By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Alex KantrowitzWhy Can't AI Make Its Own Discoveries? — With Yann LeCun
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Yann LeCun on why LLMs can’t discover—and what’s next
- Yann LeCun argues that large language models mainly learn to produce text matching training statistics, making them strong at retrieval but fundamentally limited at inventing new solutions or asking novel research questions.
- He says current “reasoning” approaches like chain-of-thought are limited because they lack an internal search-and-verification mechanism and operate in token space rather than abstract mental-model space.
- He describes diminishing returns in scaling LLMs due to constrained text data and costly post-training, concluding that human-level AI will not come from LLM scaling alone.
- He outlines four missing capabilities for more general intelligence—understanding the physical world, persistent memory, reasoning, and planning—and says they require learning from video and building world models.
- He presents JEPA (including video-focused V-JEPA) as an alternative training approach that learns abstract representations for prediction, enabling detection of physically impossible events and supporting action-conditioned planning.
IDEAS WORTH REMEMBERING
5 ideasLeCun draws a hard line between “AI” broadly and LLM chatbots specifically.
He describes LLMs as statistical text systems optimized to reproduce patterns from training data; they can be highly useful for retrieval and fluent answers but still produce false statements and, in their “purest form,” don’t create genuinely new ideas.
Token-by-token “reasoning” isn’t the same as internal search-and-check reasoning.
He argues chain-of-thought can make models appear more thoughtful by spending more tokens/compute, but it doesn’t add the core mechanism typical of reasoning: searching through a space of candidate solutions with a way to verify constraints.
The LLM paradigm can improve, but with shrinking payoff relative to cost.
He says scaling LLMs is already showing diminishing returns because natural text is effectively exhausted, while patching gaps via synthetic data and mass human labeling becomes increasingly expensive for limited gains.
A system that feels expert via recall is not the same as a system that can invent.
He rejects the idea that simply scaling LLMs will reach “human-level AI,” and says “a PhD next to you” is more like massive memory/retrieval than true invention of new solutions to new problems.
The hardest part is the last mile: reliability, integration, and real-world deployment.
He says impressive demos don’t guarantee production reliability, drawing parallels to earlier AI cycles (expert systems) and describing how deployment can fail when the final reliability/integration hurdles are hard.
WORDS WORTH SAVING
5 quotesChatbots, LLMs, are trained on an enormous amount of, of knowledge, which is purely text, and they're trained to basically regurgitate, to retrieve, to, uh, essentially produce answers that are conformed to the statistics of whatever text they've been trained on.
— Yann LeCun
We don't think in language. We think in, uh, you know, mental representations of a situation. We have mental models of everything we think about.
— Yann LeCun
There is no mechanism at all in LLMs for this search mechanism. What you have is you have to kind of build this on top of it, right?
— Yann LeCun
We are not going to get to human-level AI by just scaling up LLMs. This is just not going to happen, okay?
— Yann LeCun
The problem of understanding the world does not go through the solution to the, to, to generating video at the pixel level, okay?
— Yann LeCun
SPEAKERS
Alex Kantrowitz
hostTechnology journalist and host of the Big Technology Podcast.
Yann LeCun
guestMeta Chief AI Scientist and Turing Award–winning researcher in deep learning.
High quality AI-generated summary created from speaker-labeled transcript.
