October 17, 2019

Y Kant AI Read?

Hi there,

I just spent the last two months researching a neural network from Google named BERT. By certain measures, BERT can now "read" better than humans. My new longform article for Quanta tells the story of how it got to that point.

Or did it? I think this GIF captures the spirit of what's actually going on inside BERT:

OK, it's no spoiler to tell you: BERT can't read. Not really. The story really is: why not? And what, if anything, does whatever-BERT-is-doing have to do with "real" reading? (Extra dork points if you got that reference in my subject line.)

The story also includes an extended analogy between neural networks and pie. Please enjoy 🥧

-J