Interview, Conference Presentation
Yann LeCun: Human-Level Artificial Intelligence | AI Podcast Clips
- Building a sentient AI system resembling "Her" remains distant with no specified timeframe, as researchers face an unknown number of potential obstacles that could number "50 mountains" or more beyond the current visible "first peak."
- Future autonomous systems are projected to require a "predictive model of the world" capable of representing uncertainty in "three-dimensional continuous spaces," a capability that currently lacks a known representation method.
- Development plans involve creating self-supervised learning systems that mimic human and animal developmental trajectories, targeting milestones such as distinguishing animate from inanimate objects at "two, three months" and understanding gravity around "eight or nine months."
- A proposed system architecture integrates three core modules: a model of the world, an objective predictor, and a policy network to optimize actions, utilizing model predictive control by testing sequences of hypotheses to observe results.
- Behavioral drive is expected to rely on an objective function rooted in biological analogies to the "basal ganglia" to compute contentment or miscontentment, enabling systems to predict negative outcomes like burning their hands to avoid harm.
- Failure modes for autonomous systems are categorized into three distinct types: possessing a "wrong model of the world," maintaining an "objective not aligned with actual goals," or failing to "figure out a course of action" despite having accurate models and objectives.
- Past research is characterized as "overly optimistic" due to premature focus on the initial difficulty peak, while future progress depends on overcoming the significant unknowns of multi-stage obstacles and advanced spatial representation.