Interview, Fireside Chat
Daniel Kahneman: Deep Learning (System 1 and System 2) | AI Podcast Clips
- Current deep learning systems function as "System 1" processes: they are highly predictive and pattern-matching based but lack "System 2" capabilities, specifically reasoning, causality, and the ability to represent meaning or real interaction.
- Major limitation identified: Deep learning models cannot effectively reason or represent causality without significant architectural transformation, leading to a consensus among experts that current neural networks will eventually hit capability limits.
- DeepMind transitioned from solving Chess to Go, and then to AlphaZero, at a speed described as "bewildering" and faster than anticipated.
- Gary Marcus highlights a fundamental gap between human and machine learning: humans learn quickly from few examples (e.g., children require two or three), whereas current machines lack built-in expectations to enable similar rapid learning.
- Demis Hassabis (DeepMind) and Yann LeCun (OpenAI) are attempting to integrate reasoning and knowledge assembly into neural networks, though temporal causality remains largely out of reach.
- Yann LeCun contends that current accomplishments are not the ultimate solution, utilizing the metaphor that the field has only "seen one or two mountain peaks" rather than the full landscape of what is possible.
- A consensus view suggests that without "grounding" (sensory input, physical interaction, or a body), AI systems lack true understanding of the world and treat language as unmoored symbols.
- Grounding requires either a physical body or a robust perceptual system that allows the machine to accumulate knowledge through active interaction and action.
- Current autonomous vehicle systems treat pedestrians as static obstacles to be avoided rather than agents to be engaged with in a game-theoretic interaction.
- Predicting pedestrian behavior involves complex non-verbal signals, such as eye contact when crossing and looking away to signal commitment, which machines currently struggle to interpret without a model of human mortality or intent.
- Autonomous vehicle perception is aided by a "system multiplier" effect where data from all vehicles feeds a collective learning system, improving the ability to anticipate human actions in traffic.
- The speaker notes that while anticipating outcomes (like in Go) does not require understanding the opponent, accurately navigating real-world human interactions likely requires a model of the human mind and social dynamics.