After more than two hours of announcements at Google I/O that included flashy new models, AI agents, a smarter search box, tools for work and video creation, and smart glasses, Google DeepMind CEO Demis Hassabis delivered a poetic, if slightly abrupt kicker:

“We’re at the foothills of the singularity,” he said on stage in Mountain View Tuesday.

The decision to end the show that way was a deliberate one. “We debated it back and forth,” Hassabis said in an interview a few hours after the event. “I was closing, and I wanted to be authentic about what I’m thinking with AGI,” he said. “The singularity, at least my interpretation of that word and that term, means the era that we’re in.”

When ChatGPT burst onto the scene nearly four years ago, it sparked a kind of existential reckoning about the role humans will have when machines get powerful enough to take on some of our thinking. Since then, as happens with all technology, people have gotten more used to AI in everyday life, and improving large language models have become essential tools across industries.

Still, the singularity reference was a reminder of a larger reality lurking underneath the hood of every new product announcement: Every seemingly incremental feature has become a proxy for the steady march of AI capability.

For instance, Google showed off how its Antigravity 2.0 product can autonomously build a computer operating system for less than $1,000 (a task that would have required teams of software engineers months to complete in the pre-AI era).

These “harnesses,” which coordinate and corral AI models to give them direction and reliability, seem less like a breakthrough and more like infrastructure that makes AI models more useful.

But Hassabis sees the increase in machine autonomy as one of the key steps on the gradual march toward the singularity.

“This year, I really felt … that it’s the beginning,” he said. “Agents are starting to work, becoming useful harnesses … coding is starting to work properly. Areas of science and math are being accelerated.”

Even Gemini Omni, which the company marketed on stage as a fun consumer product that allows people to use language prompts to turn real-life videos into surreal ones, is actually a step toward AI models that can better understand the world.

Already, Hassabis said, Alphabet’s self-driving car division Waymo is testing AI models that would give autonomous vehicles a kind of “imagination” to react to unpredictable or dangerous situations.

Text-to-video models could be the key to general purpose robotics and artificial general intelligence, he said, noting that “an AGI is going to have to understand the physical world.” Musing about potential applications, he said, “I could even imagine planning with visuals, not in text tokens but in visuals. Humans certainly do that. We’ll see if AIs need to or can get around it. But in my view, it’s almost certainly going to be needed.”