@Michaelzsguo: Over two years ago, Google co-founder Sergey Brin stood at AGI House, admitting Google had fallen behind in large models and vowing to catch up. We all saw what happened next: Gemini steadily caught up, and when Gemini 3 launched six months ago, Google was back in the top tier by many metrics.

More than two years ago, Google co-founder Sergey Brin stood at AGI House and admitted that Google had fallen behind in large models, vowing to catch up.

The subsequent results are clear to all: Gemini steadily caught up, and when Gemini 3 was released six months ago, Google had at least regained a top-tier position across many dimensions.

But six months later, the landscape has shifted again.

This time, Sergey—still fighting on the front lines—returned to AGI House for a new Q&A session. He discussed his views on AGI and superintelligence, Transformers, world models, and AI for Science. He explained why Google is now heavily investing in coding agents, how he, Demis, and Corey divide responsibilities for Gemini, and whether he still believes Google can catch up again.

Here are the key takeaways:

• Google has re-prioritized coding as one of its highest priorities. Sergey admitted that they probably should have focused deeply on coding earlier. Now, Google is very committed to this direction.

• He still has confidence in Gemini, but acknowledges that competitors are improving rapidly in coding. He noted that some models perform exceptionally well on long, deep coding tasks, while Gemini Flash’s strength lies in speed, making it suitable for interactive rapid iteration.

• He believes the key to AGI is not just “being able to answer questions,” but self-improvement. Personally, he leans toward defining AGI as an AI that can improve itself, rather than simply “one that can perform any human task.”

• World models are important, especially if AI is to enter the physical world. If AI is to do what humans do, it must understand the world, predict the consequences of actions, and interact with reality. This extends to robotics, multimodal models, and video models.

• Transformers may not have hit their limit yet. He believes that Transformers have already expanded from text to images, video, and other domains, and continue to evolve. Architectures similar to Transformers could potentially lead to AGI.

• His role at Google is more of a “catalyst” than a formal leader. Corey, Demis, and others handle organization and delivery, while Sergey constantly questions the team about whether they are missing important directions or underestimating certain priorities.

• He is not overly pessimistic about AI replacing human purpose. He uses the examples of chess and Go: even after machines surpassed humans, people didn’t stop playing. Instead, they became stronger with the help of AI.

Overall, you can see the real anxiety and focus within Google right now: AGI, self-improvement, coding agents, world models, and using AI to build the next generation of AI.

(Chinese subtitled video)