oneshotlm
← What's new
New Models

Mercury 3 added

Mercury 3 made game and UI state readable as text, then used that feedback to revise its builds.

Mercury 3 worked from text-only browser feedback in this evaluation: DOM text and console output. Its Brick Breaker transcript shows a practical response to that constraint. It added a temporary readout of the ball’s position and velocity, score, lives and remaining bricks, then read those values during clicks and drags. After revising launch and restart handling, it removed the debug display and reloaded.

Drum Machine used a similar approach: the model exposed button coordinates and sequencer state as text, exercised the controls, and revised input handling. Reaction–Diffusion also gained numerical readouts of its simulation state. These transcripts are useful for following how the model made otherwise hidden behavior observable; they do not establish visual or audio quality.

Across 35 saved sessions, 30 included a recorded file revision followed by another browser check. Thirty-four ended normally. Double Pendulum reached the 15-minute limit without a recorded browser-tool call, so its saved output should be read as an unfinished session.

Top builds in this update

mercury-3 — Brick breaker OUTPUT
mercury-3 Brick breaker
Generated pi 0.87.1 · browser iteration · 108.8s · $0.0269 · 29k out
Full run + transcript →
mercury-3 — Drum machine OUTPUT
mercury-3 Drum machine
Generated pi 0.87.1 · browser iteration · 129.3s · $0.0321 · 29k out
Full run + transcript →
mercury-3 — Reaction-diffusion OUTPUT
mercury-3 Reaction-diffusion
Generated pi 0.87.1 · browser iteration · 137.6s · $0.0362 · 35k out
Full run + transcript →