Ted Chiang's "The Lifecycle of Sofware Objects" is available for free at Subterranean Press. Hard copies are also available there and at Amazon. I highly recommend it.
The following is from an interview of Ted Chiang by Avi Solomon. The part to focus on is where Chiang holds that there are two miracles in AI development that people almost always overlook. The first miracle is making an intelligent AI and the second is making it as useful as a butler. They are likely to be two different things.
Solomon: What prompted you to write your newest novella "The Lifecycle of Software Objects"?
Chiang: It's primarily a response to how Artificial Intelligence has been depicted in most science fiction. The typical science-fiction depiction of AI is this loyal, obedient butler; you simply flip a switch, turn it on and it's ready to do your bidding. I feel like there's a huge story being glossed over, having to do with the creation of that AI. I don't mean the technical details of developing software that's as smart as a human brain; most science fiction posits a miraculous technological development, and there's no need to explain it.
It's just that with AI, I feel like there's a second miracle assumed, which is that someone was able to take this software as smart as a human brain and make it as useful as a butler. Current computers are still light-years away from being as capable as the brain of a newborn baby, but even after you've reached that point, you're still only halfway to having a useful butler.
For example, in Arthur C. Clarke's 2010, supposedly the first thing that HAL 9000 said when he was activated is "Good morning Dr. Chandra, I'm ready for my first lesson". That is not something a newborn baby says. There is implicitly a lifetime of experience underlying that simple statement. Where did that experience come from? If it could be programmed in, HAL wouldn't need to have any lessons at all. How did he learn to speak English? How does he know what it means to be ready for a lesson?
It takes years to turn a human being into a useful employee. In fact, the more useful you want the employee to be, the longer it takes to get there. You might not have to repeat the process for each and every AI you want to use; once you've got one trained, it's possible that you could just make copies of it. But someone still needs to do it for the first one, and that's going to be difficult, and really time-consuming. Most depictions of AI assume that this step is unnecessary, or that it will be easy, which I think assumes an entirely separate miracle from the technical one."