DeepMind Founder Demis Hassabis Discusses AGI Timeline and Challenges

On April 8, a new half-hour interview with Demis Hassabis, founder of DeepMind, was released. In the interview, Hassabis expressed a strong belief that the possibility of achieving AGI within the next five years is very high. He noted that approximately 90% of the key breakthroughs supporting the modern AI industry over the past decade or even fifteen years have come from Google Brain, Google Research, or DeepMind. He stated, “If there are any missing key breakthroughs in the future, we have the capability to achieve them.”

Regarding the commercialization of model capabilities, Hassabis believes that the gap between leading labs is starting to widen, making it increasingly difficult to extract value from the same ideas. Labs capable of inventing entirely new algorithmic concepts will gain a significant advantage in the coming years, as previous ideas have been “squeezed dry.”

In the video, Hassabis engaged in a deep discussion with host Harry Stebbings on core topics such as the timeline for AGI realization, technical bottlenecks, model commercialization, the future of open-source, the post-large language model era, and whether AI can truly solve drug development issues. He shared insights on the reasons behind DeepMind’s progress and future plans, as well as his first impressions of meeting Elon Musk.

01. Achieving AGI in Five Years: The Biggest Bottleneck is Computing Power

Host: What does AGI mean to you today? This can serve as our starting point for discussion.

Demis Hassabis: Our definition has been very consistent: AGI is a system that possesses all cognitive abilities of the human mind. We use this standard because the human brain is the only known instance of general intelligence in the universe. So for me, this is the benchmark AGI must reach.

Host: How far are we from AGI? There are various opinions in the industry, with some predicting it could be achieved as early as 2026 or 2027. What do you think?

Demis Hassabis: The possibility of achieving AGI within the next five years is very high.

Host: Is this closer than you initially thought? Has your judgment changed over time?

Demis Hassabis: Not really. My co-founder and DeepMind’s chief scientist, Shane Legg, often predicted the timeline for AGI back in 2010 when we founded the company. At that time, very few people took AI seriously; it was seen as a dead end. Those blogs are still online, and anyone can check them. We extrapolated based on computing power and algorithmic advancements, predicting it would take about 20 years from the start. Now, it seems we are progressing as planned.

Host: From today’s perspective, what is the biggest technical bottleneck?

Demis Hassabis: I believe computing power is the biggest bottleneck. This is not only due to the “scaling law”—you need to continuously build larger architectures with more parameters to achieve smarter systems. Another area that requires substantial computing power is experimentation. Computers and clouds are our workbenches. If you have a new idea and want to test it, you need to validate it at a reasonable scale. Therefore, if you have many researchers and many new ideas, you need abundant computing power.

Host: You mentioned the “scaling law.” Many believe we are reaching the limits of this law, and performance improvements are starting to show platform effects. Do you agree?

Demis Hassabis: No, I don’t think so. I believe the reality is more nuanced. Of course, when leading companies began building large language models, each new system brought significant performance leaps. This exponential growth will inevitably slow down at some point. However, this does not mean that further scaling existing systems will not yield good returns. We and other leading labs are still obtaining very considerable returns from scaling power. Although obviously less than during the early scaling phases, the returns remain substantial.

Host: In what areas have we actually fallen behind your initial expectations?

Demis Hassabis: Honestly, in most areas, we are ahead of what I expected. You can look at video generation models or even our latest system, Genie, which is an interactive world model. If someone had shown me these things five to ten years ago, I would have been very shocked. So, in most areas, we are ahead of the initial expectations. However, there are still significant gaps, such as “continuous learning,” meaning that current systems do not learn new things once they are trained and deployed in the real world.

02. Continuous Learning is One of DeepMind’s Next Steps

Host: Nowadays, when researching and preparing new programs, DeepMind has become my first choice. But this was not the case two or three years ago. What do you think has driven this acceleration and progress at DeepMind?

Demis Hassabis: We have indeed made some organizational adjustments. In fact, Google and DeepMind have always had the deepest and broadest research reserves in the industry. If you look back over the past decade or even fifteen years, about 90% of the breakthrough results supporting the modern AI industry have come from Google Brain, Google Research, or DeepMind, such as AlphaGo, reinforcement learning, and of course, the Transformer architecture. These are all key milestones.

Therefore, I believe that if there are any missing key breakthroughs in the future, we have the capability to achieve them. We have essentially gathered all the top talents in the company to work towards the same goal. Additionally, we have consolidated all computing resources to build the largest models rather than parallelizing two or three different versions within the company. So I believe that, to a large extent, we are assembling all the elements we already possess and pushing forward with a focus and speed akin to a startup, thus returning to the technological forefront and maintaining leadership in many areas.

Host: You said that if anyone is to make a breakthrough, it should be DeepMind. Do you see continuous learning as your most anticipated next breakthrough?

Demis Hassabis: I think there are still many things missing. Continuous learning is one of them. Additionally, researching different memory systems has significant potential. Currently, we mainly rely on long context windows to shove all information in, which is somewhat “brute force.” I believe there are many interesting architectures that can be invented in this area. There are also long-term planning and hierarchical planning. Current systems are not good at handling long-term planning, such as matters spanning many years. However, the human mind can do this. So there are many problems to overcome. Perhaps the biggest issue is that they perform very well when asked specific questions in a certain way, but if you change the way you ask, they may even fail at very basic things. General intelligence should not be like that. I call this phenomenon “Jagged Intelligence.”

03. Strong Support for Open-Source Models

Host: Many in the industry are discussing the commercialization of model capabilities. Do you think we will see that situation? Or will one or two labs continue to accelerate, leaving others behind?

Demis Hassabis: I believe that currently, three or four leading labs—of which we are one—are starting to widen the gap between each other. The reason is that many existing tools (such as coding tools and mathematical tools) will help build the next generation of systems. I also believe that extracting value from the same ideas will become increasingly difficult. Therefore, labs capable of inventing entirely new algorithmic concepts will gain a significant advantage in the coming years, as previous ideas have been “squeezed dry.”

Host: My next question is, over the years, many of your studies at DeepMind have been quite open, and we have seen many high-quality open-source models. What do you think about the future of open-source?

Demis Hassabis: I believe it will likely resemble what we see now. We have always been strong supporters of open science and open-source models. From the early Transformer to AlphaFold, we have done a lot of work to share these results with the world to help the research community. We plan to continue doing this, especially in application areas, such as applying AI to science, which is clearly my personal passion. However, I also believe that you will increasingly see that open-source models may lag behind the cutting-edge models. Typically, the open-source community takes about six months to reimplement and understand those new ideas. Nevertheless, we are also actively promoting an open-source model called Gemma, determined to make it the best in its class at its respective scale. It is an ideal choice for small developers, scholars, or nascent startups, and is also suitable for edge computing. Thus, we are very optimistic about open-source models for certain types of applications.

04. Future AGI Requires Global Regulation

Host: Next, I would like to ask you how you view the world after large language models. Different scholars have very different opinions, such as Yang Likun, who holds a very different view.

Demis Hassabis: Frankly, I do not agree with Yang Likun on some issues. I believe there is still a 50% chance that there are some missing key elements, and we still need breakthroughs in areas like world models. However, I am very confident that foundational models have proven their tremendous success. They can perform impressively on extremely challenging tasks, and I do not believe that this capability will disappear. We continue to obtain returns from the scaling law. Therefore, the real question is: when we look to the future of AGI systems, are LLMs (large language models) the only key component, or are they part of the overall system? My judgment is that they will not be replaced but will become the foundation for building upon, just like what we are doing with world models.

Host: As you mentioned, AGI may emerge at that time. So when we look five years into the future, what will that world look like? Many people have expressed concerns from different angles. Let’s start with the positive side. What do you think that world will be like?

Demis Hassabis: I believe the positive aspect, which is also the original intention of my entire career dedicated to building AGI, is that it will ultimately become the most powerful tool in the fields of science and medicine. We urgently need such technology to drive scientific discoveries and find treatments for diseases. Therefore, I hope that in a little over five years, we will usher in a golden age of scientific discovery.

I believe we will soon be able to approach that goal. First, after completing the AlphaFold protein folding project, we spun off a company—Isomorphic Labs—that is currently doing very well. Its core idea is to focus on solving the remaining aspects of the drug discovery process, including extensive chemical work, compound design, toxicity testing, and various property evaluations required for drug safety. We expect that within the next five to ten years, the entire drug design engine will be ready.

The next bottleneck is clinical trials, which still require many years. However, I believe AI can help, such as simulating certain parts of human metabolism and precisely stratifying patients to ensure that specific patients receive drugs best suited to their genomic makeup. Therefore, AI can also play a valuable role here. However, I believe the real revolution may occur after several AI-designed drugs successfully complete the entire process. At that point, governments and regulatory agencies will see these results and have enough data to backtrack and verify model predictions. Perhaps another ten years later, we can truly trust these model predictions, allowing us to skip certain steps, such as no longer needing animal trials or accelerating dosages because the reliability of the models has been validated. Thus, I believe we must take a two-step approach: first tackle the drug design problem, then address the timing issues in the regulatory process.

Host: Speaking of regulation, AI safety is undoubtedly a major topic that has raised widespread concerns. I remember Stephen Hawking once said: we must get this right because we may not have a second chance. Do you agree with his view?

Demis Hassabis: I completely agree. I think this is precisely the risk we face. I am primarily concerned about two things: first, malicious actors abusing these systems. Second, the technical issues: in a year or two, when these systems become more embodied and autonomous, as we gradually move towards AGI, can we keep them on the expected safe track? I believe appropriate regulation can help ensure that all leading providers meet at least minimum safety standards, but ideally, this requires unified standards at the international level.

Host: So what kind of regulation is “appropriate”? Quoting your words in the documentary, you mentioned, “We need more global coordination,” which makes me worried because, in fact, we are doing worse in global coordination.

Demis Hassabis: Indeed. We are in an extremely special period. This technology could be the most influential technology in human history, while the international system is highly fragmented. This is clearly not an ideal state. However, we must still do our utmost to at least establish a set of minimum standards and several benchmarks to test the adverse properties of systems, such as “deception.” No one should build systems capable of deception, as that would allow them to bypass other safety measures. If all goes well, we can establish some sort of certification mechanism, similar to a “quality mark,” indicating that the model has specific safety protections and performance guarantees, allowing consumers and companies to build safely on top of it. I believe this is the ideal direction for development. Moreover, this must be international, as these systems inherently possess transnational and cross-regional characteristics.

Host: So who will serve as the arbitrator?

Demis Hassabis: I believe the ultimate responsibility must lie with governments. However, organizations capable of undertaking specific technical work can be institutions like AI safety research institutes. The UK has a very excellent AI safety research institute established during the tenure of former Prime Minister Sunak, which I believe is doing well. The US also has similar institutions. Perhaps those major countries with top research capabilities should establish equivalent institutions, equipped with high-quality researchers capable of assessing and auditing these systems against specific benchmarks and independently verifying whether they meet appropriate standards.

Host: If I could give you a magic wand applicable only to AI safety, what kind of ideas or plans would you implement?

Demis Hassabis: I believe we need some sort of international organization, perhaps similar to the International Atomic Energy Agency. AI safety research institutes from various countries can provide input to it, and the research community must also participate in determining which benchmarks are appropriate and which properties and capabilities need to be checked.

Additionally, there may be other safety measures, such as not allowing AI systems to output non-human-readable markers, like some machine language we cannot understand. I believe that would introduce new security vulnerabilities. Then, these international agencies would test for the above matters. I believe this will give the public confidence, and the academic community and civil society can also participate to ensure that those systems that will become extremely powerful receive independent scrutiny.

05. Overhyped and Undervalued in AI

Host: When you see the true capabilities of these systems, how do you view the issue of labor replacement? What does this mean for the labor market?

Demis Hassabis: Undoubtedly, every revolutionary new technology in history has led to significant disruption of jobs. This is certain, and I believe this time will be no exception. Many old jobs will disappear or become economically unviable. However, history also tells us that a whole new set of professions will emerge as a result. These professions were previously unimaginable and are often high-quality, high-income jobs. This is a conventional evolutionary process. Of course, we must be very cautious in judging whether “this time is truly different.” People like Marc Andreessen believe that this time is not fundamentally different from the past ten major breakthroughs like the internet and mobile communications. However, I do believe that the impact this time will be greater than any previous technological breakthrough, with a scale ten times that of the Industrial Revolution and a speed that is also ten times that of the Industrial Revolution. In other words, it will unfold within a decade rather than a century. I have read many books about the Industrial Revolution; that revolution brought significant turmoil but also great progress. Ideally, this time we should better mitigate those negative effects than during the Industrial Revolution.

Host: Someone told me that we always overestimate what can be done in a year and underestimate what can be done in ten years. Does this judgment still hold here, or is the pace of technological development actually faster than we imagine?

Demis Hassabis: No, I believe this judgment still holds. I mean, perhaps the time scales for both the short and long term are a bit closer than for other technologies. However, I do believe that, looking at today and the next year, there is some overhype in the AI field; from certain perspectives, there is no more room for hype. Interestingly, on the other hand, I think that in the ten-year time scale, its revolutionary nature is still severely underestimated. We can call this the long term. Even in today’s AI field, this dichotomy still exists.

Host: Besides concerns about the labor market, there are also worries about income inequality and wealth concentration among a few participants. Given your comments on the Industrial Revolution, how do you think this will evolve?

Demis Hassabis: I believe there are different possible paths. For example, pension funds should buy shares in all major AI companies to ensure everyone benefits. Perhaps every country should establish a sovereign wealth fund to do this. This is an investment-level solution.

Additionally, I think we need to consider how to redistribute and ensure that everyone benefits if this tremendous increase in productivity occurs only in a narrow field. I can think of various ways, such as using the additional productivity gains to provide infrastructure and other public services. In the five to ten-year time scale, there may be incredible breakthroughs, such as perhaps we have solved the nuclear fusion problem, which we are working on with Commonwealth Fusion partners. I believe AI will lead us to entirely new possibilities: excellent superconductors, more efficient batteries, and breakthroughs in materials science. All of these will fundamentally change the nature of the economy.

Host: So how do we address the energy crisis brought about by the AI revolution? Its scale in energy demand is unprecedented.

Demis Hassabis: I believe that in the medium to long term, AI will pay for itself in energy costs, even exceeding that. We are undertaking a series of projects to optimize existing infrastructure, such as optimizing the power grid. I believe we can enhance the national grid’s efficiency by about 30% to 40%. Additionally, there are climate and weather modeling projects; we have the best weather modeling system in the world, which helps pinpoint where impacts will occur and take mitigation measures. Finally, the most exciting breakthroughs may involve nuclear fusion, new batteries, and superconductors, and AI is crucial to helping us achieve these goals. By then, humanity will enter an entirely new energy landscape, which will certainly help address climate and environmental issues and ultimately help us access space at lower costs. Because if you have nuclear fusion as an incredible energy source, you essentially have almost unlimited rocket fuel, simply by distilling and catalyzing seawater.

Host: I’ll bring out that magic wand again. What would you do to cultivate a growth mindset, a capability to establish trillion-dollar companies that we do not yet have today?

Demis Hassabis: We are very good at generating entrepreneurial ideas and bringing them to a certain level, just like we did with DeepMind. However, if you really want to cross that chasm and become a trillion-dollar global player, where does the multi-billion dollar financing round come from? That enables you to truly challenge existing legacy companies. I believe this was certainly missing when I was fundraising for DeepMind ten years ago, and I think it is still somewhat missing today: that level of ambition and the amount of funding that capital markets can support.

06. First Meeting with Musk Was Great

Host: Let’s do a quick Q&A. What was your impression when you first met Elon Musk?

Demis Hassabis: It was great. That was at an event hosted by Founders Fund. At that time, both SpaceX and DeepMind were part of the same investment portfolio. I think we were both invited guests; that should be my first time attending an investment portfolio meeting, probably around 2011 or 2012, when we were still an inconspicuous upstart, only getting a small speaking slot. Musk was the core figure in that portfolio, and he gave the keynote speech. Later, we met after the meeting. He joked that we had greeted each other while passing in the restroom. We immediately hit it off, like those ambitious thinkers who also love science fiction. I was very eager to visit his rocket factory, so I tried to secure an invitation to SpaceX. By the end of that meeting, he indeed extended an invitation. Our second meeting was at the SpaceX factory in Los Angeles.

Host: So, what medical revolution are you most looking forward to?

Demis Hassabis: Honestly, I want to truly cure cancer. I know this sounds cliché, but what we are building at Isomorphic is general. We are trying to create a drug design platform applicable to any therapeutic area. So ideally, it will cover all areas, from neurodegenerative diseases, cardiovascular diseases, immunology to cancer. These are our current priorities, but ultimately, it should be applicable to every disease.

Host: Is there anything you are thinking about that others have not yet read or discussed?

Demis Hassabis: Many people are concerned about the economic issues brought about by AGI that we discussed earlier. However, I am very worried about the philosophical issues behind it. For instance, assuming we get the technology right and handle the economic aspects, then what remains are philosophical questions: What is meaning? What is purpose? We will explore what consciousness is and what it means to be human. I believe these questions are about to confront us. We need some great new philosophers to help us find direction.

Host: Finally, here’s a somewhat difficult question. There are many ways to describe what you are doing. How would you most like to be remembered? What kind of legacy do you hope to leave behind?

Demis Hassabis: I hope my legacy is advancing scientific progress and creating technologies that bring great benefits to the world, such as curing those terrible diseases.