Test yourself, then get feedback
Retrieving from memory beats rereading. Feedback nearly doubles the effect.
Research note · October 2026
Meetnoo
AI does the work. We want you to understand what was done. noo is a teacher that sits alongside everything you do with AI, and it doesn't hand you the answer: it helps you find it.
An AI that answers for you sets you back. This has been measured.
The strongest studies compare students working with AI against students working without it, then give everyone the same test with no AI. The results converge: left unconstrained, AI raises exercise scores and erodes learning.
Tightly constrained AI tutor, physics1Harvard · Sci. Reports 2025 · n=194
+0.63
Optional coach-mode tutor2Khanmigo, Tennessee · NBER 2026 · 18 middle schools
+0.07
GPT with guardrails, mathematics3Bastani et al. · PNAS 2025 · ~1,000 high school students
≈ 0
Assistant that gives the answers4University of Maryland · IZA, Oct. 2026 · n=2,379
−0.37
Learning a library by delegating the code5Anthropic · controlled trial, Jan. 2026 · n=52
−0.74
In Bastani et al., unguarded GPT scored 17% lower on the final exam, while practice scores rose 48%. The feeling of progress and actual progress come apart. The same tool yields −0.74 or +0.63 standard deviations: the difference isn't the model, it's what the model is asked to do.
The question is no longer whether AI can do the work. It's whether, once the work is done, anyone still understands it.— THE NOO THESIS
Produce, get it wrong, get corrected.
We build noo on the mechanisms with the best-established effects, drawn from meta-analyses6. Each one becomes a rule for how the teacher behaves.
0.73g
0.55–0.64g
Productive failure
Work on it alone before being shown the method. You get the authors' problem first, not their solution.
0.55g
Self-explanation
Putting your reasoning into words exposes its gaps. You explain to the Noo; it looks for the flaw.
≈ 0.54d
Spaced review
Coming back to an idea a few days later, just as it was about to fade.
≈ 0.40d
Generate before you see
Predict a result before it's shown. The bet always comes before the reveal.
—
Fade the support
Plenty of guidance for a novice, less and less as they progress. Otherwise the help ends up doing harm.
Mean effect sizes (Hedges' g or Cohen's d), measured mostly on short material. Whether they carry over to weeks of work on a research topic is our working hypothesis, not an established result.
AI teaches. You think. Never the other way around.
Everything mechanical goes to the machine. Everything that builds understanding stays with you, and that is precisely what the tool protects.
What the AI does
- Reads the paper, the code, the data
- Builds the simulations, figures and examples
- Runs the experiments, measures the results
- Picks the next question at the right level
- Remembers what you understood, and where you got stuck
What you do
- Propose an idea before seeing the authors'
- Bet on a result, and explain why
- Put it in your own words
- Find the flaw in your own reasoning
- Decide where to go next
An optional guardrail is a bypassed guardrail. So noo has no “just give me the answer” mode. A direct explanation exists, but you earn it, after a real attempt.
One screen. Almost no buttons.
No dashboard, no menus, no syllabus, no settings. An open area where the teacher places what you should look at, a Noo that points to it, and a single reply line. Every element we removed is a decision the AI makes for you, so that all that's left to you is the thinking.
show
Put it on screen
For every central idea, the teacher builds a living diagram you set in motion. Or it lays down a highlighted passage, some code, a number.
point
Point at it
The Noo turns toward the exact element (a word, a line, a bar in a chart) and looks at it.
ask
Ask
A question, a choice or a numeric bet, which waits for your answer before going any further.
These three gestures are the only ways the teacher acts on the screen, and the reply line is your only one. No navigation: you move forward by answering.
One idea, one small world to set in motion.
Understanding means being able to vary a cause and predict the effect. So for every idea that matters, the teacher builds a custom screen during the session, where you act on the mechanism. You don't have to ask. Try it: change the number of assistants and services, then switch to a common plug.
Each world is drawn by the AI, rendered, looked at by it and fixed before you see it, in a minute or two. Then it joins a shared library.
One shape, two liquid eyes. A Noo of your own.
From the Greek noûs, mind. The Noo replaces the interface: where an app would show a spinner, a tooltip or a “next” button, it thinks, looks at what matters, waits for your answer. Every posture is one less line of text.
listening
thinking
pointing
reading
getting it
doubting
surprised
asleep
a shape of its owngaze · follows the cursortwo drops that fuse
SEED → GENOME
7 GENES
SCALE 1:1
7 GENES
SCALE 1:1
Two liquid eyes
Two dark drops that merge when it focuses, spin into an ∞ when it works, let a tear run when something is hard: every expression reads at a glance.
A Noo of your own
When you sign up, a seed draws its genome: colour, character, body, eyes, liquid, ink, sometimes a quirk. Some are extremely rare. You don't choose it: it's yours for good.
It says what it's doing
When the teacher reads your document, runs numbers or builds a screen, the Noo says so (“Reading the document”) instead of a spinner.
Six design choices, and what they rule out.
i
Courses, one memory
Each subject gets its own course, which you pick up whenever you like, right where you left it. The memory of what you understood runs through all of them.
ii
No syllabus, no promises
The teacher keeps a map of the terrain behind the scenes. It is never imposed on you as a curriculum.
iii
Genuinely hard subjects
No simplification that hollows out the subject. When a prerequisite is missing, we go down and fill it in, then climb back up.
iv
Quiet proactivity
The Noo only writes to you when it has something worth saying: “you were stuck on this, I found an angle.”
v
A memory of what you understood
Not of what you read. Your phrasings, your mistakes, what finally gave way. Readable, versioned, yours.
vi
Convictions
The teacher tells you when your idea is wrong. A fluent but mistaken paraphrase doesn't get a pass.
The same teacher, alongside any kind of work.
One screen, one memory of you, one teacher that judges when to step in. Only the source changes. And what you understood while reading a paper comes back when you code.
Research
The paper you couldn't crack
Paste a link. The teacher doesn't summarize: it poses the problem the authors faced, then has you rediscover their idea.
How would you have done it?
Code
What your agent just did
While your coding agent works, noo follows its session and helps you understand the change, at the depth you choose.
Before I show you the diff: where would you have put this cache?
Practice
Learning by doing, with a real score
A data science competition, a real dataset. You bet on each experiment before running it; the score settles it.
How much will this move the score? Why?
Concepts
An idea you never quite understood
Diffusion, backpropagation, entropy. Worked through in a simulation until you can explain it to someone else.
Move this slider. What just happened?
Time spent is not a measure. Understanding is.
We judge noo by the one criterion serious studies rely on: what you can do without help, later.
01
The unaided test
Without the paper, without the teacher: can you explain the core of the idea in your own words?
02
Transfer
Answering a question you've never seen, on a neighboring case. That's the difference between understanding and recognizing.
03
Recall at day 7
A week later, unannounced. Understanding that doesn't survive a week wasn't understanding.
04
The teacher's behavior
Simulated students: did it make the student do the work, or give the answer away? Did it cave to a persistent mistake?
96%
Current models give away the answer in their very first message more than 96% of the time when asked to be Socratic7. Teaching behavior doesn't come from an instruction; it is measured and corrected, like code.
You don't share what you consumed. You share what you understood.
Work with AI without handing over your thinking.
noo is in private beta. We're testing it first on our own papers, our own code, our own blind spots.