Public transcript

#416 – Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI

By: MindVoice Production
Length: 2 h 54 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this conversation, Yann LeCun explains why autoregressive LLMs, while useful, are not a path to human-level intelligence because they lack understanding of the physical world, persistent memory, reasoning, and planning. He argues that world models must be learned from video via joint embedding predictive architectures (JEPA) rather than generative pixel prediction, and outlines an objective-driven, energy-based blueprint for future AI systems that plan answers by optimization. He also defends open source as the only way to ensure diverse, unbiased AI assistants, rejects doomer scenarios about AGI as relying on false assumptions, and expresses optimism that AI will make humanity smarter, likening it to the printing press.

Key points

Yann LeCun argues autoregressive LLMs lack understanding of the physical world, persistent memory, reasoning, and planning — four essential characteristics of intelligence.
A 4-year-old receives roughly 10^15 bytes of visual data in 16,000 waking hours, vastly more than the ~2×10^13 bytes of text LLMs train on, so most knowledge comes from observation, not language.
Generative models trained to predict pixels from video failed for 10 years; joint embedding predictive architectures (JEPA) that predict abstract representations instead work well.
Hallucination is inherent to autoregressive prediction: errors accumulate and the probability of a nonsensical answer grows exponentially with the number of tokens.
LeCun's blueprint for future AI is objective-driven: systems that plan answers by gradient-based optimization in abstract representation space before decoding to text.
No AI system can be unbiased because bias is in the eye of the beholder; the only solution is diversity, enabled by open source platforms anyone can fine-tune.
AGI will not be an event but gradual progress; rogue AI will be countered by good AI, and systems won't inherently want to dominate since domination drives are hardwired only in social species.
LeCun compares AI to the printing press: it will make humanity smarter, with some disruption but an overall positive effect.

Questions it answers

09:07Limits of autoregressive LLMs

Why does Yann LeCun think autoregressive LLMs won't reach human-level intelligence?

He says LLMs lack understanding of the physical world, persistent memory, reasoning, and planning, and that language alone is too low-bandwidth to build a world model.

Four essential characteristics of intelligence are missing from LLMs.
A 4-year-old sees ~10^15 bytes vs ~2×10^13 bytes of LLM training text.
Most knowledge comes from observing the world, not language.

18:10JEPA and world models

What replaces generative models for learning from video?

Generative pixel-prediction approaches failed for a decade; joint embedding predictive architectures (JEPA) that predict abstract representations instead, using non-contrastive methods like DINO, I-JEPA and V-JEPA, work well.

Reconstruction-based self-supervision produces poor image features.
JEPA predicts representations, not pixels, lifting the abstraction level.
V-JEPA learns video representations good for action classification and detecting impossible physics.

48:00Planning and model predictive control

How can AI systems plan like humans?

A world model plus model predictive control allows planning action sequences to satisfy objectives, but hierarchical planning across learned abstraction levels remains unsolved.

World models enable planning at inference time.
Hierarchical planning (New York to Paris) is necessary but nobody knows how to learn it.
LLMs can only regurgitate plans similar to training data.

57:20Hallucination and drift

Why do LLMs hallucinate?

Because each token has a probability of error, mistakes compound so the chance of nonsense grows exponentially with answer length, and fine-tuning can't cover the long tail of possible prompts.

Errors accumulate exponentially in autoregressive generation.
LLMs behave like giant lookup tables over trained prompts.
Computation per token is constant regardless of problem difficulty.

1:12:00System 2 and energy-based models

What blueprint could give AI systems real reasoning?

Objective-driven systems would optimize an energy-based model by gradient descent in abstract representation space before decoding to text, mirroring human System 2 deliberation.

Answers would be planned via optimization before being uttered.
Energy-based models score compatibility of an answer with a prompt.
RLHF's reward model is similar but currently used only for fine-tuning, not inference.

1:36:00Open source and bias

Why is open source the answer to AI bias and control?

Because no system can be unbiased — bias is in the eye of the beholder — the only solution is diversity of AI assistants, which open source platforms enable for any language, culture, or value system.

AI assistants will mediate all digital interaction and can't be controlled by a few West Coast companies.
Examples: Llama fine-tuned for India's 22 languages, Senegalese local languages.
Meta's business model (ads, business customers) is not impaired by open sourcing base models.

2:01:00AGI timelines and doomers

Why does LeCun reject doomer scenarios?

He says superintelligence won't be an event, guardrails will be built iteratively like turbojet safety, and rogue AI will be countered by good AI; domination drives are hardwired only in social species.

AGI arrives gradually, not as a single event.
Objective-driven AI can include guardrails like obeying humans.
Breakthroughs disseminate quickly, so no single actor stays far ahead.

2:33:00Hope for humanity

What gives LeCun hope for humanity's future with AI?

He compares AI to the printing press: it will amplify human intelligence, cause disruption but not mass unemployment, and ultimately empower the goodness in people.

Everyone will have a staff of smart AI assistants, like managing smarter people.
The printing press analogy: transformative despite 200 years of religious conflict.
Economists expect gradual shifts in professions, not mass unemployment.

Notes

Why LLMs Are Not the Path to Human-Level Intelligence

Yann LeCun lists four essential characteristics of intelligent systems: understanding the physical world, persistent memory, reasoning, and planning — LLMs can do none or only primitive versions of these.
LLMs train on roughly 10^13 tokens (~2×10^13 bytes), which would take 170,000 years to read at 8 hours a day. Yet a 4-year-old child absorbs about 10^15 bytes through vision in 16,000 waking hours, showing most knowledge comes from observing the world, not language.
He cites the Moravec paradox: computers pass the bar exam but can't learn to drive in 20 hours like a 17-year-old or load a dishwasher like a 10-year-old.

Joint Embedding Predictive Architectures (JEPA)

Attempts to train generative models to predict video pixels failed for 10 years (GANs, VAEs, regularized autoencoders, MAE-style reconstruction).
JEPA instead runs a full input and a corrupted/masked version through encoders and trains a predictor to predict the representation of the full input from the corrupted one — predicting in abstract representation space, not pixel space.
FAIR's I-JEPA, V-JEPA, and DINO use non-contrastive, distillation-based methods. V-JEPA masks temporal tubes across 16-frame videos and produces representations good enough for action classification and detecting physically impossible videos.

Planning and Objective-Driven AI

A world model predicts the state of the world at time t+1 given a state and an action, enabling model predictive control — how rocket trajectories have been planned since the early 1960s.
Hierarchical planning (New York to Paris example) is essential but nobody knows how to learn the multiple abstraction levels automatically.
Future dialogue systems should plan answers by optimization: an energy-based model scores whether an answer fits a prompt, and gradient descent in abstract representation space refines the answer before decoding to text — System 2 versus the System 1 behavior of current LLMs.

Hallucination and LLM Limits

Each generated token has some probability of leaving the set of reasonable answers; errors compound so the probability of nonsense grows exponentially with answer length.
Fine-tuning covers the common 80% of prompts, but the long tail of possible prompts is too large, making LLMs effectively giant lookup tables.
Computation per token is constant regardless of problem difficulty, unlike humans who spend more time on harder problems.

Open Source, Bias, and AGI

Bias is in the eye of the beholder; an unbiased system is impossible. Like a free press, AI assistants must be diverse, which requires open source platforms that any group can fine-tune (e.g., Llama for India's 22 official languages, or Senegalese local languages).
AGI won't be an event; progress will be gradual, and rogue AI will face good AI. Desire to dominate is hardwired in social species, not a property of intelligence itself.
LeCun compares AI to the printing press: transformative, disruptive (200 years of religious conflict), but ultimately making humanity smarter — and notes the Ottoman Empire banned Arabic printing partly to protect the calligraphers' guild.

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