Start from wrong.Leave understanding.
Most AI tools stop at a three-second answer that feels like mastery. AskWrong turns your question into a course at your level, with a flaw to catch and reviews until it sticks.
Built on how people actually learn
people actually learn
One question in
A whole livingcourse out.
Start from a question.
Name what you want to understand. The coach scopes a roadmap, then writes the course at your level, chapter by chapter.
learn / gradient-descent / roadmapRoadmap: 4 chaptersTopic titleGradient descentCreate topic- 01The gradient
- 02The update rulewriting
- 03Choosing the step size
- 04Why it diverges
usage clip coming soonOr bring your own sources.
On Pro and Max, a PDF, pasted notes or a photo becomes the course: chapters, drills and a schedule.
learn / gradient-descent / documentsLibrary: 2 sourcesDrop a PDF, paste text, or add a photo- PDFConvex optimization, chapter 9
- IMGWhiteboard, learning-rate sweep
usage clip coming soonAsk in one box.
A real answer to your exact question. Then take it to practice at your level, or straight to the theory.
learn / gradient-descent / askYour questionMy loss goes to NaN after 5 steps. Is my gradient wrong?
askwrongThat pattern, drop then growing overshoot, is the signature of too large a step, not a wrong gradient.
Practice thisNow get the theoryusage clip coming soonOne course, three depths.
Depth tabs on one course: the hiker's intuition, the update rule, the proof it is optimal.
learn / gradient-descent / theoryDepthBeginnerPractitionerPhDChapter 01
The blindfolded hiker
Feel the slope underfoot, step downhill, repeat.
From “downhill” to a vector
θ ← θ − η∇L
Why −∇L is the steepest way down
Of all unit steps d, ∇Lᵀd is lowest at d = −∇L / ‖∇L‖.
usage clip coming soonDrills from your own bug.
A fresh multiple-choice drill built from your bug. Pick, then get the why.
learn / gradient-descent / practiceExercise 1: multiple choiceYour loss explodes to NaN within a few steps. Most likely cause?
- learning rate too small
- learning rate too large
- too few features
- batch size of 1
Why: each step overshoots, so the error grows until it diverges.
usage clip coming soonCatch the planted flaw.
A confident 'fix' with two real bugs: the flipped sign and the runaway learning rate.
learn / gradient-descent / interrogate2 of 2 caughtdescend.pyclick what is wrong1def descend(grad, x, lr=0.1):2 for _ in range(100):3 x = x + lr * grad(x)4 return x56best = descend(grad, x0, lr=10)- 3sign flipped, this climbs
- 6lr far above the 2/β stability bound
usage clip coming soonExplain it back, get graded.
Explain it in plain English. The rubric ticks direction and step size, flags the missing stability bound.
learn / gradient-descent / feynmanYour explanationGradient descent walks downhill on the loss. The gradient points uphill, so you step the other way. The learning rate is how big the step is.
64not accepted yet16 short of the pass line- ✓Step against the gradient
- ✓The learning rate sets the step
- ✕Curvature bounds it, η < 2/L
usage clip coming soonQuestioned until you pass.
One graded question at a time on chapters you pick, pressing where you are thinnest. A pass ends it early.
learn / gradient-descent / grillGrill me: learning rate2/5 turnsWhat sets the largest learning rate that still converges?
partial1 of 2 key points
Curvature L = 4: does η = 0.6 converge?
No. Each step scales the error by 1 − ηL = −1.4. It needs η < 2/L.
pass2 of 2 key points
usage clip coming soon
Six phases. Fully connected.
Any phase,
any time.
No fixed staircase. The coach sends you wherever the gap is: back to theory, into practice, out to a Feynman check.
The graph
Hover or tap a phase.
Memory model, try it
Play the forgetting curve.
Each review flattens the decay. The coach brings a concept back before your recall drops below 90%.
Hover for memory level, click to review
Desirable difficulties
Friction that teaches
Having read an explanation is not being able to rebuild it. AskWrong keeps the effort that makes knowledge stick and takes the logistics off your hands:
- Course
- Clear, at your level
- Source
- Generated, or your PDF, notes or photo
- Schedule
- What to review, and when
- Tuned for
- Retention, or your exam date
Learning from error
To spot a flaw you must hold the right model and compare. One surface, Interrogate the flaw, plants real mistakes for you to catch.
The Feynman technique
Re-explaining a concept in your own words is laborious, and that labour is what anchors it. The coach checks your explanation against the key points and schedules what you missed.
Retrieval, just before you forget
Recalling is harder than rereading, and that effort is what makes a memory last. The schedule brings a concept back as it fades; you produce it, not recognise it.
Start free.Then scale your learning.
Plans
- The full learning loopAsk, Theory, Practice, Interrogate the flaw, Feynman, Grill
- Theory at beginner and practitioner depth
- 80 credits per monthUp to 10 a day on the house models: we cover the inference. One credit is one question on the fast model; stronger models cost more
- Reviews just before you forgetSpaced repetition scheduled for you, synced across devices
- 1 new course per monthand 1 minute of voice dictation in Feynman and new courses
- 10 saved sessionsEvery item kept 30 days after you last use it
- Everything in Free, plus:
- Stronger modelsGPT, DeepSeek, Qwen, GLM and MiniMax, each priced on screen before you send
- Deeper reasoningThe reasoning dial past medium, as far as the model goes
- PhD depthFormal definitions and theorems in LaTeX, cited sources, a proof on each statement
- Briefed exam papersA full paper written from your brief: style, duration, number of exercises
- Code exercisesWrite Python in the exercise card and run it against tests in your browser, with hints and a code review.
- Several sources with citationsGround a course in your own documents. Each chapter cites the passage it drew on, and a citation opens it beside the course.
- The quick checkA few questions on what a course takes as known. It ends in suggestions, and nothing changes until you press.
- Document importYour PDFs, photos and pasted text: 200 OCR pages per month, PDFs up to 20 MB and 100 pages
- 7,000 credits per monthUp to 700 a day
- 10 new courses per monthand 120 minutes of voice dictation in Feynman and new courses
- Unlimited history1,000 saved sessions, notes and highlights synced across devices
- Everything in Pro, plus:
- The frontier modelsOpen only on Max: Kimi, the largest Qwen and DeepSeek Pro
- 30,000 credits per monthUp to 3,000 a day, 4.3x what Pro gets
- 50 new courses per monthand 360 minutes of voice dictation in Feynman and new courses
- 100,000 saved sessions
- Bigger imports800 OCR pages per month, PDFs up to 50 MB and 400 pages
Questions, answered properly
Will it just give me the answer?
Yes, with the why attached, a flaw to catch and a question back to prove you got it. This is for the answers you want to keep.
How deep does the theory go?
As deep as you ask. Each topic is one living course at beginner, practitioner or PhD depth. PhD develops the subject formally: definitions, properties and theorems in LaTeX, with cited sources and a proof button on each statement. Beginner and practitioner are on every plan; PhD comes with Pro and Max. A proof click is an answer and spends credits like one.
Why is it called “AskWrong”?
Because being wrong is where learning starts. One step hands you a deliberately flawed answer to interrogate: spotting what is wrong locks in what is right.
Do I need an account?
Yes. You create a free account, so your course and your schedule follow you from laptop to phone. An email and a password are all it takes.
Do you train on my questions?
No. We do not train on your questions. If you pick a free model, the company that runs it may train on what you send; the picker shows each host's terms before you send. No telemetry, no tracker in your browser, and we never sell your data. Delete your account and we erase it; copies a model host received stay under its terms, and our encrypted backups age out within 35 days.
How do the limits work?
Every account gets a monthly balance of credits, with a daily cap so one heavy day cannot spend the whole month. One credit is one ask on the cheapest fast model; heavier models and deeper thinking cost more, priced on screen before you send. Your plan also sets how many new courses you can start and how many minutes of voice dictation you get each month. Which models you can pick is set by your plan, not by your balance: stronger ones stay locked until you upgrade, and credits never unlock them. Hit a limit and nothing is lost: what you wrote, said or already got back stays saved. The control locks with the reason, so you are not left retrying, and you see exactly what you hit and when it refills: your balance and the date. Upgrading lifts it: Pro and Max raise the balance, open the stronger models and add PhD depth and code exercises, with more courses, voice dictation and document import.
Not answered here? Ask the person who builds it.
Write to me