Welcome back, magazine reader— …hmm, we need a cool name for our much loved and obviously very cool and tasteful audience, don’t we?
…well, I’ll get back to you when a decision is made. In the meantime, I hope you’ll enjoy the selection of stories with which Mythaxis is closing out the year. As always, this issue has a bit of everything: scifi, fantasy, horror; beats of comedy, beats of emotion; and, entirely unintentionally, a theme of “family” (one way or another) is shared by almost every story in the issue. Also, as you might guess by glancing at this issue’s cover and the art for each of the stories, we once again have turned to Artificial Intelligences for some visual flair, this time under the guidance of Micah Hyatt—check the bottom of this page for a proper thank you!
Whether AI-art is something we use again is a topic for debate, however. As a tool, AI is obviously both powerful and convenient, but the much publicised question of how ethically those training the underlying technologies have treated artist copyright holders still hangs around it. Your Humble Editor usually creates all the story art and does so from rights-free sources, so arguably no other artists are being out-competed there; but, as a justification for my still accessing striking and distinctive AI-crafted material to use, that smacks a bit of sophistry to me—there’s still the question of whose work an AI learned its trade from. If a demonstrably ethical AI-based system is to arise, that will be another matter… but whatever else happens, we will return to platforming flesh-and-blood artists for our covers in 2023.
Anyway, enough about AI, let’s talk some more about AI—our quest for an artificially intelligent editor, that is!
In our previous issue, I revealed that the Mythaxis team was training up what I will confidently assume is the world’s first AI slush reader, the Slushbot, with the goal of one day publishing the first magazine edited by an AI. Our first “live” attempt was with version 4.0. All it needed to do to win its independence was evaluate our magazine submissions and match my selection performance with such frightening accuracy that I fled our luxurious offices in terror. It didn’t do too well, however, rejecting seven of the eight stories I considered best of the window, thereby proving its limited appreciation for real quality.
Since then we’ve revised our training strategy. v4.0 gave a percentage rating as to whether each story was “accepted” or “rejected” by me using a one-shot classifier to analyse just the first 2048 characters of the text (we’re not exactly running Google’s Cloud Machine Learning Engine here). For reasons sooner or later to be revealed, v5.0 now looks at the first and last 1024 characters instead. Both versions used the stories published in Mythaxis since 2020 as the “accepted” standard. So, what did we we discover?
Well, at first glance, it did a lot better. Slushbot v5.0 also accepted five of the six stories I picked during our last submissions window, and the one that it rejected was only by a 1.4% margin. Pretty good, right?
What’s that you ask? “Did it also reject all the stories you rejected?”
In fact, of the 178 stories it evaluated, Slushbot v5.0 accepted 98. More than half. And I didn’t mention this, but it liked some of my rejections a lot more than my acceptances. Only one of those six had more than a 10% swing in its favour, 17 of the rejected stories beat that, and many more beat the other five. Most of my accepted stories enjoyed the kind of winning result that only a Brexiteer would ever call “a landslide”.
So, although we saw a notable improvement in its evaluation of my preferred stories, overall the Slushbot still performs slightly worse than a coin-toss. Fair to say, we still have a long journey ahead. But that doesn’t mean there isn’t other interesting analytic news to share though. For example, take a look at these graphs:
Interesting, right? That’s the averages of the accepted vs the rejected stories for our last submissions window. I expect you’re hungry for more, so how about these:
Here you can see the performance of the six accepted stories individually—clear as mud, I’m sure you’ll agree. If it wasn’t for those meddling question marks! Yet fear not, all will be revealed…
…in the first editorial of 2023—until then, I wish you all a very Happy New Year!