While this is a great project from both an academic and performance perspective, the resulting images are substantially larger than ones produced using Mozjpeg. From the article:

We at CloudFlare make sure that our servers run at top notch performance, so our customers' websites do as well!.

Correct me if I'm wrong, but it does seem to me that this is of greater value to CloudFlare in terms of electricity savings than to CloudFlare's customers. Less computing time for CloudFlare at the expense of larger output files for CloudFlare's customers to serve, files which will potentially be served again and again and count against the customer's transfer quota. It might not seem like much, but it adds up when you consider how many times a single image in the wild can be downloaded. Consider the following example:

  # Original size is 21796912 bytes
  # jpegtran with SIMD
  jpegtran -copy none -optimize -progressive bigjpeg.jpg > bigjpeg_jpegtran.jpg
  # bigjpeg_jpegtran.jpg resulting size is 19771224 bytes

  # mozjpeg 3.1
  mozjpeg -copy none -optimize -progressive bigjpeg.jpg > bigjpeg_mozjpeg.jpg
  # bigjpeg_mozjpeg.jpg size is 19328832 bytes

That's a difference of 442KB! I'm not emphasizing the computing time here. I do not dispute that this is orders of magnitude faster than Mozjpeg, but isn't it worth doing the extra work to get a smaller file? That file could be served billions of times. My argument is simply that the extra computing effort in the beginning leads to much grander savings in the long run.

In my (quick) testing Jpegtran's squishing performance is consistently less than that of Mozjpeg. So what, actually, is right thing to do in the end?