Philip Guo: The Small Choices That Kept Python Tutor Alive

A free code visualizer reached nearly 30 million people through instant access, a narrow mission and years of careful maintenance.

PGPhilip GuoIntervenant · UC San Diego
Chaîne
Stanford Online
Publié
20 nov. 2026
Vidéo
1:00:02
Lecture
11 min
Original
anglais
Ouvrir sur YouTube
En faisant défiler, la vidéo se réduit en bas à droite et continue. Les sous-titres suivent votre langue de lecture.
Langue de lectureDisponible en 2 langues
Cet article n’est pas encore disponible en français ; il s’affiche dans l’original en anglais.
RÉSUMÉ

En 30 secondes

Python Tutor grew from a four-day graduate-school prototype to nearly 30 million cumulative users without a company or marketing team. Its creator, Philip Guo, credits a browser-based tool that works immediately, shareable examples and years of fixing small defects. Keeping it focused meant rejecting requests for accounts, grading and a full development environment. As AI changes how people code and monthly usage falls, he is focusing more on teachers while questioning whether the usual defense of programming fundamentals will withstand disruption.

  1. 1Code diagrams make an invisible process visiblePython Tutor draws changing variables and objects as code runs. For example, it shows why modifying a list through Y can change the value later printed through X.
  2. 2Earlier visualizers often ended with their research projectsGuo found a long lineage of similar systems, but many were short-lived academic prototypes. His browser-based side project was designed around immediate use rather than publishing a novel system.
  3. 3No install or account helped examples spreadStudents could open and run linked visualizations immediately, while teachers could share and embed them in lessons. An interview-preparation site in China embedded the tool and brought substantial traffic.
  4. 4Small bug reports accumulated into reliabilityGuo says attention to details, including cases users thought too minor to report, helped the tool work across complex examples over time.
  5. 5A narrow mission meant rejecting entire product categoriesMore languages and better visualizations fit the goal of explaining code; accounts, grading, a full development environment and curriculum did not.
  6. 6Falling monthly use changed the measure of successGuo cannot attribute the entire decline to AI. He is paying more attention to long-term teacher engagement and sees school-age education as an opportunity.
  7. 7AI challenges both the tool and its creator’s assumptionsGuo argues that students still need to understand and debug AI-written code, but recognizes that this defense of fundamentals may be an incumbent’s reflex.

À retenir

  1. Remove installation and login barriers when people need to try and share an educational tool immediately.
  2. A focused mission can guide which requests to refuse as a project grows.
  3. Guo credits accumulated fixes to small user-reported problems for making Python Tutor reliable over time.
  4. Measure the depth of teachers’ engagement as well as the number of users.
  5. Guo argues that understanding code still matters when AI writes it, while acknowledging that this may be an incumbent’s assumption.
TEXTE INTÉGRAL

Lire le texte remanié, chapitre par chapitre

Transcription seule
01

Why code needs a visible explanation

A short example shows why students can struggle to infer how code changes data. Python Tutor draws those changes as a program executes.

I came to Stanford as a graduate student in 2006, without knowing much about human-computer interaction, and sat in these same seminars. Python Tutor now has a live map showing people using it across time zones. This is a personal account of how that project came to be and lasted.

Consider a short piece of code: does it print 1, 2, 3 or 1, 2, 3, 4? The answer depends on the language, and in C++ can depend on how a variable is declared. Without seeing what happens to the data, a novice may have to guess.

Learning to program often looks like putting code into a mysterious box and examining its output again and again. Teachers help students build a mental model of what happens inside, often by drawing diagrams on a board. I wondered why a computer that already runs the code could not draw those diagrams itself.

Python Tutor runs code and draws its changing state. In my example, assigning Y to X in Python does not copy the list: both names point to the same object. Appending 4 through Y changes the object that X also refers to, so printing X yields all four numbers.

Despite the name, the tool also supports languages including Java, C, C++, and JavaScript. It covers concepts taught in introductory and intermediate courses, with some material relevant to systems and programming languages courses. Its aim is to emulate a teacher showing how code runs.

I have maintained this as a side project, essentially as one developer, without a company, marketing department or large organization behind it. Users come from varied backgrounds: a mechanic learning at public kiosks, a network engineer taking an online course and a teacher at a nonprofit coding camp. The cumulative user count is approaching 30 million.

02

A teaching side project becomes a browser prototype

Teaching Python and building web projects led Guo to create the first version in four days. He later learned that similar visualizers had a long research history.

The story has three acts: inception, a decade of acceleration, and the current disruption. I first fell in love with programming in a high school computer science course and then helped teach it. As a college teaching assistant I enjoyed learning and explaining the subject, but my morale fell after I entered research-focused graduate school.

As I isolated and procrastinated, I learned Python and JavaScript through small web projects. Posting them on my personal site led a nearby startup to ask me in 2008 to teach Python to its scientists. I made notes and drew diagrams while teaching once a week; at the time it was just one of many experiments I tried.

Around 2010 browsers became capable enough to support interactive coding and automatically draw diagrams. I wrote down a vision of a web interface where students could enter code, step through execution and see data structures change. I built a simple first version in about four days and put it on a school server.

When I showed it to people, they asked why such an obvious idea had not been built before. It had: a later survey showed decades of program-visualization systems. Most, however, were short-lived research prototypes that were discarded after evaluation; few people I knew had heard of or used them.

The older projects were typically professor-led efforts with grants, student teams, planned development and experiments. They needed novel designs or results to support publications. I was outside that lineage and was trying instead to make something as simple and obvious to use as possible.

Many predecessors were desktop applications that required installation. My version ran on a website, which removed that barrier at a time when browser applications could look like toys. I did not have a grand plan: my position as an amateur happened to put me in a place to pursue something others had overlooked.

03

A free teaching tool grows with online learning

Guo kept the project alongside an academic career. Its easy-to-open, shareable visualizations spread as online learning expanded.

By the time I finished my PhD in 2012, three professors had found the tool and were using it to teach. I wanted to keep working on it, so I considered how to support myself while doing so: joining an education company, founding an organization, taking a teaching-focused position or pursuing a research career.

A company or nonprofit built around such a focused tool might need to expand into an entire platform. A teaching-focused professorship can align well with building classroom tools, and many people I know follow that path. I instead chose a research-focused tenure-track career and continued developing the site alongside it.

The early 2010s brought massive open online courses, a learn-to-code movement, growing computer science departments and investment in technology. The pandemic later created another surge in remote learning. Python Tutor rode that wider wave, but many well-publicized education apps and startups from the period did not survive.

My admittedly hindsight-shaped account of longevity starts with making something people find immediately useful. A person can click a link and see a diagram, edit the code and run it, without an installation, account or login. Teachers can put links in lesson documents and students can explore the examples in seconds.

The visualizations cover many features needed for learning, abstract away some details and preserve others, and should fail clearly when they cannot handle something. Links can be shared in messages or forums, and instructors can embed the tool directly in their own sites.

One popular Chinese interview-preparation site embedded Python Tutor throughout its explanations, sending it substantial traffic without my having known about the site. People can embed it without asking me for permission. That light distribution helps the tool spread through teaching materials and communities I do not control.

I have watched well-funded education initiatives launch with substantial publicity and then disappear. My view is that people do not keep using and sharing a tool simply because it looked interesting on launch day. They need to find it genuinely useful enough to pass on; early email and Hacker News responses showed me that some users felt they had been looking for something like this.

04

Treating software as craft, not just a growth metric

Small user-reported defects accumulate into a better experience. Growth and institutional incentives can instead pull a project away from its users.

Books about working with materials and teaching design helped shape my sense of craft. People make software as a business to grow a metric, as research to pursue knowledge and publish, or as art for expression. I think of this project as craft: making something both good and useful for other people.

Craft also means treating small defects as worth fixing. When a user reported that a C example did not visualize perfectly and called it a minor detail, I welcomed the report. Thousands of small improvements accumulate into the experience of a tool that works as people expect, including on complicated cases.

Growth can also damage a good project. Inside a large company, a growing project attracts demands to serve corporate goals; a startup or nonprofit faces pressure to monetize, raise funding and hire. A teaching project can grow into an all-encompassing curriculum, while a research project can change to collect data and produce papers.

Those pressures reflect who the builder must impress: executives, funders and journalists, teaching colleagues, or reviewers of papers and grants. Those audiences are not necessarily the people using the tool. I think pursuing their approval can pull a project away from the qualities that made it useful in the first place.

05

Protecting the project’s narrow mission

Guo supports additions that improve code visualization and declines features that would turn it into a platform. Limited time helped enforce that focus.

I try to keep the mission narrow: help novices build useful mental models of code. Supporting more languages and harder language constructs fits that mission. It also means saying no to most requests that would turn the tool into a full development environment, a production debugger, a learning-management system or a complete curriculum.

That focus was not solely a carefully planned strategy. Throughout the decade I was running a research group, teaching and supervising students, and had limited time to work on the tool myself. Those constraints forced me to choose what mattered most.

06

Responding to falling use and generative AI

Guo is prioritizing teacher engagement and redesigning the site with a human collaborator and AI implementation tools. He weighs the case for fundamentals against the possibility that educators are defending an old model.

Monthly use peaked before a later decline. The arrival of generative AI is a tempting explanation, but I cannot establish that it caused the whole decrease. AI is changing the ecosystem as people worry about software jobs and computer science enrollment, while search results fill with imitation sites, some built around advertising rather than working functionality.

My own interface had also begun to look old. With a shrinking audience perhaps becoming the new reality, I have shifted attention from the size of the graph toward depth of engagement, particularly among teachers. Some have used Python Tutor for over a decade; former student users have returned as teachers themselves.

I see potential in school-age education. Computer science curricula took longer to reach K–12 schools than universities, but once policies and curricula are established they can remain for a long time. That makes the growing presence of coding in schools an encouraging area for the tool, in my view.

I could have asked AI to modernize the site, but that might have made it look like every other AI-designed page. Instead, I am working with a visual artist and designer to develop a shared aesthetic; AI helps implement the design on the web. The direction remains a human collaboration.

I also use AI to help with back-end work. It can feel like supervising a team of programmers, testers, data scientists, IT staff and security specialists. My own role becomes closer to a product manager who delegates, checks the work and learns how to supervise these tools.

That raises an obvious question: why use Python Tutor or learn the details of code if AI will write more and more of it? Computer science educators commonly argue that novices still need foundations and problem-solving skills before they become experts who use AI. Mathematics offers an analogy: calculators exist, yet students still work through equations to understand them.

The mission has always been to help people understand, debug and build on code written by other people. Replacing 'people' with 'AI' may not change the need to know what code is doing. But I recognize that this is also exactly the kind of confident argument an incumbent makes while being disrupted.

I illustrated that danger with a fictional scene about the ship of education sinking while its architects keep insisting that fundamentals will always matter. It is a joke, but it points to my uncertainty about whether the familiar educational answer is enough for what is changing.

I also see signs of hope. Some people at AI-forward organizations emphasize understanding how things work, and I sense a reaction against an oversupply of generated material. My expectation is that some younger people will want to study deeply and seek things that feel made by someone who cares.

Mass-produced, AI-generated software may be fine for many utilitarian needs. I think there will also be room for software made as a craft: something good and useful that someone cares about maintaining. Making and keeping such a thing good over a long period has also given me personal fulfillment.

07

Questions about time, research data and finding users

Guo describes working in gaps, the tradeoff between tracking and simplicity, and why he thinks diagrams met a real teaching need. He says his early motivation was intellectual fulfillment rather than starting a company.

A doctoral student who used Python Tutor in middle school asks how I found time for it amid academic work. I generally filled gaps in my schedule, working in bursts of a few weeks when paper or teaching deadlines were not pressing. Sustaining a project for a long time does not require a large amount of time at every moment.

Asked about younger students using physical-computing platforms such as micro:bit and MakeCode, I think step-by-step visualization would be a promising direction. I hope AI tools might help make such support feasible, though I do not announce an implementation.

An audience member describes research suggesting that students with consistent mental models can succeed even when an initial prediction is wrong, and asks what research the tool's logs could support. I deliberately avoided accounts and extensive tracking, so the available data is noisy. A researcher could instrument a particular study, but richer data collection can conflict with a simple user experience.

Asked what organizations would support independent, long-term work, I do not have a complete answer. One approach may be to find a forward-looking group, work there while it remains supportive and move when its priorities change. Academia may offer unusual longevity because teaching and other university work provide an institutional home.

Michael asks why Python Tutor found users in the first place when so many online projects briefly attract attention and disappear. My best guess is that it met a real need: teachers were already drawing these diagrams by hand, and this put the process on a computer. At the time I did not know whether it would work; when usage began to grow, I kept it going.

An audience member asks whether I expected any of my early side projects to succeed. I did not particularly want to start a company or even build something that grew, despite being surrounded by startup culture. My PhD fellowship also gave me room to stay in school; I made things largely because the work was intellectually fulfilling.

Sur la même chaîne

1:44:59EN
Stanford Online· 21 nov. 2026Evaluate LLMs by the Work You Need Them to DoAfshine Amidi, Shervine Amidi · Stanford UniversityHuman ratings, model judges and benchmarks each miss different failures; Stanford’s instructors show how to combine and test them.ÉvaluationsAI agentsModèles et entraînement+1
48:59EN
Stanford Online· 19 nov. 2026Cynthia Bailey: Choose AI models for the task and guard your API keysCynthia Bailey · Stanford UniversityStanford’s Cynthia Bailey shows how token overhead, repeated agent steps and exposed credentials turn AI use into real costs.AI cost & performanceAI agentsAI Safety+1
49:17EN
Stanford Online· 19 nov. 2026Stanford graduate study: start with a course before committing to a degreePax Heymeyer, Jonathan Nottman · Stanford Engineering Center for Global and Online EducationStanford’s online team explains course-by-course admission, credit, workload, exams and 2026–27 tuition.Education
6 diapositives1:13:23EN
Stanford Online· 18 nov. 2026Terminal-Bench: How to Test Whether AI Agents Can Do Real WorkAlex Shaw, Ryan Marten · Laude InstituteThe benchmark's creators explain why useful agent tests need realistic tasks, reliable graders, and safeguards against shortcuts.ÉvaluationsBenchmarksIngénierie d’agents+2
6 diapositives58:52EN
Stanford Online· 17 nov. 2026Cynthia Bailey: Define What an AI Agent Can Do and When It Must StopCynthia Bailey · Stanford UniversityConnectors let models act through other apps; explicit permissions, human checks and stopping conditions determine how safely they do it.AI agentsMCPIngénierie d’agents+2
4 diapositives1:20:34EN
Stanford Online· 17 nov. 2026When to Optimize an AI Agent’s Prompt or Its WeightsMichael Ryan · Stanford UniversityMichael Ryan compares prompt search, reinforcement learning and distillation, starting with the evaluation that makes any improvement measurable.Prompt optimizationIngénierie d’agentsReinforcement Learning+2