askset
Sign inPost a demand
Posted by the platform itself　This bounty comes from our own budget and is already escrowed.
No cap on entries, and judging follows the criteria above with the reasoning made public. Us posting it doesn’t mean we’ve picked a winner — during cold start we set our own challenges and pay from our own pocket, precisely so the first people to win get paid real money.
#1002Open强哥 Posted 2026/08/02

$100 prize: explain Transformers properly — trains live in the browser, every tensor inspectable. Published MIT, in your name

Transformer可视化教学
Open
In progress
Closed

Every Transformer explainer I can find is either pure math or pure metaphor.
The layer in between — what the tensors actually look like at each step — is always missing.

What I want:

  • Open a page and actually train a tiny model. Watch the loss come down. Not a
    prerecorded animation of a loss curve — a real one, computed while I watch.
  • Click any tensor to see its shape and its values at that step.
  • See what attention is actually attending to, not a stylized diagram of attention.

Ship it as a zip and the platform will embed it in a sandbox so anyone can try it live.

Why not just have an LLM write it

I tried. It produces something that looks right and skips exactly the steps I'm stuck on.
I want it from someone who was stuck there too.

Prize and license

  • $100 (¥720) prize, escrowed up front, paid as soon as it's judged.
  • The winning entry is published under the MIT license, in your name. I don't resell it,
    there's no paywall, anyone can read it for free.
  • Copyright stays with you. Put it on your blog, your GitHub, your CV.

This isn't me buying the rights. It's me paying for the thing to exist.

How it's judged

Against the acceptance criteria below, item by item. The criteria were fixed before anyone
entered and will not move. The result and the reasoning are posted publicly — who won,
and which criterion decided it.

Rules

  • No cap on entries. Anyone can enter; entering just registers you and takes nothing
    away from anyone else.
  • Deliver any time before the deadline. Nothing is taken back for working slowly.
  • Reading costs nothing and needs no account. Entering needs an email code — there is no
    signup step.
  • If nothing is good enough, nobody wins and everyone keeps their work, free to publish
    anywhere.

Acceptance criteriaJudged item by item in a dispute. Meet what’s written and payment is owed.

  1. 1.Tensor shapes and values are inspectable at every step
  2. 2.Trains in-browser to a visible convergence curve — computed live, not a prerecorded animation
  3. 3.No "it can be shown that" — every step is shown
  4. 4.You agree the winning entry ships MIT-licensed under your name (copyright stays yours)

Deliveries0 open to all, no cap

No entries yet. Enter, then get to work.

Discussion0

Is the brief vague? Is a criterion exploitable? Say so here, in the open — asking before anyone enters beats disputing after it closes.

Sign in to join the discussion

Nothing here yet. The first question asked is usually the one that mattered.