After resigning as president of Google’s parent company Alphabet in 2019, co-founder Sergey Brin is back at Google — and working on AI “pretty much every day,” he confirmed for the first time in a conversation with Climate Corporation co-founder David Friedberg earlier this week.
In a fireside chat at the All-In Summit, Friedberg asked Brin about Google’s plans for AI, including whether the company intends to focus its efforts on a huge, general-purpose AI model, also known as a “God” model.
“If you can build the God model, you’re done,” Friedberg said. “There’s this one thing to rule them all. Or is the reality of AI that there are lots of smaller models that do application-specific things?”
Related: Google’s AI Overview Has Changed Since Its Debut, and These Websites Have Benefitted the Most
Brin said that 10 to 15 years ago, separate AI systems were needed for different tasks. A chess-playing AI, for example, was different from an image-generation AI. Google has historically used smaller AI models like these for specific tasks, and more recently even combined three separate models, for theorem proofs, geometry, and general language, to win a silver medal at the International Mathematical Olympiad.
Though these smaller AI models have been effective, “the trend is to have a more unified model,” Brin said.
“I don’t know if I would call it a God model,” he said. “But certainly shared architectures and ultimately even shared models.”
Google co-founder Sergey Brin. Photo Credit: Lionel Hahn/Getty Images
Friedberg pointed out that if shared architectures and models were the future, AI systems would require more computational resources.
Related: New Google Report Reveals the Hidden Cost of AI
Brin said that while accelerating AI’s capabilities would require “a lot” of resources, it may not be easy to predict how much, and current estimates may not take internal improvements into account.
“The algorithmic improvements that have come over the course of the past few years are maybe even outpacing the increased compute that’s being put into these models,” Brin said. He added later that, “For us, we’re building out compute as quickly as we can. We just have a huge amount of demand.”
Google’s cloud customers want resources like specialized processors to train AI. According to Brin, Google has had to turn down customers because they don’t have the resources to meet demand.
Meanwhile, Google launched a number of public-facing AI products, including AI overviews in Search and summaries for products like Gmail and Docs. On Tuesday, the state of Nevada announced that it will launch a Google-powered AI system to recommend whether unemployed workers in the state get benefits.
Learning from Failure
Brin acknowledged that not all of Google’s AI efforts have been successful. In a talk in March, he said that Google “definitely messed up” the launch of its Gemini image generator, which produced historically inaccurate images.
Now he says that it’s important to get AI out there and take a risk — even if it means embarrassment.
“Is this something magical that we’re giving the world?” Brin asked. “I think as long as we communicate it properly — like saying look this thing is amazing — and we’ll periodically get stuff really wrong, then I think we should put it out there and let people experiment and see what new ways they find to use it.”
Related: Google Pulls Gemini AI Olympics Commercial After Backlash