Demis Hassabis and the Lab That Never Sleeps
If you’re trying to keep up with who’s sprinting ahead in AI right now, you might want to take a glance towards Google DeepMind. The London-based lab has been on a dizzying streak, churning out more breakthroughs in weeks than most companies manage in a year.
Genie 3, Gemini 2.5 Pro, AlphaEarth, Aeneas—the list of shiny new models keeps growing. Throw in tools like Storybook, Kaggle Game Arena, Jules, AI Mode for Search in the UK and NotebookLM Video Overviews, and you begin to see why the company’s CEO, Demis Hassabis, recently joked, “Now you know why I don’t get much sleep, too busy pushing the frontier.”
Oh, and let’s not forget Gemma—the open model that’s already been downloaded over 200 million times. Google has even rolled out Gemma 3 270M, a compact version built for fine-tuning and running smoothly on devices. In other words, it’s small, fast and designed to be practical beyond the lab.
For most companies, that’s a year’s roadmap. For DeepMind, it was just a fortnight.
From Answers to Thinking
Most AI companies are busy training machines to spit out answers. DeepMind has a daring goal: building machines that can actually think.
In a recent podcast, Hassabis explained that DeepMind’s new line of work goes beyond matching patterns. Enter ‘Deep Think’, a system inspired by AlphaGo and AlphaZero. Instead of blurting out the first answer, the model pauses, reasons and plans. That makes a massive difference when the task is solving maths problems, writing code or tackling scientific puzzles.
As Hassabis put it, “Once you have thinking, you can do deep thinking or extremely deep thinking—and then have parallel planning.”
It sounds almost human. And that’s the point.
Building New Worlds
And then there’s Genie 3. Not your run-of-the-mill model, but one that understands how the physical world works. Solids, liquids, reflections—it can whip up entire environments from scratch.
DeepMind has already put it to work to create SIMA, an AI agent that can drop into video games and play like a human. That might sound like fun and games, but the real prize is endless training data for robotics and general intelligence.
In an exclusive interview with AIM, Deedy Das, a VC at Menlo Ventures, called Genie 3 “one of the coolest tech demos he’s ever seen in his life”. Now, coming from a guy who sees startups all day, that’s saying something.
The long-term play? An omni model—a single AI brain that does everything as well as today’s specialised models. Coding, science, gaming, reasoning—one stop shop.
Can Google Save Chrome?
DeepMind wasn’t the only one on a roll. Perplexity made its own splash with a jaw-dropping $34.5 billion bid for Google’s Chrome browser. Yes, you read that right.
The twist? Perplexity itself is valued at just $18 billion. Yet, the company told The Wall Street Journal that investors, including heavyweight investors, are ready to back the deal.
Meanwhile, while the browser wars play out, Parag Agrawal has quietly returned to the spotlight. Nearly three years after leaving Twitter, he’s launched a new AI startup, Parallel Web Systems, with $30 million in backing from Khosla Ventures, Index Ventures and First Round.
Parallel is building what could be AI’s first real browser, one where agents can fetch live data, check it, and even rate their own confidence. The kicker—its research APIs are already outpacing humans and GPT-5 on tough benchmarks.
Google, Don’t Ship Blog Posts
DeepMind’s pace is electric. But inside Google, the ride isn’t always smooth. Das argues that Google has a habit of announcing research without giving people much to play with. “Ship products, not words, if it’s not ready, ship nothing,” he said.
He’s equally blunt about Google’s slow decision-making and clunky user experiences. The Veo video model is his case in point—tucked away inside a product called Flow, buried under a pile of clicks.
However, when it comes to ambition, few labs come close. DeepMind now processes more than a quadrillion tokens every month. That’s a scale most rivals can only dream of.
The Benchmark Problem
When progress is this fast, it begs the question: how do you even keep score?
Most benchmarks test static tasks. DeepMind thinks that’s outdated. Its Game Arena pits AIs against evolving challenges, so they can’t just memorise answers. Hassabis also wants safety tests that catch behaviours like manipulation or deception. Because let’s face it, no one wants a clever liar for an AI.
So if Demis Hassabis looks a bit tired, it’s not caffeine withdrawal. It’s the weight of pushing AI into the unknown—one sleepless launch at a time.
If you found this newsletter helpful, share it with a friend, a colleague or that one person who still thinks DeepMind is falling behind.
If you haven’t subscribed to AIM Tv yet, now’s the perfect time. We break down the world of AI in real time, crisp, clear and always ahead of the curve.
While you are at it, here’s a quick look at some of the top stories of the week.
Now, let’s explore some exciting collaborations and exclusive insights from the AIM ecosystem, brought to you with a unique twist outside our standard editorial content.
India’s 100 Most Influential People in AI
AIM presents its annual compilation of the most influential leaders in AI in India for 2025, a carefully curated list of individuals who are helping shape the future of artificial intelligence in India. Check out the full list here.