CEO of Google DeepMind: AGI is still 5 to 10 years away, and AI development opportunities and risks go hand in hand

In the finale of the “AI+SF Summit” held by Axios in San Francisco, Demis Hassabis, CEO of Google DeepMind, outlined DeepMind’s research directions, technological advancements such as multimodal and world models, and discussed the development of AI agents and related risks. He also assessed the AI competition between the United States and China, and gave a rare estimate of the AGI timeline, believing that there are about 5 to 10 years left before “AI systems with human cognitive capabilities”.

With the blessing of the Nobel aura, scientist thinking dominates DeepMind

At the beginning of the event, host Mike Allen introduced Hassabis as a 5-year-old chess prodigy and a 48-year-old Nobel Prize winner. Hassabis admits that winning the award still feels super unreal, but the actual impact is obvious.

Because when he talks to government officials or cross-border decision-makers who are not familiar with AI, the “Nobel Prize” is like a key that can quickly open any door, making them more willing to listen to him talk about issues such as AI safety and responsible use, and he plans to make more active use of this title in the future.

When it comes to his daily work and management style, Hassabis emphasizes that he “always comes first as a scientist and a CEO comes second.” In his eyes, the scientific method is one of the most important inventions of mankind, and he directly applies the process of “formulating hypotheses, designing experiments, and updating opinions based on results” directly to product development and organizational management.

DeepMind’s advantages come from three levels at the same time, namely “world-class research, world-class engineering capabilities, and world-class computing infrastructure”. He believes that only when these three levels are carried out simultaneously can DeepMind be qualified to stand at the forefront of AI development.

Layout for the next 12 months: multimodal evolution, world model, and agents

Talking about the specific progress of AI in the next 12 months, Hassabis pointed out that Gemini was designed as a multimodal model from the beginning, which can process text, images, video and audio at the same time. For example, his latest image model, “Nano Banana Pro,” can produce very accurate infographics, indicating that the model’s visual understanding capabilities are rapidly improving.

The second focus is on the world model (World Models). Genie 3, developed by DeepMind, can generate interactive videos that allow users to not only watch the video but also walk into the screen as if they were entering a game, maintaining consistency and coherence in the world for about a minute. This type of model is seen as a key step in AI’s understanding of real-world appearances and rules.

The third is AI agents. Hassabis admitted that current AI agents cannot be relieved to throw a whole package of tasks directly to it, ensuring that it is done well from scratch. But he expects that in a year, the trust of AI agents will definitely increase. Google’s goal is to make Gemini a “universal assistant” that not only exists on mobile phones and computers, but can be by the user’s side at any time through wearable devices such as glasses, becoming a regular assistant for daily life and work.

( test: Gemini 3 Nano Banana Pro automatically generates humorous cartoons after thinking, turning Trump back into a small fresh meat )

The future holds promise for cosmic exploration, but security risks are as critical as video understanding

Speaking of the best scenarios that AI can bring, Hassabis hypothesizes that AI can help humanity break through several key bottlenecks, such as nuclear fusion or new batteries, new breakthroughs in materials science and semiconductors, and solutions to major diseases, and human society will have the opportunity to move forward into space exploration with more abundant resources.

But he also pointed out the worst-case scenario, which is divided into several levels:

Malicious actors use AI to design or enhance pathogens.

AI accelerates cyberattacks by foreign forces on critical infrastructure such as energy and water resources, and such things are likely to be happening, but the AI used is not yet advanced.

Highly autonomous AI agents deviate from their original instructions and human expectations, so they must invest considerable resources and attention to prevent them.

In terms of ability, he believes that the underestimated piece of the outside world is AI’s deep understanding of video. Hassabis shared that he once asked Gemini to analyze the scene, and the model not only understands the picture, but also gives a very deep interpretation of symbols and emotions, rather than just describing superficial actions.

He also mentioned that Gemini Live allows you to get instant repair assistance by pointing your phone camera at mechanical equipment, but he believes that the truly ideal vehicle will be glasses, because your hands must be empty during on-site operations to work and interact with AI at the same time.

There are only a few months left in the gap between the United States and China, and the AGI is still one or two miles away

When it comes to international competition, Hassabis believes that the United States and the West are still leading China as a whole in terms of model capabilities and innovation, but China’s latest batch of models, such as DeepSeek, are already very strong, and most of them are catching up quickly on the basis of existing technologies. He judged that in the past, the United States and the West may have led in years, but now there are only a few months left ahead of China.

Hassabis defines AGI quite clearly, namely:

“You must have all the major cognitive abilities of human beings, including long-term planning, long-term memory, continuous learning, real reasoning and creativity, etc.”

He pointed out that although LLMs at this stage already have the ability to approach top doctors in some fields, they will still make mistakes in many scenarios, and there is still a gap between true AGI and is estimated to take 5 to 10 years. Hassabis added that even if the scale of existing LLMs is pushed to the limit, it is still not enough to cross the AGI threshold, and the AI field may need one or two major technological breakthroughs that can greatly improve capabilities, such as transformers, before there is a chance to truly achieve AGI.

(IBM CEO: The AI industry is a gamble that is “difficult to recover”, with LLMs only having a 1% chance of successfully creating AGI )

The post Google DeepMind CEO: AGI is 5 to 10 years away, AI development opportunities and risks go hand in hand appeared first on Chain News ABMedia.