Stanford CS25: V3 I Generalist Agents in Open-Ended Worlds

hi everyone thanks for coming to our cs35 lecture today uh so today we're honor to have Jim B from Nvidia we'll be talking about generous agents in open ended worlds and he's a senior AI research scientist at Nvidia where his mission is to build generally capable AI agents with applications to gaming Robotics and software Automation and it

research spans from condition models multimod AI reinforcement learning and open-ended um learning he obtained his PhD degree in computer science from here Stanford advised by Professor p and previously he did research internships at openi Google AI as well as um mil Quebec AI Institute uh so yeah we're um um give it up for

Jim yeah thanks for having me so I want to start with a story of two kittens um it's a story that gave me a lot of inspiration over the career over my career so I want to share this one first you know back in 1963 there were two scientists from MIT helden Hine they did this ingenious

experiment where they put two newborn kittens in this device and the kittens have not seen the visual world yet so it's kind of like a Maro round where the two kittens are linked by a rigid mechanical bar so their movements are like exactly mered and there's an active kitten on the right hand side and

that's the only one able to move freely and then transmit the motion over over this link to the passive kitten uh which is confined to the basket and cannot really control its own movements and then after a couple of days helden high kind of take the kittens out of this marry Go Round And

then did visual testing on them and they found that only the active kitten was able to develop a healthy visual motor Loop like you know responding correctly to approaching objects or like visual Cliffs but a passive kitten did not have a healthy visual system so I find this experiment fascinating because it shows the