Autistic people are often good at certain things. They are also almost universally poor at others.
When thinking of the later category, social skills are the first thing that usually comes to mind. Speech is another... but language isn't, as exhibited by autistic persons such as Amanda Baggs. While it'd be accurate to say that there are times when she has difficulties with language, it certainly wouldn't be accurate to say that she's bad at it!
Thinking about this issue led me to think about what common elements the things in those two categories have in common. While it's pure speculation on my part (hence the title of this post), I suspect that the reason for the difference lies in the way that they're taught.
As I see it, teaching methods can be rougly grouped into three broad categories. There may be a fourth, but I haven't thought of it. Remember that this is mainly speculation on my part.
First, there are those methods that break down a skill into its component parts or steps and teach those before teaching the integrated whole. For instance, when I teach origami, I typically start by teaching how to perform basic folds. Subsequent to this, I teach the steps of a basic pattern, composed of those folds. Later, I move on to more and more complex patterns, teaching them step-by-step. We can call these methods teaching by component parts.
If you have to wonder where I got that name, please look at the first sentence of that paragraph. Yes, I know it's uncreative. Feel free to sue.
Math, for instance, is typically taught this way. So are most subjects related to science and engineering, as well as some craft skills.
Secondly, we have those methods that teach an approximation to the desired skill before modifying it into a closer approximation which is then modified again and again until the desired skill is mastered. We can group these teaching methods together and call them teaching by successive approximation.
Social skills are typically taught in this manner in modern Western cultures. Speech is another example of a skill taught like this.
If you see where I'm going with this, give yourself a cookie.
The third category -- methods that mix the two methods describe above -- is largely unimportant in this analysis. I will, however, note that my experience has been that autistic people are typically better at these skills than they are at skills typically taught entirely through methods that fall under the second category and worse than at skills taught through the first.
In other words, I suspect that this plays rather nicely into some of my earlier speculation.