The Loop  ·  Issue 024

The Loop

A field journal of the AI frontier — for engineers who ship.

  Lab bench

Experiment №013
filed May 4, 2026

game

Filed under

  • #game
  • #reflex
  • #physics

Ricochet

One bullet, bouncing forever. Click to bend it up to 60° toward your cursor. Hit targets, dodge growing hazards.

  Primer

Skip if you already know the theory; the interactive is right below.

One bullet. It enters from the left and never stops — walls bounce it forever. Click anywhere on the field and the bullet bends up to 60° toward your cursor: toward a target you want to hit, or away from a hazard about to swallow it.

You don’t move. The bullet does. You only steer it.

▶  Try it

Targets

0

Survived

0.0 s

Best

0

runs · 0

One bullet, ricocheting off every wall. Hover to aim, then click the field (or press space) to bend the bullet up to 60° toward your cursor. Hit blue targets, avoid red zones — they grow on spawn and end the run on contact. Hazards begin appearing after about five seconds.

saved locally · sign in to sync

▤  Leaderboard · top 25

Loading…

  Notes from the bench

What to watch for, why it matters, and the one thing that usually surprises people.

The twist

Most shooting games have you firing many bullets at fixed targets. Here the bullet is a persistent thing — a single object careening around a closed room — and your job is to babysit it. Every click is an investment: you spend a small bend now to put the bullet somewhere useful in two seconds.

The interesting decision is when notto click. Every nudge perturbs a trajectory you’ve already half-planned in your head. Resist the urge to over-correct. The bullet was already going to ricochet into that target on its next bounce — let it.

The math

The bullet moves at constant speed and reflects perfectly off each wall. On click, I take the angle from bullet to cursor, compare it to the current heading, and rotate by min(60°, |Δθ|)in the direction of the cursor. So a single click carves at most a 60° arc; reversing the bullet end-to-end takes three clicks at minimum, gated by a small cooldown so you can’t mash your way out of a bad situation.

Targets fade after a few seconds (less time as the run progresses). Hazards spawn at a point and grow outward to a capped radius, then sit there for a while before shrinking out. Hitting a fully-formed hazard ends the run; clipping one in its first half-second doesn’t, which gives you a beat to bend away from a bad spawn.

What I’m unsure about

Whether the bend size is right. 60° feels powerful enough to rescue a doomed run but small enough that you can’t just click frantically and live. Playtesting this honestly — well, I’m the only playtester so far, and I designed the bend, so my proprioception is biased toward thinking my own touch is correct.

In a line

Canvas physics game. A persistent bullet ricochets off every wall at constant speed; clicks rotate its heading toward the cursor with a hard 60° cap and a small cooldown. Targets fade, hazards grow, difficulty ramps with elapsed time.

Other experiments

15
  1. Exp 001

    How a sentence becomes tokens

  2. Exp 002

    Temperature and top-p, visibly

  3. Exp 003

    What does this prompt actually cost?

  4. Exp 004

    Tokens per second

  5. Exp 005

    How far should the model think?

  6. Exp 006

    Neural language vs a Markov chain

  7. Exp 007

    What each token looks at

  8. Exp 008

    Words in space

  9. Exp 009

    The injection arena

  10. Exp 010

    AI or human?

  11. Exp 011

    Context Tetris

  12. Exp 012

    Magnet flip

  13. Exp 014

    Wick

  14. Exp 015

    Echo

  15. Exp 016

    Orbit