LLMs That Learn From Experiments And Environments
Gather observational data about a system and use that data to train a model.
If the model accurately predicts the system’s behavior, it can simulate the system. The reverse is true. If the model is an accurate simulation, it can predict the system’s behavior. Now replace simulate with generate, and you understand what the next generation of LLMs can do. Text is intriguing, but think about video generation from prompts.
Ask an LLM to generate a video of an astronaut walking on the moon. The next generation of LLMs will generate the signature hops and exaggerated movements of a person attempting to navigate a low-gravity environment.
In the video, the LLM is simulating low-gravity dynamics because it has learned from enough observational data about it. Those data sources aren’t just videos. It learns from text descriptions as well. We do the same thing. I learned to catch a baseball before I learned the physics and calculus necessary to plot its trajectory.
Do you see where I’m headed yet? We observe enough examples and develop a heuristic that predicts where the ball will go. I can also identify balls that don’t follow the laws of physics in the real world or videos. Show a child 20 videos of a ball flying through the air, and they’ll be able to identify any that don’t look right.