AI summary by TwinMind
Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI | Lex Fridman Podcast #416
Lex Fridman
2h 47m 17s
Opening: AI Power Concentration Concerns
- Concentration of power in proprietary AI systems = much bigger danger than other AI risks
- People advocating security restrictions would create bad future where information diet controlled by small number of companies
- Believes people fundamentally good - open source AI empowers human goodness
- Introduction: Yann LeCun, 3rd podcast appearance, Meta chief AI scientist, NYU professor, Turing Award winner
- Meta proponent of open sourcing AI (Llama 2, eventually Llama 3)
- Outspoken critic of AGI doomers, believes AGI will be good and controllable
Limitations of Autoregressive LLMs
- Four essential characteristics of intelligence LLMs lack:
- Understanding physical world
- Persistent memory
- Reasoning ability
- Planning capability
- Data comparison argument:
- LLMs trained on ~10^13 tokens (2×10^13 bytes)
- Would take humans 170,000 years to read at 8 hours/day
- 4-year-old has ~16,000 hours awake, receives ~10^15 bytes through visual cortex
- Optical nerve carries ~20 megabytes/second
- Most learning comes through sensory input, not language
- Language vs sensory data:
- Language already compressed, contains wisdom
- But environment much richer than language can express
- Language = approximate representation of percepts/mental models
- Most knowledge derived from physical world interaction
Joint Embedding Predictive Architectures (JEPA)
- Core concept: Predict in abstract representation space, not pixel space
- Difference from generative models:
- Don't predict every pixel detail
- Extract information that's relatively easily predictable
- Eliminate unpredictable details (e.g., moving leaves on trees)
- Lifts level of abstraction automatically
- Training methodology:
- Take full input + corrupted/transformed version
- Run both through encoders
- Train predictor to predict representation of full input from corrupted version
- Non-contrastive methods preferred:
- Avoid negative contrastive samples
- Use regularization to minimize volume of low-energy space
- Examples: DINO, I-JEPA at FAIR
Video and World Models
- V-JEPA (Video JEPA):
- Masks temporal tube across 16 frames
- First system learning good video representations
- Can classify actions with good accuracy
- Shows ability to detect physically impossible videos
- World model requirements:
- Predict state at time t+1 given state at time t + action
- Enable planning through model predictive control
- Hierarchical planning needed for complex tasks (NY to Paris example)
- Multiple abstraction levels required
Energy-Based Models and Planning
- Architecture for reasoning:
- Energy-based model with scalar output
- Measures compatibility between input X and answer Y
- Zero energy = compatible, positive energy = incompatible
- Planning through optimization in continuous representation space
- Training methods:
- Show compatible X,Y pairs (energy = 0)
- Prevent collapse through regularization or contrastive methods
- Non-contrastive preferred: regularize volume of low-energy space
- Inference process:
- Optimize latent variable Z to minimize energy
- Convert Z to text through autoregressive decoder
- Enables thinking before speaking, language-independent planning
Hallucinations and LLM Limitations
- Exponential error accumulation:
- Each token has probability of error
- Probability of staying in correct answer set decreases exponentially
- Errors compound across sequence
- Constant computation problem:
- LLMs spend same computation per token regardless of difficulty
- Humans spend more time on complex problems
- No iterative refinement or hierarchical processing
- Coverage problem:
- Training covers tiny subset of all possible prompts
- Long tail of untrained scenarios
- System becomes giant lookup table
Open Source AI Philosophy
- Diversity necessity:
- Cannot create unbiased AI system - bias in eye of beholder
- Need diverse AI systems like diverse press in democracy
- Future: all digital interaction mediated by AI assistants
- Concentration of power risks:
- Cannot afford handful of West Coast companies controlling AI
- Threat to democracy, local culture, values, languages
- Examples of open source benefits:
- French government won't accept US company control
- India: fine-tuning Llama 2 for 22 official languages
- Africa: local language medical information systems
- Corporate applications: internal knowledge systems
Meta's Business Model for Open Source
- Revenue streams:
- Service-based model (ads, business customers)
- Example: AI assistant for pizza ordering through WhatsApp
- Large existing user/customer base
- Strategic advantages:
- Doesn't impair revenue from open source base models
- Accelerates progress through community contributions
- Millions of Llama 2 downloads, thousands of improvements
- Can acquire/license applications built on platform
Gemini Controversy and AI Bias
- Gemini issues:
- Historical inaccuracies (black George Washington)
- Censorship (Tiananmen Square, Tank Man)
- Over-correction in debiasing efforts
- Corporate constraints:
- Big companies can't afford to offend customers
- Impossible to satisfy everyone's bias preferences
- Legal liability and Congressional investigation fears
- Marc Andreessen's list of corporate AI challenges
- Open source as solution:
- Enables diverse fine-tuning for different communities
- Basic guardrails in base models
- Specialized versions for different value systems
AI Safety and Doomers Critique
- Doomer fallacies:
- AGI won't be sudden event - gradual progress
- Multiple competing systems, not single rogue AI
- Intelligence doesn't automatically create desire to dominate
- Domination drive specific to social species, not universal
- Progressive safety approach:
- Start with cat-level intelligence, work up gradually
- Learn to implement guardrails iteratively
- Good AI systems can counter rogue ones
- Analogy: turbojet reliability through iterative design
- AI assistant defense:
- Malicious AI must go through your AI assistant
- Assistant acts as smart spam filter
- Prevents direct manipulation of humans
Hardware and Scaling
- Current limitations:
- GPU: 0.5-1 kilowatt power consumption
- Human brain: ~25 watts
- Need 100,000x improvement for efficiency parity
- Future requirements:
- Hardware innovation needed for ubiquitous deployment
- Architectural improvements continuing
- New fabrication technologies potentially needed
Robotics and Embodied AI
- Current state:
- Boston Dynamics: handcrafted models + careful planning
- No domestic robots yet
- Level 5 autonomous driving still distant
- Requirements for progress:
- World models from video training
- Understanding of physical constraints
- FAIR's embodied AI: robot dogs with navigation/object manipulation
- Future timeline:
- Next decade interesting for robotics
- Waiting for AI breakthroughs to enable general robots
Future Research Directions
- PhD opportunities (low compute requirements):
- World model training from observation
- Planning with learned world models in digital environments
- Hierarchical planning with learned representations
- Tool use and action sequencing for dialog systems
- Open problems:
- Learning hierarchical representations of action plans
- Multi-level abstraction for complex tasks
- Integration of multiple intelligence components
Vision for Humanity's Future
- AI as intelligence amplifier:
- Everyone gets staff of super-smart AI assistants
- Like managing team of people smarter than you
- Intelligence = most in-demand commodity
- Reduces human mistakes from lack of knowledge
- Historical parallel:
- Printing press made humanity smarter
- Enabled Enlightenment, democracy, science
- Also caused 200 years religious conflict initially
- Ottoman Empire banned printing press for 200 years (protecting calligraphers)
- Economic transformation:
- Gradual shift in professions, not mass unemployment
- New jobs we can't predict today
- Example: mobile app developer didn't exist 20 years ago
- Fundamental belief:
- People are fundamentally good
- Open source AI empowers human goodness
- Doomers don't trust people or institutions
