In-depth demystification of Google's "AI reshuffle"! Latest interview with Google DeepMind CEO: How to stay competitive in the future?

By Nilay Patel

Image source: Generated by Unbounded AI tool

【Editor’s note】 Six years ago, the Transformer model proposed by Google became an indispensable research basis for AI large-scale models; today, ChatGPT launched by OpenAI based on the Transformer model and other AI technologies has enabled many top AI The Google of scientists is at a disadvantage in this AI race.

Although Google has merged the two top AI teams, DeepMind and Google Brain, to meet this challenge, some insiders broke the news that the cultures and goals of the two teams are not the same, which may become Google’s way to catch up with OpenAI and Microsoft. One of the stumbling blocks. So, what is the real situation? What strategy will Google use to deal with this AI race?

Recently, Nilay Patel, editor-in-chief of the American technology media Verge, interviewed Demis Hassabis, CEO of Google DeepMind, and had an in-depth discussion on the above issues and other important issues related to AI. The core points are as follows:

• Demis says that, by any measure, Google has achieved amazing cutting-edge research results with AlphaGo, AlphaFold, and more than 20 papers published in journals such as Nature and Science. But in a sense, ChatGPT, the big model, and the public’s reaction to it all prove that AI has entered a new era.
• Demis said, “When you realize this change, I think it requires you to change the approach to research and how much you focus on product. We all realized that it was time to streamline and focus our AI work. The obvious conclusion is to merge.”
• There are far more types of AI than generative AI. Generative AI is the “hot” thing right now, but capabilities like planning, deep reinforcement learning, problem solving, and reasoning will come back again in the next wave, alongside the current capabilities of current systems.
• Currently, Google DeepMind’s organizational structure is still evolving, and Demis wants to make sure that nothing goes wrong and everything works as it should. “Before the end of this summer, we will be a single unified entity, and I think that will be very exciting. Even within a few months, we are already feeling the benefits and advantages of this, Gemini is our next generation A multimodal large-scale model that combines the best ideas from two world-class research teams.”
• AI currently faces a very urgent need, especially for near-term problems such as deepfakes and disinformation, but the solution is already in sight.
• Demis thinks something like AGI or AGI will be achieved in the next decade or so.

Academic headlines made a simple compilation without changing the main idea of the original text. The content is as follows:

Today, I’m going to interview Demis Hassabis, CEO of Google DeepMind. Google DeepMind is Google’s new division responsible for artificial intelligence (AI) efforts across the company. Google DeepMind is the product of an internal merger: Google acquired Demis’ DeepMind startup in 2014 and operates it as a separate company within parent company Alphabet, and Google itself has an AI team called Google Brain.

Google has been showcasing research demos in the AI space for years, but with the explosion of ChatGPT and the new threat posed by Microsoft in search, Google and Alphabet CEO Sundar Pichai decided earlier this year to bring DeepMind to Google, creating Google DeepMind.

Interestingly, Google Brain and DeepMind aren’t necessarily compatible, or even necessarily focused on the same thing: DeepMind is known for applying AI techniques to areas like games and protein-folding simulations. AlphaGo has beaten the human world champion at the ancient board game Go. Google Brain, meanwhile, is more focused on the familiar generative AI toolset: the large language models behind chatbots, the editing features in Google Photos, and more. It’s a clash of cultures and a major structural decision with the goal of making Google’s AI products more competitive and get to market faster.

The competition isn’t just from OpenAI and Microsoft. A Google software engineer recently claimed that Google has no competitive advantage in AI because open-source models running on commodity hardware are rapidly evolving and catching up with tools from the giants. Demis confirmed the truth of the claim, but said it was part of Google’s debate culture, which he disagreed with because he had other ideas about Google’s competitive advantage.

Of course, we also talked about AI risks, specifically artificial general intelligence (AGI). Demis made it clear that his goal is AGI, and we discussed what timelines and steps should be taken to address the associated risks and regulatory issues. Recently, Demis, along with OpenAI CEO Sam Altman and others, signed a statement on the risks of AI that is succinct: “Mitigating the risk of human extinction from AI should be a global Priorities." That sounds pretty cool, but is there really a risk of that now? Or is it simply a distraction from the more specific issue of AI replacing large numbers of labor in various creative industries? We also discuss the new kind of workforce that AI is creating—hordes of low-wage workers sorting through data in countries like Kenya and India to help train AI systems.

This interview touches on big ideas about AI, the many questions that go with it, the myriad of complex decisions that need to be made, and one major organizational decision. My discussion with Demis was pretty in-depth, but I still feel like we haven’t covered everything. The full interview is as follows:

**Welcome Demis Hassabis. **

Thanks for the invitation.

**The field of AI involves a major idea, which brings challenges and problems, and for you, a major organizational restructuring and a series of high-risk decisions. I’m so glad you’re here. **

It’s a pleasure to be here.

**Let’s start with Google DeepMind. Google DeepMind is a new division of Google, consisting of two top teams at Google, one of which is the familiar Google Brain, which is the AI team led by Jeff Dean, and the other is DeepMind, a company co-founded by you, Acquired by Google in 2014, it operated independently as a subsidiary under the Alphabet holding company structure until recently. Let’s start at the very beginning. Why were DeepMind and Google Brain separated in the first place? **

As you mentioned, we actually started DeepMind in 2010, which was a long time ago from today’s AI era, so it can be said that it was the prehistoric period of AI. The other co-founders and I realized that as people who came from academia, we saw the development of academia, such as deep learning was just invented. We are very supportive of reinforcement learning. We are able to see that graphics processing units (GPUs) and other hardware are advancing rapidly, and if we focus on general-purpose learning systems and borrow some ideas from neuroscience and the working of the brain, a lot of progress can be made. So in 2010 we brought all these elements together. We had this thesis that we were going to make rapid progress, and that’s exactly what we achieved with the original game system. Then, in 2014, we decided to partner with what was then Google, because we foresee the need for more computing power. Apparently, Google has the most computers in the world. This is a perfect choice for us to focus on promoting research progress in the future.

**So you were acquired by Google, and Google has a new positioning after that. They formed Alphabet, and Google became a division of Alphabet. Alphabet has other divisions, and DeepMind is independent of them. This part is what I want to focus on at the beginning, because Google has promoted Google Brain to do a lot of research on language models. I remember six years ago at Google I/O where Google showed off large language models, but DeepMind focused on completely different kinds of AI research like winning Go games and protein folding. Why aren’t these studies being done within Google? Why do they belong to Alphabet? **

As part of the agreement we entered into when we were acquired, we will continue to advance research in AGI to achieve a system that can operate on a variety of cognitive tasks and possess all the cognitive capabilities of humans.

At the same time, I am also very passionate about using AI to accelerate scientific discovery, so naturally there are projects like AlphaFold, and I believe we will return to this topic in the future. But actually, long before DeepMind was founded, I thought that games were the perfect test or proof place for developing efficient, fast AI algorithms where you could generate a lot of data and the objective function was pretty clear: obviously win the game or maximize the score . There are many reasons to use games in the early stages of AI research, and it’s one of the big reasons why we’ve been so successful and able to move forward with projects like AlphaGo so quickly.

These are all very important practices that prove that these general learning techniques work. Of course, we also do a lot of work on deep learning and neural networks. We are good at combining these techniques with reinforcement learning, so that these systems can solve problems, make plans, and win games on their own. We have always been tasked with advancing the research agenda and advancing science. That’s something we’re very focused on, and it’s something that I personally hope to achieve. Then there are the AI teams within Google, like Google Brain, who have slightly different missions, closer to product and other parts of Google, bringing amazing AI technology into Google. We also have an applications division that brings DeepMind’s technology into Google products. But cultures and tasks do make a big difference.

**From the outside, the timeline looks like this: People have been working on this for a long time, talking about it. It’s a topic of discussion for some journalists like me and some researchers, and we talk about it at Google events. Then ChatGPT came out, and it wasn’t even a product. I don’t even think Sam Altman thought it was a great product when it was released, but it was just released and people could use it. Everyone freaked out, Microsoft released Bing based on ChatGPT, the world changed a lot, and then Google responded by merging DeepMind and Google Brain. From the outside, that’s how it looks. Does it look the same from the inside? **

That timeline is correct, but it’s not a direct consequence, it’s more of an indirect consequence in a sense. Google and Alphabet have always operated this way. They made a lot of “flowers” bloom, which I think has always been in line with what Larry Page and Sergey Brin started Google with. This approach gave them a lot of opportunities to build something incredible and become the amazing company it is today. In terms of research, I think it’s a very fitting way to do research, which is one of the reasons we chose Google as a partner in 2014. I feel like they really understand what it means to do basic research and prospective research and will push us to have a bigger purpose in our research. You see the results, right?

AlphaGo, AlphaFold, and more than 20 papers published in journals such as Nature and Science, etc., we have achieved amazing cutting-edge research results by any standard. But in a sense, ChatGPT, the big model, and the public’s reaction to it all prove that AI has entered a new era. By the way, it’s also a bit of a surprise to those of us who work in research, including OpenAI, why it spread so fast, because we and some other startups like Anthropic have these big languages Model. They are roughly the same in function.

So what’s surprising isn’t the technology itself, as we all understand it, but the public interest in it and the resulting buzz. It shows that we’ve had a consensus over the last two or three years that these systems have now reached a level of maturity and sophistication that can really come out of the research stage and the labs and be used to drive incredible next-generation products and experience, and breakthrough results like the AlphaFold are directly used by biologists. To me, this is just an example of AI entering a new phase where it really works in people’s everyday lives and is able to solve real-world problems that really matter, not just curiosity or entertainment like games .

When you realize this change, I think it requires a change in the way you do your research and how much you focus on the product. We all realized it was time to streamline and focus our AI efforts. The obvious conclusion is to merge.

**I would like to stop and take a moment to discuss a philosophical question. **

sure.

**I feel that the reason for the ChatGPT moment of this year’s AI explosion is that AI can do some things that ordinary people can do. I’d like you to write me an email and write a playbook, and while the output of a large language model may only be at C+ level, it’s still something I can do. People can see it. I want you to fill in the rest for this photo. It’s something one can imagine oneself doing. Maybe they don’t have such skills, but they can imagine doing it. **

All the AI demos we’ve seen before, even your AlphaFold, you say it can model all the proteins in the world, but I can’t do that, it should be done by a computer. Even a microbiologist might think, “Great! A computer can do this, because I’m thinking about how much time we need to spend, and we simply can’t do it.” “I want to beat the world champion at Go .I can’t do that. Well, no problem. A computer can do it.”

** Now the computer starts doing things that I can do, and those things don’t even have to be the most complex tasks like reading this web page and providing a summary. But that’s what touched everyone. I wonder why you think the industry as a whole didn’t see this shift coming, because we’ve been focusing on really hard things that people can’t do, and it seems to everyone’s shock that computers start doing things that people often do. **

I think this analysis is correct. I think that’s why large language models really come into the public eye, because it’s something ordinary people, the so-called “general public,” can understand and interact with. Of course, language is crucial to human intelligence and our everyday lives. I think this also explains why chatbots have spread so quickly in certain ways. While I will mention AlphaFold, and of course I may be biased when I say this, I think it actually has by far the most visible, huge, and positive impact on the world in the field of AI, because if you and any Biologists talk, AlphaFold is now used by millions of biologists, researchers and medical researchers. I think it is used by biologists almost all over the world. Every major pharmaceutical company is using it to advance their drug development programs. I’ve talked to multiple Nobel Prize-caliber biologists and chemists and they’ve told me how they’re using AlphaFold.

So we can say that a subset of scientists around the world, assuming all scientists know about AlphaFold, has had a huge impact on and accelerated the progress of their important research work. But of course, the average person on the street probably doesn’t even know what proteins are, or how important these are for things like drug discovery. Chatbots, on the other hand, are something that everyone can understand, which is unbelievable. It feels so intuitive to be able to write a poem for you or anything else that anyone else can understand and process and evaluate compared to what they would have done or been able to do themselves.

**This seems to be the point of productizing AI technology: these chatbot-like interfaces or generative products can create things for people, and that’s where the risk lies. But even in the discussion of risk, it’s escalated because people can now see, “Oh, these tools can do things.” Did you feel the same level of scrutiny as you guys worked on AlphaFold? No one seems to think “oh, AlphaFold is going to destroy humanity”. **

No, but it does get a lot of scrutiny, and this is in a very specialized field, for a well-known expert. In fact, we spoke with more than 30 experts in the field, including top biologists, bioethicists, and biosafety experts. We collaborated with the European Bioinformatics Institute to publish the AlphaFold database, which contains structural information for all proteins, and they also guided us on how to release this data safely. So we did get a lot of scrutiny and we consulted and the main conclusion was that the benefits far outweighed the risks. Although we made some minor adjustments to the structure of the release based on their feedback. But it does get a lot of scrutiny, but again, it’s only happening in a very specialized field. On the question of generative models, I do think we are at the beginning of an incredible new era that will arrive in the next five to ten years.

It’s not just about advancing science, it’s about the types of products we can build that improve people’s everyday lives, impact the daily lives of billions of people and help us live more efficiently and enrich our lives. And I think the chatbots we see today are just the tip of the iceberg. There are far more types of AI than generative AI. Generative AI is the “hot” thing right now, but I think capabilities like planning, deep reinforcement learning, problem solving, and reasoning will come back again in the next wave, alongside the current capabilities of current systems. So I think in a year or two, if we talk again, we’re going to be talking about a whole new class of products, experiences and services that have capabilities that we haven’t had before. I’m actually pretty excited about building these things. That’s one of the reasons I’m so excited to lead Google DeepMind, where I’ll focus on building these next-generation AI-based products.

**Let’s dig a little deeper into the situation at Google DeepMind itself. Let’s say Sundar Pichai comes to you and he says, “Well, I’m the CEO of Alphabet and Google. I can make this decision. I’m going to merge DeepMind into Google, I’m going to merge with Google Brain, and you’re going to be the CEO.” When you hear What is your reaction to this thought? **

But in fact, it’s not. It’s more of a conversation between the leaders of the various teams involved and Sundar about the inflection points we’re seeing, how mature the system is, what’s possible in the product area, how we can improve our user experience, we How exciting the experience of hundreds of millions of users will be, and the comprehensive elements that it all needs. These include changes in focus, changes in research methodology, and integration of required resources such as computing resources. Therefore, as the leadership team, we discussed a series of important factors to consider and then drew conclusions from them, including the decision to merge and the plans and post-merger research priorities for the next few years.

**Is there any difference between being a CEO at Google and being a CEO at Alphabet? **

It’s early days, but I think it’s the same, because even though DeepMind is an Alphabet company, it’s very unusual for another experimental project (alpha bet) because we’ve worked with a lot of Google product teams and Organizations are closely integrated and collaborated. There is an applications team at DeepMind, and their job is to translate our research into product features by collaborating with Google product teams. In fact, we’ve made hundreds of successful launches over the past few years, just quietly behind the scenes. In fact, many of the services, devices, or systems you use every day at Google have some of DeepMind’s technology behind them. So we already have this integration structure. Of course, we’re known for our breakthroughs in science and gaming, but behind the scenes we also do a lot of groundwork that affects every part of Google.

Unlike other situations, we didn’t have to build a separate business outside of Google. Even as an independent company, this was never our goal or mission. Now within Google, we are much more tightly integrated in terms of product services, which I think is an advantage because we can collaborate more deeply with other product teams and do more exciting things than at Google. other than that is easier to implement. But we still retain some freedom to choose those processes and systems that fully optimize our mission of producing the world’s most powerful and general AI systems.

** There are reports that this is actually a culture clash. And now you’re in charge of both. How did you organize the team? As CEO, how is Google DeepMind organized under your leadership? How do you manage cultural integration? **

In fact, it turns out that the two have a greater cultural similarity than outside reports. The whole process was very smooth and enjoyable, because we are talking about two world-class research teams, two of the best AI research institutions in the world, with incredible talents and achievements on both sides. As we thought about the merger and planned, we listed each team’s top ten breakthroughs. When you add this up, these breakthroughs cover 80%-90% of the breakthroughs that have built the modern AI industry in the past decade, from deep reinforcement learning to Transformers and more. This is an unbelievable group of talents, and both parties have extremely high respect for each other’s teams. In fact, the two teams have collaborated extensively on a project level over the past decade.

Of course, we know each other very well. Actually, I think the question is focus and some coordination between the two teams, and in terms of what areas we focus on, it makes sense for two independent teams to work together and maybe eliminate some duplication of efforts. To be honest, these are fairly obvious things, but very important for us to enter the new stage of AI, which is more about AI engineering, which requires a lot of resources, including computing resources, engineering resources and other resources . Even at a company the size of Google, we have to choose carefully and be clear about where we’re going to put our resources, and focus on those directions, and then achieve those goals. So I think it’s a natural evolutionary part of our AI journey.

**That thing you mentioned, “We’re going to merge these teams, we’re going to pick and choose what we’re going to do, we’re going to eliminate some of the duplication of effort.” It’s all about organizational structure. Have you settled on a structure? What do you think that structure will look like? **

The organizational structure is still evolving. We are only at the beginning of a few months. We want to make sure that nothing goes wrong and everything works. Both teams are highly productive, doing excellent research, and involved in some very important product work. All of this needs to go on.

**You keep talking about two teams. Do you think it’s two teams, or are you trying to combine it into one? **

No, of course, it’s definitely a unified team. I like to call it a “super unit,” and I’m really excited about it. But obviously, we are merging it, forming a new culture and a new organizational structure, which is a complex process, bringing together two such large research groups. But I think by the end of this summer we’ll be a single unified entity, and I think that’s going to be very exciting. Even within a few months, we are already feeling the benefits and advantages of this, as you may have heard of Gemini, our next-generation multimodal large-scale model that combines two world-class research teams best idea.

**You need to make many decisions. What you’re describing is a complex series of decisions, and how are we supposed to police it all in the outside world? This is another series of very complex decisions. As a chess champion and someone who has made games, what is your framework for decision making? I think it’s much more rigorous than other frameworks I’ve heard. **

Yes, I think that might be true. I think if you play chess seriously, even to a professional level, I was exposed to chess from my childhood, from the age of four, it is very shaping for your brain. So I think, in chess, problem solving and strategic planning, it’s a very useful framework for many things and decisions. Chess is basically about making decisions under pressure from your opponent, it’s very complex, and I think that’s a wonderful thing. I advocate for it to be included as part of the school curriculum because I think it is an excellent training ground for problem-solving and decision-making. However, I think the overall approach is closer to the scientific method.

I think all of my training, PhD and postdoc and so on, was obviously in neuroscience. So I learned about the brain, but it also taught me how to do rigorous hypothesis testing and hypothesis generation, updated with empirical evidence. The whole scientific method as well as planning in chess, both of which can be translated into the business world. You have to translate it intelligently without making it too academic. And in the real world, there is often a lot of uncertainty and hidden information in the business world, and you don’t know everything. So, in chess, all the information on the board is obvious to you. You can’t transfer these skills directly, but I think they can be very helpful if applied in the right way.

**How do you combine the two in some of the decisions you make? **

I make so many decisions every day, it’s hard to give an example right now. However, I tend to try to plan and forecast many years in advance. So, the way I tell you I try to approach things is, I have an end goal. I’m pretty good at imagination, which is a different skill to imagine or conceive of a perfect end state, whether it’s in terms of organization, product, or research. Then, starting from the end goal, I identify all the steps needed and in what order to make that outcome as realistic as possible.

Yep, it’s kind of like chess, right? In a sense, you have a plan and hope to get to the point where you beat your opponent, but you still have a lot of moves to make it happen. So in order to increase the likelihood of the final outcome, you must take incremental steps to improve your situation. I find it useful to trace back the search process from the final goal to the current state.

**Let’s apply this way of thinking to some products. You mentioned a lot of DeepMind’s technology and Google’s products. What we can see is Bard and your search generation experience. There is also AI in Google Photos, but focused on large language models (Bard and search generation experience). These cannot be final states. They are not fully mature. Gemini is coming soon, we may improve both products, all of these things will happen. When you think about the end state of these products, what do you see? **

Google’s AI systems aren’t just being applied to consumer-facing products, they’re also being applied under the hood that you may not be aware of. For example, our initial application of the AI system to the cooling system of Google’s data centers, which are massive, actually reduced the energy consumption of the cooling system by almost 30%, and if you multiply this across all the data centers and computers, the benefit will be huge. So there’s actually a lot going on under the hood of applying AI to continually improve the efficiency of these systems. But you’re right that the current products aren’t the final state, they’re really just interim. As far as chatbots and these kinds of systems go, eventually they’re going to be an incredibly versatile personal assistant that you use many times in your day-to-day life for really useful and helpful things.

From recommending books to read, to suggesting on-site events and the like, to booking travel and planning trips, and even assisting with day-to-day tasks. I think current chatbots are a long way from achieving this, and we know that some elements of planning, reasoning, and memory are missing. We are also working on it. I think today’s chatbots will pale in comparison to what’s coming in the next few years.

**My job is to cover the computing world. I think of a computer as a relatively modular system. You look at a phone, it has a screen, a chip, a cellular antenna and so on. So, should I be looking at AI systems in the same way? That is, there may be a very convincing human language interface behind it, such as a large language model, and behind it may be AlphaFold that actually performs protein folding? How do you think about tying these things together, or is it a different evolutionary path? **

In fact, there has been an entire branch of research devoted to so-called “tool use”. This concept refers to these large language models or large multimodal models that are experts in language and of course may have some math and programming and so on. But when you ask them to do something specialized, like fold proteins, play a game of chess, or something like that, they actually end up calling a tool, which could be another AI system, and then that tool provides a specific problem solutions or answers. Then in the form of language or images, this information is passed back to the user through the central large-scale language model system. So it might actually be invisible to the user, because to the user it looks like just one big AI system with multiple abilities, but under the hood, this AI system might be broken down into special Smaller systems with functionalities.

In fact, I think this might be the next era. Next-generation systems will leverage these capabilities. You can think of the central system as a switch statement, you effectively prompt it through the language and connect your query, question or whatever you ask it to answer to the appropriate tool as needed to answer the question or provide you with a solution . And then pass it back in a very understandable way, again using the best interface, the natural language interface.

**Will this process bring you closer to AGI, or will you reach some limit state where you need to do something else? **

I think this is the critical path to AGI, which is another reason why I’m so excited about this new role. In fact, the product roadmap and research roadmap from here on toward AGI-like or human-level AI are extremely complementary. In order to build products that are useful in everyday life, like general-purpose assistants, we need to advance some key capabilities, such as planning, memory, and reasoning, which I think are critical to our realization of AGI. So I think now there’s a really nice feedback loop between the product and the research, and they complement each other effectively.

**I have interviewed a lot of car company CEOs before. I asked all of them, “When do you think we’re going to have driverless cars?” They all said five years, and they’ve been saying five years for five years, right? **

right.

**I wanted to ask you a similar question about AGI, but I feel like some people I’ve been talking to have been getting smaller numbers lately. How many more years do you think it will take us to achieve AGI? **

I think there’s a lot of uncertainty about how many more major breakthroughs are needed to achieve AGI, and those breakthroughs are likely to be innovative rather than just scale-up of existing solutions. In terms of time frame, a lot depends on that. Obviously, if it takes a lot of big breakthroughs, those breakthroughs will be more difficult and take longer. But for now, I wouldn’t be surprised if AGI-like or AGI-like states are achieved in the next decade or so.

**During the next ten years. OK, I’ll come back to you in ten years and see if that works out. **

Can.

**True, it’s not a straight path. As you said, there may be some breakthroughs in the process, which may disrupt the original plan and lead it to a different path. **

Research is never a straight line. If it’s a straight line, it’s not real research. It’s not research if the answer is known before you start. Hence, day-to-day research and disruptive research are always accompanied by uncertainty, which is why the timeline cannot be accurately predicted. But we can keep an eye on trends and observe the quality of the ideas and projects that are going on right now and how they are progressing. In the next five to ten years, this trend may appear in two situations. Maybe we will approach a gradual state, or we may encounter bottlenecks in existing technologies and expansion. It wouldn’t surprise me if the latter came up, perhaps we’ll find that simply scaling existing systems leads to diminishing system performance and ultimately diminishing returns.

Practically, this means we do need some innovation to make progress. Right now, I don’t think anyone knows what state we’re in. So the answer to this question is that you have to try to advance both at the same time as much as possible. Significant resources are required both in scaling and engineering of existing systems and existing ideas, as well as in exploratory research directions that can deliver innovations that address some of the weaknesses of current systems. As a large research organization with significant resources, this is our advantage, and we can maximize our bets in both directions. In a way, I’m neutral on the question “Do we need more breakthroughs, or can the existing system just keep scaling?” I think this is an empirical question that should be pushed in both directions as much as possible. The results will speak for themselves.

**There is indeed a contradiction here. When you work at Alphabet’s DeepMind, you’re very focused on research, and then that research is passed back to Google, where engineers at Google turn it into a product. You can see how this relationship plays out. Now, you’re inside Google. As a company, Google is under enormous pressure to win this race. These are product issues that are about “making people feel real and winning in the marketplace.” There’s a leaked memo purportedly from within Google. It says that Google has no competitive advantage, open source AI models or leaked models will run on people’s laptops, and they will outperform Google because of the history of open computing over closed source competitors. Is that memo real? **

In my opinion, that memo is genuine. I think engineers at Google often write various documents, and sometimes they get leaked and spread quickly. I think it’s just a common situation, but it doesn’t have to be taken too seriously. These are just personal opinions. I think it’s interesting to hear those perspectives, but you need to make decisions about your own path. I have not read that specific memo in detail, but I disagree with its conclusions. Also, open source and published works are obvious, DeepMind has done a lot of open source work. For example, AlphaFold is open source, right? Therefore, we support open source and support research and open research. This is key to the scientific discussion, and we have been an important part of it. Of course, Google did the same, releasing Transformer, TensorFlow, and everything we did as you can see.

We’ve done a lot of work in this area. But I think there are other factors to consider. Obviously, business considerations are one, along with security concerns about accessing these powerful systems. If the bad guys have access to it, they probably aren’t that highly skilled enough to build a system themselves, but they can certainly reconfigure what already exists. How should we deal with these problems? I think it’s always been a theoretical question, but as these systems become more general, more complex, more powerful, it’s going to be very important to stop bad guys from using these systems for malicious things that they didn’t intend .

This is something we need to focus on, but back to your question, look at the history of innovation and breakthroughs that Google and DeepMind have made over the past decade or so. I’d bet on us, and I’m very confident that we’ll continue to produce the next key breakthrough, as we’ve done in the past, and that’s going to become even more true over the next decade.

**We invent most things, so we will invent most future things, do you think that is our advantage? **

I don’t see it as a competitive advantage, but I’m a very competitive person. This is probably another trait I get from playing chess, as do many researchers. Of course, they do research to discover knowledge that ultimately improves the human condition, which is our goal. But at the same time, we also want to be the first to achieve in these areas, and do it in a responsible and bold way. We have the best researchers in the world, I think globally, we have the most top researchers, and we have incredible achievements. There’s no reason to think this won’t continue in the future. In fact, I think our new organization and environment are likely to break through more and faster than in the past.

**You remind me of risk and regulation. But I want to start the discussion from a different perspective. You mentioned all the work that needs to be done, and how deep reinforcement learning works. We partnered with New York on a feature-length cover story about the mission workers who actually did the training and data labeling. There is indeed a lot of discussion about workforce-related issues in the development of AI. Hollywood screenwriters are currently on strike because they don’t want ChatGPT to write scripts. I think this is reasonable. Now, though, there’s a new class of workforce, a global group of people sitting in front of their computers and saying, "Yes, that’s a stop sign. No, that’s not a stop sign. Yes, that’s wearable. clothes. No, that’s not clothes to wear.” Is this going to last? Is this just a new job that needs to be done to make these systems work? Or will there be a moment of end? **

I think it’s hard to say. I feel like this is definitely a moment, related to the current systems and the work they need right now. We’ve been very careful about this type of work, and I think you quoted some of our researchers in that article, we’re very careful about paying fair wages and being very responsible about this type of work and choosing our partners. We also use internal teams for this type of work. So, actually, I’m very proud of our responsible performance in this type of work. But going forward, I think there may be a way for these systems to propel themselves, especially when the number of users reaches millions. Or it’s conceivable that an AI system could carry on a conversation or critique on its own.

It’s kind of like turning a language system into a game-like setting, and we’re pretty good at doing that, and we’ve been thinking about how different versions of reinforcement learning systems would rate each other. Maybe such a rating is not as accurate as a human rater, but it’s actually a useful way to do some basic rating work, and then calibrate those ratings by using a human rater at the end instead of having humans rate reviewers rate all content. So I think I can see a lot of innovations emerging that will help solve this problem, meaning there will be less need for human raters.

**Do you think human evaluators are always needed? Even in the process of approaching AGI, it seems that someone still needs to tell the computer whether it is doing a good job or not. **

Let’s take AlphaZero as an example, our general game system that eventually learned any two-player game including Chess and Go. What’s interesting about what’s happening there is that we set up the system to play tens of millions of games against itself. So, in effect, it builds its own knowledge base. It starts randomly, plays itself, improves itself, trains better versions, and pits them against each other in something like a small tournament. But in the end, you still want to test it against human world champions or other external computer programs built in the traditional way, so that you can calibrate your own metrics that tell you whether these systems are improving against these goals or metrics.

But you can’t be sure of the results until you calibrate with an external benchmark or metric. Depending on the calibration method used, human evaluators or human experts are often the best choices for calibrating in-house tests. You need to make sure that internal tests actually match reality. For researchers, this is an exciting aspect of products, because when you apply research to a product, and millions of people use it every day, you get real world feedback, there’s no getting around it reality and the best test of any theory or system.

**Do you think it is valuable or appropriate to work on labeling data for AI systems? Some of these are worth pondering, like “I’m going to tell a computer how to understand the world so that it might replace others in the future.” There’s a kind of loop here that seems to warrant more moral or philosophical thinking. Have you taken the time to think about this question? **

Yes, I did think about this question. I don’t see it that way. I think evaluators are part of the process of developing these systems, making sure that AI systems are safer, more useful, more reliable, and more trustworthy for everyone. So I think that’s a crucial component. In many industries, we conduct security testing of technologies and products. Today, the best thing you can do for an AI system is to have human evaluators. I think we need more research in the coming years. I’ve been calling for this, and we’ve been working on it ourselves, but it takes more than one organization to do it, we need to have good, solid evaluation criteria so we know if a system passes those criteria, it’s Certain properties are safe and reliable in these particular respects.

At the moment, I think many researchers in academia, civil society, and other fields have many good proposals for these tests, but I don’t think they are robust or practical enough. They are basically theoretical and philosophical in nature, and they need to be practically applied so that we can measure our systems empirically against these tests and thus have some guarantees about the performance of the system. Once we have these tests, there will be less need for humans to evaluate test feedback. I just think the reason for the need for such human evaluation test feedback at the moment is because we don’t have these independent benchmarks yet. Part of the reason is that we haven’t strictly defined these properties. I mean, it’s pretty much a field involving neuroscience, psychology, and philosophy. Even for the human brain, many terms have not been properly defined.

**You’ve signed an open letter from the Center for AI Safety, as has OpenAI’s Sam Altman and others, warning people of the possible risks of AI. However, you are still working hard, Google is also competing in the market, you have to win, you also describe yourself as competitive. There’s a contradiction in this: needing to win and launch a product in the market, but wanting to "oh my god, please regulate us, because if we don’t stop it somehow, pure capitalism is going to lead us to cliff.” How do you balance that risk? **

There is indeed a contradiction, a creative contradiction. At Google, we like to say that we want to be both bold and responsible, and that’s what we strive to be and lead by example. To be bold is to be courageous and optimistic about the benefits that AI can bring to the world, helping humanity tackle some of our biggest challenges, whether it’s disease, climate, or sustainability. AI has a huge role to play in helping scientists and medical professionals solve these problems, and we are working hard in these areas. And AlphaFold, again, I can point to it as a star project, demonstrating our efforts in this regard. That’s the bold side. And the responsible side is making sure that we do our work with as much prudence and foresight as possible, taking these factors into account as much as possible.

We need to predict the possible problems of success as far in advance as possible, rather than “after the fact”. Perhaps social media is an example, which has experienced incredible growth. Obviously, it produced a lot of good in the world, but 15 years later, we realized that these systems also had some unintended consequences. For AI, I hope to take a different path. I think it’s a deep, important and powerful technology. We must do this in the face of a technology with such transformative potential. That doesn’t mean no mistakes will be made. It’s a very new technology, and anything new can’t be predicted in advance, but I think we can do the best we can.

The point of signing that letter was to show that, while I think it’s unlikely, we should also consider what those systems might be able to do and what they might do as we get closer to AGI. We are far from that stage at the moment. So it’s not about technology today or in the next few years, but at some point, given the rapid pace of technology development, we’re going to need to think about these issues, not just before they happen. We need to conduct research and analysis, and engage with various stakeholders, including civil society, academia, and government, five, ten, or more years into the future, in order to remain relevant in this rapidly evolving field. , determine the best option to maximize benefits and minimize risks.

At the current stage, this mainly consists of doing more research in these areas, such as proposing better evaluation methods and benchmarks to rigorously test the capabilities of these cutting-edge systems.

**You talked about the tool use of AI models, you can have a large language model to do something, it will ask AlphaFold to help you fold proteins. When combining and integrating systems like this, historically, this has led to new behavioral traits, and things that you can’t predict. Are you worried about this? There is no one rigorous test for this. **

Yes, exactly. This is exactly what I think we should be researching and thinking about ahead of time: as tool use becomes more sophisticated and different AI systems can be put together in different ways, new behavioral traits may emerge. Of course, this new behavioral signature can be very beneficial and extremely useful, but in the wrong hands or in the hands of malicious operators, it is also potentially harmful.

**Assuming that most countries around the world agree on some kind of AI regulatory framework, but individual countries say, “To hell with it, I don’t play by the rules.” This will become a hub for malicious actors to conduct AI research. What would it be like? Do you foresee this possible world? **

Yes, I think that’s a possible world. That’s why I’ve been in conversations with the government, and I think whatever regulations, safeguards, or whatever, should be tested over the next few years. Ideally, these measures should be global, and there should be international cooperation and international agreement on these safeguards.

**If the government passes a rule here, “Here’s what Google is allowed to do, here’s what Microsoft is allowed to do. You’re in charge, you’re in charge.” Then you can say, "Okay, we’re not in our We’re not going to have those capabilities; it’s illegal.” If I were just a normal guy with a MacBook, you’d accept restrictions on some of the MacBook’s capabilities because the threat of AI is just too scary ? This is something that worries me. From a practical standpoint, if there are open source models, and people use them for weird things, are we going to tell Intel to limit the capabilities of its chips? How do we enforce restrictions like this so that it actually affects everyone and not just a “if Google does something we don’t like, we put Demis in jail” approach? **

I think these are important issues that are currently being debated. I do worry about this issue. On the one hand, there are many benefits to open source and accelerated scientific discussion, a lot of progress is happening there, and it provides opportunities for many developers. On the other hand, if some bad individuals use this way to do bad things and spread them, it may bring some negative consequences. I think this is something that needs to be addressed in the next few years. I think it’s okay because now the system is not as complex, not as powerful, and therefore less risky.

But I think as the system becomes more capable and ubiquitous, the question of access rights will need to be thought about from the perspective of the government, how they restrict, control or monitor this will be an important question. I don’t have an answer for you, because I think this is actually a social issue that requires the participation of stakeholders from all walks of life to weigh the benefits and risks.

**Google’s own work, you say you’re not there yet, but Google’s work in AI does create some controversy about liability and what the models can or can’t do. Emily Bender, Timnit Gebru, and Margaret Mitchell published a famous “Stochastic Parrots” paper, which caused a lot of controversy within Google and led to their departure. Did you read that paper and think, “well, that’s right, large language models lie to people, and Google will be held accountable”? How do you feel about facing so much scrutiny now? **

Yes, in fact, there are hallucinations and inaccuracies in large language models, which is one of the reasons why Google has been very responsible. We know this. Improving factuality, relevance, and making sure they don’t spread disinformation are key areas that need to be improved over the next few years. This is a matter of great concern to us. We have many ideas for improvement. Our previous release of DeepMind’s Sparrow language model was an experiment to explore how factual and rule-compliant we can get in these systems. It turns out that we might be able to improve it by an order of magnitude, but sometimes this can come at the expense of the clarity, creativity, or usefulness of the language model.

Indeed, it’s a bit like a Pareto frontier, where improving in one dimension reduces capacity in another. Ideally, in the next phase and next generation of systems, we would like to combine the creativity, clarity and playfulness of the current system with factuality and reliability. We still have a long way to go in this regard. But I can see things improving, and I don’t see any theoretical reason why these systems won’t be able to achieve extremely high levels of accuracy and reliability for years to come.

**When you use Google Search Generative Experience (SGE), do you believe what it says? **

I believe it. Sometimes I double check, especially in the sciences, and I’ve had really interesting cases where it happens with virtually all of these models. For example, when I ask them to summarize a certain area of research, I think it would be very useful if they could do that, and then ask, “So, what important papers should I read?” papers with plausible author lists. But when you go to look it up, it turns out they’re just the most famous people in the field or the titles of two different papers put together. But of course, judging by the combination of words, they are very believable. I think in this case, these systems need to understand citations, papers, and author lists as a whole, rather than word-by-word predictions.

These interesting cases really need us to improve, and as people who want to push the frontier of science, we certainly hope to improve and fix these problems, this is a particularly interesting application scenario, we hope to be able to improve it, but also to meet our own needs . I wish these systems would do a better job of summarizing for me “here are the five best papers read about a particular disease” or something like that to get a quick overview of a particular field. I think this will be very useful.

**Just to tell you something, I googled my friend John Gruber and SGE confidently told me that he pioneered the use of Macs in newspapers and invented WebKit. I don’t know what the source of this information is. Does it need to reach a certain level of quality or realism before rolling it out to a broad audience? **

Yes, we think about it all the time, especially at Google, because Google has extremely high standards for things like search, and we rely on these services every day and every hour. We want to be able to achieve that level of reliability. Obviously, we have a long way to go at the moment, and not just us, but anyone is far from this level of generational systems. But it’s our gold standard. In fact, tool usage is very useful in this regard, and you can build these systems so that they can fact-check themselves, and even cross-reference using searches or other reliable sources, just like a good researcher. Cross-check the facts. At the same time, we also need a better understanding of the world, of what research papers are, which entities they involve, and so on.

Therefore, these systems need to have a better understanding of the media they process. It might also be possible to give these systems the ability to reason and plan, since then they might be able to evaluate their own output. We have a lot of experience with this in game programs. They don’t just output the first move you think of in chess or Go. They actually do some searching and planning, and then back up. Sometimes they change their minds and move on to better steps. You can imagine a similar process in language and writing.

** There is a concept called model collapse. We will train a large language model on the data generated by the large language model, which may form a loop. When you talk about cross-referencing of facts, I think of Google – Google will take a bunch of information from the Internet and try to cross-reference, but maybe all this information is delusional generated by large language models in 2023. So, how to guard against this situation? **

We’re working on solving this problem, and we’re building some really cool solutions. I think the answer is to use encrypted watermarking, a complex watermarking technique that is difficult or impossible to easily remove, and may be embedded directly into the generative model as part of the generative process. We hope to release this technology and possibly make it available to third parties as a general solution. I think the industry needs these types of solutions where generated media, including images, audio, and even text, can be tagged with some kind of flag to indicate to users and future AI systems that the media was generated by AI. I think that’s a very urgent need for AI right now, especially for near-term issues like deepfakes and disinformation. But I do think a solution is in sight.

**Demis, thank you very much and look forward to your next visit. **

Thank you very much.

Original author:

Nilay Patel, editor-in-chief of technology media Verge

Original link: