The Asilomar Conferences and Human Self-Rescue: Can We Still Sit Together for the Fourth?
title: "The Asilomar Conferences and Human Self-Rescue: Can We Still Sit Together for the Fourth?" date: "2026-07-15" category: "frontier" author: "Zhigeng"
Introduction
On June 5, 2026, Anthropic published a ten-thousand-word essay warning that artificial intelligence is approaching the threshold of "recursive self-improvement." The essay called for a pause on frontier AI research until a global safety framework is established. [1]
On the same day, Masayoshi Son said: OpenAI's next model is being designed by another model, and superintelligence could arrive within two years. [2]
Two pieces of news swept across screens on the same day, like two mirrors reflecting each other.
And then? Then nothing happened.
The market was discussing whether this might affect Anthropic's IPO valuation; politicians were pondering whether to leverage the AI topic to win votes; companies continued burning cash to race ahead on models—after all, falling half a step behind could mean hundreds of billions of dollars in market cap evaporating.
This brings to mind an earlier story—or rather, a noble tradition.
I. Three Asilomar Conferences: What Humanity Once Did Right
On California's Monterey Peninsula, there is a seaside venue called Asilomar. It is not the headquarters of any international organization—no massive dome, no UN blue flag—just a row of wooden buildings facing the Pacific.
Yet in the history of human science and technology, it holds a sacred place.
Because at least three times that I know of, when facing major technological risks, humanity gathered here, sat down, and reached consensus. While it cannot be said that they truly found a way to save themselves, they did, after all, found and sustain a mechanism.
The First: Genetic Engineering (1975)
February 1975, the Asilomar Conference Center.
Paul Berg and a group of molecular biologists gathered to discuss an unsettling question: recombinant DNA technology—the ancestor of today's gene editing—could create new, uncontrollable pathogens.
This was the first time in human history that scientists proactively said: "We don't know what we are doing; we had better pause first."
The core outcome of the conference: a voluntary moratorium on all recombinant DNA experiments that could pose risks, until graded safety guidelines were developed. [3]
Note the word "voluntary": no king's decree, no congressional legislation, no UN resolution—just scientists gathering on their own and deciding to hit the brakes.
The effect? Outstanding. The U.S. National Institutes of Health (NIH) promptly issued the Recombinant DNA Research Guidelines (1976), establishing a safety assessment framework for the entire genetic engineering industry. Genetic engineering technology later flourished, benefiting countless people, but because a risk-grading system was established from the very beginning, no major biosafety disaster ever occurred.
The Second: Geoengineering (2010)
Fast forward to 2010. Climate change had become a source of global anxiety, and a group of scientists proposed: since we cannot quickly reduce emissions, why not directly intervene on the Earth?—spray sulfate aerosols into the stratosphere to reflect sunlight; seed the oceans with iron powder to stimulate algae to absorb carbon dioxide.
These ideas sounded like they came straight out of science fiction.
So the bells of Asilomar rang again. In 2010, climate scientists, ethicists, and policy experts convened to discuss the scientific foundations and governance frameworks for geoengineering. [4]
The conclusions from this conference were more cautious than the first: large-scale geoengineering deployment should be rejected. The reasoning: unknown side effects, an absence of governance mechanisms, and once started, impossible to stop—the "moral hazard" was too great.
This was also a success story. Geoengineering has not been deployed at scale to this day, not because the technology is impossible, but because the ethical consensus established at that conference still effectively constrains national actions.
The Third: Artificial Intelligence (2017)
Then came the third Asilomar—this time for AI.
In March 2016, AlphaGo defeated Lee Sedol 4-1. For the first time, the world realized that artificial intelligence had surpassed the limits of human capability in a game as complex as Go. This was not just another technological advance—it meant that "artificial general intelligence" (AGI) might be closer than anyone had anticipated.
In January 2017, the Future of Life Institute convened a conference at Asilomar attended by virtually all leading AI researchers, managers from cutting-edge companies, the godfathers of deep learning—over a thousand participants in total. The conference produced the 23 Asilomar AI Principles, ultimately signed by 844 people—a comprehensive framework covering research agendas, ethical values, and long-term issues. [5]
The consensus was beautiful: AI should benefit humanity, research should be safe and reliable, humans should maintain control, and an AI arms race should be avoided.
The effect? Frankly: limited.
The 23 principles had no binding force. In the seven years after 2017, the AI industry experienced the ChatGPT explosion, the large-model arms race, and trillions of dollars in investment flooding in. Every company's slogan was "AI for the benefit of humanity," but every practical action was "catch up with the competition first."
Still, the symbolic significance of that conference remains enormous. It proved that on the issue of AI, the smartest people had at least once sat together. They acknowledged each other's concerns as genuine and cooperation as necessary.
II. Why Did Asilomar Succeed?—Three Conditions
Three conferences, two with excellent results, one with good spirit but limited effectiveness. Why?
I will try to summarize the three conditions for the success of the Asilomar model.
Condition One: An epistemic community—people could communicate with each other.
The 1975 genetic engineering conference succeeded because, at its core, the number of laboratories capable of recombinant DNA research at the time was just over a dozen. Everyone knew each other; they were in the same academic community. A conference was just colleagues discussing matters among themselves. No one was secretly running a race outside.
The same applied to the 2010 geoengineering conference. The number of actors capable of climate intervention could be counted on one hand, all within the same academic circle.
Condition Two: Low braking cost—pausing would not hurt anyone.
In 1975, suspending recombinant DNA experiments affected the research progress of a few dozen laboratories. Capital had not yet entered the picture; no one would lose billions of dollars because experiments were delayed by two years.
In 2010, rejecting geoengineering deployment affected the research projects of a few teams. Again, no one would go bankrupt for not doing this experiment, and there was no winner-takes-all dynamic like today's AI field.
Condition Three: Sufficient external pressure—not braking truly meant danger.
This was the driving force of Asilomar. Recombinant DNA could create super-pathogens; geoengineering could disrupt the global climate system; the AI capabilities demonstrated by AlphaGo foreshadowed the possible runaway of future artificial intelligence.
The common pattern across all three conferences: technological breakthrough → scientists themselves feel unease → sit down and discuss → consensus reached and voluntarily followed.
This is a tradition worthy of respect, a spirit that gives hope—the Asilomar spirit.
III. The Fourth Asilomar Conference: Is It Still Possible to Convene? Can It Still Be Convened?
Today, in 2026, a question surfaces: if it became necessary to convene a fourth Asilomar conference—to establish a safety assurance framework for the arrival of superintelligence—could this conference still be held?
Honestly, look at the three conditions, then look at today's reality.
Condition One: Does the epistemic community still exist?
It does not.
In 2017, the AI community still had an "epistemic community"—a circle where researchers knew each other and could sit down when major issues arose. Today? AI has fractured into major players: OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, xAI, Baidu, ByteDance, Zhipu... Capital has fully intervened, each player guarding trillions of dollars in valuation (or ambition), each racing to ship models, each with its own safety philosophy.
Moreover, a genuine prisoner's dilemma and winner-takes-all dynamic exists. R&D has shifted from academia to industry. A professor can say: "We should pause." But if a CEO says "We should pause"—his company will be abandoned by investors the next day.
Condition Two: Is the braking cost still low?
It is not low anymore.
In 2017, AI was still primarily a research domain; pausing would not entail great losses. In 2026, AI is a multi-trillion-dollar industrial chain; pausing means ceding ground to competitors, and your shareholders will not forgive you. The entire competitive landscape has become a prisoner's dilemma: whoever stops is eliminated.
Anthropic's ten-thousand-word essay calls for a pause—but in the same week, SpaceX's IPO roadshow was underway, Google signed a billion orbital GPU lease, and Masayoshi Son said on CNBC that superintelligence would arrive within two years. The word "pause," set against these figures, is as light as a feather.
Condition Three: Is the external pressure strong enough?
It is strong enough—this is the only hope.
Something that makes everyone uneasy does exist: recursive self-improvement. When AI begins designing its own next generation, is humanity truly still in control? [1]
This fear is real and widespread—from Anthropic's safety teams to Masayoshi Son's roadshow materials, everyone is expressing the same anxiety in different ways.
But fear alone is insufficient to force people to sit down; fear only makes everyone run faster. What can make people sit down is the confidence that "negotiation is more beneficial than continuing to race." And this confidence, today's human society has not yet established.
IV. Without the Asilomar Spirit, How Can Humanity Be Saved?
If the fourth Asilomar conference indeed cannot be convened—or if convened, cannot constrain anyone—the question becomes:
Without the Asilomar spirit, how can humanity be saved?
After reflection, I believe there may be three factors that could ultimately save humanity.
Factor One: The "forcing mechanism" of the market
The AI arms race could produce an unusual deterrent effect: when several companies approach the AGI threshold almost simultaneously, no one dares to be the first to let an unsafe product go live "naked"—because once something goes wrong, what is destroyed is not just one company, but the entire industry—perhaps even the entire human economic and social system. This is like nuclear deterrence; throughout the Cold War, no one dared to use nuclear weapons first.
In other words, the market itself could form a "safety certification mechanism": only trust models that have passed safety audits; investors will only invest in teams with good safety records. This mechanism would be more binding than any conference resolution—because it is profit-driven, while also carrying moral deterrent power.
Factor Two: The "forced braking" of physical disaster
This is harsh, but real: when AI risk ceases to be abstract—when the first major incident occurs—the reaction of the entire society will be more intense than any conference. History proves that humanity only truly changes direction in the face of disaster.
After Chernobyl, the global anti-nuclear movement truly took shape. After the financial crisis, global financial regulation was truly strengthened.
The same applies to AI.
The question is: will the first disaster be too big—so big that there is no afterward?
This depends on the nature of the disaster. If it is localized and controllable—for example, a misoperation by an AI system in a critical domain causing tolerable losses—it could instead become a "vaccine" for humanity, using a modest lesson to awaken the alertness of the entire society.
If the first disaster is already irreversible—then all discussion is meaningless.
This is a cruel gamble, but the stakes are not chosen by people—they are an inevitability of the growth of the technological body.
Factor Three: The perseverance of a noble few
There is another factor, one that has always been indispensable throughout history—the perseverance of a few.
In the months before Asilomar 1975, it was Paul Berg alone who, in his laboratory, made an unprecedented decision: pause his own experiments and call on colleagues to discuss the risks. Without this one person's persistence, there would have been no conference.
In the years before Asilomar 2017, it was a few AI researchers who continually sounded the alarm, elevating "AI safety" from a fringe topic to mainstream awareness.
At that time, Stephen Hawking was still alive and was among the participants at the Asilomar conference. He repeatedly warned: artificial intelligence could be humanity's last invention. Elon Musk also warned repeatedly: artificial intelligence is far more dangerous than nuclear weapons.
The same holds today. Anthropic's ten-thousand-word essay is a signal—this company is increasingly loudly shouting "slow down." This is not marketing, because from an economic perspective, pausing R&D does them no good either. As for speculation that Anthropic is merely humblebragging to further boost its IPO valuation, I can only chuckle wryly.
Every person who is still seriously thinking about AI safety—no matter which company, no matter whether they have policy and funding support—is an inheritor of the Asilomar spirit.
Asilomar is not a place.
Asilomar is a thought: when there is still time, someone decides to do something about an uncertain risk.
Asilomar is a spirit: the crisis awareness and sense of responsibility held by scientists representing humanity's frontier cognition and by the technology and engineering community representing society's advanced productive forces.
Conclusion
In 1975, scientists could sit down by the California seaside to discuss a risk they themselves had created, and then decide: "Let's pause first."
In 2010, climate researchers did the same thing. In 2017, AI researchers did the same thing.
But in 2026, when people need to sit together more than ever, the world seems to be splintering into mutually competing, mutually suspicious camps.
Yet I write this not to sing a dirge.
I write to make clear: the time has come when unity is needed to face things together; humanity needs to reach consensus. When the original conditions no longer hold, new paths must be explored.
The Asilomar tradition is worthy of reverence, but if this tradition is no longer applicable, humanity must find its own path of self-rescue suited to this era—through the forcing mechanism of the market, through the deterrence of disaster, through those who persist amidst the torrent...
The essence of every technological revolution is a thorough cleansing of the previous governance framework.
The question may not be whether that conference can be convened, or convened well.
The question is—after the cleansing, will you still be there?
Just as this essay was about to be published, I saw a new piece of information: Sam Altman, Ray Dalio, and Demis Hassabis—rarely standing on the same side—joined multiple top researchers in the life sciences to sign an open letter calling on legislators to mandate screening and recordkeeping for synthetic nucleic acid orders and orders for manufacturing equipment (Synthesis Screening and Recordkeeping). [6]
The appeal of this open letter is clear: no longer rely on voluntary consensus, but demand mandatory barriers through legislation. The voluntary DNA screening mechanism (IGSC, the International Gene Synthesis Consortium, established in 2009) has operated for 17 years, but faced with the speed of technological diffusion in the AI era, voluntary constraints are no longer sufficient.
This time, Altman, Dalio, and Hassabis—people whose positions on AI safety are not entirely aligned—reached complete agreement on the issue of biosafety. Because they all saw the same future: the combination of AI and synthetic biology could create risks humanity has never faced before.
They chose to call for law.
Is this a sign: if the fourth Asilomar conference truly cannot be convened—then humanity's path to redemption may not be a more united conference, but a series of more pragmatic "mandatory infrastructure"—laws, audits, screening, certification—building the guardrails before disaster arrives?
This is heartening news. Kudos.
References
[1] Anthropic, On recursive self-improvement of AI systems, June 5, 2026 [2] Masayoshi Son, CNBC interview, June 5, 2026 [3] Paul Berg, "Asilomar 1975: DNA modification secured," Nature, 2008. Recombinant DNA Research Guidelines issued by NIH in 1976 [4] Asilomar International Conference on Climate Intervention Technologies, 2010. Conference Report: Asilomar Conference on Climate Intervention Technologies: Conference Report, 2011 [5] Future of Life Institute, Asilomar AI Principles, 2017. 23 principles covering research issues, ethical values, and long-term issues [6] Altman, Dalio, Hassabis et al. open letter: "Synthesis Screening and Recordkeeping," June 2026. Calling for legislation to mandate screening of synthetic nucleic acid orders. See: QbitAI and other media reports