|
Contrarian hypothesis for you :) The more I work on AI/ML the more I think the fears around AGI have been misplaced. General intelligence almost by definition involves general motivations (rewards). Very general rewards inevitably make the AI / ML model hard to control, unpredictable and difficult communicate with (just like a human). If you find managing humans to be like herding sheep, try herding sheep, then try herding a non-biological general intelligence that perceives the world in a totally different way. Therefore, what humans want AGI to be is largely a contradiction. We want something so clever and general it has the complexity and nuance of a human but so reliable and compliant as to be like a machine in a factory i.e very not human. These two things just don't go together. Take the classic AGI that goes all out making paper clips and destroys everything. Are we seriously suggesting we have made something so nuanced and clever as to understand building regulations, supply chains, the art of the deal, the entire manufacturing process, human incentives, business, finance, HR, taxes and everything else that goes with exclusively running a paper clip manufacturing operation. Yet it is also so utterly single minded and blinkered that can't conceive of anything beyond more paper clips, even the obvious inevitable consequences of such a pursuit. In reality it really seems wants and skills are very much too sides of the same coin, you just don't learn about things you don't want and so the only way to learn about lots of things is to want lots of things, not just paper clips. So I think maybe we could make AGI relatively soon and even now could have a good stab at lower level intelligence. The reality I see though is that we just actually don't want to because it wouldn't actually be very useful. What we want, is what are doing, extremely capable but ultimately intellectually dumb factory machines / cars that do as we wish on repeat. |