By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Alex Kantrowitz

Why Can't AI Make Its Own Discoveries? — With Yann LeCun

Yann LeCun is the chief AI scientist at Meta. He joins Big Technology Podcast to discuss the strengths and limitations of current AI models, weighing in on why they've been unable to invent new things despite possessing almost all the…

Alex KantrowitzhostYann LeCunguest

Mar 19, 202559mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Why “memorizing all text” doesn’t produce scientific discovery
Alex asks why systems trained on massive amounts of human knowledge haven’t generated major new discoveries. Yann LeCun argues that this expectation is misplaced for large language models because their core behavior is producing text that matches training statistics, not inventing new ideas.
- •Alex relays a question about why generative AI hasn’t made new scientific connections despite absorbing huge corpora
- •Yann LeCun distinguishes “AI” broadly from chatbots/LLMs specifically
- •LLMs are described as excellent at producing statistically likely text but prone to producing false statements
- •Yann LeCun says LLMs, in their pure form, are not capable of inventing truly new things
Why “asking the right questions” is a different skill than solving them
Alex brings up the idea that an “Einstein in a data center” would need to ask questions nobody else would ask. Yann LeCun says current LLMs can’t do that and explains that humans don’t primarily think in language, but in internal situation models, with language as a separate output layer.
- •Alex references the idea that breakthroughs require asking heretical or novel questions
- •Yann LeCun says LLMs won’t become that kind of system in their current form, though they could be a component
- •He compares LLM-like text production to a small specialized part of the brain, not the seat of intelligence
- •He explains research involves multiple skills: retrieving knowledge, framing good questions, formulating equations, and solving them
- •He characterizes LLMs as strong at retrieval but weak at new problem-solving and especially weak at question formation
Reasoning vs. “chain of thought”: why token-by-token isn’t real search
Alex asks whether “reasoning models” and step-by-step approaches can enable contrarian thinking. Yann LeCun argues that current approaches mainly bolt procedures onto LLMs, and that genuine reasoning requires searching a solution space—something LLMs don’t intrinsically do.
- •Alex asks if reasoning-style prompts can induce deeper thinking or self-questioning
- •Yann LeCun questions what “reasoning” means and describes current methods as incremental tweaks
- •He describes chain-of-thought as generating extra tokens to spend more computation, with limited benefit
- •He defines classical reasoning as searching through candidate solutions with a way to evaluate success
- •He criticizes approaches that generate many candidate outputs and select among them as inefficient
Humans reason with mental models, not text: cubes, planning, and “regurgitation”
Yann LeCun explains that humans and animals reason in internal, abstract spaces rather than in language tokens. Alex connects this to how impressive “novel” outputs can still be recombinations of training material, and Yann LeCun agrees much of it is regurgitation.
- •Yann LeCun says humans don’t reason by producing many textual solutions and picking the best
- •He uses a cube-rotation visualization to illustrate non-linguistic internal modeling
- •He argues internal reasoning happens in abstract/continuous representations, not text space
- •Alex cites an example where a model’s chain-of-thought “insight” sounds like a remix of popular books
- •Yann LeCun agrees that much model output is essentially regurgitation
Diminishing returns for LLM scaling and the need for a new architecture
Alex raises the idea that predictable improvements from scaling may be ending. Yann LeCun argues that text data is being exhausted, synthetic and human-generated patches are costly, and that a new paradigm is needed—especially for agentic systems that can plan and solve new problems.
- •Alex frames a potential “wall” in LLM progress; Yann LeCun reframes it as diminishing returns
- •Yann LeCun says systems have been trained on extremely large amounts of text and data is running out
- •They discuss synthetic data and hiring humans to fill gaps, with diminishing payoff
- •Yann LeCun says many systems still don’t understand basic logic
- •He says true agentic behavior needs planning and solution-searching, not regurgitated plans
Investment timelines: “PhD next to you” vs. systems that can invent
Alex challenges the mismatch between massive investment in LLM-first companies and the need for a new paradigm. Yann LeCun says scaling LLMs won’t reach human-level AI, criticizes claims of near-term “genius data centers,” and contrasts retrieval-heavy assistants with the inventive work a PhD does.
- •Alex lists major investments and asks what happens if LLM progress slows before a new paradigm arrives
- •Yann LeCun says human-level AI will not come from simply scaling LLMs
- •He rejects the idea that such capability is imminent and calls “country of genius in a data center” unrealistic
- •He predicts people may feel like they have a “PhD next to you,” but it’s really memory/retrieval at scale
- •He explains infrastructure spending is largely about serving many users, so it can still be justified
Consumer adoption vs. enterprise deployment: reliability and the “last mile”
Alex contrasts consumer usage with enterprise needs, where small hallucination rates can be unacceptable. Yann LeCun says deployment difficulties are not new, pointing to autonomy demos and the hard “last mile” of reliability, integration, and trust.
- •Alex discusses why consumer adoption can grow even if systems remain flawed
- •He describes enterprise issues: high costs, fallibility, and few pilots reaching production
- •Yann LeCun compares this to autonomous driving: impressive demos vs. the difficulty of full reliability
- •He emphasizes integration into workflows and achieving near-perfect reliability as the hard part
- •They discuss how deployment friction is a recurring pattern in tech and AI
Historical parallels: Watson, expert systems, and the risk of backlash
Yann LeCun reviews earlier AI waves where hype exceeded real-world deployment, including IBM Watson’s medical ambitions and the 1980s expert-systems boom. Alex asks directly about the risk of another AI winter, and Yann LeCun argues the bigger mistake is believing scale alone delivers human-level AI.
- •Yann LeCun describes IBM Watson’s deployment ambitions and why real-world rollout failed
- •He recounts the expert-systems era and the “knowledge engineer” concept
- •He describes how broad societal transformation didn’t materialize for most expert systems
- •Alex asks if today’s investment and expectations could trigger a new AI winter/backlash
- •Yann LeCun says investors betting on “just scale LLMs” for human-level AI are making a mistake
Four missing capabilities: world understanding, memory, reasoning, and planning
Yann LeCun outlines what he considers core capabilities current systems lack: understanding the physical world, persistent memory, reasoning, and planning. He says progress will be continuous and community-driven rather than a single “AGI event,” and he argues open research sharing accelerates advancement.
- •Yann LeCun names four needed characteristics: physical-world understanding, persistent memory, reasoning, planning
- •He says these require learning from natural sensory data like video, not only text
- •He notes progress exists in research settings but practical scalable architectures will take time
- •He argues there will not be a single moment where AGI suddenly arrives; advances will be continuous
- •He says broad sharing of research helps the field move faster than secrecy
The “paper test” and why video generators still don’t prove comprehension
Alex revisits a past physics question where a model hallucinated, and notes the current system now answers correctly. Yann LeCun argues this could be fine-tuning coverage rather than true understanding and says video generation systems still fail in unanticipated scenarios, revealing brittle learning.
- •Alex recounts the earlier “paper held horizontally” question and how it once produced a wrong answer
- •He tests it again and gets a correct explanation; Yann LeCun suggests it may reflect targeted fine-tuning
- •Yann LeCun argues you can’t train for all possible questions because the long tail is effectively infinite
- •Alex points to improved video generation (e.g., plausible joints/limbs) as evidence of progress
- •Yann LeCun counters that failures like impossible fluid behavior show a lack of underlying understanding
Why prediction should be abstract: JEPA and learning world models for planning
Yann LeCun explains why pixel-level generative prediction is mathematically hard in high-dimensional continuous spaces and argues it’s the wrong path to world understanding. He introduces JEPA (and video variants) as a way to learn abstract representations that support detecting impossibilities and enabling action-conditioned prediction for planning.
- •Yann LeCun argues understanding the world doesn’t require predicting every pixel-level detail
- •He compares human prediction limits (e.g., exact fluid paths) with our ability to predict outcomes abstractly
- •He describes self-supervised learning via corruption-and-reconstruction and says it works poorly for video world modeling
- •He explains why predicting distributions over high-dimensional frames is intractable, motivating non-generative approaches
- •He outlines JEPA: predicting representations (embeddings) rather than reconstructing inputs, and describes detecting “impossible” events via prediction error
- •He describes action-conditioned models that predict next abstract state from current state plus action, enabling search over action sequences (planning)
DeepSeek and open source: faster iteration, broader talent, cheaper deployment
In a time-limited closing exchange, Alex asks about DeepSeek and whether open source has overtaken proprietary models. Yann LeCun says open source progresses faster through global talent and shared ideas, notes deployment advantages (cost and control), and argues strong research capability exists worldwide.
- •Yann LeCun says “overtaken” is too strong, but open source innovation moves faster
- •He argues proprietary companies benefit from open source progress because they can learn from it too
- •He says small teams with freedom can generate new ideas and, when shared, accelerate global progress
- •He reports partners saying prototypes may use proprietary APIs, but deployments often use open models for cost and control
- •He argues geopolitical fears about open source are misguided and emphasizes strong scientific talent globally
Wrap-up and sign-off
Alex closes by thanking Yann LeCun and reflecting on how these discussions help cut through hype. The episode ends with a brief sign-off and note about the next show.
- •Alex thanks Yann LeCun for returning and for providing grounding amid industry hype
- •Yann LeCun thanks Alex
- •Alex signs off and previews the next episode timing

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.