Google’s Sergey Brin Sees A Path To AGI But Not Beyond It

In a recent AGI House interview, Sergey Brin described Gemini as a system whose capabilities are not just evolving but integrating world knowledge across languages and modalities. He said the software that AI runs on has also evolved beyond what it was originally designed for, and while Brin can envision Gemini achieving AGI, he also couldn’t see what comes next.

AGI: Artificial General Intelligence

AGI is a level of AI that can learn, understand, and apply knowledge across tasks in a manner similar to humans. Today’s AI can produce useful answers, write code, analyze images, and solve many narrow problems, but it does not yet understand the world or independently apply knowledge across domains the way a human can.

OpenAI, Google DeepMind, and Anthropic are all developing AGI, but they emphasize different reasons for what they want to do with it. OpenAI focuses on economic benefits, Google DeepMind emphasizes scientific discovery, and Anthropic prioritizes human progress.

Next Big Thing: AI Capabilities Are Converging

Brin said that Google’s earlier AI progress relied on specialized models that were built for specific tasks. But he said that Gemini is increasingly achieving state-of-the-art performance across multiple domains like mathematics and scientific reasoning. What Google is seeing is that capabilities that used to rely on models trained to do specific things are now giving way to model families that can do it all: convergence.

He also said that convergence was something that happened; it wasn’t something he expected when Google began developing AI.

The context of his answer was a question about what the next big thing is, with his answer being convergence.

Brin responded:

“I think the exciting thing is that all of these things are converging to the same general models.

In the past, we would have to have specialized models. And in the case of protein folding, we obviously still do.

But increasingly, our main Gemini LLMs can be the state-of-the-art for math, for example, and for other kinds of scientific questions. So that convergence is, I don’t know, I guess it’s not something I really would have predicted at the outset. But it’s been kind of incredible to see.

And I guess baked into that is this concept of transfer, just the idea that when you train for a certain class of problems, let’s say you’re training for coding, that that actually can help your math reasoning and vice versa.

And that’s been really exciting to see… the multimodal capability also is an example of that. Like, can you actually get a transfer from being able to process images to actually being able to think through kind of geometric text problems too.”

Transfer learning is one reason convergence is happening. Transfer learning is where you train a model in one thing and it turns out that it has benefits in accomplishing tasks in something else that’s seemingly unrelated. So what’s happening now is that Google is finding that combining…