Inkitt has defined itself as a publisher "without an acquisition department." It invites willing authors to put their manuscripts on its platform and promises to publish the best-performing ones as fee-yielding books. Performance, it claims, is measured by an AI-inspired algorithm.
 |
| James Beamon |
In January this year, James Beamon, one of these authors, engaged in a dialogue with the platform about said algorithm which yielded little but obfuscation and gobbledygook Inkitt-wise. I thought about this and sent him the following letter (mildly redacted):
I have posted two or three stories about Inkitt and had a chance to observe the phenomena that you describe in your post (regarding the relationship between reading behavior and their analytical engine).
Before I started to write fiction, I taught Artificial Intelligence at the University of Amsterdam, the discipline whose name Inkitt invokes as its unique sales proposition (“our algorithm is AI”).
My hunch is that this algorithm is mostly ballyhoo.
Why?
The algorithm supposedly links reading behavior to sales success, so it either (a) knows, or (b) has learned how reading behavior predicts book sales.
(Ad a) Imagine that you are the programmer, or the team of programmers hired to code the algorithm. You will have some hunches as to how the reading behavior re successful novels differs from reading behavior re less successful novels (and, perhaps not coincidentally, these hunches surface in the answers we get from Inkitt (“readers unable to put the novel down”)). There's some obvious plausibility to this, but initial hunches are not Artificial Intelligence. They represent the natural intelligence of a bunch of kids (mostly/usually), who spend their nights with a cold pizza on their lap hired to write the code. In other words, Inkitt’s AI-touting sales proposition does not hold, or at least: it did not hold at the outset.
If Inkitt has an advantage NOW over traditional (human) intelligence (agents, editors), it would be on the data side. Agents or editors don't have data about the reading behavior of a manuscript that hasn’t been read by anybody except them, whereas Inkitt, 2.5 years into its existence, can claim to possess such data.
To wit, we are talking about data that MIGHT predict sales success. To turn the “MIGHT” into a “WILL,” one would have to TEST the algorithm---meaning one has to set up experiments in which the MEASURED READING BEHAVIOR is correlated with MEASURED SALES. Unless Inkitt (or its pizza-eating coders) are GOD, they don’t know for sure in advance; that's why they, like everybody else, must TEST a hypothesis before claiming its confirmation. That’s why SCIENCE is different from TRUMP-touting MEGA-CHURCHES.
Now, such tests would have to look at a broader range than the 61 titles published by Inkitt in the meantime; in particular, they would also have to look at the darker side of the distribution, namely at PUBLISHED books that ought to do poorly according to Inkitt’s algorithm. And this Inkitt did not do.
(Ad b) Which brings us to learning. Modern AI relies on machine learning; eventually machines will know MORE than humans because they LEARNED MORE than humans. That's the unique sales proposition of Artificial Intelligence. So, Inkitt might claim that its algorithm is learning from the contraposition of the measured reading behavior of its members ("cannot put the book down") versus sales. But how does the algorithm "learn" if it looks only at the outliers, namely those books selected on the hunches of our pizza-eating coders? How about those books not selected? Perhaps they would sell even better? We're back to square one. THERE IS NO PROPER LEARNING.
Now, as to those books selected: if you look at Inkitt's book store, you see that almost all books are listed as bestsellers---no, let me retract, ALL BOOKS are listed as bestsellers. Apart from some simple hunches as to the plausibility of miracles (how plausible is it that there's not a single dud, that the algorithm ALWAYS works), there is another consideration: The one-time top ranking on Amazon cited by Inkitt as evidence means little.
I published my last book on Amazon last week, and on a total of FOUR sales it shot to bestselling position #12. The way Amazon's algorithm works, you stay in that position for ONE hour, and your rank degrades if you don't keep selling books at the best clip. So, I'm now at #74, since I didn't sell any books in the meantime.
Have another look at Inkitt's book store. The first book listed as of the time of this writing (Aug 20, 2018), titled “The Black Notebook”, reached rank #45 (of all Amazon books); the second title, “The Rogue’s Fate” reached rank #78, and so on. We are comparing apples with oranges here since my rank of #12 was in its own category, so rank #78 against all Amazon books means much more in absolute sales---but---HOW MUCH MORE? In my days, publishers touted their bestsellers with absolute numbers (“More than a million copies sold”). INKITT KNOWS THOSE NUMBERS. I would challenge them to make them public.
There’s another little miracle here. ALL their books are BESTSELLERS, but not for a SINGLE BOOK are absolute sales numbers provided. Are Inkitt’s sales so bad that they have to keeps their numbers under wraps?
This isn't our first post on Inkitt. For the previous one, go here.
No comments:
Post a Comment