AI has really arrived, can the economy withstand it? — A heated debate among "big short", "AI giants" and "top tech influencers"

Recently, three heavyweight figures from different fields sat down at the same table: the contrarian Michael Burry, who foresaw the 2008 financial crisis; Jack Clark, co-founder of Anthropic with cutting-edge models; and Dwarkesh Patel, the host who has interviewed all the big names in Silicon Valley.

They posed a straightforward and piercing question: As the AI revolution rolls in, can the global economy smoothly navigate this drastic shift? Is this wave of AI investment a necessary path to realizing the future, or a historical capital misallocation unfolding before our eyes?

Key Points Summary:

• The "Cognitive Correction" from 2017 to 2025: The true explosion of AI is large-scale pre-training, not training agents from scratch. The industry previously bet on "blank slate agents," believing that task-based environmental training would yield general intelligence, but ended up with AI that was "superhuman" only in specific tasks. What truly changed the world were the large-scale pre-training paths brought by Transformers and Scaling Laws. Today's return to agents is due to the massive pre-trained models behind them. The industry consensus has shifted to: what we currently see is the "floor" of capabilities, not the "ceiling."
• Chatbots trigger a trillion-dollar infrastructure race: Investment logic is completely abnormal. ChatGPT initially seemed like just a writing, searching, and homework tool, but unexpectedly ignited a global trillion-dollar investment in hardware and infrastructure. While application layer revenues have yet to materialize, capital expenditures have exploded, forcing traditional software companies to transform into capital-intensive hardware enterprises. This "first lay down infrastructure and then wait for demand" model is extremely rare in investment history.
• "Who is winning" cannot be simply answered: Competitive advantages in AI are not durable. The AI industry is not like a "winner-takes-all" platform economy, but more like a "highly competitive field": leaders are constantly changing, and companies like Google, OpenAI, and Anthropic struggle to maintain long-term advantages. Talent mobility, ecosystem diffusion, and reverse engineering quickly weaken barriers. The current landscape resembles "the top three taking turns," with leadership potentially reversing at any moment.
• Does AI enhance productivity? The key is not how powerful the tools are, but the lack of real, quantifiable metrics. Existing data is conflicting: METR research suggests that using AI coding tools decreases efficiency, while Anthropic's user survey reports a 50% efficiency increase. Both sides acknowledge that the lack of a refined "process dashboard" leads to unclear real ROI. A subjective sense of "ease" does not equate to actual efficiency gains, and the industry urgently needs a reliable system for quantifying productivity.
• Why is AI's capability astonishing yet has not led to large-scale replacement of white-collar jobs? Theoretically, current models far exceed expectations from 2017, but high error rates, weak self-correction, and complex responsibility chains make seamless integration into real workflows difficult. Only scenarios like software development, which are "naturally closed-loop," can be quickly utilized; More industries need to build verification and automation loops to unleash true productivity potential.
• Burry's core concern: it's not whether AI is useful, but whether the capital return can hold up. Burry focuses on financial structural risks such as ROIC, depreciation cycles, and stranded assets. Data centers and chips are updated rapidly, and many assets should not be treated according to traditional long-term depreciation logic; otherwise, profits may be overstated. If the growth rate of terminal AI revenue is far below infrastructure investment, there may be a huge backlog of "construction in progress," potentially triggering a private credit risk chain.
• The indicators that will "change their views" in 2026: turning opinions into verifiable bets. The three guests all proposed indicators they are willing to be "slapped in the face" by: whether AI application revenue can exceed $500 billion, whether frontier lab revenue can exceed $100 billion, whether chip lifespan can be extended, whether continuous learning can be conquered, and whether scaling faces bottlenecks. The health of the industry in the coming year will be revealed across four dimensions: "revenue, capability, efficiency, and capital return."
• Consensus: the ultimate bottleneck of the AI revolution is not algorithms, but energy. Regardless of how the technological route evolves, the demand for computing power continues to rise, and electricity supply becomes an absolute hard constraint. Small nuclear power, independent grids, and energy infrastructure will determine whether AI can enter the broad real economy. The real limitation is not the model, but "where does the electricity come from." The AI revolution may be written in the power grid, not in the code.
• How to judge whether the AI boom has gone off track: look at five key variables. The real value of this debate is that it provides "hard indicators" to determine whether AI is moving towards healthy development. Technology will continue to advance, but that does not mean the business loop has been established; productivity improvement requires reliable data, not self-perception; the risks of depreciation and term mismatch in the capital cycle will gradually become apparent; employment impact is not obvious, as real workflows are far more complex than demos; energy and infrastructure are the ultimate limits. To judge whether AI is off course, look at these five lines: capability, efficiency, capital return, industrial loop, and energy supply.

The following is the original interview text:

The Story of AI

Host Patrick McKenzie: You have been hired as historians for the past few years. Please briefly narrate what humanity has built since the publication of the paper "Attention Is All You Need." What situations in 2025 would surprise audiences from 2017? Which learned predictions have fallen short? Please tell this story from the perspective of your field (research, policy, or market).

Jack Clark: Back in 2017, most people were betting that the path to a truly general system lay in training agents from scratch through a progressively difficult task curriculum, thereby creating a general capability system. This was reflected in the research projects of all major labs at the time (such as DeepMind and OpenAI), which were trying to train players that surpassed humans in games like "StarCraft," "Dota 2," and AlphaGo I see it as a "blank slate" (tabula rasa) gamble—starting from a blank intelligent agent, honing it repeatedly in certain environments until it becomes smart.

Of course, as we now know, this did not truly lead to general intelligence—but it did produce superhuman-level agents within the trained task distribution.

It was at that time that people began to try another approach: conducting large-scale training on datasets and attempting to build models that could predict these distributions and generate from them. It turned out that this method was extremely effective and was accelerated by two core factors:

1. The Transformer architecture from "Attention Is All You Need," which made this large-scale pre-training incredibly efficient;
2. The roughly synchronous development of "Scaling Laws," or a core insight: you can simulate the capabilities of pre-trained models in relation to the foundational resources (data, computing power) you invest.

By combining the insights of Transformers and scaling laws, a few individuals accurately predicted that by massively scaling data and computing power, general systems could be obtained.

Now, interestingly, things have come full circle: people have started building agents again, but this time, they are endowed with all the insights from pre-trained models. DeepMind's SIMA 2 paper is an excellent example, as they created a general agent for exploring 3D environments, which is precisely based on the underlying pre-trained Gemini model. Another example is Claude Code, which is a coding agent whose underlying capabilities come from a massive pre-trained model.

Patrick: Given that large language models (LLMs) are programmable and ubiquitous (including open-source versions that are stronger than in 2017 but slightly weaker than the current top levels), we have now reached a stage where any future development regarding AI capabilities (or other interesting things) no longer needs to be built on a cognitive foundation that is worse than what we currently have. This perspective of "what you see today is the floor, not the ceiling" is, I believe, one of the things that industry insiders understand most thoroughly, while decision-makers and the outside world misunderstand the most.

Every future StarCraft AI will surely have read the original Chinese version of Sun Tzu's Art of War, unless its designers assess that this would weaken its ability to defend against Zerg rushes.

Jack: Yes, we at Anthropic often tell policymakers, "This is the worst time ever!" But it's really hard to make them understand the importance behind that statement. Another intuitively difficult concept to grasp is the speed of capability enhancement—currently, a typical example is that many people trying out Opus 4.5 in Claude Code will say things like, "Wow, this thing is so much stronger than before." If your impression of LLMs is still stuck in last November, then your judgment of cutting-edge technology is seriously off Michael Burry: In my opinion, in 2017, AI did not equal LLM. AI refers to Artificial General Intelligence (AGI). I believe that people at that time did not think of LLM as AI. I grew up reading science fiction, which predicted many things, but none imagined "AI" as some kind of search-intensive chatbot.

As for "Attention Is All You Need" and the Transformer model it introduced, those were developed by Google engineers using Tensor. By the mid-2010s, AI was not a foreign concept. Neural networks and machine learning startups were everywhere, and AI was frequently mentioned in conferences. Google already had large language models at that time, but they were for internal use. One of my biggest surprises was that Google did not manage to stay ahead throughout the process, considering its dominance in search and Android, both in chips and software.

Another surprising point was that I thought Application-Specific Integrated Circuits (ASICs) would be adopted earlier, and small language models (SLMs) would become popular sooner. It is truly astonishing that NVIDIA has maintained such a dominant position in inference for so long.

What surprised me the most was that ChatGPT actually triggered this investment frenzy. The use cases for ChatGPT were very limited from the start—search, student cheating, and programming. Now there are better programming LLMs. Yet, it was a chatbot that sparked trillions of dollars in spending.

Speaking of this spending, I think the most brilliant moment in Dwarkesh's interview with Satya Nadella was when the latter admitted that all major software companies are essentially hardware companies, capital-intensive. I'm not sure if the analysts tracking these companies really understand what "maintenance capital expenditure" is.

Dwarkesh Patel: Great perspective. It is surprising how fleeting the lead in the AI field has been so far. Of course, in 2017, Google was far ahead. A few years ago, OpenAI also seemed to be leading by a large margin. But there seems to be some force (perhaps talent poaching, rumors circulating, or reverse engineering) that neutralizes any overwhelming advantage a single lab might have. Instead, the three giants are rotating positions on the podium every few months. I'm curious whether "recursive superintelligence" can really change this situation, or if we should get used to this long-term fierce competition.

Jack: Regarding recursion, all leading labs are leveraging AI tools to accelerate their developers' progress, but this is not an overnight success. It seems to have a "barrel effect"—for example, if you can now produce 10 times the code, but your code review tools have only improved by 2 times, then you won't see a significant overall speedup. A major unresolved question is: is it possible to completely close this loop? If so, you might see some kind of compounding R&D advantage

Can AI tools really improve productivity?

Dwarkesh: The core question at the "million-dollar level" is: which metric can more accurately reflect the speedup achieved by lab researchers and engineers? Is it the productivity study from METR (which shows that using coding tools actually decreased the efficiency of merging PRs by about 20% when developers worked with familiar libraries)? Or is it the human-equivalent time span for self-contained coding tasks (which has now reached several hours and doubles every four to seven months)? I have no direct experience, but I guess it's closer to the former, as there is currently a lack of good feedback validation loops, and the criteria are very broad (maintainability, taste, etc.).

Jack: I agree, this is a crucial question—and the data is conflicting and sparse. For example, we surveyed developers at Anthropic, and among the 60% of respondents using Claude, they self-reported a 50% increase in productivity. But studies like METR seem to contradict this. We need better data, especially regarding tool monitoring for developers inside and outside AI labs, to see what is actually happening. Looking at the bigger picture, the unprecedented and large-scale adoption of coding tools does indicate that people are subjectively experiencing significant gains—if more and more developers are keen on making themselves less efficient, that would be very counterintuitive.

Dwarkesh: Don't blame me for being pedantic, but the METR study precisely predicted this: self-reported productivity is far higher and may even be in the opposite direction of actual productivity.

Jack: Yes, I agree. Without revealing too much, we are specifically considering how to monitor and figure out the "truth" here, as people's self-reports may ultimately be very different from reality. Hopefully, we can produce some relevant research results by 2026!

Which company is winning?

Michael: Do you think the podium will continue to rotate? From what I've heard, Google is gaining favor among developers at AWS and Microsoft. And it seems that the company's "search inertia" has been cleared.

Dwarkesh: Interesting. In my view, the competition is more intense than ever. The feedback on Opus 4.5 and Gemini 3.5 Pro on Twitter has been great. I can't judge which company will win, but it's definitely not settled yet.

Jack: I also think it's more competitive than ever!

Dwarkesh: I'm curious about everyone's thoughts: how many failed training runs or "dud" models can Anthropic, OpenAI, or Google withstand? Considering they constantly need to rely on revenue and so-called "vibes" for funding (by the way, what exactly is the funding for?).

Michael: The secret of Google Search has always been its low cost, so those information searches that cannot be monetized (accounting for 80% or more) won't turn into huge losses for the company I think this is the fundamental issue with generative AI and LLMs today—they are too expensive. It's hard to understand what their profit model is, or what the competitive advantage of any given model might be—whether it's able to charge higher fees or has lower operating costs.

Perhaps Google will ultimately become the one with the lowest operating costs and win this war that will eventually turn into a commoditized economy.

Dwarkesh: Great point. Especially if you think that much of the progress over the past year has been the result of inference scaling, which requires exponential growth in variable costs to maintain.

Ultimately, the price ceiling of things is determined by their replacement cost. Therefore, only if progress continues rapidly, and as Jack said, ultimately achieves self-compounding, can foundational model companies maintain high profit margins (which they currently seem to be doing).

Why hasn't AI taken all our jobs yet?

Dwarkesh: The complexity involved in automating work and simulating human behavior is surprising. We have crossed so many common-sense definitions of AGI—the Turing Test is hardly even worth mentioning now; we already have models that can reason and solve complex, open-ended coding and mathematical problems. If you had shown me Gemini 3 or Claude 4.5 Opus in 2017, I would have thought it would lead to half of white-collar jobs being lost. However, the impact of AI on the labor market, even if it does exist, requires a microscope to see in spreadsheets.

I also find the scale and speed of private capital investment in AI surprising. Just a few years ago, people were still discussing that AGI had to be a government-led "Manhattan Project," as that was the only way to transform the economy into a computing and data engine. But so far, it seems that old-fashioned market forces can fully support AI investments that account for several percentage points of GDP.

Michael: Dwarkesh's point about the Turing Test is good—it has indeed been discussed for a long time. But in the past, during the Industrial Revolution and the service revolution, the impact on the labor force was so significant that it necessitated the establishment and expansion of compulsory education to keep young people out of the labor pool for longer. We clearly haven't seen anything like that happen yet.

Jack: Yes, Dwarkesh and Michael, one norm in the AI community is that they keep constructing so-called challenges to measure real intelligence, and then AI systems break through these benchmarks, only to find yourself facing a system that appears very capable on the surface but can still make mistakes that seem extremely absurd or unbelievable to any human. A recent example is that according to benchmark tests, LLMs scored at "superhuman" levels on a range of so-called difficult cognitive tasks, but when they make mistakes, they cannot self-correct. This is improving now, but it illustrates how counterintuitive the weaknesses of AI models can be. And you often find these weaknesses while witnessing significant progress Dwarkesh: I'm wondering if the opposite is also true—do humans consistently make mistakes that LLMs find extremely absurd or unbelievable? Are LLMs really more "uneven" than humans, or is it just a different kind of unevenness?

Patrick: Borrowing an observation from Dwarkesh's book: One ordinary way LLMs exhibit superhuman abilities is that they can speak more languages than any human—by an unimaginable amount—and their proficiency exceeds that of almost all multilingual learners. It's incredible that this happens by chance, and even the labs didn't specifically train for it. One of the most astonishing demonstrations I've seen is an LLM that was originally intended to be trained only on English documents, yet it could translate a CNN news article into Japanese at a level roughly equivalent to a professional translator. From this perspective, an LLM that hasn't been trained on "politeness" might say, "Humans are really strange and specialized; look at how many of them live in a world with books but can't speak Japanese."

Why Many Workers Haven't Started Using AI (Currently)

Patrick: Programming seems to be the vanguard of large-scale industrial applications of AI, with companies like Cursor experiencing skyrocketing revenues, and discerning tech experts beginning to use Claude Code and OpenAI Codex, with the term "vibe coding" becoming widely circulated. This has led to a noticeable asymmetry in enthusiasm for AI, as most people are not programmers. What will be the next industry to change? What kind of change will manifest in financial reports, employment, or prices, rather than just in demo presentations?

Jack: Programming has a nice property in that it is relatively "closed-loop"—you generate or adjust code using LLMs, then validate it and push it into production. Until a broader set of tools emerges, LLMs have this "closed-loop" property outside of programming—such as the establishment of web search capabilities and the emergence of features like Model Context Protocol (MCP) connections, which significantly extend the "closed-loop" utility of LLMs beyond programming.

For example, I've recently been researching cost curves for various things (like the cost per kilogram to orbit or the cost per watt of solar energy). Before having these tools, you could also research with LLMs, but the friction was enormous, forcing you to switch back and forth between LLMs and other tools. Now that this friction has disappeared, you'll see a significant increase in adoption rates. Therefore, I expect what is happening with programmers will soon happen with a broader range of knowledge workers—this should manifest in a diffuse but widespread way in fields like scientific research, law, academia, and consulting.

Michael: Ultimately, someone has to pay for AI. Someone outside pays for goods or services, and that is GDP. The growth rate of this spending is at the GDP level, 2% to 4%—companies with pricing power may see some uplift, but this seems unlikely in the future AI landscape The "cake" of the economy does not magically expand but is constrained by arithmetic. There is no mysticism. The entire software cake—SaaS software that runs various business and creative functions—is less than $1 trillion. This is why I keep returning to the ratio issue between infrastructure and applications—NVIDIA has sold $400 billion worth of chips, but the corresponding revenue from terminal AI products is less than $100 billion.

AI must improve productivity and create entirely new categories of spending that do not cannibalize other categories. This is very difficult. Can AI improve productivity enough? That is debatable. The current capital expenditure cycle is based on faith and "fear of missing out" (FOMO). No one has been able to provide truly viable data. At least not yet.

There is a very simple narrative out there that believes AI will make everything so wonderful that spending will explode. But it is more likely that it will reduce spending. If AI replaces a $500 seat license with a $50 one, that is good for productivity but deflationary for productivity spending. Moreover, the productivity gains achieved are likely to be shared among all competitors.

Dwarkesh: Michael, isn't this the "lump of labor fallacy"? That is, the belief that the total amount of software to be written is fixed, and using this as the upper limit of AI's impact on software?

Michael: New markets will indeed emerge, but their development will be slower than what interest-driven futurists believe. It has always been this way. Demographics and total addressable market (TAM) are often just marketing gimmicks rather than grounded in reality. The population in Europe is shrinking. The U.S. is the only major Western country that is growing, and that is due to immigration, but the immigration issue has also been politicized. FOMO is a potent drug. Look at some comments from Apple or Microsoft; it seems they are also aware of this.

Dwarkesh: By the way, it's quite ironic that when AI emerges, we happen to need it to save us from the demographic pit that the economy would have fallen into over the next few decades.

Michael: Yes, Dwarkesh. In healthcare, there is a real shortage; the number of human doctors in the future will not be sufficient. High-quality healthcare must become cheaper, and technology is needed to expand the reach and coverage of real medical expertise.

Will engineers be unemployed?

Patrick: The "Big Five" (AppAmaGooFaceSoft, namely Apple, Amazon, Google, Facebook, Microsoft) currently employ about 500,000 engineers. Please provide a number for 2035 and explain your logic—or argue that employee count is a flawed variable and point out the balance sheet or productivity metrics that should be focused on.

Michael: Since 2000, Microsoft has added 18,000 employees while its stock price has remained stagnant for 14 years. In fact, despite experiencing severe stock market crashes, the employee counts of Cisco, Dell, and Intel have hardly changed So I think this is a flawed variable. It doesn't tell us anything about value creation, especially for companies that are cash-rich and in a monopoly, duopoly, or oligopoly position. I believe this number will be lower, or not much higher, because I think we are heading towards a very long-term recession. When the stock prices of the ultra-large cloud vendors fell in 2022, they laid off employees, and when the stock prices rose, they hired most of them back. This is just a few years of volatility.

I will track the total cost of "stock-based compensation" (SBC) before I dare to say that productivity is growing at record levels. At NVIDIA, I calculated that about half of its profits were offset by stock-linked compensation, which transferred value to employees. Well, if half of the employees are now worth $25 million, where is the productivity gain for these employees reflected? Not to mention that if SBC costs were accurately accounted for, the profit margins would be much lower.

The core metric for measuring everything is the return on invested capital (ROIC), and these software companies used to have very high ROIC. Now that they are turning into capital-intensive hardware companies, ROIC will surely decline, which will depress stock prices in the long run. In the market, nothing can predict long-term trends better than the direction (up or down) and speed of ROIC. Right now, these companies' ROIC is plummeting, and this situation will continue until 2035.

In an interview with Dwarkesh, Satya Nadella said he hopes software can maintain ROIC during heavy capital expenditure cycles. I don't see that happening, and even for Nadella, it sounds like just a hope.

Dwarkesh: A naive question, why is ROIC more important than absolute returns? I would rather have a large company that can continuously grow (even if the return on investment shrinks) than a small company that makes money like a printing press but has a cap on its scale.

Many large tech companies have seen their ROIC decline, but their potential market over the next 20 years has expanded from advertising (with annual revenues of $400 billion) to labor (with annual revenues in the hundreds of trillions).

Michael: The return on invested capital—more importantly, its trend—is an indicator of how much opportunity a company has left. In my view, I have seen many "consolidation" cases where companies grow through debt acquisitions. This makes ROIC incredibly clear. If the returns from these acquisitions ultimately fall below the cost of debt, the company will collapse like WorldCom.

At some point, the expenditures on AI construction must yield returns above the cost of investment; otherwise, there will be no incremental economic value. If a company grows merely by borrowing more money or spending all its cash flow on low-return projects, it is not attractive to investors, and the price-to-earnings ratio will decline. Many non-tech companies generate a lot of cash, but apart from buying ready-made solutions, they have no real growth prospects, and their trading price-to-earnings ratios are only about 8 times

Where is all the money flowing?

Patrick: From the perspective of the capital cycle, at what stage do you think we are in AI development—early over-investment, mid-cycle reshuffling, or some stage that is fundamentally different from past tech booms? What would change your view?

Michael: I do believe this is different from previous booms, the difference being that capital expenditures are extremely short-lived. Chips are iterating every year now; today’s data centers will not be able to support chips a few years down the line. One could even argue that a large portion of this should be expensed rather than capitalized. Or it should be depreciated over two, three, or four years.

Another huge difference is that private credit is financing this boom as much as, if not more than, the public capital markets. Private credit is a murky area, but the maturity mismatch is very pronounced—most of these assets are being securitized as if they could last 20 years, while giving hyper-scale cloud providers an exit opportunity every four to five years. This is just asking for trouble. There will be a lot of stranded assets.

Of course, the ones spending the money are the richest companies on Earth. But whether it comes from cash or capital markets, massive spending is massive spending, and the planned expenditures have already overwhelmed even the balance sheets and cash flows of large cloud providers like today.

Moreover, "Construction in Progress" (CIP) is now an accounting tool, and I believe it is being utilized. Capital equipment that is not "in use" does not start depreciating and does not reduce income. They can stay there indefinitely. I foresee a lot of stranded assets being hidden in CIP to protect profits; I think we are already seeing signs of this.

In an interview with Dwarkesh, Nadella mentioned that he pulled back on some projects and slowed down construction because he didn’t want to be trapped by the four to five years of depreciation on a generation of chips. This is akin to a form of "self-supply."

We are currently in the mid-cycle—having passed the stage where the stock market rewards investors for further construction, we are entering a period where real costs and income scarcity begin to show.

In past cycles, the stock market and capital markets typically peaked about halfway through, with the remaining capital expenditures being completed under an increasingly pessimistic (or realistic) view of the relevant assets.

Dwarkesh: I think it entirely depends on whether AI continues to advance rapidly. If you can really run the smartest human brains on B200 chips, then we are clearly under-invested. I believe the current revenue information at the application layer is not as informative as the original predictions about AI capability advancements themselves.

Jack: I agree with this—the degree of capability advancement in recent years has been surprisingly significant, leading to an explosive growth in AI usage. Future model capabilities may see further leaps, which could have extremely significant impacts on the economy.

Where did the market go wrong?

Patrick: In the AI supply chain, where has value accumulated? How is this different from recent or historical technological advancements? Who do you think the market is most mistaken about right now? Michael: Historically, value accumulates in the hands of those with enduring competitive advantages across all industries, manifested as pricing power or insurmountable cost or distribution advantages.

It is still unclear whether the spending here will lead to such outcomes.

Warren Buffett owned a department store in the late 1960s. When the department store across the street installed escalators, he had to install one too. In the end, neither benefited from that expensive project. There were no lasting improvements in profit margins or cost, and both remained in a state of perfect competition. Most AI implementations will likely be similar.

This is why the lack of a clear path to tangible economic applications despite trillions of dollars in spending is so concerning. Most will not benefit because their competitors will benefit to the same extent, and no one will gain a competitive advantage from it.

I believe the market has the deepest misjudgment regarding the two benchmark companies in AI—NVIDIA and Palantir. These are two of the luckiest companies. They have adapted well, but their luck comes from the fact that when all this started, no one had designed products specifically for AI. They just happened to be used as AI products.

NVIDIA's advantage is not enduring. For most AI use cases, SLM and ASIC are the future. If necessary, they will be backward compatible with CUDA. NVIDIA is a power-hungry, inelegant stopgap solution, merely holding the fort before competitors with entirely different solutions come in.

Palantir's CEO compared me to a [bad guy] because of an imagined short bet of $1 billion against my company. This is not how a confident CEO should behave. He is desperately marketing to maintain the status quo, but this will decline. After accounting for equity incentives, the company has almost no earnings.

Dwarkesh: Whether AI labs can achieve lasting competitive advantages through recursive self-improvement effects remains to be seen. But if Jack is right and AI developers should have already seen significant productivity gains, then why is competition now fiercer than ever? Either this internal "eating your own dog food" cannot sustain a competitive advantage, or the productivity gains from AI are smaller than they appear.

If the outcome proves that (1) no one in the AI supply chain can earn crazy profits, and (2) AI is still a big deal, then clearly the value has accumulated in the hands of customers. That sounds great to me.

Michael: In the example of the escalator, the only value indeed flowed to the customers. This is always the case if producers or providers cannot charge monopoly rents.

What Would Change Their Views

Patrick: What kind of headline in 2026 (whether technological or financial) would surprise you and prompt you to recalibrate your overall view on AI progress or valuation? Looking back, what has been the biggest surprise or cognitive recalibration so far? Michael: The thing that surprises me the most and prompts me to recalibrate is that autonomous AI entities have replaced millions of jobs in large companies. This would shock me, but it wouldn't necessarily help me understand where the lasting advantages lie. It's another example of Buffett's escalator.

Another is that application layer revenue reaches $500 billion or more due to the proliferation of killer applications.

Currently, we will see one of two scenarios: either NVIDIA's chips can last five to six years, thus reducing demand for them; or they can only last two to three years, leading to a collapse in cloud vendors' profits and the destruction of private credit.

Looking back, the biggest surprises so far are:

1. Google has not been leading all the way— all eight authors of "Attention Is All You Need" are Google employees; they have search, Gmail, Android, and even LLMs and chips, but they messed up and gave much weaker competitors an opportunity. Google is chasing a startup in the AI field: this is unbelievable.
2. ChatGPT— a chatbot has sparked a multi-trillion-dollar infrastructure race. It's like someone made a prototype robot, and then companies around the world started investing in the future of robots.
3. NVIDIA has managed to maintain such a long dominance in the inference era. I thought ASICs and SLMs should have already taken the lead by now, and we should have moved beyond the prompt engineering phase. Perhaps the obsession with NVIDIA has hindered participants, or it could be due to NVIDIA's anti-competitive behavior.

Dwarkesh: The biggest surprise for me would be:

1. The cumulative revenue of AI labs in 2026 being below $40 billion or above $100 billion. This would mean that things are accelerating or decelerating significantly compared to my expectations.
2. The "continuous learning" problem being solved. Not in the sense that context learning is "solved" like GPT-3, but rather that GPT-5.2 has an understanding of context that is almost human-like. If working with the model feels like replicating an old employee who has worked with you for six months, rather than extracting labor from it in its first hour of onboarding, I think this would be a huge release of AI capability.

I believe the timeline for AGI has significantly shortened since 2020. Back then, you might assign a certain probability to "scaling GPT-3 a thousand times would achieve AGI," and you might also assign a certain probability to "we are completely off track and must wait until the end of the century." If progress breaks through the trend line and points to the emergence of truly human-replaceable intelligence within the next 5 to 20 years, that would be my biggest surprise.

Jack: If "scaling hits a wall," that would be truly surprising and have extremely far-reaching implications for the underlying research paradigm and the broader AI economy. Clearly, the construction of infrastructure, including massive investments in future AI model training facilities, indicates that people are betting that scaling will not hit a wall Another thing that surprised me is: if a technological breakthrough occurs that can improve the efficiency of distributed training, and a group of actors gathers enough computers to train a very powerful system. If this happens, it means that not only can you have open-source weight models, but also an "open model development" model, where training a cutting-edge model no longer requires a large single entity (like a company). This will change the political and economic landscape of AI and have extremely significant policy implications, especially regarding the diffusion of cutting-edge capabilities. Epoch has a good analysis on distributed training that everyone can refer to.

How They Actually Use LLMs

Patrick: What was your most significant professional interaction with an LLM recently? Please omit sensitive details if necessary. In that interaction, what was your relationship with the LLM?

Michael: I now use Claude to generate all my charts and tables. I find the raw data, but I no longer spend time creating or designing professional tables, charts, or visuals. I still don't trust those numbers and need to verify them, but that kind of creative work has become a thing of the past for me. Relatedly, I specifically use Claude to find raw data because much of the current data no longer exists solely in SEC filings or mainstream reports.

Patrick: I think people outside the finance circle cannot understand how many of the highest-paid and best-educated individuals are hired to act as Microsoft PowerPoint and Excel experts. While this still has value, perhaps the social status value of pivot tables and VLOOKUP() will outlast their functional value, but in my speech at the Bank of England, all the charts were also created using LLMs. In hindsight, it was absurd that we used to ask humans to spend hours fine-tuning them.

Dwarkesh: They are now my personal one-on-one tutors. I tried to hire human tutors for different topics I was preparing, but I found that the response speed and pace of LLMs made the experience qualitatively much better. I am getting a digital equivalent experience: just like people are willing to pay a premium far above Uber for Waymo. This leads me to believe that the "human premium" for many jobs will not only not be high, but may actually turn negative.

Michael: On this point, many people believe that blue-collar skilled work is immune to AI. Given that I can now complete many home electrical and other repairs myself with Claude by my side, I'm not sure that's correct. If I were middle class facing an $800 bill for a plumber or electrician, I might just use Claude directly. I love the feeling: take a picture and then figure out all the steps needed to fix it

Risks, Power, and How to Shape the Future

Patrick: The views of relatively informed individuals on AI risks vary widely, from "it may cause discomfort on social media" to "the downside risks include the complete destruction of everything humanity values." What keeps you up at night? Additionally, if you had five minutes to speak with senior decision-makers, how would you advise them to reallocate their attention and resources?

Jack: What worries me most is whether people can successfully achieve "AI building AI"—completely closing the loop on AI development (sometimes referred to as recursive self-improvement AI). To be clear, I believe that by January 2026, the likelihood of recursive self-improving AI systems emerging on Earth is essentially zero, but we are indeed seeing very early signs of AI becoming better at handling components of AI research (from kernel development to autonomously fine-tuning open-source weight models).

If these things continue to improve, and you eventually build an AI system that can build itself, then the development of AI will dramatically accelerate and may become harder for humans to understand. This will bring about a series of significant policy issues and could lead to unprecedented leaps in world economic activity driven by AI systems.

In other words, if I had five minutes with policymakers, I would tell them: "Self-improving AI sounds like science fiction, but there is no evidence in existing technology to suggest it is unachievable. If it really happens, it will be a monumental event, and you must pay attention to it. You should demand transparency from AI companies about what they are seeing here and ensure you have trusted third parties to test these characteristics of AI systems."

Michael: Jack, I guess decision-makers would listen to your advice, and I hope they do.

As for the current AI, it doesn't worry me too much regarding the risks faced by humanity. I think chatbots have the potential to make people less intelligent—doctors who overuse them may start to forget their innate medical knowledge. That's not good, but it's not catastrophic.

The catastrophic concerns involving AGI or artificial superintelligence (ASI) are not as frightening to me. I grew up during the Cold War, when the world could explode at any moment. We practiced drills for that in school. Helicopters sprayed malathion over all of us while we played soccer. I saw "The Terminator" over 30 years ago. "Red Dawn" also seemed quite likely at the time. I believe humanity will adapt.

If I had the chance to advise senior decision-makers, I would request them to allocate $1 trillion (since throwing around trillions now is like tossing millions) to bypass all protests and regulations, sprinkle small nuclear reactors across the nation, and build a brand new, world-class power grid for everyone. Complete this as soon as possible and protect it with the latest physical security and cybersecurity measures; they could even create a dedicated nuclear defense force to protect each facility, funded by the federal government This is our only hope as a nation to achieve sufficient growth to ultimately repay our debts and ensure long-term security—energy must never become a limiting factor for our innovation.

Jack: I strongly agree with the energy part (although our subjective concerns about other matters may differ!). AI will play a significant role in the economy, and it fundamentally relies on infrastructure to be delivered efficiently and cheaply to businesses and consumers—similar to how countries in the past decided to undertake large-scale electrification, build roads, and construct sewage systems (large capital expenditure projects!). We urgently need to do the same in the energy sector now.

I also believe that large-scale AI data centers are very useful test customers for new energy technologies, and I am particularly excited to see the future integration of AI energy demand with nuclear technology. More broadly, I believe that "economic security is national security," so ensuring we have the infrastructure needed to build an AI economy will have a cascading positive impact on our industrial base and overall robustness.

Risk Warning and Disclaimer

The market has risks, and investments should be made cautiously. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk