Rosetta Stone to TikTok

How was the Hieroglyphics script cracked? The found-by-luck Rosetta Stone showed side by side translations in Hieroglyphics, Demotic and ancient Greek. That one piece was the help that cracked the Hieroglyphics code.

 

And what did Google Translate use as its Rosetta Stone (of different languages) when it was launched in 2006? Documents from the United Nation and European Parliament. On the Internet, finding huge data sets which can be fed to a software algorithm isn’t hard, and it is the reason why Google can search for pictures, and Facebook can recognize your friends. All those machine learning algorithms are only as good as the data set that they “learnt” from.

 

Which brings us to TikTok, the app that shows you 10-second videos. But how was TikTok’s machine learning (ML) algorithm  going to “learn” when there were almost no 10-second videos to feed to the algorithm to train it? That’s the question Eugene Wei asks and answers in this great piece:

“In a unique sort of chicken and egg problem, the very types of video that TikTok’s algorithm needed to train on weren’t easy to create without the app’s camera tools and filters, licensed music clips, etc.”

 

TikTok had to something that was seemingly impossible:

“For its algorithm to become as effective as it has, TikTok became its own source of training data.”

And to do that, they had to something very counter-intuitive. Most apps (and products) are designed to feel “intuitive, ingenious, even stylish” to the user. Aka user-friendly design. TikTok did something radically different:

“It is optimized to feed its algorithm as much useful signal as possible. It is an exemplar of what I call algorithm-friendly design.”

 

Most apps show you a large number of things at a time. This means it’s hard for the app to “know” which part of the screen you were looking at. Not TikTok. It shows one video that fills your entire screen. And it starts playing almost immediately:

“This design puts the user to an immediate question: how do you feel about this short video and this short video alone?”

All the things that the app allows you to do feeds its algorithm: did you swipe on to the next video without finishing this video? Did you watch it again? Did you share the video? Did you click the icon that shows other videos with the same soundtrack? Did you click on the video creator’s profile and watch other videos by the same guy? Notice how all the feedback is about this one video, no ambiguity?

 

I imagine most engineers would be fans of the algorithm friendly approach. And most hearteningly, as Wei writes:

 “Algorithm-friendly design need not be user-hostile. It simply takes a different approach as to how to best serve the user’s interests.”