Comment by HarHarVeryFunny
7 hours ago
Well, humans developed language. Language is just a tool that let's us leverage our innate intelligence.
I'm sure that we will eventually build artificial brains, capable of bootstrapping communications and language for themelves (if run en-masse in a simulation where the benefit of communication would emerge). An LLM can't do this since it is by definition/construction something only capable of learning a pre-existing language.
An artificial brain, just like a wet jiggly one, is always going to be more intelligent than a one-trick pony like an LLM - a language processor, but it is notable how intelligent that one-trick pony nonetheless appears to be.
I think it's interesting that you think we could bootstrap an artificial brain with no inputs from human culture. I disagree, but am open to an existence proof of this kind. Such an artificial brain would be totally alien to us, of course. I wonder how differently it would perform versus something more grounded in "real" culture and writing?
I'd say that human babies and LLMs are both existence proof that prediction and prediction error feedback is all you need to learn. The artificial brain/baby would be designed to learn just like us by prediction, and should therefore be capable of learning language from scratch just as we do.
You could choose to lock it in a virtual or physical basement with printout of the Common Crawl dataset and raise it like an LLM that learns language with zero real world grounding, in which case it may feel a bit like an LLM (but smarter - able to learn, etc), or you could let it interact with the real world and learn everything, including language, that way and be a lot more grounded.
It's hard to guess how the grounded version would feel to talk to vs the CommonCrawl one - I think it would mostly come down to how far we wanted to go in making the artificial brain have all the moving parts of a human one. In an odd way the more human we tried to make it, the more alien it might feel, but not for the "uncanny valley" reason you might imagine...
The thing with an LLM that makes it feel so human is that they are designed to 100% copy humans - their output is 100% driven by the training goal of trying to exactly match the samples in the training set. As soon as we start to try to build something more brain-like then it's behavior is going to be a lot less predictable - not just auto-regressive "auto-pilot speech", but driven by it's own internal thoughts, emotions, innate traits, etc - depending on how much of our brain we tried to copy in the artificial one. I expect it would feel a lot more organic, less robotic, to talk to, but at the same time perhaps less human since unlike the LLM it's not built just to mimic human speech.