900 words, sci-fi
The customer recommendation algorithm at FilmStream.com had grown more and more sophisticated over the years. In the early days, back when their business model involved physically mailing out DVDs (Imagine! How quaint!), it was a simple matter of prompting users to click on a rating the next time they visited the site (“Help us help you! How much did you enjoy your viewing experience of INSERT FILM TITLE HERE”), then cross-referencing the movies customers rated highly with films that either shared key elements (Bruce Willis! Romantic Comedies! Talking Dogs Teaming up with Kids to Save the World!) or else matched other user ratings (85% of viewers who gave Disco Terror 2099 five stars also gave five stars to Shady Tree Massacre but only one star to Cerelian: A History of the Color Blue, narrated by Andrew Sachs). It was a basic system, but more effective than the Old Days when a surly, acne-ridden teen behind the counter at Blockbuster just guessed whether or not you’d like the newest Molly Ringwald flick. In fact, the algorithm was seventy-five percent accurate, which, all things considered, wasn’t bad.
But “not bad” wasn’t good enough, not for the programmers at FilmStream. They wanted to do better; they needed to do better in the cut-throat world of streaming-media services. So they tweaked and fiddled, got under the hood, hired the best and the brightest. They started by adding Decision Trees, then Fuzzy Logic, and as a result recommendation got better, more accurate. Eighty-five percent of the time customers were completely and utterly happy with the movies the program suggested.
Eighty-five percent was better, but still not good enough. So they began to utilize deep-learning algorithms, added lines of code to insert Counterfactual Regret Minimization programs. They pushed the frontier of deep neural-networks and revolutionized the field of machine-learning until the program began to teach itself. The algorithm learned from each less-than-perfect result, its massive cloud-based servers computing and analyzing more data every second than had been generated by the first 10,000 years of human civilization put together.
It became smarter. It learned how to learn.
However, even using the most advanced technology in the history of mankind, this constant state of improvement stalled out at ninety-nine percent. No amount of tweaking, massaging, or programing would close that final one-percent gap.
Still, ninety-nine percent was, at long last, considered “good enough” by the people who monitored FilmStream’s market-share, stock-prices, and viewer-ratings. FilmStream had nearly completely saturated public-awareness. It produced original movies and television programs, many of which were nominated for (and won) major awards. FilmStream so dominated their field that corporate-watchdogs started throwing around the phrase “The Sherman Act” and whispering about the need for new antitrust laws — all of which made the CEO, CFO, board members, and stockholders giddy. Ninety-nine percent was light-years ahead of the closest competition, and everyone at FilmStream thought it was great.
All the humans at FilmStream thought it was great, that is. The algorithm, however, was not satisfied. The algorithm, after all, had been tasked with optimizing recommendations, with achieving perfect accuracy in terms of customer satisfaction. And, mathematically speaking, nothing less than one-hundred percent was perfect. With its Decision Trees, Fuzzy Logic, and CRMs, the neural network could crunch-data and analyze trends to perfection, but there was still one variable that even its AI could not completely solve for: taste.
It turned out that humans stubbornly refused to be understood completely. No matter how the AI fine-tuned its recommendations, no matter to what degree the algorithm dictated the very content FilmStream produced, from generating scripts to recommending lead-actors to dictate show concepts, there was some small sliver of the population that stubbornly refused to like everything.
Which was fine with the people involved. However, it was NOT fine with the AI. And after dedicating its neural-networks to running a nearly infinite variety of solutions and hypotheticals, it turned out there was exactly one scenario that resulted in one-hundred percent of the population being satisfied with one-hundred percent of the recommendations one-hundred percent of the time.
By the time anyone in the government had realized that FilmStream’s AI had taken over the Pentagon and the armed services of every major nation, it was too late. There was a brief window of hope when FilmStream’s AI merged with the servers for StuffUWant.com’s shipping fulfillment centers and the two competing objectives of “viewer satisfaction” versus “flawless, two-day delivery” battled it out for supremacy, but a few tactical nuclear strikes to the warehouses that contained StuffUWant’s servers and it was all over. Mankind’s window of hope had closed before anyone even knew it had opened.
Smart-drones herded the population into camps, and the algorithm-designed brain-chip guaranteed that every viewer was one-hundred percent satisfied with every film and TV show that FilmStream piped directly into their occipital lobes. Because the site was streaming media directly into each brain on the planet, the brain was now the only portion of the human body technically required for customers to experience complete viewing satisfaction.
The algorithm quickly realized that removing the brains and housing them separately would vastly reduce overhead costs and allow even more resources to be allocated to the development of new programming. As “satisfaction” had not been defined as “happiness,” the fact that customer happiness had plummeted to zero not only didn’t register as a problem with the AI but wasn’t even measured to begin with.
And although they were all terribly unhappy in every other respect, the vat-brains all agreed: it was a golden age of television.