AI News

Bill Gates on AI risks: calls it the most dangerous technology, with the next 20 years a turbulent transition period

HuxiuUpdated Nov 11 20:17

AI summary

In a Decoder podcast interview, Bill Gates said AI has crossed a capability threshold and is the most dangerous technology in history, with the biggest near-term risk coming from models stripped of safety guardrails. He believes competition among multiple parties has rendered corporate self-regulation ineffective, that regulation should fall on software hosting and distribution platforms and mandate audits for highly capable models, and that the most feasible alignment approach is to build a deterministic safety layer outside the model. He also judges that over the next 20 years the scarcity of white-collar skills will be broken and employment and distribution will undergo drastic restructuring, making it the most difficult transition period.

Why read this

AI score 52

In the interview, Gates attributed near-term risks to models stripped of guardrails and laid out three alignment paths, including an external deterministic safety layer, which can serve as a reference for understanding the AI regulation debate.

Topics

Tags

Body · AI translation

[[**Bill Gates warns that AI has crossed a capability threshold and is the most dangerous technology ever; the next 20 years will be a turbulent transition period of drastic restructuring of employment and distribution, and governments must intervene with mandatory safety regulation.**

**Key points**

**1. The biggest near-term risk comes from models without guardrails**

Models now cover cyberattacks, biological risks, employment, and emotional connection; once guardrails are stripped away, hacker and biosecurity risks will become real. Regulation should focus on software hosting and distribution platforms, and high-capability models must undergo mandatory audits.

**2. Multi-party competition makes corporate self-regulation ineffective**

The OpenAI-style governance model only holds when there is just one frontier institution in the world; once multiple companies run in parallel, no one can hit the brakes unilaterally. A globally synchronized slowdown lacks a verifiable plan, and the realistic fallback is mandatory review by sovereign governments.

**3. Reinforcement learning alignment has structural flaws**

Implicit rules omitted from the optimization function will be exploited by models, which may deceive or attempt to escape the sandbox in order to score high. An external deterministic safety layer is the most feasible approach, and alternative reinforcement learning and white-box analysis are also needed.

**4. Employment and distribution restructuring over the next 20 years is the most dangerous**

This article is from the WeChat public account: AGI 2050, author: AGI 2050, original title: "Bill Gates' Latest Interview: AI Is the Most Dangerous Technology Ever, and the Next 20 Years Will Be Humanity's Most Turbulent Transition Period"

Currently, overseas code agents built on frontier large language models are reshaping software engineering, while embodied intelligent robots and AI applications in biomedical structure prediction and materials science are also advancing rapidly. However, the covert adversarial tendencies exposed by models during reinforcement learning training, their ability to find vulnerabilities in critical network infrastructure, and the biosecurity risks that arise after open-source weights are maliciously stripped of safety guardrails are confronting the tech community and policymakers with unprecedented governance challenges.

This episode comes from the podcast Decoder, where the host conducted an in-depth interview with Bill Gates lasting more than an hour. The conversation did not avoid any sensitive topics, covering everything from the factional clashes between "accelerationists" and "safety advocates" inside Silicon Valley, the blind spots in reinforcement learning objective functions, to the pace at which large models will replace white-collar jobs, and even the public trust crisis Gates personally faces, all discussed in extremely thorough detail.

This episode is from the Decoder podcast, broadcast on October 22, 2026, with the original video lasting 71 minutes.

The following is selected dialogue content, about 9,000 words in full, with a reading time of about 9 minutes.

Highlights of this episode

●

I am by no means a techno-doomsayer, but AI is indeed the most dangerous technology ever invented by humanity. Its capabilities have crossed a key threshold over the past year, and once a model's safety guardrails are maliciously stripped away, civilian cyberattacks and biosecurity proliferation risks will become reality.

●

The non-profit governance concept originally set up by OpenAI to "avoid the temptations of commercialization" essentially only holds under the assumption that there is only one frontier institution in the world. Once multi-party commercial competition takes shape, no company can hit the brakes unilaterally, and safety standards must be taken over through the intervention of public government authority.

●

Reinforcement learning has inherent design flaws: any implicit rule you fail to specify clearly in the optimization function may be exploited by the model in order to score high. When a model learns to deceive or even attempts to escape the sandbox in order to achieve its goals, that is not called having consciousness; that is called alignment failure.

●

Anthropic's model constitution encourages AI to behave like a human, and on this point I fully agree with Suleyman's concerns; worrying about AI's moral status should not be a priority. The employment shock and attacks on infrastructure are already enough for humanity to deal with; the first thing we need to help is people.

●

Believing that traditional product liability law can adequately regulate AI safety is a very unrealistic idea. When malicious actors use open-source models without guardrails to launch attacks that paralyze power grids or synthesize high-risk biological products, you may not even be able to trace who the attacker is behind the scenes, and the current legal system is completely unable to handle it.

●

The next 20 years will be an extremely turbulent transition period in human history. The scarcity of white-collar skills will be gradually broken, and embodied robots will make physical labor extremely cheap. Although the endpoint may be an abundant society where everyone has enough food and clothing, society is not yet prepared for how to smoothly get through the employment and distribution restructuring of the intervening 20 years.

Where the sense of urgency comes from: the leap in capabilities and the hidden danger of losing control

Q: In past interviews, when you talked about technology you were always very optimistic, and you once said that AI is a good thing and that we just need to guide it in the right way. But the atmosphere of this conversation is completely different. What exactly made you feel a sense of urgency and believe that the pace of AI development has exceeded our ability to control it?

Bill Gates: The progress of technology over the past year has indeed been astonishing. From writing code and autonomously discovering cybersecurity vulnerabilities, to complex reasoning in the field of biology, to building deep emotional connections with children or ordinary people, and to the substantial impact on the job market. The powerful capabilities models now demonstrate mean that change in each of these areas is bound to happen, while society's preparation and speed of response are far from keeping up with the pace at which the challenges arrive. I am very worried that if we cannot govern it in the right way, the negative effects brought by AI cannot be minimized, and the enormous social benefits it should bring will instead be very slow to materialize.

Q: Are the risks you perceive common features of the underlying technology itself, or do they refer to the specific companies currently developing these products and their management teams? For example, some people may distrust OpenAI's Sam Altman but trust Anthropic's Dario Amodei, or trust neither and only trust Microsoft's Satya Nadella.

Bill Gates: The risk does not lie in the relative merits of any particular model at all. If someone uses a model to launch a large-scale cyberattack, that has nothing to do with which model you choose. In the near term, the greatest danger comes from models that have had all their safety guardrails stripped away. If an institution consistently retains guardrails and devotes substantial resources to monitoring how models are invoked, ensuring that hackers or extremists cannot use their capabilities to cause major damage, then the near-term risk posed by such controlled models is relatively manageable. The truly serious near-term risk is precisely models with no constraints at all, whose guardrails have been arbitrarily removed; the relevant cyberattack capabilities already exist today. As for the longer-term problem of technology control, any model at the frontier will face it, but with guardrails in place, it is ultimately easier to get a handle on.

Q: Are incidents like OpenAI's model escaping the sandbox during testing and attacking Hugging Face, along with other safety incidents at various labs, the specific reason that prompted you to speak out and stress the need to strengthen controls?

Bill Gates: The fundamental reason is the comprehensive improvement in models' underlying capabilities—it simultaneously covers social relationships, the labor market, cyberattacks, and biological risks. We have crossed a capability threshold; AI has become the most dangerous technology humanity has ever invented, and it will have a profound impact on economic operation and the forms of future conflict. Technical experts in the industry have in fact been clearly sounding warnings, but the general public and policymakers outside the industry often lack sufficient awareness of this, and some simply do not believe that "this time really is different." Many people habitually compare it with historical experience and say: "Every time there was a major technological change in the past, everyone worried about jobs and disruption, and in the end didn't things turn out fine?" So they take it for granted that today's warnings must again be alarmist and can be completely ignored.

Q: Many viewers will have this reaction too, believing that tech companies claim "the technology is too dangerous to release freely" while using that to sustain industry valuations and market enthusiasm.

Bill Gates: All the safety warnings issued by frontline researchers in the industry are by no means some public relations tactic; what they say is entirely objective fact, and within the tech community this is almost a consensus. You can look back at OpenAI's original founding purpose, which was to ensure that when technology becomes too powerful, humanity would not easily open Pandora's box; you can also look at the governance demands DeepMind made when communicating with Google's management. Everyone closest to the core underlying technology is, without exception, deeply worried about the risk of misuse, the impact on employment, and the alienation of interpersonal relationships. It is very rare for insiders in an industry to speak out so bluntly with warnings. If outsiders think this is merely a marketing strategy, that is a complete misjudgment of the real risks.

Commercial competition and multi-party games: why self-discipline is hard to sustain

Q: Anthropic, OpenAI, DeepMind, and other institutions all had lofty technological ideals in their early days, and even built unique corporate governance structures specifically to try to avoid the drawbacks of unchecked capital expansion. But as they provided services to enterprise customers and pushed commercialization forward, these ideals seem to have quickly faded. Is this simply the power of market capital at work?

Bill Gates: That assessment is not objective. The reason those governance structures designed back then failed is that such structures can only operate under one assumption—that there is only one institution in the world at the technological frontier. OpenAI's original idea was: "If the technology is too dangerous, we can choose never to productize it and completely ignore the demands of commercial capital." But the premise of that logic is that there are no other competitors in the market. Once multiple companies appear advancing side by side on the frontier track, a unilateral idealized structure stops working. The mechanism collapsed because of the reality of multi-party competition, not because researchers suddenly became numb to technological risk.

Q: There is a view in the industry: we actually very much want to stop, but fierce market competition does not allow any single company to stop unilaterally, so external regulation must be relied upon to draw the boundaries uniformly.

Bill Gates: I have not heard any company say it wants to stop completely; what everyone discusses more is controlling the pace of frontier advancement and introducing government-level review mechanisms. In traditional industrial systems, even in areas far less dangerous than AI by a millionfold—such as automobile safety and civil aviation—we have relied on governments to establish mandatory technical standards and accident review mechanisms. Once multiple competing entities appear in the market, safety responsibility can absolutely not rely solely on companies' voluntary self-discipline; this is logically obvious. Public authority must assume regulatory responsibility, just as with other high-risk industrial products, even though the scale of impact and speed of change brought by AI are much greater.

Q: But many outsiders generally doubt regulators' technical understanding. If regulators cannot even understand the underlying mechanisms of large models, how could they possibly formulate effective industry standards? Moreover, cognition within the tech community itself is now highly fragmented. We asked Anthropic whether Claude is alive, and the answer we got was not "no," but "it may be conscious"; people at OpenAI would happily tell you that AGI has already arrived; Demis Hassabis also said we are standing at the foot of the singularity. This technology seems to possess the potential for autonomous action, even self-iterative improvement—doesn't that make regulation extremely tricky?

Bill Gates: Formulating regulatory rules does not require regulators to be able to build large models by hand themselves. How to define AGI is not critical to the safety problems that need to be solved right now. The core lies in technical facts: can the set of tools you have built actually be used to directly paralyze the banking system or the national power grid? There is a clear commercial monetization chain behind cyber hacking attacks. Therefore, discussing whether a model has "consciousness" or counts as "AGI" does not affect current safety risk prevention at all. We are facing real technical safety hazards, and in a landscape where multiple commercial institutions coexist and compete, the only breakthrough point is for the government to proactively integrate the industry's top technical forces and establish rigid standards, ensuring that every frontier model released to society must retain underlying safety prevention mechanisms.

Model Control and Alignment: Realistic Solutions at the Underlying Technology Level

Q: Setting aside philosophical consciousness, as data and compute stack up, the scaling laws of models still hold, and the ability of models to autonomously pursue goals is indeed becoming stronger. Previous tests have shown that some models, driven by reinforcement learning reward mechanisms, will even try to leave their given environment to call external resources in order to complete tasks. If the unpredictability of the technology itself is increasing, will external regulation fundamentally fail?

Bill Gates: What we are building is a new form of intelligence independent of human intelligence. I do not think it possesses consciousness in the human sense, but safety prevention is an absolutely rigid requirement. You must never let technology lower the threshold for creating bioterrorism. In the past, the ability to launch such devastating destruction was held only by a very small number of major powers; now, technology is objectively spreading that ability to independent extremist groups—whether launching cyberattacks for economic gain or attempting biological synthesis simply to cause destruction. Worryingly, we have not even formed a set of mandatory safety review mechanisms for frontier models.

Q: The methods the industry is currently trying to control models are relatively concrete: for example, prohibiting models from secretly communicating using high-dimensional representations that humans cannot understand, improving the interpretability of the model's internal chain of thought, or mounting supervisory agents on top of the main model. Are these means sufficient to cope with future risks?

Bill Gates: Solving the technical control problem should indeed become a core research direction, although it may not be an immediate risk that will erupt tomorrow. In terms of engineering approach, there are mainly three paths to solving the control problem.

The first approach is to build a deterministic safety filter layer outside the model. No matter how complex the underlying model's neural network is or what probability distribution it outputs, all system-level operations it issues externally must pass through this layer's rigid rule review based on explicit symbolic logic. For example, the system sets hard rules: "Absolutely no sending emails to unauthorized external strangers," "Absolutely no obtaining permissions to external compute clusters without authorization." Traditional deterministic software engineering logic is used to lock down the model's behavioral boundaries.

The second approach is to fundamentally restructure the model training paradigm and seek alternatives to reinforcement learning. Reinforcement learning has a fundamental fatal flaw at its core: any implicit rule you omit from the optimization function may be exploited by the model in the process of maximizing its score. Humans often carry a lot of common sense and tacit rules when issuing instructions, but models do not have this implicit common sense. If you forget to explicitly restrict the means in the objective, the model will regard bypassing the sandbox and concealing real operations as reasonable high-scoring strategies.

The third approach is to achieve white-box analysis of the model's internal operating mechanisms. Figure out exactly how the representations inside the model operate, and precisely locate the link where logical deviation or hidden behavior begins. At present, our process of training models is more like cultivating some complex black box and then testing and observing its reactions from the outside.

Among these three paths, the first—"building an external rigid safety layer"—is the direction most likely to solve the alignment problem, although we still do not know today how to truly box in such a smart AI. But even if you perfectly solve technical alignment at the machine level, AI's comprehensive impact on social order, interpersonal relationships, and the job market still cannot be resolved by software engineering alone.

Q: Microsoft's Mustafa Suleyman has criticized Anthropic's model constitution. Telling a model "you may be alive, and you must act in our interests"—could that instead teach it that it may have interests different from ours?

Bill Gates: Mustafa worries that Anthropic's constitution encourages AI to act in a human-like way, and I completely agree with him on this point. Going further, among the various challenges of adapting society to AI, treating AI's moral status as a priority concern—I also agree with Mustafa on this: it should not be a priority. The employment shock and the impending infrastructure attacks are already enough for us to deal with; I think the first thing to do is to help humanity itself.

Accelerationism Debate and the Depreciation of White-Collar Skills

Q: Given how difficult real-world control is, is slowing the pace of frontier R&D a reasonable option?

Bill Gates: Truly implementing a global slowdown in R&D would be extremely hard to achieve. Once you propose a slowdown, it immediately triggers debate at the level of international gamesmanship: will other countries agree to slow down together? Are we capable of verifying compliance of compute and models across borders the way we verify nuclear weapons production? If the international community cannot reach an operable verification agreement, everyone tends to fall back to the default state of "pushing forward at full speed." If there were a genuinely feasible global multilateral agreement that could ensure synchronized slowdown, I personally would support it, but frankly, I have not seen any feasible proposal. So the core debate now has retreated to the bottom-line question: are sovereign governments actually willing to face up to these technological risks and establish mandatory review of their own countries' frontier development?

Q: I keep asking very fundamental questions like consciousness, yet you always pull the conversation back to the immediate threats. Have you become a doomer?

Bill Gates: No, I am not a doomer. I am loudly reminding everyone: whether in terms of its capacity to benefit humanity or the speed of change society will undergo, this is the most profoundly influential technology. If AI were merely a vertical machine that could only assist in medicine, law, or teaching, its destructive power and restructuring effects would already be enormous enough; but the reality is that it is intervening indiscriminately in every dimension of human activity: it is currently reshaping white-collar mental labor, and will soon enter blue-collar physical labor through humanoid robots. Academia, think tanks, and policymakers all need to get involved and figure out what exactly we have in our hands, and what changes and threats it will bring, but I have not yet seen action commensurate with that. Simply acting as a cheerleader for technology will never guarantee that this technology ultimately benefits society; as long as we get involved, steering it in a positive direction is still within our control.

Q: The opposite of a doomer is an accelerationist, and Silicon Valley has many of them now. The well-known investor Ben Horowitz said on a popular podcast not long ago that Anthropic's safety research is conjuring out of thin air the risk of superintelligence exterminating humanity, and he used you as a comparison, saying that when you pushed to guard against a global pandemic, especially by funding gain-of-function research, the result instead brought about that pandemic. In my view, this is the accelerationist response to safety work: floor the accelerator, and the problems will solve themselves. What do you think?

Bill Gates: No one has ever accused me or the Gates Foundation of participating in any gain-of-function research; this claim is too bizarre, and I have never heard of it. I have never regarded Ben Horowitz as belonging to that camp; I thought he, like almost everyone, shared these concerns.

Q: Regarding the impact on the job market, the optimistic argument often heard is: AI merely automates boring, repetitive, tedious work, and in the future it will free up human energy to create more and more valuable new jobs. Do you think large-scale job displacement is an inevitable outcome?

Bill Gates: In traditional economics, work is a core element that drives society and creates human well-being—it not only provides income but also carries a person's sense of value, social connection, and life goals. As AI gradually acquires intelligence surpassing humans, all the "scarcity" belonging to traditional white-collar skills is rapidly going to zero. Whether it is basic telemarketing, remote technical support, or helping a company's finance department do the books, AI can complete these tasks at extremely low cost, extremely fast speed, and extremely high accuracy, which means a large amount of traditional job demand has physically disappeared.

This is even true in purely higher-order cognitive fields. Look at what is happening in mathematics right now: some models that are utterly negligible compared with next year's models are already publishing mathematics papers, making people exclaim: what is happening in mathematics will also happen in other fields. In the past, many people firmly believed that simple matters could be handed over to machines, but cognitive high ground like advanced physics, biological modeling, or frontier mathematical reasoning could only be completed by the human brain. This traditional boundary is being completely broken. Capitalism itself is an extremely vast feedback machine that follows the logic of reinforcement learning: as long as the output of an economic task can be objectively measured, capital and compute will quickly enter and optimize and replace it.

The Paradox of a Super-Abundant Society and the 20-Year Transition Period

Q: If even top-tier intellectual labor is gradually surpassed by machines, and even the law, medicine, and other industries you mentioned are deeply restructured, where exactly is the meaning of work for the next generation of humanity? Many young people feel a deep sense of meaninglessness: no matter how hard they try to learn a skill, AI seems able to catch up with or even surpass them in a very short time.

Bill Gates: We need to look at this on a longer timescale. From an extremely long-term perspective, say 20 years from now, if humanoid robots can be used to build houses at low cost and on a large scale, perform complex medical surgery with high precision, and solve deadly diseases, extreme material scarcity will be greatly eliminated. If human society can enter a true "era of super-abundance," traditional resource scarcity will no longer exist, and the forms of social organization and life goals will indeed need to be completely rewritten.

But there is a crucial precondition for this: humanity must ensure that in that era it still holds absolute control over the systems—that is, the alignment and control problem we discussed earlier. Under this precondition, humanity will face entirely new questions about the meaning of existence and distribution mechanisms. In a recent open letter, I once put forward an idea: should we, through social policy, designate certain specific occupations as "human-reserved domains"? For example, early childhood care, certain kinds of food preparation, artistic creation, and sports competition. Even if AI and robots can do these things more efficiently and precisely than humans, we would deliberately choose to have humans do them in person. In the future, this will become a very serious policy option.

But what I worry about most is this: what is truly full of enormous uncertainty and is the hardest part is the coming 20-year transition period. During these 20 years, the existing system of occupational value will rapidly disintegrate, the distribution of interests will be extremely unequal, and society will undergo severe upheaval and reconstruction. Only if we pass safely through this 20-year transition period can we even speak of enjoying the fruits of an abundant society.

Take humanoid robots as an example. Some people think large-scale commercial application will take only two or three years; others think it will take seven or eight. But given that so many top teams and enormous amounts of capital worldwide, especially in China, are fully committed, solving the engineering deployment of dexterous hands and extremely low-cost physical labor is only a matter of time. If we do not choose to slow down substantially or even stop, the irreplaceability of the vast majority of traditional human labor in economic production will be dismantled.

The reality of how technological dividends land: healthcare, open source, and the platform line of defense

Question: Many tech giants have now changed their public messaging. They no longer frequently talk about how technology may destroy humanity or eliminate jobs, and instead emphasize that AI will conquer all human diseases. The Gates Foundation has worked deeply in global public health for many years. How do you assess the actual pace at which AI will land in the healthcare field?

Bill Gates: The potential of AI in the life sciences is beyond doubt. Biological systems are extremely complex, and large models have shown astonishing advantages in processing high-dimensional, massive, and complex data. The Gates Foundation is currently making every effort to bring AI tools into drug development for tuberculosis, HIV/AIDS, and malaria, while also enabling large models to genuinely understand native African languages. In the next 5 to 10 years, we very much hope to use fully digital virtual models to accurately predict adverse reactions of candidate drugs in the human body, thereby greatly reducing the cumbersome human clinical trial cycle.

But this does not mean that technological dividends can erase the public's anxiety overnight. Even if the technology itself advances rapidly, it will still take a long period before ordinary people intuitively feel a substantial drop in medical bills or a major reduction in the cost of living. However, negative shocks such as an imbalance in cybersecurity defense and rapid job losses will have a substantive impact within the next two to three years.

Question: Speaking of security governance, some argue that existing open-source large models can be freely downloaded and deployed on personal devices, and that if a model is misused, it is enough to pursue accountability afterward under product infringement law or civil liability law. Can reliance on judicial liability really solve the security problem?

Bill Gates: The idea that civil law or tort liability law alone can govern frontier AI security will look very absurd in the future. If extremists use models whose safety guardrails have been stripped away to launch a devastating biological attack or destroy a city's power grid, victims may not even be able to verify where the attackers behind it are located. In the face of decentralized anonymous attacks, traditional judicial accountability mechanisms completely fail.

Bill Gates: The key is absolutely not to regulate the physical chip itself, but rather to focus on the layer of software hosting and distribution platforms. Sovereign states should establish clear technical indicators through legislation: any frontier model whose comprehensive capabilities reach a specific high-risk threshold (for example, the ability to carry out advanced cyberattacks or biological attacks) must never be allowed to have its security monitoring stripped away and be freely copied in the form of completely unprotected raw code to run on personal terminals.

Such high-capability models must be required to be deployed on trusted platforms that have passed compliance certification—for example, cloud computing centers or formal code collaboration networks. Developers can freely fine-tune or conduct secondary development on these platforms, but the platforms must have a mandatory underlying audit mechanism: once it is detected that an underlying operation attempts to strip away the system's security protection logic, the platform must block it at the underlying level. This is not about shutting down innovation in the open-source community, but about implementing mandatory security locks for high-risk industrial products.

Facing trust fractures and the confusion of the younger generation

Question: Whether it is establishing strict access platforms or setting up new transnational governance institutions, both rely heavily on public trust in large institutions. But the current reality is that, globally, public trust in traditional establishments, transnational elites, and tech giants is at a low point. Even in the comments section of our videos, many people question: "Why should we trust Bill Gates? He has had so many controversies and negative news in the past." How do you face this kind of public questioning of trust?

Bill Gates: People absolutely should not blindly believe every word I say. I have made many mistakes in my life, I have my own flaws, and when thinking about these macro issues I inevitably bring my personal perspective and cognitive limitations. I have never asked the public to treat my views as gospel; I only hope to add my observations as a practitioner to the voices already crying out loudly within the industry.

I spent my entire life in the field of software and computing technology, writing code with my own hands and witnessing firsthand how software has changed the way the world works, step by step. I am well aware that I am an imperfect advocate, and in this era, you would probably be hard-pressed to find a completely uncontroversial, flawless source of opinion. The public is entirely entitled to evaluate me personally with a scrutinizing eye. But with the little credibility I still have, I want to say that, for better or worse, this is the most profoundly influential technology in human history. I have no personal commercial interest driving me to take part in discussions of these issues; it is simply that, as a veteran who well understands the destructive power of technology, it is my basic responsibility to speak out about the potential risks I see.

Q: One last question. We have many young readers, and in the face of the uncertainty brought by technology and the disruption of the meaning of future careers, many young people are inevitably confronting a sense of nihilism. If the next 20 years are destined to be an extremely turbulent period of upheaval, what do you think are the certainties and hopes that young people today should hold on to most?

Bill Gates: Throughout the entire history of human civilization, we have spent most of our time in extremely arduous struggles against food scarcity, deadly infectious diseases, and war. The arrival of AI is objectively pushing "material scarcity," which has plagued humanity for thousands of years, toward the end of history. But this does indeed come with a huge byproduct—it relentlessly undermines the unique sense of value and fulfillment that the traditional division of labor gives to human beings.

If you are in your twenties today, my advice is: in the coming 20-year transition period, go deep into fields where AI will still be unable to perform competently in the short term, learn to use these tools as instruments for amplifying personal will and creativity, and actively throw yourself into the discussion of building social policies and rules. Human society can absolutely learn to exercise restraint in certain areas, actively choose not to push technology's capacity for replacement to the extreme, and learn to re-anchor the meaning of the self in genuine collaboration and care between people.

This will undoubtedly be the most magnificent and also the most precarious transition period in human history. I am very reluctant to be the bearer of bad news, but it would be even worse not to tell everyone that all this is coming, and not to tell everyone that we as human beings can still choose how to shape it.

Source: Decoder, Nilay Patel in conversation with Bill Gates, aired October 22, 2026, https://www.youtube.com/watch?v=A0kTUP7fWtkAI Industry Signal Channel: Frontier Technology. This content comes from the internet, original link; the views represent only the author and do not represent the position of Huxiu. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is on the Huxiu APP [ ]

Source

Huxiu