Google’s Sergey Brin on the Limits of AGI

Google co-founder Sergey Brin has said that artificial intelligence appears to be progressing towards Artificial General Intelligence (AGI), but he is uncertain about what the next stage beyond that might look like.

In a recent interview at AGI House, Brin discussed how Google’s Gemini models are evolving beyond their original design. He explained that modern systems are no longer limited to isolated tasks, but are instead beginning to integrate knowledge across languages, formats, and different types of data. While he believes AGI is becoming more achievable, he admitted that the future beyond it is difficult to predict.

What AGI actually means

AGI, or Artificial General Intelligence, refers to a level of AI that can understand, learn, and apply knowledge across a wide range of tasks in a way that resembles human reasoning.

Unlike today’s AI systems, which are highly capable but still task-specific, AGI would be able to transfer knowledge between domains and solve unfamiliar problems more independently. Current models can write code, generate text, analyse images and assist with complex queries, but they still lack a true general understanding of the world.

Major AI developers such as Google DeepMind, OpenAI, and Anthropic are all working towards AGI, although their motivations differ. OpenAI tends to focus on economic productivity, DeepMind on scientific discovery, and Anthropic on building safer and more aligned systems.

The shift towards convergence in AI models

Brin highlighted a key shift in AI development: the move from specialised systems towards unified models capable of handling multiple types of tasks.

In earlier stages of AI, different models were trained for specific functions such as language processing, mathematics, or scientific reasoning. However, with systems like Gemini, Google is increasingly seeing strong performance across multiple areas within a single model family.

This trend, known as convergence, was not something Brin said he originally anticipated when Google first began developing modern AI systems.

He explained that improvements in one area often translate into gains in others, a process linked to transfer learning. This means that training a model for one task, such as coding, can unexpectedly improve its performance in areas like mathematics or reasoning. The same applies across modalities, including text, images and other forms of data.

Transformers and their unexpected flexibility

When asked whether transformer models could play a role in achieving AGI, Brin pointed out how adaptable they have become.

Transformers, the architecture behind most modern large language models, were originally designed for text-based tasks. However, they are now used across a much broader range of applications, including image and video processing.

Brin noted that while the underlying concept has remained influential, it has also evolved significantly over time, including the introduction of variations such as mixture-of-experts models, which route tasks to specialised internal components for better efficiency.

In his view, it is possible that something closely related to transformers could ultimately support AGI, although he acknowledged that the technology continues to change rapidly.

World models and Gemini’s direction

Another area of focus in the discussion was the idea of world models, which are AI systems designed to simulate and understand aspects of the physical world. These models help predict outcomes, reason about actions, and support more advanced decision-making.

Brin connected this idea to Google’s Gemini Omni model, which was introduced at Google I/O and is designed to handle multiple input and output formats, including generative media such as video.

He explained that achieving human-level capability would require AI to better understand and interact with the physical world, particularly in areas such as robotics and planning. In this sense, world models are seen as an important step towards more general intelligence.

He also described how these capabilities are increasingly being built into Gemini itself, rather than treated as separate systems. This blending of reasoning, perception, and generation reflects the broader theme of convergence running through Google’s AI development.

Different interpretations of AGI

Brin also noted that there is no single agreed definition of AGI. Some view it as systems that can perform any task a human can do, while others interpret it as AI capable of improving itself.

He suggested that understanding and interacting with the physical world will be essential for more advanced definitions of AGI, particularly if robotics and real-world interaction become central to its development.

What comes after AGI remains unknown

When asked what might follow AGI, Brin admitted he does not have an answer. He compared the current AI wave to previous technological shifts such as the rise of the internet and mobile computing, both of which led to major industry changes.

However, he suggested that the stage beyond AGI is still unclear and may only become apparent once AGI itself has been achieved. In his view, identifying what comes next could itself represent a major opportunity for future innovation.

The bigger picture

Overall, Brin’s perspective highlights a clear direction in AI development: increasing convergence across models, tasks, and data types.

Key ideas shaping this progress include:

The merging of specialised models into general-purpose systems
Transfer learning improving performance across unrelated tasks
Continued evolution of transformer-based architectures
The growing importance of world models in understanding reality
The gradual progression towards AGI rather than a sudden breakthrough

Google’s Gemini system, as described by Brin, reflects this shift. It combines reasoning, multimodal understanding, and generative capabilities within a single framework, pointing towards a future where AI systems are increasingly unified.