Why Google DeepMind CEO Demis Hassabis Can’t Sleep At Night

Google DeepMind may well be the most underrated AI lab today. In the past few weeks alone, the company has shipped more real-world breakthroughs than most AI companies manage in an entire year.

In just two weeks, DeepMind and Google have rolled out a wave of big AI launches, including Genie 3, Gemini 2.5 Pro, AlphaEarth and Aeneas, alongside tools such as Storybook, Kaggle Game Arena, Jules, AI Mode for Search in the UK, and NotebookLM Video Overviews. Their open model, Gemma, has already been downloaded over 200 million times.

“Now you know why I don’t get much sleep,  too busy pushing the frontier,” wrote Google DeepMind chief Demis Hassabis in a post on X.

“DeepMind is my favourite lab. I love how they are not limited to research just about LLMs, but healthcare, Physics, robots, vision, world models, Biology, etc,” said an AI student on X.

Building Thinking Models 

In a recent podcast with Logan Kilpatrick, Hassabis said that the pace of progress inside the company has reached a level where the company is pretty much releasing something every day. “It’s hard to keep up, even internally.”

“The biggest achievement in the past two years is that Google and DeepMind now trust each other,” said Shixiang Shane Gu, senior staff research scientist at Google DeepMind, in a post on X. 

He further said that Google has committed to using AI to reimagine its core businesses and internal tools, and that DeepMind has proven it can deliver industrial-grade models beyond research.

As AI competition intensifies globally, DeepMind has positioned itself as both a research leader and a contributor to Google’s commercial AI offerings. Hassabis discussed the lab’s recent announcements, including Deep Think, the International Mathematical Olympiad (IMO) gold medal model and Genie 3. 

Hassabis further said that Deep Think draws on DeepMind’s long history with agent-based systems such as AlphaGo and AlphaZero. “We’ve always worked on systems that can complete a whole task,” he said. These systems now combine multimodal inputs with reasoning and planning capabilities.

“Once you have ‘thinking’, you can do ‘deep thinking’ or ‘extremely deep thinking’ and then have ‘parallel planning’,” he explained during the discussion. This allows AI to plan before producing output, especially for complex domains. “For things like Maths, coding, scientific problems and also gaming, you’re going to need to process and plan and basically do this thinking and not just output the first thing that the model comes up with.”

Genie 3 and World Model 

Genie 3, DeepMind’s demonstration of a world model, understands the structure and behaviour of the physical world, including solids, liquids and reflections. Hassabis said generating realistic worlds is one way to demonstrate the depth of understanding such a model can achieve.

Hassabis said the same technology is being applied in Scalable Instructable Multiworld Agent (SIMA), a simulated agent that can take control of and play existing computer games. This, he explained, can produce unlimited training data for robotics and artificial general intelligence research. He also said that simulation-based training could have potential applications in other domains where physical testing is expensive or slow.

Deedy Das of Menlo Ventures told AIM that DeepMind’s recent work on Genie 3 and generative environments was one of the most incredible tech demos he had ever seen in his life.

Regarding the model’s applications in robotics, he added, “It’s still too early to know whether that will work. Startups continue to operate across different layers of the robotics stack—from training environments and data labelling to software, intelligence, and hardware—with several attempting to tackle multiple parts of it.”

According to Hassabis, work is also moving towards an omni model capable of performing a wide range of tasks. “We’re starting to see convergence of those models together into what we call an omni model, which can do everything,” he said, adding that the goal is for an AGI system to handle all tasks at the same quality as today’s specialised models, but within a single, unified model.

Setting New Standards for AI Evaluation

On evaluation, Hassabis said there is a need for new benchmarks that go beyond current tests, which mostly measure performance on specific, static tasks. DeepMind’s Game Arena is one such attempt, with games that evolve in complexity and are unique to each match to prevent overfitting.

“Evaluation is an unsolved problem,” Hassabis said. “You need to be able to measure reasoning, planning, memory and physical intelligence, not just pattern matching.”

He also discussed the importance of safety benchmarks. These would test for behaviours such as deception or manipulation, ensuring systems remain reliable when deployed in real-world environments.

Anticipating Technology Shifts

On the topic of integrating tools into AI systems, Hassabis said the decision depends on impact. “It’s very much an empirical question. Does adding that capability help the other capabilities? If it does, then do it. If it harms the other general capabilities, then maybe consider using it as a tool.”

Hassabis said that product managers and designers must think ahead about where technology will be when a product launches. “You’ve got to be really close and understand the technology world to intercept where that technology will be in a year’s time,” he said.

He noted that DeepMind has crossed the milestone of processing more than a quadrillion tokens each month. While not a direct measure of capability, it signals the scale at which the company is now operating.

What Google Could Do Better

Das said that he genuinely believes that Google DeepMind is one of the “least evil” and most engineering-centric companies in the world. “They push the frontier of innovation, have tremendous results on difficult problems with no clear revenue, they make AI models cheaper, they’ve built the best video model by far.”

However, he sees gaps in how Google ships and shares innovation. “They launch blog posts often without a product users can play with, and I feel strongly this should just be reprimanded by the entire tech community,” he said. “We’re in 2025. Tech companies should ship products, not words. If it’s not ready, ship nothing.”

He also pointed to speed, decision-making and ecosystem engagement. “They move slower than I’d like,” Das said. “They hesitate to make bold decisions while trying to appeal to everyone and maintain their brand image…They are less entrenched in the external ecosystem, and I wish they would partner more closely with startups and other companies where they can do experimental stuff.”

Sometimes, he noted, the culture can be insular. “Their researchers are more Google-minded than they are tech-minded because of how high the metaphorical walls around the city of Google are.”

Das suggested that Google release a larger open source model than Gemma. “People would love it,” he said. He also believes Google’s marketing and product packaging could be improved.

Citing Google’s Veo 3  video model launch as an example, he described the friction in accessing it. “It was packaged in an unrelated product called Flow and you had to click 10 things to get to it,” Das said. 

The post Why Google DeepMind CEO Demis Hassabis Can’t Sleep At Night appeared first on Analytics India Magazine.

Generated by RSStT. The copyright belongs to the original author.

Source