Thank you for having me. It is an honor to be here. To start, I think I should list my relevant qualifications and experience. I am a human, a one-of-a-kind version of myself, and it took me nearly 40 years to accumulate the necessary experience for this speech. I will return to those points later.
My entire career has been dedicated to finding information that seems important to the world's future, and then conveying that information to the public. Two decades ago, I started as a journalist, covering data centers. It was that work that made me so fascinated by artificial intelligence that by 2016, I left journalism to join OpenAI. A few years later, I co-founded Anthropic.
Since then, I have been trying to tell the world that one day, AI systems would become powerful enough to alter national security, change the economic trajectory of the entire planet, and eventually reach the ability to build new AI systems themselves. And now that moment has arrived. AI is changing national security. AI is changing the economic development trajectory of the world. As we have heard in many sessions here, AI is also contributing to its own development process.
Many of you already know this. You are having to track AI as one of the countless destabilizing forces shaping your world and your work. And what I want to explain today is that AI is on the verge of creating changes even greater than what we currently anticipate. We need to prepare commensurately for that. I will explain why.
For many years, researchers and science fiction authors have predicted a future where AI systems would sharply increase the speed of scientific and technological progress in general, making everything happen so fast that humans can hardly keep up. They call this phenomenon the technological singularity. And I believe that this year we have obtained compelling evidence that this process has truly begun.
In 1993, a writer predicted the technological singularity. He said, 'When human-surpassing intelligences drive progress, the speed of that progress will be much faster.' In fact, there seems to be no reason to assume that this very process of progress would not lead to the creation of even more intelligent entities in increasingly shorter periods of time. He described the singularity as a point where we are forced to abandon existing predictive models because a new reality has emerged and dominates everything.
The technological singularity has arrived. Its impacts will reach every aspect of life. But the most important consequence is quite simple. We should prepare for the possibility that in the coming years, the overall pace of progress—not just in AI, but in science in general—will accelerate. This will leave us feeling bewildered and disoriented, bringing immense prosperity, but potentially accompanied by serious dangers.
Therefore, today I will focus on the former while not downplaying the remaining risks. I will talk about the benefits that rapid AI progress can bring, as well as what we need to do to share those benefits together. Many of the benefits that humanity will receive from AI will come from science. I believe that AI-driven scientific breakthroughs will shift the trajectory of the global economy, change how nations plan and build infrastructure, and help us all be healthier and have more time with our loved ones.
But at the same time, that progress will also force us to ask profound philosophical questions about the choices we must make regarding the future development of technology. To clarify this point, for the past half-century, most countries have faced declining research productivity. Although the number of researchers has increased rapidly, finding new ideas and inventions seems to be becoming increasingly difficult.
But now, something important has changed. We have built AI systems capable of supporting, augmenting, and automating parts of human work. This is most evident in the field of software development, where AI systems now have the capability and performance that surpass most software engineers in the world, and the result is that the field of software development is changing. Now, an individual can write software for anything they imagine, creating products that previously required a whole team working for months.
We have seen concrete evidence of this right inside Anthropic. As you can see here, our researchers are now producing nearly 10 times more code than they did in the company's first four years. This increase in code volume essentially correlates directly with advancements in AI capabilities.
However, new AI capabilities will also bring new challenges. Improvements in coding ability mean that AI systems have become much better at cybersecurity, but also better at carrying out cyberattacks, as we have heard. This is leading the world into a race to deploy those powerful capabilities to build cyber defense systems. And malicious actors are also racing to use these same AI systems to conduct attacks and unauthorized intrusions.
And it is important to realize that these two aspects are closely linked and inseparable. The advancement of AI systems in a general field like programming will create both positive and negative impacts in areas of international importance, such as cybersecurity. This is a foreshadowing of what is to come. What is about to happen is that progress brings new capabilities but also creates new challenges, and these capabilities do not stop at programming. The same will happen in science and robotics. In fact, that process has already begun.
This means that some key features of the global economy could change. For instance, the pace of innovation or the frequency with which we invent entirely new general-purpose technologies, capable of shaping entire industries around them, just like the internet, the telephone, and artificial intelligence.
So what makes me believe this statement? And what will I do to help you believe in it as well? In 2026, we witnessed the first signs that AI systems could conduct scientific research at a pioneering level. In mathematics, AI systems have been used to propose solutions to long-standing problems, such as the Navier-Stokes millennium prize problem, the Jacobian conjecture, and several unsolved Erdős problems. In many cases, it was the AI systems themselves that provided the core breakthrough ideas that made finding solutions possible.
In biology, researchers have used AI to design viruses from scratch and identify new drug candidates for testing. There are now drugs in the testing process for which the majority of the design work was performed by AI systems. In the field of robotics, AI systems are increasingly capable of operating laboratory equipment autonomously. This allows us to continue automating experiments across many fields, from biology to materials science, thereby accelerating progress in areas like battery or fuel cell development.
Taken individually, each example is remarkable, but what is truly profound is that they all relate to the same underlying technology. And in most of these applications, they even use the same advanced AI model as their foundation. Now, we have single entities—AI systems like Claude, ChatGPT, or Gemini—capable of performing research work at the forefront of many different scientific disciplines. We possess multidisciplinary polymaths in digital form.
We can also look at further quantitative evidence. One of those metrics is the scores of AI systems on a test called Humanity's Final Exam. I am not the one who gave it that name. That somewhat intimidatingly named test was developed after witnessing the remarkable strides AI made in previous years, as these systems continuously conquered many difficult and diverse scientific challenges. The test covers a wide range of subjects, from chemistry, computer science, and mathematics to classical literature and ecology.
When released in January 2025, the highest-performing AI systems only achieved a score of 8%. As of this week, the highest-scoring system is Claude Opus at 57%.
Another example is FrontierMath. The test was released in November 2024 to evaluate the AI's ability to solve hundreds of difficult math problems. These problems were selected and vetted by mathematicians. When the test first launched, the top systems only solved 2%. Now, that figure has reached 94%.
And looking at the bigger picture, the point I want to emphasize is that we are seeing this same pattern everywhere. I have a file on my computer where I have collected 100 charts just in the past week alone. This morning, before coming here to speak, I even uploaded more of these charts. If you select different scientific fields, you will see the same trend emerge. The rate of progress is accelerating, and that acceleration has begun in the last two years, sometimes even just within one year.
I will conclude this evidence presentation with the latest example from a research paper published last week. This is what I usually do for fun. I read research papers in the time between my conference presentations. This research paper is titled 'Theorem of Evil'. This is a study evaluating the ability of AI systems to solve previously unpublished theoretical physics problems. The 50 questions in this study were compiled by physicists around the world from their own research, with answers that had never been published.
It is no surprise that AI systems achieved high scores. But what is surprising is the way they achieved those scores. Now I would like to quote one of the authors who created this test. Indian physicist Suvrat Raju, who posted his commentary on the results on Facebook. He wrote: 'I don't know any individual who could solve a significant proportion of our final problem set without AI assistance.'
He said our initial assessment showed GPT-6 Astra scored above 80%, closely followed by Opus 5.5. But when we started checking for errors the models might have made, the models—especially Astra—insisted they were right and the answers provided by the experts were wrong. When we went back to consult the experts, it turned out that in many cases, the experts themselves discovered very subtle errors hidden deep in their calculations. And as they corrected the answers, Astra's score began to climb, eventually reaching 98%. 49 out of 50 answers.
Think about this. Some of the world's leading scientific experts constructed a test from problems that the AI had never had the chance to see the answers to. And when they tested the AI, the system not only scored highly but also boosted its score by finding errors in the very answers the humans had created.
In movies, about two-thirds of the way through, there is usually a scene where the young protagonist finally proves to their mentor that they are capable enough to step out into the world and succeed on their own. It is a familiar storytelling trope where the teacher becomes the student and the student becomes the teacher. And that is exactly what happened last week.
So why am I telling you this story? Why are AI systems that are increasingly good at scientific research so important? I believe the reason is that truly conducting pioneering scientific research capable of generating breakthroughs that expand the boundaries of knowledge is also a test of one of the final frontiers in AI capability. That is the ability to be creative.
In my view, if there are AI systems that can solve problems creatively to achieve scientific success, we should see that as a sign that they are capable of expanding their capabilities into many other fields, certainly including AI research. This could create an explosive impact on innovation. Therefore, if AI systems become increasingly good at science, they will become a powerful force shaping the growth rate and structure of the future economy. That means starting today, you can make decisions to create the conditions for your nation and its people to benefit from this trend in the future.
Now, let's talk about growth. One of the main drivers of growth is research activity that generates new ideas, and one of the key growth accelerators is the human capital that we possess. People all over the world are always asking questions, conducting scientific research, and helping the world move forward. But we also know that, as some economists have noted, good ideas are becoming harder to find, the cost of generating breakthroughs is rising, and we have to constantly invest more research resources just to maintain the same pace of progress as before.
So, the question I want to ask you is: what would happen if this reality changed? What if this technology allowed us to change the number of researchers on the planet as well as the time required to train them? As I said at the beginning when introducing myself, there is only one version of me, and it took me about 40 years to be here. What would happen if we could run hundreds or thousands of versions of me at the same time?
What would happen if new versions of me could be created instantly instead of having to wait 40 years? And how would my work change if I could create copies of myself and assign them to do the work under my supervision? That is the prospect that many broad sectors of the economy will experience in the next few years. We will have the ability to augment human labor in frontier fields like science in ways we simply have never been able to before.
And this will likely change the speed at which we find valuable ideas, because we can add hundreds or even thousands of times more researchers to solve any given problem. Thereby multiplying the power of our scientists. We can also do that very quickly. If you want to train a new generation of researchers in a certain field, it might take 10 to 15 years of intensive education and millions of dollars just to train a small group.
In contrast, these AI systems can be created instantly in any quantity, provided there is enough compute. They can also be applied across many different fields, almost immediately leveraging new scientific insights, and they can combine existing knowledge in new, unexpected ways to generate discoveries. That means we can quickly augment everyone on the planet with virtual scientists who can work side-by-side with us.
So, in the coming years, every human scientist on the planet will become more important and capable as they gain the ability to oversee hundreds or even thousands of aligned experts to work on any problem they care about.
Our team of economists at Anthropic has built models that simulate different economic scenarios based on various trajectories of AI technology development. And if we assume what people working in AI labs believe—that AI models will be capable of autonomously completing many different types of tasks and their prevalence in the economy will continue to grow in the coming years—then under conditions approaching this technological singularity, one model forecasts the economy will undergo a massive transformation with U.S. GDP potentially reaching 44 trillion dollars by 2030.
I know this number sounds hard to believe, but it is the projected outcome if technology continues to improve its capabilities. And if GDP really grows that strongly, then according to economic theory, there must certainly be new industries and inventions created to drive such massive economic expansion. That is exactly what is implied by the prospect of AI systems capable of autonomously conducting scientific research.
So, what should we do in the face of this prospect? How can we enable the entire world to benefit from these increasingly developed capabilities? And how do we balance the opportunities for progress with the necessary caution? We need to plan for success, helping humanity leverage this technology to create abundance and prosperity. We also need to prepare to make wise choices about how to use that very technology. Just because AI can become a talented scientist doesn't mean we should use it for everything.
Regarding preparing for success, I want to start with a story from my own experience. As I have said, over the last 10 years, I have met with governments around the world and tried to talk to them about advances in AI. I told them that because AI systems are becoming much more powerful and useful, the best thing nations can do to prepare to take advantage of the coming revolution is to build enough power capacity and data center space to run the powerful AI systems we are discussing, as well as invest in the AI supply chain.
Not many policymakers listened to that advice, and I think we are now facing the consequences of this. In today's world, data center capacity is concentrated mainly in the US, China, and increasingly in the Middle East. Supply chains are also highly concentrated and are under such intense competition that prices are constantly rising.
Currently, expanding that infrastructure is still very important, but that is only the first phase of the economic and scientific revolution that unstoppable development is unleashing. Now we have even more to build, and I believe many more countries can participate if they choose to do so. Because the next phase of the AI revolution will be an expansion into the physical world. Specifically, because AI systems are directly conducting scientific experiments and operating robots. Sometimes combining both of these activities.
Therefore, similar to computing power and electricity, we will need to deploy more scientific equipment and robotic systems worldwide. Otherwise, we will quickly fall into a capacity deficit to meet demand, and history will repeat itself. So, we need to build that capacity now to be able to take advantage of what is coming.
According to UNESCO estimates, there are currently about 10 million scientists in the world and even with those 10 million scientists, we are still short of personnel to oversee the current workload. There is a lack of laboratories for materials science analysis, a lack of biology labs to develop and test new drugs, as well as a lack of large-scale scientific equipment both in space and on Earth to conduct the experiments we desire.
Tell me what will happen if instead of just 10 million scientists, we have 1 billion virtual scientists, or even 10 billion, coordinated by the very same 10 million human scientists we have today. That is the world we can anticipate emerging quickly, and we must be ready.
If we prepare well, we will be in a better position to translate the intelligence of AI systems into benefits for everyone. That means longer, healthier lives, more energy, more food, and more wealth. And just as with data centers, the countries that build and develop this supply chain will be the ones that benefit the most. That is a choice. This is an opportunity for everyone.
If you come from a developed country with a rich scientific ecosystem, you will have the chance to help all your scientists increase their productivity. And if you come from a developing nation that provides the basic inputs for the AI industry and scientific research, such as raw materials, then you have the opportunity for a leapfrog development. You can build scientific testing facilities where human scientists collaborate with digital scientists, thereby developing your own scientific research community. You can also continue to expand your contribution to the supply chain. A sector that will become increasingly valuable.
But having infrastructure alone is not enough. If scientific discoveries become abundant, we need to avoid creating bottlenecks in the subsequent stages. A massive wave of proposed new drugs won't provide much benefit if, in the end, only a few are put into clinical trials. Policymakers in this room will need to help us find ways to accelerate the process of moving discoveries from the lab into application in medical facilities, factories, and production, as well as promoting the adoption and transfer of those achievements between nations.
The better we prepare for this world, the more capable we will be of turning the achievements of AI into real benefits for the entire world. But ultimately, we need to prepare as a global community to make choices about the future. Because as I said, there are risks here. The ability to use AI in science also means the potential for science to be misused. Such as the development of biological weapons, cyber weapons, and things we cannot even imagine yet.
We have many choices to make, and only by correctly perceiving the nature of this technology, understanding its current power as well as the power it will achieve in the future, will we see the urgency of those choices.
The first choice concerns which fields we direct AI toward. Just because we have an AI system capable of scientific research doesn't mean we should let it conduct research on its own without oversight. One example is biological research, a field where we need to apply the strictest human oversight and ensure transparency to maintain safety. Another area is AI research itself. Labs like Anthropic will have to decide the level of oversight needed for research conducted by AI.
This could help us achieve more technological progress, but it also entails risks. We will need to proceed with extreme caution; otherwise, we could lose control of the technology. In all likelihood, we will also need rules to govern these activities as well as to govern ourselves.
Earlier this week, I was in Tokyo. A university student there told me they were working in the physics department. The student said that AI has begun to have the ability to conduct many of the experiments they are studying. So, as humans, which experiments should they spend their time on, and which should they let AI systems perform on their own? This is a very subtle question that I do not have an answer to, but it is one that every branch of science will have to answer quickly.
The second choice concerns the level to which we need to understand what AI systems are doing. As we have seen this year with advancements in AI in mathematics, we have entered a world where AI-generated achievements are so creative that humans must now spend time decoding them, trying to understand their meaning, and finding ways to apply those insights themselves.
So, how much do we need to understand? Again, the answer will be very complex and relate to fundamental questions about meaning in the era of the technological singularity. In which areas of scientific progress do we want to truly understand the nature of discoveries? In just a few years, AI systems will not only be capable of conducting scientific research but also of inventing their own scientific devices designed for the way AI operates, not for human thinking. Whether we need to understand those devices and whether we should understand them is another choice entirely.
And ultimately, if we take AI progress seriously, we need to allow ourselves to imagine a world that is fundamentally different. As a teenager, I read a lot about the world and felt afraid. I think many other teenagers felt the same way. I used to believe that the 21st century would be extremely difficult. I believed that population growth, climate change, resource depletion, and a host of other crises would converge to lead to a Malthusian collapse of humanity. And I believed that as we rose to invent computers and the internet, we might eventually destroy ourselves through overconsumption.
But in the years that followed, I consistently saw technology helping us change those outlooks. I saw solar panels emerge and become much cheaper at a pace far faster than anyone had predicted. I saw a deepening understanding of the human body, helping us invent cures for diseases that had once taken the lives of many, including members of my own family. I also saw us do things I had never imagined before, but found absolutely amazing. For instance, how we are using AI systems to decode animal communication or how companies are building rockets that can launch and then return to land on Earth.
I work in AI because this technology can help us do much more of those things. And I work in AI because whether this technology actually achieves those results depends on our choices. These are the choices that those in this room will make about how we guide and govern AI. And they are also the choices that I myself must make.
I work in AI because it is the most important technology humanity has ever created. If used correctly, this technology will allow us to solve problems that, if left unaddressed, could push humanity toward extinction. But this is also a technology with dangers deeply embedded in its very nature.
To navigate this phase, to find a path through the technological singularity, we will need to cooperate to build special and strict rules for this technology and the companies developing it. And we need to demonstrate one of the most difficult, but in my opinion the noblest, of human qualities: restraint.
We will need to build the infrastructure to take advantage of this technology. But we also need to develop the technology itself slowly and deliberately. To regulate or control the pace of its development. We need to constantly distinguish between what we can do and what we should do. And starting today, we need to invest in the scientific capacity to turn those advancements into reality.
All of this reminds me of the great astronomer Carl Sagan in his book Pale Blue Dot. He once said that one day, the humans who could reach the stars would still be a species very much like us, but possessing more strengths and fewer weaknesses, more confident, more farsighted, more capable, and more cautious. That is exactly what AI demands of us. That is what AI demands of me and the companies I represent.
And if Sagan once thought we had hundreds of thousands of years to achieve that—to become better and wiser humans—we only have a few years, certainly less than a decade. I believe we can do it, but only by viewing AI with the seriousness it deserves can we all succeed. Thank you very much.
