We recently released Drone-Bench, measuring how effective AI is at navigating drones and identifying targets.
Many people had strong reactions to our announcement. We think this is good; we wanted to start the discussion of where we do/don’t want AI in society. That frontier AIs are now capable of navigating drones, and are improving rapidly at this capability, is concerning.
After it went viral on twitter, the most common sentiment was: “Why would you build the scary thing which we were warned not to build?”
First, our benchmark does not help AI navigate drones. It is a measure of how good AI models already are at doing so. Whether AI can do it or not, does not depend on the existence of our benchmark.
Once AI can competently navigate drones, we believe it is important for the general public to know, not just leaders of tech companies. Our benchmark shows clear improvement between every model release; it is only a matter of time before frontier AI has superhuman drone navigation capabilities. Just six months ago, the frontier models struggled and crashed; Astra now beats our human baseline on every task in the benchmark.
Our hope for the benchmark is that it will help convince people to take this seriously as soon as possible, before we get to the point of superhuman AI drones. How these capabilities should be allowed to be used, if at all, is for the general public and lawmakers to decide, not just the AI labs. Accurate public information of current model capabilities is essential to allow people outside the labs to make informed decisions.
Finally, we are not letting AI companies hill-climb this benchmark. No lab has access to it; we run all evaluations ourselves. Its only purpose is to tell the world how fast AI autonomy in the physical world is advancing.