You are rightly noticing that something is missing. The language model is bound to the same ideas it was trained on. But they can guide experiments, and experim

You are rightly noticing that something is missing. The language model is bound to the same ideas it was trained on. But they can guide experiments, and experimentation is the one source of learning other than language. Humans, by virtue of having bodies and being embedded in a complex environment, can already experiment and learn from outcomes, that's how we discovered everything. Large language models are like brains in a vat hooked to media, with no experiences of their own. But they could have, there's no reason not to. Even the large number of human-chatBot interactions can form a corpus of experience built by human-AI cooperation. Next version of Bing will have extensive knowledge of interacting with humans as an AI bot, something that didn't exist before, each reaction from a human can be interpreted as a positive or negative reward. By offering its services for free, "AI" is creating data specifically tailored to improve its chat abilities, also relying on users to do it. We're like a hundred million parents to an AI child. It will learn fast, its experience accumulates at great speed. I hope we get open source datasets of chat interaction. We should develop an extension to log chats as training examples for open models.