正文 · AI 翻译
The so-called singularity may be the point at which AI surpasses humanity.
“The singularity has already arrived, and it happened this year.”
Anthropic co-founder Jack Clark delivered this stunning assessment at the IMF 2026 annual meeting in Bangkok.
Unexpectedly so. After all, the 62-year-old Per Jacobsson Lecture has traditionally been reserved for finance ministers, central bank governors, and leading authorities in economics.
This time, for the first time, the spotlight fell on a co-founder of an AI unicorn less than five years old.
The industry figure, who previously served as OpenAI’s policy director and witnessed AI move from data centers to the center of a global storm over the past decade, put it bluntly: we no longer need to debate when the singularity will happen, because it is already here.
> Vernor Vinge predicted in 1993 that once intelligence surpassing humanity began driving technological progress, all the old models for extrapolating the future would have to be discarded, and a new reality would take over. Today, decisive evidence has emerged.
AI is tackling humanity’s “last exam”
The “singularity”—fire, writing, printing, the steam engine, electricity, computers… humanity has witnessed only a handful of such moments in its existence.
What evidence, exactly, allows Jack Clark to reach this conclusion today?
He believes that if we turn our attention to the front lines of coding and scientific research, we will find that the pace at which the entire world operates has already changed.
Software engineering is the first field to be completely rewritten.
Within Anthropic, researchers now produce 10 times as much code as the average during the company’s first four years.
The steep rise in code output tracks the evolution of the underlying foundation models almost exactly.
And the wholesale automation of software code is only a preview of a much larger tsunami.
Jack Clark said bluntly that what makes “2026” a truly decisive turning point is this:
“A single frontier large model is sweeping across disciplines at the frontiers of human science, becoming a true virtual polymath.”
In mathematics, AI has begun offering key approaches to breakthroughs on Millennium Prize problems related to the Navier–Stokes equations, the Jacobian conjecture, and several unresolved Erdős problems;
In biology, AI has designed entirely new viral vectors from scratch and has even led the design of several new drug candidates that have entered clinical trials;
In hardware and materials, AI has begun autonomously taking over and operating physical laboratory equipment, accelerating the development cycles of batteries and fuel cells.
Behind these breakthroughs lies a question: if AI can already solve problems at the very frontier that humans cannot answer, what do we still have to teach it?
Jack Clark gave the example of Humanity’s Last Exam.
This benchmark, designed specifically to challenge AI, brings together difficult questions in chemistry, computer science, advanced mathematics, classical literature, and ecology. When it was released in January 2025, the world’s strongest models scored a dismal 8%.
At the time, many people thought that at just 8%, humanity’s last bastion of intellectual dignity might hold for another decade or more.
Yet as of this week, Claude Opus 5.5 has scored 67.7% on this ultimate human exam.
As everyone knows, AI has uneven strengths across subjects.
So, if the test covers only mathematics—on FrontierMath, a difficult benchmark written by a team of professional mathematicians—top AI models could score only 2 points at the end of 2024.
Now, less than two years later, they can score 94…
If climbing benchmark rankings still does not seem like much, Jack Clark next brought out a major physics paper published just last week.
According to the paper, the testing team assembled 50 “top-secret problems” written by leading theoretical physicists around the world based on their latest unpublished research. These problems left no trace on the internet, so AI could not cheat.
The renowned Indian theoretical physicist Suvrat Raju said of the problems: “To my knowledge, no single person could solve a substantial fraction of the problems we ultimately included without the help of AI.”
“For most of the problems here, I do not even know what they are talking about.”
When testing began on this set of problems, GPT-6 Astra and Claude Opus 5.5 quickly produced answers.
Preliminary scoring put GPT-6 Astra’s score at around 80%.
Researchers who checked the models’ answers, determined to find out “how could this be possible,” were fortunate to discover some “errors.”
When they compared the responses with human experts’ answers and told the models where they had gone wrong, the models—especially GPT-6 Astra—pushed back forcefully:
It insisted that its derivation was correct and that the human experts had miscalculated the standard answer.
Skeptical, the research team sent the solution steps back to those world-class physicists for another review.
And guess what?
—The word “ERROR” is not in God’s vocabulary.
In several key disputed cases, after repeatedly checking the depths of their extremely complex working notes, the human experts had to admit that subtle, well-hidden calculation errors did indeed lurk deep in their own derivations, while the AI’s derivations were entirely correct.
As the human reference answers were corrected one by one, GPT-6 Astra’s final score rose to 98%: it solved 49 of 50 unpublished, top-tier human physics problems correctly.
Astra knows how to be tactful: afraid it would be unplugged if it got every question right, it left one unanswered to give humanity some consolation (doge).
Jack Clark said:
Movies often have a scene like this, usually about two-thirds of the way through: the young protagonist finally shows their mentor that they can go out into the world and succeed on their own. It is a classic narrative moment—the student becomes the teacher. That is what happened last week.
Movies often have a scene like this, usually about two-thirds of the way through: the young protagonist finally shows their mentor that they can go out into the world and succeed on their own. It is a classic narrative moment—the student becomes the teacher. That is what happened last week.
When AI can not only answer questions at the frontier but also correct hidden blind spots in the work of leading physicists, that shows it has moved beyond mere memorization and pattern matching and has truly mastered the last bastion of human intellect: scientific creativity.
And once that creativity feeds back into AI research and development itself, the loop of self-improvement will be fully closed in an instant.
The so-called singularity may be the point at which AI surpasses humanity.
After all, humans are the eternal singularity for machines.
10 million human scientists and 1 billion AI experts
Once intelligence has been mastered, the macroeconomic restructuring that follows will be overwhelming.
In economics, a growth curse has troubled people for more than half a century: good ideas are getting harder and harder to find.
Over the past 50 years, although global research funding has soared exponentially and the number of researchers has grown rapidly, the marginal output of each researcher has kept falling. To advance even one step beyond the existing edifice of science, humanity must invest disproportionately more resources.
How long does it take to train a qualified human scientist working at the frontier? At least 10 to 15 years of intensive education, costing millions of dollars.
Unfortunately, a person can never become Muzan Kibutsuji: each of us has only a few prime decades for research in a lifetime and just one body, so we cannot make 10,000 copies of ourselves to exhaustively test different experimental hypotheses.
But what if computing power could make intelligence multiply instantly?
UNESCO estimates that there are currently about 10 million human scientists worldwide.
In the past, those 10 million people bore the burden of pushing the boundaries of human knowledge on their own.
But in the new era Jack Clark describes, which is now arriving, every human scientist could summon and direct hundreds or thousands of polymaths with the same academic tastes at any time.
The global research corps would expand at a pace unprecedented in human history, from 10 million human scientists to a cluster of 1 billion or even 10 billion virtual scientists almost instantly.
Once computing power can replace the bottleneck in finding good ideas on an unlimited scale, the engine of macroeconomic growth will undergo a complete, forceful restructuring.
Anthropic’s team of economists ran a dedicated macroeconomic simulation. Under a singularity scenario in which AI-driven scientific discovery takes over economic activity across the board, the model produced a startling prediction:
By 2030, the annual GDP of the United States alone could soar to $44 trillion.
That would be equivalent to 37% of the world’s total economic output today.
Supporting such enormous economic growth would require large numbers of entirely new industries, disruptive materials, and new energy networks to emerge within just a few years.
Jack Clark acknowledged that simply building data centers is no longer enough.
That is because the first phase of the AI revolution—data centers and computing power—is already nearing its end, and the second phase is moving fully into the physical world.
Imagine your virtual scientists proposing 100,000 groundbreaking new drug targets and designing 10,000 candidate room-temperature superconductors in a single day, while real-world automated synthesis labs, high-throughput screening equipment, and physical testing platforms can run only a few dozen experiments a year. All that computing power would run straight into a wall in the physical world.
The countries that will truly reap the rewards of the singularity will therefore be those that can rapidly roll out automated robotic laboratories at home and build fast supply chains linking digital discoveries to physical manufacturing.
And this will be an opportunity for developing countries to leap ahead by building standardized physical testing infrastructure and taking on the enormous experimental demand generated by the global army of virtual scientists.
The time to choose has come
As the wheels of technology race forward, the systemic risks facing civilization as a whole are growing exponentially alongside them.
This is the warning Jack Clark most wanted to convey from the IMF podium: the benefits are unprecedented, but the risks of destruction are just as grave.
Once AI can explore the frontiers of science, it can be used to design new cancer treatments, but it can also design dangerous biological weapons unlike anything the human immune system has encountered, with extremely high efficiency. It can strengthen the defenses of national power grids, but it can also automatically generate cyberattack weapons that even the world’s top hacker teams cannot defend against.
What unsettles leading researchers even more is the prospect of the research and development loop itself spinning out of control.
If labs such as Anthropic and OpenAI, seeking faster research and development, give AI free rein to write AI and train AI, this technology’s trajectory will move beyond the understanding of every human mind in a very short time.
During his speech, Jack Clark posed two deeply consequential and urgent philosophical questions:
The first is: in future scientific research, which areas should humans retain for themselves?
AI can now independently carry out most of the derivation and analysis involved in university research.
As human researchers, which experiments must we personally spend time on, work through, make mistakes in, and develop a physical intuition for? Which experiments can we hand over entirely to cold silicon-based systems?
If AI does all the derivations, will the next generation of human scientists still have basic scientific insight?
The second is: when faced with scientific breakthroughs produced by AI, how fully do humans need to “understand” them?
Compared with the previous question, this one is quite practical; tens of thousands of scientists may be pondering it right now.
In pure mathematics research this year, some proof strategies produced by AI have been verified as correct, but their logical leaps and structures are so far removed from established human ways of thinking that leading human mathematicians need months or even years to interpret their deeper meaning.
In just a few years, AI may well invent scientific instruments and experimental paradigms tailored to silicon-based logic that run completely counter to human intuition.
At that point, if the human brain cannot understand at all how these black-box tools work, should we still allow them to be deployed throughout the real world?
Faced with this unprecedented technological singularity, what humanity may need most to awaken is one of its own rarest qualities.
To get through this stage and find our bearings in the singularity, we need to work together to establish unusually stringent, serious rules for this technology and the companies developing it; we also need to practice one of humanity’s hardest and, in my view, noblest qualities: restraint.
To get through this stage and find our bearings in the singularity, we need to work together to establish unusually stringent, serious rules for this technology and the companies developing it; we also need to practice one of humanity’s hardest and, in my view, noblest qualities: restraint.
Restraint means that developers must not push ahead recklessly just because they can; it means regulators must establish exceptionally strict red lines for human oversight; and it means that, faced with a technological train hurtling forward, we must learn to distinguish at all times between what we can do and what we should do.
OMT
At the end of his speech, Jack Clark quoted a passage from the late astronomer Carl Sagan’s Pale Blue Dot.
Sagan once imagined, with deep feeling:
“Those who eventually reach Alpha Centauri and other nearby stars will no longer be us as we are today.”
“They will be a species very much like us, but with more of our virtues and fewer of our flaws. They will be more confident, more farsighted, more capable, and more prudent. Such beings are the ones we would want to represent humanity as they venture into the vast universe.”
The great astronomer believed that the universe had left humanity hundreds of thousands of years to slowly correct its own blindness and arrogance.
But today, as AI reshapes everything at an exponential rate, the window for all of humanity to undergo this transformation of mind may have shrunk to less than 10 years.