Earlier quoted context omitted.
Same thinking, I expect, that gave us XSLT.
XSLT was proposed as a limited transformation language, and actually (still) does a very good job of that. And since we're doomed to repeat the past we now see various XSLT-ish languages being developed for JSON/YAML/etc, relearning all the edge cases that drove XSLT into its current form. On the other hand, students of the past will notice that Jsonette looks awfully similar to Ant build files, which sucked. JSON/XM…
I think we are finally starting to accept that Turing-complete data is a bug, despite what the Lisp people have been saying for 50 years, and is not a feature on the Internet. Microsoft served as the poster child in the 90's.
The tension, I strongly believe, comes during debugging. When you are trying to figure out why the data doesn't look right, the less the code resembles the data being emitted, the harder it is to locate the source of the error. Templating engines get pretty close to addressing that concern, in ways that transforms or generators don't. So we keep trying to invent a better one. But once you let a little logic in, everybody wants all the logic.
Ant found the Uncanny Valley, in a way that few other tools have managed. I think I'm okay with your implication that it was actually worse than XSLT, which is saying a lot.
[edit] Ant is also the reason I shift uncomfortably whenever someone announces proudly that a tool was 'designed and implemented on a plane trip'. I'd rather have code you wrote when drunk than code you made while stuffed in a sardine can, subjected to noise levels that OSHA would have issues with, and suffering from altitude sickness. You literally designed this while your brain wasn't getting enough oxygen.