | |
| Daniel Kahneman on World Representations, System 1, and Symbolic Thought | 2 | 4 | 1 | 0 | Kantrowitz invites Kahneman to unpack System 1 and System 2. Kahneman explains world representations and poses a fundamental challenge to LeCun regarding whether generic learning can acquire symbolic certainty without innate structure. | |
| Common Sense and Observational Learning in Human Infants and Animals | 1 | 5 | 1 | 0 | LeCun explains why driving or intuitive physics takes humans only hours compared to millions of hours in reinforcement learning. The host takes a back seat as LeCun breaks down developmental milestones in human babies. | |
| The Nature vs. Nurture Debate in Neural Architectures | 1 | 5 | 2 | 0 | Kahneman introduces nativism and hierarchical Bayesian models as counterpoints. LeCun partially rejects nativism, arguing that edge filters and face recognition do not need hardcoding because neural algorithms learn them rapidly from minimal sensory input. | |
| Defining Symbolic Reasoning and Discrete Representations | 2 | 3 | 2 | 1 | The host asks for definitions of symbolic reasoning as LeCun questions whether animals reason symbolically. Kahneman defends discreetness and logical categories while acknowledging LeCun's challenge. | |
| Information Theory, Noise Resistance, and the Evolution of Language | 2 | 5 | 1 | 0 | LeCun provides a detailed masterclass on information theory, discrete error correction, and why digital signals evolved in neural communication and human language. Kantrowitz briefly asks for clarification on discrete representations. | |
| Limits of Hand-Engineered Bayesian Systems vs. Learned Neural Concepts | 2 | 5 | 2 | 1 | Kahneman asks whether discretization must be innate, and Kantrowitz requests a definition of hierarchical Bayesian systems. LeCun explains explaining-away logic in Bayesian networks and critiques hand-engineered expert systems for failing to learn representations. | |
| Reasoning as Mental Simulation and Predictive Coding in the Brain | 1 | 4 | 1 | 0 | LeCun and Kahneman discuss non-symbolic reasoning as analogical simulation and predictive coding. LeCun notes conscious perception is actually a forward prediction compensating for neural transmission delay. | |
| Evolutionary Specialization vs. Learned Face Recognition in AI | 2 | 4 | 1 | 1 | Kahneman queries whether face identification requires innate specialized systems. LeCun points out that deep learning solved face identification without pre-engineered innate structures, exceeding human capacity. | |
| Eliminating Absurd AI Errors and Overfitting to Context | 2 | 4 | 1 | 0 | Kahneman probes whether machines can be designed to avoid absurd errors. LeCun explains spurious correlations in vision models (like failing to recognize a cow on a beach) and argues model-based common sense is required. | |
| Self-Supervised Learning in LLMs and the Grounding Problem | 1 | 5 | 2 | 0 | LeCun explains masked-token self-supervised learning in LLMs while critiquing their lack of physical grounding. Kahneman and LeCun agree that text-only prediction leads to superficial understanding and hallucinated absurdities. | |
| Overcoming Video Prediction Failures with Joint-Embedding Architectures | 1 | 5 | 1 | 0 | Kahneman asks if adding video to language models will yield a qualitative breakthrough. LeCun details why direct pixel prediction fails and introduces joint-embedding architectures with non-contrastive methods to prevent representation collapse. | |
| Latent World Models, Uncertainty, and Retrodictive Sense-Making | 1 | 4 | 2 | 0 | Kahneman asks whether current prediction failures stem from scale or fundamental architecture, and highlights retrodictive sense-making. LeCun agrees that representing multimodal uncertainty in latent space is the core theoretical hurdle. | |
| Yann LeCun's Cake Metaphor: Self-Supervised Learning as the Dark Matter of AI | 1 | 4 | 1 | 0 | Kahneman brings up LeCun's cake metaphor. LeCun elaborates on self-supervised learning as the cake, supervised learning as icing, and reinforcement learning as the cherry, dubbing unsupervised world modeling the dark matter of intelligence. | |
| Industry Growth, Conceptual Innovation, and the Risk of Future AI Winters | 2 | 3 | 1 | 1 | Kahneman asks about exponential progress and the host manages time constraints. LeCun assesses industry resilience against a traditional AI winter while highlighting the race to crack machine common sense before funding fatigue sets in. |
Statements from this episode (28)
LeCun: Engineers cannot directly build intelligence, machines must learn it
“There is no intelligence without learning. And in the engineering world, it's almost true as well. And it's probably because I'm either lazy or not smart enough to, that I think that as human engineers, we cannot actually directly conceive and construct an int…”
LeCun: Supervised and reinforcement learning do not reflect biological learning
“The type of learning that we are currently able to reproduce in machine, which is supervised learning and reinforcement learning do not seem to Reflect what we observe in humans and animals. There is another type of learning, another paradigm of learning that …”
LeCun: Current machine learning replicates System 1 thinking without reasoning
“Currently what we can do in machine learning is more like the system when the stuff that, you know, here is an input, here is an output that does not re require reasoning if you want.”
Kahneman: Human intelligence requires causal world representations to make predictions
“When we describe human intelligence, we describe, we speak of a representation of the world. And it's the representation that leads to prediction. That is, there is no shortcut to the prediction from the data. You go through a representation, which includes Ho…”
Kahneman: Human minds make sense of events retrospectively, not by constant prediction
“It's not so much that we have specific expectations about what's going to happen next. What is happening most of the time is that things happen and then we make sense of them. That is, we actually go back and fit them into what happened before.”
Kahneman: Internal world models and feeling unsurprised are System 1 functions
“The representation of the world That we have in our ability to anticipate or to feel unsurprised by what happens, which is I think more than anticipating, ah, that is all system one. That's all, you know, that's all automatic. It's effortless and it's very qui…”
LeCun: Current AI systems are brittle due to task specialization
“Well I mean, I agree with Danny that current current AI systems are, you know, very specialized, and that makes them very brittle because they're trained for one task or maybe a collection of tasks.”
LeCun: Reinforcement learning cannot safely or efficiently train self-driving cars
“If we were to use, let's say, reinforcement learning to train a self-driving car to drive itself,
It would have to drive itself for millions of hours and cause, you know, until thousands of accidents and destroy itself multiple times before it learns to drive …”
LeCun: Face Detection in Cortex Is Rapidly Learned, Not Innately Hardwired
“Face detection can be learned in minutes. If you're a baby, your vergence is bad. Your focus is, is basically fixed at a relatively short range. So the only thing you see during the first weeks of your life are faces and nipples, essentially. So, you know, and…”
LeCun: Brain mechanisms for symbolic reasoning do not need to be hardwired
“Is it necessary to have explicitly hardwired mechanisms in our brain that allows us to do things like symbolic manipulation or reasoning? And I find that hard to believe that, that they need to be hardwired.”
LeCun: Nearly all animal brains have discrete symbolic representations
“If by symbol, we mean the ability to form sort of discrete categories represent sort of discrete categories in our, you know, mental representation system. Then I think pretty much every brain has some sort of symbolic representation.”
LeCun: Brains use digital neural spikes for energetic efficiency and signal regeneration
“Our brains actually use digital communication internally. A neuron, you know neurons send spike to communicate with each other. And the reason is it's easier to regenerate a binary signal than it is to regenerate an analog signal, and it's more efficient energ…”
LeCun: Discrete conceptual entities in biological minds are completely learned
“And I certainly do not believe that those, the meaning of those discrete entities are predetermined in the human mind, certainly, and the animal mind either. Those are completely learned.”
LeCun: Deep neural nets are hierarchical Bayesian systems when viewed properly
“Multi-layer, you know, deep neural nets are hierarchical Bayesian systems. If you kind of view them the right way, and there are certain forms of them that are actually explicitly Bayesian.”
LeCun: Hand-engineering systems from scratch caused decline of classical AI
“All those things would be built entirely by hand. Okay. Completely engineered. And that was essentially that, that necessity was one of the reason of the kind of decrease in interest in sort of good old fashioned AI.”
LeCun: Human and animal reasoning is mental simulation, not logic
“A lot of reasoning in certainly in animals and in humans is not logical reasoning. It's basically simulation. Or analogical reasoning, which kind of similar.”
LeCun: AI's missing key component is predictive world models
“The key element in this form of intelligence is your ability to build models of the world, predictive models of the world. And that's what we're missing. Okay. That's what we need to figure out how to do with machines.”
LeCun: Conscious perception is a 100-millisecond forward prediction
“In fact, your perception of the world is not the world as it is. It's the world as it's going to be, because there is, you know, about a hundred millisecond delay between what you see and how your brain interprets what it is. So so in fact your brain predicts …”
LeCun: AI Systems Can Identify More Faces Than Any Human
“Identify faces and have a system that recognize, recognizes more faces than any human can with you know, the same level of accuracy. You know, maybe not with the same sort of robustness to you know, different changes in pose and facial hair and things like tha…”
LeCun: World models, not more data, will solve AI common sense
“So I think this is not going to be solved in my opinion by you know, more tweaks on the architectures and more training data and things like this. I mean, it may be mitigated, but I think it's not going to be fixed by that. I think it's going to be fixed by sy…”
LeCun: Most Human Knowledge Is Not Represented in Any Text
“Like most of human knowledge is not represented in any text in existence.”
LeCun: True AI Intelligence Requires Grounding in a World with Physics
“So I'm one of those people who believe that truly intelligent systems will need to acquire knowledge will need to be grounded in some reality. It could be a simulated reality. It could be a virtual world. But it has to be an environment that has its own logic …”
LeCun: Existing video prediction AI models fail beyond a fraction of a second
“There's a lot of systems that attempt to do video prediction and basically attempt to learn representations of the world that are in an abstract way can learn to predict what's going to happen in the video in the long term. They all work within a few frames of…”
LeCun: Joint-embedding self-supervised learning is AI's best shot for world models
“I'm really excited about this area of research because I think it's the germ. I think it's our best shot as to you know, a path towards learning, you know, getting machines to learn abstract representations you know, and more predictive models. Where those pre…”
LeCun: Self-Supervised Learning Forms the Bulk of Natural and Machine Intelligence
“Most of what we learn as humans and animals and in, in the future that machine will learn is learned in this kind of self-supervised manner, basically by watching the world go by and by, you know, taking an action once in a while, but in a, you know, non-task …”
LeCun: AI Research Focuses on Icing and Cherries Before Baking the Cake
“We, we're currently focusing on the icing and the cherry, but we haven't figured out how to bake the cake yet.”