248 | Yejin Choi on AI and Common Sense

Episode Summary

Executive Summary: Sean Carroll and Yejin Choi discuss how large language models work, why they are impressive yet brittle, and why AI may be in a rapid “phase transition.” The conversation centers on word vectors, transformers, common sense, creativity, symbolic reasoning, education, misinformation, and the limits of current AI compared with human cognition.

Main Topics: AI as a phase transition (Priority: 5/5): Carroll frames current AI progress as a sudden, rapid shift rather than a slow evolution, arguing that society is already seeing major behavior changes in how people use language models. How transformers and word vectors work (Priority: 5/5): Choi explains that modern LLMs represent words as continuous vectors and use attention within transformer architectures to predict the next token, enabling fluent language generation at scale. Common sense and symbolic reasoning gaps (Priority: 5/5): A major theme is that LLMs can appear intelligent yet fail on basic common-sense or symbolic tasks, suggesting the need for additional modules or hybrid architectures. Creativity, interpolation, and human uniqueness (Priority: 4/5): The discussion contrasts LLM remixing/interpolation with human creative extrapolation, with Choi arguing that current models can imitate but not reliably originate in the way exceptional humans do. Education, overreliance, and changing workflows (Priority: 4/5): The speakers debate whether AI should be treated like a calculator or search engine, versus a tool that may undermine learning if students outsource too much thinking to it. Misinformation, deepfakes, and platform solutions (Priority: 5/5): Both speakers worry about AI accelerating misinformation, but Choi argues that detection alone is insufficient and that broader platform-level verification and AI literacy are needed. Human cognition as a model and a contrast (Priority: 4/5): The conversation repeatedly returns to the idea that AI may illuminate human habits, biases, templated reasoning, curiosity, and even cognitive dissonance, while still lacking embodiment and internal motivation.

Key Arguments: Current LLMs are trained primarily to predict the next word/token, yet can produce long, coherent text because scale and attention make local prediction look like planning. Word vectors were a breakthrough because meaning depends on context; vector representations allow analogies and similarity-based reasoning that discrete word IDs could not capture. LLMs often perform interpolation over seen examples rather than true extrapolation, which is why they can sound creative while failing in unfamiliar situations. Common sense remains a major weakness because LLMs lack robust world models, theory of mind, and reliable handling of counterfactual or unusual scenarios. A hybrid approach may be necessary: neural networks plus symbolic reasoning or other modules, since purely monolithic transformers are not enough for certain tasks. Human creativity may remain superior in the foreseeable future because people can intentionally break from prior examples, ask questions, and pursue novel goals with motivation and curiosity. AI misuse is not entirely new—misinformation existed before—but AI can scale it faster, making AI literacy and verification infrastructure increasingly important. Automatic fake-news detection is politically and technically hard; labels like “true” and “false” often require social consensus, not just model inference.

Data Points: Context window size: 82,000 tokens - Choi says the largest available context size for current models is about 82,000 tokens. Child theory-of-mind development age: 4 or 5 years old - Choi notes that children acquire basic theory-of-mind reasoning by around age four or five. Time reference for classroom AI change: 1 year - Carroll says that about a year earlier, his students were not likely to use AI for papers, but now it is almost inevitable. Model comparison: GPT-4 - Choi says GPT-4 became much better at common-sense questions than earlier ChatGPT versions.

Pivotal Quotes: "This is the time to be open, to watch things develop, to imagine what could happen, but not try to be too definite about what will happen until it actually is so that you can correctly adapt to this brave new world that we're entering in." — Sean Carroll: Carroll frames the AI moment as uncertain and fast-moving, urging intellectual openness rather than dogmatism. "We might need something more modularized, but at the same time, messier in some sense, broadly speaking for this to go to the next level." — Yejin Choi: Choi explains why current transformer models may need hybrid or modular additions to improve common sense and reasoning. "I do think that humans have this capability to push further. Good. In some ways." — Yejin Choi: Choi contrasts human creativity with LLMs' tendency to remix prior patterns rather than truly originate.

Implications: AI is already changing education, media, and public trust. Listeners should expect powerful but brittle tools, verify outputs carefully, and anticipate hybrid systems plus stronger governance as AI use expands.