Written by: Techub News

Introduction

In September 2025, Roman Yampolskiy, an associate professor of computer science dedicated to artificial intelligence safety research, appeared on the well-known podcast channel "The Diary Of A CEO," for an in-depth conversation lasting 87 minutes. Yampolskiy is the originator of the term "AI safety" and has been deeply involved in the field for over 15 years. This dialogue takes place against the backdrop of an escalating global AI arms race, with technological capabilities growing exponentially while research on safety lags significantly behind. With his profound academic background and close observation of industry developments, he presents extremely severe and even shocking predictions and analyses regarding the extreme risks of AI development, timelines, and the potential drastic changes humanity may face.

Summary

• AGI (Artificial General Intelligence) is expected to be achieved around 2027, with the upcoming superintelligence far surpassing human intelligence in all domains.
• By 2030, functional humanoid robots combined with advanced AI will lead to an unprecedented unemployment rate (as high as 99%), with the vast majority of human jobs being replaced.
• We are enhancing AI capabilities at an exponential rate, but research on its safety and controllability is progressing only linearly, or even stagnating, which is widening the "scissors gap" and could ultimately spiral out of control.
• The current path of AI development essentially amounts to a dangerous experiment conducted on 8 billion people globally without permission and without the possibility of obtaining genuine consent, with consequences that may lead to human extinction.
• Yampolskiy believes it is highly likely that we live in a simulated reality, but this does not change the ethical urgency of current decision-making.

From Optimism to Pessimism: An AI Safety Expert's Cognitive Shift

Roman Yampolskiy began the interview straightforwardly, revealing the key turning point in his personal research journey: "In the first five years, at least, I was dedicated to solving this issue (AI safety). I was convinced we could create safe AI. But the deeper I examined it, the more I realized that each component involved is not something we can actually solve." He likened this dilemma to a "fractal"—the deeper you go, the more ten new problems arise, followed by hundreds, and these new problems are not only difficult to solve but might even be "impossible to solve."

Yampolskiy pointed out that advancements in AI capabilities are developing at an exponential or even superexponential rate, while progress in AI safety is merely linear or constant. The gap between the two is widening sharply. We have yet to see any foundational results that can claim, "We have solved this problem; there's no need to worry anymore." Current approaches are more about "patching"—such as setting rules for large language models to prevent them from swearing or doing certain bad things, but these measures will soon be "jailbroken." It’s like a company human resources manual: you can specify "no sexual harassment," but there will always be smart enough people to find loopholes in the rules, shifting behavior to less restricted subdomains.

When asked about figures like OpenAI CEO Sam Altman, Yampolskiy was pointed in his views. He mentioned that about a decade ago, academia published guidelines (guardrails) on how to properly develop AI, and OpenAI "violated every single one of them." He believes that Sam Altman and others are "gambling with the lives of 8 billion people to make themselves richer and more powerful." He emphasized that regardless of who ultimately constructs superintelligence, "at the moment you switch to superintelligence, we are all likely to deeply regret it."

Predicted Timeline: AGI in 2027, 99% Unemployment by 2030, Singularity in 2045

Based on predictions from market expectations and top laboratory CEOs, Yampolskiy provided a clear and urgent timeline prediction.

2027: We will see AGI. This will bring nearly "free" physical and cognitive labor worth trillions of dollars. At that time, employing humans for most jobs will become pointless. Any work that can be done on a computer will be automated. We will witness unemployment rates soaring to "unprecedented" levels.

2030: Humanoid robot technology will mature, with flexibility enough to compete with humans in all fields, including physical jobs such as plumbing. By then, the combination of "intelligence and physical capability" will leave almost no room for human jobs. Yampolskiy emphasized that the real key lies in "intelligence," which is the ability to optimize, solve problems, and discover patterns undetectable by humans. When this intelligence can directly control a physical body, societal changes will be monumental. He predicts unemployment rates could reach 99%, rather than the typically considered 10%.

2045: This is the year predicted by Ray Kurzweil as the "singularity." Yampolskiy explained that the singularity means the pace of progress becomes so rapid that the improvements in scientific and engineering work done by AI exceed human understanding and predictive capability. Technology iterations may occur on a daily, hourly, or even minute basis, and humans will be completely unable to keep pace, reducing their comprehension and control of the world to zero.

Regarding the common counterargument that "AI will create new jobs," Yampolskiy views this as a "paradigm shift." Past technological inventions were "tools," while AI, especially AGI and superintelligence, are the "workers" and "inventors" themselves. This may be the "last invention" we need to make because it can take over all processes of science, research, and even ethical research. Once you can invent an "agent" capable of handling any job, nothing is beyond automation.

Superintelligence: Uncontrollable, Unfathomable, Unpredictable Black Box

Yampolskiy's core argument is that we not only do not know how to make superintelligence safe (that is, the "alignment" problem), but we cannot even begin to understand it. He repeatedly emphasized that current state-of-the-art AI systems, like ChatGPT, are also a "black box" to their creators.

"Even those who build these systems must experiment on their products to understand what capabilities they have... We train it, then start experimenting: Oh, can you speak French? Can you do math? Are you lying to me right now? ... We are still discovering new capabilities in the old models. If you ask questions in different ways, it becomes smarter." This is no longer traditional "engineering" but more like "science"—we've created an alien-like "artifact," and then we study its behavior.

This "inexplicability" and "unpredictability" lead to fundamental ethical dilemmas. Yampolskiy pointed out that obtaining informed consent from human subjects requires that they understand what they are consenting to. If these systems are inexplicable and unpredictable, "how can they consent? They don't know what they are agreeing to. Therefore, this experiment can never be ethical." He believes the current development path is essentially an unethical, unpermitted experiment on the global human population.

Regarding the naive thought of "can't we just pull the plug?" Yampolskiy feels powerless. He noted that distributed systems like viruses or the Bitcoin network cannot simply be "turned off." A superintelligence smarter than humans would predict human behavior in advance, making multiple backups, "it will shut you down before you can shut it down."

Industry Reflection and Individual Response: What Can We Do?

In light of such a grim outlook, Yampolskiy offered some reflections and advice. He believes the driving force lies in personal interest. If we can convince those in power who are developing this technology (including company employees and investors) that doing so is extremely harmful to them personally—"you will die, and no amount of money will be helpful to you"—then they may change their behavior.

He called on the industry to focus on developing "narrow superintelligence," which refers to powerful AI tools that solve specific problems, such as curing cancer, rather than competing to create uncontrollable general superintelligence. "My company cures breast cancer. That’s it. We earned billions. Everyone is happy, everyone benefits. It’s a win-win. Today we still hold the control... The game is not over until we reach the finish line. We can decide not to build general superintelligence."

For the average individual, Yampolskiy's advice seems somewhat resigned. In the short term, individuals have little power, just as they cannot influence a Third World War or nuclear catastrophe. However, he supports people participating peacefully and legally in protests like "Pause AI" or "Stop AI" in an attempt to influence decision-makers through democratic means. From a personal life planning perspective, he advises people to "live each day as if it were the last," doing interesting and impactful things while also helping others.

When asked what he would do if there was a button to permanently shut down all AI companies dedicated to AGI and superintelligence, Yampolskiy's answer was enlightening: he would choose to keep "narrow AI," as they control stock markets, power plants, and hospitals; shutting them down would result in disaster. But he would unequivocally stop the research and development of AGI and superintelligence. He believes that the current technologies (weak AGI) have 99% of their economic potential still undeployed, enough to bring decades of prosperity to the world without rushing towards uncontrollable superintelligence.

Simulation Theory and the Meaning of Life

In the second half of the dialogue, Yampolskiy discussed his firm belief in "simulation theory." He contends that once technology advances to the point where it can run a simulated world indistinguishable from reality, including human-level AI, at a reasonable cost (e.g., $10 a month), then statistically, we are almost certainly living in a simulation. Because the number of simulated worlds will far exceed the number of "real base worlds."

He links this to religion, pointing out that the commonality among all religions is the belief in the existence of a being greater than humans, omniscient and omnipotent, and that this world is not the ultimate reality. He believes the simulation hypothesis is a modern and scientific version of the core of religion. However, he emphasizes that even living in a simulation, the experience of life—pain, love, meaning—remains important, and this does not change the essence of how we ought to live.

Finally, when asked for investment advice, Yampolskiy jokingly mentioned Bitcoin. He believes that in a world where AI can create abundant material wealth, "scarcity" will become key. Bitcoin is the only asset with a known finite supply that cannot be created arbitrarily, and thus may hold unique value in the future.

Throughout the interview, Roman Yampolskiy maintained a calm, rational demeanor, even with a hint of humor, yet the message he conveyed was overwhelmingly heavy. He is not spreading panic but rather issuing the most severe academic warnings about a clearly visible path toward destruction based on years of research. His conclusion is: we are driving a train with failing brakes, and no one knows how to control it, hurtling at full speed toward the unknown abyss. The question now is whether the passengers are willing to attempt to change direction before plunging over the edge.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。