Long before OpenAI was wowing the world with ChatGPT, there was DeepMind.
Founded in 2010 in London, it built a team of researchers plucked from the UK’s top universities, who have since pioneered some of the world’s most high-profile breakthroughs in AI, including the protein structure prediction system AlphaFold in 2020 and the champion-beating board game player AlphaGo in 2016.
In 2014, it was scooped up by Google for $400m — one of the largest European tech acquisitions ever, at the time.
And it has, until recently, operated largely independently — enjoying access to the financial and hardware resources of its parent company, and the freedom to conduct blue sky research across generative models, reinforcement learning, robotics, safety and protein folding. In 2021, the company spun out Isomorphic Labs, an independent lab dedicated to applying protein folding techniques to drug discovery.
But now, as other Big Tech companies like Meta, Microsoft and Amazon are betting the house on AI, Google has realised the race is on. In April, it announced its internal AI lab, Google Brain, would merge with DeepMind. Its goal: to win the race to build the world’s first artificial general intelligence (AGI).
“Now, with the competition of OpenAI — and the realisation that AGI is going to be perhaps the world’s most profitable product ever — it’s not a sure slam dunk that it’s going to be Google that gets there,” one former DeepMind research engineer, who asked to be kept anonymous, tells Sifted.
The first result of this pooling of resources to stay ahead of the pack looks set to be Gemini — a large language model that’s powered by some of the problem-solving techniques that went into AlphaGo. It’s expected to be released in the coming months.
At the same time, Google is facing a new AI economy where the best AI researchers have more options than ever — to build their own thing or join one of several other well-funded AI labs with huge resources — and are, increasingly, choosing to explore them.
From research to profit
Google’s merger with DeepMind is a big transformation for a company that’s unlike any other in the field of AI and spent much of the 2010s hiring the brightest minds in machine learning from Europe’s top universities.
“What DeepMind did was it bought academia… It took so many of the best professors and graduates — where they all would have gone into academia otherwise — and it built this research hub,” says one former employee who worked with the ethics team. “The early premise was that you’d only be researching, it wouldn’t be about making money.”
In 2022, DeepMind was responsible for 12% of the most-cited AI research papers published globally, putting it ahead of Microsoft, Stanford and UC Berkeley, with only Meta and Google creating more research impact, according to research from AI search startup Zeta Alpha.
DeepMind generates revenue from selling services internally within the Alphabet Group, as well as through external contracts — such as a partnership with Britain’s National Health Service. It’s been profitable since 2020 but saw its margins squeezed in its latest company accounts.
This is where Gemini comes in. With OpenAI on track to make more than $1bn in revenue in 2023 from its LLMs, Google wants to release something bigger and better.
The fact that Gemini will be built using techniques from AlphaGo — the game-playing AI that beat a human Go champion in 2016 — suggests it could end up being more powerful, and useful, than OpenAI’s GPT-4. That’s because the model will combine the brute-force statistical prediction capabilities of LLMs, with the problem-solving capabilities of reinforcement learning (the machine learning approach used in AlphaGo).
Google also has a lot of computing power (known as “compute” in the AI industry) resources at hand. Access to specialist chips for AI training is a key factor in training powerful models, and semiconductor news site SemiAnalysis recently described Google as the “most compute-rich company in the world”.
The publication estimates that the company’s compute infrastructure will be five times more powerful than OpenAI’s by the end of this year, and 20 times heftier by the end of next year.
But while Google DeepMind flexes its language model and reinforcement learning chops to build Gemini, question marks hang over what the merger means for researchers who are focused on more foundational research that’s further from commercialisation.
Former employees tell Sifted that it’s still unclear how the push for productisation of DeepMind’s research will affect teams in the long run, but some would rather leave and start their own thing than wait and see.
“The move towards a more product focus meant morale was low among some people more on the frontier research side,” says Sid Jayakumar, founder of GenAI startup Finster AI, who spent seven years at DeepMind.
“We hired a lot of really good, really senior engineers, researchers who we basically asked to replicate an academic setting within industry, which was unique at the time and what was needed to build things like AlphaGo and AlphaFold.”
“It's no longer just an academic setting and rightfully so, in my view. But if you came from that [academic] perspective, you go, ‘This isn't great — what we were hired to do is no longer the priority,’” Jayakumar adds.
One former research scientist tells Sifted that one of the reasons he recently left DeepMind was that he wasn’t sure if the projects he was working on would survive the push to productise the lab’s research.
“We were working on quite fundamental stuff and it’s not always clear how that survives a change,” they tell Sifted. “My personal thoughts were, ‘What’s going to happen to these fundamental research programmes when we’re asked for more commercial impact?’”
Outflow of talent
For many AI engineers, DeepMind remains a killer place to have on the CV — but top researchers are leaving to found their own ventures, in apparently increasing numbers. Sixteen former DeepMinders launched their own ventures in the last twelve months, compared to seven in the previous year, according to Sifted analysis of LinkedIn.
Recent leavers include Cyprien de Masson d’Autume, cofounder of AI research and product company Reka AI, and Michael Johanson, cofounder of Artificial.Agency, a Canada-based AI startup that’s currently still in stealth mode. Both de Masson d’Autume and Johanson served as senior researchers at DeepMind.
The outflow of top researchers is a trend that mirrors Google’s own track record on AI talent retention, as many of the researchers behind its biggest breakthroughs have now left the company. In the past eight years, twenty top researchers who worked across milestone papers have moved on to found companies including Character.AI, Cohere and Adept, or to work at big AI labs like Meta, Hugging Face and Anthropic.
The company’s most high-profile loss is likely Arthur Mensch — cofounder of Mistral AI, the Paris-based AI startup that recently raised a massive €105m seed round and is seen as one of Europe’s brightest contenders to build LLMs like GPT-4.
He recently told Sifted he’d left DeepMind because the company was “not innovative enough” — with Mistral going on to release its own language model in just three months.
Another former DeepMind researcher-turned-founder also told Sifted that — given the rapid progress in AI — they left the company this year to launch a company that could be more agile.
“As a large listed organisation, I think there’s a lot of worry around releasing something to users that’s not perfect,” they tell Sifted. “You can iterate much faster and get feedback faster outside [of Google] and I think that was my main motivation.”
Those who haven’t left are getting constantly approached by recruiters.
“There’s lots of people who are biding their time working on ideas and intending to leave. You’ve got to understand — DeepMind researchers are being called up by recruiters who are saying 'I can easily get you a $700k or $800k salary,’” says one investor that’s close to the company.
But there are also plenty who want to stay, says former employee Jayakumar.
“Google DeepMind’s got the best AI team and has had consistently. Google has never moved faster and I don’t remember urgency being shown like it is now… I would actually be more worried if they were still focusing the most on that very open blue sky research and hadn't moved towards productionising.”
Sifted reached out to DeepMind asking for an interview and responses to the points made in this piece. The company declined an interview, but Dex Hunter-Torricke, head of communications at Google DeepMind, says that the work the company does “reaches billions of people through Google’s products and delivers industry-leading breakthroughs in science and research”.
“We’re proud of our world-class team and delighted to continue attracting the best talent,” he adds.