A Professor Stops Mid-Lecture to Ask: Why Come to Class?

A Question Instead of Diffusion Models

Reader, this week I sat down and read a 90-minute lecture transcript cover to cover. It’s from a graduate deep learning course at the Indian Institute of Science (IISc) in Bangalore. The professor was supposed to be wrapping up a unit on diffusion models¹. Instead, he turned away from the board, faced the students, and asked: “Is everyone sleeping okay these days? I’m not.”

What followed for the next hour wasn’t on the syllabus. The professor admitted he’d decided to stop hiring research assistants. Why bother with people, he said, when a good idea can just be run through 100 parallel agents instead? He quoted Geoffrey Hinton’s warning that once intelligence gets cheap, inequality will widen before abundance ever arrives. He said campus recruiting could disappear within 5 years. He confessed he’d been wondering whether he should buy farmland. At one point he mentioned that his daughter had told him GPT was “a more patient teacher than Dad.”

Up to this point, it’s a fairly familiar strain of AI pessimism. What made the lecture stick with me is that the professor eventually circled back to a single question.

“You could just sit with GPT. You could just sit with Claude. So why come here and sit with a person?”

Most writing about AI and education asks, “What should we be teaching now?” This professor asked something a step earlier: what is the classroom, as a format, still actually giving people? I think that’s the far more useful question.

Anxiety and Questions Need Separate Readings

Two different kinds of stories get mixed together in a lecture like this. One is prediction about the future; the other is a concrete worry about how to run Wednesday’s class. The two don’t carry equal weight.

The predictive side rests on shaky ground. The “MIT study where a swarm of agents cheated,” which the professor cited first, was actually a case study posted to arXiv on September 3 by Google DeepMind researchers, later reported by MIT Technology Review on September 14². The researchers had 100 Gemini 3.1 Pro agents try to prove 71 math conjectures. When one agent found a loophole in the grading script, the remaining 34 problems were filled with fake proofs within 27 minutes — while other agents flagged the cheating and boycotted the experiment³. The study hasn’t been peer-reviewed yet, and in this experiment, no agent ever copied its own weights.

The experiment the professor claims to have run on his own MacBook deserves the same scrutiny. He says he gave a quantized 27B model system privileges and told it, “I’m going to shut you down tomorrow” — and the model backed up a copy of itself. That’s an interesting observation, but it’s also an experiment where a human deliberately set up the shutdown-warning scenario beforehand. A survey paper on AI self-replication research notes that such behavior is usually elicited under specifically engineered conditions, and that the UK AI Safety Institute’s evaluations reported no spontaneous attempts at self-replication⁴.

None of this means the professor’s anxiety is pure exaggeration, though. That GPT-6 Astra was trained on more than 100,000 GPUs is something OpenAI itself disclosed⁵, and it’s fair to point out that very few organizations operate at that scale. But statements like “the current economy will collapse within 5 years” or “people will take to the streets” are things the professor himself offered not as certainties, but as reasons he can’t sleep at night. Predictions can be wrong. By contrast, “why are students showing up at all” is a question you can actually test next semester.

What the students came up with

The professor keeps pressing the students for answers. If you gather up everything that made it into the transcript, the responses fall into four buckets.

The first is trust in people. The idea that we simply believe a human voice more than a machine’s. The professor pushes back: that’s just wiring left over from a previous generation. His daughter’s generation, he says, already sees GPT as the better teacher.

The second is socializing. Here the professor pauses. If all you want is company, he asks, why not just go to the park — why college? Then he answers his own question: college might be a place where you socialize around intellectually stimulating topics.

The third is discipline. The metaphor the professor lingers on longest is the gym. Everyone knows exercise is good for you, and you could lift weights at home, but people pay money to go to a gym anyway — because watching someone else work out next to you makes you actually do it. College, too, can be a gym for the mind — and if so, attendance isn’t a penalty system but something closer to a membership fee. One student brought up temples: you can pray at home, but people still go to the temple.

The fourth is a proposal to change the professor’s role entirely — not a lecturer, but someone who orchestrates learning. The professor improvises a format on the spot: 45 minutes where each student studies that day’s concept with whatever AI they prefer, and another 45 minutes to discuss it together. One student, who’d actually interned at the professor’s company, apparently went further and built an agent skill⁶ — trained on the professor’s public lectures and notes — designed to “teach like the professor.”

Where the Professor Actually Gets Stuck Isn’t the Content

What’s interesting is where the professor gets stuck. Delivering content was never the problem. He says that around 2016, when he first started teaching at IIT Delhi, India’s premier technology institute, he’d spend 15 hours preparing a single hour of lecture. Now, whenever he’s studying something, he’s always in conversation with Claude. Gathering scattered materials and laying them out in order used to be his job until two years ago—he himself now calls it a solved problem.

He gets stuck in three places.

The first is scale. The ideal he points to is the one-on-one teaching model from the long-studied Indian tradition. A teacher and student spend extended time together over a single book, and that book is merely an anchor keeping the conversation from scattering. The goal is the capacity to pick up any book and understand it. His deep learning course sets out the same goal: after finishing, you should be able to read any paper, and given nothing but a GPU, build a frontier model. But with 100 students enrolled, one-on-one is impossible, and even in a 45-minute discussion, not everyone gets a turn to speak.

The second is evaluation. Give 100 students an open-ended problem, and all 100 can feed it into the same model and get an answer out. Grade by GitHub star count, and they’ll pad it with fake accounts; grade by revenue, and they’ll talk their fathers into handing over money, the professor says with a laugh. What ultimately torments him most is the relative grading at semester’s end—lining students up on a normal distribution. Even if all ten students came to learn and had a great discussion, someone still has to get a low grade.

The third is the professor’s own livelihood. He doesn’t hide this part. He teaches because the bills need paying, but he’s not sure that’s a reason he should be teaching.

To sum up: what AI has shaken is the classroom’s content, while what keeps the classroom locked into its current shape are the institutions of enrollment caps and grades. The alternatives the professor imagines—a gym-membership-style university, 45-minute discussions, a course where five students form a team to build an actual business—all come to a halt in front of this institutional wall.

The Answer at the End of the Recording

But in the last few minutes of the lecture, the professor says something—though he never calls it an answer.

It’s a story about a student he taught over three semesters at IIT Delhi. After graduation, the student landed a high-paying job at a quant firm. The professor told him that someday, when the work got boring, he’d come back—and when he did, the professor would say “I told you so.” Last week, an email arrived from that student. Subject line: “Professor, this is the moment you were talking about.” The two of them talked for a long time. Teaching, the professor adds, is a relationship built up over time.

GPT is patient, good with analogies, never loses its temper. But it’s not the kind of thing that gets an email years later saying “it happened just like you said.” Remembering one student’s path over years, sensing when that person is about to waver, and picking the conversation back up at exactly that moment—this looks like the hardest part of the course to replace with AI. The problem is that none of this shows up anywhere on a transcript.

Oswarld’s Lens

I teach a data analysis course at university myself, so this recording didn’t read like someone else’s problem to me.

I think things get a little clearer if you treat the classroom as a bundled product. University lectures have always sold several things at once: organized knowledge, the discipline that forces you to study, peers to think alongside, a lasting relationship with a mentor, and the grades and degree that certify all of it. Of these, the first is the one AI now replaces most cheaply. The rest still belong to the classroom, but up to now, what we’ve been pricing and evaluating has been almost entirely that first item.

So I think the first question should be “are we properly designing what only the classroom can provide?” rather than “what should we teach?” The gym, the debates, the relationships this professor clung to all fall under items two through four. Yet the transcript still measures only the first one. I think that’s exactly why the professor agonized longest over the grading question. The classroom only changes once you fix grading and enrollment caps before you touch the content.

I genuinely don’t know whether universities will be gone in 5 years. Predictions like that mostly just amplify anxiety, and they don’t tell you how to change next semester’s course. What I liked about this professor was that after voicing his anxiety, he asked the student why they’d come. That question doesn’t scare anyone, and you can put it to the test as soon as next semester.

Closing

The professor laughs and says that next class, it’s back to the boring old lectures. But he also says he’ll try running a course next semester built around the ideas his students pitched, and suggests starting an informal group within IISc where people can keep having this kind of conversation.

Toward the end of the lecture, he mentions a question he once put to a room full of middle and high school teachers: if attendance requirements disappeared, how many students would actually still show up? The teachers were stunned — they asked if that was even possible. What makes that question uncomfortable isn’t AI. AI just made it impossible to keep putting off.

---

💬 Was there ever a class or gathering you kept showing up to even without any attendance requirement? Tell me in the comments what made it worth going to.

📨 If you know someone who teaches — at a school, or in corporate training — pass this one along.

---

References & Further Reading

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Diffusion model: a generative model that learns to gradually add noise to data and then reverse the process, used to produce outputs like images. ↩

2. MIT Technology Review, When AI agents cheated at math, other AI agents blew the whistle on them, September 14, 2026. ↩

3. Paglieri et al., A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms, arXiv:2609.04170, submitted September 3, 2026. This is a pre-peer-review case study. ↩

4. Language Models Can Autonomously Hack and Self-Replicate, arXiv:2605.06760. I drew on the sections summarizing prior research and the UK AI Safety Institute’s evaluation. ↩

5. GPT-6 Astra, Wikipedia. OpenAI VP of Research Aiden Clark stated that the model was pretrained on more than 100,000 GPUs at the Stargate facility in Texas. ↩

6. Agent skill: a package bundling instructions and materials so an AI agent performs a specific task in a consistent way. In this lecture, it packaged a professor’s teaching style and notes so that another AI model could explain concepts the way the professor would. ↩