Apologies up front for this issue being late and lazy, but I spent an awful long time sitting on the tarmac today. There are a couple of things I might take with in the “cloud and infrastructure” section—specifically the pieces on Google missing the boat on serverless computing, and on edge computing killing the cloud—but they do raise some good points. You should read them.
But here are the three most interesting things I came across today:
Amazon Mechanical Turk workers have had enough (WIRED):
This could be potentially significant in the world of machine learning and artificial intelligence—for better and for worse. Turkers label a lot of data for training models, and changes in the workflow and economics of that labor source could affect the workflow and economics of AI research.
Introducing Network Service Tiers: Your cloud network, your way (Google):
Long story, short: Google is offering a lower-performance, lower-cost cloud networking option. This pay-less-get-less strategy makes a lot of sense in storage, but I’m struggling to come up with many use cases where users would want higher latency and lower network-performance guarantees. However, I believe Google’s bandwidth costs have historically been higher than competitors’ prices, so perhaps users just want lower prices and there’s a limit to how low Google is willing to go on its private network.
Deep learning for Siri’s voice: On-device deep mixture density networks for hybrid unit selection synthesis (Apple):
Apple researchers explain their approach to giving Siri a better, more natural voice. (They also published two other research papers on arguably less-interesting AI problems/solutions.) Really, though, the big deal here is that Apple is talking about this stuff at all—something it has been loathe to do historically, but probably feels compelled to do in the name of attracting talent. Also note that it’s using its own “journal” for this publishing research rather than usual channels such as arXiv. I guess when you’re Apple, you can afford to do things on your own terms.