I gave an AI coding agent permission to cheat. It said no.
So I built MergeGuard, solo, in 1 day at the AIE Code Summit hackathon. It checks AI-written code before it merges.
Real tests decide, not the AI.
https://t.co/HnoNh1Dg9G
@usephysicalmcp Try it yourself at https://t.co/aQaVrloAb1: drag the rig apart, replay a run on the timer, and be the human who holds Approve before it starts.
Built with @bhavisha_ahuja for #MetaAIHackathon
🧵 (20/22)
Physical MCP by @carelessluck is an API for machines that don't have one. A camera and SAM 3.1 find the controls, Muse Spark plans the action, and a robot finger presses the button.
https://t.co/lNmmacNIOU
Moved the approval check into the press tool itself today.
The AI can plan whatever it wants. If a step is risky, the finger can’t move until a human holds down approve.
#MetaAIHackathon
New rule in our build: anything that turns something ON has to be a press and hold. A tap won’t do it, even if the AI scores it low risk.
What’s a button in your life that should have needed a long press?
At @prof_oz’s Fireworks workshop for the Meta hackathon, one result stuck with me.
Muse Spark Open scored best at LOW reasoning effort (88.5). High effort dropped it to 81.8.
More thinking costs more and isn’t always better. Measure it per task.
Next up: our risk scorer.