Isbell, Charles

Dean of the College of Computing, Georgia Tech

Charles Isbell and Michael Littman: Machine Learning and Education

Lex Fridman Podcast · Dec 26, 2020

• Charles Isbell
• Michael Littman
• machine learning versus statistics
• ICML versus NeurIPS culture
• productive struggle in education
• Bell Labs and AT&T Labs history
• Georgia Tech OMSCS
• college experience versus content delivery
• connection in remote learning
• beginner programming pedagogy
• AI risk as mundane misuse not superintelligence

Core ideas

Isbell and Littman open with a long-running friendly disagreement over whether machine learning is "just" (computational) statistics. Littman argues machine learning is fundamentally a computational/programming discipline — concerned with debugging, software engineering, hyperparameters, and system-building — more than a statistical one, comparing the relationship to how chemistry is "just" physics but is more usefully treated as its own level of abstraction. Isbell partly agrees but pushes toward a framing centered on rules and symbols, and both illustrate the distinction historically by contrasting the "computer science" culture of ICML in the late 1990s/early 2000s with NeurIPS's more statistics-oriented, "trying to impress statisticians" culture, arguing this cultural difference shaped what kinds of questions each community asked and what outcomes it valued. They extend this into a discussion of how neural networks and "software 2.0" have shifted machine learning practice toward treating hyperparameter choices, data curation, and cleaning as core parts of "programming" rather than incidental to it.

On teaching, Isbell describes his (and Littman's) shared machine-learning course assignment design philosophy: students are given working implementations of algorithms (KNN, SVMs, boosting, decision trees, neural networks) for free and told not to bother re-implementing them, since the pedagogical point is to select and analyze two datasets that reveal meaningful differences between algorithms — forcing students to engage with the primacy of data over algorithmic mechanics, a lesson he says students often only appreciate well after the course ends, not during it. Both argue that productive "struggle" (as opposed to hopeless "suffering") is essential to real learning — echoing Georgia Tech's now-discontinued "drown-proofing" swim requirement and its "build tomorrow the night before" ethos — while cautioning that if learners lose all hope of eventually succeeding, the value of struggle collapses into demoralization. They trace their own personal and professional history: meeting during Littman's job interview at AT&T Labs (a descendant of the legendary, monopoly-subsidized Bell Labs, whose "pure research" culture and serendipitous hallway/whiteboard interactions both remember as irreplaceable), being laid off together and briefly building an automated calendar assistant under a boss (Ron Brachman) who later helped launch the DARPA program that led to Siri, and eventually reconnecting professionally when Isbell recruited Littman to co-teach Georgia Tech's Udacity-based Online Master's in Computer Science (OMSCS) — a partnership that produced the "Smoove and Curly" persona videos and a lecturer/student teaching format designed specifically to preserve the energy of live audience interaction that Isbell found himself missing when recording a solo MOOC.

The conversation turns to fictional AI narratives (Westworld, 2001: A Space Odyssey): both are skeptical that Hollywood-style "robot uprising" scenarios represent the real danger of AI, arguing instead that the actual risk lies in mundane, poorly-debugged systems using data to make consequential decisions more efficiently and thus more dangerously than before — a distinction they frame as the difference between "true intelligence surpassing us" (unlikely and distant) versus "tools making us more efficient at doing terrible things" (already happening). They briefly discuss simulation-hypothesis-style questions, both landing on skepticism that we live in a deliberately constructed simulation while finding the underlying computational framing of physical reality (e.g., the unintuitive nature of quantum mechanics) a genuinely interesting lens.

On the future of higher education — sharpened by COVID-19 — Isbell argues (drawing on a piece by journalist Ian Bogost) that the "disaggregation" thesis (that online tools would separate content delivery from physical campus attendance) fundamentally misunderstands what students are paying for: not classes, but the "college experience" — social belonging, independence, and a rite of passage — meaning remote learning won't replace campuses but will instead expand access for a different population (particularly older, working adult learners) rather than substituting for the traditional undergraduate experience. He cites Georgia Tech's Online Master's in Computer Science (OMSCS) as evidence: despite costing roughly $7,000 versus $46,000 on campus and admitting a majority of applicants, its graduates show unusually high loyalty (over 80 alumni volunteering as paid teaching assistants for near-minimum wage) because they organically built their own social structures (self-organized regional groups, shared identity markers like class T-shirts) to satisfy a need for belonging that the campus experience meets more automatically. Both express concern that credential inflation and unequal access to the "networked" campus experience could deepen social stratification even as online options expand overall access, and Isbell connects this to how university faculty incentives work: research and teaching pull against each other in tension, but framing teaching as tied to an institution's core "mission" — rather than treating it as a mere distraction from research — helps resolve this tension, illustrated by an alumni survey where the people alumni most credited with impacting their lives were teaching-focused lecturers, not high-profile research faculty.

They discuss "connection" as the actual variable driving classroom effectiveness rather than physical co-presence per se — noting internal survey data showing students report equal perceived learning across different class formats, but consistently prefer classes with a stronger sense of connection to instructors and peers, regardless of delivery medium — while acknowledging that shared, synchronous experience (the shared unspoken reactions of a room full of people, akin to a comedy audience or a concert crowd) does something distinct that pure content delivery does not replicate, without concluding this makes remote/asynchronous formats obsolete, drawing an analogy to how recorded music didn't eliminate concerts but changed their role. The conversation closes with career and pedagogical advice: both encourage pursuing genuine passion and tolerating uncertainty about long-term outcomes (Littman citing "choose your parents wisely" as a joking nod to the role of luck, and reflecting that his own career worked out well but wasn't chosen for practical reasons), and offer concrete beginner programming advice — start with small, well-scoped exercises rather than ambitious end projects, expect fundamental concepts like variable assignment (illustrated by a simple "x=3, y=4, y=x, x=7, what is y?" thought experiment revealing how spreadsheet-trained intuitions can create persistent confusion about references versus values) to be genuinely difficult and not a sign of personal inadequacy, and that the specific programming language matters far less than finding one with strong tutorials and community support to ease the on-ramp.

Excerpts