Joscha Bach: prediction

Transcript context

…To do some of the “prompt engineering” for you. They create these cognitive architectures that do the prompt engineering and you’re just doing the high, high-level meta prompt engineering. There are limitations in a language model alone. I feel that part of my mind works similarly to a language model, which means I can yell into it a prompt, and it’s going to give me a creative response. But I have to do something with those points first. I have to take it as a generative artifact that may or may not be true. It’s usually a confabulation, it’s just an idea. Then I take this idea and modify it. I might build a new prompt that is stepping off this idea and develop it to the next level or put it into something larger, or I might try to prove whether it’s true or make an experiment. This is what the language models right now are not doing yet, but there’s also no technical reason for why they shouldn’t be able to do this. The way to make a language model coherent is probably not to use reinforcement learning until it only gives you one possible answer that is linking to its source data, but it’s using this as a component in the larger system that can also be built by the language model or is enabled by language model structured components or using different technologies. I suspect that language models will be an important stepping stone in developing different types of systems. One thing that is really missing in the form of language models that we have today is real-time world coupling, right? It’s difficult to do perception with a language model and motor control with a language model. Instead, you would need to have different type of thing that is working with it. Also, the language model is a little bit obscuring what its actual functionality is. Some people associate the structure of the neural network of the language model with the nervous system. I think that’s the wrong intuition. The neural networks are unlike nervous system. They are more like 100-step functions that use differentiable linear algebra to approximate correlation between adjacent brain states. It’s basically a function that moves the system from one representational state to the next representational state. So if you try to map this into a metaphor that is closer to our brain, imagine that you would take a language model or a model like DELI that you use… For instance, this image-guided diffusion to approximate and camera image and use the activation state of the neural network to interpret the camera image, which in principle I think will be possible very soon. You do this periodically, and now you look at these patterns, how when this thing interacts with the world periodically look like as in time, and these time slices, they are somewhat equivalent to the activation state of the brain at a given moment. How is the actual brain different? Just the asynchronous craziness?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.