On Built This Week, AI in education has come up in three places: math, workplace training and study tools. Carina, founder and CEO of Axiom Math, is building an AI mathematician whose proofs a computer can check, written in a language more universities now teach. We used AI to write, film and build a compliance training course for our own team. And both hosts see AI adapting to each student, with one condition Sam keeps coming back to: students still have to think for themselves.
It's one of these things that obviously can be hyper customized to each student and their learning abilities and what their interests are, and that sounds like right up the alley of, you know, proper education, I would say.
— Jordan Metzner, Episode 3
How companies we talked to use AI in education
Axiom Math: an AI mathematician, checked in Lean
Axiom is building a model that proposes new problems, proves them and auto-formalizes existing math into Lean, a programming language for proofs. A Lean proof either compiles or it doesn't, so nobody has to trust an AI judge or hunt for a hidden bug in a 3,000-line proof written in English. The hard part, Carina said, is data: there are more than a trillion tokens of Python code and only about 10 million tokens of Lean. (Episode 22)
We just run it like you run a Python computer program. I think that's the biggest difference between these formal models and the large language models.
— Carina, Axiom Math, Episode 22
Lean in the university classroom
Jordan asked whether math students now write their theorems in Lean, the way calculators once changed math class. Carina listed the University of Washington, Carnegie Mellon, Stanford's automated reasoning lab, and universities in Europe and the UK. Mathlib itself began as a teaching effort by Kevin Buzzard at Imperial College London, who wanted his freshman proofs class to think more rigorously. The students who jumped in were mostly advanced ones who didn't need the class, and two or three of them now work at Axiom.
Students and teachers: notes in, podcasts out
In Episode 26, Sam passed on what one of his daughter's teachers told him: many students put their week's notes into NotebookLM, generate an audio overview and listen to it during a workout or a walk. Sam turned a plain-text onboarding Google Doc for new recruiters into a three-phase infographic in about five seconds. Jordan has seen NotebookLM used by elementary school students, college students, doctors and scientists, and pointed to its flashcards, mind maps and custom quizzes.
What we built
- AI compliance training and certification platform: our alternative to training software that, in Jordan's words, feels like 1990s videos. AI organized the curriculum and built the platform, and every scenario video came from Google Veo 3.1 Fast. Each exam has 10 questions, learners must watch at least 95% of each video, and Jordan set the pass mark at 100%. Sam pointed out the same setup could run a new-hire orientation with a certificate at the end. (Episode 20)
- LLM Math Roaster: Jordan's build for Axiom. It sends a problem to Gemini 2.5 Pro, GPT-5, Claude Sonnet 4.5 and Grok 4 Fast for Lean 4 proofs, then ChatGPT scores them on a leaderboard. On Fermat's Little Theorem, Gemini scored 98. Carina noticed one model had proved it for natural numbers instead of integers, and suggested compiling the proofs instead of trusting an LLM judge.
What's still hard
- Critical thinking. Sam's biggest fear for AI in the classroom is that it replaces students' critical thinking. Getting an answer doesn't mean the approach is right, and he wants students to keep a critical eye on anything AI generates and the habit of checking for themselves.
- Kids gravitate to what they're good at. Sam sees AI as an accelerant for kids with different learning profiles. His daughter likes math but prefers the parts she already does well, when she should probably spend more time on the parts she struggles with.
- Verified math is short on data. The 100,000x gap between Python and Lean is why Axiom generates synthetic Lean by converting informal math. Carina said some companies hire large numbers of human Lean programmers instead, which is expensive and doesn't scale.
- AI grading AI. Our Math Roaster leaned on an LLM judge. Carina's point was that a compiler is a better source of truth than another model.
- Generated lessons need a review pass. About 60 to 70% of our compliance course content was right on the first try. The rest took a few more passes.
Both hosts still land on the hopeful side. Sam compares AI in school to the internet twenty years ago: his daughter already used iPads and online tools in third and fourth grade, so the real question is how to use it safely. His five-year-old picked up a stack of space facts from a few conversations in ChatGPT voice mode, and Jordan sees AI as a companion for children's learning as a real opportunity. Jordan is also quick to say he's no education expert.
Episodes on AI in education
- Episode 3: Auto-Marketing for Podcasts: the hosts react, as parents, to the American Federation of Teachers' AI training hub with Microsoft, OpenAI and Anthropic (July 2025).
- Episode 20: Compliance, Pomelli, and the Rise of the Robots: our AI-built compliance training platform (November 2025).
- Episode 22: Let's Do the Math! with Carina of Axiom Math (November 2025).
- Episode 26: NotebookLM for studying and onboarding (January 2026).
FAQ
How is AI used in education?
In our episodes: Axiom Math is building an AI mathematician whose proofs are checked in Lean, a language several universities now teach; we built a compliance training platform where AI wrote the curriculum and Veo 3.1 Fast made the videos; and students use NotebookLM to turn their weekly notes into audio overviews.
Will AI hurt students' critical thinking?
That is Sam's biggest fear as a parent of two. He worries students will copy and paste an answer without asking whether the approach is right. His stance is to be proactive and proceed with caution, because AI is already in the classroom the way the internet was twenty years ago.
Can AI build an employee training course?
We built one. AI organized the curriculum and built the exam platform, and every scenario video came from Veo 3.1 Fast. About 60 to 70% of the generated content was right on the first pass. Jordan said comparable platforms usually cost $10 to $100 per user per month.
Do universities teach Lean?
Yes, according to Axiom Math founder Carina. She named the University of Washington, Carnegie Mellon, Stanford and universities in Europe and the UK. Mathlib, Lean's math library, started as a teaching effort at Imperial College London.