2024-08-20
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Machine Heart Report
Synced Editorial Department
“I feel more comfortable when the CEO of an AI company acts more like a computer scientist than a salesman.”
For DeepMind, 2023 is a year full of changes. In April of this year, Google announced that it would merge Google Brain and DeepMind to form a new department called Google DeepMind. The new department will lead the research and advancement of groundbreaking AI products while maintaining ethical standards.
Google Brain and DeepMind - one created Transformer, the other created AlphaGo, AlphaFold... The two departments joined forces to create Gemini, a benchmark for ChatGPT, at the end of 2023. Today, Gemini often ranks among the top three in the LMSYS Chatbot Arena, a large model ranking. It can be seen that the merger of the two has achieved certain results.
So, what is the future of Google DeepMind? In a recent conversation with Hannah Fry, associate professor of urban mathematics at the University College London's Center for Advanced Spatial Analysis, Google DeepMind CEO and co-founder Demis Hassabis revealed some of the company's plans and expressed his views on some current issues in the field of AI.
视频链接：https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2650930939&idx=2&sn=00d72f97f26fc7acc3b2a2fd39434048&chksm=84e43a85b393b393d7a9bd7caeafce2fcd71b6299e195df3e5a716cb840a401c85dc9efff669&token=899618486&lang=zh_CN#rd
Hassabis’ core ideas are as follows:
After watching the interview, some people commented that the interview made them feel comfortable because Hassabis sounded more like a computer scientist than a salesman. Others said that acquiring DeepMind and letting them develop freely was the best AI decision Google has ever made, and they hope that Google will let them continue their work and not disturb them as much as possible.
The following is the interview content compiled by Machine Heart.
AI is developing beyond expectations
Fry: Looking back, when we started planning this podcast in 2017, DeepMind was a relatively small, focused AI research lab that had just been acquired by Google and given the freedom to pursue its own unique research projects from a safe distance in London. But things have changed dramatically since then. Since last year, Google has reorganized its entire structure to put AI and the DeepMind team at the heart of its strategy.
Google DeepMind continues its quest to give AI human-level intelligence, the so-called artificial general intelligence (AGI). It launched a series of powerful new AI models, called Gemini, and an AI agent called Project Astra that is capable of processing audio, video, images, and code. The lab has also made huge leaps in applying AI to multiple scientific fields including the prediction of the structure of all molecules in the human body, not just proteins. In 2021, they also spun out a new company, Isomorphic Labs, dedicated to discovering new drugs to treat diseases. Google DeepMind is also working on powerful AI agents that can learn to perform tasks on their own through reinforcement learning, and continues the legend of Alpha Go defeating humans in the game of Go.
Today we have invited DeepMind co-founder and CEO Demis Hassabis.
I'm wondering if your job has become easier or more difficult since the surge in public interest in AI?
Hassabis: I think it's a double-edged sword. The hard thing is that right now there's so much scrutiny, so much attention, so much noise in the field. I prefer when there are fewer people and we can focus more on the science. But on the bright side, it shows that technology is ready to impact the real world in many different ways and impact people's daily lives in positive ways, so I think that's exciting as well.
Frye: Were you surprised at how quickly the public imagination was captured? I guess you expected it to end up like this, didn't you?
Hassabis: It is. Those of us who have been working on this for decades, eventually at some point the public will realize how important AI is going to be. But it’s still a little surreal to see it actually happen and to see it happen in this way. I think it’s really because of the emergence of chatbots and the development of language models, because everyone uses language and everyone understands language, so it’s an easy way for the public to understand and measure the level of development of AI.
Frye: I heard you describe these chatbots as “extraordinarily effective.” What does that mean?
Hassabis: I mean, if you look back 5 to 10 years ago, people might have thought that in order to achieve AI development, you needed to build some amazing architecture and expand on it, without having to specifically solve specific problems like abstract concepts. In a lot of discussions 5 to 10 years ago, people thought that there needed to be a special way to deal with abstract concepts because that's obviously how the brain works. But if you give AI systems enough data, like the data from the entire internet, they do seem to be able to learn from it and generalize some patterns, not just memorize them, but actually understand to some extent what they're processing. It's kind of "extraordinarily effective" because I don't think anyone would have thought 5 years ago that it would work as well as it does now.
FRYE: So, it was a surprise…
Hassabis: Yeah, we talked about concepts and grounding earlier — grounding language in real-world experience, maybe in simulation or robotic embodied intelligence. Of course, these systems aren’t there yet, they make a lot of mistakes, they don’t have a real model of the world yet. ButJust by learning the language, they have come further than they expected.。
Frye: I think we need to explain the concept of grounding.
Hassabis: The grounding problem is a problem that classical AI systems built in the '80s and '90s at places like MIT had. You can think of these systems as giant logical databases where words have relationships to each other. The problem is that you can say "dogs have legs," and that's in the database, but when you show the system a picture of a dog, it has no idea what that bunch of pixels have to do with that symbol. That's the grounding problem — you have this symbolic, abstract representation, but what do they really mean in the real world, especially the messy real world? They tried to solve this problem, but never quite succeeded.
And today's systems, they learn directly from the data, so in a sense they're forming that connection from the beginning, but what's interesting is that if it's just learning from language, in theory it should be missing a lot of the grounding that you need, but it turns out that a lot of the grounding information can somehow be inferred.
Frye: Why do you say that?
Hassabis: In theory, because these initial large language models didn't exist in the real world, they weren't connected to a simulator, they weren't connected to a robot, they weren't even multimodal initially -- they weren't exposed to vision or anything else, they just existed in the language space. So, they were learning in an abstract domain. So it was surprising that they could infer something about the real world from that domain.
Frye: If grounding is gained through people's interaction with the system, that makes sense...
Hassabis: Exactly. So certainly if they get something wrong, like early versions got it wrong on questions like what a dog sounds like in the real world because of the lack of grounding, people correct them through feedback. And part of that feedback comes from our own knowledge of reality. So some of the grounding creeps in that way.
Frye: I remember seeing a very vivid example about the difference between "crossing the English Channel" and "walking across the English Channel."
Hassabis: Well, that example does work. If it gets the answer wrong, you tell it that it's wrong, and then it has to figure out that you can't walk across the English Channel.
Is AI overhyped or underestimated?
Frye: I want to ask you a little bit about the hype. Do you think AI is overhyped, underhyped, or just hyped in the wrong direction right now?
Hassabis: On the one hand, in the short term, AI is overhyped. People claim it can do a lot of things, but it can't, and there are a lot of startups and venture capital chasing some crazy ideas that are not mature enough.
On the other hand, I think AI is still underestimated. Maybe people don't fully understand what will happen when we reach AGI and how great the responsibility is.
Frye: You’ve been in this field for decades, and it’s easy for you to tell which of these startups and venture capitalists are chasing realistic goals and which are not. But how can others tell?
Hassabis: Obviously you have to do some technical due diligence and have some understanding of the technology and the latest trends.At the same time, you also have to look at the background of the person making the comment, how tech savvy they are, did they switch to AI last year from something else? Were they working on cryptocurrency last year? These may be some clues that they may be following the trend, which does not mean they will have some good ideas, and even if they do, it may be like a lottery.
I think this always happens when a field suddenly gets a lot of attention, then the money follows and everyone feels like they can’t miss out.
That creates an environment that is, we could say, opportunistic, which is a bit contrary to those who have been working on deep tech, deep science for decades, which I think is the way we should continue to approach AGI.
Gemini: The first lighthouse project after the merger of Google Brain and DeepMind
Frye: Let's talk about Gemini. In what ways is Gemini different from other large language models that have been released by other labs?
Hassabis: From the beginning, we wanted Gemini to be able to handle multiple modalities, so it can handle not only language, but also audio, video, images, code, and other modalities. The reason why we want to do this is, first of all, I think this is the way to allow these systems to really understand the world around them and build better world models, which goes back to the previous grounding problem.
We also had a vision of having a universal assistant. We made a prototype called Astra that not only understood what you were typing, but actually understood the context you were in. An intelligent assistant like that would be much more useful. So we built multimodality in from the beginning. This was another thing that only our model was doing at the time, and now other models are catching up.
Our other big innovations in memory, like long context, can actually remember about a million or two million tokens. So you can give it War and Peace or the entire movie and have it answer questions or find things in a video stream.
Frye: At Google I/O, you used an example of how Astra could help you remember where you put your glasses, right? But I wonder if this is just a fancy version of those old Google Glasses.
Hassabis: Of course, Google has a long history of developing glasses devices, actually dating back to about 2012, which was way ahead of its time. But they were maybe just lacking the technology, and the agent or the smart assistant can really understand what you are saying. So, we are very excited about the digital assistant that can always be with you and understand the world around you. It seems like a really natural use case when you use it.
Frye: Next I want to go back a little bit to the origins of Gemini, which came from two different research departments at Google.
Hassabis: Yes, last year we merged Alphabet's two research divisions, Google Brain and DeepMind, into Google DeepMind. We call it a super unit, bringing together the best talent from across the company. This means that we combine the best knowledge gained from all research, especially in language models.
So we launched models like Chinchilla, Gopher, and built PaLM, LaMDA, and other early models. Each of these models had its own strengths and weaknesses, so we combined them into Gemini, which became the first Lighthouse Project launched after the merger. Then, another important thing was to integrate all the computing resources so that you can do super-large-scale training runs. I think these are great.
Frye: In many ways, Google Brain and DeepMind have slightly different focuses. Is that okay?
Hassabis: Google's various departments are obviously focused on the forefront of artificial intelligence, and there has been a lot of collaboration at the individual research level, but there are differences at the strategic level. With the merger of Google DeepMind, I like to describe it as Google's Engine Room, which is running very well. I think there are many more similarities than differences in the way we work, and we will continue to maintain and strengthen our advantages in basic research and so on.
For example,Where does the next Transformer architecture come from?We want to invent it. Google Brain researchers invented the Transformer architecture that is popular today. We combined that architecture with deep reinforcement learning that we pioneered. I think there is still a need for more innovation. I support doing this, just like the Google Brain and DeepMind teams have done over the past 10 years. It's exciting.
Future direction: Combining AlphaGo with Gemini
Frye: I want to talk about Gemini, how is it performing? How does it compare to other models?
Hassabis: This question involves benchmarks.I think the whole field needs better benchmarks. There are some well-known academic benchmarks, but they are saturated now and don't really distinguish between the subtle differences between different top models.。
In my opinion,There are currently three types of models at the top and forefront: our Gemini, OpenAI’s GPT, and Anthropic’s ClaudeThere are also many models that perform well, such as the Llama series and Mistral series models launched by Meta and Mistral, which are good at different tasks. It depends on what type of task you want to perform. Choose Claude for encoding, GPT for reasoning, and Gemini for memory, long context, and multimodal understanding.
Of course, companies will continue to improve models. For example, Gemini is only a model that has been launched for less than a year. I think we are on a very good trajectory, and hopefully Gemini will be at the forefront the next time we talk.
Frye: Yes, the big models still have a long way to go. Does that also mean that these models are not very good in some ways?
Hassabis: Absolutely. In fact, this is the biggest debate right now. Many of the models now are derived from technologies that were invented five or six years ago. So, these models are still missing a lot of things, they produce hallucinations, and they are not good at planning.
Frye: What kind of plans?
Hassabis: For example, some long-term planning, models can't solve long-term problems. You give it a goal, they can't really take action for you. So,The model is very similar to a passive question answering systemYou ask a question, and they give you some sort of response, but they don't solve the problem for you. For example, you want a digital assistant to book your entire vacation to Italy, and book all the restaurants, museums, etc. Unfortunately, it can't do that.
I think this is the next era of research, and we call them (more) agent-based systems or intelligent systems that have agent-like behavior. Of course, this is what Google is good at. Google has built the game-playing agent AlphaGo and other agents in the past. So,A lot of what we are doing is combining established projects with new large-scale multimodal models and becoming the next generation of systems, such as AlphaGo combined with Gemini。
Frye: I think AlphaGo is very good at planning.
Hassabis: Yes, AlphaGo is very good at planning. Of course, it is only in the field of games. So, we need to generalize it to general areas such as daily work and language.
Frye: You mentioned earlier that Google DeepMind has now become the engine room of Google. That's a pretty big shift. So is Google making a big bet on AI?
Hassabis: I think so. I think Google has always understood the importance of AI. When Sundar took over as CEO, he said that Google is an AI-first company. We talked about this early in his tenure, and he believed that AI has the potential to be the next big paradigm shift after mobile Internet, and it has greater potential than before.
Maybe in the last year or two, we've really started to experience what this means, not just from a research perspective, but also in terms of product and other aspects. It's very exciting, so I think it's the right choice for us to align all of our talents and then do our best to advance AI progress.
Fry: We know that Google DeepMind attaches great importance to research and scientific things. But as it becomes the engine room of Google, does it mean that it must care more about commercial interests and no longer the purest things?
Hassabis: Yes, we are definitely more concerned about the commercial interests within our remit. But actually, I have a few things to say. First, we will continue the scientific work related to AlphaFold, and AlphaFold 3 was released a few months ago. We are also doubling down on investment in this. I think this is a unique work done by Google DeepMind.
You know, even our competitors think this is going to be a general purpose AI product. We started a new company, Isomorphic Labs, to do drug discovery. That's all very exciting, and it's going really well. So we'll continue to do that. At the same time, we're also doing a lot of work on climate prediction and other things.
We have a large team, so we can do multiple things at once. We're building our large-scale models, Gemini, and so on. We're building a product team to bring all of this amazing technology to all of the areas where Google is. So in a way, that's an advantage we have, to be able to plug in all of our technology at any time. We invent things that can be used by a billion people immediately, and that's really inspiring.
Another thing is,We now need a much greater degree of integration between AI technology developed for products and work done for pure AGI research purposes.Five years ago, you had to build some special AI for a product. Now you can separate out the main research, and of course you still have to do some product-specific work, but it's probably only 10% of all the work.
therefore,There is no longer a contradiction between developing AI products and building AGI.I would say 90% is the same research program. So if you launch products and you put them out into the world, you learn a lot from that. People use it, and you learn a lot about things like your internal metrics don't quite match up with what people are saying, and then you can update that. That helps a lot with your research.
How to test GenAI technology
Frye: I wonder if there’s a tension between the breakthroughs that come from applying AI to science and the right time to release these things to the public. Inside Google DeepMind, tools like the Large Language Model are used for research rather than being viewed as potential commercial products.
Hassabis: We’ve taken responsibility and safety very seriously from the beginning. Google incorporated some basic ethical principles into its AI guidelines as early as 2010. We have always been consistent with the entire Google and hope to deploy responsibly as one of the leaders in this field.
So it's interesting to start shipping real products with GenAI capabilities now. There's actually a lot to learn, and we're learning quickly, which is good. For current technologies, our risk is relatively low because they're not that powerful yet. But as the technology becomes more powerful, we have to be more careful.
The product team, as well as other teams, are learning how to test GenAI technology. These technologies are different from normal technology because it doesn't always do the same thing. It's almost like testing an open-world game, and the things you can try to do with it are almost endless. So it's fun to figure out how to red team it.
Frye: So, the red team test here is you guys are competing against each other?
Hassabis: Yes. Red team testing is when you take a dedicated team out of the development technology team to stress test the technology and try to break it in any way possible. You actually need to use tools to automate the testing, and even if there are thousands of people doing it, it's not enough compared to billions of users.
Also, I think we have to do it in phases, with an experimentation phase, a closed beta phase, and then release again, just like we've released games in the past. So you're learning at every step. I think what we need to do more is use AI itself to help us with red team testing internally, to actually automatically find some bugs or do triple screening. So that our developers and testers can really focus on those tricky cases.
Frye: There's something really interesting here, where you're in a much larger probability space. So even if there's a small chance that something will happen, if you try it enough times, eventually it's going to go wrong. I think there have been some public mistakes.
Hassabis: As I mentioned, I think product teams are used to all kinds of testing. They know they've tested these things, but in a random and probabilistic way. In fact, in many cases, if it's just a normal software, you can say that you've tested 99.999% of the things. And then infer that, that's enough.
But that's not the case with generative systems. They can do all sorts of things that are a little bit outside the norm, a little bit outside the realm of what you've seen before. If some smart person or adversary decides to test these systems in some way, like a hacker.
These systems might exist in a combinatorial way, with everything you've ever said to it in it. And then it's in some special state, or its memory is filled with special things, and that's why they need to output something. It's complicated, and it's not infinite. So there are ways to solve this problem, but it's a lot more nuanced than rolling out normal techniques.
Frye: I remember you saying, I think it was when I first interviewed you, that we actually have to recognize that this is a completely different kind of computing. You have to move away from the deterministic things that we understand perfectly and move toward something much messier, like probabilistic. Do you think the public also needs to change their mindset a little bit about the kind of computing that they think about?
Hassabis: Yeah, I agree. Maybe that's another thing we need to think about, and it's interesting thatBefore you release a system, you can actually release a principles document or something like that., to clearly show the intended use of this system, what is it designed to do? What is it good for? What can't it do? I think there really needs to be some kind of recognition here, like, if you use it in these ways, you'll find it useful, but don't try to use it for other things, because it won't work.
I think this is something we need to do in some areas, and users may also need experience in this area. It's actually very interesting, and this may be why the chatbot itself was somewhat unexpected, even for OpenAI, including ChatGPT, they were surprised. We also have our own chatbots, and we also noticed that these bots still have flaws, such as hallucinations and other problems.
But what we don't realize is that despite these flaws, there are actually a lot of really good use cases for chatbots. People are finding some really valuable uses for things like summarizing files and long documents, writing emails, filling out forms, and so on. Because of the wide range of use cases, even if there are some small mistakes, people don't actually mind, humans can easily fix them, and it saves a lot of time. I guess that's the amazing thing that people are finding, when they use it, people are finding these valuable use cases, even though these systems are flawed in all the ways that we know they are.
About open source: Once released, it cannot be withdrawn
Frye: That brings me to my next question, which is about open source. As you mentioned, when things are in the hands of people, really remarkable things can happen. I understand that DeepMind has open sourced a lot of its projects in the past, but that seems to have changed over time.
Hassabis: Yes, we are very supportive of open source and open science. As you know, we openly publish almost everything we do, such as Transformer, AlphaGo and AlphaFold, which are all published in Nature and other journals, and AlphaFold is also open source. By sharing information, technology and science can advance rapidly. So we almost always do this, and we think it is a very beneficial thing, which is the way science works.
The only exception is that AI, AGI, and strong AI have two sides. The question is who is using it, and scientists and technicians who really act with good intentions can make constructive and critical suggestions, which is the fastest way for society to progress. But the question is how do you also limit access to people with bad intentions who may use the same systems for bad purposes, misuse them, such as weapons systems, but we can't predict these in advance. And the general systems themselves can be repurposed in this way. We can still grasp it today because I don't think these systems are that powerful.
Over the next two to four years, especially as we start developing systems with agent-like behaviors, there is a possibility that these systems could cause serious harm if they are misused by someone. While we don't have specific solutions, as a community we need to think about what this means for open source.
Maybe the leading edge models need to go through more review and then be open sourced a year or two after release.This model is what we are following because we have our own open source models called Gemma. These models are smaller and not cutting edge, so their functionality is still very useful to developers, easy to run on a laptop, and have fewer parameters. The functionality is now well understood. However, the performance of these models is not as good as the latest cutting edge models such as Gemini 1.5. The approach we may eventually take is,We will have open source models, but they will be about a year behind the state-of-the-art, so that we can truly evaluate how users use these models in public and understand the capabilities of cutting-edge models.
The main problem with open source is that once you release it, you can’t take it back.Unlike proprietary models, developers cannot simply shut down an open source model if it is used in an inappropriate way.Once open source is used, it is like walking through a one-way door, so you need to be very cautious before open sourcing.
Frye: Whether AGI can be confined to the moat within an organization.
Hassabis: That’s an open question. We don’t know how to do this yet, because that’s something we need to think about when we start talking about high-level, human-like AI.
Frye: What about the middle layer?
Hassabis: At the middle level, we have some good ideas for dealing with these issues. For example, testing can be done through safe sandbox environments. This means testing the behavior of an agent in a game environment or a partially connected version of the internet. There is a lot of safety work going on in this area and in other areas like fintech. We might borrow some of these ideas and then build systems accordingly, and that's how we test early prototype systems. But we also know that these measures may not be enough to constrain AGI, a system that may be smarter than us. So we need to understand these systems better so that we can design protocols for AGI. By then, we will have better ways to control it and perhaps also leverage the next generation of AI systems and tools to monitor AI systems.
How to regulate AI
Frye: On the topic of security, many people seem to think that the word regulation solves all problems. How do you think regulation should be structured?
Hassabis：The government is accelerating its understanding and involvement in AI technology, which is a positive phenomenon.I believe international cooperation is necessary, especially in areas such as regulation, safety measures and deployment norms.。
As we get closer to AGI, we need to recognize that because technology is advancing rapidly,Our regulatory approach also needs to be flexible and adapt quickly to the latest technological developmentsIf you had regulated AI five years ago, you would have regulated something completely different. Today we see generative AI, but five years from now it may be something different.
Currently, agent-based systems may pose the highest risk. Therefore, I suggest strengthening existing regulations in areas where there are already regulations (such as health, transportation, etc.) and adapting them to the AI era, just as regulations were previously updated for mobile and the Internet.
The first thing I would do is stay focused and make sure we understand and test cutting-edge systems. As things become clearer, regulations need to be developed around these situations, perhaps in a few years when it will make more sense. What we are currently missing is benchmarking, proper capability testing, including what the entire industry wants to know, at what point our capabilities may pose a significant risk. There is no answer to this at the moment, and the agent-based capabilities I just mentioned may be the next threshold, but there is no recognized way to test for them yet.
One possible test is to see if the system has the ability to deceive. If there is deception in the system, then nothing else it reports can be trusted. Therefore,Testing for deception should be a top consideration for emerging capabilitiesThere are many other capabilities that are worth testing, such as the ability to achieve specific goals, the ability to replicate, and there is a lot of work going on. I think these are basically where government agencies are playing a role. I think it would be very good for them to push hard in this area, and of course, the laboratories should also contribute what they know.
Frye: Where do institutions fit into this world that you're describing? Even if we get to the stage where we have AGI that can support all scientific research, will institutions still have a role?
Hassabis: I think there is. In the journey to AGI, I think it will be a collaboration between the community, academia, government, and industrial labs. I really believe that this is the only way we are going to get to this end stage.
Hassabis's test criteria for AGI
Hassabis: If you're asking about what happens after AGI, one of the reasons I've always wanted to build AGI is so that we can use it to start answering some of the biggest, most fundamental questions about nature, reality, physics, and consciousness. Depending on what form it takes, it could be a combination of human experts and AI. I think that will continue for some time in terms of exploring the next frontier.
Currently these systems cannot generate conjectures or hypotheses on their ownAt the moment, they can help you prove certain problems, win gold medals at the International Mathematical Olympiad, and maybe even solve famous mathematical conjectures, but they are not yet capable of proposing hypotheses like the Riemann Hypothesis or the Theory of General Relativity.This has always been my test for true general artificial intelligence.— It will be able to do these things and even invent new theories. We don’t have any systems yet, and we may not even know how to theoretically design systems that can do these things.
Frye: Computer scientist Stuart Russell has expressed to me his concern that once we reach that stage of AGI development, we might all end up living lives of unfettered luxury and without any purpose in life, lives filled with material comforts but lacking any deep meaning or purpose.
Hassabis: That's a really interesting question. This is probably beyond AGI and more like what people sometimes call ASI. We should have enormous resources by then, assuming we can ensure that these resources are distributed fairly and equally, then we will be in a position where we can freely choose how to act, and "meaning" will become a big philosophical question. I think we will need philosophers and maybe even theologians, and social scientists to start thinking about this question now. What brings meaning? I still think self-actualization is important, and I don't think we will all just indulge in meditation and maybe we will play computer games. But even so, is that really a bad thing? It's a question worth exploring.
Although AGI will bring about huge changes, such as curing many diseases or even all diseases, solving energy and climate problems, it may also make us face a deeper question: What is the meaning of life? Just like people climbing Mount Everest or participating in extreme sports, these activities seem meaningless on the surface, but in fact they are people's pursuit of challenging themselves. With the development of AGI, we may have everything on the material level, but with it comes a rethinking of the meaning of life. This question has been underestimated in both the early and late stages of technological development, and we need to re-evaluate the so-called hype and its real impact on our future.
Frye: Let's get back to the question about AGI. I know your big mission is to build AI that benefits everyone. But how do you make sure that it actually benefits everyone? How do you take into account everyone's preferences and not just the designers' preferences?
Hassabis: I think it is impossible to include all preferences in one system because people cannot agree on many issues. I think we may have a secure architecture on which personalized artificial intelligence can be built, and then people can decide what the AI system can and cannot be used for based on their preferences, usage purposes, and deployment purposes. In general, the architecture needs to ensure security, and then people can make some variations and increments based on the architecture.
So I think as we get closer to general AI, we're probably going to have to collaborate more ideally internationally and then make sure that we build general AI in a safe environment.
Once we accomplish this, everyone can have their own personalized pocket API if they wish.
Frye: Okay. But my point is that there are some bad behaviors that AI can exhibit.
Hassabis: Yes, bad emerging behaviors, capabilities. Deception is one example. We have to understand all of these issues better.
There are two things to worry about: one is that humans might misuse AI, and the other is AI itself (as it gets closer and closer to AGI, its performance goes off track). I think these two problems require different solutions. Yes, this is the problem we have to deal with as we get closer and closer to building AGI.
Going back to your point about benefiting everyone, using AlphaFold as an example, I think we could cure most diseases in the next year or two if AI drug design works. And then they can be turned into personalized medicines to minimize side effects for the individual, which is related to the person's personal disease and personal metabolism and so on. So these are all amazing things, you know, clean energy, renewable energy, technology is going to have huge benefits, but we also have to mitigate the risks.
Frye: You said one way you wanted to mitigate risk was that one day you'd basically do a scientific version of "Avengers Assemble"?
Hassabis:certainly.
Frye: So, how do you know when the right time is?
Hassabis: Well, that's a big problem. You can't do it too early because you'll never be able to win over some of the naysayers. Today, you see some very famous people saying there are no risks in AI. And then people like Geoffrey Hinton say there are a lot of risks.
Frye: I wanted to talk to you a little bit more about neuroscience. How much does it still inspire what you’re doing? Because I noticed the other day DeepMind unveiled a virtual mouse with an artificial brain, which is helping to change our understanding of how the brain controls movement. I remember we talked a lot about how we can get direct inspiration from biological systems, is that still core to your approach?
Hassabis: No, it's now developed, and I think we're already in the engineering phase, like large systems, large-scale training architectures. Neuroscience has had a little bit of an impact on that. Neuroscience is one of the sources of ideas, but when the engineering is large, neuroscience takes a back seat. So now it's probably more about applying AI to neuroscience. I think as we get closer to AGI, understanding the brain will be one of the coolest use cases for AGI.
Frye: I wonder if you're also envisioning that there will be things that are beyond human comprehension that AGI will help us discover, understand?
Hassabis: I think it's possible that an AGI system could understand higher levels of abstraction than we do. I think an AI system could effectively have any kind of prefrontal cortex, so it's conceivable that higher levels of abstraction and patterns could be seen in the universe that we can't really understand or remember right away.
And then I think that from an explainability perspective, we can't infinitely scale our own brains, but in theory given enough time, SPEs, and memory, AGI could understand anything computable.
Frye: You said DeepMind was a 20-year project. How close are you to getting there?
Hassabis: We are on the right track.
Frye: Will AGI be achieved by 2030?
Hassabis: I wouldn't be surprised if it comes out within the next decade.
Collect and organize the latest news information and make it available for free for everyone to read
