Real Time Learning: The School That Upgrades Every Time the Model Does

Bloom's most famous result sounds like a misprint. Tutor one child at a time, with corrective feedback, and the average child scores about two standard deviations above a child in an ordinary classroom. Put plainly, the tutored child beats about 98 per cent of the conventional class. Benjamin Bloom published it in 1984 as a problem rather than a victory: if one-to-one tutoring is that strong, how do we build group instruction that comes close? Bloom's "The 2 Sigma Problem" has been unsettling schools ever since.

I write this from Nairobi, where I build AI systems, and I start with an admission. I have been trying to answer Bloom's question with a machine, and I have not answered it. What I have is a theory with a name, Real Time Learning, and a story about a girl and a project. The theory is the founder's conviction and an untested hypothesis. Keep that label in view as you read.

A Box Arrives in a Junior School

Imagine a junior school in a county far enough from Nairobi that the matatu fare is a negotiation. A box arrives, with laptops and interactive smart boards inside. The box is imagined; the programme behind it is not. The Ministry of Education reported in July 2026 that 10,382 laptops and an equal number of smart boards were going to Junior Schools across 47 counties under the Kenya Digital Economy Acceleration Project (Capital FM via AllAfrica). This article is about what a child does with it, and what the school does with the child.

Let us call the child Wanjiru. She is a composite, a character built to stand for many children I have not met, and every scene with her is imagined. Her project is a question she cares about: why the price of petrol at the trading centre seems to jump on the days the radio says it will fall. She keeps a log of prices in a school exercise book. Then she gets a laptop, and an agent that lives on it, and the exercise book begins to talk back.

Five Definitions, Stated Plainly

A theory that cannot define its terms is a mood. So here are mine. These are the founder's definitions, not the terms of any existing literature.

Real Time Learning is a way of learning in which a student improves continuously by doing real projects alongside a persistent agent. Each better model upgrades the feedback she receives, without waiting for a new curriculum or exam. "Real time" refers to the feedback loop, not to speed.

A portfolio is the set of real projects a learner owns. Each has a stated goal, a dated record of decisions, and tests that the learner wrote or agreed to. It is not a folder of finished essays. It is a collection of things that can be wrong in public.

The persistent personal agent, which I also call the harness, is the software that holds the learner's history and surrounds the model with memory, tests, rules and logs. It tutors, it tests and it extends. Because it persists, it knows what the learner tried last year.

Re-audit on model update means running a learner's old projects, and their tests, against a newer model to find gaps the older model could not see. A project that passed in January may fail in October. Nothing in the student changed.

Finally, the student is the director and the agent is a tireless junior. The director decides what is built and signs the result. The junior drafts, checks and never tires, but it does not decide what counts as understanding. That rule matters most.

The Mechanism: Old Projects Re-Read by Better Engines

Here is the mechanism as I believe it works. It is a hypothesis, and I will say where it could break.

A school built on courses and exams asks one question each term: did the student absorb the syllabus by the date the examiner set? A school built on Real Time Learning asks a different question, continuously: what can this learner now do, in a project she chose, that she could not do before, and what is the next gap?

The persistent agent keeps Wanjiru's price log, her decisions about what counts as a price, and the mistakes she made when she averaged across market days. It tutors by asking, not telling. It tests by generating cases her tests missed. It extends by proposing the next question, and the director decides whether to take it.

Now the engine changes, and each upgrade triggers a re-audit of the learner's old work. Here are three imagined versions of what that might reveal.

Imagine that the first upgrade notices a units problem. Her log mixes shillings per litre with shillings per kilogram of maize, a mismatch she never saw because the old model did not look at the units column. The re-audit flags it. She fixes it, and an earlier conclusion is revised.

Imagine that the second upgrade is better at statistics. It reads her claim that petrol rises on radio days and asks the question she had not asked: what do the other days look like? Her project passed every test she had written. It now fails the audit, because the claim has no comparison group. Her school report never showed that gap, since her marks measured the arithmetic, not the claim.

Imagine that the third upgrade is good enough to read her log in Kiswahili and ask her to explain the finding to her grandmother, who sells paraffin at a roadside stall. She can compute the average. She cannot yet say what it means. The re-audit reveals a gap in explanation rather than in arithmetic, and the next project is built around it.

Notice what did not change. The project stayed hers, and so did the decisions. Only the engine's ability to see the gap improved. That is the core of the theory.

Why is this different from today's AI tutoring? In the tutors I have seen, the session is the unit of memory, and the child is forgotten when it closes. That is my observation, not a measured finding. My judgment is that a tutor living inside a project the child owns, with memory that persists across years and model releases, changes the incentive. The child is not asked to produce answers. She is asked to keep a project honest.

The child is not asked to produce answers. She is asked to keep a project honest.

What the Evidence Can and Cannot Carry

No study tests Real Time Learning itself. What exists is evidence about nearby questions, and it points in several directions.

The ceiling is Bloom's. The one-to-one advantage is large in the 1984 studies, which were small. Bloom framed the result as a target rather than a method. The question is whether a persistent, project-bound agent can carry some of that advantage into a classroom of forty.

The nearest positive result is a study of AI tutoring in Nigeria, reported in a World Bank working paper titled "From Chalkboards to Chatbots". Fix The News reported in May 2025 that the Nigerian students' gains were equivalent to one and a half to two years of ordinary schooling, and that female students and students with stronger starting results benefited most (Fix The News). Judgment: that figure is a cost-effectiveness comparison, not two years of measured learning, and I would not repeat it as a headline without the working paper in front of the reader.

The Khanmigo study taught me the most about access. Oreopoulos and Low ran a trial across 18 middle schools in Hamilton County, Tennessee, in which grade levels were randomly assigned to Khanmigo, set to coach rather than give answers, during a daily remedial maths block (NBER working paper). The authors report that assignment raised maths achievement by about 1.3 national percentile ranks per term. The surprise was in engagement. Ninety-six per cent of students tried the tutor at least once. The median student messaged it on only a third of the days they practised, and only 17 per cent of practice sessions with a mistake produced a message. The authors' summary is that "the binding constraint appears to be engagement." Chalkbeat reports that the researchers thought the benefit came from Khan Academy overall, not from the AI component (Chalkbeat). The study is a working paper and has not been peer reviewed.

Then the warnings. Kosmyna and colleagues at the MIT Media Lab, Wellesley College and MassArt published a preprint, "Your Brain on ChatGPT". It compared essay writing with an LLM, a search engine, or the brain alone, across 54 participants in sessions one to three, 18 per group (arXiv:2506.08872). They report that LLM users displayed the weakest brain connectivity in their EEG measures, and that over four months they consistently underperformed at neural, linguistic and behavioural levels. The sample is small and the task is a single essay. Judgment: a study this size cannot settle whether any effect lasts, and I do not read it as proof of lasting harm.

The warning I take most seriously is that generative AI used without guardrails can harm learning. That is the reason I keep a human check inside the theory. Judgment: this is a working assumption the test proposed below should check, not a finding I can cite here.

What does this evidence say about Real Time Learning? In my judgment, it supports one thing with some confidence: access is not use. It also suggests a method, which is to measure unassisted performance rather than the chat log. It does not show that a persistent project agent, re-run across model updates, produces lasting learning. That claim is what schools must test.

The Dangers Are Not Hypothetical

The first is dependence. If the agent does the thinking, the child becomes excellent at getting help. The MIT preprint found that LLM users underperformed on neural, linguistic and behavioural measures over four months, in a small study of one essay task. The remedy is structural: an unassisted check, a test the agent cannot touch, marked by a teacher on paper. Judgment: a school that cannot run that check should not run Real Time Learning at all.

The second is inequality. The laptop box is real, and so is the connectivity gap. UNICEF-supported work reported 363 of a targeted 1,000 public primary schools connected to reliable internet by December 2025, reaching over 277,000 learners (Capital FM via AllAfrica). Judgment: an agent that lives in the cloud is of little use in a school without a signal. The theory must therefore be offline-first, or it is a theory for the connected children of Nairobi.

The third is the child's data, and it is the sharpest. A persistent agent that re-audits a child's projects keeps a history: names, marks, photographs, mistakes. Judgment: that history should not be created without a parent's or guardian's written consent, and the law on children's data in Kenya should be checked against this design before any pilot starts.

One more danger belongs to the examiner. A portfolio is only evidence if someone can tell whose hand made it. A Kenyatta University lecturer, quoted in the Capital FM report via AllAfrica, put it more sharply: "The biggest mistake would be to assume that because an answer is well written, it is necessarily correct." That sentence belongs on the wall of every staff room that adopts this theory.

Next Monday, in a Kenyan School

Suppose a headteacher reads this on a Monday and wants to act before lunch. None of the following needs a new law or a new laptop.

Choose one project per learner. The learner chooses it, the teacher approves it, and it goes in a log with a date and a question the learner can answer in her own words.

Run an unassisted check every fortnight. No device, no agent, and a teacher who marks it against the same rubric each time. In my judgment, this is the most important element of the whole theory, because it is where dependence shows up.

Write a data policy before the first login. Get parental or guardian consent in writing. Collect the minimum. State plainly who holds the history, where it lives, and how it is deleted. Judgment: treat a written impact assessment as the default for any system that profiles children. If the answer to "where is her history kept?" is "on a server I cannot name", the pilot does not start.

Then test it properly. Take two classes in one school for a term. One runs Real Time Learning; the other keeps the course-and-exam method, with the same teacher and syllabus. Before the term starts, pre-register three measures: the unassisted test, a transfer task in an unfamiliar context, and an engagement log counting the gaps the learner named herself. Re-run the re-audit on old projects under two model versions, and record how many gaps appear and whether closing them shows on the unassisted test. Judgment: this design can falsify the theory, and that is why I want it run.

I do not know whether Real Time Learning works. I believe it is worth testing, because the alternative is a school that asks a child to recite what an examiner wrote years ago, while the engines on her desk keep changing and nobody re-reads her work.

Imagine the first unassisted test of the term. Wanjiru has been working on her petrol project for weeks, and the audit has run twice. The agent is silent, because the test is hers. She sets her exercise book aside, reads the question, and types her own answer. Then she presses run. The test fails on a comparison group she never built. She reads the red line twice. For the first time, she can see the gap.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing (parallel Claude Haiku 5.5 agents) and checked against the linked sources on 30 October 2026. Convictions, proposals and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Bright Kwame Ayisi (Kwameghana), Wikimedia Commons user; file credited as own work. CC0 1.0. Original source. Illustrative image: students gathering around a laptop in a school room. Photograph: Bright Kwame Ayisi (Kwameghana), Wikimedia Commons, CC0 1.0 (public domain dedication). This image does not document the events described.