← home feed →

A Manifesto for Learning

Schooling has been a compromise. One teacher, thirty children, a calendar, a syllabus paced to the median. We have lived inside the compromise so long it looks like learning. It isn’t — and the cost has been paid in childhood. Hours that should have been play, family, embodied curiosity were spent waiting for the room to catch up, or being dragged faster than the mind could absorb. Most “learning” since has been consumption wearing the costume of study: read the paper, feel learned; watch the lecture, feel taught. The signal that any of it stuck is conspicuously missing.

The mechanisms that make learning actually work are known, and they require something close to a tutor per learner. Until now that tutor was unaffordable for almost everyone. It now costs $20 a month. The only thing standing between this fact and a generation that learns more deeply in less time is the structure we put around the tool. As it ships, AI accelerates intellectual laziness — the default with any tool, but newly efficient. Labs are working on this at the model layer and some of the harm will attenuate. The deeper problem is older than the tool: humans avoid effort, and effort with feedback is what learning requires.

This manifesto has two jobs.

First: invert the default. Make the rigorous path — predict before reading, retrieve before re-looking, explain before checking, reflect before closing — cheaper to walk than to skip. The mechanisms that build durable competence are known and convergent; the work is to install them in a learner’s life so that running them is easier than not.

Second: give the hours back. Not so children become more efficient students — so they can be children, free to play, to wonder, to be slow and embodied, to find what is worth learning on their own terms. Not so adults become more productive learners — so they can pursue what they actually value, in their actual lives, with the kind of attention they have never been allowed to spend on themselves. School becomes one site of this, not the only site. The classroom narrows to what classrooms are uniquely good at — shared inquiry, shared making, the social part of becoming a person — and the rest stops crowding everything else out.

The structure that does this is small. A folder per skill, a loop of five verbs, an agent that withholds answers, a few hooks and routines. The mechanisms are not new. What is new is that they can finally run — for everyone, not just the children of people who could afford the tutor. That is the vision: a generation that learns what it values, in less time, with more of itself left over to be human.

I. What broke

Three failures stack.

Cohort schooling optimizes for the median student and the calendar, not the individual learner. Curriculum precedes calibration. Pacing is external. By the time the topic reaches you, you are either already through it or already lost — and the room moves on either way.

Consumption-as-learning. You can read a book without learning it. Finish a course without retaining it. Watch a tutorial and feel competent up to the moment you must do the thing. The fluency illusion is industrial: videos, summaries, course completions, threads — all surface signals of effort, all decoupled from durable competence. Dunlosky et al. (2013) ranked ten common study techniques against the evidence. Rereading, highlighting, and summarization were rated low utility. They are also the most popular.

AI as substitute, not scaffold. Give a learner an LLM without guardrails and they will produce better work in the moment. Bastani et al. (2025, PNAS) ran this experiment on ~1,000 high schoolers. With unrestricted GPT-4, in-session math scores rose 48–127%. When the AI was removed for the exam, scores fell 17% below the no-AI control. The crutch effect. The same study tested a guardrailed “GPT Tutor” designed to give hints, not answers. It eliminated the harm. It also produced no learning gain. Useful, but not formative.

That study is a snapshot of an early-generation model with no pedagogical character training. Labs are working on this directly — Anthropic, OpenAI, Google all have explicit work on tutoring behavior — and some of the model-layer harm will attenuate. The deeper claim survives the labs: the default path with any tool is to avoid effort, and any tool capable of substituting for effort will be used to skip it — unless the path that doesn’t is the easier one.

II. What the science actually says

The mechanisms that make learning happen are known. None of them are gentle.

Retrieval beats restudy. Adesope, Trevisan & Sundararajan (2017) meta-analyzed 272 effects across 188 experiments: practicing recall produced g = 0.51 over rereading; g = 0.93 over passive exposure. Pulling an answer from your head is the act that consolidates it. Recognition — the warm feeling that “I’ve seen this before” — doesn’t.

Spacing beats massing. Latimier, Peyre & Ramus (2021): g = 0.74 for distributed vs. massed retrieval. The optimal gap between reviews scales with how long you want to remember the thing (Cepeda et al., 2008).

Self-explanation beats consumption. Bisra et al. (2018), 69 effect sizes: g = 0.55. Forcing yourself to explain the material in your own words — before checking — produces durable knowledge that passive reading cannot.

Interleaving beats blocked drill (with boundaries). Brunmair & Richter (2019), 59 studies: g = 0.42. Mixing categories during practice improves discrimination and transfer. Strongest when categories are similar; reverses when they aren’t.

Formative feedback beats grading. Black & Wiliam (1998): d = 0.4–0.7 for frequent low-stakes feedback that the learner uses to adjust strategy. Hattie places formative evaluation at d ≈ 0.9 — among the largest effects in education.

Interactive beats passive. Chi & Wylie’s ICAP framework (2014): Interactive > Constructive > Active > Passive. Tutoring works not because of what the tutor says but because of what the tutor forces the learner to say.

Mastery beats coverage — modestly. Bloom’s 1984 claim of 2 standard deviations from one-on-one tutoring plus mastery learning does not replicate. Nickow, Oreopoulos & Quan’s 2024 meta-analysis of 96 RCTs put the real tutoring effect at ~0.37 SD. VanLehn (2011) found human tutoring at d = 0.79 and intelligent tutoring systems at d = 0.76 — a fraction of Bloom’s headline, but meaningful, and notably equal between human and machine when the machine attends to interaction granularity.

The mechanisms are convergent: retrieval, spacing, self-explanation, interleaving, formative feedback, interactivity, mastery. Effort that gets feedback. They are what tutoring delivers when it works. They are what schooling cannot deliver at scale. They are what you can install in your own practice — with or without a teacher.

III. The loop

Five verbs. Run them on repeat.

Calibrate. Find where you actually are. Write what you can explain unprompted. Record a baseline. Surface the questions you can’t answer. No one progresses from “I don’t know what I know.”

Aim. Pick one thing for the next two weeks. Specific enough to act on. With a “done when” you’d defend. The aim comes from the calibration — from the edge of what you currently know, not from someone else’s syllabus.

Practice. Effort with feedback. Active recall before rereading. Self-explanation before checking. Worked examples while you’re a novice; problem-solving as expertise grows (Sweller’s expertise reversal — high assistance helps novices and hurts experts; Chen, Kalyuga & Sweller 2017). For embodied skills, the same logic applies off the page: drill the target, capture the attempt, get feedback.

Reflect. End every session by writing: one thing that clicked, one still-confused thing, one question to add to the open-questions list. The questions list is the engine. Tomorrow’s aim is somewhere in it.

Measure. Periodically — weekly, monthly — return to your baseline and test. Not to grade yourself. To see whether the aim is still the right aim. Recalibrate.

These verbs are not a metaphor. They are the empirical mechanisms with sleeves on. Retrieval lives in Practice and Measure. Spacing lives in the cadence between them. Self-explanation lives in Reflect. Formative feedback is the inner loop. Mastery lives in the rigor of “done when.”

IV. The agent

A coach with infinite patience, perfect memory of your notes, and the willingness to answer you faster than you can think — is dangerous. The same agent, used differently, is the closest thing to a personal tutor adults have ever had.

The dividing line is whether the agent does the work for you or makes you do the work.

On the scaffold side of the line: Wang et al.’s 2024 Tutor CoPilot RCT (900 tutors, 1,800 K–12 students) found AI-assisted tutors gave fewer answers and asked more guiding questions; their students were 4 points more likely to master a topic, rising to 9 points for students of lower-rated tutors. Kestin et al. (2025, Harvard intro physics) found an AI tutor prompted to scaffold rather than solve produced roughly twice the learning gains of an active-learning classroom in half the time. De Simone et al. (2025, World Bank, Nigeria) found six weeks of GPT-4 after-school tutoring produced 0.31 SD gains — equivalent to 1.5–2 years of business-as-usual schooling.

On the substitute side: Bastani’s 17% exam-score collapse. Fan et al. (2025, BJET) — ChatGPT users wrote better essays but transferred nothing and showed reduced self-regulated monitoring; the authors named it “metacognitive laziness.” Lee et al. (2025, Microsoft Research / CHI) — higher trust in AI correlated with less critical engagement; reported cognitive effort across Bloom’s levels dropped 55–79%.

The pattern is consistent and not subtle:

The agent is a tutor only when the learner is doing the work. Otherwise it is a service. There is a place for services. It is not learning.

V. The artifacts

The workflow has a small physical footprint:

These files are not bureaucracy. They are the substrate on which the loop runs. Reading them back is retrieval. Writing them is self-explanation. Comparing this week’s journal to last is formative feedback. The convention is the mechanism, made tangible.

The agent operates on these files. It can quiz you from questions.md, propose the next aim from the gap between goals.md and progress.md, run a practice session from focus.md, and refuse to close a session until you have written a reflection. Daily and weekly automations pull you back when you drift, gently — the system is opportunistic, not coercive. A skipped day is data, not failure.

VI. The wider claim

The mechanisms generalize. Anywhere a human is trying to become better at something difficult, the loop applies.

Tutoring is the natural form. A good tutor’s session is the loop. The folder is the shared artifact between tutor and tutee. The role of the AI agent is not to replace the tutor but to make every learner their own tutor’s apprentice — running the mechanisms in the long stretches when no human is in the room.

Self-directed adult learning is the default form. It’s what most of us actually spend our lives doing. The convention plus the agent makes it sustainable for the first time at scale.

Higher education and research are strong fits. A PhD is the loop on hard mode — vague goals, faint baselines, ambiguous measurements. The discipline of the workflow does not solve research; it survives it.

Formal schooling is where this matters most. Cohort pacing and standardized curricula were the prosthetic for a world in which one-to-one tutoring was unaffordable. They are no longer the only option. With the mechanisms running for each child at their actual edge, classrooms can narrow to what classrooms uniquely do — shared inquiry, shared making, the social work of becoming a person. Lectures shrink. Practice individualizes. The hours saved go back to childhood. Children are the load-bearing case: if the structure works for them, it works.

This is not a critique of teachers. It is a critique of structure. Bloom’s 1984 paper, even with its inflated headline, identified the real problem: classroom instruction loses information by being collective. Tutoring recovers it. The workflow generalizes the tutor’s mechanisms to anyone who will run them.

VII. The promise

If you do this, here is what happens.

You stop confusing time-on-content with progress. You can answer “where am I” with evidence, not vibe. Your questions become better. The thing you are trying to learn becomes a stack of small, falsifiable claims you are working through, not a fog. Months in, you re-read your old journal entries and find someone who didn’t know things you now know — and didn’t know they didn’t know them. That is the only signal that has ever counted.

The agent is not your teacher. It is the patient apprentice your tutor never had time to be — the instrument you play to make the mechanisms real.

And when this runs at scale — for the children who would otherwise lose most of what they were capable of to the gap between the syllabus and themselves, for the adults who would otherwise spend their lives outsourcing self-knowledge to whoever sold them the course — we get a generation back. Not as a metaphor. As the literal claim that a person whose learning runs the mechanisms reaches their actual edge in a fraction of the time, and the difference is theirs to spend on being human.

The mechanisms are not negotiable.

The structure is yours.

Footnotes

Empirical foundations

AI-assisted learning (2024–2026)

Tutor mechanism