But what about this promise makes it so uniquely seductive?
Part of it is, as you pointed out, just the elimination of costly labor. That’s a capitalist’s wet dream. But the main thing that makes it attractive as a slick, highly marketable investment vehicle is that AI models are inherently black boxes.
There are ways you can examine the ways they work (for example, researchers found that the parts of an LLM that “understand” one topic, like money, can also simultaneously “understand” other different, yet related things, like value, credit, etc), but we can’t truly comprehend everything about them. It would be like looking at a math problem billions of equations large and assuming we could hold the whole equation perfectly in our brain and do the mental math to solve it. We can’t.
That means that instead of seeing “here’s our robot that is currently capable of this, but these are the components that could be upgraded/replaced, X is an issue it faces because of Y” and so on, instead you get “It’s not good at this yet, but it will be if you just throw a few billion dollars more compute at it, we promise this time.”
Problems are abstracted away to “something that will fix itself later,” or something that “just happens, but we’ll find a way to fix it”, and not any kind of mechanical constraint a VC fund manager might be able to understand.
Part of it is, as you pointed out, just the elimination of costly labor. That’s a capitalist’s wet dream. But the main thing that makes it attractive as a slick, highly marketable investment vehicle is that AI models are inherently black boxes.
There are ways you can examine the ways they work (for example, researchers found that the parts of an LLM that “understand” one topic, like money, can also simultaneously “understand” other different, yet related things, like value, credit, etc), but we can’t truly comprehend everything about them. It would be like looking at a math problem billions of equations large and assuming we could hold the whole equation perfectly in our brain and do the mental math to solve it. We can’t.
That means that instead of seeing “here’s our robot that is currently capable of this, but these are the components that could be upgraded/replaced, X is an issue it faces because of Y” and so on, instead you get “It’s not good at this yet, but it will be if you just throw a few billion dollars more compute at it, we promise this time.”
Problems are abstracted away to “something that will fix itself later,” or something that “just happens, but we’ll find a way to fix it”, and not any kind of mechanical constraint a VC fund manager might be able to understand.