Sunday, July 8, 2018

"The algorithm merchants"

TH:

In 2011, Michael Fowler, a 20-year veteran of the t-shirt business, began to experiment with ways to generate more designs.

At the time, his company, Solid Gold Bomb, had a catalog of around 1k t-shirts, each conceived by a human. But Fowler knew that the t-shirts were “a numbers game, a quantitative culture” — and to scale, he needed to dramatically increase his output.

So, he wrote a simple computer code that performed the following:

Start with a phrase (e.g. “Kiss me, I’m a ____”)
Scan a database of digital dictionaries, compile hundreds of thousands of words
Formulate a massive list of phrase variations with these words
Generate product images of t-shirts with phrases on them
Using a wide range of starting phrases, the algorithm could spit out an endless array of t-shirts. And by printing on demand, he could maintain a virtual inventory without printing shirts until they were actually ordered.

In short order, Solid Gold Bomb’s catalog ballooned to more than 22 million t-shirts.

On Amazon, the company listed 550k+ t-shirts. The hyper-specific phrases were hit or miss with customers — but small sales, aided by targeted Facebook ads, added up to a cumulatively large sum.

Soon, he was selling 800 shirts a day.

...

But as it turns out, the key to these operations (huge volume) can also be its curse — and oftentimes, these “algorithmically-generated” products can go terribly, terribly wrong.

...

a few months later, Solid Gold Bomb went out of business. Fowler, a father of 3, was left wondering how his algorithm had betrayed him.