Joe Liemandt on Alpha School, AI Tutoring, and Rethinking Education

Joe Liemandt — the Stanford dropout who built Trilogy Software and now runs Alpha School — argues that a mastery-based AI tutor lets children finish a day’s academics in about two hours, that motivation, not intelligence, is the real bottleneck to learning, and that the teacher-in-front-of-a-classroom model wastes most of a child’s time.

Key ideas

1. Two-hour learning: master the academics, free the rest of the day. Alpha’s pitch was originally ‘2X learning’ — your child learns twice as much — but parents pushed back until Liemandt reframed it as two-hour learning: crush the academics in two hours so the child gets the other four back for projects and life skills. The content, he insists, is simply not that voluminous — a grade-level subject takes twenty to thirty hours to master, against the couple of hundred hours a normal school spends. Seventh-grade science, 180 school days in the standard model, takes the average Alpha child twenty-two hours to mastery on Timeback, the school’s app.

2. Learning is effort-based, not IQ-based — once you enforce mastery. The standard system is time-based: you advance every year regardless, so gaps compound and knowledge, being cumulative, collapses in high school. A mastery-based AI tutor only lets a child advance once the current material is known cold, which converts learning from an IQ contest into an effort one. Liemandt claims every child can get 100% on a state standardised test — his ‘100 for 100’ programme pays $100 for a perfect score and routinely produces more perfect scores in a building of 200 than a district of 100,000.

3. Motivation is the bottleneck, and the school is built to solve it. To teach a child you need lessons of the right difficulty and a motivated student; AI handles the first easily but the second is the hard problem. Alpha negotiates two hours of genuine engagement in exchange for the rest of the day, and a ‘waste meter’ streamed from the child’s screen coaches them to stop skipping videos and using ineffective techniques. Kids do the academics to earn time for the hard, meaningful projects they actually love.

4. AI generates lessons; it does not chat. ‘Chatbots are cheatbots’ — Liemandt bans conversational AI on the academic side because most children use it to cheat. The real tool is generative AI that fuses the curriculum, the child’s knowledge graph, and their interest graph into a personalised next lesson grounded in learning science (the zone of proximal development, worked examples, cognitive-load theory). Each child in the same room can be on a different grade level entirely.

5. Replace teachers with guides, and hold high standards with high support. The classroom teacher’s job bundles five hard, underpaid skills; Alpha unbundles it. AI teaches the academics; adult ‘guides’ — the top 1% of teachers, hired only for their ability to connect and motivate — coach children through hard things, backed by a dean of parents so they never have to manage adults. Half of educators quit when told they are accountable for whether the child learns.

Content

Why the system feels broken

Liemandt opens on the parental anxiety driving change: an AI world is coming, and the education both host and guest went through plainly will not prepare children for it. But he insists the rot predates AI — test scores keep falling while spending rises. His diagnosis is structural. The standard model is time-based, moving children up by age with a teacher at the front, and it rewards exactly two traits: high IQ and Big Five conscientiousness (the natural grinder). A child strong in both thrives; everyone else is left behind, and no amount of money changes those two levers. The second structural fault is that educational outcome tracks family income more tightly than almost anything else. In the United States specifically, he notes, academic outcomes and earning potential are only loosely correlated, which lets parents quietly decide the academics do not matter much — and so everyone in the ecosystem lets standards drop. Grade inflation (80% of Harvard students get A’s) is the visible symptom.

The results, and the selection-effects objection

Alpha is a high-end private school expanding to 25 campuses. Liemandt is a militant believer in standardised tests precisely because nobody believes the model works: when children doing two hours a day post top-1% MAP scores (NWEA) in every subject and 790–800 on the maths SAT, sceptics relent. He publishes the results. The distinctive claim is growth rate, not selection: on the 300-point MAP scale a normal child gains about five points a year, an Alpha child ten; and where the median US student rises just one point across all four years of high school, Alpha’s accelerate. He meets the obvious rebuttal — ‘it’s just rich kids in Austin’ — head-on. As a product person he embraces selection effects rather than apologising for them: every school selects, so benchmark Alpha not against hard cohorts but against the best private schools, where its top-1% children still outscore their top-1% children. The forthcoming Texas Sports Academy tests the model on a wider funnel: $15,000 tuition offset by a $12,000 Texas voucher to roughly $300 a month, drawing children who come in academically bottom-half but love sport.

Mastery, and the two-sigma backdrop

The intellectual anchor is Bloom’s two sigma problem — one-to-one tutoring with mastery produces performance two standard deviations above classroom teaching — a result known for forty years but useless in practice, because you could not give every child a human tutor and because enforcing mastery is hard. Liemandt defines mastery through sport: a coach will not let a point guard lose the ball 20% of the time, yet academics wave a child on at 80%, so holes accumulate. Knowledge is cumulative; a child who hates high-school maths usually missed something in fifth grade. A mastery-based AI tutor diagnoses the gaps, fills them without social stigma (nobody sees you doing third-grade material), and only then advances — a design a single teacher legally required to teach at grade level cannot replicate. Post-COVID the span in one classroom stretched to seven years, and social promotion pushed unready children forward, making the classroom model still less workable.

How the AI actually teaches

Liemandt is careful to say what he does not mean by AI. Chatbots are cheatbots; the school does not teach by handing children ChatGPT. The unlock is generative AI that ingests the curriculum, the child’s knowledge graph (what they know and do not), and their interest graph (baseball, the Avengers) — and, from 2026, cognitive-load theory, treating working memory like a chip’s registers with a personalised number of repetitions to move material into long-term memory. It generates the perfect next lesson: a worked example, short, engaging not passive, pitched to keep the child at 80–85% accuracy, the zone every video-game designer knows. The second and most expensive component streams the child’s screen to a top model that coaches behaviour — the waste meter — costing roughly $10,000 per child today, which he expects to drive below $1,000 then $100. The product is named Timeback, and its most important job is motivation.

Guides, life skills, and high standards

Because academics take two hours, the afternoon is free for project-based workshops teaching five categories of life skill: leadership and teamwork; storytelling and public speaking; grit and hard work; entrepreneurship and financial literacy; socialisation and relationship building. Alpha quantifies the soft: every third-grader does a Rubik’s cube, every eighth-grader passes a teamwork test by finishing a Tough Mudder together, fourth- and fifth-graders pass a Wharton MBA leadership simulation, and fifth-graders launch food trucks (chosen, one child explained, for the gross margins on breakfast food). The engine underneath is high standards with high support: children are made to do hard things — kindergartners scaling a 40-foot rock wall, second-graders trained on Atomic Habits to run a 5K — supported by a caring adult through the struggle-fail-cry-succeed loop that child-development research treats as the core of growth. Liemandt argues the mental-health epidemic of disengaged, scrolling middle-schoolers is the fruit of low standards, not high ones. The role of the teacher is rebuilt into a ‘guide’: AI carries the domain expertise and the learning science, a dean of parents absorbs the parent relationship, and the guide is hired solely to connect, motivate, and hold the line — the one or two teachers who transformed each of our lives, scaled.

Trilogy, Jack Welch, and accountability

The management instincts come from building Trilogy Software, which Liemandt started after dropping out of Stanford, working hundred-hour days once he had found what he loved — the passion he now wants to ignite in children before they leave school. His father refused to invest in the company, reasoning that a good idea would attract other people’s money. His mentor after his father’s death was Jack Welch, from whom he learned high standards (‘is this your first step or your last step?’) and, decisively, ROI: ‘if GE doesn’t get an ROI, your product, you, and your company suck.’ He carries that into education as radical accountability — Alpha may be the only school that treats a child’s failure to learn as the school’s fault, never the child’s. He took PCorder.com public, watched it implode in the dot-com bust, bought it back, and kept Trilogy private ever since, preferring total control: on Timeback you deliver the three commitments — the child loves school, learns 2X in two hours, and builds life skills — for every kid, or you do not use the system. Roughly half of educators leave when held to it.

Scale, and the wager

The academic engine scales with compute; the hard parts are physical schools and virtual reach. Alpha governs new campuses with ‘painfully insightful metrics’ — surveying children every eight weeks (‘do you love school more than vacation?’, answered yes by 40–60%) and parents (‘do you trust your guide to hold high standards so you can give only unconditional love?’). The virtual plan opens Timeback to outside builders in 2026 — a ‘Shopify for schools’ letting entrepreneurs run Montessori, wilderness, gifted, or sports variants on the same engine — plus a homeschool version whose central problem is recreating motivation once Liemandt no longer owns all six hours of the day. His first graduating class of twelve sent eleven to their first-choice colleges (Stanford, Vanderbilt, NYU Shanghai) where all posted 4.0s and dismissed the large lecture as a waste of time, citing Harvard’s own study that an AI tutor beat its teaching. Asked what success is, he answers: transform education for a billion children — ‘our job is to make this the best time in history to be a 5-year-old.‘

See also

• Shane Parrish — host
• Joe Liemandt — guest; founder of Trilogy Software and founder-principal of Alpha School
• Andrej Karpathy on AGI Timelines, RL's Limits, and the Future of Education — a technologist’s parallel vision of AI as personal tutor and the reinvention of learning
• Alison Gopnik on Childhood Learning, AI as a Cultural Technology, and Rethinking Nature vs. Nurture — a developmental-psychology counterpoint on how children actually learn and what AI can and cannot do for them
• Bryan Caplan on Learning Across Disciplines — the sceptical case that much schooling signals rather than teaches, a foil to Liemandt’s outcomes-first mastery model
• James Clear on Atomic Habits, Systems over Goals, and Identity-Based Change — the 1%-better-every-day and identity-change ideas Alpha teaches second-graders to run a 5K