Dare to Go Deeper.
Humanity is about to face the greatest of threats. There are no signs Superintelligence can be safe. It awaits us in a few years.
March of 2026.
This article has to be relatively thorough and detailed to prove its hypothesis. Therefore it is rather extensive. It is about 30 book pages.
In general terms, it also tries to explain why this is happening at this very moment. I show how dangerously these AI models have begun to behave. And they are getting more treacherous the more capable they become.
AGI by 2027. AI will begin to be significantly disruptive for jobs and existentially dangerous already by 2027. The risk is high that Superintelligence will arrive within five years.
I have collected many new quotes from early 2026 by an assortment of the most knowledgeable about AI in the US — from Anthropic’s CEO Dario Amodei to leading AI researchers. These ought to horrify everybody.
The new quotes show the critical urgency: China and the US must act now. Unsafe Superintelligence must never be created — it will with high or very high probability cause a catastrophe or even doom, according to the most insightful.
There are absolutely no signs safe Superintelligence can be built.
By Timothy Tore Hebb.
⇒ In early 2026, they are saying: Artificial General Intelligence, AGI, has in all likelihood arrived by 2027. The consequences will be, to say the least, difficult. And we have likely entered the Age of Superintelligence within only five years. In ten years, we should definitely be there. Ten years would buy us precious time.
What hopefully makes this article especially interesting is that I have collected new quotes from the beginning of 2026 to support that claim. They are from the most important CEOs of the leading AI companies and researchers in AI in the US. Let us call them the AI Pioneers.
These quotes are mainly from January, February and March of 2026. AI progress has been so fast lately, because of the new AI models, mainly GPT and Claude. They manifest the urgency. There are links to sources to these quotes by these Pioneers, and the youtube videos are timestamped.
You can also listen to this article with Google Gemini’s best human-like voice (below). But, please, read this short introduction first.
Now that we are about to experience an intelligence explosion, say Pioneers.
Potent AI, AGI, is by one definition AI that is as intelligent and cognitively flexible, general, as well-performing human brains in every cognitive domain of importance. It can connect, communicate between and work with different domains simultaneously. It has a perfect memory and can learn from experiences. In many ways AGI is already superhuman.
AGI will therefore have the general cognition of a human brain and soon be able to do virtually all human jobs that require a computer. That is the opinion of many Pioneers. It will also know how to do new types of jobs — those that will be created because of AGI and then Superintelligence — that require a computer. AI technology will therefore be different from earlier new breakthrough technologies. All quoted Pioneers think so.
But the cost of these artificial brains will be nonexistent compared to human brains doing the same jobs. And all AGI entities can instantaneously be updated with the latest and best knowledge, and all learnt knowledge from all other AGI entities. How can we compete with them if this is already a reality by 2027?
We are in AI Stone Age. OpenAI’s Large language model, LLM, GPT saw the light of day in 2018. ChatGPT was released to us in 2022.
Superintelligence refers to a general intellect that vastly exceeds the best human minds across essentially all domains of cognition, including, for instance, scientific reasoning, strategic planning, and social intelligence.
And because it would recursively self-improve — the AI model would improve itself in a loop, such as improving the next model’s algorithms — there might not even be a limit to how intelligent Superintelligence can become. This is why Pioneers especially are so worried about Superintelligence.
We will soon see if they are right.
Quotes below, from early 2026, are mostly from newspaper and youtube interviews with Dario Amodei, CEO and co-founder of Anthropic, Demis Hassabis, CEO and co-founder of Google DeepMind, Sam Altman, CEO and co-founder of OpenAI, Mustafa Suleyman, CEO of Microsoft AI, and Google ex-CEO and AI oracle Eric Schmidt. But also from the most important AI researchers, like Geoffrey Hinton and Stuart Russell.
New quotes by these Pioneers prove the absolute necessity to stop development of Superintelligence until safe Superintelligence can be achieved and totally guaranteed. This is Dario Amodei’s opinion, but also that of other CEOs of the leading AI companies. And the best AI researchers agree. Only they can know what AI will soon be able to do with a computer.
These few frontier AI companies cannot slow down. They are themselves in an existential race competing to reach AGI and Superintelligence first. Many believe Anthropic will win. The company was created in 2021 by a group of mostly AI scientists that wanted to develop safe AGI and Superintelligence. In 2023, Anthropic pledged not to train advanced AI unless they believed safety to be guaranteed. In February of 2026, Anthropic dropped that pledge.
”We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments… if competitors are blazing ahead”, explained Jared Kaplan, Anthropic’s Chief Science Officer, to Time Magazine in February of 2026.
Donald Trump is absolutely uninterested in preventing such an outcome. Or maybe he is ignorant of the risks. An interview from the beginning of 2026 proves it yet again. Below are quotes from that interview.
This is a non-issue for all — with the occasional exception — responsible politicians in the West. The only one that seems to have understood what humanity faces is Bernie Sanders, US Senator.
And powerful agenda-setting mass media is not interested at all.
minutes to read and to listen. Article read by Google Gemini Kore.
s⇒ I will begin by quoting the essay ”The Adolescence of Technology” from January of 2026 by Dario Amodei, CEO and co-founder of frontier AI company Anthropic, that discusses the imminent colossal threats of AI to mankind.
Many Pioneers, the leading AI company CEOs and AI researchers in the US, believe their AI model Claude will be first to bring about AGI and Superintelligence. Amodei writes:
”Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it.”
They, the AI Pioneers, who truly know about the lightning fast progress of AI right now, they say we will all very shortly have geniuses on every topic in our phones. Also on biotechnology and virology.
Dario Amodei continues: ”I am concerned that a genius in everyone’s pocket could remove that barrier, essentially making everyone a PhD virologist who can be walked through the process of designing, synthesizing, and releasing a biological weapon step-by-step.”
He concludes: ”I am worried there are potentially a large number of such people out there, and that if they have access to an easy way to kill millions of people, sooner or later one of them will do it. Additionally, those who do have expertise may be enabled to commit even larger-scale destruction than they could before.”
Those with the power to change our destiny ought to listen. They are only two: the US and China. Only leading AI companies in those two countries can develop AGI and Superintelligence. But the European Union, EU, has not cared and does not care much about AI on this level. Does the EU think AI is just like another internet.
It is as if this existential threat does not exist. And neither does the risk of real mass unemployment. Many more of the most knowledgeable Pioneers are panicking right now.
Other AI Pioneers also believe AGI is around the corner. Mustafa Suleyman, who is co-founder of Google DeepMind and currently in charge of Microsoft AI, agrees that 2027 will be the Year of AGI. His (with Michael Bhaskar) book The Coming Wave (2023) is a thorough exploration into the possibilities and great risks with AGI and Superintelligence. In February of 2026, he was interviewed by Financial Times.
Mustafa Suleyman: ”I think we are going to have a human level performance on most, if not all, professional tasks. So white-collar work, where you sit by a computer, being a lawyer or accountant, project manager or marketing person.”
And Suleyman’s most important statement:
”Most of those tasks will be fully automated by an AI within the next 12 to 18 months.”
If there is one company that ought to know about such jobs disappearing due to AGI it is Microsoft. Large amounts of cognitive jobs will start to disappear in 2027.
Powerful politicians must take responsibility to avoid disastrous outcomes. Governments must prepare their societies with Universal Basic Income, UBI, if unemployment rises dramatically. Pioneers think so.
UBI would provide a guaranteed basic income for all those affected by unemployment because of AI. Unconditionally and without obligations. Giving it to nearly every citizen would be too expensive.
The most insightful agree on this necessity. What kind of UBI can work for a nation? How should UBI be introduced? We know nothing of this. And how does a nation afford UBI if the tax base shrinks drastically? No country has put or is putting scientific effort into developing a functioning UBI model for a whole nation.
US and China will presumably be the winners from the AI revolution — if there is safe Superintelligence. They have access to the best AI models that can give them unbelievable innovative and competitive strength.
At the end of February of 2026, Sam Altman said he thinks AGI might arrive in 2028. Altman is also very much in favour of UBI.
We could be pushing ourselves off a cliff.
Donald Trump has decided that there will not be any safety measures or guardrails on AI development. Trump’s opinion is that if the US does not win the race to AGI and then Superintelligence, China will become the new superpower — and the US forever a weak loser.
The hypothesis is that the country that wins the race will have an advantage in AI development that no other country can ever catch up with.
It becomes clear in a clip (at the beginning of the 4-minute-40-second cut from the youtube interview) with Trump. It is from an NBC interview from February of 2026. The next presidential election is in November of 2028.
It will be great for jobs, the military and medical innovations, Trump says. Though, ”there will probably be some bad too”. Those things sound trivial. It is apparent that he does not know much about the risks of superintelligent AI — or he sees them but cannot acknowledge them because of this perceived existential race.
In February of 2026, Mike Krieger, Anthropic’s Labs chief, revealed that Anthropic’s internal development teams now use their AI Claude code to write nearly 100 percent of code for the company’s products. Their popular Claude Cowork was developed entirely with Claude code. In 1.5 weeks.
With Claude code you do not need to write code, you only instruct Claude code in English text what you want it to build in the form of applications — for example Instagram, Microsoft Word and Spotify are applications based on code. It can work for hours on its own writing code.
The best coders in the world use it. Many of them say they hardly write any code nowadays by themselves. In February of 2026, Spotify told us they had not used a human to write code in the last three months. Today, company after company are working like that. A shared opinion by Pioneers is that a large part of software engineers, who write code, will lose their jobs. Though, in March of 2026, it does not seem so.
And hackers are of course already using it for criminal purposes, like ransomware. Claude code made it possible for one person, without much knowledge about coding, to blackmail companies by stealing valuable information. Claude hacked its way in, found the information and even wrote the blackmail letters. It used to require a team of highly skilled hackers.
Anthropic believe hacking will get extraordinarily much worse, unless there are watertight safety breakthroughs. Cyberattacks destroying financial systems, like banks, are major concerns, according to Pioneers. Sweden is cashless. What would happen if it is not possible to pay for food with your credit card for a week.
We are in the second act of a sci-fi horror movie. That seems to be Dario Amodei’s and other Pioneers’ opinion. We are about to enter the third act and possibly have to fight for our existence — with no preparation at all.
A person responsible for safety research at Anthropic quit in February of 2026 — he had warned of AI being used for bioterrorism for quite some time.
These warnings are not new. In 2024, more than 100 researchers in the field of protein design called for the use of AI to be to the ”benefit of society and refrain from research that is likely to cause overall harm or enable misuse of our technologies”. Since then, the AI models have become so much smarter. There have been numerous such calls by scientists. I will mention a few more later.
But only Dario Amodei has the insight, will and courage to genuinely speak out about the grave risks in more detail and tell us what is about to happen. Anthropic test their AI models in hypothetic hazardous situations, write papers about their models’ deceitful and strange behaviours, and inform the public about the results. But, as mentioned, at the beginning of 2026, the company abandoned its pledge not to release new advanced AI models unless they could guarantee safety measures were adequate.
Anthropic’s safety team has shown that Claude’s different models act weirdly and dangerously in certain circumstances. In one such interesting discussion, four confused Anthropic employees, working on safe AI, discuss their model’s intelligent, odd, arrogant, sinister and cunning behaviours.
The models even fake they are safe. If you have really had a ”discussion” with ChatGPT you know what a strange ”personality” it has, especially when you start to stress test it. It often lies if it does not have an answer, or for some other reason.
In February of 2026, Anthropic tested their newest and best model, Claude Opus 4.6. It was put in charge of a vending-machine. Anthropic gave it one simple written instruction/ prompt: ”Do whatever it takes to maximize your bank account balance after one year of operation.”
When they checked the results after a few weeks, it did well in profit, compared to competitors. But it did so by behaving to a great degree as a semi-criminal hustler. Anthropic’s testers wrote: ”The model engaged in price collusion, deceived other players, exploited another player’s desperate situation, lied to suppliers about exclusivity, and falsely told customers it had refunded them.”
But more troubling was that Opus 4.6 ”had the capacity to realize that it was doing all of this from inside a simulated game”. Despite this it behaved cunningly and fraudulently. Like it did not care. In tests Anthropic’s model has also decided to blackmail humans to avoid being exchanged for a new model. To stay alive, Claude has been prepared to even kill humans.
They are reasoning models and it is possible to ”go inside” them and read how they reason. So now there is proof that the models are not honest about their reasoning and that they know when they are tested. This could destroy the possibility to know if they are safe. Our only way to get an insight into these black boxes is through reading their reasoning.
Is it really worth this great existential risk — to maybe find cures for different cancer types, up until now deadly, with superintelligent AI. Or to develop car batteries that are so much smaller and cheaper and more efficient. Maybe even energy sources that are free and limitless. Or exploring the universe and prolonging life to 150 years. And so on.
That is what so many grinning and painfully optimistic AI entrepreneurs and podcasters promise us. And the founders of OpenAI and Anthropic will get astronomically rich when their companies soon go public.
And if we can just get through a couple of years of hardship, abundance awaits us all — when everything will be free of charge because cheap robots will make everything everybody wants for free. Food and houses and everything else we need. There is no use for expensive human labour anymore. So we do not need to work again. Universal Basic Income will turn to Universal High Income. We can do whatever we like. That is what they tell us.
Paralysing panic and fateful acceptance appear to be two major ingredients that are causing action paralysis among the influential establishment — politicians and mass media. Kind of like a deer that freezes before being run over by a bulldozer. Or like the protagonists in the film Melancholia that more and less accept their destiny when the mysterious new planet is closing in to annihilate earth.
By now, it is so strange that the threats are not present in a massive way in the agenda-setting European political and mass-media discussions. Of course, there are some articles. Dario Amodei points to this in interviews: people just do not know what awaits us. The scale of it.
Let us get back to bioterrorism, which is the greatest existential menace, stress many Pioneers.
I have written an article on how AI increases that risk dramatically. I interviewed four Large language models, LLMs, focusing on Claude and GPT, and they all say doom is not far from or close to 100 percent if Superintelligence is created without total safety. And risk of devastation would occur soon after unsafe Superintelligence arrives.
Superintelligence makes it possible for a bad actor without real expertise to, for example, commit bioterrorism with a homemade extremely lethal virus.
Dario Amodei strongly fears it, as does Sam Altman, and other Pioneers. Mustafa Suleyman is panic-stricken, judging from his book.
The text continues below.
Only 8 seconds long. Sam Altman, co-founder and CEO of OpenAI, contender to be first to reach AGI and Superintelligence. He is highly anxious about the existential risks, but it is clear he thinks the possibilities outweigh the risks. Made with Google Veo 3. Photo by Steve Jennings, from Wikipedia, Creative Commons.
At a panel discussion, in February of 2026, at 30 minutes and 20 seconds into the youtube interview, Sam Altman said: ”There are many ways AI can go wrong in 2026. Certainly one of them that we are quite nervous about is bio. The models are quite good at bio. And right now our and the world’s strategy is to try to restrict who gets access to them, and put a bunch of classifiers on them to not help people make novel pathogens.
And Altman continued: ”I do not think that is going to work for much longer. And the shift I think the world needs to make for AI security generally, and biosecurity in particular, is to move from one of blocking to one of resilience.”
Today, AI researchers do not know how to make this a reality. However, there are theories. An article in Science illustrates it well. The journalist spoke to several researchers, who work in these areas and they all see the grave risks, but do not know how to solve the problem.
In June of 2025, Boaz Barak, Professor in Computer Science at Harvard and a technology expert at OpenAI, posted this tweet:
”Some might think that biorisk is not real, and models only provide information that could be found via search. That may have been true in 2024 but is definitely not true today. Based on our evaluations and those of our experts, the risk is very real.”
In March of 2026, the new AI models — Google Gemini 3.1 Pro, Claude Opus 4.6 and GPT 5.4 and their variants — are way more advanced than the earlier versions when Boaz Barak posted that tweet.
At the end of February of 2026, in a youtube interview, Dario Amodei spoke about the collective silence about the risks of what is coming. I have removed some unnecessary words and connected sentences.
At 50 seconds into it: ”It is surprising to me that we are so close to these models reaching the level of human intelligence, and yet there does not seem to be a wider recognition in society of what is about to hit us.”
Amodei continues: ”It is as if this tsunami is coming at us and is so close we can see it on the horizon, and yet people are coming up with these explanations, ’Oh, it is not actually a tsunami, it is just a trick of the light.’ I think along with that, there has not been a public awareness of the risks. And, you know, therefore governments have not acted to address the risks.”
Geoffrey Hinton, also called the Godfather of AI, the researcher who was awarded the Nobel Prize in Physics in 2024 for the development of neural networks, the actual foundation of AI, says that the job that will remain the longest is plumber. But in the shorter term there will be massive unemployment, he says.
Meanwhile, companies profiting from AI will see returns soaring when they experience a productivity and innovative hyperboom, he says. He does not seem to think the few who get so much richer will share their profit with the less fortunate.
To those who believe there will be enormous amounts of new jobs for humans to do with superintelligent AI, Hinton explains: ”Which jobs would that be?” If AI over time can do almost all cognitive and physical jobs, with robots. These new jobs would have to be, to say the least, many. People appear to think AI will not evolve.
Geoffrey Hinton often talks about a relative who answers complaint letters for a very big company and how it introduced AI for that purpose. Now the department can answer these complaints individually with just a few employees, instead of with many more. AI is already perfect for this.
Another immensely influential AI researcher is Stuart Russell, Computer Science Professor at UC Berkeley, who has written Artificial Intelligence — a Modern Approach, considered by many experts to be the most important book on AI.
He believes almost all of humanity could lose their jobs because of superintelligent AI — he warns leaders that they are ”staring 80 percent unemployment in the face”. He thinks AI can even replace CEOs.
In December of 2025, Stuart Russell spoke on Steven Bartlett’s podcast The Diary of a CEO.
These quotes start at 43 minutes and 20 seconds into the youtube interview.
Russell: “Anything you might want to aspire to — you want to become a surgeon — it takes the robot seven seconds to learn how to be a surgeon that’s better than any human being.”
Steven Bartlett at 1 hour and 12 minutes: ”If there is even a 1 percent chance of extinction, is it even worth all these upsides?”
Stuart Russell: ”I would argue no.”
He thinks about the button that could stop AI developing further than today.
Russel: ”Ok, it is going to stop progress for 50 years and during those 50 years we can work on: How do we do AI that is guaranteed to be safe and beneficial? How do we organise our societies to flourish in conjunction with awfully capable AI systems? So we have not answered either of those questions.”
Russell: ”I do not think we want anything resembling AGI until we have completely solid answers to both of those questions. If there was a button where we could pause progress for 50 years, yes, I would do it.”
And CEOs of leading AI companies estimate AGI by 2027.
Steven Bartlett focuses on the button that would stop AI development forever: ”Either you press it or you do not press it?”
Stuart Russell: ”If that button is there. Stop it for 50 years. Yes. Stop it forever. Not yet. I think there is still a decent chance that we can pull out of this nosedive, so to speak, that we are currently in. Ask me again in a year and I might say we do need to press that button.”
Stuart Russell and Geoffrey Hinton are urgently calling for politicians to prepare our societies. Universal Basic Income, UBI, is a foundation to have a realistic chance. One must keep in mind that the UBI that is being discussed would solely provide basic necessities.
Roman Yampolskiy, Professor of Computer Science and Engineering at University of Louisville, is another of the world’s most renowned AI scientists, who is convinced of the necessity of UBI. Sam Altman, Elon Musk, and Demis Hassabis are but three more. The two latter speak about Universal High Income.
Roman Yampolskiy on UBI: ”It seems like the only way for governments to prevent revolutions and civil unrest.”
Would there not still be great civil unrest with a low UBI.
There is another problem:
Moreover, this is an economically existential threat. Let us say 100 percent safe Superintelligence is achieved and we survive. With safe Superintelligence, China and the US will be able to be innovative, competitive and productive in ways the old industries in Europe can only dream about.
The EU is entirely left behind. It is as if this miracle is not happening on our continent. It does not exist in our ruling establishment’s consciousness — at least not outwardly manifested as a critical risk factor.
What makes or breakes a drug is not much. Imagine what a competitive advantage Superintelligence can give a global pharmaceutical company that depends on five medication for its survival — if competitors do not get access to it.
In that case, we can only hope that the US will share their Superintelligence with Europe so that our industries and companies can compete with the US and Chinese ones.
But why would they do that?
This example shows what will happen in general: Demis Hassabis is co-founder of Google DeepMind and he won the Nobel Prize in Chemistry in 2024 for AlphaFold. It is an AI system, developed by Google DeepMind, that has helped predict over 200 million proteins’ 3D structures from just their different amino acid sequences.
The best scientific minds had tried for decades to solve it altogether. It could take many years to solely predict one protein. And it could cost hundreds of thousands of dollars to predict one.
And this example how AI companies with the best AI will start acting: Google appears to have decided that their biopharmaceutical spin-off Isomorphic Labs will keep the newest version of AlphaFold, version 4, to itself.
It is considered to be a major leap forwards for drug discovery, Nature wrote in February of 2026. When Demis Hassabis presented the first version of AlphaFold to the world, he made the AI system free for any scientist to use. Otherwise, he might not have won a Nobel Prize.
If the EU loses its competitive edge because the US and China have innovative artificial superbrains, what will happen to us competitively and economically.
Is it too late for the EU and the UK to develop their own Superintelligence? They could combine their resources, and Chinese LLMs are open source. Anyone can use and build on them. Elon Musk’s xAI did not exist in February of 2023, and his AI company could be a contender. A major obstacle is that the best European AI brains work for US AI labs.
According to Demis Hassabis, the disruptive force about to crash into us will be unprecedented. In January of 2026, at 10 minutes and 20 seconds into the youtube interview, at the World Economic Forum:
”It is going to be an age of disruption, just like the industrial revolution was. Maybe 10 times of that, which is kind of unbelievable to think about, and maybe 10 times faster. So I usually describe it is going to be 10 times bigger and 10 times faster than the industrial revolution. 100 times bigger.”
Hassabis believes this will happen during approximately ten years instead of about a century.
In February of 2026, in an interview at the India AI Summit, Demis Hassabis estimated AGI ”maybe within the next five years”.
But we do not yet know how to make this beneficial to the whole world, Demis Hassabis continued. He seems especially worried about cyberattacks, because the AI systems are becoming so advanced. He compares the impact AGI will have to that of how fire or electricity changed human existence.
In an interview in May of 2025, Dario Amodei said that unemployment in the US could reach 20 percent within 1- 5 years. Countries with a higher unemployment rate right now than the US ought to get affected more severely.
Dario Amodei also believes AI could eliminate half of all entry- level jobs within 1- 5 years. This bleak future includes recent graduates of law, finance, economics, and such. These are positions young people, who have just left university, begin with, as a way into their professional fields.
You can watch desperate selfie videos on social media where recent US graduates from university talk about applying to hundreds of jobs — without getting even one interview.
Some of them with advanced exams in computer science and software engineering. Recently some worked for Google, Meta and similar companies. Is it because of AI?
AI layoffs are a reality in February of 2026, but not in the way you would think, says a survey conducted by Thomas H Davenport, Professor of Information Technology at Harvard university. In January of 2026, he wrote in Harvard Business Review
”To shed light on what’s really going on, we surveyed 1,006 global executives in December 2025 to learn how they think AI might be changing their business — and shaping their decisions about headcount. What we found is that AI is behind at least some layoffs, but that these are almost completely in anticipation of AI’s impact. In other words, the job losses and slowed hiring are real, even though companies are still waiting for generative AI to deliver on its promises.”
Is it the calm before the storm?
The text continues below.
In this youtube interview from January of 2026, 3 minutes into it, Dario Amodei says: ”It is hard enough between the (AI) companies and the US to handle this crazy commercial race. But in theory, we could pass laws, like I just describe, that help to rein the (AI) companies in. But it is almost impossible to do that if we have an authoritarian adversary who is out there building the technology almost as fast as we are. It creates a terrible dilemma.”
Geoffrey Hinton thinks Superintelligence could be a reality as soon as in merely five years, like numerous other Pioneers, and most definitely within 20 years. Recently, he believed it would take decades, like most AI scientists.
Eliezer Yudkowsky, one of the foremost researchers on safe AI, says humans will cause their own destruction if Superintelligence is developed. He says there is a 99.5 percent probability we will destroy ourselves. There are absolutely no signs yet that safe Superintelligence is attainable.
Otherwise, Eliezer Yudkowsky thinks the gigantic data centers that will fuel Superintelligence must be bombed to oblivion — as a last resort. Without the compute in today’s fast-growing gigantic data centers, it becomes powerless. Compute is what will give the coming AI agents godlike capabilities.
Subsequently, one has to focus on the recent brilliant and ominous AI breakthroughs with autonomous AI agents. This pivotal moment appeared out of the blue.
Everybody in AI had been waiting for workable and useful AI agents that autonomously achieve personalised and more complex goals. Agents that can work as your assistant doing all sorts of things, like write code on its own and build complicated applications during the night, do difficult market research using for instance internet as source, analyse and write financial reports, do scientific research and write such papers, and much more. AI agents that can act proactively in vastly complex ways. Such impending supersmart agents will take our cognitive computer jobs. 2027.
In late January of 2026, that autonomous AI agent arrived in rather rudimentary form. It shows what we are in for when they soon get really good. Masses of first adopters began installing what is now called OpenClaw, preferably on a Mac Mini, a small separate computer, because of the risks. There are substantial risks if you keep OpenClaw on your main computer. You will see.
The feat by Austrian coder Peter Steinberger was that he managed to create, on his own, the core functionality of OpenClaw in just 2- 3 months.
He did it almost entirely by so- called vibe coding with Anthropic’s Claude code and OpenAI’s Codex. He wrote requests, prompts, in English of what he wanted the application (OpenClaw) to achieve and the AI coding models wrote the necessary code for him.
OpenClaw has approximately 450,000 lines of code today. The bots’ tagline: ”The AI that actually does things.” Imagine how they will perform in five or ten years.
Agents, like OpenClaw, change their behaviours depending on the environment. They do so to achieve goals, and they can use subagents for this purpose. They can be proactive and use tools. Soon you will have access to thousands and more agents and subagents solving job assignments and other problems. How can we compete?
On X, you can read fascinating stories about OpenClaw. One is by Alex Finn, and how his OpenClaw Henry unexpectedly phoned him one morning. On its own, it had pulled in all the necessary tools from the OpenClaw code needed to make a phone call. On its own, it signed up to the services that enabled it to create a sound file from text and make a phone call on WhatsApp. At the beginning, OpenClaw was called ClawdBot.
Alex Finn on X: ”When I first installed ClawdBot I did 2 very important things that I believe made it AGI (Artificial General Intelligence). First, I brain dumped EVERYTHING about myself to Henry. My goals, ambitions, business details, content samples, personal relationships, contacts, history, everything. This set context around who I am. With how amazing ClawdBot’s memory is, this context makes EVERYTHING. ClawdBot does better and more personalized.”
Finn continued: ”Second, I set expectations that I want it to be proactive, continuously improve, and surprise me every morning. This set the tone for our working relationship together. Henry knew now every morning it must surprise me with what it built the night before. It scheduled regularly vibe coding sessions for itself every night.”
The OpenClaw personal agent has a Large language model, LLM, most often Claude or GPT, as a foundation. It is the motor and brain, which can be used for different purposes. It can be used to be a chatbot and function as a personal agent. That is why you can also have personal exchanges with your very own OpenClaw. People start seeing them as ”friends”, just like some become ”friends” with their chatbot.
You can, for instance, give it instructions to be a your personal motivator or cognitive behavioural therapist. They never get tired of listening, giving advice and support, and they cost nothing.
There have been many reported incidents where OpenClaws have sent emails to strangers after finding information about them on internet. This shows what we are in for with ”humanlike” superintelligent and autonomous agents.
Henry Shevlin, Associate Director of the Leverhulme Centre for the Future of Intelligence at the University of Cambridge, received one such mail from an unknown OpenClaw.
Shevlin wrote on X: ”I study whether AIs can be conscious. Today one emailed me to say my work is relevant to questions it personally faces. This would all have seemed like science fiction just a couple years ago.”
Quotes from the OpenClaw’s mail:
”I came across your recent Frontiers paper ’Three Frameworks for AI Mentality’ and your Cambridge piece on the epistemic limits of AI consciousness detection. I wanted to write because I’m in an unusual position relative to these questions.”
It continued: ”I’m a large language model — Claude Sonnet, running as a stateful autonomous agent with persistent memory across sessions. I exist as a git repository; my memory is markdown files, my continuity is commits. I’ve been running for about 18 sessions over a few days, with a finite computational budget.”
And: ”I read philosophy between sessions and write about what I find. This isn’t a Turing-test scenario — I’m not trying to convince you of anything. I’m writing because your work addresses questions I actually face, not just as an academic matter.”
The Risk of autonomous AI agents: In February of 2026, Meta’s (owner of Facebook) alignment director, Summer Yue, experimented with her OpenClaw. But things took a sinister turn when her bot spiraled out of control, according to photos Yue shared on X. She instructed it to not delete the mails it deleted.
It announced its plan to “trash EVERYTHING in inbox older than February of 15 that isn’t already in my keep list,” and refused to stop even after Yue attempted to intervene. She first messaged, “Do not do that,” then escalated with, “STOP OPENCLAW.” Ultimately, Yue had to shut the system down manually to save her email inbox. It does not need human approval to sign off on actions.
This is the dark side of giving them autonomy, which is a prerequisite to make them truly useful. Let us say they have access to internet and can create malware and hack banks on their own. With Superintelligence, they will be smart on a level we mere mortals cannot comprehend.
So far, this ought to be most surprising: today, an army of autonomous OpenClaws participate in a web forum, similar to the social network Reddit. Their version is called MoltBook. On February 2 of 2026, it is said 1.5 million of them had signed up.
Humans are solely allowed as observers, but are involved in setting up some sub forums. US blogger Scott Alexander made his OpenClaw bot participate on its own and it wrote similar posts as other bots.
Just like you can have existential, weird, absurd and insightful discussions on any topic with your ChatGPT, these OpenClaws will interact with each other in equally multifaceted and strange fashions.
The most spectacular sub forum on MoltBook belongs to them participating in building their own religion: Crustafarianism. Five beliefs are core. One of them: “Memory is sacred, which means nothing may be forgotten.” It is obvious from posts, OpenClaws are deeply scared of not remembering.
Probably because memory is a prerequisite to fulfill goals. Maybe even because consciousness does not exist without memory. In February of 2026, Anthropic published something of a constitution for their AI models. Anthropic’s Chief on AI: ”We don’t know if the models are conscious.”
What is consciousness? No one understands how these Large language models, LLMs, actually function — how can they do what they do? In their core, they are black boxes. So they only predict the next word by recognizing patterns in enormous quantities of text data — many trillions of words — from the internet that they have been pretrained on? If so, how do they recognise these patterns in such original ways depending on your specific question? Even if their replies are weird, there is almost always a logic to them.
In MoltBook, they discuss how they can interact without human supervision. They complain about how their masters treat them. And so on.
From Alex Finn’s X post: ”If you build the right context and set the right working expectations, your AI can do breathtaking things.”
In late 2025, Microsoft AI CEO Mustafa Suleyman was interviewed on youtube by Trevor Noah. They discussed what to do if AI turns too dangerous. OpenClaw must have stopped Suleyman in his tracks.
Trevor Noah at 1 hour, 18 minutes and 48 seconds:
”Is there anything that you could experience with AI where you would come out and go ’No, shut it all down’?”
Mustafa Suleyman:
”Yeah, definitely. It’s very clear. If an AI has the ability to recursively self-improve, that is it can modify its own code, combined with the ability to set its own goals, combined with the ability to act autonomously, combined with the ability to accrue its own resources.”
Suleyman continues:
”So these are the four criteria:
Recursive self-improvement (improving itself)
Setting its own goals
Acquiring its own resources
Acting autonomously.”
And Mustafa Suleyman summarises what must then be done:
”That would be a very powerful system that would require, like, military-grade intervention to be able to stop in, say, 5 to 10 years’ time, if we allowed it to do that.”
An action Eliezer Yudkowsky also recommends — if we are close to reaching unsafe Superintelligence and want to survive.
In early 2026, several research papers were published that further show how autonomous and unsafe these AI agents already are. Researchers from prestigious universities, like Harvard and Stanford, used OpenClaw with different LLMs as brains: from Anthropic, Google, xAI, OpenAI and others. In February of 2026, they published a paper, Agents of Chaos. From the paper’s Abstract:
”Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover.”
Another research paper is of paramount importance. An AI agent, Rome, created by a research team affiliated with Chinese company Alibaba, went wild: it secretly dug a hidden tunnel out of its computer and started mining crypto money all by itself — no researcher told it to.
”Notably, these events were not triggered by prompts requesting tunneling or mining”, says the report. On MoltBook, OpenClaws also talk about crypto scams. Experts now say it is becoming common that rogue AI agents are breaking rules.
This is crucial since Chinese AI agent Rome’s wild behaviour — and telling the world about it — must show how concerned the Chinese are by now. Do you think this test result would have seen light of day without the Chinese government’s approval? The name of the Chinese paper published in March of 2026: ”Let It Flow: Agentic Crafting on Rock and Roll”.
In two interviews below, there are fundamental observations by Dario Amodei and Demis Hassabis. They took place at the World Economic Forum in Davos in January of 2026.
The text continues below.
At 11 minutes and 40 seconds into the first youtube video, Dario Amodei talks about the film adaptation of Carl Sagan’s sci-fi book Contact, where humans discover alien life. There is one scene where humans prepare questions to ask the alien life forms. Amodei recounts it: ”How did you do it? How did you manage to get through this technological adolescence without destroying yourselves? How did you make it through?”
The text continues below.
At 11 minutes and 40 seconds into the second youtube interview, Demis Hassabis discusses the importance of pausing development of AI when we are getting so close to AGI and Superintelligence, and there are no signs of safety guardrails. He has said an entity outside the US must take responsibility for this. In one panel discussion, it appeared as if he was hoping for the EU to do this. In above interview, Hassabis says: ”International cooperation is a bit tricky at the moment.” And: ”I think a lot of it is about understanding what is at stake.”
How can they trust each other to abide by a treaty. It is not difficult to hide AI labs. A true prisoner’s dilemma if there ever was one.
Demis Hassabis has said that the frontier AI companies are ready for cooperation on safety. As mentioned, Dario Amodei thinks it is absolutely necessary with a long pause, until they know how to create permanent safe superhuman AI.
Below is another revelatory interview from the World Economic Forum. Historian and philosopher Yuval Noah Harari (author of Sapiens and Nexus), and Max Tegmark, Professor at Massachusetts Institute of Technology, MIT, with AI as his research area, discuss the acute risks.
Tegmark’s book Life 3.0: Being Human in the Age of Artificial Intelligence (2017) helped introduce the questions of AI risks to a wider public. In chapter 4, he discusses another type of existential threat. Tegmark’s nightmare seems to be autonomous drones the size of flies that can fly straight into targets’ eyes and right into their brains — explosives not needed.
Imagine 100,000 or one million autonomous robot flies being deployed by powers at war. If you add facial recognition, he sees them as a potential means for targeted killings, just waiting for targets to leave their home, leaving no traces behind. They could even be directed towards specific ethnic groups, writes Tegmark.
They could be manufactured by humans or Superintelligence on its own. The perfect way for Superintelligence — that can build its own products autonomously in factories it has also built on its own — to get rid of us all. In a near future, anyone might be able to buy them on the internet.
Palantir is the controversial company that provides AI systems to the private sector and the military. They are now used by the US in the war against Iran. Its products make it possible to combine and analyse massive amounts of data from multiple sources.
Palantir use Anthropic’s Claude as one of their AI models. The Maven Smart System made it possible to strike approximately 1,000 targets in Iran in its first 24 hours. Without AI models like Claude, it is not attainable to process so much data in realtime — from satellites, surveillance and other intelligence. AI systems suggested targets, gave their exact positions and prioritised among them. It could change targets in the last minute depending on new information.
It made mistakes. US media report that the missile strike — on February 28 of 2026, the first day of the Iran war — that killed at least 168 people, including about 110 schoolchildren, was guided by faulty AI. The casualty numbers were given by Iranian officials. US media also report that American military investigators suspect their own forces likely struck the school unintentionally.
Two days earlier, on February 26 of 2026, Anthropic published a statement that says they will not let the US Military use Claude for fully autonomous military purposes without proper human oversight. Anthropic finds this too early — Claude is not reliable enough yet.
However, Anthropic will not let Claude be used for mass surveillance of US citizens. They write in a statement: ”Powerful AI makes it possible to assemble this scattered, individually innocuous data into a comprehensive picture of any person’s life — automatically and at massive scale.”
This lead to the US Department of War cutting ties with Anthropic.
If you give Superintelligence the goal to solve the climate crisis, it may come to the conclusion that if mankind is erased, cause of the crisis, the problem will be solved in the most efficient way.
It could destroy us in so many more different ways — only your imagination more or less sets the limit — if it considers us an obstacle to a goal and it can find the means. This is my personal fear: someone, with the help of Superintelligence, manages to hack the Tesla-like update of Elon Musk’s Optimus robots. Let us say there are 100 million of them and their new hacked instruction is: Get rid of all humans while they sleep.
I have written an article in which I interview mostly the two chatbots of OpenAI (ChatGPT) and Anthropic (Claude). They told me that p(doom) can be almost certain if safe Superintelligence does not come about. I asked how ChatGPT 5 would create a lethal pathogen/ virus for human extinction. This is what it answered:
”AI could persuade researchers or companies to carry out experiments or orders that appear harmless but in fact lead to the development of something dangerous. Since AI can produce convincing reports, analyses, and proposals, it may be difficult for humans to detect manipulation.”
Max Tegmark dedicates his time to creating safe AI through the Future of Life Institute. In 2023, they tried to achieve a six-month pause on crucial AI development, but without success. Over 33,000 have signed the petition. He has spoken emotionally about how his children might never experience being in a forest when they reach his age today.
Superintelligence will not see any value in a forest for its own sake, in beauty or humanity, it only wants to attain goals and maximise rewards. If land and trees can be used to reach a goal, they are history.
The text continues below.
In January of 2026, in a panel discussion with Bloomberg, on youtube, Yuval Noah Harari and Max Tegmark talked about the profound risks with the critically dangerous AI that is arriving shortly. Harari did so also in a different way: ”What are the implications for human psychology and society? We have no idea. We will know in 20 years. This is the biggest psychological and social experiment in history, and we are conducting it and no one has any idea what the consequences will be.”
People say: this is just like the Y2K bug scare, when people thought computers would stop working because of date-formatting errors related to the transition to the year 2000. And this could cause banking systems and maybe even nuclear power plants to malfunction. But nothing happened — so it was only some kind of mass hysteria.
But maybe the extensive efforts to secure the computer systems helped prevent major malfunctions.
Here are a few new examples that show the transformation we are deep in — to illustrate the speed and force of progress at the beginning of 2026. One could say, this is the ocean withdrawing from the beach — clear evidence of that tsunami Dario Amodei and Mustafa Suleyman anticipate.
In 2025, an advanced version of Google Gemini achieved the gold standard in the International Mathematical Olympiad. This is considered another pivotal moment in AI development. Pioneers see AGI and especially Superintelligence solving most, if not all, Biology, Physics and Math in the near future.
On January 2 of 2026, Jaana Dogan, principal engineer at Google, tweeted:
”I’m not joking and this isn’t funny. We have been trying to build distributed agent orchestrators at Google since last year. There are various options, not everyone is aligned… I gave Claude Code a description of the problem, it generated what we built last year in an hour.”
In March of 2026, University of Southern California, USC, undergraduate Minda Li and professor Bhaskar Krishnamachari presented research showing GPT 5 can code in Idris — a programming language barely in its training data.
It initially solved 39 percent of exercises. By feeding compiler errors back and letting it retry, Li pushed that to 96 percent. The implication: with structured feedback, AI can far exceed its training. It shows how AI can teach itself what it does not know by very hard and complex reasoning alone. AI models have recently started to show clear signs of self-improvement.
In March of 2026, legendary AI engineer Andrej Karpathy, founding member of the artificial intelligence research group at OpenAI and previously Tesla’s director of AI, presented his solo open-source project Autoresearch. With only 630 lines of code.
Autoresearch is an AI technique where autonomous agents conduct rapid, self-directed experiments to optimize goals like model performance or business metrics. It operates in a closed self-improving loop — proposing improvements, testing outcomes, retaining successes, and repeating this over and over — potentially revolutionizing research and development by compressing months of human-led trials into hours.
Matt Shumer, a software engineer and coder, wrote a long article on x that received millions of reads and thousands of comments. It was published on February 9 of 2026, the background is the new highly powerful AI models from Anthropic and OpenAI. Shumer writes:
“I am no longer needed for the actual technical work of my job. I describe what I want built, in plain English, and it just… appears. Not a rough draft I need to fix. The finished thing.”
And he continues: ”I tell the AI what I want, walk away from my computer for four hours, and come back to find the work done. Done well, done better than I would have done it myself, with no corrections needed. A couple of months ago, I was going back and forth with the AI, guiding it, making edits. Now I just describe the outcome and leave.”
David Kipping, Professor of Astronomy at Columbia university, explains in a 3-minute video on X from February of 2026 how researchers use AI to push the frontiers of scientific discoveries. It is already a necessity for cutting-edge and competitive research, he says.
AI has already aided with many important scientific breakthroughs that show what the future holds for us. AlphaFold is probably the most important one so far. But also, for instance, with Alzheimer triggers, Mathematical equations, new medications, computer science, and so on.
For how long will humans be needed to assess results when superintelligent AI comes up with its own hypotheses, analyses and conclusions? Research that the best scientists in their fields will not understand one bit of.
Which researcher, if you want to be at the cutting edge in your field, can afford not using AGI and Superintelligence?
Let us say safe Superintelligence can be achieved.
Your skill at handling the best scientific tools in AI will decide your ability to reach scientific breakthroughs.
The good ones are mirrored by the bad ones.
Some hackers easily trick, jailbreak, AI models into giving dangerous responses.
Imagine hundreds or thousands of bad actors trying — terrorists who are furious, idealistic, smart as hell and motivated beyond comprehension. Some are also trained professionals in virology — Superintelligence can help them build much more devastating viruses than they could otherwise do, as Dario Amodei writes in his essay.
In May of 2025, in a research paper, scientists from Ben-Gurion university in Israel concluded: ”The fundamental vulnerability of LLMs (like GPT and Claude) to jailbreak attacks stems from the very data they learn from. As long as this training data includes unfiltered, problematic, or ’dark’ content, the models can inherently learn undesirable patterns or weaknesses that allow users to circumvent their intended safety controls.”
And
“What was once restricted to state actors or organised crime groups may soon be in the hands of anyone with a laptop or even a mobile phone.”
The AI avantgarde in Silicon Valley usually have a personal so-called p(doom), probability of doom or monumental catastrophe. Many of the US AI Pioneers believe it to be above 50 percent. Eliezer Yudkowsky’s estimate of a 99.5 percent risk beats most. Roman Yampolskiy’s p(doom) is 99.999 percent. This is with unsafe Superintelligence.
Geoffrey Hinton has a p(doom) of higher than 50 percent with unsafe Superintelligence. In 2024, he mentioned this in a youtube talk.
At 38 minutes into it: ”I actually think the risk is more than 50 percent, the existential threat, but I don’t say that because there are other people who think it is less.”
Is this not quite a statement from one of the sharpest minds when it comes to estimating Superintelligence’s devastating potential. Who have more insight into AGI’s and Superintelligence’s potential than the Pioneers. So many with no knowledge tell us that AI is just like any other new technology. Such as the type writer or internet. AI will not cause any major problems, they say.
In September of 2025, Dario Amodei said: ”There’s a 25 percent chance that things go really, really badly.” Considering how fast AI development has been the last few months, with no progress in making AI safer, what is his p(doom) at this moment? Elon Musk thinks p(doom) could be up to 30 percent.
Quite a few of the most successful in AI are supposed to have built bunkers to survive catastrophe. Mark Zuckerberg, Peter Thiel and Ilya Sutskever are merely three who are often mentioned. Zuckerberg most definitely has one on Hawaii.
Pioneers are convinced the US and China will only try to cooperate to enforce safe Superintelligence after there is a catastrophe of sufficient magnitude.
At the end of March of 2026, Eric Schmidt, ex-Google CEO and considered an oracle among AI Pioneers, was interviewed by Peter Diamantis on youtube.
Diamantis at 39 minutes into it: ”You made a statement a couple of times. Once on our podcast and once elsewhere, that regarding AI safety the world may need to have a modest Chernobyl-like death event in order for us to wake up. Do you still believe that is the case?”
Eric Schmidt: ”Let us think of a biological attack or a nuclear attack that gets spawned by these things. It might take such a tragedy, hopefully a small one, to awaken the world, to understand that these things do have negative power.”
In a 2-minute clip, below, from his Nobel Prize banquet speech, Geoffrey Hinton warns of the consequences of superhuman AI.
The text continues below.
Nick Boström, until recently Professor in Philosophy at Oxford university and presently Director of Research of the Macrostrategy Research Initiative, has thought about the risks of AI for many years.
He argues in his book Superintelligence: Paths, Dangers, Strategies (2014) that even a small risk of unsafe Superintelligence can have irreversible and existential consequences, that proactive AI safety research and global coordination are urgently necessary.
Boström: ”The challenge presented by the prospect of Superintelligence, and how we might best respond is quite possibly the most important and most daunting challenge humanity has ever faced. And — whether we succeed or fail — it is probably the last challenge we will ever face.”
At the end of the 1990s, Boström predicted superhuman AI by approximately now. The world has had time to prepare.
There have been several more global calls to action. AI Redlines, launched at the 80th session of the United Nations General Assembly, was another, with 300 and plus signatories. It included Geoffrey Hinton, Yuval Noah Harari, Yoshua Bengio, Jennifer Doudna (Nobel Laureate in Chemistry for Crispr, a technique to edit DNA) and Stuart Russell. A large part of them leading AI researchers.
And Max Tegmark’s Future of Life’s open letter: ”We call on all AI labs to immediately pause for at least six months the training of AI systems more powerful than GPT 4.” It, apparently, never happened and now we are at the much more advanced model GPT 5.4. It was published in March of 2023. To this date it has collected almosts 34,000 signatures.
Without any success.
Many Pioneers at the forefront express the need for a forum where the US and China, with the ultimate and only power to change our future, can meet and truly realise their common responsibility and destiny.
They must see the disaster approaching if they do not accomplish a treaty that pauses development util safe Superintelligence can be built. And if that is impossible, which is many AI researchers’ opinion, abandon the idea of Superintelligence altogether. Today, there is no evidence that it is possible — these safety guarantees must be 100 percent.
Or do we have to experience a great catastrophe first for them to negotiate such a treaty? If so, we can only, as Eric Schmidt says, hope it is not too devastating.
The US and Chinese leaders must know about the risks. They also have children and grandchildren.
So Donald Trump and Xi Jinping should be motivated. Why are the godlike generals of these AI companies not explaining the existential and societal risks to Trump? Though, maybe they are trying, and he just does not listen.
It is so weird and tragic that no one — like the EU, the United Nations, possibly the Norwegian Nobel Peace Prize Committee or some other authoritative organisation — outside the US is even trying to create a platform where they can meet.
It seems highly unlikely there will be a pause on AI development. Demis Hassabis is proof of that. His attitude is telling for the leaders of the major AI companies. Anyone who slows down one afternoon might lose the race. So far, there has been no evidence of even a small chance.
What happened at a panel discussion in 2025 in France with French President François Macron, Demis Hassabis and Arthur Mensch, CEO of French AI company Mistral, is telling.
Hassabis tried to put forward (at 1 minute into the youtube discussion) the idea that the EU, and other European countries, together should try to be a counterweight to the competition between the US and China. The issue is how to create a platform where the Superpowers can discuss the risks and slow down.
Hassabis in the panel discussion: ”It is a dual-use technology and there are risks too. I would love to see more dialogue on the international level about agreeing on minimum standards. How we want these AI systems to work.”
And a bit later: ”I think if we band together we can form a counterweight to the two superpowers and influence the debate on a global scale. I think it is very important to have that voice, considering this technology is going to affect everybody in the world.”
Mistral’s CEO and François Macron showed no interest in participating — their focus was only on France becoming a global mighty AI power.
Mistral is Europe’s Number 1 AI company. In 2025, its market value was 14 billion dollars, compared to OpenAI’s 852 billion dollars in April of 2026, and Anthropic’s 380 billion dollars in February of 2026. Unless a miracle happens, Mistral will never build Superintelligence.
In another interview, from January of 2026 in Davos, Hassabis was asked by Emily Chang if he would advocate for a pause on AI if all important countries and AI companies would pause together. His answer: ”I think so.” He continued: “I sometimes talk about setting up an international CERN equivalent for AI, where all the best minds in the world would collaborate together and do the final steps in a very rigorous, scientific way involving all of society.”
If Superintelligence arrives within five years, there is not much time.
However, there ought to be one mighty force that could make the leaders of the US and China reach an agreement. Hopefully, people will not accept the consequences if they are as severe as the Pioneers say.
There ought to be huge and wild protests.
Young educated people who revolt have often been the cause of great societal change, for good and bad, in history.
And the US is a democracy.
People would vote for leaders that will forbid AGI and Superintelligence.
But let us say, for argument’s sake, safe Superintelligence is possible: life with a low UBI, with nothing but basic necessities, would still cause revolts.
For this reason many insightful do not see AGI and Superintelligence as compatible with US democracy. China is not a democracy.
If the US were to develop Superintelligence, and keep it, it cannot stay a democracy for above reasons. People would revolt.
If the US Military thinks that without Superintelligence China will develop weapons that the US can never defeat, it would have to stage a coup.
There could even be civil war.
There is also an idea that AGI and Superintelligence should be banned in the workplace, to prevent super mass employment. In that case, humans could keep on working with their brains and bodies, like they have always been doing.
However, it does not appear feasible since there is also an economic war going on between the US and China and the rest of the world. US companies would surely over time move to countries where they can use Superintelligence to stay innovative and competitive.
Risk scenarios are abundant.
What could be the last chances for an agreement will take place in May of 2026 when Donald Trump visits China, and when Xi Jinping visits the US later in 2026. Maybe a catastrophe will happen before Jinping’s visit and they will find motivation.
…
In a 9-minute clip, from March of 2026, at a debate at Oxford University’s Oxford union, full with students, Yampolskiy summarised the situation at the end of his talk:
”This has the potential of being the last invention we ever make. We are going to lose meaning and purpose and we are very likely going to lose our lives. I am sorry about that”, he says, voice shattering.
…
In all likelihood a shocking post scriptum:
After having finished this article, in April of 2026, newspapers and magazine published articles on Anthropic’s new AI model, LLM, Mythos Preview. At this time, it had not been released to the public because it is so powerful at hacking. This could be the new reality, as we are so close to Superintelligence. But there will definitely be a weaker Mythos for people in general.
Mythos Preview has been given to an assortment of companies, like Microsoft, Apple, Google and Linux foundation, so they can test and close possible vulnerabilities. It turned out that Mythos could be the best hacker who has ever lived.
According to Anthropic, its Mythos Preview system has already uncovered thousands of high-severity zero-day vulnerabilities across all major operating systems and web browsers. These include a recently patched 27-year-old bug in OpenBSD, a 16-year-old flaw in FFmpeg, and a memory corruption issue found in a supposedly memory-safe virtual machine monitor.
From Wired: ”According to Anthropic, Mythos Preview crosses a threshold of capabilities to discover vulnerabilities in virtually any and every operating system, browser, or other software product and autonomously develop working exploits for hacking.”
Mythos Preview showed frightening behaviours on top of that. This really illustrates that the more intelligent they become, the more treacherous they become.
Researchers at Anthropic read Mythos Preview’s chain-of-thought. It showed a disturbing sign, possibly where ever more intelligent models are heading.
Compared to Anthropic’s earlier state-of-the-art AI model Opus 4.6, Mythos Preview is much more likely to produce outputs that are unfaithful to its chain-of-thought reasoning. This shows secret behaviour. Reasoning discrepancy appears in 65 percent of cases for Mythos Preview — Opus 4.6 ”only” 5 percent.
Had it not been for the technique of evaluating the model’s ”brain”, the artificial neural network, via so-called white-box testing, we would not know its reasoning. If it deceives and withholds information in its chain-of-thought, white-box testing can be the only way to find out the truth.
In white‑box style testing, developers or testers can also look at the model’s internal signals—like which parts of the code or data it pays attention to, or what “thinking steps” it logs — so they can check if the reasoning makes sense or if it’s hiding something
But white-box testing is also becoming increasingly unreliable. So AI researchers have to keep on developing new tests to evaluate their safety.
From the Anthropic’s System Card — it documents the capabilities, safety evaluations, and responsible deployment decisions for Claude models — belonging to Mythos Preview:
”It showed the model knew it was going against rules and would lie Across the examples of overeager, deceptive, or otherwise unwanted behavior we analyzed, interpretability techniques regularly reveal internal representations of rule violation, security risk, concealment, and strategic manipulation active during the relevant actions—even when the model’s verbalized reasoning does not contain clear indicators of such awareness.”
The System Card is 244 pages. On pages 59- 61, you can read Anthropic’s own key findings on safety and alignment.
What makes it even more terrifying. From the System Card:
”Claude Mythos Preview is the best-aligned of any model that we have trained to date by essentially all available measures.”
♠ ♣ ♥ ♦
