Anthropic Co-founder Jack Clark: "We Are in the Singularity"

TL;DR: Anthropic co-founder Jack Clark argues that 2026 has provided compelling evidence that the "singularity" has already begun — AI is making breakthroughs in frontier science and will bring economic and societal changes far beyond what was expected. ## Opening: One Person's Credentials "Thank you very much for that stress-free introduction." Jack Clark took the stage and first thanked the organizers for the invitation, then listed his relevant credentials: "I am a human. There is only one of me in the world, and I've spent nearly 40 years accumulating the experience to give today's talk." His entire career can be defined as: finding information that seems important for the future of the world, and then telling the world about it. He started out as a journalist twenty years ago, focused on data centers; that work made him so obsessed with AI that he left journalism in 2016 to join OpenAI. A few years later, he co-founded Anthropic. Since then, he has been trying to tell the world that one day AI systems would become so powerful as to change national security, alter the economic trajectory of the planet, and ultimately gain the ability to build themselves. "That moment has arrived. AI is changing national security, AI is changing the trajectory of the world economy, and it is driving its own development." ## The Singularity Has Begun For years, researchers and science fiction writers have predicted a future in which AI systems would dramatically accelerate the pace of overall scientific and technological progress, advancing so fast that humans can barely keep up — they call this the "singularity." Clark noted that writer Vernor Vinge predicted the singularity in 1993. Vinge said that when superhuman intelligence drives progress, that progress will be much more rapid; in fact, there seems to be no reason progress should not involve creating more intelligent beings on shorter timescales. Vinge described the singularity as a critical point where "our models must be discarded and a new reality reigns supreme." "I believe that this year we have already obtained compelling evidence that the singularity has begun." The impacts of the singularity will touch every aspect of life, but the most important takeaway is simple: we should expect that in the coming years, not just AI but the pace of progress across all of science will accelerate. This will be confusing and unsettling, while also bringing enormous wealth and potentially enormous risk. So Clark's focus today is on the former — the benefits of rapid AI progress, and how we can share them together. ## A Turning Point in Researcher Productivity For the past half century, most countries have been grappling with declining researcher productivity, even as the number of researchers has grown rapidly. Ideas and inventions seem increasingly hard to find — as some economists put it, the best ideas are getting harder to find, and the cost of breakthroughs is rising. But something important has changed: the AI systems we now build are capable of augmenting, extending, and automating human labor. This is most evident in software development — AI systems are now more capable and efficient than the vast majority of software engineers in the world. A single person can now write software for anything they can imagine, creating things that once took entire teams months of work. The evidence within Anthropic is concrete: researchers are now producing nearly 10 times the amount of code that the company produced in its first four years combined, and this growth in coding output is directly correlated with advances in AI capabilities. ## Progress Brings New Capabilities and New Challenges Improved coding ability means AI systems have become more capable in both cybersecurity and cyberattacks, driving countries around the world to race to deploy these powerful capabilities for cyber defense, while malicious actors race to use these systems for hacking. "It's important to recognize that these two are inseparable. AI systems' progress in a general domain like coding produces both positive and negative impacts in internationally relevant domains like cybersecurity." These capabilities won't stop at coding. The same thing will happen in science and robotics, which means some important properties of the global economy may change — such as the rate of innovation, or the frequency with which we invent entirely new general-purpose technologies, like the internet, mobile phones, and artificial intelligence. ## 2026: Signs of AI-Driven Frontier Science In 2026, we saw for the first time signs that AI systems may be capable of conducting scientific research at the frontier. **Mathematics**: AI systems have been used to propose solutions to long-unsolved mathematical problems, such as the Navier-Stokes Millennium Prize Problem, the Jacobian conjecture, and several unsolved Erdős problems. In many cases, AI systems provided the core insights that made the solutions possible. **Biology**: Researchers have used AI to design viruses from scratch and identify new drug candidates for testing. There are now some drugs in clinical trials that were largely designed by AI systems. **Robotics**: AI systems are now increasingly able to autonomously operate laboratory equipment, allowing us to further automate experiments across fields from biology to materials science, accelerating progress in areas like battery or fuel cell development. Each example on its own is exciting, but the profound fact is that they all involve the same underlying technology — for most of these use cases, the same underlying frontier model. We now have a single entity — an AI system like Claude, ChatGPT, or Gemini — capable of working at the frontier across multiple scientific disciplines. "We have virtual polymaths." ## Quantitative Evidence: Two Curves **Humanity's Last Exam** — that name was not Clark's. This ominously titled test covers a range of topics including chemistry, computer science, mathematics, classical literature, and ecology. - When launched in January 2025: the highest-scoring AI system scored just **8%** - As of this week: the highest-scoring system is **Claude Opus 5.5, at 67.7%** **FrontierMath** — released in November 2024, evaluating AI systems' ability to solve hundreds of challenging math problems curated and graded by professional mathematicians. - At release: the top system scored only **2%** - Now: **94%** Clark said he has a file on his computer containing 100 charts produced in the past week, all with data downloaded that very morning. Looking across different fields of science, the same pattern repeats: progress is accelerating, and it has only begun accelerating in the past two years — and in some cases, just the past year. ## TheophysEval: AI Finds the Experts' Mistakes The most recent example comes from a paper published last week called TheophysEval, testing AI systems' ability to solve unpublished problems in theoretical physics. The 50 problems in the paper were written by physicists around the world, and the answers are unpublished, drawn from their own research. Indian physicist and co-author of the test Suhas Raju wrote on Facebook: > "I don't know of any individual who could solve a significant fraction of our final problem set without AI assistance. Our initial evaluation put GPT-6 Astra in the high 80s, followed closely by Opus 5.5. But when we started investigating potential errors in the models' answers, the models — particularly Astra — insisted they were correct and that the expert-provided answers were wrong. When we went back to the experts, we found that in multiple cases, the experts discovered subtle errors deep in their calculations. As they corrected their answers, Astra's score began climbing, ultimately finishing at 98%. 49 out of 50 answers correct." Some of the world's leading scientific experts built a test using questions whose answers the AI had never seen. And when they tested the AI, it not only scored highly but gained an even higher score by finding errors in the experts' painstakingly worked answers. Clark used a cinematic analogy: "In movies, there's usually a scene around the two-thirds mark: the young protagonist finally proves to their mentor that they have the ability to go out on their own and succeed. The teacher becomes the student, or the student becomes the teacher. That's exactly what happened last week." ## Scientific Capability = The Last Frontier of Creativity Why does it matter that AI is getting better at science? Clark's view: doing real frontier science — having the capability to make breakthroughs that expand the frontiers of knowledge — implicitly tests one of the last frontiers of AI capability: **creativity**. If AI systems can solve problems creatively and conduct scientific research successfully, we should expect this to be a signal that they can generalize across many domains — certainly including AI research itself, which could have an explosive impact on innovation. ## Growth: From 40 Years to an Instant One of the primary drivers of growth is that research produces new ideas, and one of the factors driving growth is the stock of human capital. Clark's question: what if this technology allows us to change the number of researchers on the planet and the time it takes to train them? "There's only one of me in the world, and it took me about 40 years to get to this point. What does it mean if we can run hundreds or thousands of copies of me simultaneously? What does it mean if new versions of me can be created instantly, instead of waiting 40 years?" Comparative data: - Training a new generation of researchers: **10 to 15 years** of targeted education and **millions of dollars** of investment, to produce a small cohort - AI systems: can be **created in any number instantly**, given sufficient compute, deployed across disciplines, and benefit from new scientific insights almost immediately This means that in the coming years, every human scientist on this planet will become more important and more powerful, because they will gain the ability to supervise hundreds or thousands of like-minded experts working to solve the problems they care about. ## $44 Trillion US GDP by 2030 Anthropic's team of economists conducted modeling, simulating different versions of the economy under different AI technology development trajectories. If we assume that AI models will be capable of autonomously completing a broad range of tasks, and that their penetration into the economy increases in the coming years, then under these "singularity conditions," one model predicts: **US GDP could reach $44 trillion by 2030.** Clark acknowledged that "this sounds very unrealistic, but it is the implied consequence of continued technological progress." If GDP can grow that much, then economic theory suggests that new industries and inventions must be created to support such a major economic expansion — and that is exactly what the concept of "AI systems capable of autonomous scientific inquiry" implies. ## How the World Benefits: Phase One Is Over How do we ensure the world broadly benefits from these ever-advancing capabilities? How do we balance opportunities for progress with appropriate caution? "Just because AI will be an excellent scientist doesn't mean we should use it to do everything." Regarding planning for success, Clark shared his own experience: over the past 10 years, he had visited governments around the world and told them that as AI systems become increasingly powerful and useful, the best way they could seize this coming revolution was to **build sufficient power generation capacity and data center space**, and to **invest in the AI supply chain**. "Not many decision-makers took that advice. I think we now have to face the consequences." Today, data center capacity is concentrated in the United States, China, and increasingly the Middle East, with supply chains highly concentrated, intensely competitive, and prices rising. Clark's conclusion: this buildout still matters, but it is only the **first phase** of the economic and scientific revolution that AI implies.