A $9 Tutor That Refuses to Give the Answer

Shawn Jansepar is director of engineering and AI product lead at Khan Academy, the nonprofit whose mission is a free, world-class education for anyone, anywhere. The organization started with Sal Khan making tutoring videos for his cousin and now employs more than 150 people across learning science, instructional design, and software. Jansepar joined in 2017 after teaching a university course on building web applications. Nathan Labenz asks him to rewind to the GPT-4 story, before the public had any idea the model existed.

Around October 2022, in the evening, Jansepar was suddenly added to a Slack channel where a bot would talk back. An email went to a handful of key people at Khan Academy: they were chatting with OpenAI about a new model, GPT-4. ChatGPT had not launched. He had barely used the company's earlier APIs. After thirty minutes he felt as if he were interacting with an omnipotent being that knew everything. He had to go for a walk around the block to take it in. Over time the limitations showed up. For those first thirty minutes, he says, it felt like AGI. It was not AGI. It was going to change a lot.

At a company onsite shortly afterward, the few people with access used GPT-4 in a hackathon. At first they were hesitant to put anything in front of learners. They were more interested in generating content and lesson plans for teachers. Safety was the objection. Over time they found mitigations, including OpenAI's moderation API, and in December they started working with OpenAI more closely. Someone on OpenAI's side named Jessica built a quick demo: what if the model were a smarter search, guiding students to the right content? That concierge idea opened their eyes. Weekly status meetings had been going nowhere. They switched to morning and evening check-ins and rapid prototyping.

The first version was a Chrome extension hanging off every Khan Academy content page. A student could say "I don't really get this," and the model knew what "this" meant. Jansepar calls that feeling magical. Leadership was impressed even though the tutor still made plenty of math mistakes. The next week a group flew to Mountain View, prototyped in OpenAI's offices, and tested with students at Khan Lab School — a related but separate organization whose campus sits under Khan Academy's offices, with parental consent already in hand. The button students loved most was "Why should I care about this?" That popularity, he says, is a report card on education: adults rarely bother to explain why a topic exists.

ChatGPT still had not launched, so the students expected a basic chatbot and were blown away. In January the whole company was brought under NDA for a second hackathon, ideas were converged, and engineering, design, and product worked through March 14, the GPT-4 announcement. OpenAI was shipping weekly candidate models. Some partners liked a storytelling-heavy model that was weaker at math. Math was non-negotiable for Khan Academy. The production candidate arrived over a weekend; leadership, key engineers, and designers tested it hard. It had, as OpenAI had hoped, the best of both worlds. The launch went smoothly.

The rush left a pile of tech debt, and Jansepar does not regret it. OpenAI treated Khan Academy as a key education partner. Once the product was out, a lot of potential competitors looked at the brand, the trust, and the execution and asked why they would even bother entering the space. Being first, he argues, paid the interest on the debt.

The dream of a personal tutor for every child is older than the technology that can support it. Labenz asks whether Khanmigo is chasing Benjamin Bloom's two-sigma study, a pillar of Khan Academy's pedagogy from the beginning: one-on-one tutoring plus mastery learning produced gains on the order of two standard deviations over a conventional classroom. Jansepar adds the less-quoted half. Even at thirty students per teacher, mastery-based learning — not moving on until the previous concept is at least proficient — is worth about one standard deviation.

Real school does the opposite. A C in algebra 1 still sends you to algebra 2. Gaps compound until the student's knowledge looks like Swiss cheese. Khan Academy's videos already attacked that: you can pause, rewind, and there is no judgment if you missed a day. Large language models, he says, let them emulate the one-on-one end of Bloom's curve. A human expert still wins on advanced material. An AI tutor is on demand, does not judge, and from day one can know the student's full learning history — something a newly met human tutor usually does not.

Whether the AI will match a great human tutor's results remains to be seen. Khanmigo, powered by GPT-4, still makes math mistakes and gets confused on harder concepts. Getting math perfect is, in 2023, an unsolved research problem for language models. Jansepar believes it will be solved. If it is, he can imagine the tutor being as powerful as a really great human one.

Khan Academy is a small organization. Without GPT-4's zero-shot and few-shot behavior landing in their lap, Jansepar doubts they could have built this: training their own model would have been a huge effort. The product's core constraint is Socratic tutoring — guide the student, do not hand over the answer. GPT-3.5 does not follow that instruction well. GPT-4 does. That gap made GPT-4 essential, and it is why tutoring inside content is still a no-go on anything else.

The mix today is still the vast majority GPT-4. Around the tutor sit "activities," which Jansepar compares to the demo disc that used to come with a game console: craft a story, talk to a historical figure. Those can live on GPT-3.5 because instruction-following is less critical. Interest extraction for students who opt in, and conversation insights for teachers — is the student off topic, where are they stuck — are also fine on 3.5. LLaMA 2 looks promising; they have not tried it. Opportunity cost on the roadmap is the real constraint.

Labenz's cost advice matches theirs: get the thing working on GPT-4, shave cost later. Jansepar puts it in startup language. Go from zero to one. Sort monetization after you have something amazing. Better that ten people absolutely love the product than that you build something mediocre for ten thousand. The phrase he uses with his team is "focus on finding the magic." They have barely scratched the surface of what sits beyond back-and-forth chat. An essay tool in progress will highlight grammar and spelling in the UI rather than forcing a paste into a chat box.

Building Khanmigo did not follow Khan Academy's older waterfall — PM writes requirements, design draws, engineering implements. There was no perfect spec for a tutor. They locked designers, engineers, product managers, and learning scientists in a room, built a prototype, demoed it, and iterated. The process was inspired by *Creative Selection*, a book about how the iPhone was built. Demo-driven development, Jansepar says, worked.

They are not fine-tuning GPT-4 in production; that API was not available, and every turn already burns tokens on article context. Before March 14, OpenAI gave them direct access to an internal fine-tuning tool. Sal Khan and people on the content team went in and labeled what a good math-tutor response looks like versus a bad one. A typical failure was getting lost: the model would ask "what do you think the next step is," the student would reply with an equation, and the model would treat the equation as a new problem instead of an answer. After that labeling, tutoring quality improved significantly. They did it for roughly 100 questions, each with a ton of variation.

Labenz asks whether this was really a role in OpenAI's reinforcement-learning process rather than Khan Academy fine-tuning its own weights. Jansepar says yes: Sal's voice is now encoded in the "great tutor." The work did not stop the model from getting 2+2 wrong. It made the model better at tutoring math. That distinction matters because GitHub supplies millions of lines of good code, while the public internet barely contains traces of what a good tutoring turn looks like. Someone had to feed it by hand.

Khan Academy, Jansepar says, used to be too data-driven. Data-driven means long feedback loops: you wait until a dashboard assures you. Data-informed means you still look — they have a labeling infrastructure and try to score as many interactions as they can — but you let experts short-circuit the wait. Efficacy studies that claim the product works still have to be data-driven. Rapid product iteration does not. Labenz compares this to the then-recent "GPT-4 is getting worse" panic, which collapsed once people read the raw transcripts instead of a collapsed benchmark score.

Internally they adopted Amazon's one-way-door versus two-way-door frame and wrote it into org decision records. A two-way door you can walk back through: move fast, shorten loops, use intuition. Khanmigo launched under a Khan Labs banner as an experimental feature for people willing to be testers, including district pilots. A one-way door is different. Before Khanmigo they were moving a Python monolith to Go services because Python 2 was being end-of-lifed on Google Cloud. Language choice was a numbers exercise — developer productivity versus server cost — and they took months.

Labenz asks which tools they wish they could have brought back in time. Jansepar names LangChain, which they are now rewriting toward so they can swap models and get JSON out reliably, and a real labeling UI to replace a custom Google spreadsheet. They also built an internal interface for prompt engineers — a mix of engineers, PMs, and content people — with regression testing next. Those tools are becoming table stakes for every tech company. The price of being ahead of the curve was inventing them while the clock ran.

They turned Khanmigo on everywhere rather than fencing off hard subjects. Khan Academy teaches through roughly second-year university courses. If string theory goes badly, that is a two-way door: they can turn it off. The non-negotiable is harm. An extra moderation layer tries to stop the tutor going off-topic or into inappropriate territory and alerts teachers. Younger students who can handle first- and second-year material, Jansepar thinks, will notice when it is weak and can send feedback; they opted into an experiment.

Labenz cares about a subtler failure than hallucination: noticing when the *user* is confused and pushing back on a false premise. Khanmigo's edge over raw ChatGPT, Jansepar says, is the wrapper. If a message looks like math, a first call asks the model to think out loud: look at what the student said, look at the question, did they get it right, and why. That inner monologue is not shown — it is the tutor collecting themselves before answering — then fed into a second call that actually speaks to the student. The gain on math was large. They would run the monologue on every reply if it were not so expensive. They reserve it for places where accuracy really matters.

On exercises the model already has the question, the answer, and the content team's written hints, so it does best there. "Tutor me STEM" lets a student paste any math problem without a known answer, and accuracy drops. They are considering solving it first with Wolfram or a Python backend and injecting the steps. Embeddings showed up as a hallucination patch: users hit a 200,000-token daily cap branded "AI power," then asked the tutor what AI power was, and it did not know. A knowledge-base lookup now injects an answer when similarity clears a threshold. An earlier "related links" feature sent Algebra 2 questions to an MCAT course and was pulled.

They run on a dedicated OpenAI instance, not the shared pool. The vendor's advice was to put as much text as possible that is common to every user at the front of the prompt and park personal variables at the bottom. Doing that raised the cache hit rate substantially. On jailbreaks, Labenz tried ignore-previous-instructions tricks and a then-new universal attack developed on LLaMA 2 that was said to transfer to GPT-4; none of them landed. Jansepar credits wrapping every user message with a reminder to check that the student is on track, plus running every message through the moderation API for sexual and other flagged content, on top of GPT-4's own improved refusal behavior.

What is in production is not very personalized. Students can pick a reading style — simple, complex, professorial — and on an exercise the tutor knows whether they are unfamiliar, familiar, proficient, or mastered on *that* item. Nothing spans items. A student complained that she had just discussed sine, switched to tangent, and was asked what she knew about sine as if the prior conversation had never happened. On the backend each question starts a new thread. Users expect a continuous relationship. Continuity and opt-in interest collection are on the roadmap, so "why should I care" can mention soccer without asking first.

They already launched an AI course in March that teachers can assign, and support articles — some of them in the embeddings database — explain how the system works. The UI says at the bottom that Khanmigo makes mistakes sometimes, with a link to a Zendesk article on why. Logs are visible to teachers and to parents who granted access. Trust, Jansepar says, depends on that transparency.

Text-to-speech sits high on the roadmap. A lot of the product is still a wall of chat, and even "explain it simply" is a lot to read. Plenty of seventh graders have fifth-grade or even second-grade reading comprehension. They are trying multiple TTS vendors and may let students unlock voices as an engagement mechanic. Vision is less far along. He can imagine scanning homework from a phone that does not yet run the app, or watching a student draw on a Chromebook exercise and interrupting a repeated mistake with "why are you doing that step?" They have not thought that through deeply, because students sometimes need to make the mistake. The roadmap is crowded.

Near-term metrics are softer: more time on Khan Academy, more work completed. Story-crafting still counts, he notes, as reading comprehension. The stake they actually want in the ground is a standardized test. Years ago they partnered with NWEA on MAP Growth. Students tested, teachers assigned per-student goals in measurement, geometry, and other strands, students worked at their own pace, then tested again. An efficacy team found a statistically significant improvement for Khan Academy users versus non-users — without Khanmigo. The hoped-for sequel is the same design with a Khanmigo cohort against content-only. That comparison is the holy grail. They are a little ways away. The logistics are heavy.

The mission is still free world-class education for anyone, anywhere. Tokens are not free, though the product is still much cheaper than a human tutor. For now they sell to school districts with a per-student, per-month premium, and they do not hunt the richest districts. They target schools with a high share of historically under-resourced learners, using free-and-reduced-lunch rates as a proxy. If a district cannot pay, they look for local corporate sponsors — a Kansas company sponsoring Kansas students. A longer-term idea is a museum-style membership in which people who pay more offset licenses for kids who cannot. Labenz's own access was a nine-dollar monthly donation, less than half of ChatGPT Plus.

Jansepar runs the learner experience; another director owns teachers. Teachers, contrary to the fear that AI would look like a job threat, have taken to it. They report saved time. Students like a 24/7 tutor that does not judge, even when it is occasionally wrong. A major unexpected win: students whose first language is not English can talk to Khanmigo in their native language. The team had worried teachers would see replacement. They do not envision classrooms disappearing. School is social. If Khanmigo takes most of the one-on-one load, maybe ten raised hands become two, and the teacher can run project-based work instead of triage.

The one-on-one tutor is only the first shape of the product. On the later roadmap are multi-user activities: the model notices that five kids are stuck on one concept and ten on another, and asks the teacher whether to make a breakout. Students struggling with the same idea could chat with the AI as a group. A class might take turns contributing to a single story — Mad Libs, but with a model — or have the AI facilitate and grade debates. New technology's first instinct, Jansepar says, is to emulate the old world. The magic is still mostly undiscovered.

Asked what story he tells himself about global scale, he says the question gives him goosebumps. He joined in 2017 after watching Sal's TED talks about finding and filling gaps. He assumed the software already did that. It did not. At best, finding a gap was O(log n): jump to the middle of the skill list, chop, repeat — or walk backward through it. Nothing pinpointed the specific skill that was blocking grade-level work. Preventing that Swiss-cheese pattern is the opportunity. Not everyone has a small classroom or parents who can help with every subject.

Sal's videos leveled one playing field: everyone in the world could have the same teacher, even if Sal was not the single best person for every topic. Khanmigo is the next step, back toward the effect sizes of one-on-one tutoring. Jansepar cannot wait for the efficacy studies, because he thinks they will show that this actually helps students — and that it can change the world. We will see.