Reading · AI + Education = Simplified

About the Interview series

This series seeks to highlight the folks and their work of teaching and learning in the age of AI. If you have experiences around AI and education in higher education classrooms that you would like to share, consider being interviewed for this Substack.

A man faces the camera with a slight smile, wearing black rectangular glasses, a patterned blue-gray shirt, and a dark blazer. His long hair is pulled back, and a blurred bookshelf fills the background.

David Bachman is a Professor of Mathematics and Computer Science based in Claremont, California. He teaches regularly at Harvey Mudd College, Claremont McKenna College, and Pitzer College, where he holds tenure. He has published more than 20 research papers, written three books, and received two National Science Foundation grants supporting his work in topology. David also speaks frequently about artificial intelligence, leads workshops on AI-assisted software development for education, and writes about AI and the future of higher education in his Substack, Entropy Bonus.

Lance Eaton: What does your own AI use actually look like these days?

David Bachman: Right now my use really comes down to two main applications. One is as a writing partner. I write a weekly Substack, and I always write the first draft myself. I resist all urges to use any AI to write the first draft, except for some background research. So I write the first draft, and then I run it through either ChatGPT or Claude, often both. It’s interesting to see the different kinds of feedback. I’ll revise according to one’s suggestions and then send that revision through the other. As a writing partner, I find it really useful.

The other use, since January or early February, is that I’ve gone full-on board with agentic coding. And I’m really thinking about it in the context of education — how to leverage what they’re calling “vibe coding,” (but I hate that term, so I’ll just call it AI-assisted coding) — to create educational experiences tailored for my students and my classroom. Either standalone apps where people get in and interact with the content, or modules to go along with classes I’m planning to teach in the fall. I mostly use ChatGPT’s Codex, and I’ve been making tons of web apps that students log onto and engage with the material in different ways.

Lance: I was thinking about your recent post on this both/and exploration around AI. I get it, because it’s something I’ve been sitting with too, but I’m curious to hear from you: how do you hold the different views about AI at once: its value and its problems? How do you find that balance in your work, your views, and your conversations with others?

David: The analogy I made in that post was driving. When you drive, I think everyone is aware, at least subconsciously, that cars are really problematic. They’re bad for the environment, bad for noise, bad for your mental health sitting in traffic, and they kill people on freeways. But people don’t sit and weigh the pros and cons every time they get in the car. So to answer your question: I know these things on a subconscious level, but on a conscious level I just use it. I’m making that implicit choice, and it’s really not an explicit one.

But I do know there are serious problems with AI, and I think a lot of the pro-AI people don’t want to talk about that, or don’t want to acknowledge it. As an academic, it’s something I have to talk about with my students, because some of my colleagues are downright hostile toward AI, and some of the students are too. If I want to engage with them, I have to acknowledge the legitimate concerns they have at the same time that I use it. So there’s that internal balance, that internal struggle, that a lot of us are dealing with. But on the other hand, when I want to look something up, I go to ChatGPT, and I don’t tend to think about how much water or energy I’m using.

Lance: I like the comparison with the car. A few years in the AI discourse and I can get frustrated by the inconsistent application of those concerns of AI because they often stop at AI. I recognize that people are worried about the concerns AI represents, but if I’m hearing about the bias of AI from somebody shouting it from the rooftops on social media, or in an hour-long YouTube video that’s been watched a bajillion times, there’s a consistency problem. And for me, a lot of these problems predate AI, so it doesn’t matter if we fix AI if we’re not also going to deal with everything else. I’m just sharing as somebody else who is grappling with it.

David: I’m very interested in your thoughts on this too. There are so many ways to think about it. I do think the more conscientious people have to struggle with this exact question: if you’re not going to just dismiss the things that are problematic, if you’re going to acknowledge them and yet still use it, then on some conscious or subconscious level you’re making the call that the benefits outweigh the risks.

Lance: So let me transition, maybe clunkily, from there. You’ve been doing some programs and events, “creating bespoke learning opportunities with AI,” walking folks through how to build these learning apps. Can you walk us through what a session like that looks like? What has somebody made?

David: To back up: I’ve done a number of general talks for a lot of different audiences, most recently a retirement home and a large academic conference with a couple hundred people. Next week, I’m talking to a group of IT professionals. They all want to know something different, but one thing I’ve been emphasizing over the last few months is that AI development underwent a step change last December. People active in the conversation are probably aware of it, but most of the audiences I talk to have no idea things changed in such a significant way. With, I think it was, Claude Opus 4.6 and Claude Code becoming a reliable autonomous coding agent, then GPT-5.5 and Opus 4.8 shortly after, and soon Mythos, Fable, and GPT-5.6. You can build more stuff and do more stuff.

In particular, if you’re in education, this has created a lot of opportunities that most educators have no clue about. It’s not appropriate for all educators. I’ll be the first to acknowledge that if you’re a sculpture professor, it’s not clear how you’d use this to teach a sculpture class. But it’s more appropriate for more groups of people, especially in STEM, than some people will acknowledge. I teach three subjects: mathematics, data science, and computer science. In all of them there’s potential for AI to create new learning opportunities.

In these workshops, I’ll say, “Here’s a topic you might want to teach,” and I’ll pick one from a math, chemistry, or physics class. Then we’ll ask, what would it look like for students to engage with an app that isn’t just going to tell them the answer? Depending on how long I have (I’ve done workshops from 20 minutes to two hours), I might solicit ideas from the audience, then fire up Codex or Claude Code in real time, and we’ll write the prompt together, watch it build the app, and start iterating, seeing what works and what doesn’t. I’m very clear with the audience that these apps take anywhere from an hour to three or four weeks to build. It’s not generally an instant process, like a lot of people think.

One of my favorites, just as a demonstration, comes from a course I taught years ago on cartography. It was really an excuse to teach trigonometry to math-phobic students. Maps are visceral; they’re everywhere. And when you look at different depictions of the Earth, there are a lot of mathematical questions that aren’t that hard to understand but open up a very visually engaging way to learn.

I made an app where you see different map projections of the same land area. You can pick two points and trace the shortest path and see how it’s actually a curved line, depending on the map. Then you can think about how the map is created using geometry, manipulate the geometry, and watch the map change in real time. It uncovers both an understanding of these maps we see all around us and some idea of the mathematics behind them. It creates a unique learning opportunity that would have been tailored for that class I taught ten years ago.

Lance: That’s so cool. As you’re sharing this and having them explore, what have been some of their report-backs? “This is giving me an idea for X,” or “check out this prototype”?

David: It was interesting. In one of the last ones, people had their laptops and were following along in real time and that was an audience full of mathematicians. Two or three of them, by the end of the talk, showed me apps they’d made during the talk that they said they’d have had no idea how to make beforehand. It’s amazing to see these real-time applications popping up within an hour of describing the process. There’s an initial learning curve, but it’s only about an hour; this isn’t “take a semester course” or “become a software engineer.”

Share AI + Education = Simplified

Lance: A bit of a related aside. I’ve been thinking about agentic AI, one thing I keep coming back to in the learning space is friction. A lot of agentic AI’s uses are about efficiency, but in learning we want friction. I keep thinking: what if students were shown how to use agentic AI to build their own learner workbench? Imagine the first week of every course, students are pulling their different classes together, building systems of reinforcement and learning across them, finding where the connections are. It’s no longer just the instructor, it’s the student. I get nerdily excited about that as an interesting ownership of learning.

David: That’s a great idea; I hope you don’t mind if I steal it. Everyone learns differently; some students want to rely on their notes, some on recorded lectures. Now there’s the potential for each student to design their own customized learning platform that pulls that information together in a way tailored specifically for them. That’s a really good idea.

Lance: I’ve been trying to frame and think through how we capture whatever value there is for the teaching-and-learning side of agentic AI, because if we don’t find those conversations, it’s all going to be about efficiency, and that’s not what we need here. So, you mentioned you don’t like the term “vibe coding.” Tell me why.

David: The “coding” part turns off people who’ve never coded in their life. They assume “vibe coding” is a kind of coding, when the whole point is that now you can make things without knowing how to code. On the other hand, the “vibe” part makes it seem like it’s instant, something you do by the seat of your pants, which is also the wrong connotation. Both words have the wrong connotation for what it actually is.

Lance: Particularly for the population entirely new to it, it’s more exclusive; it’s internal code-talk, as it were.

You had a piece (the New Digital Divide) about how agentic AI is the skill of the future, but real agentic work can cost a lot of money. So much of the conversation is about who gets access to what, and how much. I think about this a lot, because even with institutional agreements, I’m seeing AI tools move from “use as much as you want” to capped usage. So what does an instructor actually do in a semester with a room where some students can afford good agents and some can’t or where the institution provides access, but a student might run out halfway through the semester?

David: There’s a serious equity issue in our classrooms now, and in one sense nothing can be done about it: a student who can afford a $200-a-month model will always have access to more powerful AI than anyone who isn’t paying that. With that said, I wrote a follow-up post, and I’m a little more hopeful now than I was even a month ago, because I’ve been playing more with the open models, like Qwen 2.5 and GLM 5.3

They’re not as good as the current GPT-5.6 or Opus 5, but they’re good enough that for most undergraduate tasks, they’ll perform the same. If you’re doing things beyond undergraduate work, you’ll want the power of Codex or Claude Code, but for undergraduate work, some of these open models are just as good. So I’m a lot less worried than I was a month ago, because I think that’s a viable path forward to resolve the equity issues in the classroom. A student with a $200 model may still be at an advantage, but I’m hoping it’s just not as much of one.

The schools are struggling. Here at the Claremont Colleges, I’ve talked to the deans at all five undergraduate colleges, and they’re all coming up with different solutions. Some, like Harvey Mudd, are trying to provide access to Gemini or Copilot but for any sort of agentic coding like I was describing, Gemini is not a good model for those purposes right now, which is shocking, but that’s another story. I was at a conference a couple of weeks ago sponsored by the National Science Foundation, and they have a program, the National Research Platform (NRP), providing NSF-hosted open models that essentially any educator can use. I have a free account; I just emailed the organizers at NSF and they created a workspace for me that I can use for any of my classrooms. So through that program, my students will have free access to Qwen, GLM, and all these open models that are now getting competitive with ChatGPT and Codex. This is all within the last couple of weeks for me; it’s really changed my thinking on the equity issue.

Lance: Out of curiosity, is it a portal they go to, rather than a download?

David: They have both. There’s a chat interface: you can use it as if you’re using ChatGPT, but you can also get an API access key. There are open agentic coding harnesses like Codex and Claude Code; those are harnesses for these models. The one I’ve been using is Cline, a plugin for VS Code. You give Cline an API key for one of these NSF-hosted models, and now you essentially have Claude Code on a free platform. It’s been pretty effective. I’ve hit some walls. Some things I can’t do that I know I could with Codex, but for anything in my intro computer science class, it would be usable.

“There are so many ways to think about it.”

Comments