Interview, Conference Presentation
Yann LeCun: Human-Level Artificial Intelligence | AI Podcast Clips
- Building an AI system with human-level intelligence, such as the fictional entity "Her," remains a distant goal characterized by multiple unseen challenges, metaphorically described as a series of mountains where only the first peak is currently visible.
- Historical optimism in AI, exemplified by Newell and Simon's General Problem Solver, was misplaced because it failed to account for the exponential complexity of real-world problems.
- The primary immediate obstacle for achieving this intelligence is self-supervised learning, where machines must learn world models through observation and interaction, mimicking the developmental trajectory of infants and young animals.
- Cognitive science research by Emmanuelle Dupoux indicates specific developmental timelines for concept acquisition:
- Babies distinguish between animate and inanimate objects at 2–3 months.
- Infants understand object stability around 4 months.
- Conceptual understanding of gravity develops at 8–9 months, often demonstrated by the active experimental behavior of throwing objects to test physical laws.
- An autonomous intelligent system requires a predictive model of the world to forecast state changes (from time t to t+1) resulting from specific actions, though representing uncertainty as distributions in 3D continuous spaces remains a technical limitation.
- System autonomy relies on three distinct functional components:
- A model of the world to simulate future states.
- An objective function rooted in human biological drives, specifically the basal ganglia's computation of "contentment" or "discontent."
- A policy network designed to calculate the optimal sequence of actions to maximize the objective based on the world model.
- Autonomous failure occurs in three specific scenarios:
- An incorrect or flawed model of the world.
- An objective function misaligned with the actual desired outcome (analogous to psychopathy in humans).
- Possession of an accurate model and objective but an inability to compute the optimal course of action (execution failure).