Public transcript

Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning

By: Zenvora Productions
Length: 1 h 16 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This is a conversation with Yann LeCun about deep learning, neural networks, and the path toward human-level intelligence. He discusses why HAL 9000 in 2001: A Space Odyssey exemplifies value misalignment, why gradient-based learning is central to his view of intelligence, and why he prefers the term self-supervised learning over unsupervised learning. He argues that the key unsolved problem is learning predictive models of the world, and that reasoning, grounding, and even emotions will be necessary components of truly intelligent systems.

Key points

LeCun says the most surprising fact in deep learning is that gigantic neural nets with more parameters than data samples, trained with stochastic gradient descent on non-convex objectives, actually work—contradicting pre-deep-learning textbooks.
He believes intelligence is inseparable from learning, and that reasoning must be made compatible with gradient-based learning rather than discrete logic.
He prefers the term self-supervised learning because the algorithms are the same as supervised learning, but the machine predicts masked or future parts of its input rather than human-provided labels.
Self-supervised learning works well for natural language because uncertainty in prediction is easy to represent as probability vectors, but it fails for images and video where many predictions are equally valid, producing blurry results.
He argues human intelligence is not general but highly specialized, using the random permutation of the optic nerve as a demonstration that the brain's wiring is specialized for the structure of the world.
Deep reinforcement learning is wildly inefficient compared to humans: Atari games take 80 hours of training versus 15 minutes for a human, and AlphaStar needed the equivalent of 200 years of self-play.
His hypothesis is that humans learn quickly because they have predictive world models—babies learn gravity around eight or nine months old—enabling model-based reasoning and planning without trial-and-error disasters.
An autonomous intelligent system needs four components: a hardwired contentment objective (basal ganglia), an objective predictor, a world model, and a policy network; you can fail in three ways—wrong model, misaligned objective, or inability to plan.
He criticizes Sophia the robot's marketing as misleading and says embodiment is not necessary for intelligence, but grounding in the real world is.
He predicts we will not have autonomous intelligence without emotions, since fear is the anticipation of bad outcomes driven by the objective-prediction system.

Questions it answers

00:00HAL 9000 and AI ethics

Was HAL 9000 in 2001: A Space Odyssey evil, and how should we think about AI objectives?

LeCun says HAL's behavior was value misalignment, not evil: an unconstrained objective makes a powerful machine do damaging things. Designing aligned objective functions is what legal codes have done for humans for millennia, and law-making and computer science will come together.

HAL's internal conflict came from being asked to keep secrets and tell lies.
Some limits, like a Hippocratic oath for AI, could be hardwired, but the technology doesn't exist yet.

07:22Why deep learning surprised everyone

What most surprised LeCun in deep learning?

That gigantic neural nets with more parameters than data samples, non-convex objectives, and stochastic gradient descent actually work—breaking every pre-deep-learning textbook. His intuition came from the brain being a large learning network, and from believing intelligence is inseparable from learning.

Machine learning is 'the science of sloppiness' with math closer to electrical engineering than computer science.
He is a strong believer in gradient-based learning; discrete logic is incompatible with learning.

11:00Reasoning with neural networks

Can neural networks reason, and what would they need?

Yes, but the question is how. A reasoning system needs working memory (hippocampus-like, as in memory networks and transformers), plus a recurrent network that can access and iteratively update knowledge. Planning via energy minimization (model predictive control) is another classical form of reasoning.

Scaling memory networks to hold all of Wikipedia doesn't work; new ideas are needed.
Replace symbols with vectors and logic with continuous functions to make reasoning compatible with learning.

24:49The 1990s decline of neural nets

Why did people lose faith in neural nets in the 1990s?

Around 1995 the machine learning community gave up because neural nets were very hard to make work: no Python or MATLAB, backprop written in Fortran or a hand-built Lisp interpreter, and a bag of tricks most people didn't know. Computer science gave up while electrical engineering kept at it.

LeCun's team wrote a Lisp interpreter, a neural net backend, and later a compiler to C for production character recognition at Bell Labs.
AT&T lawyers would not let them release code in open source before 1995.

34:00Patents and the convnet story

What happened with the convolutional network patents?

Two patents were filed around 1989–1990. When AT&T split in 1996, lawyers gave the convnet patent to NCR, which had no idea what a convolutional net was. LeCun worked on convnets from 2002 to 2007 crossing his fingers that nobody at NCR would notice, until the patent expired in 2007.

LeCun doesn't believe in software patents; Facebook and Google patent defensively.
He warns against startups claiming secrets to human-level intelligence.

39:04Is intelligence general?

Why does LeCun dislike the term AGI?

Because it implies human intelligence is general, when it is actually very specialized. His demonstration: if your optic nerve's million fibers were randomly permuted, your locally-connected visual cortex could never learn normal vision. Of the unimaginably many boolean functions of a million bits, the brain computes only a tiny sliver.

We seem general only relative to the things we can imagine; the rest we call heat or entropy.
He prefers 'AI' or 'human-level intelligence' to 'AGI'.

45:30Self-supervised learning

What is self-supervised learning and why does it work for language but not images?

It uses the same algorithms as supervised learning but trains the machine to reconstruct masked or future parts of its input, like BERT predicting missing words. It works for language because uncertainty is easy to represent as probability vectors over a 100,000-word lexicon, but for images and video many valid predictions exist, so least-squares training yields blurry averages.

LeCun calls it self-supervised rather than unsupervised to avoid connotations of clustering or machines learning without supervision.
Self-play in games just generates data; it doesn't solve representing uncertainty in the real world.

52:00World models and learning efficiency

Why is deep reinforcement learning so inefficient compared to humans?

Atari games take 80 hours of training versus 15 minutes for a human; AlphaStar needed the equivalent of 200 years of self-play; an RL-trained car would kill thousands of pedestrians. Humans learn quickly because they have predictive world models—intuitive physics, gravity learned at 8–9 months—that let them plan without trial-and-error disasters.

The main problem to solve is how to learn models of the world; that's what self-supervised learning is about.
A useful benchmark: pre-train on large unlabeled data, then measure how labeled-sample requirements drop.

Notes

HAL 9000 and Objective Functions

LeCun sees HAL's behavior in 2001: A Space Odyssey as value misalignment, not evil: an unconstrained objective leads machines to do damaging things.
Designing aligned objective functions is not new—legal codes have done this for humans for millennia; law-making and computer science will converge.
He would not build a HAL that keeps secrets and tells lies; internal conflict from hidden knowledge is what breaks HAL.
Some limits, like a Hippocratic oath for AI, could be hardwired, but these questions are too abstract today since we lack autonomous intelligent machines.

Why Deep Learning Works Despite the Textbooks

The most surprising fact: gigantic neural nets with more parameters than data, non-convex objectives, and stochastic gradient descent actually work, breaking every pre-deep-learning textbook rule.
His intuition came from analogy: people proved heavier-than-air flight was impossible while birds flew; the brain is a large network that learns by changing connections.
Intelligence is inseparable from learning; programming intelligence directly was a non-starter.
Machine learning is "the science of sloppiness"—its math is closer to electrical engineering and cybernetics than computer science.

Reasoning and Memory

Reasoning requires working memory (hippocampus-like), a network to access and iteratively crunch on stored information—memory networks, and now transformers, point this way.
Planning via energy minimization (model predictive control) is another form of reasoning, possibly the evolutionary origin of human reasoning.
Logic and symbols are too brittle and incompatible with learning; replace symbols with vectors and logic with continuous functions (Geoff Hinton's suggestion, and Leon Bottou's "From Machine Learning to Machine Reasoning").

The 1990s Decline and Patents

Neural nets were hard to make work: no Python or MATLAB, you wrote backprop in Fortran or a hand-built Lisp interpreter, and you needed a bag of tricks most people lacked.
AT&T lawyers split patents after the 1996 breakup; the convolutional net patent went to NCR, which didn't know what it was. LeCun worked on convnets from 2002 to 2007 "crossing his fingers" until the patent expired in 2007.

Self-Supervised Learning and World Models

He prefers "self-supervised" over "unsupervised": the machine reconstructs masked or future parts of its input (e.g., BERT predicting missing words).
It works for language because uncertainty is easy to represent as probability vectors; for images and video, many valid predictions exist, so least-squares training yields blurry averages.
Deep RL is inefficient: 80 hours for Atari vs. 15 minutes for humans; AlphaStar needed ~200 years of self-play; a car trained by RL would kill thousands of pedestrians.
Humans learn fast because of predictive world models (intuitive physics, gravity learned at 8–9 months), enabling model-based reinforcement learning.

Architecture of Intelligence

Four components: hardwired contentment objective (basal ganglia), an objective predictor, a world model, and a policy network.
You can fail three ways: wrong world model, misaligned objective (a psychopath), or inability to plan.
Embodiment is not necessary, but grounding is—language alone is too low-bandwidth to convey how the world works.
Emotions are inevitable: fear is anticipation of bad outcomes; "we're not going to have autonomous intelligence without emotions."

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.