Why AI is incredibly smart -- and shockingly stupid | Yejin Choi

Episode Summary

Executive Summary: Yejin Choi argues that AI should be made smaller and safer, not just larger. She warns that extreme-scale models concentrate power, obscure inspection, and still fail basic common-sense reasoning. Her solution is to prioritize open, transparent data, human norms, and new algorithms that teach AI to understand the world more directly.

Main Topics: Critique of extreme-scale AI (Priority: 5/5): Choi frames today’s large language models as powerful but overgrown systems whose size creates cost, opacity, and safety risks rather than solving core intelligence problems. Common sense as the central missing capability (Priority: 5/5): She argues that robust common sense is still a moonshot for AI and is essential for safe behavior, because models can ace tests yet fail trivial reasoning tasks. Societal risks of concentration and opacity (Priority: 4/5): Because only a few companies can train frontier models, power becomes centralized and outside researchers cannot adequately inspect or audit them. Open, inspectable data for AI safety (Priority: 4/5): Choi advocates for publicly available common-sense knowledge graphs and moral norm repositories so AI training content can be reviewed, corrected, and diversified. Algorithmic alternatives to brute-force scaling (Priority: 5/5): She proposes symbolic knowledge distillation and other methods that directly teach world understanding, producing smaller models with inspectable representations. Human learning as a better model for AI (Priority: 3/5): In the Q&A, she emphasizes that children learn by forming hypotheses, experimenting, and abstracting world rules—not merely by consuming more text.

Key Arguments: Extreme-scale AI is expensive, environmentally costly, and accessible only to a few companies, which concentrates power. Large language models can appear intelligent while still making obvious common-sense errors, showing that scale alone is insufficient. Common sense is not a solved problem; it remains a major research challenge despite recent progress. Training on raw web data is inadequate because it contains racism, sexism, and misinformation. AI should be taught human norms and values through open, inspectable datasets rather than proprietary black-box systems. New algorithms should aim for direct knowledge acquisition and symbolic representations, not just next-word prediction. There may be a 'Goldilocks zone' of scale, but beyond that, progress likely requires different methods, not just bigger models.

Data Points: AI training scale: tens of thousands of GPUs - Choi describes recent extreme-scale language models as being trained on massive compute clusters. Training corpus size: a trillion words - She cites speculation about the amount of text used to train the newest large language models. Universe composition analogy: 5% normal matter / 95% dark matter and dark energy - Used to explain how visible text is only a small part of the knowledge needed for language understanding. Common-sense benchmark example: 5 clothes dry in 5 hours; GPT-4 says 30 clothes take 30 hours - Illustrates failure to reason about proportionality and shared drying conditions. Common-sense benchmark example: 12-liter jug and 6-liter jug; GPT-4 gives an incorrect multi-step answer - Used to show nonsensical reasoning on a basic measurement task. Common-sense benchmark example: bridge suspended over nails, screws, and broken glass - GPT-4 incorrectly predicts a flat tire, showing failure to reason about physical separation. Universe composition analogy: 95% invisible - Dark matter analogy for unspoken rules and common sense in language use.

Pivotal Quotes: "AI today is like a Goliath. It is literally very, very large." — Yejin Choi: Opening critique of extreme-scale models and the scale-first mindset. "We need to make AI smaller to democratize it, and we need to make AI safer by teaching human norms and values." — Yejin Choi: Core thesis of the talk, summarizing her proposed direction for AI development. "You don't reach to the moon by making the tallest building in the world one inch taller at a time." — Yejin Choi: Metaphor used to argue that brute-force scaling is not the right path to robust common sense.

Implications: The talk pushes AI builders toward transparency, open datasets, and new learning methods that prioritize common sense and human values. For listeners, it suggests that bigger models are not automatically better or safer, and that governance and research diversity matter.