See how AI really works by building it. Buffalo, 1958: the Navy has paid for a machine that learns. Write the Python that teaches it, from the first perceptron through backpropagation and deep learning to attention, the idea inside today's chatbots. A programming puzzle game about machine learning.

Sign in to add this item to your wishlist, follow it, or mark it as ignored

This game is not yet available on Steam

Planned Release Date: Q1 2027

Interested?
Add to your wishlist and get notified when it becomes available.
 

About This Game

Every AI you've heard of grew out of a handful of ideas. Here you build each one yourself, in the order history found them.

The Niagara Aeronautical Laboratory has a new machine: a grid of photocells, a bank of dials, and one lamp that lights when it says yes. Nobody tells it the rule. You set its dials until it gets the examples right.

Then the motors arrive, the dials go behind glass, and the job changes. Now you write the trainer, a short Python program that shows the machine its examples and turns its dials until it learns. Every solution goes to the field, where it trains from scratch on examples it has never seen. It has to work every time.

Follow the lab for sixty years, one memo at a time: through the AI winter of 1969 and the breakthrough of 1986, a board that learns to read zip codes, a lander that learns from nothing but a score, memory, deep boards, and in 2017 a board that reads every memo the lab ever wrote and writes the next one.

Features

  • Write real Python. Trainers are plain Python with loops, lists and functions, and no libraries: anything the machine runs, Python runs the same way. New words arrive as each era invents them.

  • Watch it learn. Step through training one example at a time and see which dials turn, and why.

  • Field tests on examples it has never seen. Five runs that train from scratch, the same five every time, so a solution has to actually work.

  • Optimize your way: fewest dials, least computer time, shortest trainer or best accuracy, each with its own histogram.

  • The real ideas behind today's AI: neural networks, backpropagation, convolution, reinforcement learning, memory, deep learning and attention, across seven chapters from 1958 to 2017.

  • Play with what you built. Draw letters and digits for your own boards to read, fly the lander yourself and then hand it over, or start a memo and let your board finish it.

  • A break room with MENACE, the 1961 machine of 304 matchboxes that learns tic-tac-toe from colored beads. Play it, or let it practice a thousand games.

  • Music that learns as you do. One theme, played on each era's instruments, comes closer to right as you pass missions.

  • No machine learning background needed. An operator's manual is built in, and hints are there when you want them.

  • The true history behind every chapter, in archive notes with sources.

System Requirements

    Minimum:
    • OS: Windows 10 64-bit
    • Processor: 2 GHz dual core
    • Memory: 4 GB RAM
    • Graphics: Integrated graphics (Intel HD Graphics 4000 or better)
    • DirectX: Version 11
    • Storage: 400 MB available space
    • Sound Card: Any
    • Additional Notes: 1280x720 display or larger. Mouse and keyboard.
    Recommended:
    • OS: Windows 10 or 11 64-bit
    • Processor: Quad core 2.5 GHz or faster (Intel Core i5 / AMD Ryzen 5)
    • Memory: 8 GB RAM
    • Graphics: Any DirectX 11 graphics card or modern integrated graphics
    • DirectX: Version 11
    • Storage: 400 MB available space
    • Sound Card: Any
    • Additional Notes: 1920x1080 display. A faster processor shortens training runs and field tests.