Experiment 001
We taught our AI to play Super Mario Kart (and learned a few interesting things).
What a 16-bit racing game can teach us about agents, decision-making, and unexpected behavior.
RACER 01 MARIO POS 04 LAP 02/03 FRAME 04218 DELTA +00.74s SPEED 71.8 ITEM NONE ACTION steer_left THROTTLE .82 BRAKE .00 TARGET node_118 CHECKPOINT 05/12 TRACE apex_err=.14 wall_clear=6.2 lookahead=18 PLANNER route=inside recovery=false confidence=.81 INPUT left=.42 accel=1.00
Super Mario Kart · 1992 · SNES
A small game. A useful question.
We taught our AI to play Super Mario Kart. A 16-bit racing game gave us a place to explore agents, decision-making, and unexpected behavior.
Three things that stayed with us.
One frame isn't the whole story.
A single snapshot tells you where the kart is. A short history helps explain where it's going. Useful decisions need context, not just a clear picture.
The next move isn't always the best move.
A hard turn can solve the corner in front of you and create a problem at the next one. The interesting question is what an action sets in motion—not just whether it works right now.
Recovery is part of the job.
A clean lap is satisfying. Getting back on track after things go sideways is useful. We care about agents that can recognize a mistake, adjust, and keep the work moving.
Beyond the finish line.
The same questions follow us out of the game: does an agent have enough context, can it see the consequences of its next move, and what happens when it needs to recover?
Serious questions. Room to play.
Our place for experiments, field notes, and the occasional detour—all in service of making AI do more useful work.
Explore The Lab
We'd like to be the company that handles your AI.
Tell us what's taking too much time, getting lost between systems, or any cool ideas you have that would make your business run better.
Let's talk about your AI