Google DeepMind CEO Demis Hassabis talks about ``Why did Google and DeepMind integrate?''



Google had an AI development department called ' Google Brain .' At the same time, a company that independently conducted AI research and development under the umbrella of Google's parent company Alphabet, and developed the strongest Go AI '

AlphaGo Zero ' that even beat professional Go players and the protein three-dimensional structure prediction AI ' AlphaFold ', announced the development of ' DeepMind'. ' was also present. In April 2023, Google announced that it had acquired DeepMind and merged it with Google Brain, creating a single AI research division called ' Google DeepMind .' Demis Hassabis, a co-founder of DeepMind and an AI researcher, has been appointed as the CEO of Google DeepMind. In an interview conducted by IT news site The Verge, Mr. Hassabis clarifies the reasons and circumstances behind the integration of Google and DeepMind.

Google DeepMind CEO Demis Hassabis on ChatGPT, AI, LLMs, and more - The Verge
https://www.theverge.com/23778745/demis-hassabis-google-deepmind-ai-alphafold-risks



The Verge：
Google DeepMind is a new division of Google that is comprised of two existing divisions of Google. Google Brain is Google's AI team, and the company you founded was DeepMind. DeepMind was acquired by Alphabet in 2014, but is outside of Google and until recently operated as a separate company within the holding company Alphabet structure. Start at the very beginning. Why were DeepMind and Google Brain separate in the first place?

Demis Hassabis CEO (hereinafter referred to as Hassabis CEO):
We founded DeepMind in 2010, which is a long time ago in the history of AI. So it's like prehistoric times. Myself and my co-founders were from AI academia and had seen what was going on there and were aware of something called deep learning that had just been invented. We were big proponents of reinforcement learning. GPUs and other hardware are coming online, and if we focus on general learning systems and incorporate some ideas from neuroscience and how the brain works, we can make a lot of big advances. I understand.



So we put all this together in 2010. We had this assumption that we would make rapid progress. That's our earliest story. Then in 2014, we realized we needed more computing, so we decided to partner with Google. Google apparently owned the most computing resources in the world. It was clear to us that Google was a home base where we could focus on advancing our research as quickly as possible.

The Verge：
DeepMind was acquired by Google, and then Google changed direction midway through the process. DeepMind's parent company changed to Alphabet, and Google became a division of Alphabet. Alphabet has other divisions, but DeepMind was outside of them. That's exactly what I want to focus on here first. This is because much of what Google was doing with Google Brain was research on large-scale language models (LLM). If you remember back in 2017, Google was showing off LLM at Google I/O, but DeepMind was working on a completely different kind of work that was done completely outside of Google: 'winning at Go' and 'protein folding.' It focused on AI research. Why would it be done outside of Google?

Hassabis CEO:
Part of the deal at the time of the acquisition was to promote research into artificial general intelligence (AGI). Another personal passion of mine is using AI to accelerate scientific discovery, like AlphaFold. And since DeepMind's inception, and even before DeepMind was founded, games have been the perfect testing ground for developing AI algorithms efficiently and quickly, can generate a lot of data, and can be used for any purpose. The functions to do were also very clear.

Of course, we've also done a lot of research on deep learning and neural networks. And I think our specialty was combining those with reinforcement learning to allow these systems to actively problem solve, plan, and do things like win games. The difference between DeepMind and Google Brain is that we've always had the mandate to drive the research agenda and advance cutting-edge science. Google's internal AI teams, like Google Brain, had a slightly different mandate and were clearly focused on commercialization like other parts of Google, injecting great AI technology into Google. There was also an applied department that introduced DeepMind's technology to Google's products. However, DeepMind and Google Brain have completely different cultures and have completely different roles to play.


by George Gillams

The Verge：
From an outsider's perspective, everyone panicked when OpenAI released ChatGPT, the world changed when Microsoft released Bing based on ChatGPT, and Google responded by integrating DeepMind and Google Brain. It looks like this is the flow. Was this what it looked like from the inside?

Hassabis CEO:
I think that flow is correct, but it is not a direct result, and in a sense it is more indirect. Google and Alphabet have always operated like this. I think that's what we've always done since Larry Page and Sergey Brin founded Google. It worked so well that we organically created something incredible that allowed us to become the great company we are today. That's why DeepMind chose Google as a partner. They really understand what basic research is and what ambitious research is, and I felt that they would be able to push forward our AI research, even though it was super ambitious. The result is what you see, right?

We achieved all the usual metrics we use to deliver truly great cutting-edge research, including AlphaGo and AlphaFold, as well as more than 20 natural and scientific papers. But what ChatGPT and LLM, and the public response to them, have confirmed is that AI has entered a new era. That's because startups like us, Anthropic, and OpenAI were all developing LLMs with similar capabilities.

Since we understood the LLM technology, we were surprised not so much by the technology itself, but by the enthusiasm and buzz among the general public about the technology. LLM has moved beyond the research stage to a level of maturity and sophistication that has led to incredible next-generation products and experiences, such as AlphaFold being a breakthrough technology that helps biologists. has been reached. For me, this marks a new phase in which AI can help people in their daily lives, not just for fun or for fun, but to really solve important real-world problems. It means that it has become. Now is the time to streamline and focus your AI efforts. The logical conclusion was the merger of DeepMind and Google Brain.

The Verge:
CEO Sundar Pichai comes to you and says, 'Okay, I'm the CEO of Alphabet and I'm the CEO of Google. I just made this call. I'm going to bring DeepMind into Google and integrate it with Google Brain. How did you react when he said, 'We intend to appoint Mr. Hassabis as CEO?'

Hassabis CEO:
There was no such conversation. I think there was a discussion between the various relevant group leaders and CEO Pichai about the tipping point we're seeing, the maturity of the system, what could be commercialized, and how do we go about commercializing it? . We thought about how exciting it would be to improve the experience for billions of users, and what it would take overall. There are many factors to consider, such as a change in focus, a change in approach to research, and the combination of required resources such as computing resources, which we all discuss as a leadership group and whose conclusions inform the merger. This led to actions including: Of course, it also includes plans and content for the next few years.



The Verge:
Do you think there is a difference between Google's CEO and Alphabet's CEO?

Hassabis CEO:
It's still early days, but I think it's pretty similar between Google and Alphabet. That's because, although DeepMind was part of Alphabet, it was very closely involved and collaborated with Google's product teams and groups, known as 'Alphe Bet.' DeepMind has an application team that works with Google's product teams to translate our research findings into product features. That's why we've had hundreds of successful launches behind the scenes in recent years. In fact, DeepMind technology is embedded in many of the services, devices, and systems you use every day at Google. Of course, DeepMind is famous because of AlphaFold and AlphaGo, but behind the scenes there was a lot of work going on that affected every part of Google.

While other Alphabet companies aimed to create businesses independent of Google, we were different. Even though we are an independent company from Google, that has never been our goal or mission. And now it's more tightly integrated within Google. We're able to do deeper, more exciting, and more ambitious things in close collaboration with other product teams than we could when we were separate from Google. However, we still have the freedom to choose the processes and systems that optimize our mission to create the world's most capable and versatile AI.

The Verge：
There are also rumors that different cultures are colliding with Google and DeepMind. As CEO, how do you organize Google DeepMind and manage its cultural integration?

Hassabis CEO:
In fact, it turns out that Google and DeepMind have much more similar cultures than people would say. It's obviously a great group of people and talent, and there's a lot of respect for both groups. And integrating the two was actually surprisingly smooth and fun. Because we have world-class research groups, two of the best AI research organizations in the world, great talent from both sides talking about AI. When we were considering the Google DeepMind merger, we read a document that summarized each group's top 10 breakthroughs. To summarize, 80% to 90% of the breakthroughs that underpin the modern AI industry, from deep reinforcement learning to Transformers , have occurred in the last 10 years.

Of course, we all know each other very well. However, I think it's a matter of concentration and a little coordination across both groups. More specifically, what do we focus on? What does it mean to have two separate teams working together? Can we eliminate overlap in effort between the two teams? Honestly, it's pretty obvious, but now that we're moving into the AI engineering stage, it's important to move to this new stage, and it requires tremendous resources. Even a company as big as Google has to choose its bets carefully, know where to focus its efforts, and focus on those areas to achieve great results. So I think it's part of the natural flow of where we are on the AI journey.

The Verge：
'What do we focus on? What does it mean to have two separate teams working together? Can we eliminate the overlap in effort between the two teams?' Have you found the answer? What do you think will be the system to proceed with this?

Hassabis CEO:
The team structure is still in its infancy. It's only been a few months. We wanted to make sure nothing was broken and everything was working properly. Both teams are very productive, doing some really great research, and also working on very important products that are underway. All of this needs to continue.

The Verge:
So are you thinking of it as two teams or are you trying to make it one team?

Hassabis CEO:
No, it's certainly one unified team. I like to call it a 'super unit' and I'm really excited about it. But clearly we are creating new cultures and forming new groups, including organizational structures, as we combine them. Bringing together two large research groups like this is complex. But I think by the end of summer 2023, we'll be one, and it's going to be very exciting. As you may have heard, even after several months of integration, we are already feeling the benefits and strengths of a project like 'Gemini'. Gemini is our next generation multimodal large-scale model. There's very exciting work going on there. It combines all the best ideas from both world-class research groups.

In the interview, CEO Hassabis also talks about AI risks, regulations, the future of LLM, etc., so if you are interested, please check it out.

Related Posts:

in Software, Posted by log1i_yk