272 | Leslie Valiant on Learning and Educability in Computers and People
What this covers
Leslie Valiant, a computer scientist and Turing Award winner, joins Sean Carroll to argue that human distinctiveness rests not on intelligence—a term too murky to define rigorously—but on educability, a three-part computational framework. The conversation traces how humans (and machines, and evolution itself) learn through PAC Learning, the integration of knowledge across contexts, and the absorption of explicit theories transmitted by others. Valiant positions this framework as answering ancient philosophical puzzles in concrete, measurable terms and as the true engine driving both human cognition and artificial intelligence development.
The discussion ranges across immediate practical stakes and foundational science. Valiant emphasizes that PAC learning offers only weak guarantees for individual cases, making AI unreliable for safety-critical applications, and that educability itself is morally neutral—a capacity easily bent toward harmful ends, whether in humans or machines. He addresses how humans, lacking built-in verification for transmitted knowledge, fall prey to false and conspiracy theories, a vulnerability he traces deeper than surface problems like deepfakes. On the evolutionary side, he proposes that Evolution as Learning can be formulated as a learning process where mutation and survival play the roles of algorithm and feedback. He also contests confident claims about large language models, arguing they demonstrably do only token prediction and that attributing reasoning or understanding to them reflects observer generosity rather than proven capability. Throughout, he advocates for educability as a scientific concept with genuine robustness—and for education research grounded in theory rather than best practices.
Valiant argues that the defining feature of human uniqueness is not intelligence but 'educability'—the integration of learning from examples, chaining learned knowledge through error-tolerant reasoning, and absorbing others' explicit theories—and that this same computational framework (PAC learning) explains AI, evolution, and the human mind.
- Learning from examples (PAC learning) is the computationally feasible core capability shared by humans, machines, and evolution
- Educability adds chaining of learned knowledge and absorption of transmitted theories on top of basic learning
- Intelligence lacks a rigorous definition while educability can be explicitly defined and potentially measured
The host introduces the episode and its guest to the audience.
- The host introduces the episode and guest.
“Hello everyone. Welcome to the Mindscape Podcast.”