In 2006 Netflix released 100 million movie ratings, one to five stars, and offered $1 million to anyone who could beat its own predictions by 10 percent, a line its own scientists hadn't crossed in a decade. More than 40,000 teams from 186 countries tried. For almost three years, nobody got there alone.

So rivals merged. Three teams from Austria, Canada, Israel and the United States joined forces, and their entry was a blend of hundreds of different models, each one wrong in its own way. Their closest rival was another merged crowd, called The Ensemble. The prize went to the crowd, entered 24 minutes before the close.

Every model is wrong in its own way. Blend enough different ones and the errors cancel, the same reason 787 guesses weighed Galton's ox. What a blend can't fix is the error they all share.

Netflix used parts of it, but never shipped the final blend: the extra accuracy, it wrote, didn't justify the engineering. The lines on screen are our own simulation, not Netflix's data.

So, one genius or a crowd? What did you say?

Sources: Netflix press release, 21 September 2009; Sill, Takács, Mackey and Lin, "Feature-Weighted Linear Stacking", arXiv, 2009; Netflix Tech Blog, April 2012, as quoted by BGR.
Sounds: our own, made in code (earshot, CC0).