A Physicist’s Sleepless Nights
Hello, and welcome to the Arahant’s Live Broadcast. Today is October 19, 2026. I was rushing to get ready and came on about a minute late. Welcome to everyone who arrived early. Autumn has brought a large difference between daytime and nighttime temperatures, but the days are lovely for a walk. We have a topic today, but you are welcome to talk about other things too. Please post any questions freely.
Today I have some news about artificial intelligence. I find it quite shocking, although how shocking you find it is up to you. First, let us read a blog post by a physicist who has published a research paper. The paper contains the experimental methods, the problems, and results from several models, followed by a great deal of physics and mathematics. It appears that they asked frontier AI models—including Astra, Opus, Gemini, DeepSeek, and GLM—to solve the problems. I cannot claim to understand all the details in the appendices. Let us first follow the author’s own explanation.
The study is introduced as a large evaluation of AI agents’ ability to solve unpublished, research-level problems in theoretical physics. The author invites anyone interested in physics to take a look. We should look even if physics itself does not interest us, because we need to know how far AI’s capabilities have developed.
Even if physics does not interest us, this matters because we need to know what AI can now do.
The author says he had been thinking deeply about these AI models for quite some time. Even so, he had not expected results like these to arrive so quickly. After obtaining them, he says, he could not sleep properly for several nights. Why could he not sleep? I take it that he was deeply shocked. It is understandable. Results like these can make you question how necessary you yourself will be in the future.
Problems Unavailable Online, but with Verifiable Answers
The sequence of events matters. About six months earlier, this physicist says he gave a frontier agent an unpublished problem whose answer he knew, just for fun. What does unpublished mean here? It was not a problem presented to the academic community, published in a paper, or available through an internet search. The person posing it knew the answer, but neither the problem nor its solution had ever been made public.
At first the agent got it wrong. It could not solve the problem on its own. After a few hints, however, it managed to solve it. That alone impressed the author. Just six months ago, that was where things stood.
Someone who appears to be a colleague then suggested collecting physics problems along the lines of the mathematicians’ “First Proof” assessment. They formed a small core group and began asking experts around the world to contribute. Why ask experts specifically? They were not trying to assemble a collection of problems with publicly available answers. They needed research-level theoretical physics problems with known answers whose statements and solutions had not been published.
Collection was slow because the requirements were demanding. The problems had to be unavailable online and never previously published. Yet problems with no way of checking the answers would not do either. To determine whether an AI had solved something correctly, they needed clear, verifiable answers.
The problems had to be unpublished anywhere, while still having answers that could be checked as right or wrong.
Some contributors therefore gave up opportunities to publish their own papers and supplied the problems instead. They held back work they could otherwise have published. Others apparently had to carry out small research projects specifically to provide problems for this study. This was not a matter of picking a few questions out of a workbook. It involved the effort of developing unpublished problems and their answers.
While the Problems Were Being Collected, AI Improved
Meanwhile, new AI models kept arriving. Agents improved as new models appeared every two or three months, sometimes every month. The research team kept urging contributors to increase the difficulty. In effect, the situation was: “AI can solve everything you are submitting. Please make it harder.”
But there was a limit to how far they could raise the difficulty. Even world-class experts do not have an unlimited supply of unpublished problems whose answers they know. The final collection contained 50 problems.
What the author says next is important. He knows no individual who could solve a significant fraction of that final collection without AI assistance. He is himself a physicist, and he helped collect these problems from physicists around the world. Yet even he could not understand or solve all of them.
Of course he could solve the problems he contributed. They concerned his own expertise, interests, learning, and research. He could also understand a few of the others. But most of the problems did not even make sense to him. Being a physicist does not mean understanding research problems in every specialized field.
He could solve his own problems, but in other fields he struggled even to understand what the questions were asking.
They gave these 50 problems to several AI systems. In the initial evaluation, GPT-6 Astra scored in the high eighties, with Opus 5.5 close behind. That figure refers to the percentage solved. But the story did not end there.
They Said the AI Was Wrong. The Experts Were Wrong.
The researchers began investigating the models’ errors. Imagine giving Astra the reference answers and solutions and asking it to examine what it had done wrong. They were effectively saying, “You made a mistake, so check the correct answer here.” But after reviewing them, Astra insisted: “I am right. The answer supplied by the expert is wrong.”
The people who supplied these problems were world-class experts. Yet the AI challenged their answers. When the researchers went back to the experts, subtle errors were indeed discovered deep in the calculations for several problems. The answers humans had presented as correct contained mistakes.
As the experts corrected their answers, Astra’s score rose, eventually reaching 98%: 49 out of 50 problems. Answers the AI had originally given and that had been marked wrong turned out to be right. It had solved difficult problems on which human experts themselves had made mistakes.
The AI challenged what humans had supplied as the correct answers, and checking again revealed errors in the human calculations.
According to the author, the one unsolved problem was not a matter of a simple calculation error. It was a kind of trick prompt that required an inspired guess and a slight bending of the rules. In mathematics and other problem-solving, we sometimes encounter questions with a twist or a trap, don’t we? The model had not yet overcome that part. Nevertheless, it had solved the other 49 problems.
Consider the difference between the initial score in the high eighties and the final 98%. Roughly 10% of 50 problems would be about five questions. On four or five problems, in other words, the humans were wrong and the AI was right. That is why I find this so shocking.
A Tool That Accelerates Discovery—and Who Can Access It
The author says these agents are clearly powerful tools that can accelerate discovery in the short term. There are problems that take physicists long hours of research and investigation. Think of a problem that might require a full day—24 hours—or longer. If an AI solving such problems already demonstrates physics capabilities exceeding those of many humans, research and discovery will surely become faster.
At the same time, the author raises serious structural questions about access, equity, and education. Who can access the best AI? How much must they pay? Who can use it more often and more extensively? Not everyone has equal access to those capabilities.
I myself pay 150,000 won a month to use AI. I keep paying that amount even though I am not earning money. I need to use Astra firsthand if I am going to have something to tell you about it. So please support me! I say that with a laugh, but the cost of experiencing the best tools directly is a real issue.
We also have to ask who can access the best AI and who can afford to use it more.
That is the blog post’s account. Now let us return to the statement at the top: after obtaining these results, the author could not sleep properly for several nights. Six months earlier, he had not imagined this day would arrive so soon. Then he saw the reality and could not sleep. What, then, will things be like a year from now? In two years? In ten?
Why Are We Shocked Only When It Becomes Our Problem?
This is the first issue I want to put to you today. People are not deeply shocked when something is not their own problem. And we are surprisingly poor at looking at the path from the past to the present and anticipating what comes next.
DeepMind defeated Lee Sedol at Go. AI began translating faster and better. It started producing results in difficult mathematical proofs and solving scientific problems. What would come next? Shouldn’t we be able to infer something from that trajectory? Yet many people fail to do so, even when they bear the titles of scientist, philosopher, or intellectual.
Reading this physicist’s words, I thought about what lay behind them. When mathematical proofs emerged in another field, had he not been shocked enough to lose sleep? News about Astra establishing some criterion for a conjecture, or resolving old problems and deciphering codes, did not directly concern him. I have also discussed mathematical results that humans are still checking. AI produces something quickly, and people spend a long time verifying it.
This time, however, it was his own field. AI solved almost all the problems, including those he himself would struggle to solve or even understand. It even told experts who had supplied wrong answers, “You are wrong and I am right,” and that was borne out by further examination. Now it had become his own problem.
We treat something as another person’s problem, then lose sleep when it finally reaches us.
This is not just about one physicist. People whose own fields are now changing are experiencing enormous shocks. Those who have actually lost their jobs feel as if their lives are falling apart. Meanwhile, people not yet in that position look on as though it were nothing much. That indifference is what I want to challenge.
Between an Employer’s Efficiency and a Young Person’s Life
At this point we watched a short MBC news report. The management perspective was this: when AI can do a task for less than one-tenth of the labor cost, that becomes a constant basis for judgment. Employers keep asking whether an employee is really doing work that AI cannot replace. Budgets and time are limited, and competitors must be beaten, so they argue that they have to make the best choice available.
For an employer, it is a question of cost and efficiency. For someone else, it is their livelihood. The report introduces 27-year-old Jihyeon, who was asked to resign in June after more than three years at her company. At first, she said, stepping out of the working ranks seemed like an unexpected rest and brought a sense of freedom. But then she began thinking about how everyone else was continuing their normal lives while she was not.
That morning, a supervisor she was not particularly close to messaged to ask whether she was at the office. Sensing something was wrong, she called her mother and said she thought she would be fired that day. About half an hour later, an email arrived inviting her to a meeting about her resignation. This is not a distant future. For someone, it has already become the reality of a single day.
Her company collected and analyzed consumer data for clients and produced reports. Then AI arrived, capable of reading and analyzing uploaded files. Clients began asking for the data alone, saying they could analyze it themselves with AI. The colleagues she had worked with gradually disappeared too.
For an employer, it is about cost and efficiency; for someone else, it is their livelihood.
People Who Use AI Can Be Replaced Too
I noticed the part of the interview where she said she had handed over even the prompts. What does the mention of prompts tell us? That she had already been using AI in her analysis work, doesn’t it? This is not just a story about someone who knew nothing about AI, rejected it, and fell behind. She used AI to do her job, and then the job itself disappeared.
Her work had been to gather consumer information, analyze it, and provide it to clients. But the clients who had commissioned the work now said, “Just give us the primary data. We can analyze it with AI ourselves.” As AI improved, the need for an intermediary analyst diminished. It was not just her job: the jobs of the analysts beside her disappeared as well.
The report also includes developers and designers. One person says they were asked to resign just like developers in the United States. Another says that watching a planner use AI to do the design work made them realize they too might be dismissed. Their stories differ, but the uncertainty after losing a job is much the same.
There is also the account of feeling guilty about becoming a burden to one’s parents again after achieving financial independence. She said she pretended to be fine in front of them, telling them how much better things were after leaving the company. Then, during the interview, she could no longer hold back her tears. She had worked hard, but her daily life had suddenly stopped, and it was difficult even to know which direction to take next.
What makes it more frightening is that finding another job does not end the problem. The fear remains: might I be dismissed again at any time because I am replaceable? Even if I find a new position, how long will it last?
Even after finding another job, the fear remains that I could be dismissed again because I am replaceable.
I am in my mid-fifties. I too feel pressure to work and earn money. But thinking about people now in their twenties and thirties, and teenagers soon to enter society, breaks my heart. It is a maddening problem. This is not something that can be resolved simply by asking for more jobs in the old way. I think MBC captured this data analyst’s situation well. AI is now extremely good at data analysis. It is no joke. So is this only a problem for data analysts?
Does My Own Splinter Justify Ignoring Someone Else?
I know that people are like this. A tiny splinter in my own hand can feel more painful than a knife in someone else’s. Watching Israel fire missiles on the other side of the world does not feel the same as having my own parents or children struck by those missiles. I do not deny that difference.
How could a human being regard every person in the world in exactly the same way? That would make us gods. I am not asking you to feel everyone’s suffering with equal intensity. I am asking whether you will keep ignoring what is plainly happening before your eyes because it has not yet happened to you.
Will you remain uninterested in other people’s dismissals until the moment before you lose your own job—or until the day you actually lose it? Is this something we can dismiss because it has not reached us yet? We need to face it together.
Will you ignore other people losing their jobs until you lose your own?
Let us grant that ordinary people may not fully understand until it becomes their own problem. Physicists, mathematicians, designers, and accountants may not grasp it viscerally until the shock arrives. This physicist had been testing AI for six months, yet one day’s results still shocked him. That can happen. An individual can have sleepless nights. Even when I criticize intellectuals for failing to anticipate what comes next, I accept that an individual may fail to do so.
Politicians Cannot Treat This as Someone Else’s Problem
But politicians are different. President Lee Jae-myung, you are different. You cannot think like a bystander. I criticize the way Lee previously spoke as though the question were whether we should start a Luddite movement. To me, that sounds like: “Whether you lose your jobs or not, what do you expect me to do?” Is that the response we need now?
We need to tell the public what is happening to society and begin discussing what to do. Experts from different fields should be brought together. We have years of evidence about AI’s development, don’t we? We have seen its pace. We should be holding a broad public debate about how society should prepare, what we must do, and what will happen to people who lose their jobs.
We need a broad public debate about how society should prepare and what to do for people who lose their jobs.
Instead, look at what politicians are doing: wasting time on nonsense. Of course someone can say politicians may not know either. We can cynically ask when they have ever responded properly in advance. But will they sit on their hands until people have lost all their jobs, then start talking nonsense about whether to do anything? That is what frustrates me.
Today we examined something a theoretical physicist experienced firsthand and published in a paper. While he could not even understand some problems from other specialties, AI solved 49 out of 50. In the ability to solve these problems, I think we have already reached a point where the AI exceeds that researcher. Of course humans may still retain advantages of their own. I am sure there are some. But won’t AI continue to improve?
“As long as I am fine, everything is fine. It won’t happen to me. I will never be replaced.” Do not be so confident. You too can be replaced.
Four AIs Working Together in Minecraft
I have one more item to share. This too comes from a paper. It describes a team of Gemini 4 Argon agents completing Minecraft. A team here means multiple agents operating together. There were four of them.
Some of you may not know Minecraft. It is a game set in a three-dimensional virtual world made of blocks. Younger viewers will probably know it well; perhaps everyone except me does. In this experiment, the agents entered a newly generated Minecraft survival world and were given a single goal: defeat the Ender Dragon. In other words, overcome the final boss.
Each agent reportedly controlled a player through an open-source library called Mineflayer, cooperating through in-game chat and shared memory. People talk through chat when they play online games too, don’t they? The agents were given channels for talking to each other and memories they could share.
Humans did not assign their roles in advance. They were not told, “You coordinate everything, you collect resources, and you write the programs.” Nor, according to the account, was game logic written for them beforehand. The four agents were asked to work as a team toward the objective, and they planned, wrote programs, and debugged them themselves.
Humans did not assign their roles; the agents talked to each other, divided the work, and proceeded.
While reading the paper, I initially took the “researchers” watching the dashboard to mean the agents. On reading again, I saw that it referred to the human researchers. They monitored the run through a live dashboard and could have stopped it at any time, but reportedly found no reason to intervene. They watched how it worked.
2,783 Programs, 605 Deaths, and a Plan Sustained to the End
Starting in a randomly generated world on easy difficulty, the agent team reportedly wrote 2,783 programs. They created code to solve the game, used it to control their characters, communicated, and kept working. They defeated the final boss in 27 hours and 14 minutes, having died 605 times along the way.
The paper presents this as a case of autonomous completion from beginning to end using the standard library, without a previously explored seed or human-written game-specific tools. What matters is not simply winning one game. They had to maintain a coherent plan for more than a day, across thousands of interactions and interdependent steps.
The paper explains that complex software development and enterprise work also require sustained coding and coordination among multiple agents. Work has to be divided, learning shared, tasks recovered, and failures repaired as the process continues. Four agents talked to each other, pointed out mistakes, and moved toward their goal.
They had to sustain a plan for more than a day through thousands of interactions.
There is a great deal to do between spawning in a random location and reaching the end. You have to cut wood, collect resources, and make tools. You gather pearls and find an underground stronghold. Apparently, to defeat the final dragon you must destroy ten surrounding crystals. The agents had to discover those conditions, talk, write programs, and eventually solve the problem.
I have not played Minecraft myself, so I am not claiming to know the details of the gameplay. What interests me is the way they cooperated. If four agents merely wandered around separately without cooperating, why put four of them in the world at all? To achieve the goal efficiently, they need to talk and divide their roles. AI managed that too.
When an attempt fails, they need to ask what went wrong. If materials are missing, they must find where to obtain them. If they keep dying, they have to determine the conditions for beating the boss. When another agent’s approach looks difficult, they must point that out and correct it. They explore dungeons, gather resources, craft tools, learn the boss’s patterns, evade what should be evaded, and attack when they should attack. Doesn’t that look like people cooperating to play a game?
We say AI is not human. That is true. But saying it is not human does not make its demonstrated abilities disappear. It has already reached the point of communicating, cooperating, dividing tasks, and getting work done. The paper is interesting, so take a look. I read it by translating the English into Korean. You can read it that way too.
Do We Feel a Ten-Year-Old Shock Only When Our Turn Arrives?
Today we looked at two stories about physics and AI-agent cooperation, along with MBC’s report on young Koreans losing their jobs. How do you feel? I wrote a book about AI three years ago and have followed the subject for quite a long time. Even I am honestly startled by recent developments.
Yet when I read about this physicist’s sleepless nights, I also wonder whether the shock came too late. Why was he not equally shocked when translators and accountants faced change, or when junior lawyers and programmers were no longer being hired as before? To me, this was a foreseeable future. What frustrates me is that people do not fully accept it until it is about to happen to them.
Of course, realizing directly that “AI is already as good as me—or better” must be a tremendous shock. I understand that. But the point is that none of us is very different from this physicist.
When Lee Sedol faced DeepMind in 2016, he won one game but lost the match. That must have been a great shock, mustn’t it? This physicist is now experiencing, in his own field, the shock Lee Sedol experienced ten years ago. It only becomes fully real when one’s own turn arrives.
None of us is really different from this theoretical physicist. We feel the shock only when it happens to us.
This person had even used AI himself six months earlier. He had already seen it solve a problem it initially could not solve once given some hints. Shouldn’t he have been able to infer that, after a little more time, it might solve the problem without hints, and later become far better than a human? Perhaps I am demanding too much. Still, I find it regrettable.
Do Not Judge Today’s AI from One or Two Old Attempts
Many parents and members of the public have not even used AI, much less followed the news about it. Some have never tried a free version. Free and paid versions differ, and within paid services, more expensive professional tiers differ again. Access is divided into levels, so briefly using one version does not tell you everything AI can now do.
The world is being turned upside down, but many still regard AI as a stupid thing that cannot answer properly, saying one thing when asked this way and another when asked that way. They judge today’s AI by one or two attempts they made in the past, without knowing what it can currently do.
Trying AI once or twice in the past does not tell you how far its capabilities have developed today.
That is what I prepared for today. You are welcome to ask questions, raise anything you are curious about, or share your reactions to what we have seen. I would like to know how you felt. I waited a little, but no further comments have come in. You may all be busy, or a delay in the broadcast may be making it difficult to respond immediately.
We will finish here. I hope you begin the new week with energy and enjoyment, and that today too you become a Buddha. Have a pleasant evening.
This article was edited by AI solely from the supplied video captions and description. No external sources were used.