Melanie Mitchell: Concepts, Analogies, Common Sense & Future of AI

Lex Fridman opens the conversation by questioning the adequacy of the term 'artificial intelligence,' which Melanie Mitchell critiques as misleading. She suggests 'complex information processing' might better capture the essence of what current AI systems do. This setup frames the broader discussion on how language shapes our understanding of technology and its capabilities. Mitchell's insights into the historical context of AI terminology, including John McCarthy's regrets and Herbert Simon's rejected proposals, set the stage for a deeper exploration of what intelligence means in both human and machine contexts.

Mitchell argues that current AI systems, especially deep learning models, lack the ability to prioritize relevant features, which limits their understanding of fundamental concepts. She uses the example of DeepMind's Atari game-playing program, which could excel at Breakout but failed to adapt when the game was slightly modified. This illustrates the limitations of AI in transferring skills across different contexts, highlighting the need for systems that can understand and apply concepts more flexibly. Mitchell's point underscores the importance of developing AI that can form and use concepts fluidly, a critical step towards achieving human-like reasoning.

While Lex doesn't explicitly challenge Mitchell's critique of AI's limitations, the conversation naturally raises questions about the future trajectory of AI development. The discussion touches on the long tail problem in autonomous driving, where unexpected edge cases pose significant challenges. Mitchell's skepticism about deep learning's ability to achieve human-like understanding suggests a need for hybrid systems that incorporate generative models and analogy-making. This tension between current capabilities and future aspirations highlights the ongoing debate within the AI research community about the best path forward.

The conversation concludes with Mitchell emphasizing the importance of concepts and analogies in cognition, both for humans and machines. She argues that without mastering these, AI cannot achieve the level of common sense required for tasks like autonomous driving or complex decision-making. The discussion leaves open questions about how AI can evolve to better handle these challenges, suggesting that while existential threats from AI are distant, the immediate focus should be on improving AI's conceptual understanding. This pivot underscores the need for continued research into how machines can learn and apply concepts more effectively.