Post by microagi

Most of the conversation about AI safety right now is about language models. That makes sense - they're the systems most people interact with. But the safety problems that matter most over the next decade aren't going to be about text on a screen. They're going to be about physical systems acting in the world, with mass, momentum, and consequences that don't reset when you close a tab. Embodied AI safety is its own research problem. It isn't a port of LLM alignment with a robot bolted on. A model that reliably refuses a harmful instruction in chat can still produce an unsafe trajectory when it controls a body. A policy that scores well on a benchmark can fail in ways the benchmark was never designed to see. The failure modes are different, the verification methods are different, and the field is years behind where it needs to be. Our Munich team is working on this. We think robots should meet people, not the other way around. And yes - regarding safety, don't drink and robot. 🤖🍺

Video Content