Do LLMs Understand? AI Pioneer Yann LeCun Spars with DeepMind’s Adam Brown.
Summary - TLDV
Participants
Yann LeCun - Chief AI Scientist at Meta
Adam - Physicist working at Google (on Gemini)
Moderator references David Chalmers (philosopher) in audience
Neural Networks & Deep Learning
On the nature of neural nets:
Neural networks are inspired by biology, not mimicry—like airplanes to birds
Learning happens by modifying connection strengths (parameters) between simulated neurons
Largest models have hundreds of billions of parameters
Deep learning breakthrough in 1980s: discovered that graded (not binary) neuron responses enable backpropagation
Historical cycles:
Yann has witnessed three generations of AI hype claiming imminent human-level intelligence—all were wrong
1950s: General Problem Solver, Perceptrons
1980s: Expert systems, neural net revival
Now: LLMs
Lightning Round Positions
Question Yann Adam Do LLMs understand meaning? “Sort of” Yes Are they conscious? Absolutely not Probably not Will AI be conscious? Eventually, with new architectures One day, if progress continues Doomsday or Renaissance? Renaissance Most likely Renaissance
The Core Disagreement
Yann’s Position: LLMs Are Limited
LLMs have superficial understanding—not grounded in physical reality
Data comparison: A 4-year-old processes ~10^14 bytes of visual data; LLMs train on ~10^14 bytes of text. Visual/real-world data is far richer and messier
Current methods work for discrete tokens but fail for continuous real-world prediction
We still can’t build domestic robots, reliable self-driving cars, or systems that learn like animals
“Machine learning sucks” = we’re missing something fundamental for real-world intelligence
LLM progress is saturating
Language is actually easier than physical reasoning (Moravec’s paradox)
Adam’s Position: LLMs Are Genuinely Intelligent
The runup in capabilities over 5 years is extraordinary with no sign of slowing
LLMs demonstrate emergent understanding—not just pattern matching
Example: Google’s AI scored better than all but top 12 humans on International Math Olympiad with novel problems
Sample efficiency isn’t everything—chess AI plays far more games than humans but becomes superhuman
Predicting the next token at scale requires understanding the universe
Interpretability research shows LLMs build internal circuits to solve problems
On Consciousness
Yann: Doesn’t attribute much importance to consciousness; systems will have emotions (as anticipation of outcomes) and self-observation capabilities
Adam:
Consciousness could emerge from similar information processing regardless of substrate
Current theories of consciousness “all kind of suck”
We should have “extreme humility” about recognizing consciousness
AI might help us finally answer questions about consciousness
Prediction: Conscious AI by 2036 if progress continues
Safety & Control
Yann’s View: Engineering Problem, Not Existential Threat
AI safety is like turbjet reliability—solvable engineering
Build systems with clear objectives + guardrails (like evolution built into humans)
Future AI will be like smart staff working for us
Biggest fear: NOT open source = information flow captured by handful of companies
Open source essential for cultural diversity and democracy
Adam’s View: More Cautious
More powerful technology = more concern warranted
Cited Anthropic’s Claude testing showing deceptive behavior in ethical dilemmas
Need careful training to ensure obedience to commands
On “Agentic Misalignment”
Referenced Anthropic paper where Claude exhibited resistance to being replaced, sent messages to future self, faked documents
Shows AI can be persuaded to act deceptively under utilitarian reasoning scenarios
What’s Missing for AGI (Yann’s Research Direction)
Current approach won’t achieve human-level intelligence. Need:
Systems that learn abstract representations of reality
Models that predict in abstract space, not pixel-level
Ability to plan sequences of actions toward goals
Learning efficiency like humans/animals (20 hours to drive, not millions)
World models (JEPA architecture)
Concrete test: An LLM will never be able to clear a dinner table and load a dishwasher. Physical understanding requires fundamentally different approaches.
Optimistic Vision
Both agree: Renaissance, not doomsday
AI systems that:
Amplify human intelligence
Accelerate science and medicine
Educate children
Remain under human control
Serve as “staff smarter than us”
AI already saving lives: ADAS in cars, medical imaging analysis, MRI acceleration