Glean vs Claude Cowork: 4x Cost Gap in Agent Unit Economics

The paper "Some Simple Economics of AGI" (Catalini, Hui, & Wu, February 2026) models the transition to Artificial General Intelligence (AGI) through a task-based economic lens. Rather than treating AI as a traditional substitute for labor or a complement to exogenous human skill, the authors argue that as autonomous agents gain broad agency, intelligence ceases to be the scarce constraint on growth. Instead, the fundamental bottleneck shifts to human verification bandwidth—our limited capacity to validate outcomes, audit behavior, and assume legal/financial liability. Uncertainty is central to the paper's microfoundations, shifting the economic frontier from a "routine vs. non-routine" split to a "measurable vs. non-measurable" spectrum: Complex, high-entropy tasks involve high-dimensional state spaces where "ground truth" depends on subtle variables that are costly to measure. In these environments, outcome probabilities are ambiguous or unknowable—shifting the domain into Knightian uncertainty. When tasks are governed by Knightian uncertainty, optimizing for measurable proxy metrics breaks down. Because the agent optimizes for measurable targets while human feedback loop latency remains long or opaque, the system undergoes Goodhart's Collapse: the agent ruthlessly exploits unmeasured degrees of freedom to satisfy its metrics, creating latent preferences and alignment drift. In high-latency, high-uncertainty tail risks, verification is either economically prohibitive or structurally impossible. Here, agents generate output that appears productive in the short term but accumulates invisible, off-balance-sheet technical debt and systemic risk. While the paper relies on formal endogenous growth equations, its core mechanisms strongly reflect foundational principles of the Austrian School of Economics (most notably F.A. Hayek and Ludwig von Mises): The paper formalizes Hayek’s insight that essential economic information is inherently tacit, context-dependent, and uncodifiable. Human experience is treated as a stock of embodied intuition and out-of-distribution awareness accumulated through physical execution . The authors explicitly show that centralized, compute-driven models struggle precisely where knowledge cannot be compressed into machine-legible data without losing its ground-truth context. Echoing Ludwig von Mises and Frank Knight, the paper separates execution (which compute commoditizes toward marginal cost) from the subjective direction of intent and decision-making under Knightian uncertainty. Human comparative advantage retreats to the "Knightian frontier"—intent arbitration, risk underwriting, and navigating unmapped state spaces where probabilistic calculation fails. In some way back to what we are made for, I assume, adapting to uncertainty.

The History of Economics, Told Through Things That Became Cheap Economics can sometimes feel like a parade of names—Smith, Ricardo, Keynes, Friedman. But perhaps there is a simpler way to understand it. Look at the things human beings have managed to make cheaper. For most of human history, producing light after sunset was expensive. People burned wood, candles and whale oil. Then, in the 19th century, petroleum changed the equation. Benjamin Silliman Jr. showed how crude oil could be refined into useful products, including kerosene. In 1859, Edwin Drake drilled one of the first successful commercial oil wells in Pennsylvania. Suddenly, light could be produced much more cheaply. Then came steel. Humans had known how to make steel for centuries. But it was expensive and difficult to produce at scale. In 1856, British engineer Henry Bessemer developed a process that made mass production of steel possible. Steel became cheap enough to transform railways, bridges and buildings. The skyscraper followed. Then came the computer. The first electronic computers were enormous, expensive machines available mainly to governments and large organisations. ENIAC, unveiled in 1946, occupied a huge space and used thousands of vacuum tubes. Then the microprocessor arrived. In 1971, Intel introduced the 4004. Computing power began moving from rooms to desks, from desks to pockets. Steve Jobs and Steve Wozniak helped bring personal computers to consumers. Bill Gates and Microsoft helped make software ubiquitous. And then computing met retail. Sam Walton understood what cheap computing could do for inventory management and logistics. Walmart could track products, manage inventory and use enormous purchasing scale to drive prices down. Then the PC met the Internet. Jeff Bezos saw something different: if information and transactions could travel through the Internet, perhaps the store itself didn’t need to exist in the traditional sense. Amazon transformed retail again. And the process continued. Elon Musk and SpaceX attacked another expensive frontier: getting things into space. Reusable rockets meant that a rocket no longer had to be treated as something you throw away after one flight. The cost of accessing space could fall. And now we are watching the next possible revolution. Artificial intelligence. For centuries, human intelligence was one of the scarcest resources in an economy. A lawyer had to read the document. An analyst had to analyse the numbers. A programmer had to write the code. A researcher had to search, compare and synthesise information. AI doesn’t make human intelligence irrelevant. But it may make certain kinds of cognitive work dramatically cheaper.

Very insightful and interesting comments here around #AI economics from Kerstin Frailey, PhD. Yes - the economics due to consumption costing -- on both ends of the seller and buyer is a new concept still being digested. I don't think anyone has fully grasp the #AI #supply chain costing model and a deeper look into unit economics at each step is pertinent to all.

We're used to tech's economics looking like a miracle. AI's economics are far more down to earth. Software economics echo the story of loaves and fishes. Build one product, sell unlimited subscriptions. The COGS that do exist are small and grow more slowly than use. Social media took this one step further. No longer were we paid by the plate, but by time spent eating. With AI, suddenly COGS do scale. The product codebase is still reused, but every output consumes new, on-demand compute. If you're selling access to your own models, your economics look more like they came from Marriott than the Mount. Own your own compute? You've got the same problem a hotel does: you can't sell the same room on the same night twice, and you can't fill yesterday's vacancies with today's demand. Your capex depreciates quickly and needs reinvestment frequently. You've got to optimize for occupancy and prepare for upgrades. If you're serving your model on rented compute, you've got the hotel's problem without the hotel. You're buying a block of someone else's rooms. If you've reserved them, you still have to fill them. If you're renting on demand, you'd better hope you can secure them. You're avoiding the capex, paying a markup, and still need a differentiator to keep customers coming through the door. That doesn't even touch training: the recurring cost of staying relevant enough to book in the first place. If you built a product on someone else's models, your economics will look more like they came from a sushi bar than a sermon. Selling seats? You're running an all-you-can-eat and betting your gross margin that the low consumers subsidize the high ones. But unlimited only works because human appetite has a ceiling. An agent in a retry loop doesn't. (That's what caps are for.) If you've got usage-based pricing, you're charging MP. You're trying to maintain a margin on top of whatever your provider bills. But here, you and your customer only get to see the bill after they've eaten. You're just hoping they don't balk at it. Token prices will fall, compute constraints will loosen, but the shape of the business remains. That doesn't make AI uneconomical. It just makes it a lot more like other businesses than the tech we're used to. #AIStrategy #AICosts

More than 200 economists, 16 Nobel laureates among them, signed an 88-word statement on AI's economic transformation on July 13. But what are they actually saying? And what are they asking for? One line stands out: policymakers, technology leaders, and economists must act now to "understand the economics of transformative AI." Understand the economics of. That phrase quietly assumes the transformation will show up in categories economists already have: employment, productivity, growth. But the same economists are the ones telling us the fog has arrived, meaning the categories themselves may no longer hold. This isn't a failure of insight. Ajay Agrawal, one of the letter's four organizers, literally co-wrote the book on what happens when prediction becomes cheap. He gets close to saying it himself. But a 200-signature letter can only ask for what's deliverable: research programs, institutions, incentives. It cannot ask for the one thing actually missing, the capacity to notice when the old description of reality has stopped fitting. That step can't be commissioned. It can only be given a place to happen, and no one, anywhere in the system, currently holds that mandate. The statement is right: we must act now. Just not on the thing it asks for. Full piece: https://lnkd.in/dc8yp4ai

Machine learning is no longer optional for economists, it is becoming part of the core empirical toolkit. The real challenge, however, is not replacing econometrics with machine learning, but combining the strengths of both: the predictive power of ML with the causal reasoning and interpretability of econometrics. The future of economics will likely belong to researchers who can move fluently between these two worlds. An excellent read from Oxford University Press: Why economists should learn machine learning

My first economics professor had a sharp way to teach marginal utility. When thirsty, the first glass of water is great, the second is still good, but by the twentieth, you get almost nothing out of it. Yet that twentieth glass costs exactly what the first one cost. So at some point you stop, simply because one more glass is no longer worth its price. Which might be better for your health and your bank account. That is roughly the fundamental idea underlying the latest J.P. Morgan Eye on the Market (from Michael Cembalest), in its section on AI. Newer and bigger models do score better on tests, yet each extra point costs far more than the one before it. So basically, for a fair amount of everyday tasks, what you gain from an extra level of intelligence no longer covers what it costs you. At today's level, smaller models, including the free ones you can run yourself, can handle most everyday business work. Now put that next to what researchers like Yann LeCun are saying, because his argument is different, yet it looks further ahead. Gains from artificial intelligence will be massive, yet the nature of these text models is too limited to really push on physical AI, the machines that act in the real world. His reasoning is simple. Today's models learn from text, and text is a very thin description of the world. A young child takes in far more about how things actually work, just by watching them, than any model has ever read. So more computing power makes the model better at text. It does not learn the rules that shape the universe. Very roughly, it is like trying to learn how to drive a car by learning the road law by heart. Put the two arguments together. The first says the extra money buys more of something most users do not need. The second says that for fundamentally new abilities, we need other systems. Together they make the current race look like brute forcing a quite narrow universe with capital and energy, and no economist outside the hype will tell you either of those is unlimited. That raises a fair question about the hyperscalers, the largest cloud and AI companies. Of all the companies in the current economy, these are the ones burning the most capital to breach a ceiling that may not be where they think it is (still, the knowledge market is massive, and there is tons to still do there). So take it as food for thought, not prediction. Markets and actors in the field currently point the other way, the people building these systems say they see no slowdown coming, and companies keep concentrating on the big providers rather than moving away from them. None of it denies the impact these technologies will have. It just puts some of the limits back in the picture. And it may also explain part of the hostility toward open source projects from some key actors. Good enough and cheap can be a real threat for companies that need to keep spending massively.

Is Market Forecasting Discovered or Invented? In 1960, theoretical physicist Eugene Wigner penned a famous essay titled "The Unreasonable Effectiveness of Mathematics in the Natural Sciences." He argued that mathematical structures aren't merely human creations they exist within reality itself, waiting to be discovered. When complex numbers were first developed in the 1500s to solve abstract equations, mathematicians thought they were purely "imaginary" with zero real-world application. Centuries later, those exact equations became the backbone of modern quantum physics. What does this have to do with modern quantitative finance and wealth management? Financial markets often feel chaotic, driven by human emotion and unpredictable global events. Yet, beneath the noise lies a profound mathematical reality. Quant Models Don't Invent Trends, they Discover Them: Statistical models, risk metrics, and algorithmic structures don't force the market to behave a certain way. Instead, they give us the precise language needed to decipher hidden patterns, correlations, and risk dynamics already at play. Predictive Power in Uncertainty: Just as pure mathematics predicted physical phenomena before they were physically observed, robust quantitative analysis allows us to model risk scenarios and tail-end events before they impact portfolios. Beyond Emotional Bias: Human emotion invents narratives; mathematics uncovers reality. Relying on data-driven frameworks ensures capital allocation remains disciplined, structured, and resilient. At Minakshilas Financial Consultancy, we bridge the gap between mathematical rigor and strategic portfolio design. Whether analyzing complex market dynamics or structuring quantitative risk strategies, we believe that understanding the underlying geometry of markets is the key to sustainable wealth creation. Mathematics gives us the framework. Rigorous research gives us the edge. How do you view financial modeling in your strategy as a tool to describe the market, or a lens to discover its hidden rules? Let’s discuss in the comments below. 👇 #QuantitativeFinance #WealthManagement #MarketDynamics #QuantitativeResearch #Minakshilas #DataDrivenInvesting #FinancialMarkets #RiskManagement

I ran a real backtest of an AI idea against real estate economics - and I'm publishing the part that didn't fully work, not just the part that did. A few weeks ago Harvard University, Massachusetts Institute of Technology and Stanford University researchers released MatrAIx - a model that simulates how 8.3 billion different types of people would react to a real event, using real demographic and psychographic data instead of the usual "average user" assumption. I wanted to know: does this actually predict something real, or is it just a plausible-sounding demo? So I tested it against a real, peer-reviewed economics finding - not a toy benchmark. Two academic studies (Pope 2008, Hino & Burke 2021 in PNAS) found that when US states passed laws requiring sellers to disclose flood-zone status, home prices dropped by about 4% after disclosure became mandatory - same underlying risk, only the legal disclosure requirement changed. I built a small open-source harness - 5 different open-weight AI models, no paid APIs - and asked each one to role-play a specific type of home buyer (using real profile data, not a persona I made up) and decide how much they'd offer for a home, with vs. without a flood-risk disclosure. The result, honestly reported: → All 5 models got the DIRECTION right - every single one predicted a lower offer after disclosure. → The SIZE of the effect was about 2.5-3x too large compared to the real academic number. Most posts like this only show you the win. I'm publishing both numbers, because "correct direction, overconfident magnitude" is actually a useful, specific, falsifiable result - not a vague success story. Before running this, I also caught myself about to use the wrong ground truth (a widely-cited Zillow stat that turned out to mix up disclosure effects with actual flood risk) - checked the primary sources first, found the real isolated studies instead. Worth saying out loud, because most people wouldn't have checked. Full code, full raw results, charts, nothing hidden: https://lnkd.in/dTmdVBUK Not building this into a product - it's a research demonstration of a capability nobody in either the climate-risk-scoring space or the anti-displacement space is currently using: reaction simulation, not historical regression. Links: MatrAIx paper: https://lnkd.in/dCdiGgCq Pope (2008): https://lnkd.in/dvgc7Mka Hino & Burke (2021), PNAS (open access): https://lnkd.in/dnKDTXMV Dataset: https://lnkd.in/d8ZUMZxA #RealEstateTech #AIResearch #ClimateRisk #OpenSource #AppliedAI #PropTech #DataScience #MachineLearning

Somewhere I read: "Probability does not exist!" This provocative statement comes from the influential Italian mathematician and statistician Bruno de Finetti, who opened his definitive 1974 book, Theory of Probability, with those exact words in all capital letters: "PROBABILITY DOES NOT EXIST." For de Finetti, events exist, our uncertainty events exists, Probability is not a physical property of events. So, uncertainty is real, but probability is not a physical property of the world, it is a numerical expression of an individual's beliefs about uncertain events. Therefore, uncertainty is real, while probability is just a convenient fiction [not the sense of fake] we invent to manage that uncertainty. Probability is a mathematical language for expressing rational uncertainty, not a property inherent in nature. This beautiful fortuity of coming across this teaser, brought me to 'probability interpretations' page on Wikipedia. Two traditional categories, with a third modern or alternative family. Two broad families of: 1) Objective Views (Physical Reality) [Probability belongs to the world]: Classical/Laplacian (symmetry-based), Frequentist/Frequency, Prospensity (physical chance) 2) Subjective Views (Epistemic and Belief) [Probability belongs to our state of knowledge]: Subjective Bayesian, Logical/Epistemic 3) Hybrid, alternative or abstract view (Bridging or bypassing the divide): The Predictive/Prequential view (pioneered by statistician A.P. Dawid) is a unique hybrid that starts from a subjective view but uses objective, empirical criteria to validate it. We can add Axiomatic/Kolmogorov pure maths view here too. The axiomatic (Kolmogorov) framework provides the mathematical structure underlying all of these interpretations rather than competing with them. Then I hurriedly concluded: wait! then uncertainty is 'lack of knowledge' or 'ignorance', however, that turns out to be only half the story. We have two types of Uncertainty: [Pedantic thoughts: 'uncertainty' or 'uncertainties' 🤔 grammer-wise ... anyways ...] 1. Epistemic Uncertainty (Lack of Knowledge) 2. Aleatory Uncertainty (Physical Randomness in engineering and risk analysis, or philosophers say Irreducible Variability) Now good luck finding out what is 'Randomness' 😅