#135 – Charles Isbell: Computing, Interactive AI, and Race in America — Summary & Takeaways

This Issue's Overview

Charles Isbell is the Dean of the College of Computing at Georgia Tech, an AI researcher and educator. This episode explores the essence of computing (the trinity of models, language, and machines), the direction of interactive AI, and the structural predicament of race in American society. The most consequential judgment in the entire episode: Charles Isbell argues that "the core of computing is not technology, but a way of thinking—understanding that models, language, and machines are equivalent, and that this equivalence is truly meaningful only when humans participate in it."

1. The Essence of Computing: The Trinity of Model, Language, and Machine

Charles Isbell argues that what distinguishes computing from other disciplines is that "model, language, and machine are equivalent" — this is both the unique way of thinking in computing and the key message that education should convey.

Discipline as a Way of Thinking: Different disciplines impart different "mindsets" — science is discovery and empirical methods, mathematics is abstraction and stable truth, engineering is trade-offs and construction. The core of computing is understanding that "model, language, and machine are the same thing" — the equivalence of regular expressions and finite state automata is no coincidence; it permeates all computing activities, including debugging.
Dynamism Distinguishes It from Mathematics: Mathematicians pursue static truth ("Truth is a thing, not an ongoing process"), whereas models in computing are "executable and dynamic" — the system itself can reflect on itself, which constitutes the unique contribution of computing.
Humans Sit Inside the Triangle: Isbell emphasizes that the above trinity "matters only because there is a human inside the triangle". The value of computing lies in "human interaction with it" — data and computing power exist, but "the interesting and meaningful parts exist only relative to humans". The meaning of data and information is assigned by humans, so curriculum design must convey both "tools and skills" and the premise that "humans are in it".

Deduction: Isbell believes that computing is "permeating everything" — by 2030, a history PhD will also need to understand data science, because "the way history will be studied is through analyzing data", and the same applies to psychology and philosophy. But this also brings a risk: computing may "become everything and lose itself", just as "teaching engineers Fortran" loses the core of the discipline.

2. Interactive AI: Intelligence Is Not Isolated, but Coexistent with Others

Isbell argues that "intelligence" in isolation is meaningless — "I don't care if a tree falls in the forest with no one around to hear it, because I don't think it matters"; true intelligence lies in "intelligence in interaction with others."

Why interactive AI matters: Humans are innately wired to "look for intent where there is none" — animism is a human instinct. Meanwhile, interactive AI enables "faster learning" because you can "import experience from history" and "transfer it efficiently." As an individual, you are "trying to move the species forward together with a larger group."
Core challenge: adapting to a changing counterpart: Interactive AI is not just solving a single task, but must "continuously adapt over a long period." The counterpart is different from oneself, and "you are different from yourself 15 minutes ago, or even 15 years ago" — the system needs a "drift adaptation mechanism," assuming that the user will drift over time (preferably without discontinuities).
Why lifelong learning is hard: Isbell admits that the machine learning community "doesn't spend much time on lifelong learning" because "everyone is overfitting to specific tasks." Even those who work on "transfer learning" end up "looking for the keys under the lamppost because that's where the light is" — because the incentives (papers, products) all point to incremental progress. He suggests: "At least go from N=1 to N=2 or N=7," and "be willing to deploy a system, let it live in the messy world for months, accept that it needs five years — instead of repeatedly running the same experiment and tuning the machine a little better."

Isbell's response to the "GPT-3 brute-force approach": He believes "it won't succeed that easily." Take Google as an example — Google did not solve the information retrieval problem; it "changed the problem": from "finding relevant answers" to "minimizing false positives on the first page." Because "when there are 10 million answers, the problem is not to give relevant answers, but to avoid giving irrelevant ones." "You train yourself to learn which keywords bring you to that page," but the problem itself has been changed.

3. Predictability and Counter-Intuition: Humans Are Highly Predictable but Hate Hearing It

Isbell found through infrared data experiments at his own home: human behavior is highly predictable, but "people hate being told they are predictable."

Experiment Method and Data: He recorded all infrared remote control signals in his home (TV, lights, etc.) and used "simple counting statistics" to predict the next button press, achieving 93% accuracy. When predicting by "group" (the overall behavior of a set of buttons), accuracy rose to 99%. He found that the behavior patterns of "number keys" were similar, so they were automatically clustered—"number keys have similar behavior for you, so you naturally cluster them together."
Individual Differences and Clustering: Each person is different, but "any given individual is extremely predictable because you repeatedly do the same things." Even if different individuals vary, they still "act like each other"—a "10% outlier" behaves similarly to the other 10% of people.
True "Humanity" Lies in Anomalies: Isbell agrees that "when faced with a situation never seen before, or when forced to make a difficult decision, that is what defines you"—"not simple problems, but things that hurt you. You know the outcome will be highly suboptimal, but you still do it."

Deduction: The implicit falsification condition here is—if the experiment were conducted with larger-scale, longer-duration data, would it reveal that the "unpredictable part of humans" is larger than Isbell believes? But he explicitly says, "two days of data are enough (as long as you choose the right two days, e.g., weekends)."

4. University Rankings, Reputation Systems, and Structural Problems: From "Minimizing False Positives" to "Structural Inertia"

Isbell uses university rankings and hiring as examples to illustrate how the system perpetuates inequality by "minimizing false positives."

Rankings are "100% reputation": The US News computer science ranking is entirely based on reputation surveys ("two people in each department give scores to everyone"). This means "how to improve the ranking" is "how to improve reputation"—which requires "demonstrating leadership in times of crisis." Georgia Tech moving from outside the top 20 into the top 10 is the only case, relying on visible, impactful initiatives such as "online master's programs, undergraduate education reforms."
The "IBM effect" in hiring: "No one ever lost their job for buying an IBM computer," and similarly "no one ever lost their job for hiring a PhD from MIT." If that person performs poorly, "he came from MIT, what more could you expect?" Isbell's data shows: In 2017, about 60% of faculty PhDs in the top 4 US computer science departments came from the top 4; about 65% in the top 10 departments came from the top 10; about 50% in the top 55 departments came from the top 10. This "cannot mean that all the best professors come from 10 schools"—it merely reflects the systemic pressure to "minimize false positives."
Consequences: "The decision for you to become a Cornell professor is made when you are 17 years old—where you went for undergraduate studies." Isbell believes the solution is to "expand the pool" or "change the loss function"—allowing the system to "evaluate itself using broader metrics than it currently does."

Implications: Isbell acknowledges that "injecting randomness" is theoretically feasible, but in practice "the cost is too high, it would ruin people's lives." He suggests that real change requires "structural" adjustments—just as "changing the US News formula would cause the entire university to change its behavior."

5. Race, Structural Predicament, and the Mechanism of "Empathy"

Isbell argues that the race issue is essentially a "structural" problem — people are placed into different "clusters," using the same words but referring to different things, and therefore "talk past each other."

Separation of language and perspective: He describes an example — two groups view the same room from different angles; one group sees a clock, the other does not, so they use different names to refer to the same room. When they come back together to discuss, "they don't know they're talking about the same room" — "the problem isn't different ideologies, it's that you use the same words but refer to completely different things."
Empathy ≠ sympathy: Isbell distinguishes between empathy (understanding where the other person is coming from and how they feel) and sympathy (feeling sorry for the other person). He believes that "AI can find the overlap between two people's distributions, and then introduce you based on that common ground, making it easier for you to take the step toward empathy." But "the harder part is getting both sides to be willing to ask the question" — "maybe what AI should do is persuade you to ask 'Is there more similarity between us than difference?' rather than directly telling you the answer."
Police, violence, and structural narratives: Isbell describes his experience of being pulled over at gunpoint by police — "the worst feeling wasn't that he might shoot, but that if he did, he would get away with it." He attributes this to "a huge narrative structure that makes it easy for people to put themselves into different jars and forget that people in other jars are essentially the same as themselves." The solution is to "build structures that allow people to always talk safely and understand that others have different experiences, but in the end, we are the same."
Historical perspective: Martin Luther King's early failure in Georgia was because "the sheriff loaded everyone into trucks, took them far away, and locked them up; the media had no news to film, and nothing changed." In Birmingham, "Bull Connor used fire hoses on children; the media captured it, outrage erupted, and change happened." Isbell believes that "violence is the voice of the unheard," but it "needs to be structurally exposed" — "if television hadn't existed, the civil rights movement might not have happened, or would have taken much longer."

Inference: Isbell's answer to "why 2020 is now" is "enough time has passed, a new generation is angry enough, but doesn't remember what happened last time." He implies that structural change requires "leaders selling an illusion of optimism," but currently "a non-zero probability of violent civil unrest" is real.

Mentioned Targets

This section contains no investment-related targets. The podcast content is academic and personal opinions, and does not involve investable companies or assets.

Judgments Worth Remembering

1. "The essence of computing is the equivalence of models, language, and machines" — Isbell argues that this is the unique way of thinking that distinguishes computing from other disciplines, and this equivalence is truly meaningful only when humans are involved. (Support: The equivalence between regular expressions and finite state automata "is not trivial," and it permeates all computing activities.)

2. "Humans are very predictable — two days of data can achieve 93% prediction accuracy — but people hate being told that." (Support: Isbell's experiment with infrared remote controls in his home; simple counting statistics can reach 93%; set-based prediction can reach 99%.)

3. "The core of interactive AI is not solving a single task, but long-term adaptation to a person who changes over time, and that person is also different from others." (Support: Isbell believes that "lifelong learning" requires a system to "survive for months in a chaotic world"—academia lacks incentives to do this.)

4. "GPT-3's 'brute force' approach will not solve AI problems — it will only change the problem, just like Google changed information retrieval." (Support: Google's problem of "minimizing false positives" differs from the original "finding relevant answers." Isbell argues that "you change the problem, not solve it.")

5. "University rankings are 100% reputation — a positive feedback loop, and your fate as a professor is decided when you are 17 years old." (Support: 2017 data shows that 60% of professors in the top 4 departments come from the top 4; 65% in the top 10 come from the top 10. Isbell believes this "cannot be real" and is merely a product of "minimizing false positives.")

6. "Empathy is not sympathy — understanding where the other person is coming from does not mean you agree with him; AI can help you find common ground, but the harder part is getting you willing to ask the question." (Support: Isbell distinguishes empathy from sympathy, arguing that "people in the system always use the same words but refer to different things." AI can "find overlap between distributions," but "it requires you to be willing to ask questions first.")

7. "The racial issue is fundamentally structural — it is not solved by individual conversations, but requires building safe structures for people to keep talking." (Support: Isbell uses the examples of Martin Luther King's failure in Georgia and success in Birmingham to illustrate that "exposure" and "visibility" are key to structural change.)

8. "Death is probably like general anesthesia — no sense of time passing, so you wouldn't notice it; but I don't like the idea, I'd rather have the chance to choose whether immortality is really that bad." (Support: Isbell describes his experience during thyroid surgery where "5 hours disappeared in an instant," arguing that "in reinforcement learning, because of the discount factor, finiteness is what makes everything meaningful.")