HWIN

📣 My debut single comes out on RÜFÜS DU SOL's Rose Avenue Records’ first ever compilation LP Friday August 4th. Pre-save here: hwin.lnk.to/Metropolis 🧵(1/5)

I helped build Facebook. I watched it become a machine for addicting people. Because addiction was more profitable. Now the same logic is driving AI. I wrote about what we do about it. 🧵

Every AI investment carries an implicit short on the consumer economy. Every dollar of margin expansion from replacing workers is dollar of household income lost. Part 3 of the Global Intelligence Crisis is out. Washington has no plan for the biggest labor disruption in history. So we’re starting a conversation around one- The American Prosperity Compact The Global Intelligence Crisis · Part Three The Path Forward This is the third part of a series about AI’s potential impact on the economy. Part One detailed my experience building with agentic AI and laid out the risk of what I'm calling the Intelligence Transition: the structural shift of cognitive labor from humans to machines and the economic reorganization it forces. Part Two, which I co-authored with Citrini, painted a hypothetical downside scenario for 2028 if no policy action is taken and displacement accelerates unchecked. Last month, the essays we authored clearly struck a nerve. The response exceeded anything we anticipated, and the drop in the markets was neither our expectation nor our intent. None of the ideas we presented were individually new, but connecting the dots into a coherent narrative clearly resonated with the world we are all living in. A world where we can all feel the ground shifting beneath us. A world in which for many of us our place and purpose is a bit less clear. A world in which the human intelligence we use to do our life’s work and earn a living looks increasingly competitive with AI. Something has gone wrong when the dominant emotional response to humanity's crowning technological achievement is dread. A common refrain on social media is that you have a few years to make enough money to avoid becoming part of the permanent underclass. I feel some of that dread myself, the fear of losing a sense of purpose, and I say that as someone without meaningful personal economic worries. This shouldn't be the prevailing mood at the dawn of a new golden age. With a properly designed policy framework, we can have both rapid AI progress and social stability. The second part of this essay proposes such a framework: the American Prosperity Compact, designed for today’s political climate. I’m an AI optimist and I’m not writing as a spectator. I’ve spent fifteen years building AI companies and twenty years investing in technology. I use frontier AI models every day to analyze companies, synthesize data and build products. In my experience, today's models approximate the cognitive abilities of a 125 IQ professional working at superhuman speed and endurance. They still miss things and need reminders for things humans don’t, but the trajectory of improvement is astonishing. I believe the ideal outcome from the culmination of humanity’s technological progress is genuinely within reach: broadly shared prosperity, an explosion of scientific discovery and democratized access to expertise. But we won’t get there without a plan to navigate the Intelligence Transition. Part A: The Reality on the Ground The Synthetic Short Every investment today in an AI company or beneficiary carries an implicit short on the consumer economy. Every dollar of margin expansion from replacing human workers is a dollar of household income lost from the demand side. The lost dollars are largely earned by middle-class families who spend almost all of it, while the margin expansion and potentially higher stock prices accrue to shareholder owners. Shareholders are composed primarily of wealthier households, foreign owners and pension funds, all of whom have high propensities to save rather than spend the incremental wealth. The mismatch is large. Of course, AI investment will lead to a productivity boom. But if the layoffs come fast enough and consumer spending drops in response, the handoff between higher productivity and faster GDP growth can easily break down until the Intelligence Transition has been navigated. This would ultimately be an aggregate demand problem, and the Fed would intervene aggressively, but it’s not clear that low or even negative rates alone would suffice to spur enough incremental demand to offset the largest ever shock to the cash flows and balance sheets of the American consumer, who collectively account for 68% of U.S. GDP and 100% of the electorate. The AI complex and the consumer economy are increasingly on opposite sides of the same trade. That is not sustainable and is not something that the AI complex, investors or society should want. Markets already sense this. Even before the latest geopolitical shock, US markets had stopped going up despite strong AI progress, a healthy economy and robust earnings growth. The software sector has been under heavy pressure for six months; since January, that pressure has spread to AI-exposed businesses in intermediation and financial services. The popular explanation cites concerns about AI capex sustainability. That’s part of it, but I think there is a deeper unease: the market is beginning to price a version of what we discussed in Parts One and Two. AI is clearly the largest technology shift of our lifetimes. What happens to the economy, existing market shares and employment on the other side? No one has a good answer, and that uncertainty alone is a drag on valuations. Addressing counterarguments to Part One and Two Many readers of my two previous essays argued that this moment will be similar to previous periods of profound economic shifts across history. As in the past, labor markets will adjust to automation and new job categories will arise to take the place of lost ones. The standard framework to think about this in labor economics is Autor, Levy and Murnane's landmark 2003 paper, which sorted all workplace tasks into four quadrants: routine cognitive (bookkeeping, clerical), routine manual (assembly, sorting), non-routine cognitive (analysis, management, persuasion) and non-routine manual (driving, food prep). Logically, routine jobs are easier to automate. Routine manual work had been declining since the dawn of the factory line, but since the 1970s, computers began replacing routine cognitive work. This was offset by the fact that non-routine cognitive work continued to grow, offering a path for displaced workers. AI threatens to break this framework, as it is the first technology that targets the non-routine cognitive work quadrant itself, the refuge that absorbed every prior wave of displaced workers. There is no higher tier to escape to. If we take seriously the idea that this time could be different, that AI is a different sort of technology, the risk may also be different and larger than anything the modern economy has experienced. Previous waves of automation displaced workers from specific tasks in specific industries over decades. The Intelligence Transition threatens to compress that disruption across the entire knowledge economy on a timeline measured in years. The second key counterargument from readers was that we won’t have sufficient compute for a mass jobs displacement scenario by 2028. While compute availability and cost is certainly tightening today, the AI industry’s track record on algorithmic improvement to reduce compute required for a unit of intelligence has been stellar. Inference costs have fallen 10x or more annually. This algorithmic progress will only accelerate if the labs focus on it in a compute constrained environment. Moreover, we are mobilizing global supply chains in previously unheard of ways to build more compute. We will spend nearly $1 trillion on AI capex globally this year, and at current growth rates we will be spending over 1% of global GDP on AI compute by 2028. We've Seen This Before The optimistic refrain is that technology always creates more jobs than it destroys. Over very long periods, that has been true. But the transition periods are brutal, and the people caught in them don't live on geological timescales. The Engels Pause is the canonical example. Using Robert Allen’s classic estimates, output per worker rose sharply between 1790 and 1840 while real wages rose much more slowly. For sixty years, the gains from the industrial revolution flowed almost entirely to capital while the people doing the work captured a small fraction of the wealth they were creating. The political debt came due in labor laws, a welfare state built under duress, and a political realignment that shaped British governance for a century. The mill owners of Manchester didn't price it. Neither did the investors. The China Shock, and more broadly the era of trade-driven manufacturing displacement is a more recent example. When China joined the World Trade Organization in 2001, the resulting import surge destroyed 2.0 to 2.4 million American jobs by 2011, according to Acemoglu et al. The losses were concentrated geographically, and the communities hit hardest are still depressed. Manufacturing employment dropped from 17 million to 11 million between 2000 and 2010. Mortality rose in trade-exposed areas. "Deaths of despair" became a research category. The political backlash from those job losses is the direct ancestor of today's tariff regime. We are still paying for the policy failure of the 2000s. A typical counterargument to automation-driven employment losses invokes the ATM. When automated teller machines rolled out across America, bank teller employment actually rose. ATMs made branches cheaper to operate, banks opened more branches, and tellers shifted from cash handling to relationship banking and selling financial products. Classic Jevons paradox: make an input cheaper, and demand for the output increases enough to preserve or increase demand for the input. But the ATM story has a second act that the optimists leave out. The technology that actually killed teller jobs was the iPhone. Mobile banking didn't automate some of the teller's tasks. It eliminated the reason to visit a branch at all. Once you could deposit checks, transfer funds and manage accounts from your phone, the demand for the entire branch model collapsed. Teller employment has fallen steadily since 2010. The distinction matters. When technology automates some tasks within a role, workers can redirect to higher-value tasks and the role survives or even expands. When technology automates the underlying need, the role disappears. AI is doing both simultaneously. The Layoffs Have Started Block's 40%+ layoffs announced on February 26th serve as a vivid case study. The company cut over 4,000 of its roughly 10,000 employees, with CEO Jack Dorsey citing AI explicitly as the driver. AI is playing double duty: a convenient catch-all for unwinding Covid-era hiring excess, and a genuine structural driver. As Dorsey stated: "Intelligence tools have changed what it means to build and run a company. A significantly smaller team, using the tools we're building, can do more and do it better. And intelligence tool capabilities are compounding faster every week. I don't think we're early to this realization. I think most companies are late." Since Block’s announcement, Meta is reportedly planning to cut ~16,000 workers (20% of its workforce) to help fund $135 billion in AI capex, Oracle is reportedly preparing to shed up to 30,000 roles (12-18% of staff), and Atlassian announced cuts of 1,500 (10%). The Skill Distribution Problem AI allows top performers to produce several times the output they could have a year ago. I can see this clearly across my teams. The trouble is in the distribution. If one person augmented by AI can do the work of three, you have less need for the other two. The top 10-15% of employees become much more valuable and more in demand than ever. The bottom 20-40% are in genuine danger. Through this process, employers get the double benefit of leaner teams with far lower coordination costs. This dynamic is not limited to engineering. It applies to legal research, financial analysis, marketing, design or any domain where AI can generate a first draft that a skilled practitioner can refine. The premium on judgment, taste and domain expertise goes up. The premium on simple execution goes way down. That's an inversion of how most knowledge-work careers are structured, where you start by executing and gradually learn your way into judgment. What happens when the first rung of the ladder disappears? The Entry-Level Collapse ServiceNow CEO Bill McDermott said it plainly on CNBC recently: college graduate unemployment “could easily go into the mid-30s in the next couple of years” as AI absorbs entry-level work. The data already shows early stress. The unemployment rate for recent college graduates has risen to 5.7%, up from 3.8% in 2023, the highest level in over a decade outside the pandemic. For the first time in modern data, recent graduate unemployment now exceeds the overall national rate. Recent graduates are the most exposed because they have the least accumulated judgment and domain expertise, the very things AI is sprinting to replicate today. Recent grads are merely the first group to face pressure from AI, the same logic will extend to more experienced professionals as models improve. Labor Market Supply-Demand Balance The zeitgeist is focused on which specific jobs disappear and whether new ones appear. This misses the more important dynamic: what replacement jobs pay. The labor market is one interconnected system. When a $100K accountant gets displaced and competes for jobs in the broader economy, she takes a job in retail, putting downward pressure on retail wages. When millions of displaced knowledge workers flood into the job market, the result is wage compression across the entire distribution. The people at the bottom get squeezed hardest. This is what happened with the China Shock. The cascade doesn't stop at white-collar displacement. The same AI acceleration that threatens knowledge work is accelerating robotics and physical-world automation. Autonomous driving is no longer hypothetical. Waymo’s autonomous rideshares operate in 10 cities and rival Lyft’s market share in SF, its most mature market. Tesla launched its robotaxi in Austin last June and plans seven more cities in 2026, with the explicit long-term goal of converting every Tesla on the road into an autonomous cab via software update. Collectively, truck driving, delivery trucks and rideshare provide over 4M jobs in the US. Warehouse automation, fast-food preparation, retail checkout and last-mile delivery by drone and robot: these are all in active and growing deployment throughout the U.S. today. This entire category of work is on a clear countdown. The blue-collar jobs that would have absorbed white-collar workers are themselves disappearing on a lag measured in years. The wave ultimately hits the entire consumer economy. Part B: The American Prosperity Compact I am not arguing for a punitive "AI tax" or for slowing frontier development. I’m a capitalist; I build companies and I invest in public and private markets. Free markets created the wealthiest society in human history, and it is vital we preserve them through the Intelligence Transition. But markets require functioning consumers. An economy where a generation of Americans cannot pay their mortgages is failing, not free. My solution is called the American Prosperity Compact. It is designed to close the synthetic short. It converts the unpriced tail risk of a disorderly transition into a defined, known cost structure that capital markets can underwrite. It keeps the consumer economy functioning while the productive economy reorganizes. The Compact is structured in four cascading, contingent parts. The Foundation lays out sensible labor market reforms for today. The Circuit Breaker triggers only if job displacement accelerates past certain thresholds. The Backstop triggers only if we see profound levels of job displacement. Finally, the Accelerator is where we play offense to make the economy work better in an AI world. If machines replace humans in the production function, the math eventually leads to taxing those machines. There are many new approaches such as a token tax or higher tax rates on agents. Some may be right for the 2030s, but none can pass in 2027. The Compact is designed for today’s political economy. Every mechanism builds on existing tax code, existing benefit structures and existing institutional capacity. The goal is a framework that can be enacted in the next year and operational soon thereafter. The tax mechanisms are deliberately concentrated at the corporate level, at the source of displacement itself- the companies experiencing the most margin expansion from replacing human labor with AI. This avoids the thornier issues of raising income tax brackets or blanket wealth taxes, especially on unrealized capital gains. Corporate rates are an ideal approach because in a world where AI drives large revenue growth and cost savings opportunities simultaneously, a moderately higher rate will not meaningfully deter corporate investment when the opportunity set is so large. The companies benefiting most from the Intelligence Transition will still be wildly profitable. The question is whether a portion of that windfall recirculates into the consumer economy or concentrates at the top while aggregate demand erodes beneath it. Building the tax system for a fully post-transition economy is a broader question for future work. This framework is the first step: what our politics can actually deliver in the next year. The Foundation The Foundation removes structural distortions that made sense in a 20th-century economy but are already liabilities today. It addresses pre-existing frictions and can be designed as revenue-neutral. Stop taxing labor more than AI. The employer side of Social Security and Medicare is funded by a 7.65% payroll tax on every dollar of wages. AI-generated output carries no equivalent cost. The employer side of the tax base should shift from wages to corporate value-added broadly: revenue minus purchased inputs, which equals wages plus profits. Because the base is broader, the rate can be lower while generating the same revenue. Companies that employ many people relative to output would actually pay less. Companies generating enormous output with minimal workforces would pay a bit more. Make benefits portable. The nature of work has already changed. Roughly 38% of the workforce does some form of freelance or independent work. AI will push more people toward portfolio careers, contracting and one-person firms. AI levels the playing field for entrepreneurship, the lifeblood of our economic dynamism and a core part of the American dream. But a would-be entrepreneur who can't leave her job because her kids' health insurance depends on it is a structural failure. Portable benefits that travel with the person, not the position, would unlock the risk-taking and business formation our economy needs. The goal is not to nationalize insurance but rather to decouple core benefits from any single employer so people can carry them across jobs, transitions and periods of self-employment. Pro-rata contribution accounts, where every employer or platform pays into a worker's portable fund based on hours worked, and expanded ACA marketplace subsidies that make individual coverage genuinely affordable, can make this real without nationalizing anything. The Circuit Breaker The Circuit Breaker triggers if the job displacement is rapid. I am not proposing we enact it tomorrow. I propose we design it now so it can trigger automatically if conditions warrant. Think of it as insurance. Designing this now, before it is needed, means economic and political tail risk converts into a defined cost structure. The trigger is labor's share of GDP. Today that number sits around 54%. If it stays there or increases, nothing happens. The Circuit Breaker only activates if labor share falls below a sustained threshold, likely in the low 50s. That would represent a structural break, not a cyclical dip. Until that threshold is crossed, the mechanism lies dormant. Two mechanisms once it triggers. First, a corporate displacement tax: a tax that scales automatically with the gap between labor's share of GDP and that threshold. The companies seeing the most margin expansion pay the highest rates. If margins are expanding while headcounts collapse, the tax goes up. If the labor market recovers and labor share climbs back above the threshold, rates step back down. The data decides. Second, targeted income support for displaced workers, funded by the corporate displacement tax. Built on Earned Income Tax Credit (EITC) principles: wage insurance where a worker whose new job pays substantially less than their old one receives partial income replacement for a transition period, phased so re-employment always beats waiting. Pair it with an expanded EITC with higher income ceilings, because as displaced white-collar professionals flood into the blue-collar labor market, wage compression pushes down incomes for everyone below them too. Crucially, both preserve the core EITC insight that earned it bipartisan support in the first place: every dollar of support is conditioned on work. To be clear: the EITC only works when jobs exist. If displacement outpaces job creation, we should broaden the definition of qualifying work to include schooling, retraining, caregiving and community service. Support remains contingent on contribution and commitment, not on employment specifically. There is a deeper reason fiscal stabilizers matter here. If AI simultaneously cuts costs and displaces workers, the Fed faces a contradiction it cannot resolve: deflation invites rate cuts, but rate cuts will not create jobs in categories that no longer exist. Central bankers have minimal experience implementing negative rates. Five central banks tried between 2012 and 2024 and none went below -0.75%. Japan ran the experiment for eight years without escaping deflation. Monetary policy has a structural floor. The Circuit Breaker adds the required fiscal policy boost to maintain household income through a secular restructuring rather than a temporary downturn. The Backstop I hope we never need the Backstop. If we do, it means household debt has become unserviceable and the financial system is under stress. Displacement and credit deterioration reinforce each other in a loop, and once that loop starts it is very difficult to stop. Mortgage delinquencies rise, housing prices soften and bank balance sheets take mark-to-market hits that tighten lending further. TARP is the apt analogue. The Backstop exists to break the loop before it starts. It includes two elements, funded by higher rates of the corporate displacement tax: Comprehensive income security. Extended income replacement, mortgage forbearance and healthcare continuity, designed to prevent a household liquidity crisis from becoming a banking solvency crisis. An American AI Dividend Fund modeled on the Alaska Permanent Fund. A portion of the increased corporate taxes would be invested on behalf of every American. Alaska does this through taxing resource extraction. The analogy extends as AI is built on the corpus of human knowledge and drives displacement through the existing economic system. Rather than a handout, this is a share of returns on a collectively created asset, making citizens shareholders in the transition The Accelerator The Foundation, the Circuit Breaker and the Backstop protect the downside. The Accelerator clears the institutional bottlenecks that would cap AI's GDP contribution even if the labor transition were managed perfectly. The defensive and offensive agendas are part of the same agenda. These are mostly sensible policies that economists have supported for years. They haven't been implemented because they involve clear constituencies that win and lose. The Intelligence Transition creates a unique opening to form the coalition needed to push them through. But that coalition only forms if the first three tiers of the Compact demonstrate a real commitment to protecting workers. Deregulation doesn’t work without a safety net. Energy and grid infrastructure. Energy, not compute, is now the binding constraint on AI growth. Permitting reform for datacenters, transmission, interconnection and nuclear power is a national security priority. China is building capacity aggressively to support its AI ambitions, and every year of U.S. permitting delay is a year of competitive ground lost. Moreover, the same AI driving energy demand is accelerating breakthroughs in fusion, next-gen solar, enhanced geothermal, battery chemistry and grid optimization, making the energy buildout and the clean energy transition the same project. (Continued in next Tweet)

Co-author of the viral Citrini Research piece @alapshah1 explains the key part of its thesis in detail: "What is the thing that drives our entire economy? It's wages. Most of those wages that are ultimately driving all the discretionary spending are coming from the white-collar worker." "The problem with that is [we] made all these assumptions on loaning money to all these companies. Two mortgages, everything else. The white-collar economy is our economy." "If you all of a sudden just take a leg out of that economy, it has a contagion effect. Basically, every asset in the world. And so that, I think, is the part people haven't thought about, because when you were making these loans, no one ever conceived of a world [where] white-collar jobs are in permanent decline." "If that's at 2% a year, then I think we can skate through. But if it's at 4% or 5% a year, then we need action a lot more quickly."

Alap here, co-author of the piece w @Citrini7 We didn't have time to go through specifics in the piece, but what happens when you can ask Gemini or ChatGPT to order you a car or ramen? Let's say it's 2028 and 40% of the orders are via agent. ChatGPT can farm this out to lowest bidder and take a cut of the savings themselves. These are relatively consolidated markets but I'm sure GRUB and CART would love to compete on cost and pick up share. New startups only need to compete on supply, service and cost, can outsource user acquisition to agents

The man of the hour, live on TBPN

AI is humanity's Promethean moment- fire, stolen at last from the gods. Burning through jobs, this fire will cause a financial conflagration. We collectively possess the water and sand to tame this fire, but we must act swiftly to begin an urgent society-wide policy dialogue. Excited to partner with @Citrini7 to release the first part of our series, The Global Intelligence Crisis- The Rise of Agentic AI After years of exponential growth, the recent capability jump to agentic AI is set to upend the world as we know it beginning in 2026. Here is the uncomfortable truth: AI is no longer just a tool for economic growth; it is a near-direct substitute for human cognitive labor. In the near term, it will displace white-collar workers far faster than new markets can absorb them. In this three-part series, I draw on 20 years of experience investing in public markets and 15 years building AI companies to outline my perspective on the coming economic storm. Our entire economic system is built on a single premise: human intelligence is a scarce, expensive resource. It is the key input required to turn raw materials into goods and services that dictate our living standards. In 2026, as agentic AI comes of age, that foundational assumption is collapsing. AI is no longer a mere tool or complement; it is rapidly becoming a direct substitute for human cognitive labor. This shift fundamentally devalues white-collar work. Fueled by surging AI progress and capital, this displacement will inexorably accelerate, creating an economic shock that could eclipse the Industrial Revolution, the Global Financial Crisis and the pandemic. Without urgent policy action, this has the potential to cause a significant financial crisis in the next two years. My warning should not have you mistake me for an AI doomer. AI is humanity's Promethean moment- fire, stolen at last from the gods. Its capabilities put us less than a generation away from abundant living standards for all, limitless clean energy and the eradication of most diseases. Our challenge is not the technology itself but rather surviving the severe economic shock of its arrival and required rewiring of our financial system. The era of AI as a simple conversational chatbot is over. Over the past six months, we have crossed a critical threshold into the era of agentic AI- systems that operate autonomously to execute complex, multi-step workflows. The pace of this evolution is staggering. According to METR, a third-party that evaluates autonomous AI capabilities, the duration of tasks models can complete unaided is doubling every six to seven months, with the most recent examples suggesting a further acceleration in trend. Today, the leading models can execute 14.5 hours of continuous, autonomous work. The trendline points to a full month of unaided work by mid-2028. My Vantage Point I began my finance career as a consumer analyst at two large hedge funds, Viking Global and Citadel, where I invested through the global financial crisis and its aftermath. In 2011 I left to run my own fund called LOTUS. There, it became clear that the limiting factor on my performance was my ability to process the growing torrent of information that drives markets. My workflows were split across Bloomberg, S&P Capital IQ, Excel, Outlook, OneNote and various financial apps and websites. To centralize these disparate workflows, my brother Naman and I built Sentieo, an AI financial search engine. The results were striking–I could see the world more clearly, operate faster with a smaller team and generate better investment performance. We eventually grew Sentieo to over 1,000 investment management, banking and corporate clients before selling the business to a competitor for over $200 million in 2022. Since selling Sentieo, I have been running @lotusaifund, @LittlebirdAI - a personal AI company, and Studio Management- a startup incubator. Building Sentieo fundamentally shifted my worldview. I realized AI is the ultimate force multiplier, and I began seeking ways to use it in every facet of my companies and life. Our Use of Agentic AI At my companies, we aren’t just observing the agentic AI trendline; we have aggressively rewired our organizations to run on it. Nowhere is this shift more dramatic than in software engineering. Prototyping a new feature used to require a week of designing specifications with a product manager and designer, followed by a week of engineering iteration with a small team of engineers. Now, agentic coding interfaces allow me to write a detailed prompt and generate a functional prototype in minutes. While it isn’t usually production-ready on the first pass, it largely cuts other humans out of the initial build process and significantly increases our velocity of shipping finished product. We are seeing a similar trajectory at LOTUS. A year ago, AI models could do a decent job of answering basic financial structure and valuation questions. Today’s agentic AI is fully connected to Factset and S&P Capital IQ financial and document databases. In the old model, if I wanted to research a new idea, I would assign it to an analyst. It would take a few days for the analyst to read the relevant documents, research the key debates and data and then prepare a financial model and email with their key findings. We would then iterate on this over another few days. Today, an agentic AI can synthesize the filings and key debates, construct the financial model and generate a comprehensive memo of near comparable quality in a few minutes. This quick turnaround means I can rapidly drill down into the key issues and reach an investment decision days faster than before. The token cost for delivering this agentic work is less than 1% of the cost of a human performing the same work. Agents don’t need to sleep, don’t take vacation and can be spun up (and down) into swarms of agents as needed. Perhaps the most profound change, however, is the way we coordinate in an agent-driven organization. Human coordination is the largest, most exponential cost in any business. Economist Ronald Coase’s seminal 1937 Theory of the Firm can be paraphrased as- Firms exist because internal coordination costs are lower than market transaction costs, but only up to a point. The firm stops growing when the marginal cost of organizing one more internal transaction equals the cost of doing it via the market. Passing instructions from a founder to a product manager to an engineer is a lossy game of telephone, requiring endless messages, meetings and presentations just to keep everyone aligned. Indeed, we can think of the entire Microsoft Suite of Outlook, Word, PowerPoint and Excel as human coordination technologies. AI agents, however, share nearly perfect, continuous context. Where feasible, swapping humans for agents eliminates this massive coordination tax, collapsing friction and ramping output. We are not actively shrinking our teams today, as we run high growth early-stage businesses that are gaining market share. However, we have significantly slowed the pace of our hiring and need fewer humans than in the past. Each human can do more and is expected to adopt AI aggressively to multiply their output and impact. There are certain roles that we have concluded are better performed wholesale by agents instead of humans. These include data analytics, data migration, certain design roles, certain devops roles and certain customer service roles. This list is growing monthly as AI capabilities improve. Jobs Commentary from AI CEOs The AI threat to jobs is certainly not a new or original idea. There has been a growing chorus of warnings of white-collar job replacement emanating from the AI labs over the past year. Unsurprisingly, CEOs have been reticent to put the pieces together to go from layoff risks to downstream economic consequences. In an interview with Axios last May, Dario Amodei, CEO of Anthropic, warned that "AI could wipe out half of all entry-level white-collar jobs – and spike unemployment to 10-20% in the next one to five years." While Amodei has an incentive to talk up Anthropic’s potential, his capability improvement predictions look prophetic today. Nine months later, Mustafa Suleyman, CEO of Microsoft AI, seemed to say the quiet part not so quietly in an interview with the Financial Times last week. "White-collar work, where you’re sitting down at a computer, either being a lawyer or an accountant or a project manager or a marketing person – most of those tasks will be fully automated by an AI within the next 12 to 18 months." State of the Labor Market Agentic AI is clearly ramping up its ability to perform white-collar work and AI CEOs seem very worried about layoffs. It’s worth examining the state of the white-collar job market coming into the rise of agentic AI. In the chart that accompanies this piece, the dashed white line represents core white-collar employment, excluding employment in sectors driven by government spending, specifically Government, Health Care (government spending is half of spending) and Private Education (government loans and loan guarantees drive a significant portion of the market). We can clearly see a stagnant and declining trend in core ex-government white-collar employment since 2023. While there was certainly a 12-18 month hangover from the torrid post-pandemic hiring boom, the last 18-24 months betray a fragile balance. These core white-collar jobs are up only 4% from pre-pandemic levels over six years, compared to population growth of 5% and real GDP growth of 11% over the same period. The Information sector, which should be ground zero for AI job losses, already shows an 8% drop from its peak, with current levels below even 2020 pre-pandemic levels. Corporations are clearly doing more with fewer humans even before agentic AI enters the scene. Labor Market Supply-Demand Balance A previous section laid out agentic AI adoption in a typical startup that stands to benefit more from rapid revenue growth than from labor cost savings. This approach represents a roadmap for how larger companies will adopt agentic AI over the next year. However, larger companies have much higher coordination costs, more automatable legacy processes and most importantly, larger steadier businesses with less revenue upside and more cost savings opportunities. Ultimately this means much higher potential for corporate layoffs that have been a consistent feature of the white-collar labor market since 2023. The nexus of a weak white-collar labor market and agentic AI adoption suggests a growing risk of a white-collar jobs crisis. Agentic AI will accelerate these trends and market forces will multiply them. Critics’ objections that large enterprises will move slowly are fair, but most companies operate in competitive markets. Any company that is slow to adopt agentic AI will see a cost disadvantage and an impaired competitive position versus peers. CEOs understand this dynamic and are almost universally making AI adoption their top priority for 2026, with spending budgets to back it up. It won’t take many layoffs to upset the already fragile supply-demand balance for white-collar labor. Imagine we get 5% white-collar job losses in 12-24 months, which seems to be significantly less than what Dario or Mustafa are suggesting. These jobs are unlikely to come back as AI progress continues to accelerate. These displaced workers will be forced to seek blue-collar and gig economy jobs, putting downward pressure on wages for all workers in the economy. Employees who keep their jobs will be keenly aware of the growing risk leading to plummeting consumer confidence and spending. Contagion Risk Together with Citrini, we have written a detailed prospective timeline of how a crisis might unfold in Part Two–THE 2028 GLOBAL INTELLIGENCE CRISIS, A Thought Exercise in Financial History, from the Future. Rather than repeat all the details here, this is the high-level view of how we think it goes down. The 5% job loss estimate above assumes the economy is a closed system near equilibrium. It is not. The economy is highly reflexive, and the engine driving job losses, AI intelligence itself, continues to accelerate every quarter. First, there is no natural brake. AI capabilities improve, companies need fewer workers, displaced workers spend less, weakened companies invest more in AI to protect margins, and AI capabilities improve further. Each company’s individual response is rational. The collective result is a negative feedback loop that feeds on itself. Second, the spending damage is wildly disproportionate to the job losses. The top 20% of earners drive roughly 65% of all US consumer spending. These are the white-collar workers most exposed to AI displacement. A modest percentage decline in white-collar employment translates into a much larger hit to discretionary consumer spending, devastating the businesses that depend on it and triggering further layoffs. Third, AI agents will dismantle the vast intermediation layer of the US economy. Over fifty years, we have built trillions of dollars of enterprise value on top of human limitations: things take time, patience runs out, and most people accept a bad price to avoid more clicks. Agentic AI eliminates this friction. Software, consulting, financial services, insurance, travel, real estate and payments are all built on monetizing complexity that agents find trivial. As these sectors suffer steep revenue losses, they will shed jobs aggressively and compound the bleeding. Fourth, the financial system is one long daisy chain of correlated bets on white-collar productivity growth. Over $2.5 trillion of private credit has been deployed into leveraged buyouts underwritten against revenue assumptions that no longer hold. The $13 trillion mortgage market is built on the assumption that borrowers will remain employed at roughly their current income for thirty years. These aren’t subprime borrowers–they’re 780 FICO scores who put 20% down. The loans were good on day one. The world just changed after they were written. Fifth, the government’s fiscal position inverts at the worst possible time. Federal revenue is essentially a tax on human work. As white-collar incomes decline and payrolls shrink, tax receipts dry up just as the need for transfer payments surges. The government will need to send more money to households at precisely the moment it is collecting less from them. Where I Could Be Wrong There are a few important ways that this forecast could be wrong. The most likely way is that job losses are very gradual, allowing AI driven productivity to accelerate and boost GDP growth. A roaring economy with near stable jobs would allow for a gradual transition to an AI world. This is the market’s current base case. While certainly possible, the information sector payroll trend since 2023 looks particularly damning for this theory. Moreover, it would likely require AI progress to slow significantly which seems like a losing bet based on current trends. The second way involves comparing AI to past technological revolutions. The thinking goes that in every previous cycle when technology and automation replaced jobs it created more new jobs in new sectors. While true, every previous technology was a complement to human labor, not a near-direct and near-complete substitute. Every previous technology revolution also coincided with periods of robust job growth; US core white-collar jobs have been shrinking for over 3 years and are hopelessly off the pre-covid trend line. The third way is that decisive policy action acts to prevent a crisis. I’m not here to handicap those odds today, but I do believe that there is a credible path to align most of the electorate with many corporate, AI and political stakeholders towards this future and I’m hoping to start that dialogue in earnest soon in the forthcoming Part Three: The Path Forward. Disclosures As a fund manager and startup builder, my job is to forecast the future and allocate capital and resources accordingly. Because I see this AI-driven displacement as a likely path, my portfolios and companies are positioned for it. If my thesis plays out, my firms will benefit financially. I am stating this plainly for full transparency, but my intent in writing this is not to “talk my book” or cause panic. The societal risks of this transition are simply too massive to ignore. Even if there is a 10% chance of this specific crisis materializing, the downside is severe enough that we must start a society-wide dialogue today. Acknowledgments Thanks to my littlebird co-founders Alex Green and Naman Shah, as well as David Shor, Citrini and Josh Constine for feedback on ideas and proofing. Claude, ChatGPT, Gemini and littlebird were also each important in gathering the data, researching concepts, checking numbers, creating graphics, rewriting and finally proofing this essay. Each shined in different areas reflecting the spiky and rapidly evolving nature of the intelligence frontier. In comparison, the multiple word processors I used, and their spelling and grammar check functions felt terribly dated.

JUNE 2028. The S&P is down 38% from its highs. Unemployment just printed 10.2%. Private credit is unraveling. Prime mortgages are cracking. AI didn’t disappoint. It exceeded every expectation. What happened?​​​​​​​​​​​​​​​​ citriniresearch.com/p/2028gic

Friends in NYC! You're invited to the next installment of Founders, Funders & Friends on 9/12, featuring Scott Norton @swhnorton, founder/CEO of Sir Kensington's (acquired by Unilever). RSVP here: bit.ly/FFF-Sept

A thread about our depraved healthcare system, and a plea. In 2021, my friend Carole was diagnosed with terminal stage four cancer. A single parent with two kids, she had just turned 43. In 2017, she had watched her brother Chris die of the same cancer. He too was in his 40s.

I co-founded a healthy food company @ThistleCo. I haven't been very involved in it for years but it always brings me a smile when I see it, in this case in an IG reel. Hugh Jackman is just like me - riding around on a CitiBike & fitting his Thistle bag in the basket!

🤖 Our team is looking forward to NY @Techweek_! 🎙️ We are hosting a fireside chat on AI with @thesephist on 10/18 in Manhattan ‼️ We have 475 friends on the guest list, apply to join us on Luma! cc @alapshah1 @johnnyhwin @thesephist @katekiewel

NYC-based founders and builders: next week we're co-hosting Studio Management's Founders, Funders, and Friends event in Tribeca, featuring a fireside chat with SignalFire partner @oanaolt and veteran product leader @eugenewei. We'd love to see you! fiftyseven.nyc/fff-september

🗣️ Hosting a special talk for founders in New York City featuring Eugene Wei (Remains of the Day), Oana Olteanu l(SignalFire), & Alap Shah (Studio Management) 📅 Thursday, September 18th, 2023 from 6:30-9:30p 🍹 Drinks & light bites for early attendees  📩 Please RSVP early as space is limited cc @eugenewei @johnnyhwin @alapshah1

"Metropolis Heights" heralds a new body of work coming out this year. Crazy to think that many of these ideas were born during that summer in 2018, & that it's taken 5 years to finally get them out into the world. Patience has been a virtue, but now it's time to climb. (5/5)

and embodies the excitement from the formative days of creating what would evolve into the synth studio and community known as 57. (4/5)

📣 My debut single comes out on RÜFÜS DU SOL's Rose Avenue Records’ first ever compilation LP Friday August 4th. Pre-save here: hwin.lnk.to/Metropolis 🧵(1/5)