Latest / Geoffrey Challen: Teaching and Technology / Taking Flaky Farther
Transcript
- 0:00Geoffrey Challen: Hey everyone, my name is Geoff Challen, and I teach computing at the University of Illinois.
- 0:04Geoffrey Challen: I write about teaching and technology at geoffreychallen.com,
- 0:08Geoffrey Challen: where you can find a written version of the essay you're about to hear.
- 0:11Geoffrey Challen: What follows is the essay being read aloud, but not by me, by a clone of my voice.
- 0:16Geoffrey Challen: I hope you enjoy it.
- 0:17Geoffrey Challen: If you have any feedback, please don't hesitate to reach out to me directly.
- 0:21Geoffrey Challen: I'd love to hear from you.
- 0:22Geoffrey Challen: Taking Flaky Farther, an essay by Geoffrey Challen, read aloud by Geoffrey Challen's voice clone.
- 0:29Geoffrey Challen: In the excerpts from his AI sessions, his prompts are in this voice,
- 0:33Geoffrey Challen: and the agent's replies are in a second synthetic voice.
- 0:37Geoffrey Challen: Progress toward getting computing educators to recognize what AI coding agents can do has been slow.
- 0:43Geoffrey Challen: That’s not surprising.
- 0:44Geoffrey Challen: The good agents are new.
- 0:46Geoffrey Challen: But their capabilities are genuinely impressive,
- 0:49Geoffrey Challen: and I think many computing educators don’t use them enough to really believe it.
- 0:53Geoffrey Challen: So I was excited to see a new piece in Communications of the ACM by Kathi Fisler,
- 0:58Geoffrey Challen: Shriram Krishnamurthi, and Michael Littman on teaching novices in the agentic AI era.
- 1:04Geoffrey Challen: My colleague Cory Bart linked it on our computingeducators.org forum this week.
- 1:09Geoffrey Challen: There’s a difference between heading in the right direction and even being pointed in the right direction,
- 1:14Geoffrey Challen: and not many people are doing either yet.
- 1:17Geoffrey Challen: This piece is pointed in the right direction.
- 1:20Geoffrey Challen: Students working at the level of specifications and testing.
- 1:23Geoffrey Challen: Even the radical option of dropping most code writing from early courses.
- 1:28Geoffrey Challen: It also acknowledges, although I wish more forcefully, that reading code is not a good way to validate a system.
- 1:35Geoffrey Challen: In their words: “Tracing does not scale.”
- 1:39Geoffrey Challen: That matters, because a position I’m seeing some introductory courses begin to take is:
- 1:44Geoffrey Challen: we’ll use AI to produce the code, but students still have to learn to read it.
- 1:49Geoffrey Challen: I think that’s worse than having students read and write code.
- 1:52Geoffrey Challen: When you write code, it follows your own mental model of how to solve the problem,
- 1:56Geoffrey Challen: so reading it back, it’s much easier to see what it’s supposed to do and to spot what’s wrong.
- 2:01Geoffrey Challen: Reading somebody else’s code is really difficult.
- 2:05Geoffrey Challen: It’s so difficult that one of the things we’re discovering in this new era is how broken a lot of software is,
- 2:11Geoffrey Challen: including important code that dozens or hundreds of highly skilled engineers have read over and over again.
- 2:17Geoffrey Challen: Agents are finding bugs in code that has been deployed for decades, written and maintained by very intelligent people.
- 2:24Geoffrey Challen: One lesson of the AI era is that humans are bad at reading code.
- 2:28Geoffrey Challen: We could try to retrain everyone to be better at it.
- 2:31Geoffrey Challen: But I think it points to basic human limitations.
- 2:35Geoffrey Challen: Two Problems With the Metaphor
- 2:37Geoffrey Challen: My worry is the metaphor they build the piece around: an AI agent as a flaky compiler.
- 2:43Geoffrey Challen: I see two problems with it.
- 2:46Geoffrey Challen: The first is that it supports, maybe unintentionally,
- 2:49Geoffrey Challen: the view that we should teach students to generate code with AI and then read it to decide whether it’s correct.
- 2:55Geoffrey Challen: It ignores all of the new ways agents can take part in validation,
- 2:59Geoffrey Challen: which can put humans in a position where they don’t have to read code and yet get stronger guarantees that it works.
- 3:05Geoffrey Challen: The flaky compiler reinforces that middle ground, which I think is the worst of the options.
- 3:10Geoffrey Challen: Reading and writing code is fun.
- 3:12Geoffrey Challen: Reading code that someone else wrote,
- 3:14Geoffrey Challen: particularly a coding agent that may have its own ideas about what good code looks like, is not.
- 3:20Geoffrey Challen: That puts humans in an awkward position, and it could produce worse courses.
- 3:24Geoffrey Challen: The second problem is that the metaphor ignores nearly everything AI can do to
- 3:29Geoffrey Challen: support software development beyond compiling an English specification into code.
- 3:34Geoffrey Challen: When Cory shared the piece, I replied with a list, off the top of my head:
- 3:38Geoffrey Challen: A compiler can’t read code and make architectural suggestions.
- 3:42Geoffrey Challen: A compiler can’t generate tests and use them to correct its own mistakes.
- 3:46Geoffrey Challen: A compiler can’t generate user personas and use them to validate and improve a computing artifact.
- 3:52Geoffrey Challen: A compiler can’t take a vague bug report, explore the history of the project to pinpoint the cause,
- 3:58Geoffrey Challen: apply and deploy a local fix, and then file a bug report, complete with a local demonstration.
- 4:04Geoffrey Challen: A compiler can’t take the instruction “something’s not working right now”, analyze logs from a production system,
- 4:10Geoffrey Challen: spot a performance problem, fix the code, and redeploy the affected component, all in the space of five minutes.
- 4:16Geoffrey Challen: A compiler can’t brainstorm new project ideas, coordinate a team of other compilers to solve a more complex problem.
- 4:23Geoffrey Challen: I’ve been recording my work with AI for the past year, so I went back to find real examples.
- 4:29Geoffrey Challen: What follows are excerpts from those sessions: my prompts exactly as I typed them, and the agent’s replies trimmed.
- 4:36Geoffrey Challen: All times are Central.
- 4:38Geoffrey Challen: Chasing a Vague Bug Report Upstream
- 4:41Geoffrey Challen: Fixing Production in Nine Minutes
- 4:44Geoffrey Challen: Validating With Simulated Students
- 4:47Geoffrey Challen: Checking In on a Live Class
- 4:50Geoffrey Challen: Investigating a Storage Outage
- 4:52Geoffrey Challen: Brainstorming With Data
- 4:54Geoffrey Challen: A Research Assistant That Pushes Back
- 4:57Geoffrey Challen: When I Have to Push Back
- 4:59Geoffrey Challen: Chasing a Vague Bug Report Upstream
- 5:02Geoffrey Challen: A few students reported on our course forum that code was vanishing from our in-browser editor, Ace, while they typed.
- 5:09Geoffrey Challen: It took me a while to even understand what they were describing:
- 5:12Geoffrey Challen: as they typed, the lines being pushed downward disappeared.
- 5:16Geoffrey Challen: My agent is connected to the forum.
- 5:18Geoffrey Challen: Then we have a new bug report to chase.
- 5:21Geoffrey Challen: Check the forum.
- 5:22AI agent: Production is running Ace 1.45.0, but package.json pins ace-builds to 1.44.0, so something pulled in a second copy.
- 5:33AI agent: Tracing where it comes from and when it arrived:
- 5:36Geoffrey Challen: Our first hypothesis, that two versions of the editor in the bundle were interfering with each other,
- 5:41Geoffrey Challen: turned out to be wrong.
- 5:43Geoffrey Challen: The real cause was a change in the editor library itself, released the day before.
- 5:48Geoffrey Challen: We pinned the previous version, and the fix was live in production 17 minutes after I asked.
- 5:53Geoffrey Challen: Then I wanted to know exactly what had broken.
- 5:56Geoffrey Challen: I’m starting the dev server.
- 5:58Geoffrey Challen: Let’s try to repro this locally.
- 6:00Geoffrey Challen: Otherwise we’re just guessing about a fix.
- 6:03AI agent: I reproduced the bug locally and confirmed the fix.
- 6:06AI agent: I replayed a student’s recorded session (1,123 edits and cursor moves)
- 6:12AI agent: through our real editor component on the dev site, with Ace 1.45.0 installed locally.
- 6:18AI agent: At step 96, the Enter that takes the code from 3 lines to 4,
- 6:23AI agent: the editor grows to 88 px but the inner box Ace draws the code into stays at 66 px.
- 6:30Geoffrey Challen: That’s about 40 minutes, while I was doing other things.
- 6:33Geoffrey Challen: The agent built a minimal reproduction, I filed the bug report the next day,
- 6:38Geoffrey Challen: and the maintainers merged a fix two days later.
- 6:40Geoffrey Challen: A flaky compiler cannot do this.
- 6:44Geoffrey Challen: Fixing Production in Nine Minutes
- 6:46Geoffrey Challen: On Tuesdays my course runs new kinds of activities,
- 6:50Geoffrey Challen: and early in the semester we play a get-to-know-each-other bingo game.
- 6:54Geoffrey Challen: Students open a board on their phones, walk around the room asking each other the questions on the squares,
- 6:59Geoffrey Challen: record the answer, take a selfie, and scan each other’s codes to fill the square.
- 7:04Geoffrey Challen: The projector shows their selfies live as they’re taken, along with the number of bingos and blackouts.
- 7:10Geoffrey Challen: We ran it last spring, so I didn’t think it needed many changes.
- 7:13Geoffrey Challen: This fall’s first section had more students than any we’d run it with before,
- 7:18Geoffrey Challen: and within about five minutes it was clear something was wrong.
- 7:22Geoffrey Challen: It seems like the bingo endpoints are somewhat unstable at the moment.
- 7:27AI agent: Yes, the readiness probe is the mechanism turning this into hard errors.
- 7:32AI agent: Cause: /staff-display.
- 7:34AI agent: Every square completion calls notifyBingoStaffUpdate(), then every open staff display refetches,
- 7:41AI agent: then the server does collection.find(query).toArray() over every board with every base64 photo
- 7:48Geoffrey Challen: We need to fix this now.
- 7:51AI agent: Fixed and deployed.
- 7:53Geoffrey Challen: INCREDIBLE WORK!
- 7:55Geoffrey Challen: Thank you.
- 7:56Geoffrey Challen: That was fast.
- 7:57Geoffrey Challen: Nine minutes from my first message to a fix in production, while the class kept playing.
- 8:03Geoffrey Challen: Then it wrote up the incident and posted it to our staff forum.
- 8:06Geoffrey Challen: Again, not something a flaky compiler can do.
- 8:09Geoffrey Challen: I have other examples like this,
- 8:11Geoffrey Challen: including agents taking memory traces of our running backend containers to find performance problems.
- 8:17Geoffrey Challen: Validating With Simulated Students
- 8:20Geoffrey Challen: Another thing a flaky compiler can’t do: validation that goes far beyond reading the code.
- 8:26Geoffrey Challen: In those same activities we’ve had to design around students who don’t read the directions,
- 8:30Geoffrey Challen: arrive late, or sit at the wrong table.
- 8:32Geoffrey Challen: So now, as part of developing every activity,
- 8:35Geoffrey Challen: we have agents create simulated student personas and use them to walk through it, looking for the corner cases.
- 8:41Geoffrey Challen: Some are designed to be students who don’t follow instructions carefully or won’t read a long block of text.
- 8:48Geoffrey Challen: Let’s do adversarial review of tomorrow’s activity.
- 8:51Geoffrey Challen: Please design an approach and use Opus agents with different student personas.
- 8:55Geoffrey Challen: We’ve done this before so there may be notes in the repository.
- 8:59Geoffrey Challen: The goal is to complete the process of thoroughly testing the activity and instructions so things go smoothly tomorrow.
- 9:05Geoffrey Challen: Also take a look at the existing automated test suites and correct any gaps.
- 9:09Geoffrey Challen: Thanks!
- 9:11AI agent: Early signal from the traces:
- 9:13AI agent: 7 of 9 simulated students saw the red “Too Late for Attendance in This Section”
- 9:18AI agent: banner at :11 and were confused by it (several decided it was “for other people”,
- 9:23AI agent: one left early)
- 9:25Geoffrey Challen: Fifty-three minutes later the banner was fixed,
- 9:28Geoffrey Challen: other agents had found six server bugs and fixed four of them, and the tests had gone from 46 passing to 111.
- 9:35Geoffrey Challen: Reading the code is a terrible way to find a problem like that banner.
- 9:39Geoffrey Challen: You can’t even get a sense from the code of what’s on the screen.
- 9:42Geoffrey Challen: You have to look at the interface and connect it back to the code, and agents can do all of that quickly.
- 9:48Geoffrey Challen: Checking In on a Live Class
- 9:51Geoffrey Challen: I’m also teaching a new course this semester on conversational programming.
- 9:55Geoffrey Challen: I don’t lecture, because lectures don’t work.
- 9:58Geoffrey Challen: Students work together, and each meeting is designed differently.
- 10:02Geoffrey Challen: The course has about 20 students right now, but I’m imagining teaching it to many more,
- 10:06Geoffrey Challen: and wondering how I’d keep track of what’s happening.
- 10:09Geoffrey Challen: So I’ve been building myself a chat interface for the class:
- 10:13Geoffrey Challen: an agent that has data about the room and can help me decide that the 20 minutes we planned are done after 15,
- 10:19Geoffrey Challen: or that one student has hit a strange bug and needs me, or that a late arrival needs a group.
- 10:24Geoffrey Challen: Here’s part of one class, while it was happening:
- 10:27Geoffrey Challen: Sweet.
- 10:28Geoffrey Challen: Hey, can we start the playlist?
- 10:31Geoffrey Challen: Next!
- 10:32Geoffrey Challen: OK.
- 10:33Geoffrey Challen: How’s it going?
- 10:34Geoffrey Challen: We can also work together to decide when to call these stages.
- 10:38Geoffrey Challen: We’re seven minutes in and students have already collected a lot of ideas,
- 10:42Geoffrey Challen: although a lot of them are falling into the “something else” category :-).
- 10:46Geoffrey Challen: I guess that means students are thinking hard!
- 10:49AI agent: The ideas are specific, which is a good sign.
- 10:53AI agent: To be fair, though, most of the hard thinking so far is the agents’.
- 10:57AI agent: Only 13 of the 105 ideas were written by students; the agents proposed the rest.
- 11:03AI agent: Students are asking good questions and getting long lists back.
- 11:07AI agent: That’s a natural place to start, and the go-around is a good moment to push on it:
- 11:12AI agent: “Which one on your list would you have thought of yourself?
- 11:14AI agent: Which one do you actually trust?”
- 11:17Geoffrey Challen: This is still new.
- 11:18Geoffrey Challen: I built the system with coding agents, and used them to validate it.
- 11:22Geoffrey Challen: But you can’t talk to a flaky compiler about what’s going on in your class.
- 11:27Geoffrey Challen: Investigating a Storage Outage
- 11:29Geoffrey Challen: Agents are also good at much broader investigations.
- 11:33Geoffrey Challen: My course runs on virtual machines provided by the university,
- 11:36Geoffrey Challen: with Kubernetes on top to manage everything, and agents are great at Kubernetes.
- 11:41Geoffrey Challen: We’ve used this setup for a long time.
- 11:44Geoffrey Challen: In October the cluster suddenly became unstable: machines rebooting,
- 11:49Geoffrey Challen: database hosts going down, the whole system essentially unusable.
- 11:53Geoffrey Challen: Of course part of it happened during a quiz in our computer-based testing facility,
- 11:57Geoffrey Challen: so some quizzes were spoiled and students had to reschedule.
- 12:00Geoffrey Challen: It went on for a couple of weeks, coming and going, and we thought it had been fixed when it hadn’t.
- 12:07Geoffrey Challen: The investigation itself was not fast.
- 12:10Geoffrey Challen: But the instrumentation needed to see the problem came together quickly.
- 12:14Geoffrey Challen: On the first afternoon a team of agents went through storage history, metrics,
- 12:18Geoffrey Challen: host logs, and our database machines in parallel, and I asked the obvious question:
- 12:23Geoffrey Challen: So wait, sorry.
- 12:25Geoffrey Challen: This sounds like an underlying storage problem, correct?
- 12:28Geoffrey Challen: If it’s hitting nodes across the cluster.
- 12:30AI agent: Yes.
- 12:31AI agent: The evidence says the storage under the VMs degraded on Thursday afternoon,
- 12:36AI agent: and everything since has been happening on top of that. db1,
- 12:40AI agent: db2 and db3 run MongoDB on their own virtual disks, no Ceph, no Kubernetes.
- 12:46AI agent: Their write latency jumped 10–30× on Thursday
- 12:51Geoffrey Challen: Within a few hours of work with agents we had new metrics in the cluster,
- 12:55Geoffrey Challen: new reporting, and alerts that reach my phone through Healthchecks and Pushover:
- 12:59Geoffrey Challen: No need to alert staff here; I’m the cluster operator and only maintainer.
- 13:04Geoffrey Challen: Let’s definitely set up email and the Healthchecks.
- 13:07Geoffrey Challen: I’ll install Pushover for the mission-critical stuff.
- 13:10Geoffrey Challen: Two days later the problem came back.
- 13:13Geoffrey Challen: OK.
- 13:14Geoffrey Challen: Overall, great work on this.
- 13:16Geoffrey Challen: The alerts fired right at the right time and before (I suspect) Tech Services even knew there was problem.
- 13:22Geoffrey Challen: Nice job.
- 13:24Geoffrey Challen: The data the agents gathered helped campus IT figure out what was happening.
- 13:28Geoffrey Challen: They provide storage to my machines using a distributed storage system called Ceph,
- 13:33Geoffrey Challen: and they had made a change to it, not realizing that I also run Ceph inside my cluster.
- 13:38Geoffrey Challen: Turtles all the way down: Ceph running on Ceph produced terrible storage performance.
- 13:44Geoffrey Challen: The fix was to move my machines onto local disks,
- 13:47Geoffrey Challen: which is safe in my case because everything I need to keep is replicated by my own Ceph and backed up as well.
- 13:54Geoffrey Challen: A flaky compiler cannot investigate a strange storage problem on your cluster, gather the data,
- 13:59Geoffrey Challen: give it to the people you’re working with, and put monitoring in place so you see the next one first.
- 14:06Geoffrey Challen: Brainstorming With Data
- 14:08Geoffrey Challen: You also don’t brainstorm with a flaky compiler.
- 14:11Geoffrey Challen: Last week we grouped students at tables by what they were working on:
- 14:15Geoffrey Challen: agents had read every student’s project summary, so we could seat students with others working on similar ideas.
- 14:21Geoffrey Challen: I was trying to decide how big the tables should be, since students don’t always show up.
- 14:26Geoffrey Challen: I mean, don’t small groups work better here?
- 14:29Geoffrey Challen: We could also use a mix depending on project fit, yes?
- 14:32Geoffrey Challen: Can we incorporate the project summaries into the assignment process?
- 14:36Geoffrey Challen: What I mean is that perhaps students working on more common ideas should
- 14:40Geoffrey Challen: sit it larger groups whereas those working on more unique ideas in smaller.
- 14:44Geoffrey Challen: Does that make sense?
- 14:45Geoffrey Challen: It seems like it could complicate the assignment process.
- 14:49AI agent: Your idea makes sense.
- 14:51AI agent: Before I answer, I’m testing it against the stored embeddings, so the answer comes from data, not intuition.
- 14:57AI agent: Smaller groups do work better here, but sizing groups by how common each project is didn’t help.
- 15:04AI agent: I tested both on the stored project summaries,
- 15:06AI agent: regrouping all 1,047 students who have one, within their hours, four ways.
- 15:13Geoffrey Challen: That’s data-driven brainstorming, in 97 seconds.
- 15:17Geoffrey Challen: A Research Assistant That Pushes Back
- 15:20Geoffrey Challen: A flaky compiler isn’t a research assistant, either.
- 15:24Geoffrey Challen: While writing my last essay I remembered hearing that studies
- 15:27Geoffrey Challen: had found no relationship between research and teaching ability.
- 15:31Geoffrey Challen: Trying to work logically here.
- 15:33Geoffrey Challen: Universities might say: Hey, we have the best researchers, therefore we have the best educators.
- 15:39Geoffrey Challen: But there’s evidence against this, yet?
- 15:42AI agent: You’re right, and the evidence runs in one direction across forty years.
- 15:47AI agent: Four results, in order of strength for your purposes: Feldman 1987.
- 15:52AI agent: A review of 29 studies:
- 15:54AI agent: the correlation between research productivity and teaching effectiveness is 0.12 Hattie and Marsh 1996.
- 16:02AI agent: The meta-analysis of 58 studies: the overall correlation is about 0.06
- 16:09Geoffrey Challen: Those are Feldman’s 1987 review and Hattie and Marsh’s 1996 meta-analysis.
- 16:15Geoffrey Challen: And compilers don’t push back.
- 16:18Geoffrey Challen: Later in the same work, the agent warned me that a paper I wanted to cite “does not literally say” what I wanted it to.
- 16:26Geoffrey Challen: When I Have to Push Back
- 16:28Geoffrey Challen: It isn’t always right, and I have to push back too.
- 16:32Geoffrey Challen: In my conversational programming course we put limits on the AI we provide to students,
- 16:37Geoffrey Challen: and in the middle of one class we started hitting them.
- 16:40Geoffrey Challen: Is the proxy up?
- 16:41Geoffrey Challen: Students are getting errors.
- 16:43AI agent: The proxy is up and healthy.
- 16:45AI agent: The errors come from Azure, which is refusing our requests with 429 (rate limited).
- 16:52Geoffrey Challen: OK.
- 16:52Geoffrey Challen: This was a 492 from our ai.geoffreychallen.com proxy.
- 16:57AI agent: That changes the picture: the 429 is your gateway’s own limit, not Azure’s.
- 17:03Geoffrey Challen: The limit was one I had set myself, on a proxy the agent didn’t know about yet.
- 17:08Geoffrey Challen: I raised it, and the errors stopped six minutes after I first asked.
- 17:13Geoffrey Challen: Beyond the Flaky Compiler
- 17:15Geoffrey Challen: That’s a laundry list, but that’s the point.
- 17:18Geoffrey Challen: If we want to help our colleagues understand how these tools are going to reshape computing,
- 17:23Geoffrey Challen: we need to go far beyond a metaphor like the flaky compiler, because these tools can do so much more than that.
- 17:30Geoffrey Challen: I think some people are reluctant to acknowledge everything AI can contribute to
- 17:34Geoffrey Challen: software development because they’re worried about what that leaves for humans.
- 17:38Geoffrey Challen: But look at the technology around us.
- 17:40Geoffrey Challen: There are so many problems, so many places where technology doesn’t support human flourishing,
- 17:46Geoffrey Challen: so many problems it could help solve that we haven’t even started on.
- 17:50Geoffrey Challen: Sometimes we have solutions and they aren’t very good.
- 17:53Geoffrey Challen: Sometimes we haven’t started at all.
- 17:56Geoffrey Challen: All of the ways these tools can help create technology are good news,
- 18:00Geoffrey Challen: because they mean humans get to contribute the human things.
- 18:04Geoffrey Challen: We bring our lived experience, the things we care about, and the things that concern us.
- 18:09Geoffrey Challen: We watch the results, see what happens in our communities, and use technology to make the world better.
- 18:14Geoffrey Challen: And by using these tools well,
- 18:16Geoffrey Challen: we’ll bring in a lot of people who could never create technology that serves their needs before.
- 18:21Geoffrey Challen: That’s what’s really exciting.
- 18:24Geoffrey Challen: This is real Geoff again.
- 18:25Geoffrey Challen: Thanks for listening.
- 18:26Geoffrey Challen: You'll find more essays at geoffreychallen.com, and you can subscribe to the podcast wherever you're listening now.
- 18:32Geoffrey Challen: Until next time.
- 18:33Geoffrey Challen: Keep teaching with technology.