Surely the way one would represent this in XML, rather than `srcset="foo 1x, bar 2x"`, which strikes me as odd, would be:

   <picture>
      <srcset media="(min-width: 40em)">
         <source size="1x" src="big.jpg" />
         <source size="2x" src="big-hd.jpg" />
      </srcset>
      ...another srcset...
      <img src="fallback.jpg" />
   </picture>

Fractionally more verbose, but really a lot less fiddly.

The article states that the final problem on the Boston globe redesign (meant as a proof of concept for responsiveness) was that image prefetching feature to speed up rendering which happens before html parsing. Thus they needed a way for browsers to parse that information separately ahead of time.

I guess that it should be possible though for browser to parse a html fragment rooted on the picture tag, and then plug that tree back later on in the full document tree when it is constructed. Or is it simpler to search for picture/img attributes? Oh, there's this whole implicit tag closing business in html though...How do we know where to stop parsing a fragment? At least attributes values stops at the end of a string literal, or on a tag end. Perhaps that's the reason why they went for a dsl in attributes.

I agree with you though, it's cleaner your way, and perhaps xhtml could use that approach in the future?

Quite, and it has the enormous advantage that I can extract all data using nothing more than an XML parser, rather than having a two stage [parse XML -> parse embedded DSL] parser for special cases. Even the media aspects of the srcset could probably be better expressed (with more verbosity though) as a standard XML structure.

I really wish it was - though I'm far from a fan of XML for most cases, it does work rather well for this when used as intended...

Verbosity was a huge argument against <picture>. People were ridiculing it with complex use cases that required awful amounts of markup.

Hixie was against using elements, as it's harder to spec (attribute change is atomic). Eventually <picture> got a simplified algorithm that avoids tricky cases of elements, but at that point srcset was a done deal.

At least we've got separate media, sizes and srcset instead of one massive "micro"syntax.