DeepMind's Demis Hassabis on the 5-10 Year Realistic Scenario for AGI
At the Axios 'AI+SF Summit' event, Google DeepMind CEO Demis Hassabis spent about 30 minutes discussing the current state of AI and the next decade.
From the changes since winning the Nobel Prize to multimodal AI 'Gemini,' agents, and AGI (Artificial General Intelligence)—I will break down these technical topics in an easy-to-understand way.
1. The asset of 'having a voice' brought by the Nobel Prize
Hassabis says that even now, more than 400 days after winning the Nobel Prize, it still feels 'surreal,' but the changes are clear.
When talking to the general public or high-ranking government officials—people who are not deeply familiar with the inner workings of AI—the Nobel Prize acts as a powerful shortcut.
Because they understand instantly that 'this person is an expert,' the platform for speaking on important topics has expanded significantly.
There are two main things he wants to communicate using that platform moving forward.
Long-term AGI safety
The 'responsible use' of AI that is already underway
In other words, he is emphasizing not just the risks of the distant future, but also 'how to design the AI applications happening right in front of us' with equal importance.
2. Thinking like a scientist is DeepMind's competitive advantage
Hassabis defines himself as 'a scientist first and foremost.' What that means is that he is someone who uses the scientific method in every situation.
'The scientific method might be the most important idea humanity has ever produced.'
He goes as far as to say that, and explains that he applies the process of experimentation, verification, and updating hypotheses to
research,
business,
and even daily life.
He explains that he is applying it even to daily life.
He cited the following 'trinity' as DeepMind's strength.
World-class research
World-class engineering
World-class infrastructure
He explains that connecting these through scientific methods and ensuring "accuracy and rigor" is what gives them an advantage in the AI race, which he describes as "probably the most intense competition in tech history."
3. AI Evolution Over the Next 12 Months: Multimodal, World Models, and Agents
3-1. The Evolution of Multimodal Gemini
When asked what will happen in the next year, the first thing he mentioned was the convergence of multimodality.
-
Gemini was designed to be multimodal from the start
It can handle "text, images, audio, and video" as input
And it is becoming possible to generate those modalities as output as well
The latest image model (referred to as "NO Banana Pro" during the interview) is showing surprising capabilities in visual understanding, such as generating high-precision infographics.
In particular, he suggests that as the integration of video understanding and LLMs progresses, entirely new types of assistants and tools will emerge from the combination of "language + video + sound."
3-2. World Model "Genie 3" and Interactive Simulation
Another area he is personally focusing on is world models.
An interactive video model called "Genie 3"
Allows you to "walk around inside the generated video space like a game or simulation"
The consistency of the world is maintained over a time scale of about one minute
This goes beyond simple video generation and means that "AI is beginning to hold consistent rules of the world internally." The range of applications is extremely broad, including games, robotics, and simulation-based planning.
3-3. The Horizon of Agents and the "Universal Assistant"
Furthermore, Hassabis speaks about "agents" as follows:
Current AI agents are not reliable enough to be fully entrusted with entire tasks
However, he predicts that "a year from now, we will be much closer"
-
What DeepMind/Google is aiming for is a "Universal Assistant."
A presence that is always by your side, not just on PCs and smartphones, but also on wearable devices like glasses
It will not only increase work productivity but also handle recommendations for all aspects of life, including movies, books, and activities
His view is that while it won't be "full automation" within a one-year span, agents capable of highly advanced delegation will become a reality.
4. The best case is 'radical abundance,' but P(doom) is not zero
4-1. The 'post-scarcity' world brought about by AI
Hassabis's ideal is a kind of 'radical abundance.'.
Breakthroughs in clean and renewable energy (fusion, next-generation batteries, solar cell materials, etc.)
Significant advancements in science and technology, including semiconductors, materials science, and drug discovery
Many diseases being overcome, bringing humanity closer to a 'post-scarcity' world
Beyond that, he even discusses a sci-fi vision where 'humanity expands into space and spreads consciousness throughout the galaxy.'
At the same time, he reveals that he is seriously concerned about the philosophical question of 'what will happen to the role and purpose of humans' in such a utopia.
4-2. Since P(doom) is not zero, we must seriously reduce it
On the other hand, he is quite candid about the risks.
Malicious actors using AI to design pathogens
Cyberattacks by foreign powers on energy and water infrastructure
Highly autonomous agents 'deviating' from the designer's intent
Regarding these 'existential risks,' while he does not state the probability as a number, he emphasizes,
'P(doom) (probability of doom) is not zero. That is precisely why we should devote appropriate resources and attention to it.'
he emphasizes.
At the same time, amidst corporate competition, he also holds hope for the possibility that 'more responsible players will be rewarded by the market'.
When large companies introduce AI agents,
data handling
guarantees of behavior
The view is that by imposing strict requirements, a mechanism will work to ensure that highly safe vendors are selected.
5. '5-10 years' to AGI: Scaling + 2 breakthroughs
Hassabis also has a clear stance on AGI.
He explicitly states that 'current models are not AGI'
However, he predicts that 'it could be reached within 5-10 years'
The AGI he defines is not just about having high test scores, but sets a very high bar, such as:
possessing all cognitive abilities, including creativity and the ability to invent
demonstrating consistent, 'smooth' intelligence rather than being uneven across fields
having continuous learning, online learning, and long-term planning capabilities
as well as other requirements.
Regarding current LLMs, he describes them as:
showing 'PhD-level' or 'Olympic gold medal-class' power in some fields
while being surprisingly fragile in others
what he calls 'jagged intelligence'
Furthermore, he states that achieving AGI will require:
pushing the scaling of current architectures to the limit
and one or two more major breakthroughs on the scale of 'Transformer' or 'AlphaGo'
to be necessary.
6. Bubbles, the talent war, and human adaptability
Finally, he also touches on related hot topics.
Regarding investment in AI, while noting that "there are clearly some bubble-like areas, such as certain seed rounds,"
he views it in the long term as "the most transformative technology in human history," and as an investment that is justified in the aggregate.
As for the competition for talent, while companies making extreme offers (Meta, in the context of the conversation) are being talked about,
DeepMind prioritizes "mission-driven talent"
and appeals to them as a place where they can tackle cutting-edge challenges across the full stack (research, product, and infrastructure).
Also, to the question of whether humans can keep up with this change,
"Our brains originally evolved for hunter-gatherer life, yet we have adapted to modern civilization,"
he points out, showing strong confidence in human adaptability and ingenuity.
In the future, "human-side upgrades," such as technology that connects the brain to computers, may also become an option.
7. Games and Sports: Understanding the World as a Simulation
Hassabis is a former chess prodigy and also a game developer.
To him, games are a "microcosm" of the world.
In the real world, there are at most a few dozen "truly important decisions" in a lifetime.
On the other hand, in a game, you can practice and repeat that decision-making process thousands of times.
That is why he believes games are the "best training equipment for honing decision-making and planning skills."
Regarding sports as well, he suggests that AI will be a tool that analyzes vast amounts of data to further push the "ultimate optimization" of elite sports,
such as optimizing set plays and player positioning.
Conclusion: Creating the Future of AI While "Experimenting" with the Scientific Method
What was consistent throughout this interview was Hassabis's "attitude as a scientist."
Which approach is closer to AGI is determined by experimental results, not by belief.
If it becomes clear that LLM scaling is promising, he will shift significant resources there.
At the same time, under the premise that P(doom) ≠ 0, he is also seriously committed to safety research and governance.
AI is a technology that can bring about radical abundance, serious risks, or both.
What emerges from Hassabis's words is a quiet resolve: "we have no choice but to control its destination as a society while making full use of the scientific method."