The latest billionaire trend? Doomsday bunkers with a flammable moat

We don't really know where the plateau for the current AI techniques is. A lot of what we see looks impressive, but it's very superficial in practice. Pretty much all AI today boils down to feeding huge volumes of data into a neural network that ends up creating a compressed representation of the data, and then doing stochastic predictions based on that model. This is great for doing stuff like text or image generation, but it simply doesn't work for any applications where there's a specific correct result needed. What's worse is that use of such systems to control things in the physical world is incredibly dangerous as we're seeing with self driving cars.

Since the neural net is simply comparing numbers together to make decisions it doesn't have any understanding of what it's actually doing in a human sense. It's not able to explain the reasoning behind its decisions to a human or even guarantee to understand human instruction. And it's not aware of its own limitations.

In order to make an AI that can replace a human decision maker it would need to have an internal representation of the physical world that's similar to our own. Then we would have to teach it language within the context of the world. This is how we could build an AI that can be said to understand things and that we have a shared context with allowing us to communicate in a meaningful way. People are experimenting with this stuff, but this sort of stuff is still in very early stages, and it's not clear that techniques used for LLM models will work well for this approach.

I'd caution to be highly skeptical regarding AI claims we're seeing because most of these claims are made by people who have very little understanding of how this stuff actually works, and whose job is to sell this tech to the public. Pretty much none of the actual experts in the field share this optimism.

Of course, nobody knows what the future brings and we might make some amazing breakthroughs in the coming years. However, given what we know right now, there's little reason to expect this sort of exponential growth to continue for long. It's also worth noting that we've already gone through a wave of similar hype back in the 80s where people started getting really impressive results with neural nets and symbolic logic, but scaling that turned out to be much harder than anybody anticipated.