Demis Stepped Down. Or Up?

Demis Hassabis is no longer running Google DeepMind day to day.

Depending on the headline, he has been promoted, demoted or pushed aside.

It’s … messy? And very important for the future of Google’s role in AI. So we’ll be having a look at what happened and what this means.

The facts:

We know Demis has stepped away from the CEO job to become Chair of Google DeepMind (the company he co-founded) and Chief Scientist of Alphabet. He is still CEO of Isomorphic Labs. He is still very much inside Google.

But he has given up operational control of the company he co-founded.

At the same time, Google's previous Chief Scientist Jeff Dean has left after 27 years, taking three of the people who helped build Google's AI advantage with him.

So nope, Google has not "lost Demis".

It has split the science from the shipping. That’s good - I think. BUT a worrying amount of talent is walking out of the building. Which is a problem for Google.

Who is Demis?

Bizarrely enough, the biggest impact Demis has had on my life is a computer game called Theme Park…

He programmed much of it as a teenager, including the AI controlling hundreds of little visitors wandering around your park.

As a child I discovered I could jack up the salt in the burgers, make everybody thirsty and then charge them an absolutely exorbitant amount for drinks.

Thank you Demis. Good times. 😆 

He went on to build Republic: The Revolution, an attempt to simulate an entire country and political crisis in 2003. It did not do particularly well, mainly because the idea was about 20 years ahead of the computers trying to run it.

This has always been his thing. Big, chunky problems. 

The mode through which Deepmind has pushed to solve these problems has been via games. Running AIs through millions of simulations of reality - learning how computers learn.

DeepMind started with Atari games, mastering games like Breakout through machine learning rather than being told how to win.

The big watershed moment was when Deepmingds AlphaGo beat Lee Sedol 4-1 in 2016 in the game of Go.

Previously Go was considered to be unsolvable by a computer because (unlike chess) a computer could not rely on searching through all the possible moves alone to find the best strategy. The number of potential moves in Go (10¹⁷⁰) dwarfs the estimated number of atoms in the universe(10⁸⁰) meaning that a computer would need something more than search- something closer to “intuition”.

There’s a fantastic documentary about it that I highly recommend watching. In particular for the Move 37 moment. It’s free on Youtube: AlphaGo documentary.

Move 37 looked so strange that people initially assumed it was a mistake. Instead AlphaGo did something new. And it won.

Then came AlphaFold, which took on a 50-year protein-folding problem and helped earn Demis and John Jumper half of the 2024 Nobel Prize in Chemistry.

If you have not watched AlphaGo or The Thinking Game, do. Both are free.

The Thinking Game in particular makes it very obvious that Demis has never really wanted to spend his life deciding where the Gemini button goes.

He wants AGI. World models. Drug discovery. Curing diseases. All of them.

You know. The small stuff, basically.

What actually changed

In Google's official announcement, Demis says the new role lets him focus on long-term AGI strategy, scientific breakthroughs and Isomorphic Labs' work on drug discovery. His goal is to solve disease.

In Demi’s place Koray Kavukcuoglu becomes SVP of Google DeepMind. His remit includes model development, frontier research, the Gemini app and Google's developer teams. Basically…the PRODUCT that we consumers get.

And then there is Jeff Dean.

Jeff Dean has been the Chief Scientist at Alphabet (Google’s holding comapny) for years. Dean has left with Sanjay Ghemawat, Oriol Vinyals and Quoc Le to start Discovery Loop, a company trying to automate machine learning, science and engineering.

Importantly: Google is a founding investor and cloud partner, so this is more negotiated spinout than messy divorce.

Still. Four people who spent between 14 and 30 years building Google's products, infrastructure and AI models decided their next big breakthrough would be easier outside of Google proper.

For my money here’s a summary of what seems to have happened:

• Jeff Dean told Google he was off on his own. Which would leave the Chief Scientist position open.

• Demis Hassabis gets offered that position. It’s an amazing fit for him as he’s very much focused on the BIG questions and advances in AI rather than the day to day shipping of models to compete with Anthropic and OpenAI.

• Koray Kavukcuoglu (previously CTO at Deepmind) gets bumped up to run Deepmind day to day, with a focus on product.

Gemini needs speed. Demis did not.

Google DeepMind has been doing two jobs which run at completely different speeds.

One team needs to get Gemini models, apps and developer tools into the market. Fast. OpenAI and Anthropic are already winning a lot of the day-to-day usage, especially coding. Google are (no offence Googlers!) currently being trounced. There’s no other way to spin it.

The other half wants to spend years pursuing research which might not make a penny for a long time...or ever. That’s just not the goal.

Those jobs rub against each other.

Google needs Gemini to make money. Demis wants the freedom to attack problems where the commercial payoff is either distant or unknowable.

Jeff Dean's exit also left the Alphabet Chief Scientist role empty. Moving Demis into it, then putting Koray in charge of DeepMind's operations, does kinda make sense. My hope is that Koray can make Google faster whilst Demis gets out of some meetings and can focus onto giant science bets.

The pessimistic view

The bearish case is the people leaving. Google are bleeding talent. Google keeps losing people who created the advantage everybody says will help it win.

Jeff Dean and Sanjay Ghemawat helped build Google's foundational infrastructure. Oriol Vinyals and Quoc Le are huge names in modern machine learning. Noam Shazeer had already left. John Jumper, who shared the Nobel with Demis, has gone too.

Sure. Their code, patents and teams do not vanish when they leave. Google will have all that locked down. But their brains are gone..

Which weird idea deserves three years and a mountain of compute? Which result is a dead end? Which tiny research thread could become the next AlphaFold? We don’t know. But losing the big thinkers makes these calls much more difficult.

Axios reports Google tried to keep the Discovery Loop founders. They went anyway. Why did four people with access to Google's money, chips and infrastructure decide they needed a new company to automate scientific discovery properly?

I don't know. But it is a hella fair question. It’s THE question here.

Google can stay in the race longer than anyone

All of this said I still believe Google are in a strong position.

I have been banging this drum for months: focusing only on which model topped this week's benchmark is midwit behaviour. It’s noise.

Google owns the stack.

It designs TPUs. It owns the data centres and Google Cloud. It has DeepMind, Gemini and Gemma. Then it has Search, Chrome, Android, YouTube, Gmail and Workspace to put the models in front of billions of people.

It also has a cash engine. Alphabet reported $119.8 billion in Q2 revenue, with Google Cloud growing 82% to $24.8 billion.

The Gemini app has 950 million monthly users. Google's first-party model APIs are processing roughly 22 billion tokens per minute.

This is why I personally came away from Google I/O saying its releases were a bit meh but its position was not.

OpenAI and Anthropic need to keep raising and spending gigantic piles of money. Google can fund the race from the businesses it already owns.

If we are heading towards a financial reset (not unlikely…) then Anthropic and OpenAI are in perilous positions. They are at the cliff edge. Without more investment they die - and fast.

Google though? They can weather the storm and come out the other side dominant. Indeed they’ve done this before - sailing through past financial crashes that demolished their competitors, allowing them to each time emerge stronger.

Google has the chips, infrastructure, research, distribution and cash.

Now it needs to keep the people who built all of that...and ideally ship something coherent…

This remains Google's race to lose.

Kyle