Microsoft AI CEO sees long-term memory as key to unlocking future AI experiences

Microsoft’s revamped personal Copilot, released this week, lets users engage in natural conversations with AI using their voice. It doesn’t retain information from one session to another, but figuring the right way to implement long-term memory will be one of the most important next steps AI overall.

That was one of the takeaways from Mustafa Suleyman, the CEO of Microsoft AI, during the closing session at Madrona’s IA Summit on Wednesday afternoon in downtown Seattle. Suleyman spoke on stage with S. “Soma” Somasegar, a former Microsoft executive who is now a managing director with the venture capital firm.

Suleyman, the co-founder of DeepMind and a former Google vice president of artificial intelligence, was most recently co-founder and CEO of the startup Inflection AI. He was hired by Microsoft in March, along with other Inflection execs, to oversee consumer AI initiatives, as Microsoft struck a licensing deal and partnership with the startup.

Speaking at the event Wednesday, Suleyman gave a sense for how he sees AI unfolding from here.

“For the first time in human history, machines have learned to speak our languages,” he said. “The programming interface has already fundamentally changed. Everybody is going to get access to this tool that they can use to program the digital world around them. And obviously that programming then changes us back.”

He explained how this has shifted his mindset about developing technology.

“I don’t even think of this as an application. I think of this as fundamentally a relationship,” he said. “My team and I are now in the business of engineering personality. We’re crafting a lasting, meaningful, trusted relationship. That is the new platform, as far as I see it.”

Breaking down the components of AI systems, Suleyman noted that models are becoming increasingly factual and accurate (IQ). At the same time, he said, emotional intelligence (EQ) is coming through in more fluid interactions. And there’s also progress in the realm of AI taking action (AQ) on behalf of a user.

“On those three fronts, I think we’re seeing quite predictable improvements,” he said. “The missing piece that loops all those together, I think, is memory. … I’m pretty sure in the next 18 months, we’re going to have AIs with very good memory.”

Suleyman pointed out that much of the industry has been fascinated by the potential for artificial general intelligence, or AGI, the notion of an all-knowing and general-purpose AI on par with human intelligence.

However, he said, another definition of intelligence is the ability to focus on the right thing at the right time.

“It may be that, if we can direct processing power to the right subsystem, at the right time, that’s a kind of meta-enabler that enables us to leapfrog the challenges of just scaling it through a traditional context window,” he said. “If I have really good IQ, EQ, AQ, and memory, that’s a very powerful system — a very, very powerful system.”