Kavukcuoglu moving from London to Mountain View to lead Google’s AI effort

Google DeepMind’s chief technology officer, Koray Kavukcuoglu, will become Google’s chief AI architect.

Kavukcuoglu, a former aerospace engineer, has been with DeepMind for over a dozen years. According to a memo sent out by Google’s CEO, Sundar Pichai, Kavukcuoglu will be in charge of accelerating the introduction of new AI models into the market, with the focus being “seamless integration, faster iteration, and greater efficiency.”

The memo, seen by media outlet Semafor, claims that Pichai wants to speed up product development.

“We’re entering a new phase of the AI platform shift. It will require us to also shift into a new gear as a company to ensure our products are evolving just as quickly as our models,” Pichai wrote in the memo.

Kavukcuoglu, who now lives in London, is said to be moving to Mountain View, California, where Alphabet’s headquarters are.

A few weeks ago, DeepMind’s CTO joined the Big Technology Podcast to discuss Google’s AI research. He was asked: Is it scale that you want to throw at these models, or is it new techniques that will move them to the next level?

Scale is surely important, Kavukcuoglu replied, but just as important is the data that they put into those models, architectural elements, and algorithms.

“You want to make sure that you research the kinds of architectures that pay off the best under that kind of scaling property. But data is important, too, it’s as critical as algorithms, architectures, and modules that we put into the system,” he said.

According to Kavukcuoglu, when you have a particular model, you can multiply its reasoning capabilities by repeatedly using it through different techniques.

Talking about reasoning, in a new study titled “The Illusion of Thinking: Understanding the Strength and Limitations of Reasoning Models via the Lens of Problem Complexity,” Apple researchers say they found that leading AI models still have trouble reasoning. Moreover, they can collapse when dealing with more complex problems.

Kavukcuoglu said Google's goal was to build AGI in the "right way".