Towards RSI for the $1.8T U.S. learning economy

People learn in different ways. They have different goals and deadlines, and they get stuck in different places. A good tutor works this out for each student and plans around it, which is why families still pay for tutors when AI help is free.

This holds well beyond school. A student preparing for the SAT, a nurse learning a new protocol and an engineer moving to a new stack each need a plan that fits them.

Learning technology has improved in stages. Each stage made one more part of learning respond to the person. Planning the path has stayed out of reach.

Most products today follow a fixed course, pick the next question from right and wrong answers, or let an AI tutor improvise the next session from its memory of past chats. Tools like Gemini and Aristotle look at earlier sessions, and some sketch a study plan. That plan comes from prompting the same general model used for everything else. Nothing models how this person should be taught, and nothing learns from results which plans work.

Companies already build different strategies for different subjects. Inside a subject, every learner gets the same one. A strategy that fits a person has to be built for that person and keep changing as they change, and no team can write that by hand for every learner.

What we think is needed

A system that builds a plan for each person, changes it from their results, and gets better at building plans across everyone it teaches. When the system also improves the way it makes those changes, it is improving itself recursively. This is recursive self-improvement?Recursive self-improvement. A system that gets better at its job, and also gets better at the way it improves itself., and it is what we are building.

Here is where today’s products sit. The further right, the more personal the plan. The higher up, the more the product improves itself from results. Bigger dots cost more. No product reaches the top right yet.

See our analysis for the research behind this and a product-by-product comparison.

We want an AI system that teaches better with every learner it works with, and that also gets better at the way it improves.

Human experts define what good teaching looks like. The system will learn from their corrections, and over time it will need them less.

It works as three loops, one inside the other.

We want it to serve the whole learning economy, one step at a time.

Every step will run on the same engine, so what it learns in one area carries over to the others.

We are starting with tutoring.

This kind of system takes several stages, and possibly years of work. So we are building toward it one stage at a time.

We started with two precursors of the harness. Both are self-modifying harnesses with human approval, at two different intensities.

A high-touch version sets the bar for learning outcomes, with a detailed expert correction on every change. A low-touch version will test how close we can get to that bar with far less human time, across many more learners.

Learn 0.1a gives us detailed data from a few learners, and Learn 0.1b will give us lighter data from many more. Our work is to bring the results of Learn 0.1b up to those of Learn 0.1a.

Learn 0.1a: high-touch review of every proposed edit

Our first product is Learn 0.1a, a harness?Everything one learner gets: their AI agents (such as a tutor and a quiz maker), their plan and their lesson formats. that edits itself, with a person approving each edit. It is live today.

We started with tutoring, where an expert works with every learner and outcomes are easy to measure. Learn 0.1a runs inside LessonOrca tutoring. Early results there have shown an improvement of about 150+ points in SAT scores for our first students.

Learn 0.1a proposes edits to two things: the inner harness of each student, which is the curriculum and the agents that work with them toward their goal, and its own strategy for making those edits.

Human reviewers guard every edit. They approve, correct or reject each self-improvement diff?One proposed change to a harness, shown as before and after, so a person can approve, correct or reject it., and they also teach weekly sessions and check in with students. We save every correction, but we do not use them for training today. We plan to use them as labeled examples of good teaching once learners and their families give explicit consent.

Learn 0.1b: light supervision, with a human control center

Next, we are launching an app where the same harness runs with much less human help. One person will oversee 75 to 100 learners from a control center and correct their harnesses. Each learner will need about 10 minutes of review a week.

Learn 0.1b covers the SAT, PSAT, ACT, AP and subject tutoring. Orca, the AI tutor, teaches every lesson, and there are no live lessons.

Learn 0.1b is built for a much larger scale than Learn 0.1a, so it will collect much more data.

See the Learn 0.1b page and how it works.

With consent from learners and their families, we plan to use what we learn from Learn 0.1a and 0.1b. It will show us where human help adds value and where it does not, which core changes the harness needs, and how to build the next version.

Each stage starts when the one before it reaches its metric.

Our goal is to go from Learn 0.1a and 0.1b to Learn 1: the first recursively self-improving harness for learning anything, under any constraints on how you learn, like deadlines or personal preferences.

Learn 1 should be universal. A learner puts in their goals, and it comes up with agents, activities and progress tracking, then keeps updating itself to be a better learning system for that learner.

From there, the same engine extends to new domains, like Copilot 1 for year-round school support.

Every learner starts with two things.

The first is a foundation harness?The harness every new learner starts with in a subject. Each learner's harness begins as a copy of it. for their domain: a set of AI agents?AI helpers with one job each, for example a tutor, a planner or a quiz maker., a curriculum and delivery formats, built around their goal and deadline.

The second is a personal optimizer, called a strategy model?A second AI for each learner. After every session, it reads the learner's results and proposes changes to their harness.. It watches their progress and rewrites the harness as they learn, including its prompts, its agents and its plan. Each harness reshapes itself to fit its learner over time.

Pod 1

Above every learner will sit Pod?The part of our system we are building to learn across all learners. Pod will find the changes that help and add them to the starting harness., the distributed optimization layer that we are building.

Pod is what will make the whole system self-improving. It will pool anonymized outcomes from every learner, measure which harness changes actually help, and merge the winners back into the foundation for everyone.

We plan to build the first version of Pod from Learn 0.1a and 0.1b, with learners’ explicit consent. Learn 0.1a would feed it deep, expert-corrected data, and Learn 0.1b would feed it scale.

Each stage pays for itself. With consent, the data from each stage will help build the next one.

Learn 0.1a (live)

Families pay for high-touch tutoring through LessonOrca, with a weekly human tutor and daily AI sessions. See the prices for SAT and ACT and for other subjects.

This tutoring will also give us our highest-margin data, because every session comes with expert corrections and a clean test score outcome.

Learn 0.1b (waitlist)

A consumer app for the SAT, PSAT, ACT, AP and subject tutoring, at $35 a month. One reviewer will oversee 75 to 100 learners from a control center, at about 10 minutes per learner each week. Join the waitlist.

Infrastructure: the Orca API

We will sell our own products and the engine underneath them. LessonOrca and the Learn app are our products. The Orca API will let other companies run self-improving harnesses for their own learners.

Customers will pay a monthly fee per active learner for the persistent learner model and harness, plus metered usage per session, with the underlying model costs passed through at cost. Later, we plan to train our own learner and reward models. Those will become endpoints with a price per call.

Every API deployment will send anonymized outcomes back to the shared engine, so each customer will make the system better for the next one.

Orca21 Labs is two founders at UC Berkeley.

Abhiraam is a machine learning researcher. He is first author, with Google and MIT, on a self-supervised foundation model that lets biomedical models train with 60–90% less labeled data, and he built the recognition model behind the U.S. Treasury’s EyeNote app, which serves 100,000+ people.

Waleed works where AI meets business. At Databricks, he did AI/ML infrastructure research with the field engineering and product teams, and wrote strategy material for enterprise customers with large AI workloads. Before Berkeley, he founded and ran three companies: WEInstruct, an SAT and ACT tutoring company where he designed personalized learning plans for 30+ students; Tin2Totes, a profitable recycling operation with 50+ collection sites; and The Bag Project, which raised $10,000+. As Washington DECA State Vice President, he led 15,000+ students.

Both of us have tutored the SAT since high school.