Melanie Mitchell: Artificial Intelligence and Analogy

The Nature of Artificial Intelligence

Melanie Mitchell examines the limitations of the current field of AI, arguing that while narrow AI has achieved remarkable success, true Artificial General Intelligence remains distant. She highlights that the field often over-relies on brute-force approaches while struggling to define what intelligence actually entails.

Core Cognitive Challenges

• Analogy and Concepts: Mitchell emphasizes that analogy-making is a fundamental pillar of human cognition. She asserts that without the ability to form and fluidly use concepts, high-level reasoning is impossible.
• Common Sense: A major open problem is the lack of common sense in current AI. Most human knowledge—such as intuitive physics and metaphysics—remains "invisible" and difficult to encode through traditional approaches.
• Mental Models: Mitchell advocates for moving toward active, generative models where agents possess internal simulations of the world, fostering deeper understanding rather than just pattern matching.

Future Directions and Philosophy

"How to form and fluidly use concepts is the most important open problem in AI."

• Innate Knowledge: Mitchell suggests that learning cannot happen in a vacuum; biological systems possess innate structures that allow them to learn, a factor often ignored in deep learning.
• Existential Risk: She expresses skepticism regarding existential threats from superintelligence in the near term, labeling them as distant concerns compared to immediate, pressing societal issues.
• Embodied Intelligence: There is a growing focus on the necessity of embodiment—the idea that intelligence cannot be fully separated from social interaction, emotion, and self-preservation.