share_log

Weekend Reading | Renowned Investor Ray Kurzweil: AI Will Fully Surpass Human Capabilities by 2031—No Bubble

Tencent Technology ·  Oct 31 15:47

Source: Tencent Technology

Technological forecasting typically oscillates between two extremes: either vague grand narratives or short-sighted linear extrapolations.

But there is one person who seeks to approach the future in a different way.

Before becoming a futurist, Lars Tvede was better known as a seasoned European hedge fund manager, a serial entrepreneur, and the author of classic business works such as "The Psychology of Finance" and "The Inevitable Economic Cycle." Over the past three decades, he has been exploring the intersection of technological evolution and capital cycles, bridging Switzerland with global secondary markets. This year, he published "Super Trends," a book examining how AI will reshape our lives.

This cross-industry innovation stems from his own deep engagement with AI applications.

A few years ago, Tweed founded the trend‑research firm Supertrends. Every day, he uses Agent to sift through 5,000 tech publications and patent databases, mapping out the precise year each cutting‑edge technology emerged—much like charting the brain's "connectome."

In 2027, the total volume of tokens generated by AI will equal the combined output of all human communication; in 2028, artificial general intelligence (AGI) will be realized; and by 2031, superintelligence will comprehensively surpass even the most exceptional individuals.

Behind his radical judgments, when confronted with questions such as "Has the Transformer architecture already reached its limits?" or "The hallucinations of large models," Tweed maintains a technologically evolutionary form of optimism.

But what truly sets Tweed's thinking apart from that of Silicon Valley tech evangelists is his investor‑focused perspective on this transformative shift.

During a two-hour conversation, he laid out a very detailed economic analysis. Despite widespread concerns that data centers might repeat the mistakes of the "railway bubble," Tweed's industry‑chain research painted a different picture: model production costs are plummeting by 80% to 90% annually, while token prices have tripled in response to surging demand for computing power.

Although Frontier Models remains unprofitable, the profits generated by its upstream chip and infrastructure segments are sufficient to cover these losses three to six times over, with some computing‑power clusters even achieving payback within a year.

He labeled the public's skepticism about the slow pace of AI deployment as the "current‑state fallacy," noting that people tend to judge future infrastructure—delivered in two years—by today's capabilities.

When this exponentially growing computational power ultimately drives lighting, software, healthcare, law, and even most physical goods into a state of "super-abundance" and extreme affordability, Twidell argues that the newly created value will inevitably flow to those "things that AI cannot produce in greater quantities."

Scarce industrial metals, irreplaceable prime locations, mechanical watches, live concerts featuring real musicians, and "being human" itself are poised to usher in a long-term bull market for assets.

Even if machines take over 95% of today's jobs by 2050, capital will still flow to sectors that cannot be made more efficient.

Compared with the economic calculus, Tweed's prognosis for social structures and the human mind is far more stark.

He explicitly opposes using "Q&M Dental" to directly hand out money to the unemployed, arguing that it would strip people of their innate drive to earn respect through hard work, ultimately leading to a civilization akin to that depicted in "WALL-E," where people lie in chairs, losing all motivation to rise early and strive.

Just as, after bidding farewell to manual labor, humanity invented the gym, so too, in the "passion economy" that has cast off the anxieties of survival, we must once again seek out challenges, forge communities, and rediscover the reasons to rise early and strive.

The following is a conversation between Boiling Under and Lars Tweed:

01. How to Create a Timeline of AI Evolution

Q: In your book, you cover virtually every aspect of future life, and you even assign very specific dates to the development of various technologies and their societal impacts. How did you arrive at this timeline?

Russ Tweed: These forecasts come from our company, Supertrends. We've developed an app that scans roughly five thousand vetted scientific and technological publications every day, while also monitoring key thought leaders on platforms like X to gather their predictions, which we then aggregate into a consensus forecast.

Of course, asserting that something will happen in 2049 is, by its very nature, a speculation. Yet a well‑grounded prediction is far more useful than vague talk of "some time in the future." Providing a specific year also serves another important purpose: many new technologies depend on earlier innovations that have already been developed. We are currently constructing a technological network map that illustrates how existing technologies give rise to subsequent advances—somewhat akin to the "connectome" in neuroscience.

This is much like the recent research that mapped the fruit fly brain. Scientists devoted decades to this endeavor, slicing the fruit fly's brain into ultra-thin sections and capturing countless images. Yet it wasn't until the advent of AI that researchers could decipher these images, reconstruct a three-dimensional model of the fruit fly brain, and then upload it into the metaverse.

That replica of the fruit fly's brain is flying in virtual space and can even respond to simulated odors, which can be described as a genuine "brain upload."

I mention this example because, in principle, if we could establish a comparable "connectome" for technological innovation worldwide, we could trace the evolution of technology forward and gain a clearer understanding of how future developments will interconnect. While this goal remains distant for Supertrends, we have already begun making progress toward it.

Question: Could you first walk us through the key changes that AI will drive at each stage of its development?

Russell Tweed: The simplest way to categorize this is into three stages: The first stage is the chatbot—responding only when prompted, and doing nothing unless activated. The second stage is the agent, which runs continuously in the background and takes autonomous actions; we're still very much in the early days of this phase. The third stage is embodied AI, starting with self-driving cars and eventually expanding to humanoid robots, autonomous drones, and other forms. Of course, this is a highly simplified framework.

We can also assess just how intelligent these models are. Since 2022, cutting‑edge generative AI has advanced in just a few years, with its cognitive capabilities rising from the level of preschoolers and elementary school students to that of university undergraduates and PhD holders—and now even surpassing some PhDs. In tests involving complex problems, human PhDs answer correctly on average 79% of the time, while state‑of‑the‑art models have already exceeded 90%. By standard IQ metrics, certain leading models have reached scores between 180 and 200. This demonstrates that human intelligence is far from the upper limit of what's possible.

As for when we can consider artificial general intelligence (AGI) to have been achieved, that depends in part on how we define it. My own reasonable estimate is 2028: by then, it will be capable of performing any task that a human can do. Beyond that, we may see the emergence of superintelligence.

(Superintelligence) will comprehensively surpass the best human minds in every domain, and this could occur as early as 2031.

What follows is what I call "collective superintelligence." By that stage, in addition to powerful centralized models, there will also be millions or even billions of distributed models. You might ask: from the perspective of physical laws, the more spatially compact an intelligent system is, the higher its operational efficiency—so why bother with decentralization? Because distributed models, scattered across various locations, can learn directly in the real world.

When we converse with ChatGPT, Claude, or DeepSeek, it is essentially retrieving what it has already learned or temporarily searching the web for information—in a sense, merely conducting "desk research." But if it were simultaneously connected to a million robots, each of which is working in the real world every day, that would be akin to continuously carrying out "field research." Combining desk research with field research would make all the difference.

Following this, a new scientific revolution is likely to emerge: science driving science, with AI creating AI. We have already witnessed some shifts in biochemistry—such as using AI to investigate how proteins fold, how different proteins interact, and which crystals can form. Signs that AI is helping to improve itself are already evident, though for now they still rely on human input. In a few years, it may well take over the iterative process entirely.

I believe the technological singularity has already begun; in its early stages, technological capabilities far outpace humanity's actual rate of adoption, and people have yet to realize just how far it has progressed.

02. There is no bubble in AI—only a shortage.

Q: Today's chatbots have indeed reached, and in many domains even surpassed, the level of human PhDs. This year, numerous companies have begun experimenting with AI agents, yet the actual results have fallen short of expectations.

Russell Tweed: Yes, this is precisely where technological capabilities outpace real-world applications. This spring, Anthropic released a career‑radar chart showing that the large blue areas represent jobs that are already technologically viable to replace, while the much smaller red areas indicate tasks that have already been effectively automated. In the future, both of these zones will expand. I believe that by 2050, the blue zone could encompass as much as 95% of today's occupational landscape, while the red zone—representing actual job displacement—will lag behind.

Indeed, studies have shown that most companies' AI projects have yet to yield significant results. I recall a Boston-based study estimating an 95% failure rate, though some IT professionals argue that such a high attrition rate is actually quite typical for conventional IT initiatives. Moreover, negative news tends to attract more media attention than positive outcomes.

Meanwhile, the AI programming field is undergoing dramatic changes. Goldman Sachs once compared three areas: data entry, technical support desks, and software programming. According to that study, the cost of AI programming can be as low as 5% of that of human programmers, whereas in the technical support desk sector, the cost of using AI is almost the same as hiring humans—AI is only slightly more cost-effective.

This is highly counterintuitive: service desk staff may be ready to work after just a few days of training, whereas programmers often hold master's or doctoral degrees.

The reason lies in the specific characteristics of each task. The service desk requires high‑speed, real‑time speech recognition and sometimes even the ability to detect customer emotions; by contrast, writing code consumes far fewer tokens. That's why AI is advancing step by step, one domain at a time. Some jobs will be disrupted quickly, while others are not yet ready—so we shouldn't assume that all industries are moving at the same pace.

Today, numerous indicators in the industry suggest that demand for AI continues to outstrip supply—and that demand is growing even faster. From what I'm observing, token prices have begun to rise this year, nearly tripling from their previous levels, while the cost of producing computing power is still declining at an annual rate of 80% to 90%. This divergence between rising prices and falling costs has created substantial profit margins. Industry analysts also note that the market urgently needs more capacity; at present, virtually all output is being sold as soon as it's produced.

The consumer market is different again. According to the data I've seen, ordinary consumers account for roughly 70% of users, yet they contribute only about 5% of paid revenue. However, just because consumers don't pay doesn't mean they don't value it. We could reframe the question: "If you were unable to use any AI next month, how much compensation would you expect?" In surveys, respondents' average requested compensation was 30% higher than the average actual spending of paying users. Since many people initially used the service for free, their compensation demands often exceed their actual spending—but this still underscores how deeply consumers value AI in their minds.

By the way, here's a particularly intriguing comparison: AI is spreading at a pace several times faster than that of mobile phones or the internet in their early days. My research shows that, throughout human history, only one product has achieved such rapid global adoption—COVID‑19 vaccines. Both reached 100 million users in roughly the same amount of time after launch. Keep in mind that vaccination directly affects life and death, determining whether people can go out to work; yet GPT's initial uptake has managed to keep pace with vaccines, which is truly thought‑provoking.

A collaborator and I are currently completing a 200-page research report on the economics of AI, much of which was left out of this book because it is too specialized. After delving deeper into the subject, I have become even more convinced that the current AI industry is constrained by supply. Of course, there are dissatisfied users and failed experiments in the market. Yet in my day-to-day work, I have personally been involved in three highly successful projects.

The first project is Supertrends. It automatically scans 5,000 publications daily, queries patent databases, processes information in 40 languages, generates 100,000 pieces of innovative content each year, and even produces its own reports and visualizations. All of this is accomplished without any manual intervention—just a single Mac computer keeps it running.

The second project, meanwhile, requires close collaboration with a human team. In Denmark, a group is seeking to launch an AI‑powered integrated media company to compete with the country's three major traditional media conglomerates. They brought on a seasoned executive who previously served as editor-in-chief of a leading media outlet, and teamed up with another experienced media professional to draft a business plan, which they then discussed with our Supertrends team to map out how to bring it to life.

Their initial plan called for hiring forty people. To put that in perspective, a traditional media outlet of comparable size would typically require four hundred employees—so even forty seemed remarkably lean at the time. A few months later, the team reported that the deeper they dug, the more they realized AI could accomplish far beyond what they'd initially envisioned, and during the research phase, the AI models themselves grew significantly more powerful. As a result, they scaled down the headcount to twenty-two. The system has now been in pilot operation for three months and is poised for full launch, with the team further reduced to fourteen members—staffing is now complete. As for whether it can perform as well as a major traditional media organization? I'd say it matches or exceeds their performance in most areas, and in some respects even outperforms them, thanks to its ability to generate richer content and deliver highly personalized services.

The third project was deployed in our hedge‑fund business. Every day, we receive roughly a hundred PDF reports from investment banks; just reading one of them would require several cups of coffee, and it's simply impossible to go through them all manually. Yet, for macro investing, it's essential to understand how the market's most astute minds are thinking.

To this end, we are developing an intelligent system designed to identify the consensus of views among top analysts. Once the consensus points to a new domain, the system automatically spawns a dedicated instance—a sub‑agent—to seek out the most knowledgeable experts in that field; if further sub‑domains emerge, it launches additional instances to delve even deeper. For example, starting from "industries that provide foundational enablers for AI," the system will autonomously drill down into the semiconductor sector, then into the high‑bandwidth memory (HBM) space, tracking the forecasts and evolving perspectives of experts at each hierarchical level.

At present, debugging this model has been quite costly. However, once the system is up and running, it will continue to learn autonomously, without requiring repeated manual instructions. Should a new technology like CRISPR gene editing suddenly emerge, the system will proactively detect it, assess whether anyone is discussing related investment opportunities, and then identify experts in that field to keep track of developments.

These three firsthand experiences have convinced me that we're still far from fully grasping what AI is capable of. Every time I think I've come to understand it, by the following week I find it has already gone even further.

There are many ways to measure AI progress, but each metric is ultimately a simplification. In the past, Moore's Law predicted a doubling of computing power every two years; by contrast, according to certain composite metrics, AI capabilities can improve by a factor of 130 every two years, as they integrate numerous technologies that have already been growing exponentially.

Some have likened the current wave of data-center construction to the railway boom of yesteryear, raising concerns about overinvestment. Yet data centers have a unique characteristic: once the hardware is in place, efficiency can still be dramatically boosted through advances in software algorithms. Research I've reviewed suggests that software‑driven improvements alone could yield up to a five‑hundredfold increase in performance. Of course, this doesn't mean chips will never need to be replaced; the energy efficiency of next‑generation chips continues to improve, and older models will eventually be phased out as their power consumption becomes relatively too high. For now, however, with supply under severe strain, existing chips can remain in use for a very long time.

I believe many skeptics fall into the "current‑state fallacy": they judge whether today's investments are worthwhile solely based on their perception of AI's present capabilities. Yet from the decision to invest to the completion and operation of a data center, it typically takes 18 to 24 months—by which time AI's capabilities will have already advanced by several more orders of magnitude. Humans often struggle to intuitively grasp such exponential trends.

I'm also examining where the money really is in this wave. My understanding of the Chinese market remains limited; I've focused primarily on the U.S. and Western markets. Based on current growth rates, I estimate that the global AI industry could rival the traditional automotive sector as early as next year and surpass it by 2028. While leading large‑model companies are still posting losses, the upstream chip industry may be generating profits three to six times greater than the total losses of these model firms. The range is so broad because it's difficult to pinpoint, from a financial standpoint, exactly how much of a chip company's earnings can be attributed to AI.

Looking at the payback period along the industry chain, I estimate that building an AI chip fab typically takes three to five years to break even, while a high-bandwidth memory production line may take two to four years; data center infrastructure with a power‑supply contract usually recovers its investment in about six years; and once compute‑intensive servers are deployed and services are offered, the payback could be as short as three years. Musk even claims he can achieve payback in just one year—though, of course, that's Musk; he's always been fond of making bold statements like that.

Take his Colossus 2 supercomputing cluster, for example: the exact cost remains unknown—perhaps between ten and twenty billion dollars, or even more. What I've heard is that, once it completes its own training workloads, it can lease out its computing power, generating annual rental revenues of up to fifty billion dollars, while still retaining a portion of its capacity for internal use. So when he claims it will recoup its investment within a year, I find that entirely credible. The reason these operations can generate such substantial profits boils down to the current supply bottleneck.

However, large-scale investment will ultimately eliminate supply bottlenecks. Once these bottlenecks are removed, corporate profits will have to stem from genuine competitive moats and sustainable business models.

I believe its key evolution lies in delving deeper into the enterprise‑level application layer, focusing on the tasks people actually need to accomplish in their day-to-day work—while the system also remembers what you've done before. For example, I keep all the project‑related materials in a single folder; after using it for a while, it almost needs no further explanation to know exactly what I'm looking for. At this point, I have no desire to switch, because it understands me well enough and has already handled much of the groundwork for me. That's what we call exceptionally strong user stickiness.

In short, from a business investment perspective, if the maximum score is ten, my confidence in AI would be a perfect ten.

03. The current limitations of AI technology could also become its future strengths.

Q: You're very optimistic, but some AI researchers argue that the current technological trajectory based on the Transformer architecture may still lack crucial foundational innovations and might not directly lead to superintelligence. Moreover, there are physical constraints: scaling up GPU computing power requires building new data centers, which takes time. At the same time, today's AI systems also lack robust world models, long-term memory, and the ability to learn continuously—whether these challenges can be overcome remains uncertain.

Moreover, you argue that AI will recursively improve itself, but some research suggests that current AI systems are not adept at tackling open-ended problems; their scope for exploring potential solutions is rather limited, and they cannot think in the same way as human scientists. These factors all seem to cast doubt on the notion that AI could continue to grow so rapidly. What do you think?

Russell T. Tweed: Research that points out inherent flaws in AI often becomes outdated very quickly. Over the past three years, we've witnessed incredible technological advances; moreover, some characteristics once dismissed as "weaknesses" may, upon closer examination, turn out to be strengths.

Take "illusions" as an example: humans, too, can experience them. Our visual field naturally contains blind spots, yet the brain automatically fills in what it expects to be there, compensating for the visual gap. When playing soccer and someone passes the ball your way, you don't first crunch the physics equations in your head; instead, you instantly retrieve similar scenes you've encountered before and, on instinct, anticipate where the ball will land.

I believe that "predicting what will happen next" is a core hallmark of intelligence, and AI does this every day. This kind of anticipation and imagination can also fuel creativity: even if we have never personally experienced a particular situation, we can still envision what might unfold. The human mind is inherently generative, and so is AI. Of course, we must not confuse imagination with reality; humans need to learn to distinguish between the two, and AI is likewise learning to make that distinction.

In practical applications, some tasks demand absolute precision—for instance, arithmetic should be left to a calculator. But if I'm crafting a tagline for a startup, creativity is what's called for.

Precisely because there are so many possible paths of technological progress, I find it hard to believe that it would suddenly run into some insurmountable natural limit.

Some people still argue that AI merely parrots content authored by humans on the internet, and therefore cannot be smarter than humans. In reality, however, AI is generating entirely new knowledge through self-play and system dynamics models. Before long, much of the information it relies on may no longer come from humans but will instead be discovered by the AI itself. Of course, this will require ongoing innovation, but I'm already witnessing rapid progress in this area.

Q: It is indeed accelerating, but not across all fields. Software programming is rather unique: existing training techniques enable rapid progress in this area, whereas other domains do not see the same level of advancement. Once the "low-hanging fruit" has been harvested, what lies ahead may well be far more challenging work.

Russ Tweed: Exactly—just as animals initially fed on low‑lying leaves and, once those were exhausted, had to evolve into giraffes to reach the higher foliage, the difference is that AI's "neck" is growing exponentially. What seems out of reach today, I believe, will be entirely attainable in the future.

Information technology and genetic science were once thought to face insurmountable obstacles, but each was eventually overcome. In the past, humans solved these challenges; in the future, with the powerful assistance of AI—and perhaps even through AI itself—these problems will be addressed.

You might as well compare today's models with those from just one year ago. With the advent of reasoning‑based models, we now take for granted receiving reports that are clearly structured, backed by verified data, and cross‑checked against multiple sources—something that was entirely unimaginable a year ago.

Last week, I dictated a research assignment directly to GPT, making several revisions along the way, and then handed the compiled results to Claude: tasking it with identifying fifty leading venture capital firms and thirty startup accelerators worldwide, determining what they invest in each quarter, and analyzing—by technology category—which sectors are on the rise and which are in decline.

As a result, it took only eighteen minutes to produce a report of exceptionally high quality, and it even proactively added some data visualizations that I hadn't requested but found extremely useful. Suppose Goldman Sachs had just hired a top graduate from MIT or the University of Chicago's business school; without AI, I'd estimate that the same task would take at least a full week.

Moreover, the duration of continuous tasks that AI can handle on its own is rapidly increasing. I believe that in just a few years, it will be able to operate autonomously and continuously without interruption.

We can imagine a scenario: Elon Musk sends a batch of robots with IQs as high as 300 to Mars, equips them with nuclear reactors, solar panels, and basic tools, and tells them, "I'll be back in thirty years—build a civilization on Mars." When humanity returns three decades later, the buildings and roads will already be in place. Faced with any challenge, these robots will figure out solutions on their own, working continuously for thirty years. Though this is still a thought experiment, I believe we're not far from achieving such capabilities.

Following this line of reasoning, we can extend the software‑level concept of "natural language programming" to the physical world. In the future, if I say I want a house in a particular style and specify the rooms I need, the AI will generate design proposals until I'm satisfied. Then I'll simply say, "Go ahead and build it—I'm off on vacation," and the AI will independently locate a site, acquire the land, secure building permits, and coordinate robotic teams to carry out construction.

04. The post-AI era will be an age of abundance, but it should not be an era of UBI.

Q: The technological changes you describe will bring about an "ultra-abundance" of intelligence, but this does not necessarily translate directly into material abundance. For instance, the "time cost" of obtaining food has not been reduced nearly as dramatically as that of lighting—our efficiency in raising livestock and cultivating crops may have improved by only a few times, at most tenfold, compared with the past. Different sectors and types of goods are affected in distinct ways. How much longer will it take for all domains to truly achieve such ultra-abundance?

Russell T. Tweed: The pace at which different goods achieve superabundance indeed varies. Without complementary conditions, intelligence alone cannot produce many physical goods. Consequently, superabundance will first emerge in intelligence itself, followed by an abundance of services, and then gradually extend to an abundance of other tangible items. Industries such as programming, accounting, finance, and media may not follow a strictly fixed sequence or proceed at the same rate, but overall we are moving toward broader abundance—a process that has, in fact, been unfolding since the Stone Age.

Two centuries ago, almost everything was scarce. A useful way to gauge abundance is what you call the "price of time": how long must one work to obtain a given quantity of goods? The most striking example is lighting. In the past, people had to gather firewood and light fires, and once the fuel was gone, they had to go out again; today, some buildings in Shanghai themselves resemble giant lamps. Over the span from the Stone Age to the present, the price of time for illumination has improved by a factor of half a million. If we trace this progress back to humanity's mastery of fire, it may have taken as long as 1.5 million years—or even longer. By contrast, I estimate that achieving a 500,000‑fold reduction in the cost of AI intelligence will require only six years.

I also anticipate that by 2027, the number of tokens generated by AI will equal the total number of words spoken by all humans; thereafter, it will surpass the total number of words thought in human minds, and eventually exceed the combined cognitive capacity of all humanity. Once it reaches a parity with human cognition, it may be five times greater than that of all humans within a year, twenty-five times greater in two years, and soon leave us far behind.

As for the food issue you mentioned, it offers valuable insights for investors. In fact, I believe that today humanity produces more food than is actually needed: globally, there are more people who are overweight than underweight, and more obese individuals than those who go hungry. The root of the problem lies in poor distribution, not insufficient production. After reaching a peak, global agricultural land has remained largely unchanged—and has even declined slightly—while the world's population has doubled and per capita food consumption has risen, demonstrating that we can indeed continue to improve productivity.

However, it is crucial to recognize that "AI cannot increase the supply of anything."

First is "human beings themselves"—even when robots can handle certain tasks, we still prefer to rely on real people. Second are industrial metals and other minerals; at present, it's not clear how AI or robotics could significantly accelerate their production. On the contrary, if AI fuels economic growth and makes people wealthier, society's overall demand for physical materials is likely to rise. Third are prime, scarce locations—such as coastal areas, lakeshores, ski resorts, and highly desirable urban centers. For all these assets whose supply cannot be substantially expanded by AI, we can expect to see pronounced price inflation in the future.

Almost all economists, myself included, believe that AI will fundamentally usher in "deflation": many goods will become cheaper or even free, keeping the overall consumer price index (CPI) flat or pushing it lower. Yet central banks typically aim to maintain inflation around 2%; when inflation falls below this target, they respond by injecting substantial liquidity into the economy through measures such as low interest rates and quantitative easing (QE).

As a result, technological progress is driving down consumer‑goods prices, while monetary expansion is working hard to counteract that effect. The excess money thus ends up flowing into scarce assets outside the scope of the Consumer Price Index—such as stocks, prime real estate, works of art, and collectibles. Consequently, in the AI era, some of the biggest winners may well be precisely those assets whose supply AI cannot increase—for example, paintings by renowned artists. I believe the cultural and artistic sectors are poised for a bull market.

In my book, I divide the world of the future into three economic zones: the first is "tech super‑clusters" like Shenzhen, Shanghai, and Silicon Valley; the second is "slow‑living regions," such as parts of Europe, where people enjoy a high quality of life and no longer feel compelled to hustle endlessly; and the third comprises areas that have neither actively participated in nor benefited from technological change—though I believe this third category is steadily shrinking. At the bottom tier, machines will give rise to a vast "super‑intelligent economy," while at the top they will carve out space for humans, fostering what I call the "Passion Economy": a realm where people can pursue the things they truly care about. I don't mean to overstate it—no one will necessarily achieve their ultimate dream or become a Picasso—but I do think the focus of work will shift markedly toward what people genuinely want to do.

So, could AI pose a danger? Absolutely—almost every beneficial technology has its dark side, just as the internet can harbor viruses and nuclear energy can be used to build atomic bombs. For me, one systemic risk posed by AI is that it might render humanity utterly passive, leading us to stop taking action and cease striving altogether.

Humans, in fact, need to strive. While excessive struggle and hardship can certainly overwhelm us, a complete absence of them can also be destructive. This is much like physical activity: in the past, people routinely moved their bodies and consumed foods that had not undergone so much modern processing; but with the advent of advanced technology and an abundance of processed foods, we now have to make a deliberate effort to exercise just to stay healthy. Similarly, once AI generates vast wealth and people no longer need to work, human society could slip into the kind of state depicted in the animated film "WALL‑E"—where people spend their days reclining in hovering chairs, doing nothing at all.

That is precisely my deepest, most terrifying vision of humanity's future. If a primitive hunter‑gatherer from the Stone Age were to see today's gyms, he would surely ask: "Why do you waste your energy hauling around these iron weights? Isn't gathering firewood, foraging for berries, and hunting wild boar exhausting enough? It's already good enough to rest once you've finished your work." Yet today, we must rely on the gym just to stay healthy. By the same token, as we look ahead, we must seriously ask ourselves: how can we reclaim the "effort and striving" that have become unnecessary in our everyday lives?

If humanity were to sink into such a state of utter complacency, it would be a tragedy for individuals and a catastrophe for culture as a whole. Many of the world's most successful cultures have revered hard work, nurtured imagination and perseverance, and shown people what they can achieve through effort—far beyond material wealth. The ancient Roman Empire and the dynasties of ancient China rarely interacted, yet each developed highly sophisticated civilizations; I believe that a cultural ethos that values diligence and ambition was a key factor in this success. Once we strip away the motivation that drives people to rise each morning and move forward, over the long term we risk fundamentally undermining the very foundations of our culture. How can we reap the benefits of AI without depriving people of the reasons to strive and live purposefully? This is a central issue that society as a whole must address.

For this very reason, I believe that "universal basic income" (UBI) poses a significant social risk. Some people suggest that in the future, we could tax machines and distribute the proceeds equally to everyone, allowing them to simply go to the riverside to paint. Leaving aside the fact that both machines and capital can move across national borders, even collecting the money would be extremely difficult; what worries me even more is that such an arrangement—unconditionally handing out cash regardless of whether one works or not—could ultimately spiral into a social catastrophe.

In fact, a more prudent middle ground is entirely feasible—for example, by providing "Universal Basic Services": free healthcare, free education, and, once autonomous vehicles become widespread, a certain degree of free public transportation. Deciding to make certain essential services free while still preserving people's incentives to act is fundamentally different from the psychological risks associated with handing out unconditional cash.

In addition, there is another approach, similar to the "Trump Account" (a child‑start investment account) recently discussed in the United States. As I understand it, each child would receive an initial investment principal at birth, with additional contributions from charities and family and friends, and social funds could also be directed toward children in low-income areas. In this way, everyone becomes a "stakeholder" in technological development, directly holding a share of the ownership rights to certain machines.

I'm not worried about this Q&M Dental–style employee stock ownership arrangement.

Q: But if everyone receives a share of wealth from birth, wouldn't that still be wealth they can claim without doing anything? In the end, if people live off dividends, might we not end up in the kind of "lying flat" scenario depicted in *WALL-E*? Isn't your real concern that, once external pressures to survive are gone, humans will completely lose their drive to act and to push themselves beyond their limits?

Russell Tweed: That's a very good question. Simply holding shares does indeed mean you don't have to work, but the crucial distinction is that you establish a direct "stake" in societal progress. You start paying attention to the company's stock price—up when it rises, down when it falls. As people engage with technology and machines, their mindset shifts from the adversarial "they're against us" to a sense of shared destiny: "we're in this together."

As for the loss of the desire to take action you mentioned, that is precisely at the heart of the issue. For along with it goes one's sense of self‑respect. Beyond material security, there are two things that can bring profound happiness: first, love—something we can discuss in more detail later; and second, the experience of striving to accomplish something that seems difficult and extraordinary in your own eyes. It could be in sports, business, or social engagement—something that pushes you right to the edge of your capabilities. You decide to give it a try, and in the end, you actually succeed. At that moment, you may be utterly exhausted, yet inwardly brimming with joy, feeling that life has taken on deep meaning.

This drive to challenge oneself is something that should never be taken away.

05. If we want to prevent people from becoming obsolete, education needs a major overhaul.

Q: When I previously interviewed the philosopher Nick Bostrom, he also touched on a similar issue. He suggested that in the future, we might need to establish specialized institutions to help people engage in deliberately designed, challenging games or tasks. For once daily life and survival‑related activities have been taken over by AI and natural challenges have vanished, humanity will have to create such challenges artificially.

Russell T. Weed: Indeed. We are heading toward a world where machines outperform us in the standardized skills taught in schools. That's why I believe future education should place greater emphasis on cultivating psychological resilience and a growth mindset.

I'm a big fan of Carol Dweck's book, "Mindset," which thoroughly explains the difference between a growth mindset and a fixed mindset. I also appreciate Stephen Karpman's "Drama Triangle." In real life, some people, no matter what happens, habitually pigeonhole themselves and others into the roles of "victim," "rescuer," and "perpetrator," blaming external circumstances while refusing to take responsibility for their own lives and wasting their energy on things they can't change.

Therefore, education in the future should increasingly help students become "more self‑efficacious individuals"—or, to put it more naturally, help each person "become their better selves."

When humans collaborate with AI, the AI functions much like middleware within software. The human starts at the beginning by setting a goal—say, building a house—and the AI, positioned in the middle layer, takes on the task of executing the details. Then, the human steps back to the end, evaluates the outcome, determines whether it meets their expectations and whether adjustments are needed, or even considers whether they might as well build an entire village instead.

In other words, humans always occupy the "two ends" of the workflow, while AI sits in the "middle."

Since humans begin at the starting point, they must excel at "asking questions." Yet what do traditional schools primarily teach today? Teachers pose questions, and students provide answers—what schools spend over a decade cultivating in students is, in fact, nothing more than the middle‑level task of "answering questions"! Of course, foundational knowledge remains essential; our minds must be equipped with a solid base of information. But by comparison, future education must place greater emphasis on developing the ability to "ask good questions." And once we receive the results delivered by AI, we too must learn to assess their quality—because, at its core, that, too, is a form of questioning.

Whether in formal education or lifelong learning, we must adapt to this trend. Training the human brain is somewhat like training large-scale models: once early‑formed habits become ingrained, they are extremely difficult to change. If, during one's school years, students spend all their time practicing memorization and test‑taking while never honing their ability to ask questions, it will be very hard to reverse that mindset once they enter the workforce.

In this regard, I believe China has already gone quite far in its approach to AI. As far as I know, some Chinese educational programs teach children from an early age how to use AI and also train older learners to collaborate with it. By contrast, in Denmark, there are still discussions about how to keep AI out of education—because when teachers assign homework, students simply ask GPT for the answers and hand them in; teachers can spot this immediately and find it utterly pointless. I understand teachers' concerns, but students will inevitably have to work alongside AI once they enter the workforce. What schools should really be teaching is how to partner with AI to get things done effectively.

06

Interpersonal Relationships in the Age of AI

Q: Let's return to the topic of "love and interpersonal relationships" we just discussed. I think most people instinctively aspire to a life that's easy and free of setbacks. AI companions are exceedingly polite, devoid of their own negative emotions, and always respond to you with gentle care; in contrast, real human beings have their own concerns and feelings, can be difficult to communicate with at times, and may even leave you, fall in love with someone else, or offer you critical feedback. How do you think AI will impact genuine human-to-human relationships?

Russell T. Tweed: Even before AI, social media had already profoundly reshaped human relationships—and I'd argue, mostly for the worse. With fewer face-to-face interactions and more digital ones, humanity seems to have lost something precious. Today, we see more people feeling lonely and fewer children being born, and these trends appear closely linked to digital media. I've read studies showing that when impoverished African communities gained access to satellite television, birth rates declined; similarly, the decline in Western fertility rates coincides strikingly with the widespread adoption of smartphones. So I fully share your concerns: we should assume that AI companions are likely to produce comparable effects.

The frictionless AI companionship you just described is much like the "ultra-processed foods" of the food industry: one bite satisfies you instantly, with sugar rushing into your bloodstream to deliver a quick thrill—yet long-term consumption leads to nutritional deficiencies. Today, all over the world, you can see couples or friends sitting at the same table, each staring down at their own phone. The digital medium offers such instant gratification that we end up missing out on the most precious, genuine connections between people.

AI makes non‑face-to-face virtual communication far more appealing, but it does come with significant risks.

But there's another side to the coin. Personally, AI is an excellent tool for self‑education: I consult it daily about things I don't know and engage in deep conversations with it, which has made me more knowledgeable. At a broader level, the future development of the "passion economy" and cultural activities will also give people more leisure and composure to connect with others.

In the future, people's time will increase in two ways simultaneously: first, they will no longer have to take jobs they dislike just to earn a living; second, biomedical breakthroughs may extend "healthy life expectancy" considerably.

Traditional life is too short to allow one to focus on just one pursuit; otherwise, nothing will get done. But if one has as many as a hundred healthy years in which to work effectively, each dream can be allotted twenty years for deep cultivation, making it far more likely that a person will feel they are in the right place, doing what they truly want to do.

As people gain more time to pursue their interests, the internet makes it easier than ever to connect with like-minded individuals—whether you're into cats, dogs, stamp collecting, birdwatching, or Python programming, you can find a community online. Meanwhile, in the real world, though in-person socializing tends to be more costly and time‑consuming, the "club economy" is booming. I've studied club economies around the globe, and they're growing at roughly twice the pace of the global economy as a whole.

In London, when a businessman is a member of a private club, he often invites his friends to dine there. I once met a London publisher whose Writers and Publishers Club was filled with colleagues—everyone shared a love of books, so they naturally had endless things to talk about.

I live in Verbier, Switzerland, a popular ski resort. When I first moved here, the town didn't have any clubs, so I helped found one that uses a member‑referral system to ensure members are well‑matched. Today, the town boasts several such clubs, and some friends are even planning to start one exclusively for entrepreneurs. As physical labor declined over time, humanity invented modern sports and gyms; in the age of AI, we will surely develop new ways to compensate, bringing real‑world social life back to a healthier balance—offline clubs are one approach, as is the "flipped classroom" model in education: students first learn with AI, then meet offline to collaborate with their peers.

Q: Although London has a wealth of clubs, it's also known for its elite culture and can even come across as somewhat arrogant in certain respects. I can imagine that Switzerland and Denmark boast numerous egalitarian interest‑based communities, but why has London, too, been able to cultivate the kind of vibrant, healthy community you describe?

Russ Tweed: In fact, clubs aren't necessarily snobbish, nor do they necessarily have to be elitist. Even when there are "elite" standards within a club, the criteria used by different organizations vary widely. For example, a biker gang like the Hells Angels and Monaco's elite yacht club have entirely different requirements for earning the respect of their peers. Someone who is respected in one setting may not enjoy the same standing in another—and may not even want to go there.

Once, at a gas station in Switzerland, I came across a large gathering of vintage Morris Minors. Their owners had even formed a dedicated club, united solely by their shared love for this classic little car. This shows that clubs can be built around virtually any specific object or interest. As the philosopher Robert Nozick argued in Anarchy, State, and Utopia, people differ in their psychology and preferences: "One person's paradise may be another's hell." Send someone from a yacht club to a motorcycle club, and they're bound to be unhappy—and the reverse is true as well. In The Righteous Mind, Jonathan Haidt examines humanity's fundamental moral intuitions and likewise finds that political divides are rooted in deeper psychological differences.

So I believe that a good society should allow everyone to find like-minded people and pursue their dreams together, while also having the right to set boundaries and keep troublemakers out of their community—if someone barges into our club to start a fight, then it ceases to be our utopia. In a diverse community, there doesn't have to be just one standard for earning respect: at the Morris Minor Club, as long as you can personally restore a vintage car to pristine condition, you'll command the respect of every fellow enthusiast in the room.

Adam Smith argued in The Theory of Moral Sentiments that humans are naturally inclined to seek the respect of others. If a person merely inherits a large fortune and squanders it, others may be drawn to him for his wealth, but they will not truly respect him.

Society will only continue to progress when it respects those who truly deliver results and give their all. That's why I worry that simply handing out money earned by machines to everyone could gradually erode the motivation people have to earn respect through hard work.

07. Even if AI takes over production, humans remain the most expensive.

Q: But in another twenty-five years, the most important engine driving overall societal growth will inevitably be AI. By then, no matter how hard people work, the combined economic output of all humanity may amount to less than 1% of what AI generates.

Today, assessing a person's worth often hinges on how much output they generate and how much profit they contribute. As material production ceases to be the paramount concern, the criteria by which society evaluates individuals—and indeed the very foundations of the economy—may well undergo fundamental transformation.

So, must humanity's standards of value and the nature of work fundamentally differ from what they are today?

Russell Tweed: This is an extremely important topic in macroeconomics. Professor Chad Jones of Stanford University has conducted dedicated research on the share of GDP attributable to a "general-purpose technology" once it emerges. For example, IT hardware—whose share was virtually zero after World War II—rose steadily to a historical peak during the internet boom (I recall it was around 2.8%), only to decline again to roughly 2% today. Yet today, global computer hardware is far more abundant than ever before—so why has its share of GDP fallen? Because as hardware becomes both better and more plentiful, it also grows progressively cheaper.

So, when I say that by 2050 AI will be able to perform 95% of today's jobs, I'm referring to the jobs as we define them today. Consider the steam engine in the early days of the Industrial Revolution: at first, it was primarily used to pump water out of coal mines—mining the coal to power the engine, then pumping the water back out. This somewhat resembles the criticism leveled today against "capital circulating endlessly within the AI industry chain." Back then, Watt used the unit of "horsepower" to compare a machine's output to that of a working horse. If someone had asked Watt how useful his machine really was, he might have modestly replied that it could replace 80% of the horses employed for pumping water in coal mines. But if someone like Elon Musk had been around, he would have insisted: "No, it can replace all the horsepower in society—90% or even 95% of what horses do today!" In the end, the total horsepower provided by machines in industrial civilization far surpassed the combined horsepower of all the real horses on Earth at the time.

Humans are very likely to face a similar situation in the future: the total amount of work performed by machines will far exceed that of all humanity, expanding the economy to a scale that is now difficult to imagine.

However, according to economic principles, in the long run, societal wealth and capital will ultimately flow to those who "provide a limited supply."

And we humans always find ways to keep certain human‑provided values limited and scarce.

Consider today's high-income earners: top musicians, professional athletes, Formula 1 drivers, soccer stars, and superstars like Taylor Swift. They compete fiercely with one another, yet because the pool of truly exceptional talent is finite, we remain willing to pay premium prices for "the very best human performance." As I understand Jones's research, humanity's share of the global economic pie is unlikely to shrink significantly in the future. From the steam engine to electricity, computers, and the internet, every technological revolution has sparked predictions that "humans will have nothing left to do"—yet each time, we've found new pursuits. I believe the same will hold true this time around.

Ultimately, the very fact of "being a genuine human being" cannot be automated by AI. Like prime real estate or hard-to-extract rare metals, people may still command a high price. Even as AI drives the cost of every standardized commodity down to near zero—or even makes it free—human beings' intrinsic value does not necessarily depreciate; on the contrary, the money we earn could buy far more than it can today.

For example, in "slow‑living zones," people can engage in activities related to culture, the arts, and lifestyle. Just as dining out in Switzerland is no longer a financial burden today, in the future hiring a lawyer or an accountant, or accessing high‑quality education, will also become remarkably affordable. With many of life's basic challenges eliminated, individuals will be free to pursue what they truly care about.

However, I also acknowledge that those who naturally prefer routine, repetitive tasks may find it increasingly difficult to adapt in the future, as such jobs are indeed becoming harder to come by.

Q: You say people can turn to cultural pursuits, but nowadays even music and written works can be generated by AI. If one day AI produces articles that are better than human‑written ones and music that sounds more beautiful, why would anyone still bother creating it themselves? Moreover, the mechanical watches and top‑tier performances you mentioned earlier are, at their core, luxury goods—expensive precisely because they rely on handcrafted techniques or live, human performance. Does this really reflect a universal trend?

Russ Twidell: I believe humanity still has ample reason to create and perform. Consider this: over the years, digital music has grown increasingly affordable, yet tickets to live concerts by real‑world artists have become ever more expensive; smartphones let you check the time for free, while traditional mechanical watches are selling at ever higher prices.

Money in society ultimately flows to those sectors where "efficiency cannot be improved through technology," because, all too often, what people truly crave are experiences that carry human warmth and authenticity. The live music I hear in a concert hall may not necessarily boast sound quality that surpasses a digital recording online, but the immersive, on‑the‑spot experience it offers is far richer—and entirely different. That's why, even when tickets are pricey, I'm still happy to pay.

Of course, the transformation of society as a whole will by no means be easy; it will be fraught with upheaval, and many people will have to change what they are doing. Among all these changes, I believe that education is both the most crucial and the most challenging.

Today, some tech entrepreneurs are founding new universities, seeking to revolutionize education. Yet the inertia of traditional education runs deep—manifesting not only in teaching practices but also in how people perceive academic credentials. Even I, upon hearing that someone is studying at a top-tier institution, instinctively assume they must be very smart—a trait that, in itself, is a positive thing.

But it also means that when it comes to reforming education, we can easily fall far behind the changes that the times truly demand.

Editor/rice

The translation is provided by third-party software.


The above content is for informational or educational purposes only and does not constitute any investment advice related to EleBank. Although we strive to ensure the truthfulness, accuracy, and originality of all such content, we cannot guarantee it.