272 | Leslie Valiant on Learning and Educability in Computers and People

Episode Summary

Executive Summary: Sean Carroll and Leslie Valiant discuss computational learning theory, especially PAC (“probably approximately correct”) learning, as a rigorous framework for understanding how systems generalize from data. Valiant argues that learning, reasoning, and human educability can be modeled scientifically, and proposes that human uniqueness lies less in intelligence than in the ability to learn, chain knowledge, and absorb explicit instruction. They also explore AI limits, education, evolution, and why current systems remain powerful but narrow.

Main Topics: PAC learning and computational learning theory (Priority: 5/5): Valiant explains his PAC framework for formalizing how learners generalize from examples and achieve high-probability success with manageable computational effort. Learning versus intelligence (Priority: 5/5): The conversation contrasts the vague, hard-to-define notion of intelligence with educability, which Valiant defines more precisely as a measurable capacity to learn, chain knowledge, and take instruction. AI, large language models, and the limits of current systems (Priority: 4/5): They discuss how LLMs excel at next-token prediction and smooth prose, but remain limited in reasoning unless learning is combined with robust reasoning architectures. Reasoning integrated with learning (Priority: 4/5): Valiant argues for probabilistic, error-tolerant reasoning that can be married to learning, rather than classical logic’s brittle certainty. Human uniqueness and education (Priority: 5/5): Valiant’s new theory of educability frames human distinctiveness as the integration of learning from experience, chaining ideas across contexts, and absorbing explicit theories from others. Evolution and natural learning (Priority: 3/5): The discussion extends PAC-style learning to Darwinian evolution, with survival acting as feedback and mutation as the learning mechanism. AI risk and societal impact (Priority: 4/5): Valiant downplays singularity-style fears but emphasizes that AI will become more capable in ways we understand, requiring caution similar to other dangerous technologies.

Key Arguments: PAC learning captures the core scientific problem of generalization: a learner must do well on future examples with high probability, while remaining computationally feasible. Machine learning is broader than neural networks; neural nets are just one class of algorithms within a much larger field. The phrase "probably approximately correct" is a useful epistemic lesson: we should expect imperfect but statistically reliable predictions, not certainty. The hardest problems are not just computationally difficult but may also be hard to learn from data; cryptography is a deliberate example of making something hard to learn. Human educability is more meaningful than intelligence because intelligence lacks a clear definition, while educability can be decomposed into learn-from-experience, chain-knowledge, and accept-explicit-teaching. Large language models are optimized for next-token prediction, not proven reasoning; their apparent reasoning abilities should not be assumed without evidence. Reasoning should be redesigned to tolerate uncertainty and probability so it can work with learning rather than against it. Evolution can be interpreted as a learning process where survival serves as feedback and mutation as the adaptive mechanism. Education research should become more scientific, using measurable concepts like educability rather than relying mostly on vague notions of intelligence or best practices. AI’s biggest near-term impact will likely be a mixed human-machine economy, not a mystical singularity.

Data Points: PAC acronym: "probably approximately correct" - Valiant’s learning framework and book title Turing Award year: 2010 - Valiant’s recognition for theoretical computer science contributions Approximate age of possible human educability emergence: 300,000 years ago - Valiant speculates the capability may have emerged in a predecessor species Time scale of evolutionary spread mentioned: "the last tens of thousands of years" - He notes no evidence of a mutation spreading across humans in that period Training cost scale for large language models: "millions of dollars" - Used as evidence that modern systems require substantial computation to learn Evolutionary timescale reference: "100 million years ago" - Valiant contrasts ancient animal adaptation with human-specific educability Modeling term: polynomial time - Learning should be computationally feasible rather than intractable Performance curve shape: constant exponent / algebraic curve - Valiant says PAC learning yields error decreasing as a power law with effort

Pivotal Quotes: "“probably approximately correct”" — Sean Carroll: He highlights the phrase as the core intuition behind PAC learning "“The main downside of intelligence is that no one can define it.”" — Leslie Valiant: He contrasts vague intelligence with his more precise concept of educability "“The future is still using the same phenomenon of learning boxes, but probably in a system you’d have many boxes and you’d have some sort of reasoning capability.”" — Leslie Valiant: His view of how AI should evolve beyond standalone models

Implications: Listeners should think of AI and human cognition as probabilistic, trainable systems with real limits. The practical future is not superintelligence but more capable, carefully designed human-machine systems, while education should be studied as a measurable science of educability.