New Computing. New Pedagogy.

Abstract

Abstract

People are now building software by directing coding agents in natural language. Some are better at this than others, and we don’t understand why. Figuring out how to teach conversational programming is the most important open question in software engineering education, and in computing education more broadly. The skill is real, not reducible to classical programming, economically decisive, and absent from our curricula. This talk accompanies a position paper, Nobody Knows How to Teach Conversational Programming: Today’s Most Powerful Computing Skill, and makes two asks of the room: tell me specific things to teach, and I can put them in front of more than a thousand students next month; and bring a lens to the data, millions of student–agent messages with the code beside them, from thousands of students learning this skill.

Summary

Summary

A ten-minute vision talk, delivered in the Tuesday morning session of a small workshop on the frontiers of software engineering education. It opens with a personal failure: almost twenty thousand introductory students taught, most of whom never built something they could use and are no longer creating technology. The world, cut into a jigsaw, comes apart until one piece in ten remains. Then the dream our field has always had, in Grace Hopper’s words from 1958, and the fear it has always provoked, in Dijkstra’s words from 1975 with one word swapped. Today the dream is coming true: three live showcases of what people with no training are already building, and three levels of what they could do with more, an assignment, a course, and a degree. The heart of the talk is why this is the most positive vision of technology and human flourishing I can imagine: few have built for everyone; now everyone can build for a few, or for themselves. Nobody knows how to teach this yet. That’s the job, and software engineering is the discipline with the most to share, much of it already carried forward inside the agents. It closes on fundamental uncertainty, said plainly over a group setting off toward a brightening horizon, on what we can do together, and on the pieces of the world coming back.

I’m a computing educator, and I’m failing

I’m a computing educator, and I’m failing

I’m not a software engineering researcher. I’m here for help. In twenty years I’ve taught almost twenty thousand students to program, and most of them never realized their dreams in technology. Most are no longer programming at all. It was too difficult for some and too unpleasant for many. The slide is the world, whole, and then most of it gone.

Our community has always dreamed of this

Our community has always dreamed of this

Programming computers in natural language is a dream as old as the field. Hopper’s FLOW-MATIC team, in 1958: anyone, accountants, analysts, operations managers, can use this system with little training; detailed computer coding is no longer necessary. I wish she could have seen this year. Then a line the room will nod at, about students exposed to AI tools being mentally mutilated beyond hope of regeneration, until the word flips back to what Dijkstra actually wrote in 1975: BASIC. I learned BASIC. The room’s concerns are older than the room, and half right every time: the hard part survived.

Today that dream is coming true

Today that dream is coming true

Coding agents that can really do this are less than a year old. The leap came in late 2025; before that they made too many mistakes and couldn’t correct themselves. Gibson: the future is already here, just not evenly distributed. Amara: we overestimate the short run and underestimate the long. We’re barely into the short run, the tools are expensive, and most of the people using them already had computing training. Very early days.

But the signs are already very exciting

But the signs are already very exciting

Three live pages, scrolled during the talk: the CS 124 Duckies, 287 independent projects last spring, most of them built for someone; the final projects from Using and Understanding AI, a course whose students never read code; and the Illinois summer AI studio, where faculty across campus built their own tools. People with no training are building real, useful things, and the models are getting better fast.

Now imagine what they could do with more

Now imagine what they could do with more

More than a few hours. The shape of the old assignment, and what happens to it once the student hands the specification to an agent, is the subject of our workshop report, AI Can Do Your Homework. An assignment: CS 124’s independent project, your own app with an agent, two thousand students a year, with seventy percent of the grade still proctored and AI-free. A course: Conversational Programming, no source code at all, real users by week two, and a pedagogical agent that watches every student’s builder and carries what the instructor knows. A degree: Applied Computing, six courses in two years, fluency first and how software works last, studios and a public defense. Working beside a collaborator that carries more software engineering knowledge than any person alive, to do what they always wanted to do with technology.

Here’s why I’m so excited

Here’s why I’m so excited

I want to see what the people left out of technology will create once they have the chance. Too few people build software that affects too many, and so many of software’s harms trace to that, as do the money and the power. The animation on this slide starts as a hub and spokes, three familiar faces at the center of everyone, and ends as clusters, pairs, and singles: few built for everyone; now everyone can build for a few, or for themselves. Bertrand Meyer’s framing, extended: software itself is already an E-type technology, built by a few and lived with by everyone. If AI turns out E-type too, that’s no regression. But it offers something software never has: the chance to level the field.

Nobody knows how to teach this yet. That’s the job.

Nobody knows how to teach this yet. That’s the job.

I don’t know how to teach this. I don’t think anyone does; that’s the title of the paper, and it’s the humble truth. We spent decades building computing pedagogy, and the pedagogy for the new computing will take real time too. The slide opens tight on one person climbing a mountain alone, then pulls back: far ahead, another is riding a crystal giant up the same slope, pointing the way. New shoulders, same challenges. Software engineering is the part of computer science with the most to share with creators who will never read, write, or debug source code. Much of that knowledge is already embodied in today’s coding agents, sitting in a library of this room’s own books, and not all of it will, or must, transfer. My new course starts from nothing on purpose: no prerequisites, no code, just talk to it, and let the students, the agent, and the work tell us what needs teaching. And we need a bigger piece of the community than computing has ever let in: writers, designers, studio teachers, domain experts, anyone who has taught people to make something and judge it. They aren’t the audience for this work. They’re the colleagues.

I finally asked the AI how to work with it

I finally asked the AI how to work with it

It took me a long time to think to ask. Five of its eight answers, one at a time: say what done looks like, not how to get there; say what you saw, not what you concluded; never accept “done” without looking; correct it, out loud, early; ask it how to work with it. A lot of this room’s field, in a few lines. You’ve done an amazing job helping coding agents become good at this. Helping their human collaborators is the next great challenge. The rest are on the educators page.

What we can do

What we can do

I don’t know what my students need to learn. I don’t know how to teach them. I don’t know what happens next. Said over a small group with a crystal companion leaving mapped farmland behind, the sky brightening a step with each line. Then the call to action, four pictures and one word each. Learn: an educator and an agent at one screen; we need software engineering people to learn these tools, because you bring so much knowledge to bear and will see things I don’t. Study: what knowledge supports this new skill, and how is it acquired? I can help, with the data and with courses to deploy pedagogy in. Teach: design the new computing curriculum. Build: new educational tools and learning platforms for our students, our colleagues, and the next billion programmers. The world comes back together under the last line: I hope to meet you all in this future.

Thanks to Bertrand Meyer, Annie, and Armando Fox, to everyone at the workshop, and to the wonderful staff of the Château de Villebrumier.

How this talk was made

How this talk was made

The talk and its slides were made by conversation, in the way the talk describes: I said what I wanted, Claude built it, I looked and asked for changes, and no slide code was written by hand. The outline went through seven versions over a weekend on one machine, and the deck was built on the laptop over the day before the talk, between sessions of the workshop. Every prompt I typed to make it, from the first outline to the deployed deck, is on the prompts page, with counts of what Claude did in reply and a few highlights pulled to the top.

Credits

Credits

The illustrations were generated with Gemini from prompts in the deck’s repository. Portraits from Wikimedia Commons: Grace Hopper at the UNIVAC I console, 1960, Smithsonian Institution, no known copyright restrictions; Edsger Dijkstra, 2002, Hamilton Richards, CC BY-SA 3.0, shown mirrored; Mark Zuckerberg at the F8 2019 keynote, Anthony Quintano, CC BY 2.0; Elon Musk, Royal Society portrait, Debbie Rowe / The Royal Society, CC BY-SA 3.0; Bill Gates at the Munich Security Conference 2017, Kuhlmann / MSC, CC BY 3.0 DE.