We technically wouldnt have to in the sense that its not physically impossible for us to “stumble into it” while trying to mimic the basic brain architecture that we do know, but yes, this is one of the reasons I tend to be a bit skeptical of AI claims, we’re trying to mimic something without even knowing how the thing we’re trying to mimic works very well. Clearly we havent had zero success just given that we can make something that can talk well enough to sometimes fool people into thinking its a human, but we also clearly havent gotten all the way there. If the tech we’ve built will work with enough scaling or tweaking, or if we need something fundamentally different, I dont know.

My guess, though it is just a guess, is that we’ve probably got something that could serve as a component of some future agi system but that there’s probably more to it we havent figured out that we need to add to get there beyond just making the thing bigger and feeding it more data. Humans do learn from experience after all, and I would imagine the total data input of all one’s senses constantly working adds up to a lot, but we also clearly dont need to be fed all the information to be found on the internet just to learn how to think and talk, the fact these AI models need so much training data makes me suspect that either we’re missing something fundamental and trying to compensate with more training, or else we’ve devised a way to doing this that is really inefficient compared to however our brains do it.