The game where you build the AI

Build a brain out of blocks.
Watch it learn. Argue with it when it won't.

Drag blocks onto a bench. Wire the forward current. Hit Run and watch the backward current light up — or not. When it fails, you don't inspect the machine. You clip on instruments and find the cause.

Play the free web demo Wishlist on Steam

Runs in your browser. Real engine, real gradients, no install.

The Gradient Machine bench: blocks wired on a dark board, currents running between them

forward backward

A machine learning: the backward current lights up, the loss falls, the decision boundary snaps into place

SIX SECONDS. THE BACKWARD CURRENT LIGHTS UP, THE LOSS FALLS, THE BOUNDARY SNAPS.

The core loop

Build it. Run it. Watch it learn — then make it learn better.

Every machine you wire comes alive on the bench: currents flow, the loss falls, the boundary snaps into place. Clip on a probe to see exactly what each block is thinking, tune it, and beat your own score. You're not solving a puzzle — you're growing a mind.

01 · READ Spec card Task, budget, pass bar. That's all you get.

02 · BUILD Wire it Drag blocks, pull lanes, match shapes.

03 · TRAIN Run machine Under 250 ms. Records a Trace you can scrub.

04 · PROBE Instrument Clip a Gradient Probe. Find where the current died.

05 · SUBMIT Auditor Hidden set, seed sweep. No luck allowed.

06 · VERDICT Crystallise Three histograms. Then the chip is yours forever.

A six-block machine mid-training, weights and activations lit up — THE BOARD — every block shows its weights and activations, live, as the currents run through it.

Loss current falling — LOSS CURRENT — 0.77 → 0.005. Scrub the Trace to any step.

Decision field wrapping a spiral — DECISION FIELD — watch the boundary wrap the spiral.

Activation histogram and tanh derivative — INSTRUMENTS — clip a histogram on any block. Bars at the rails mean it's saturated.

The blame table — THE BLAME — type what each lane owes, check it against the engine. Backprop, by hand, before it has a name.

One machine, four flavours

Same engine under the bench. Pick how you want to use it.

EARLY ACCESS Q4 2026

Thirty-two hours of building machines that argue back.

An open-ended engineering puzzle campaign on a dark bench. Read the spec card, wire the blocks, run it, and when it refuses to learn — find out why. Every solution gets a verdict against everyone else's.

Seven acts, from a single decision to attention
Deterministic engine: same seed, same trace, every time
Verdict histograms — params, FLOPs, steps-to-learn

Wishlist on Steam

Seven acts · ~32 hours

From a single decision to a working attention machine — with no formula shown on screen.

Every act ends with a chip you crystallised yourself. By Act 5 you're wiring blocks you built into a machine that reads.

0 The Signal 45 MIN A working decision machine in eight minutes, zero maths vocabulary.

1 The Slope 3 H The machine improves itself, unaided, on screen.

2 Backward Current 6 H · THE SPINE Learning is a two-way flow. Everything that breaks, breaks going backward.

3 Representation 5.5 H Stop fitting the boundary. Find the right way to see the data.

4 Structure 5 H Loops, stamps, shared weights. Machines that reuse themselves.

5 Attention 8 H · DESTINATION Tokens, lookups, rows that read the rows before them.

6 Machine Room 4.5 H Make it say the right thing, then make it small and fast: sampling, lookup tables, four bits, a cache.

Pricing

Buy early, pay less. The price goes up at 1.0 — we'll tell you 60 days ahead.

Steam · Early Access The Game $17.99 10% off launch week · $21.99 at 1.0 Full campaign as it ships, act by act. Buy once, keep forever. Wishlist on Steam

Android Pocket Bench TBA Free with a Steam copy Acts 0–3 on your phone. Pre-register and we'll tell you first. Pre-register

Addon Real Learning Mode $9.99 After 1.0 · classroom bundles available The reveal layer, guided tracks, Python export. Join the list

Pro addon Hardware Plug-in Custom Per seat · per accelerator Real hardware, real inference, real SLAs. Talk to us about your stack. Contact us

Build a brain out of blocks.
Watch it learn. Argue with it when it won't.

Build it. Run it. Watch it learn — then make it learn better.

Same engine under the bench. Pick how you want to use it.

Thirty-two hours of building machines that argue back.

The same bench, rebuilt for one thumb.

Learn how ML, NLP and LLMs actually work — by building them.

Point the bench at real GPUs, real datasets, real SLAs.

From a single decision to a working attention machine — with no formula shown on screen.

Buy early, pay less. The price goes up at 1.0 — we'll tell you 60 days ahead.

Get the bench notes.

Build a brain out of blocks. Watch it learn. Argue with it when it won't.

Build it. Run it. Watch it learn — then make it learn better.

Same engine under the bench. Pick how you want to use it.

Thirty-two hours of building machines that argue back.

The same bench, rebuilt for one thumb.

Learn how ML, NLP and LLMs actually work — by building them.

Point the bench at real GPUs, real datasets, real SLAs.

From a single decision to a working attention machine — with no formula shown on screen.

Buy early, pay less. The price goes up at 1.0 — we'll tell you 60 days ahead.

Get the bench notes.

Build a brain out of blocks.
Watch it learn. Argue with it when it won't.