Taole from Aofei Temple

Quantum Bit | Public Account QbitAI

When will AGI be achieved? GoogleDeepMind CEO Hassabis predicted in a recent interview:

ten years.

ifAGII wouldn't be surprised to see it appear within the next decade.

In this interview, Hassabis talked a lot about the development trend of AI.

There are some topics that are of interest to everyone, such as:

AI is overhyped in the short term, but its potential and impact are still underestimated in the long term

Google DeepMind, as Google's "super team", needs to find a balance between commercial interests and scientific research

Open source is important for technological progress, but when you encounter technology that could be abused, you may need to wait a year before making it open source in order to evaluate and limit abuse.

For more details, please see the text version below~

Early AI “landing issues”

Q: For today’s interview, I invited Hassabis, the co-founder of DeepMind and current CEO of Google DeepMind.

Hassabis: Thank you, it's great to be with you again. Previously we talked about how concepts and language can be used in the real world, such as in simulations or as actual intelligence in robots, which may be necessary for us to understand the world around us.

While it’s important to acknowledge that these systems aren’t there yet, they make a lot of mistakes and haven’t really built a complete model of the world, they’ve made more progress than we could have ever imagined, just through language learning alone.

Q: Yes, last time we talked a lot about how to implement language in the real world. But can you briefly explain what "implementation" means? In case those who are watching our interview for the first time don't understand?

Hassabis："Landing"This concept,It originates from classic AI systems built in academic institutions such as MIT in the 1980s and 1990s.Most of these systems are huge logical systems, which can be imagined as giant databases containing many interrelated words.

However, the problem is that although you can find statements such as "dogs have four legs" in the database, when the system faces a real photo of a dog, it cannot match these pixels with the symbols in the database. This is the so-called "landing problem."

In other words, the system had symbols or abstract representations, but it could not understand what those representations meant in the real world. Since then, despite attempts to solve this problem, perfection has never been achieved.

Unlike systems of the past, today’s AI systems learn directly from data and make connections between that data and the real world from the outset.

Interestingly, even though these systems initially learn only based on language, they should theoretically lack a lot of the information needed to “go to the ground” because they are not connected to simulators, robots, or other forms of input in the real world, but only learn in the language space.

Yet, surprisingly, these systems are still able to infer some knowledge about the real world from them.

Q：This may be because, in the process of interacting with the system, people will tell the system which answers are correct and which are wrong. Through this kind of feedback, the system can receive some "grounding" information, thus gradually establishing a connection with the real world.

Hassabis: Indeed, if the system gave incorrect answers in early versions due to lack of "grounded" information, such as incorrectly answering a question like "How does a dog bark?", then people's feedback would correct it.

This feedback is based on our own "grounded" knowledge, so, to some extent,The system absorbs and learns from this feedback.

Hyped in the short term, undervalued in the long term

Q: In addition, I also want to ask you about the hype of AI. Do you think that our current situation, the current situation, is under-hyped or over-hyped?

Hassabis: I think it's more of the latter. I would say,In the near term, the hype is a bit overdone.

I think people say AI can do all kinds of things, but in reality it's not as powerful as people say it is. There are a lot of startups and venture capital chasing those unrealistic ideas, and their ideas are not mature.

But on the other hand, I think even now, it's still underestimated or not taken seriously enough. Especially as we get to AGI and post-AGI, I still feel like people don't fully understand how big of a change this is going to be and the responsibilities that come with it.

so,I think in the short term it is a bit overhyped, but in the long term it is still undervalued.

From multimodal Gemini to Astra general purpose AI agent

Q: Okay, now I want to ask a more important question in this interview. How do you think Gemini is different from large language models released by other laboratories?

Hassabis: From the very beginning of the Gemini project, we set a goal to enable it to process multiple types of data at the same time. We hope that it can not only understand text, but also parse sound, video, pictures, and code, basically all forms of information.

We firmly believe that only by enabling systems to understand and process various information in the real world can they truly understand the world and build a more accurate and comprehensive world model.

This is actually an extension of the "landing" problem we mentioned above, but this time we use language as the basis to achieve this.

Q: So “implementation” is still the key to the entire project?

Hassabis: Indeed, this is crucial.

We also have an ultimate goal, which isBuild an all-round assistantWe have developed aAstroAstro is a prototype of a project that can not only understand your input, but also perceive the context of your environment.

Imagine if your personal assistant or digital assistant could be more helpful if it had a deeper understanding of the context of your question or the situation you're in. So we've always believed that this type of system will be more useful.

To this end, we built multimodality into the system from the very beginning of the project. At the time, it was the only model that had this capability, and now other models are trying to catch up.

Q：Project Astro is an emerging general-purpose AI agent capable of processing both video and audio data.I remember at Google I/O you showed an example of how Astro could help you remember where you put your glasses. I'm interested in the origins of this technology, and whether it's just an advanced evolution of the old Google Glass?

Hassabis: Google has a long history of developing glasses-like devices, having been involved in the field as early as around 2012. Therefore, they have a significant first-mover advantage in this field. It may be that they lacked the necessary technology to enable the smart assistant to understand what it sees. But now,With this digital assistant, it can understand the world around you, which makes people feel very natural.

Q: I want to trace the origins of Gemini because it came from two independent departments within the organization. Was it done by them together?

Hassabis: Indeed, last year we moved Alphabet(Google's parent company)The two research departments ofDeepMindandGoogle Brain, merged into a new department, which we call"Super Sector"。

In this way, all the top talents in the company came together in one team, and we combined the best knowledge from all research fields, especially in language models.

We have some previous projects, such as Trin Chilla and Gopher, which have participated in the development of early language models such as Palm and Lambda. These models have their own strengths and weaknesses, and we have integrated them into the Gemini project, which is the first very important project after the merger.

In addition, there is another particularly important point, that is, weIntegrate all computing resources togetherThis allows us to do large-scale training runs, essentially pooling all the computing power together.Making Gemini more powerful and efficient。

The two teams have always been at the forefront of AI, and have collaborated on individual research, but perhaps not so closely in strategy. I would describe the combined team as Google’s"Engine Room"。

I think the working methods of the two teams are actually quite similar, with little difference. Next, we will continue to strengthen our advantages in basic research, such as thinking about the nextTransformerWe all want to figure out what the architecture will be like.

Speaking of which, the previous Transformer was developed by Google Brain, and we combined it with Deep Reinforcement Learning, but I think there needs to be more innovation. I believe that, just like the past 10 years, whether it is Brain or DeepMind, we will continue to contribute.

Google's "Engine Room"

Q: You just said that Google DeepMind is now the "engine room" of Google, which is a big change. I wonder if Google is betting big on you now?

Hassabis: I think so. I think Google has always been aware of the importance of AI. When Pichai first became CEO, he said,Google is an “AI-first” company。

We discussed this topic when he first took office, and he saw the potential of AI as the next major paradigm shift after mobile Internet, and its impact is even more far-reaching than that.

But I think in the last year or two, we've really started to understand what that means, not just from a research perspective, but also from a product and other aspects. So, it's a very exciting time, but I think this is the right choice for us to integrate all of our talents and go all in.

Q: From another perspective, for DeepMind, now becoming the "engine room" of Google, does it mean that you have to find more balance between considering commercial interests and pure scientific research?

Hassabis: It is true that we now need to consider commercial interests more, which has become part of our responsibilities. However, there are still some points that need to be clarified. We will continue to advance our scientific research work and our investment in this regard is still increasing.

I think,This is one of the unique things we do at Google DeepMind., even our competitors see these results as a general benefit of AI.

Open Source is Necessary

Q: Okay, that brings me to my next question: open source. When technology is in the hands of the masses, as you said, some really amazing things can happen. I know DeepMind has open-sourced a lot of its research in the past, but it seems that this is changing now. Can you talk about your views on open source?

Hassabis：Open source is necessary, and we have always been a firm supporter of open source and open science.As you know, we have published almost all the research projects we have done, including Transformer andAlphaGoWe have published projects like this in top journals such as Nature and Science.

AlphaFold is also open sourceThese are what we madeChoose wiselyYou are right, this approach works because technology and science advance at the fastest rate possible through the sharing of information. In most cases,Open source is universally beneficial; it’s how science works。

There are exceptions, though, and that’s when it comes to technologies that have dual uses, like AGI and strong AI.

The problem is that you want to enable all the good use cases, and you want real scientists and technologists to build and critique these ideas so that society can move forward quickly, but at the same time, you limit the bad guys who might abuse these systems, and that's the problem.

It's OK now because I don't think the systems are strong enough, but in two or three years, especially when you start getting systems with agentic behavior, it could cause serious harm.

We have our own open source Gemini models called Gemma, but they are smaller models and not the most cutting-edge models.

Their capabilities are still very useful for developers because they can run on laptops and because they have a small number of parameters. Their capabilities are well understood at this stage because they are not the latest cutting-edge models.

Probably where we'll end up is we'll have open source models, but they'll lag behind the latest cutting-edge models by a year or so so that we can really evaluate the capabilities of those models in public testing with users.

One of the problems with open source is that if something goes wrong, you can't take it back. With a proprietary model, if bad guys start using it in bad ways, you can shut it down, or in extreme cases shut down the entire system. But once you open source something, you can't take it back. It's a one-way door.

AGI will be achieved within ten years

Q: Currently, various AI models are developed by scientific researchers, but I want to know, if we enter the stage where AI supports all scientific research, will there still be room for R&D institutions?

Hassabis: I think there is still room for it. We are now in the stage before the emergence of general AGI, and I think it will require close collaboration between society, academia, government and industrial laboratories.

I really believe that this is the only way we're going to get there. If you're asking what comes after AGI, that's probably what you really want to know. AGI has always been something I've been eager to build because we can use it to explore some of the most fundamental questions about the nature of reality, physics, consciousness, and so on.

Q: Computer scientist Stuart Russell once told me that he was somewhat worried that once we achieve AGI, we might all be like the aristocrats of the past, enjoying a carefree and luxurious life without any goals or thoughts.

Hassabis: I think it's going to be very interesting. But it also gets to the "undervalued" issue that I mentioned earlier, the difference between the near-term and the long-term hype. If you want to call it hype, then it is indeed undervalued to a certain extent.

I think the changes in the future are going to be huge. I believe that eventually we will be able toCure many diseases, even all diseases, and solve energy and climate problems.

Q: Speaking of which, I remember you once said that you hope AGI can explore the mysteries of the universe. Do you think there are some possibilities that we have not yet imagined, such as phenomena such as wormholes?

Hassabis:certainly,I totally believe in this possibility.I really hope that wormholes become a reality. It seems to me that we still have many misunderstandings about physics and the nature of reality.

Obviously, there are countless unsolved mysteries hidden in the unification of quantum mechanics and gravity, the problem of the standard model, string theory, etc. I have had in-depth discussions with many friends in the physics community, and they all believe that there are many things in the existing theoretical framework that cannot fit together perfectly.

I'm personally not too keen on the multiverse explanation, so I think it would be really amazing if we could come up with new theories and test them using large-scale equipment in space.

The reason I'm so fascinated by the Planck scale of time and space is that it seems to represent the ultimate resolution of reality, like the smallest unit into which everything can be divided.

Therefore, I think we should conduct in-depth experimental exploration at this level, especially when we haveWith AGI and abundant resources, perhaps we can design or build such experimental equipment.

Q: You once said that DeepMind is a 20-year project. How far have we come now? Are you still on track?

Hassabis: Yes, we are still on schedule, which may sound strange because usually a 20-year project always feels like it will take another 20 years to complete. But we have made a lot of progress.

Our goal is to have it done by 2030, so I wouldn’t be surprised if we can get there within the next decade.

• Video link: https://www.youtube.com/watch?v=pZybROKrj2Q *