The joyless world of data-driven startups

As great as data is, it is also limited. Why? Because we can never gather enough data. All our data is just a simplification of what's actually going on.

Essentially we're just grabbing data that's generated by black-box tests on systems that are astronomically more complex then we can comprehend. In many cases the data tells only a fraction of the story. It's akin to some alien race trying to understand the a computer desktop by measuring the electrical inputs on a usb port and seeing how that effects the voltage output of the hdmi port.

--

Here's a telling example of the power of intuition by a quote from steve jobs responding to Marc Andreessen inquiring about the "critical" problem the iphone had of "not having a physical keyboard.":

‘They’ll get used to it.’

Any datapoint you gathered on keyboards back in the day would have told you otherwise!

It isn't about what he values versus what you value. What the author complains about are well known problems with recommendation engines. Take the naive reco algo - "You just bought a 50 inch Philips TV. You might also like - 50 inch Sony TV, 50 inch LG TV, 50 inch Samsung TV, 50 inch ..." - See the problem ? I already bought my fucking TV, you can't expect me to buy more & more of the same or similar goods.

Then there's the CF algorithm & its variants, with well known problems - namely, they don't actually match content to one's preferences. Typically, one POV dominates across the board due to the sparsity of the matrix, & getting the diversity required for the matrix to fill out takes a long long time and a very large number of people with diverse opinions. You mistakenly give a five star rating to Godfather & you are bombarded with mafia movies for a long time. You attempt to confuse the system by giving Pretty Woman five stars as well. Then the system tries to gamely proceed by suggesting "Those who watched Godfather AND Pretty Woman are more likely to watch - So I married an Axe murderer."

Can't win.

There are auto-complete screenplay software that basically make a composite of the top 100 best selling screenplays & do what in the industry is called a flip. Namely, change male to female, winner to loser, comedy to tragedy etc. These data-driven screenplay software might suggest that if you take the ladies from Thelma & Louise & replace them with grizzly old men, you get Unforgiven.

There are lyric generation software with the same flavor, umpteen loop generators & infinite jukeboxes, content recommendation systems along the same lines - since you just starred this code sample on angular, you will enjoy this github repo on react,...

Hopefully you see the downside.

You're pretty much suggesting that using strong feedback to force culture to stay within a tiny area of the total possible cultural phase space is just as interesting as allowing chaotic exploration of the entire space.

It's not just an argument against creativity, it's an argument against invention in general.

>My issue is that none of those examples are backed by any evidence that they are not doing a decent job of finding a global maximum.

That's the thing about global maxima - you only know that you've found a global maximum if you've explored the entire space.

Otherwise you've just stumbled across a local attractor, and you're stuck in a loop around it.

This isn't even a good analogy, because cultural attractors are contingent, and they vary over time. They're also unpredictable.

The reason they capture attention isn't because they're maxima in some analytic sense. They become found maxima because they summarise some aspect of human experience, so they appeal to a lot of people at once.

The spectrum of possible maxima is mysterious and not understood, which is how you get - say - Harry Potter coming out of nowhere and captivating a generation.

Culture is music, not a sine wave. You don't just want a single signal - you want a mix of related-but-different signals running all the time.