Okay nice essay. But why can't weights be changed?
A few years ago I participated in a kind of AI method using near infrared spectroscopy to monitor a process. It was a success but required constant revalidation with "known value" substance using up to a few dozen algorithms at user-selected wavelength regions. There is no reason this process of changing weights and methods couldn't have been automated. That my friend is a real world example involving experience. But it could have been programmed into a learning model and validated with every inplementation
Thank you ! I actually think real time weight update is theoretically possible while also avoiding messing up the network. Would be super useful if we figure out how to do that. That would enhance LLMs closer to real intelligence if we follow Sutton’s definition
A few years ago I participated in a kind of AI method using near infrared spectroscopy to monitor a process. It was a success but required constant revalidation with "known value" substance using up to a few dozen algorithms at user-selected wavelength regions. There is no reason this process of changing weights and methods couldn't have been automated. That my friend is a real world example involving experience. But it could have been programmed into a learning model and validated with every inplementation