Alphabet Spells Out YouTube Recommendations: Are Some Letters Omitted?

September 23, 2021

I have been taking a look at Snorkel (Stanford AI Labs, open source stuff, and the commercial variants). I am a dim wit. It seems to me that Google has found a diving partner and embracing some exotic equipment. The purpose of the Snorkel is to implement smart workflows. These apparently will allow better, faster, and cheaper operations; for example, classifying content for the purpose of training smart software. Are their applications of Snorkel-type thinking to content recommendation systems. Absolutely. Note that subject matter experts and knowledge bases are needed at the outset of setting up a Snorkelized system. Then, the “smarts” are componentized. Future interaction is by “engineers”, who may or may not be subject matter experts. The directed acyclic graphs are obviously “directed.” Sounds super efficient.

Now navigate to “On YouTube’s Recommendation System.” This is a lot of words for a Googler to string together: About 2,500.

Here’s the key passage:

These human evaluations then train our system to model their decisions, and we now scale their assessments to all videos across YouTube.

Now what letters are left out? Maybe the ones that spell built-in biases, stochastic drift, and Timnit Gebru? On the other hand, that could be a “Ré” of hope for cost reduction.

Stephen E Arnold, September 23, 2021


Comments are closed.

  • Archives

  • Recent Posts

  • Meta