I was curious about the claimed collection performance too, since this is an area that a lot of people have looked at, and a 5x improvement there would be pretty impressive.

Apparently, what makes jsblocks different is that it appears to be doing dynamic code generation. That's an interesting idea, though I'd have concerns about how it respects lexical scoping. There are a couple bumps apparent as well:

- the benchmark is only faster because it ends in a .reduce(). I suspect that code generation allows them to remove all collection generation here and just update the reduced value in a loop. If you remove the reduction at the end, then lodash is fastest, as expected.

- the jsblocks code in their own benchmark reports the wrong result. It appears there's either an off-by-one error or the generated code runs the filter step before the map. I'm not counting this against the method, as presumably it's fixable, but it might speak to the difficulty of code generation as a strategy.