Know what you actually understood.

GraspN turns any page, PDF, or pasted passage into questions grounded in the source—then quietly builds a map of what you understand, what needs attention, and what to revisit next.

Paste something and try it Paste something this week

No account needed to start

CHECKS · REASONINGCHALLENGE

Why does scaled dot-product attention divide by √d_k before the softmax?

ATo normalise the output so the attention weights sum to one
BTo stop large dot products from pushing softmax into a small-gradient region
CTo reduce the parameter count of the projection matrices

↳ source: passage 12Why this was asked

A real question from a real paragraph. The citation lets you check the question against the source.

Read it. Check it. Keep it.

Add a source, check what you understood, and let GraspN remember what deserves another look.

STAGE 01
Add
Add a page you're reading, a PDF, or pasted text. GraspN keeps the source organised so every question can point back to the exact passage.
The extension stays thin. It reads what's in front of you and hands off; the heavy work happens on the backend.
SOURCE PASSAGE12
We suspect that for large values of d_k, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients.
STAGE 02
Check
Questions focus on the important ideas in the text. Each one explains what it checks, how challenging it is, and why it matters.
An item that cannot explain what it measured does not ship. That is a rule, not an aspiration.
CHECK QUESTIONREASONING
Why does scaled dot-product attention divide by √d_k before the softmax?
STAGE 03
Remember
Your answer updates what GraspN understands about your learning. A repeated wrong answer becomes something useful to revisit, not merely a bad score.
The session can expire. What it taught the graph does not disappear with it.
YOUR UNDERSTANDING3 UPDATED
Scaled dot-product attention
0.72
Softmax gradient saturation
0.41
Positional encoding
0.88

Nothing is asked that the text didn't say.

Every item carries the block it came from. Open the citation and you land on the sentence — which is also how you catch the model being wrong.

BLOCK: A stable, ordered text unit. Numbering survives re-imports, so a citation keeps pointing at the same words.
EVIDENCE: Source support is stored with the concept, not with the question. The question is disposable; the link is not.

attention-is-all-you-need.pdf

10The two most commonly used attention functions are additive attention and dot-product attention.
11Dot-product attention is identical to our algorithm, except for the scaling factor of 1/√d_k.
12We suspect that for large values of d_k, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients.
13To counteract this effect, we scale the dot products by 1/√d_k.

Passage 12 produced the question in the hero. The highlight shows exactly where the answer came from.

A check is temporary. Your understanding keeps growing.

GraspN connects your sources, answers, possible misunderstandings, and areas to revisit—a map that becomes more useful every time you learn something.

NODE: A canonical concept, sized by how much evidence supports it.
EDGE: A typed relationship — requires, causes, part-of — never a guess.
AMBER: A tracked misconception. It will resurface until the evidence says otherwise.

Four ways to be asked. One place it lands.

Difficulty is not the only axis. What a question is trying to find out matters more — and it is recorded, so you can see which kind of thinking you're actually weak at.

DIMENSION	WHAT THE ITEM ASKS FOR	WHAT IT UPDATES
RECALL	Retrieve a stated fact, term, or value from the material.	Mastery on the concept, weighted low. Cheap signal, easy to fake.
REASONING	Explain why something follows — mechanism, cause, consequence.	Mastery and misconception candidates. The highest-value signal.
APPLICATION	Use the idea on a situation that isn't in the source text.	Mastery plus transfer confidence on related concepts.
DEFINITIONS	Draw the boundary — what counts, what doesn't, what it isn't.	Concept boundaries, and the relationships that depend on them.

Where you read, and where you think.

THE EXTENSION

It reads the tab you're already in.

Select a passage or take the whole page. The extension extracts and hands off — it never becomes a second app you have to remember to open, and it never carries the heavy work.

The two most commonly used attention functions are additive attention and dot-product attention. Dot-product attention is identical to our algorithm, except for the scaling factor. We suspect that for large values of d, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients. To counteract this effect, we scale the dot products.

4 concepts found

IN SELECTION · 318 WORDS

Measure what I understood

THE WEB APP

Where your map lives.

Paste or upload a source, work through questions, and see your understanding develop across everything you've read.

NEXT LEARNING ACTIONEXPLAINED

Revisit softmax gradient saturation at higher difficulty.

because: 2 incorrect answers, 6 days apart,
both selecting the normalisation explanation

Your understanding is not our asset.

Learner knowledge is the product, which is exactly why it can't be treated as a disposable cache or held hostage. It is readable, portable, and deletable — on request, not on appeal.

EXPORT: Concepts, evidence, misconceptions, and mastery, as structured data you can actually read.
DELETE: Removes the graph, not just the session. No shadow copy kept for model training.
EXPLAIN: Every item can say what it measured and why it was asked. No unexplainable scores.

Paste something you read this week.

No account, no setup, no card. Find out in about four minutes whether it stuck.

Try it free

Or drop a PDF — up to 40 pages