When I walked through Google’s AI labs in 2015, executives boasted about how research was the lifeblood of its burgeoning AI products.
John Giannandrea, who was at the time Google’s head of engineering, though he would soon oversee all of the company’s AI and search products, eventually leaving to join Apple, told me that core AI research on simple things like handwriting was essential.
“We, as a company, want to understand how people would write a word,” he said. “So that’s something we would invest in forever, even if we didn’t have a product.”
The argument was that there was so much data still yet to be collected, and so many insights for machine learning to unearth, that it would be unwise to stop researching even basic tasks like handwriting.
But now Google’s cloud business is playing to a different tune.
Andrew Moore, who rejoined Google as its head of Cloud AI in September 2018 after also serving as the dean of Carnegie Mellon’s computer science department, said that Google Cloud is no longer interested in AI research unless it leads to a product.
“Deployment is the word we’ve been using most commonly in Cloud AI at the moment,” Moore said at a press briefing April 11. “It is all about taking a project from initial inspiration all the way through to it running for your business reliably. We’re no longer interested in the world of proof of concept being the main form of AI in product.”
Google Cloud operates as almost a separate entity inside of Google, which doesn’t suggest that Google as a whole won’t be researching new AI techniques. Google products with the largest potential to change entire business, like its human-voice emulating AI called Duplex, came from its AI research, as did the algorithms currently powering Google search.
But Moore’s statements signal a new era within the company, one where business outcomes trump coding experiments. While advertising is still Google’s cash cow, it’s increasingly looking to Cloud as a new source of revenue as the online ad business matures and changes.
“The urgency we have inside Google Cloud to make sure that we can help the world safely take this stuff and actually deploy it where it can do good, and so the technical challenges have shifted recently,” Moore said.
“It’s very much more for us about making artificial intelligence safely usable, and less about inventing brand new technology. There’s so many other parts of the [software] that we need to work on to make AI deployable.”
Correction (April 12): An earlier version of this post stated that Moore rejoined Google in December, rather than September, 2018.