> In selecting for intelligence a harm is acquired: the risk of sevre autism.

Not sure what you mean here. You can have simultaneous selection for and against 'autism spectrum disorder', if it is heterogeneous. If for example it is a mix of actual social dysfunctionality/brain damage and greater intelligence/abstracting/systematizing, you could have selection against the former and selection for the former, and the net measured effect on the mixture could go either way.

(This is not as odd as it sounds. We seem to see something very similar with the Neuroticism trait: the overall trait appear bad for you, which makes sense since it's usually seen as not very good to have a lot of, but if you break it down into its 'facets', some of them are good for you: "Genetic contribution to two factors of neuroticism is associated with affluence, better health, and longer life" http://biorxiv.org/content/early/2017/06/06/146787 , Hill et al 2017. It's analogous to how there are minimal sex differences on Big Five personality when examined at the Big Five factor level, but when broken down to their sub-factors, sex differences jump out at you.)

> By selecting for intelligence, we incidentally acquire a problem.

That is directly at odds with the analysis in the Grove et al. paper; the genetic basis of increased educational attainment is distinct from that associated with 'severe' autism.

I think this is a case of wanting to see patterns that aren't necessarily there; we'd like to see a nice even spectrum of behavior, with undesirable extremes at the ends. It's a nice reductive interpretation of complicated data. The problem is that it doesn't seem to fit. At least so far; genomic studies are very useful, especially as a starting point, but ultimately mechanistic explanations that generate testable genetic hypotheses are the gold standard for biology. And for good reason.

Still, given that what evidence exists does suggest a different framework, I'd be very hesitant to conclude that there is a trade-off.