I think LLMs could provide a human friendly interface for robots. There’s a lot of interesting work happening with embodied AI now, and in my opinion embodiment is the key ingredient for making AI intelligent in a human sense. A robot has to interact with the environment and it builds an internal model of the world for making decisions. This creates a feedback loop where the robot can learn the rules of the world and do meaningful interaction, and that’s precisely what’s missing with LLMs.
Not necessarily just an LLM on its own. The key part is that the internal model is coupled with reinforcement learning where it becomes rooted in the behaviors of the physical world. Real time continuous learning is the way to get there, but it can be done using different approaches. For example, neurosymbolic AI combines deep neural networks with symbolic logic. The LLM is used to parse and classify noisy input data, while a logic engine is used to make decisions about it. My expectation is that we’ll see more of these types of approaches where different machine learning techniques are combined together going forward. LLMs will just be one part of the bigger whole.
I think LLMs could provide a human friendly interface for robots. There’s a lot of interesting work happening with embodied AI now, and in my opinion embodiment is the key ingredient for making AI intelligent in a human sense. A robot has to interact with the environment and it builds an internal model of the world for making decisions. This creates a feedback loop where the robot can learn the rules of the world and do meaningful interaction, and that’s precisely what’s missing with LLMs.
So an LLM with realtime learning/updation?
Not necessarily just an LLM on its own. The key part is that the internal model is coupled with reinforcement learning where it becomes rooted in the behaviors of the physical world. Real time continuous learning is the way to get there, but it can be done using different approaches. For example, neurosymbolic AI combines deep neural networks with symbolic logic. The LLM is used to parse and classify noisy input data, while a logic engine is used to make decisions about it. My expectation is that we’ll see more of these types of approaches where different machine learning techniques are combined together going forward. LLMs will just be one part of the bigger whole.