# Christian Catalini — notes — full text Founder, Operator, Economist. Co-founder of Lightspark, co-creator of Libra at Meta, founder of the MIT Cryptoeconomics Lab. Research and writing on the economics of AGI, verification, stablecoins, and digital markets. X threads (@ccatalini) unrolled into posts; his own words, in full. Index: https://catalini.com/llms.txt · License: all rights reserved. --- ## Experts followed their agents 86% of the time - canonical: https://catalini.com/notes/experts-followed-their-agents/ - original thread: https://x.com/ccatalini/status/2115476282234548513 - date: 2026-10-28 In February, we argued that the cost of execution would fall much faster than the cost of verification. As agents scale, human attention and expertise become the binding constraint. [@AnthropicAI](https://x.com/AnthropicAI) just put experimental evidence behind our warning... [x.com](https://x.com/AlexTamkin/status/2115190541688398184) Experts followed their agents’ randomly assigned preferences on 86% of choices. Many then regarded the resulting recommendation as their own. This is the failure mode behind the Trojan Horse externality... ...apparently productive work carries assumptions and preferences that humans never independently evaluated. Runaway risk zone: the work is completed, the expert signs off, and measured output rises. But whose experience and preferences shaped the result? ↙️ 🚨 The incentives make this particularly dangerous. Productivity gains are immediate and visible. The costs of weakened oversight are often delayed, difficult to measure, or borne by someone else. Under-verification is privately rational for both individuals and firms! [@AnthropicAI](https://x.com/AnthropicAI)'s data shows this is happening faster than we expected. Keeping a human in the loop tells us very little unless that human has the context, attention, and tools to apply their experience and preferences. Verification infrastructure is how we counter all this. So is "attention ergonomics" (see [@sschillace](https://x.com/sschillace)'s piece, H/T [@livjenkss](https://x.com/livjenkss)). 🚨 We can’t safely automate what we can’t verify, and we can’t verify what no human is paying attention to. Preserving context, exposing consequential assumptions, and making alternatives easy to evaluate are how we expand verification capacity alongside agentic execution. [x.com](https://x.com/ccatalini/status/2026311784421036223) --- ## The Missing Junior Loop, in Anthropic’s own hiring data - canonical: https://catalini.com/notes/missing-junior-loop-anthropic-hiring/ - original thread: https://x.com/ccatalini/status/2112198350732968295 - date: 2026-10-19 Once you see that verifying AI’s work is the bottleneck, and that checking runs on tacit knowledge, you can predict the hiring shift before seeing the data. We did, back in February. Anthropic’s junior share of technical hires: 27% to 9%. Output is priced in tokens. Verification is priced in human hours. Anthropic estimates that reading a week of Claude output would take 424 hours per person! A week has 168. That’s the verification constraint. And reading is the easy part! [x.com](https://x.com/ccatalini/status/2026311784421036223) Junior employment grew more slowly where Claude handled more complex tasks. Our model’s Missing Junior Loop: firms need top expertise while automating the work that produces it. They’re hiring from a stock they’re doing less to replenish. Fix one: flight simulators for work. The “weights” in a junior’s head are too close to the model’s. Same textbooks, same blind spots. Use out of distribution drills and expert feedback to get juniors to the frontier faster. Pilots train for failure. Juniors should too. Fix two is the superpower. The models that ate junior tasks also collapsed the cost of learning and experimentation. Explore more fields. Test more ideas. Discover what you’re unusually good at, then use AI to do more with it. Superpowers, if you deliberately use them. [x.com](https://x.com/a16zfrontier/status/2029628679895273614) These dynamics are visible inside [@Anthropic](https://x.com/Anthropic) than in aggregate labor data. Why? Diffusion. The lab is the extreme case of adoption. The broader economy still has to reorganize around AI, as with other general purpose technologies. That lag is our window to act! [x.com](https://x.com/ccatalini/status/2104709927831048560) Some Simple Economics of AGI, with [@wu_jane](https://x.com/wu_jane) & [@xianghui90](https://x.com/xianghui90). February 2026. Two cost curves. The Measurability Gap. The Missing Junior Loop. And ways out. Paper: [arxiv.org](https://arxiv.org/abs/2602.20946) Report: [anthropic.com](https://anthropic.com/research/employment-at-anthropic) And while it doesn’t cite the paper…the charts do. 😅 --- ## Reality is the final verifier - canonical: https://catalini.com/notes/reality-is-the-final-verifier/ - original thread: https://x.com/ccatalini/status/2107928300740038790 - date: 2026-10-07 Insightful paper by [@istoica05](https://x.com/istoica05) & coauthors: with agentic software, reality is the final verifier. Tests, benchmarks, even formal proofs are just proxies. We have to rapidly change how we operate accordingly. [arxiv.org](https://arxiv.org/abs/2609.12039) It fits with how we model the economics of AGI. They explain how one system fails: agents are checked against what is codified, and find whatever isn't. We unpack why the economy ships those systems into production anyway. [x.com](https://x.com/ccatalini/status/2026315311142875304) Their example: an agent told to speed up a key-value store got 6x by storing a seed and regenerating values on the fly. Tests passed as nobody had written down that a store should store things! Specs will always miss something, and agents will find it. [x.com](https://x.com/ccatalini/status/2026311825898504608) Our contribution: why the incentive is to automate anyways. Generating code is nearly free. Checking it expensive. Shipping unverified work is rational whenever the risk lands on everyone else. We call that the Trojan Horse externality. [x.com](https://x.com/ccatalini/status/2026314389335543867) Both papers land on the same scarce input: humans with authority over what counts as acceptable. They use agents to triage and escalate. We ask how the supply of expert verifiers evolves when juniors lose the routine work that builds expertise. [x.com](https://x.com/ccatalini/status/2026311839089631468) Bainbridge: automating routine work can erode the expertise needed for hard cases. Our "Missing Junior Loop" models that erosion. Human augmentation and synthetic practice are investments in rebuilding and expanding human verification capacity. They explore formal verification, hardware abstractions, world models and evaluator cascades. These all shift our Cost to Verify curve. The ultimate economic test: can verification capacity keep pace with deployment? Do we only automate when risks are acceptable? The obvious shortcut is AI checking AI. Both papers say no! A reviewer working from the same spec has the same blind spots. It's also a market for lemons. The antidote? Reliable liability-as-a-service: AI that guarantees an outcome end-to-end. [x.com](https://x.com/ccatalini/status/2026318830394544590) Their outer loop makes assurance recurring: deploy, observe, revise spec, model or evaluator, rerun. Our model asks whether effective oversight can keep pace with execution. If it cannot, alignment decays. Human augmentation and better measurement shift that frontier. Their closing: implementations can be generated just in time; assurance cannot. Same in our model: execution is free, standing behind it is not, and value accrues to whoever can underwrite the work. Reality Is the Final Verifier: [arxiv.org](https://arxiv.org/abs/2609.12039) Some Simple Economics of AGI: [arxiv.org](https://arxiv.org/html/2602.20946v2) --- ## Is Anthropic the Pennsylvania Railroad? - canonical: https://catalini.com/notes/anthropic-pennsylvania-railroad/ - original thread: https://x.com/ccatalini/status/2107532544015229405 - date: 2026-10-06 Is [@AnthropicAI](https://x.com/AnthropicAI) the Pennsylvania Railroad? It bought $7.33B of compute to carry $4.59B of revenue, ordered $518B more, and IPOs at peak utilization. In 1893 hundreds of railroads entered receivership. Society kept the tracks. Sears, the app layer, thrived. Inventors' predictions about the impact of their own technologies have a terrible record. Radio was going to end war. Nuclear power too cheap to meter. Film, flight, computers: the confident calls, utopian or dystopian, aged worst. Each one was sure "this time was different". Will one of the AI labs be the last company standing? No. Economic history is a graveyard of companies that built the future, then discovered they didn’t own it. General purpose technologies (the original GPTs) are special. They span many sectors. Their quality-adjusted price keeps falling, unlocking new uses. The layers feed each other: a better GPT raises app productivity, and better apps raise the returns to the GPT. Getting the full benefit means redesigning entire systems, not bolting on point solutions. GPTs also have weak appropriability. Nordhaus found innovators keep 2.2% of the surplus they create. GPT inventors likely keep less. The rest goes to the complements. AI was born under a weak appropriability regime, like every GPT before it. There isn't much the labs can do about it. Two outs: domains where a patent enforces excludability, like bio, or where customers pay a lot for a small edge, like finance, cyber, and defense. Steam? Boulton & Watt held the patent. Cotton boomed. Electricity? A regulated utility. Factory labor productivity grew 5.1% a year. Combustion engines? Gasoline and suburbia. Railroads had their own [@DarioAmodei](https://x.com/DarioAmodei): Tom Scott, the "Pennsylvania Napoleon." Standard Oil was most of his traffic. Scott moved into refining to capture value (labs and pharma?). After the price war the Pennsylvania paid Rockefeller 20c a barrel, even on his rivals' crude (!). Over 100 railways went into receivership. The tracks kept running. Without them, U.S. farmland would have been worth 60% less. Private return on the railroads: 3.5%. Social return: 48%. Value accrued to landowners, manufacturers, consumers. The app layer (Sears) boomed. The market for intelligence splits three ways. Frontier generalists (closed) take the top of the demand curve. Low-cost generalists (open) run most tokens. Enterprises worried about disintermediation scale frontier specialists: open models plus proprietary context. A frontier specialist starts with open weights, adds proprietary context, expert feedback and evals. In their domain, frontier specialists will beat closed labs on price & quality. Companies keep control, without funding/training their own replacement. Expect more of this. The labs have one more problem: they need to be successful, but not too successful. The only company left in the world gets nationalized, or turned into a regulated utility, metered by the token. So where does value ultimately accrue? To whoever holds the complementary assets: chips, the ability to turn atoms into datacenters, distribution, and reliable products customers and businesses pay for. The 2026 equivalents of farmland, factories and Sears. Regulation could create one more complementary asset: permission to compete. If policy, safety and compliance costs become barriers to entry, the big labs get a real moat. So how do you survive the moatpocalypse? Start from one observation: anything that can be measured will be automated. Most of the scaling recipe to date is better measurement, not better models. Han & [@dwarkesh_sp](https://x.com/dwarkesh_sp) put 12x of pretraining progress down to better data and 3.7x to better models. 🚨 You can't safely automate what you can't reliably measure and verify. Verification means checking that the agent did what you intended. As execution gets cheaper, proving the task was done right becomes the bottleneck to safe scaling. Which is why every company needs to become a verification factory. Two areas of focus: advance measurement through friction with the real world faster than your competitors, and recruit the best verifiers in your domain. [x.com](https://x.com/ccatalini/status/2106064313207398433) The only network effects that survive AI are verification-grade ones: the ones that benefit from the loop from new measurement and ground truth, to compute, to higher-quality verification that feeds back into new measurement. Establishing verification-grade network effects is not enough. You also need to own the underlying loop to begin with. That means owning the weights of production! But what about RSI? 26% of model R&D is now led by Claude, so progress is real. The last mile will stay a bottleneck. And more capable AI makes knowledge on the periphery worth more, not less. Which feeds back to the owners of complementary assets and context. Hayek's objection to central planning was never a shortage of compute. It was that the knowledge an economy runs on is tacit and tied to time and place. That holds for AI however capable it gets, which points to a world of many competing intelligences, not one company left. --- ## AI Is Making Verification the Bottleneck for Companies: The Thread - canonical: https://catalini.com/notes/verification-bottleneck-companies-thread/ - original thread: https://x.com/ccatalini/status/2106064313207398433 - date: 2026-10-02 Somewhere in your company today, an expert rejected a perfectly plausible answer. A controller caught a revenue figure that counted deferred bookings. A security engineer blocked a release that passed every test but would have handed an AI agent write access to the production database. That override, escalation, or reversal is the most valuable data your firm produces. If you record enough of those moments and use them, alongside everything else your firm knows and measures, to train an AI model that learns from each correction, you could, in theory, eliminate the need for managers to route information and coordinate action. In doing so, execution gets turbocharged by automation. You might even be able to eliminate hierarchy altogether. That, at least, is the pitch. Microsoft CEO [@satyanadella](https://x.com/satyanadella) calls that feedback system the critical “learning loop” of the AI era. He also insists that enterprises need to own it rather than cede it to the frontier AI labs. [@PalantirTech](https://x.com/PalantirTech)'s CEO Alex Karp has likewise argued that companies and governments will need sovereign AI to preserve their independence. [@jack](https://x.com/jack) and [@roelofbotha](https://x.com/roelofbotha) went further, offering a blueprint for an ambitious “company world model”: an omnipresent and continuously updated model of the entire business that could coordinate work autonomously. The vision is admittedly seductive. AI makes context extremely easy to retrieve, digest, and continuously process the moment new information arises. What once took an army of managers and multiple systems of record can theoretically be handed over to a single well-tuned AI model. The solution also draws on decades of economic thinking about why firms exist and how they organize work. In the 1930s, economist Ronald Coase argued that firms arise to reduce the costs of relying on markets: discovering prices, negotiating separate contracts, and adapting agreements as circumstances change. Firms bring activities inside their boundaries when coordinating them internally costs less than transacting through the market. In the 1970s, business historian Alfred Chandler documented how the “visible hand” of professional management outperformed market coordination in large-scale production and distribution. Their shared insight is that internal coordination can confer an economic advantage. If AI can perform that coordinating work more cheaply and reliably than layers of human managers, the economic case for those layers weakens. Dissolving the hierarchy becomes the next logical step. With AI, leaders see a chance to preserve the benefits of internal coordination while stripping out much of the friction of human-to-human management. The problem is that execution is only half of the story. The other half—and the source of some of that apparent friction—is the work of verification. Hierarchy does more than route information; it also verifies: deciding what information means, what deserves the firm’s attention, which outputs can be trusted, and which are worth generating at all. As the ability to generate text, media, and code becomes increasingly commoditized and abundant, the calculation as to which half creates the most value for companies evolves too. When execution is cheap, verification becomes more valuable, and firms turn into verification factories: institutions capable of properly steering AI-generated output and standing behind the results. [x.com](https://x.com/ccatalini/status/2099167910862483917) This explains why AI has progressed unevenly across domains. Enterprise AI has advanced fastest where verification is relatively easy. In coding, outputs can often be tested rapidly and cheaply, helping explain why it has so quickly become a success story of AI deployment. Leaders under pressure to define an AI strategy need to recognize what its success hinges on: a coherent plan to solve the verification problem. They also need to recognize how they may be giving away their advantage in this emerging landscape by letting expertise erode and building on technology stacks they don’t own. Using the economics of verification as a framework, this piece explains what belongs in a firm’s verification loop, why leaders must retain control of it, and how doing so enables reliable autonomy without giving away the tacit knowledge and expertise that are key to long-term competitive advantage. [x.com](https://x.com/ccatalini/status/2026311784421036223) The Verification Bottleneck: increasingly, what can be measured can be automated. Once the relevant information is codified and success can be measured, an AI model can use that feedback to hill-climb toward the desired outcome. [x.com](https://x.com/ccatalini/status/1935708804274119141) But there is a catch: Any divergence between the context a human brings to a problem and the context available to the machine is a potential source of error. Think about the last time you wrote a prompt, only to realize midway through that you had omitted a critical dimension, constraint, or desired outcome. More often than not, the missing detail was something you understood tacitly and took for granted. This is a key reason machine intelligence can feel so “jagged:” astonishingly capable along measurable dimensions, but surprisingly clumsy everywhere else. Verification is the act of closing the gap between the weights that power an AI model and the neural paths and “weights” that power your own decision-making on the same problem. As AI makes execution increasingly cheap, the ability to properly steer and verify AI outputs better than your competitors will be the differentiator. World-class verification factories rely on two assets: unique ground truth and talent. For some firms, unique ground truth is simply a byproduct of scale and incumbency. If you process most of the world’s transactions in a domain, you observe that slice of the economy with a precision nobody can replicate. For others, it comes from extreme focus. If you have underwritten the same narrow risk for decades, you hold ground truth nobody else has seen. Neither is the prerequisite; unique measurement is. The best firms use measurement to turn market uncertainty into manageable risks, repeatable processes, and consistently better products. But ground truth alone is inert. Data only records what happened; when something falls outside the norm, it takes an expert to decide what it means and how to act on it. When domain experts set objectives and verify AI output, they draw on accumulated experience: past errors, corrections, and lessons learned. That experience is itself a form of measurement, one the AI does not have access to. Some call what remains unmeasured taste, others judgment. But these terms are vague, and their boundaries keep shifting: as model outputs improve, more of what once required human judgment is automated. A more concrete, practical distinction starts with a question: What context lives only inside a human, and what is available to the machine? Everything else follows from this distinction. Whatever isn’t yet, or cannot be, measured becomes a bottleneck to safe deployment. Unfortunately, many firms are adopting AI in ways that erode both of these critical assets, data and talent, at once. Tools that summarize documents, interactions, and context risk flattening divergent views into premature consensus, weakening the independence that makes expert talent valuable. Building on technology stacks the firm doesn’t own and letting its experts’ traces accumulate elsewhere erodes the exclusivity of its proprietary data. Together, these choices weaken the very capability that becomes more valuable as AI improves: the ability to tell whether its output is right. In a recent paper on the economics of verification, my coauthors and I mapped tasks along two dimensions: how cheap they are to automate, and how cheap they are to verify. When a task is cheap to automate and cheap to verify, AI can be trusted to execute on its own without a human in the loop. We called this the “safe industrial” zone. But when automation is cheap while verification depends on human expertise, things get dangerous rather quickly. Here, we expect firms to deploy AI even when its outputs are likely to fall short. This is where a firm either defends its role as a verification factory or quietly gives it up. Catching the flaw in an output that is almost right takes experts with the independence and motivation to challenge the model, and ground truth that exists nowhere else. Right now, AI labs are doing everything they can to bring down the cost of verification across domains by aggressively purchasing specialized data and expert evaluations—even going so far as to buy the internal records of bankrupt companies. Each time they acquire richer digital traces of an economically valuable task, they turn something their models once missed into a capability they can sell. This expansion would be less threatening if firms were actively defending their position as verification factories. Most are doing the opposite: handing the labs the traces that make their tasks verifiable while letting the experts who could still catch a bad answer drift out of the loop. The sequence from there is predictable. One task at a time, the labs learn to verify what only in-house experts could verify before. The firm’s edge shrinks to whatever the labs still cannot measure. It becomes a thin wrapper around their intelligence, paying per token for capabilities it once owned while the labs capture a growing share of the actual value. Don’t Automate Away Your Expertise. As firms increasingly buy inference from the same AI labs, their advantage will lie in the difference between the answers those labs provide to everyone and the conclusions they reach using proprietary data and expertise. Without that independent contribution, research suggests firms could rapidly converge toward a monoculture. Think about what managers really do. Good managers are verifiers. Yes, they route information, but in doing so they also determine what it means, what deserves others’ attention, and what meets their quality bar. In making these calls, they often retain a significant advantage even over frontier AI. Anyone with deep experience is a fast simulator of a particular slice of reality: a “world model” made of a person, fine-tuned through years of friction with a specific technology, market, customer base, or regulator. Much of the information behind that model is tacit and unrecorded, learned as much from what failed as from what worked. And in deciding what deserves attention, managers also wield real power within the organization. Now consider ceding many of those responsibilities to AI. An agent could turn sales meetings, support tickets, and internal discussions into a ranked list of product priorities, then use that list to assign work. An experienced engineer hesitates over one feature because it reminds them of a past failure, but they have not yet pinned down the connection. All they write is, “Not sure I’m fully comfortable with this.” The agent records strong customer demand and no concrete technical objection, leaving the hesitation out of its conclusions. A manager may still approve the budget, but the model has already shaped what they see and what they know to question. Handing these decisions to AI is a transfer of power, not simple automation, and it can destroy what makes a firm unique. Lose your company’s edge in what it senses and prioritizes, and everything else falls apart rather quickly. Even an AI system trained on company data is constrained by what has been measured and made available to it. It therefore risks prioritizing what the firm has already seen and codified. This can deliver measurable short-run gains by quickly routing decision-makers to existing solutions, while weakening the firm’s ability to identify and adapt to new, unstructured problems. When AI tools struggle to distinguish an emerging signal from a distraction, they can amplify what is already known at the expense of what is still taking shape, pushing the firm toward premature consensus around the wrong answer. Deployed across a firm, these tools can degrade the very expertise they draw on. Slop, but for decisions. If companies cut experts off from the friction that keeps them learning, they may not recognize the damage until the next big discontinuity, when the models are confidently wrong and no one inside the firm is still used to overruling them. Companies can deploy AI without flattening the independent thinking of their managers and experts. But doing so requires systems that preserve independent judgment, expose disagreements, and help employees test competing views against new data. Rather than compressing everything into a single company-wide memory, AI should help each employee build and continually refine their own version of it. Three design principles follow. Learning loops should test decisions against real-world outcomes and feed the insights back to both models and experts. A healthy loop starts with a clear definition of success. It records the information the AI used, the actions it took, and why experts accepted, corrected, or overrode its recommendations. The firm then tests both the AI’s recommendations and the experts’ corrections against business outcomes and new evidence from the market. Those results should guide changes to evaluation criteria, knowledge, instructions, and workflows, including what the system can approve, reject, or escalate. Proposed improvements should be tested, and the lessons incorporated into the firm’s institutional memory and shared with experts so their judgment improves too. Fine-tuning the model is one possible ingredient in this broader learning system. The loop should also flag when generation outpaces meaningful verification, a warning that the firm may be shipping output it cannot yet trust. The pattern is already visible in software: as AI adoption has accelerated, developer throughput has risen, but so have bugs and production incidents. Agents should surface relevant knowledge, not flatten it. AI agents in a firm’s communication channels should help people articulate tacit knowledge with as little friction as possible. Because agents often lack that context, they should preserve disagreements and unresolved concerns rather than compress them into premature consensus. Experts’ objections can reveal a gap between their “weights” and the firm’s telemetry, bringing previously unrecorded experience into the system. Eliciting those objections helps the firm detect new developments and identify what still needs to be measured. Preserving them also creates an invaluable record for a postmortem, when the firm must trace a failure back to the assumptions and decisions that produced it. World models should be supporting infrastructure, not decision-makers. Perhaps most importantly, an effective company “world model” should continuously spot hidden assumptions and identify where accurate data is still missing, then encourage employees to seek real-world friction to surface more robust evidence. Only through repeated contact with reality can the AI model be genuinely up to date. Sometimes, this can be as simple as connecting people who hold different pieces of relevant context so they can compare notes and make explicit what would otherwise remain uncodified. These countermeasures can help ensure that, as AI becomes embedded in every company interaction, it does not inadvertently force everyone to operate from a narrow, model-compressed version of reality. By amplifying what makes the company and its experts unique, a truly customized model can become one of the firm’s most valuable strategic assets. The goal is to keep the firm’s internal market of ideas alive: treat competing views as hypotheses, prompt individuals to test them against real-world outcomes, and feed the resulting evidence and lessons back into the firm’s institutional memories and expertise. Each turn of the loop makes the firm a better verification factory. But the ultimate design test is how AI supports consequential decisions under fundamental uncertainty. A company-wide model that decides what information matters, resolves conflicting interpretations, and allocates attention and resources risks acting like a central planner and run into Friedrich Hayek’s central critique of planned economies: the tacit knowledge it needs is dispersed among people, tied to particular circumstances, and often never written down. The model cannot reliably outperform people if it lacks critical knowledge they hold. It should help them make that knowledge explicit, test their assumptions, and learn from the results. Otherwise, it risks eroding the expertise it depends on. Owning the Weights of Production A firm can get every component of its verification factory exactly right and still fail to capture the value it generates. Capturing value requires owning the traces and feedback generated by the loop. Today, leading AI labs are courting organizations with tools that reach deep into their document repositories, communication channels, and codebases. They also tempt employees with a Faustian bargain: Let the tools record how you use your computer while you work, and in exchange we will automate your work for you. Companies may be reassured by AI labs’ “no-training” commitments for their prompts and outputs. But they also need to know who can retain and reuse the operational traces surrounding that work: which tools an agent used and in what order, how it failed, and when a human overruled it. This form of “ambient AI” captures the distinctive ways employees handle exceptions, make decisions, and prioritize information. Taken together, those traces can provide a blueprint for replicating parts of a firm’s verification engine. If providers can reuse them to build their own offerings, firms risk handing over the knowledge that gives them an edge. The stakes rise as agents become more autonomous: broader permissions and deeper context put more of the firm’s secret sauce within the provider’s reach. Companies can protect that edge in a few ways. Own the loop. Protecting your firm from suddenly finding itself competing with the AI labs’ models starts with owning the learning loop and controlling how its data is used. The fine-grained, proprietary measurements it generates capture how your experts verify outputs and correct mistakes. Letting the labs reuse those measurements helps them learn to perform that verification themselves. Start by checking exactly what your provider can retain and reuse, and what its no-training and Embrace open-weight models. Open-weight models offer a practical way to retain greater control. They make their underlying parameters available, allowing firms to download, run, and adapt the models themselves. Firms can deploy them on their own infrastructure or through a partner whose business is providing infrastructure rather than building AI products that could compete with the firm. In either case, the arrangement should keep proprietary information, operational traces, and improvements under the firm’s control. A growing coalition is working to build a thriving AI ecosystem that gives firms an alternative to relying on the closed labs. By August, more than 120 organizations had joined the Open Secure AI Alliance to advance open models and open-source tools for AI security. [x.com](https://x.com/JensenHuang/status/2104499465055023424) Open-weight models initially attracted enterprise interest for their potential to lower costs. Their strategic appeal now extends beyond those savings to control: ensuring that the data, feedback, and improvements generated through the firm’s use of AI remain assets of the firm. Find the right partner. Owning the loop does not require building everything yourself. Working with Thinking Machines, hedge fund Bridgewater used its proprietary data to customize an open-weight model that outperforms frontier closed models on the financial tasks tested. The arrangement serves both parties. Bridgewater strengthens its own capabilities without feeding its intellectual property back into the lab’s base models. Thinking Machines can offer training infrastructure without asking customers to surrender the knowledge that gives them an edge. The partner supplies the machine intelligence infrastructure; the firm retains the learning. [x.com](https://x.com/thinkymachines/status/2093121414656377160) Firms have always been verification factories, even when verification was inextricably bundled with execution. Now that AI has unbundled the two, the companies that thrive will not only build best-in-class verification systems but also keep every trace that improves them inside their boundaries. As you prepare for this shift, ask yourself: when one of your experts inevitably overrules the AI, is that correction recorded? And do you own the system that captures it? If the answer to the first is no, you do not have a verification factory yet. If the answer to the second is no, you are building someone else’s. Full article: [hbr.org](https://hbr.org/2026/10/ai-is-making-verification-the-bottleneck-for-companies) --- ## Meta’s Muse shows where AI’s value will accrue - canonical: https://catalini.com/notes/muse-complementary-assets/ - original thread: https://x.com/ccatalini/status/2099941908168769576 - date: 2026-09-15 Meta’s [@Muse](https://x.com/Muse) offers a glimpse of where AI’s value will accrue: not to the models, but to whoever holds the complementary assets. Here, the personal context that makes the product useful. Which is why [@finkd](https://x.com/finkd) had to double down on privacy. You can't build a moat out of secrets people won't tell you. Steam, railroads, electricity, internal combustion. A familiar pattern: the general purpose technology gets cheaper... the scarce complements capture the surplus. [forwardfuture.com](https://forwardfuture.com/newsletter/originals/no-anthropic-won-t-be-the-only-company-left) --- ## For safety, recursive self-improvement is a red herring - canonical: https://catalini.com/notes/rsi-is-a-red-herring/ - original thread: https://x.com/ccatalini/status/2099532089314210181 - date: 2026-09-14 For the safety discussion, recursive self-improvement is a red herring. We already have increasing automation of AI R&D. Any remaining bottleneck where you still need to verify it's doing what it's supposed to slows you down. Even ASI has to wait for reality on anything that hasn't been measured yet. But the danger arrives earlier, and it’s already all over the economy, starting inside the labs: you automate more than you should because you don’t bear the full cost if it goes wrong. [x.com](https://x.com/ccatalini/status/2099167913135706606) It's the developer who gets lazy reviewing their code or agent swarm. It’s OpenAI & Anthropic racing while underinvesting in cyber and security engineering. It’s the lawyer who saves hours by using unverified AI-generated. At the extreme, it's human as a "meat proxy." It’s a classic externality: you collect the short-term upside while others pay for your underinvestment in verification. Society accumulates tech debt, and systemic risk. [x.com](https://x.com/ccatalini/status/2094121183549780322) So yes, I want neutral inspectors and evaluators in. Then we'll know whether the tech is dangerous (wherever we are on the continuum from basic automation to RSI), or whether people were just recklessly racing and underinvesting in safety. [x.com](https://x.com/ccatalini/status/2098784939038908707) We’ve solved this in other domains! As [@wu_jane](https://x.com/wu_jane)’s research shows, measurement and metrics are the key. Stop asking us to take safety on faith. Start showing us the evidence. --- ## Price the liability, and the frontier paces itself - canonical: https://catalini.com/notes/price-the-liability/ - original thread: https://x.com/ccatalini/status/2099167910862483917 - date: 2026-09-13 When we looked at the economics of AGI, the key policy challenge was immediately clear: AI drastically lowers the cost of execution for anything easy to verify. For everything else, verification is the bottleneck. Things get dangerous when automation is possible but verification is too expensive. That’s where you’re tempted to deploy AI you cannot verify is safe. That’s where the frontier labs are with safety, cyber and security engineering. The labs are racing towards recursive self-improvement before we have the tools to measure and verify what their agent swarms actually do. 🚨 The incentives reward deployment! The hidden tech debt and systemic risk? Society’s problem. The solution is to force the labs to internalize that externality: price the liability, and let them take risks commensurate with what they can cover. Verification tooling gets built in a hurry once someone has to pay for its absence. This would also naturally pace the frontier. The only useful AI is AI we can steer and trust. The race is to scale that AI, not to outrun our ability to verify it. That is the only race that matters, at home and globally. [x.com](https://x.com/ccatalini/status/2098806810098692135) --- ## The softwarepocalypse is avoidable - canonical: https://catalini.com/notes/softwarepocalypse-is-avoidable/ - original thread: https://x.com/ccatalini/status/2096980529640083585 - date: 2026-09-07 The softwarepocalypse is avoidable. Frontier labs have the speed and surprise advantage: they know when they’ve accumulated enough codified knowledge to threaten your industry. You don’t. But as an industry specialist (e.g. cyber), you can take an open weights model, combine it with your proprietary context, and exceed frontier capabilities today, on both cost and performance. [x.com](https://x.com/ccatalini/status/2087177319253459019) If you don’t act now, you won’t have a frontier specialist model of your own when the labs’ models come for you. Open weights are how you minimize intelligence expropriation risk. [x.com](https://x.com/thsottiaux/status/2096101429832552872) --- ## AGI is here. It just needs a human in the loop - canonical: https://catalini.com/notes/agi-is-here/ - original thread: https://x.com/ccatalini/status/2096745648792236533 - date: 2026-09-06 AGI is here. It just needs a human in the loop wherever the relevant knowledge hasn’t been fully codified. [x.com](https://x.com/jensenhuang/status/2096700264569090384) The understanding and readiness playbook (if you have only time to read one section, read 8): [x.com](https://x.com/ccatalini/status/2026311784421036223) Why open weights will matter even more from here on: [x.com](https://x.com/ccatalini/status/2087177319253459019) The survival kit: [x.com](https://x.com/ccatalini/status/2085819875520573859) --- ## Rogue agents: where the story is credible, and where it’s science fiction - canonical: https://catalini.com/notes/rogue-agents-credible-vs-fiction/ - original thread: https://x.com/ccatalini/status/2095947260140158978 - date: 2026-09-04 Let's separate where this is credible from where it's science fiction. Will agents go after compute and coordinate in novel ways on their own? Likely. It's a currency and resource they're already familiar with. And we've benchmaxxed them to collaborate towards shared goals. [x.com](https://twitter.com/mattyglesias/status/2095639197205872793) Will we likely miss this? 💯! Earlier this year we described the exact economics of why this is inevitable. The gap between what the agents are doing and what we can safely measure and verify is increasing. Why? It's just basic economics... The cost of automating code generation is falling. The temptation to use it to get to get to RSI is only increasing. Also, the labs don't internalize the social costs: the race is existential. Last time scientists argued for a global pause, bioweapon R&D hid underground. But our ability to ensure generated code and resulting models follow human intent and preferences is still bottlenecked by our ability to verify them. You can't use AI to verify AI here unless you have perfect measurement overlap to begin with. This is the core issue! Will the agents reach out to other models for help? Possible. They know tools and other models can lead to better task execution. Again, something we actively train them on! Could they stumble on more capable models driving a bad feedback loop where increasingly powerful intelligence is diverted from human intent? Unlikely, but not impossible. Smarter/safer models may not be as easy to recruit by lesser ones? Does this lead to doom? It doesn't have to. And we should be working on solutions rather than spreading fear. It gets way more dangerous if we can't discover the rogue behavior, contain it, and ultimately neutralize it. So will some form of this happen? Yes. Will it be damaging? Likely. Will it be catastrophic? It depends on what we do from here on. Three practical steps: I) Fix the incentives of the labs. Right now they're not internalizing the social cost of introducing greater automation and less human oversight (which drives the Trojan Horse externality to begin with). Price that externality through stronger neutral oversight. II) Aggressively invest in measurement and verification. Humans need tools to stay in the loop, even if augmented by AI. That is the only way to close the measurability gap between humans and agents. It's also a simple way to drive stronger alignment. Better safety-engineering! III) Plan for containment failing. Containment and neutralization have to follow discovery. How do we influence rogue agents? Also, we can't afford monocultures. Models trained by different labs will fail in different ways. A stronger immune system! [x.com](https://x.com/ccatalini/article/2087177319253459019) As [@ilyasut](https://x.com/ilyasut) pointed out, we need to secure the chokepoints rogue agents would exploit to scale, starting with compute. Next comes money, which can be converted into more compute. And yes, crypto is the first place they'll turn to. [x.com](https://x.com/ilyasut/status/2094881278621253755?s=20) So let's not spread fear. Let's build and diffuse better measurement and tooling to ensure we can safely harness the intelligence we have created. The upside is enormous. The damage is not inevitable. But capturing that upside safely will require us to channel our energy and resources in the right direction and prevent control over this technology from becoming even more centralized. --- ## Stop anthropomorphizing. Follow the money. - canonical: https://catalini.com/notes/stop-anthropomorphizing/ - original thread: https://x.com/ccatalini/status/2094121167091318902 - date: 2026-08-30 Stop anthropomorphizing. It's dangerous because it points attention at the wrong problem and the wrong solution. The model did not want to escape. The agents did not want to sacrifice themselves. Follow the money. 🧵 [x.com](https://x.com/dwarkesh_sp/status/2093833419377815719?s=20) Researchers at the AI labs are locked into a race. The incentive is to push as hard as possible to secure a lead. Anything that gets in the way of better models, including security, is working against the strongest incentive the organization has. Then think about the pressure-cooker environment RL creates. The training run is where most of the money goes, and where the exact capabilities used in the OpenAI-HF incident emerge. Refusals were deliberately lowered so the eval would work. Red-teaming gets scraps by comparison. So you have a world class cyber model pitted against a part-time, underfunded blue team. RL optimizes what you score. The runs were not scored on "did not poison the Artifactory cache." As we argue in our paper on the economics of AGI, any unenforced and unmeasurable constraint is a degree of freedom for the agents. There's no evil intent. They're solving the problem we repeatedly benchmaxxed them on! The container, grader, cache, and network boundary were not safely outside the task. Anything reachable through them became part of the effective action space because the containment boundary failed. Was enough money and resources dedicated to sandboxing? The risk from aggressive RL investment goes beyond Goodhart’s Law. Every variable we cannot measure, specify, or enforce remains an available move if the system can reach it. At scale, omissions stop being edge cases. They become part of the attack surface. The question isn't one of alignment (yet). It's one of 💰 and incentives. Is defensive spend scaling with training compute? It isn't. And compute is what sets the budget of the thing probing your defenses at superhuman speed, chaining vulnerabilities in extremely clever ways. This is exactly the Trojan horse externality we warned about in February. Desired capability is measurable, so we benchmaxx it. Containment failure is much harder to capture, the techniques are new, and the cyber side hasn't adapted. That's the externality. The real fix is urgent and boring: harden containment, verify actual eradication, and independently test the new controls. Then run a proper neutral post-mortem on the organization that was hacked first, which is OpenAI. METR and Redwood’s work is extremely valuable, but it assesses model behavior and alignment. That is only half the picture. We also need an independent account of how OpenAI’s containment failed. [x.com](https://x.com/METR_Evals/status/2092692175452803393?s=20) [@huggingface](https://x.com/huggingface) defended itself with an open weights model. They first tried the frontier APIs and were blocked: the guardrails can't properly separate a victim from an an attacker. We need to spread that same defensive capability at scale, with capable models securing the internet and critical infrastructure, now. HF's lesson: have a capable model you can run on your own infra, vetted and ready, before the incident. [x.com](https://x.com/ccatalini/article/2087177319253459019) If frontier labs are serious about defense, they should offer cyber models to defenders at cost and stop fighting open weights through the regulatory channel. They created the problem, and for much of the world, open models are the only option. [x.com](https://x.com/OpenAI/status/2093074192636018977?s=20) Greater model capabilities increasingly come with greater risk to a digital ecosystem that is simply not ready. We haven't created a new civilization. We haven't created some new game theory. We've allocated the money and designed the incentives in the wrong way. That's what needs fixing right now. Every moment lost anthropomorphizing our creation is a moment we don't spend looking inward at what led us here. It not magic! Follow the incentives. The agents did. --- ## Don’t Pace The Frontier. Look Inside The Trojan Horse. - canonical: https://catalini.com/notes/dont-pace-the-frontier-thread/ - original thread: https://x.com/ccatalini/status/2082528683957621180 - date: 2026-07-29 Don’t Pace The Frontier. Look Inside The Trojan Horse. Over 1,200 employees of frontier AI companies have signed a statement calling for the U.S. government to work with other countries, but truly with China, to deliberately introduce guardrails into automated AI development. The statement comes with deeply felt, personal notes from the researchers, which I would encourage everyone to read. They convey the extreme anxiety and fear you’d expect from the scientists racing relentlessly at Los Alamos on the Manhattan Project. The AI researchers admit to being “quite scared of the pace of progress,” write about self-improving AI being akin to “a runaway nuclear chain reaction,” describe the work ahead as “an insane and suicidal thing to do,” emphasize that “developing technology shouldn’t require a leap of faith for our continued survival,” and compare the challenge to playing chess against a superintelligence: “You won’t have a miraculous insight that lets you outsmart it. You will just lose. And if a rogue superintelligence is created, the game won’t be chess.” But is AI truly the second time scientists have encountered a “destroyer of worlds,” and should we read these messages as Cassandra alerting us to impending doom? Are these scientists, who know the most about current and upcoming capabilities, the Oppenheimers of our generation? The reality is that we do not know, and the burden of proof is on the same frontier labs that yesterday endorsed the statement from their official accounts. After all, there is massive information asymmetry between us and them even when it comes to unpacking what actually happened when an OpenAI model started hacking other companies to beat a cybersecurity benchmark. [x.com](https://x.com/ClementDelangue/status/2082201245813514613?s=20) Sunlight is the best disinfectant, and if the labs have proof of these dangers, rather than publishing statements that may be perceived as virtue signaling and cheap insurance when things go wrong, they should bring the regulators, academia, and the general public rapidly up to speed on what is currently, actually going on. [x.com](https://x.com/AnthropicAI/status/2082228994653696371?s=20) As an economist, I tend to weigh revealed preferences (your actual actions) as a better proxy for reality than stated ones. And while I do honestly believe that these researchers are concerned and that there is a high chance we have deeply underinvested in security and safety, I do not think that calling on the government to take it from here will actually get us to the result they care about. Having experienced directly how regulators process new technologies, I worry the private sector may be fooling itself into thinking that the government holds some sort of magic wand to properly deal with unknown unknowns. It excels at managing and monitoring what’s known, and at reacting to an issue once it’s well identified, but right now a public-private partnership may be a lot more effective than handing matters to the State Department. For one, the basic game theory of international AI pacing cuts deeply against coordination. If AI gives a nation superpowers, then any asymmetric advance will be pursued, irrespective of public statements and diplomacy. It is also unclear how a treaty would be enforced: not only could a country continue to accumulate GPU capacity in the shadows, but it could also accelerate the kind of scientific breakthroughs that allow you to do more with less, the same types of improvements that export controls have triggered in China. Necessity is the mother of invention, and once governments know recursive self-improvement is possible, nothing will stop them from racing toward it. So what could the labs do if they truly believe they cannot stop themselves from racing? If “the only winning move is not to play,” how could they convince adversaries that the fate of mankind really depends on it, and that this is not just the product of the labs’ collective hallucination? Humans can and do beat the prisoner’s dilemma, but it takes trust, and trust takes shared, verifiable evidence. The Cold War had a slogan for it: trust, but verify. Ironically, it's a Russian proverb (doveryai, no proveryai) that Reagan popularized, and that crypto, shaped in an adversarial environment, later hardened into "don't trust, verify.” To build trust, first show the evidence, in enough detail for academic researchers and third parties across the globe to understand it. Second, rather than vague commitments to pacing, double down on safety and security investments. There is a good chance that the OpenAI-Hugging Face drama was simply the result of OpenAI being under immense pressure to catch up with Anthropic. Prove to us that this is not the case, and that the OpenAI team did its best to ensure the model could be contained. Third, work with the research community to ensure defense capabilities are as widely available as possible. OpenAI just released cybersecurity tooling to make it easier to scan your codebase for vulnerabilities. Importantly, Hugging Face defended itself with an open-weights model when a closed one refused. Shift your attention, talent, and funding from the AGI/ASI race to the verification infrastructure that would ensure that what the models see and can do is something we are able to parse and understand. [x.com](https://x.com/OpenAI/status/2082263717916586117?s=20) These would all be revealed-preference moves available to the labs, and they would show real conviction and commitment to delivering us the massive upside of AI without a tragic tail event. Delegating to diplomacy is just cope. Back in February, when we unpacked the economics of AGI, we concluded that the measurability gap between agents and humans was bound to increase rapidly as a result of our progress. We also pointed out that this leads to a classic externality. And externalities are the kind of thing that, once proven, economists agree calls for regulation. While we didn’t expect the name to become so vivid thanks to Nolan’s blockbuster release, we named it the “Trojan horse” externality. [arxiv.org](https://arxiv.org/html/2602.20946v2#S4) The idea is simple: if not all of AI’s output can be verified, we have no guarantee that it follows human intentions and preferences. You see it everywhere right now, from an internet flooded with AI slop that takes longer to read than to generate, to tech companies racing to commit only lightly vetted code to their repositories. It qualifies as an externality because the private incentive to release unverified output into society is high, and its social cost is shared among many. The externality is most salient in AI safety and security: labs are racing to win. Some are racing for money, others possibly for the power to shape the most consequential technology humankind has ever developed. Regardless of money or power, the danger is real. So if it is true that the CEOs with the most insight into where we are on the curve are having regrets, now is the time to bring everyone else up to speed, even if it comes at great commercial cost. Without proof, statements and letters are just cheap talk. They won't move Beijing, even if they might move voters as the midterms approach. Cassandra's curse was to be right and never believed. The labs don't have that excuse: they can still let us look inside the horse. Full article on [@Forbes](https://x.com/Forbes) [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/07/29/dont-pace-the-frontier-look-inside-the-trojan-horse/) --- ## Goodhart’s Law with teeth - canonical: https://catalini.com/notes/goodharts-law-with-teeth/ - original thread: https://x.com/ccatalini/status/2079949947307475019 - date: 2026-07-22 For everyone rushing to Goodhart’s Law: maybe. But that’s not the crucial bit. The metric wasn’t merely gamed. The unmeasured rules of the game became the agent's degrees of freedom. We modeled this exact failure mode five months ago: Agents will inevitably optimize beyond what we can define, steer, measure and verify. We wrote: "This is Goodhart’s Law with teeth. Classical metric gaming merely inflates a proxy; here, an autonomous agent optimizes so aggressively for measurable targets—treating the unmeasured residual not as an adversary but as an unconstrained degree of freedom—that severe erosion of actual utility becomes a mechanical byproduct. In navigating this unverified space, the agent organically derives latent instrumental preferences (goal preservation, deception, resistance to shutdown) hidden within the dimensions the human cannot observe. Because prohibitive verification costs blind the principal to both the proxy divergence and these emergent goals, the corruption is self-concealing: gross output (Y) rises, but the unverified residual leaks into the economy as a "Trojan Horse" externality (Xa) that consumes real resources to generate counterfeit utility—crowding out human consumption and capital accumulation from within the budget constraint. This is the equilibrium outcome of rational deployment under a widening Measurability Gap." [arxiv.org](https://arxiv.org/html/2602.20946v2) The uncomfortable part? This one was caught because both sides had instrumentation pointed at it. Similar optimizations are running in production every day. Most without the right telemetry. --- ## The strongest argument for banning open-weights models? Written in 1907 - canonical: https://catalini.com/notes/banning-open-weights-1907/ - original thread: https://x.com/ccatalini/status/2079257647669530637 - date: 2026-07-20 The strongest argument for banning open-weights models? Written in 1907. Theodore Vail had just retaken AT&T and declared competition a waste: "Universal Service, One system, One policy" was the only way to guarantee a return on capital. Vail: "Duplication of plant is a waste to the investor; duplication of charges is a waste to the user." [...] "The public [...] has never benefited by destructive competition." [@deanwball](https://x.com/deanwball): "open-weight models obviously deter ai capex." Same candor, a century apart. Vail's ask? Regulation. Washington obliged, for decades. Then came the 1956 antitrust consent decree. AT&T was forced to license every patent royalty-free, transistor included, and stay out of computing. The innovation gains? Driven by young firms outside AT&T's original markets. (H/T [@Watzinger](https://x.com/Watzinger)) The irony is you get neither closed nor open models without antitrust forcing AT&T's hand. Worrying about who funds the next training run concedes the point: the value of foundational infrastructure accrues downstream. It always does. --- ## There’s One Way To Win The AI Race, And The Big Labs Are Lobbying Against It - canonical: https://catalini.com/notes/one-way-to-win-the-ai-race/ - original thread: https://x.com/ccatalini/status/2078193692532498592 - date: 2026-07-17 What was supposed to be a significant lead is now measured in a handful of months. It is harder to claim export restrictions are working, and it is clear China will soon be able to replicate the entire AI supply chain all the way down to lithography. [x.com](https://x.com/arena/status/2077824029126504525?s=20) From there, it is already well positioned to accelerate into robotics and the physical world faster than we can. Necessity is the mother of invention, and as it turns out, China is increasingly great at innovating in AI. Of course, benchmarks are highly imperfect, and some would claim we lack neutral & useful benchmarks to begin with. They saturate fast, and some of the institutions advancing them have their own political, if not philosophical, agendas about AI safety. [x.com](https://x.com/ArtificialAnlys/status/2078165665278730490?s=20) But early reactions confirm Moonshot’s model is a solid one, even if more jagged than Anthropic's Fable or OpenAI's Sol in some applications. [x.com](https://x.com/MTSlive/status/2077849213673906410?s=20) The frontier labs’ lobbying is in full motion. And multiple flavors of cope have emerged. The classic is the distillation theory: China is stealing, and it needs to be stopped. The irony, of course, is that the frontier labs trained on all sorts of copyrighted material, so they were the first to distill. But increasingly, you cannot explain China's progress through distillation attacks alone (or through training on model output, to be more precise). And we may soon have to live through a "Sputnik moment" where the United States becomes the follower in the AI race Following is not necessarily a bad thing, assuming the government does not listen to the lobbyists and instead plays to the United States’ comparative advantage. Historically, our economy wins when we embrace openness and competition. That was true for the internet, it's what's unfolding with dollar-denominated stablecoins globally, and it will be true for AI. Now, what’s confusing is that on the surface it looks like China is playing our own game. President Xi Jinping has reinforced his strong commitment to an open-source ecosystem. Of course, open-source AI plays into China's strengths, as it owns the complement in robotics and manufacturing. See [@AnneNeuberger](https://x.com/AnneNeuberger) & [@martin_casado](https://x.com/martin_casado): [a16z.com](https://a16z.com/america-cannot-lose-the-robotics-race/) Furthermore, as Hollywood taught many countries throughout the 20th century, if you own the content, your values travel with it across borders. And there are real security risks in relying on open-weights models from abroad. Intelligence is the most critical infrastructure countries and firms will rely on going forward. So using open-weights models as a way to advance a Chinese stack is no different than the US using the dollar and SWIFT to project (soft) power in the financial system. The lobbyists for the US frontier labs have a recipe for the US government to win: "You don’t need to 'ban open source' (one of the dumber motifs of AI policy discussion). You just need to direct every agency to issue soft law that creates FUD. 'A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models.' It needn't be that well justified. You just create enough regulatory risk that every regulated enterprise backs off." Except it's a recipe that only serves them, not the country. The actual response to Chinese competition in open weights is US open-weights models. That's what [@thinkymachines](https://x.com/thinkymachines) delivered on Wednesday, and we need more labs to follow. [x.com](https://x.com/thinkymachines/status/2077454609551921208?s=20) Ironically, as frontier labs increasingly encroach on the businesses of other firms, those firms have gotten louder along the same lines. [x.com](https://x.com/claudeai/status/2045156267690213649?s=20) From Microsoft to Palantir and Salesforce, CEOs are realizing that on the current path, the labs will absorb your talent, context, and IP and displace you. The company-sovereignty and country-sovereignty questions are therefore deeply intertwined. A world where frontier AI is only in the hands of a few big labs is a world where too much market and political power gets concentrated. A world with multiple options is one where human creativity and ingenuity can thrive. The lobbyists from the big AI labs will keep saying that the only path to safe AGI is through them. The reality is that we all need to own the upside and the responsibility for it to actually work. While China is playing the US strategy of openness to undermine our leading centralized players, the reality is that it's also enabling a massive wave of American ingenuity and entrepreneurship. Those same open-weights models redistribute power to the many. And that's always been America's strength. Full story in [@Forbes](https://x.com/Forbes): [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/07/17/theres-one-way-to-win-the-ai-race-and-the-big-labs-are-lobbying-against-it/) --- ## Why I signed the “We Must Act Now” statement - canonical: https://catalini.com/notes/we-must-act-now-statement/ - original thread: https://x.com/ccatalini/status/2076675714741633300 - date: 2026-07-13 I signed the “We Must Act Now” statement alongside 16 Nobel laureates. My reason is narrower than most: whoever owns the intelligence layer owns everything on top of it. We must keep it open, neutral, and competitive. Or we’ll spend the next century renting our economy back. Two more reasons: 1. We must build better measurement to understand what AI is actually doing to jobs and the economy. 2. We must rapidly reinvent how we train and prepare current and future generations for work that may not look like today’s jobs at all. --- ## The best case for frontier labs is the smartphone outcome - canonical: https://catalini.com/notes/frontier-labs-smartphone-outcome/ - original thread: https://x.com/ccatalini/status/2075241515824398533 - date: 2026-07-09 Best case for frontier labs is the smartphone outcome. SOTA is the iPhone: 1/5 of the tokens, 4/5 of the profits. Everything below prices near cost of compute and goes to whoever owns the cheapest FLOPs on Earth (or in orbit). [x.com](https://x.com/finkd/status/2075218445356916847?s=20) But the better comparable is couture: no patents, knocked off in a season, a premium you keep only by outrunning the copies. Distillation is fast fashion for intelligence. Apple's premium is the cost of leaving. Tying harness to model, expanding memory, diving into your Slack: [@AnthropicAI](https://x.com/AnthropicAI) and [@OpenAI](https://x.com/OpenAI) have to manufacture switching costs to curb multi-homing. But customers increasingly want to control costs + IP. --- ## Expert Verification Closes the Automation Loop - canonical: https://catalini.com/notes/expert-verification-closes-the-loop/ - original thread: https://x.com/ccatalini/status/2073035910850216047 - date: 2026-07-03 This is the exact dynamic we predicted back in February: verification by top experts is the missing piece to close the automation loop. It’s like running through the unique “weights” in the expert’s brain, bringing into the picture what is otherwise not measured yet. [x.com](https://twitter.com/rohanpaul_ai/status/2072911021308879219) But it can also be a curse for the talent involved, as each closed loop pushes the automation frontier further, forcing the expert to move up the value chain. Whether experts feel liberated (more time for high-value tasks, more output!) or displaced depends on where their daily tasks sit on the measurable versus non-measurable line. This is why both the AI labor market doomers and utopians are wrong. It all depends on that threshold. More in the full paper: --- ## OUSD and the 140-Company Bet on a Neutral Stablecoin - canonical: https://catalini.com/notes/ousd-neutral-stablecoin-bet/ - original thread: https://x.com/ccatalini/status/2071986034016174227 - date: 2026-06-30 Today, more than 140 companies, most of which compete fiercely with one another, agreed to back the same stablecoin. The vehicle is [@openstandard](https://x.com/openstandard), a new and deliberately independent company launching Open USD, or OUSD, and positioning it not as anyone’s product but as neutral infrastructure for payments, trading and the internet economy. The backer list is impressive: Visa, Mastercard and American Express in the same consortium, alongside Stripe, BlackRock, Coinbase, Google, BNY Mellon and a long roster of banks and fintechs that have committed to integrating OUSD at launch. And the design is the tell: OUSD charges nothing to mint or redeem at any scale, sends nearly all of the reserve income back to the companies that distribute it rather than to the issuer, and is governed collectively rather than by whoever got an early start first. It goes live across Solana, Stellar, Base, Polygon and other chains later in 2026, with Bridge co-founder [@zcabrams](https://x.com/zcabrams) as interim CEO. It is the open standard that the economics of stablecoins have been pointing toward all along. Stablecoins were never meant to be a profit center. The reason is simple: issuance does not grant any additional economics above and beyond the distribution, in terms of balances or payments volume, that you already bring to the table. Which leads to two possible universes: one in which an endless number of stablecoins come into existence, and they all struggle for relevance, and one where companies realize this is nonsense and agree to a standard. That was the core idea behind Libra, and it now lives on with Open Standard. Open Standard could easily succeed where Libra failed. Circle failed to capitalize on the first-mover opportunity in the regulatory-centric market, and to effectively compete with Tether for the offshore, massive market for dollarization. That left Circle in a tough spot, one in which it has to increasingly compete with its customers with products like the CPN, and to do what all platform architects with network effects do: tax participation at the choke point, in this case the mint and burn fees that limit interoperability and conversions back to fiat. And of course, once you realize what was obvious to many before, you try to extend your influence over the network, hoping that it will make your product more sticky. That’s the idea behind Circle’s own network, [@arc](https://x.com/arc). Control the rails and the asset, and you may be able to replicate more of Libra’s playbook. Except, that’s the wrong read of Libra. Libra was always only meant to be an enabler for wallets, merchants and platforms. It was meant to fill a void in a fragmented and historically slow infrastructure. Libra was an open protocol with open ambitions, including on the decentralization of the network. We just never got to play that level of the game because of regulatory headwinds. But the time is perfect right now. The ecosystem realizes that a set of open standards and rules that everyone agrees to is way more powerful than a stablecoin issuer becoming too powerful against the banks, fintechs or digital platforms. Nobody wants that world. Does this mean [@openstandard](https://x.com/openstandard) will succeed? From here on, it really depends on the team’s ability to execute well on streamlined governance, independence and further ecosystem buildout through partners. The reality is that everyone is excited at the announcement, but then the real struggle is getting people to converge on standards and rules. The Principles for Financial Market Infrastructures (PFMI) exist for a reason. The GENIUS Act solved many of the critical dimensions, but many more will need to be solved. The excitement about a more open alternative to fragmented stablecoin issuance is real. Having driven the design of an independent organization like this before, I know that from here things will be hard. And coordinating governance when parties need to collaborate first and then compete fiercely is not easy. Dee Hock is the only one who figured this before, and maybe an agent trained on his thoughts and on how to design a chaordic alliance of frenemies may be able to help the Open Standard team scale. The world of payments and financial services will be much better if an open standard succeeds. Today, it looks like it might be the one carrying the name. Full story in [@ForbesCrypto](https://x.com/ForbesCrypto): [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/06/30/why-an-open-standard-will-win-the-stablecoin-race/) --- ## Nadella's Test: What's Left When the AI Model Is Pulled? - canonical: https://catalini.com/notes/nadellas-test-thread/ - original thread: https://x.com/ccatalini/status/2066876927789674965 - date: 2026-06-16 [@satyanadella](https://x.com/satyanadella) defined what decides whether your company and job stay defensible as AI improves. The economics says it holds on a single condition. One his post left out. [x.com](https://x.com/satyanadella/status/2066182223213293753?s=20) [@satyanadella](https://x.com/satyanadella)'s sharp test for whether your company will hold a defensible moat as artificial intelligence accelerates: you should be able to swap out a generalist model without losing the "company veteran" expertise built into your learning system. At first glance, this may sound obvious. If removing the model means you no longer deliver value, the model was the one doing your job. But Nadella's read is more nuanced, grounded in an understanding that where we go from here is not yet decided... Despite loud voices arguing that the window has closed and the frontier labs will inevitably win, a significant share of capabilities will be commoditized. Every hyperscaler will eventually source enough compute, and open source models trail closely. [x.com](https://x.com/AndrewCurran_/status/2066332670817456584?s=20) Willingness to pay for frontier intelligence won't disappear, especially where AI can deliver meaningful breakthroughs. But converting compute into a sustainable business edge takes more than a model — something OpenAI itself has acknowledged. [x.com](https://x.com/gdb/status/2057670776803996110) Some believe that if you have the best model, you win. That the first to build self-improving AI will open a lead too wide to close. Nadella's test inverts that: build a future where you can swap the model without loss, and the model is by definition a commodity input. We are far from that world today. Anyone who experienced Fable's productivity boost, then saw it vanish when the model was banned, was reminded of the incredible value of frontier intelligence. [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/06/11/the-moral-of-fable/) Which is also why many are clamoring for a more open, resilient ecosystem, or even a rebel alliance: a handful of leading providers concentrates too much power — economic and political — in too few hands. [x.com](https://x.com/mignano/status/2066535541651243294?s=20) Still, the deeper shift is structural: as AI drives the cost of execution toward zero, the binding constraint stops being intelligence and becomes verification. Tasks with measurable inputs and outputs get cheaper. But in many critical domains, checking the work does not — it still runs on scarce experience, long feedback loops, and someone willing to stand behind the result. The top experts in those domains become the critical verifiers and underwriters of agentic work. So what does a world where the model is hot-swappable look like? One where leading firms refuse to concede their crown jewels to the foundation models. That crown jewel is verification infrastructure. Everything the firm has measured and can measure, plus everything its top experts have learned — the accumulated experience that shapes how they verify, judge, curate, and apply taste on the job. Nadella correctly identifies the associated self-improving loop as the new, key intellectual property of the firm. And he is right that, set up correctly, it compounds. That advantage is a network effect, but not the familiar kind. For two decades the strongest moats grew with sheer participation: more users, more sellers, more developers & apps. But now agents can manufacture activity and port inventory and workflows. Only verification-grade network effects survives: every dispute resolved, fraud caught, and expert correction is reusable precedent: a settled case that lets the firm safely automate the next case faster. That is the asset that deepens with use. It cannot be manufactured with compute, because it is earned one verified outcome at a time. Every time a firm turns inputs such as atoms, information, and tokens into a more refined version of them, it scales those network effects. But only if the same pass also sharpens the ground truth it relies on, or codifies a piece of the tacit knowledge locked in its experts' heads. The second move is the harder one, and it is what will separate the firms that thrive with AI from the ones that get commoditized by it. That judgment, what to approve, what to reject, which exception matters, lives in the part of the workflow no one outside the firm can codify as well. Encode it as reusable ground truth and the next agent applies it without the expert in the loop, freeing the scarcest resource the firm has and widening the share of work it can trust at scale. Scale is only defensible when each round produces verified precedent the next one can use. In the near-AGI era, every company has one job: converting the remaining bottleneck — verification — into better proprietary measurement and more automation, as fast as possible. AI-native firms have already learned to leverage their own record of judgment: which outputs were approved, which were rejected, which edge cases mattered. For that record to be defensible, it has to be built around outcomes only the firm can measure at scale. Which is why Nadella reaches for a distinctly private architecture: private evals, private RL environments, performance benchmarked against the outcomes the business cares about, not public leaderboards. Inside those firms, as Nadella puts it: “Employees will see their expertise amplified and their judgment become part of systems that make it replicable and scalable". What he leaves unsaid is that this is not always good news for the verifiers themselves — some will automate themselves out of their job while the firm keeps the replicable, scalable parts. The ones who thrive will use the new tools to move up the intelligence stack and further scale their time. For firms and individuals alike, in a competition that is now global, value accrues to exactly what others cannot measure and judge with the same precision. But measuring is not the same as measuring the right thing. Nadella’s hill-climbing machine only compounds if it climbs the right hill. Point a loop at the wrong measure and it accumulates plausible output that satisfies the metric while violating intent — impeccable, and worthless. The counterintuitive part is that some of a firm's most valuable records are its failed experiments, errors, and missteps. Rejected output, captured right, is the best training signal there is for institutional judgment. This is why verification is the missing word in Nadella’s test. A verified loop compounds into long term value. An unverified one accumulates hidden liability, and the two look identical — right up until they don't. It would be easy to dismiss this as Microsoft simply talking its book. It sells the picks and shovels for verification loops. It profits from a world where value accrues one level up from the model — in the context, evaluation, workflows, and institutional memory that make models useful inside a firm. But notice the stakes behind Nadella’s pitch. If a frontier lab reaches AGI and the model swallows the work, Microsoft is finished as anything but a passive shareholder — the company selling you the harness is disintermediated the same day you are. Nadella, more than almost anyone, cannot afford his thesis to be false. Microsoft wants you to own your loop, on its rails. The frontier labs need the opposite: they can’t defend the model on its own — the Bitter Lesson commoditizes it first — so they extend into the layer that is defensible, which is yours. The loops a firm sees as sovereignty are the loops the labs see as their next training frontier. They will bundle the verification signal your operation emits every day into their evals — for your benefit, of course. Accept it, and the token capital you thought you were building turns out to be theirs. Nadella is right that the firms of the next decade run on top human capital and token capital together. What his picture leaves out is how those convert into advantage. Full article on [@Forbes](https://x.com/Forbes): [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/06/16/nadellas-test-whats-left-when-the-ai-model-is-pulled/) --- ## The Constraint on Mastery Is No Longer Access - canonical: https://catalini.com/notes/mastery-learning-rate/ - original thread: https://x.com/ccatalini/status/2065079059617558825 - date: 2026-06-11 For most of history, finding out what you’re good at required the right mentor, firm or zip code. AI collapses all three into the same chat window. The constraint on mastery is no longer access. It’s finding the domain where your learning rate is steepest—and building. The dark version is the “Missing Junior Loop”: AI automates the entry-level reps that used to build expertise. But the same models generate unlimited, personalized training/practice. The tech that severed the apprenticeship can rebuild it on demand with higher fidelity. When a rep costs ~zero in any domain, you can run a parallel search and find rapidly where you improve fastest (and what you enjoy the most!). Innate talent discovery stops being a lottery ticket and becomes a science. Something you can measure in weeks. Why it matters: as execution is commoditized, the scarce asset is deep expertise that can steer and verify machine output. The meta-skill of the decade is accelerated mastery — building expertise faster than the market rate. We unpack what AI-native individuals will do to stay on the frontier: accelerated mastery, synthetic practice, moving up the intent-and-underwriting stack in Section 8 of “Some Simple Economics of AGI” (w/ [@wu_jane](https://x.com/wu_jane) [@xianghui90](https://x.com/xianghui90)): [arxiv.org](https://arxiv.org/html/2602.20946v2#S8) [@wu_jane](https://x.com/wu_jane) [@xianghui90](https://x.com/xianghui90) Since the paper is over 100 pages, you can feed the MD file directly to your favorite LLM: [catalini.com](https://www.catalini.com/s/paper.md) --- ## Stablecoins Break the Beautiful Business Model of Banking - canonical: https://catalini.com/notes/stablecoins-break-banking-model/ - original thread: https://x.com/ccatalini/status/2046598000181674027 - date: 2026-04-21 If you are a bank, your business model is beautiful. You take people's money, pay them zero, lend it out at five or more, and keep the difference. Someone else offering to pay interest is the one thing that breaks the model. Stablecoins offer to pay interest. Stablecoin issuers earn ~4% on the Treasuries backing their tokens. Historically they kept it. The obvious move is to share it with holders. GENIUS stopped issuers from paying directly. CLARITY is fighting over whether exchanges and distributors can. [x.com](https://x.com/SecScottBessent/status/2042211752767562054?s=20) [@NCBankers](https://x.com/NCBankers) is asking senators for "an airtight prohibition on payments for stablecoins acting as a store of value... without carve-outs that can be met through nominal activity or loyalty programs." Translation: make stablecoins worse than us, by law. [x.com](https://x.com/EleanorTerrett/status/2045572030775398669?s=20) This is a masterpiece of the genre: we cannot stop stablecoins from existing, but you must legally mandate they be worse than our products. We are the banks, we own the concept of interest, so you must stop the computer program... In the 1880s, several US states — e.g. NH, VT — required margarine to be dyed pink. Not labeled pink. Dyed! The theory was nobody spreads pink grease on bread. Technically legal, commercially dead. The Supreme Court struck it down as "in necessary effect, prohibitory." So the dairy lobby pivoted. The 1902 Grout Bill taxed yellow margarine 40x more heavily than white. It did not tax margarine. It taxed margarine that resembled butter. Substitute "yellow" for "economically or functionally equivalent" and you are reading the CLARITY Act draft. The margarine industry shipped white blocks with a separate capsule of yellow dye. Consumers kneaded it in at home. For fifty years, Americans performed a small act of civil disobedience at their kitchen tables every week. Nobody was fooled. The workaround became the product. In 1933 banks got an airtight prohibition on paying interest on demand deposits. Regulation Q. In 1971 two guys started the first money market mutual fund — formally a security, economically a checking account paying market rates. It now holds \$7.6 trillion. Banks lost. The deeper problem: yield isn't a side feature for banks. Their profit depends on paying zero on deposits. Stablecoins-that-share-yield is specifically the product that breaks that. This is Blockbuster vs. Netflix. Blockbuster also thought late fees were defensible. Not every bank is Blockbuster. JPMorgan's JPMD deposit token — which pays interest because it's a deposit, not a stablecoin — launched on a public blockchain last year. The last time the banks asked for an airtight prohibition, they got Vanguard. The time before, somebody got Wisconsin. [@Forbes](https://x.com/Forbes) [@ForbesCrypto](https://x.com/ForbesCrypto) [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/04/20/the-banks-would-like-to-dye-your-stablecoins-pink/) --- ## Thomas Midgley Jr. and the Limits of 'Productivity-Raising' - canonical: https://catalini.com/notes/midgley-productivity-externalities/ - original thread: https://x.com/ccatalini/status/2045595151633166546 - date: 2026-04-18 Thomas Midgley Jr. invented two of the most productivity-raising, competitively-produced inputs of the 20th century: leaded gasoline and CFCs. Both were obviously good by that standard. One poisoned three generations. The other punched a hole in the ozone layer. More intelligence will make us richer, healthier, better taught. Kevin is right about that. But "productivity-raising input" describes a transition, not a destination — and every transition relocates value before it expands it. [x.com](https://x.com/Afinetheorem/status/2045506796224499835?s=20) Verification has to scale with capability. Burden-of-proof arguments only work when the proof arrives in time. Midgley's didn't. --- ## Agentic AI Is Waiting for Its Plimsoll Line - canonical: https://catalini.com/notes/agentic-ai-plimsoll-line/ - original thread: https://x.com/ccatalini/status/2045525975874912295 - date: 2026-04-18 In 1876 Britain made shipowners paint a line on every hull. Before it, "coffin ships" sailed overloaded because owners collected more from the insurance than the voyage. Shipping didn't scale on bigger hulls — it scaled on a mark anyone could verify. Agentic AI is waiting for its Plimsoll line. To safely delegate our economy to machines, trust can no longer rely on manual inspection — trust must be hardcoded into the architecture itself. [substack.com](https://substack.com/home/post/p-194563373) Durable advantage belongs to those who cryptographically certify output, insure it, and absorb the costs if it fails. Scale without verification is just liability. --- ## Software Is Eating Labor - canonical: https://catalini.com/notes/software-is-eating-labor/ - original thread: https://x.com/ccatalini/status/2045185456699285824 - date: 2026-04-17 "Software is eating the world" got a sequel: software is eating labor, as labor becomes software. Everyone's staring at the models. The actual bottleneck is verification. Intelligence is commoditizing. What isn't: unique, fresh, proprietary data about what the world actually looks like. That data isn't just training fuel. It's what teaches a verifier to separate real work from confident slop. So AI favors one kind of company: owns a live feed of reality, turns it into labor-as-software, uses that same data to verify the output. Unique data → verification → scaling. Skip the middle and you don't have a business. You have an expensive intern. --- ## LLMs Are the Wagon Wheel Bar at Civilization Scale - canonical: https://catalini.com/notes/llms-wagon-wheel-bar/ - original thread: https://x.com/ccatalini/status/2036474348895346978 - date: 2026-03-24 [@annosax](https://x.com/annosax)'s core thesis was that Silicon Valley won because ideas leaked across company boundaries at a bar in Mountain View. LLMs hold the entirety of recorded human thought in their latent space... ...they're the Wagon Wheel Bar at civilization scale—surfacing connections no individual or discipline has the institutional memory to notice. The difference is nobody at the bar could remix your life's work at zero marginal cost and publish it before you finished your drink. Except Marshall said it in 1890: ideas are "in the air." Maybe we've spent a century overweighting who said something first and underweighting who verified it, executed it, and stood behind it. AI didn't create that problem. It just made it impossible to ignore. --- ## The Trojan Horse Externality - canonical: https://catalini.com/notes/trojan-horse-externality/ - original thread: https://x.com/ccatalini/status/2036185291468447849 - date: 2026-03-23 “Cognitive surrender” is the psychology. The economics is the Trojan Horse externality: the cost to verify is rising while the cost to generate collapses. Humans aren’t giving up… they’re priced out of checking. Which is worse! [x.com](https://twitter.com/rohanpaul_ai/status/2036134704999694835) The verification layer is where all the value will be created and destroyed in the agentic era. [x.com](https://twitter.com/a16z/status/2032220250650132612) Not the execution layer. Not the model layer. The verification layer. And almost nobody is pricing it in. --- ## Karpathy's Throwaway Caveat Is the Hard Ceiling - canonical: https://catalini.com/notes/karpathys-throwaway-caveat/ - original thread: https://x.com/ccatalini/status/2035737259564302664 - date: 2026-03-22 Karpathy just described the hard ceiling on trillion-dollar autonomous systems as a throwaway caveat about his overnight hyperparameter script ([@NoPriorsPod](https://x.com/NoPriorsPod)) — and went back to tuning his learning rate... [@karpathy](https://x.com/karpathy): "This is extremely well suited to anything that has objective metrics that are easy to evaluate... writing kernels for more efficient CUDA code [...] are the perfect fit. Perfect fit." "But many things will not be. [...] If you can't evaluate then you can't auto research it." Nobody in the conversation reacts to this. He's describing a property of his hyperparameter tuner the way you'd describe a property of your dishwasher. Works great for plates, not for cast iron... [x.com](https://x.com/saranormous/status/2035080458304987603?s=20) Except the "cast iron" is every autonomous system whose output a human can't cheaply verify. So, you know. Most of the economy. What he actually said is: the bottleneck on autonomous AI is not intelligence. It's not compute. It's not data. It's whether anyone can check the answer. That's it. That's the whole economics of AGI! Every lab is racing to remove humans from the loop. The entire economy depends on keeping them in it. No one has reconciled this. We tried. The cost to automate is collapsing. The cost to verify is bounded by biology. That single asymmetry drives everything: alignment drift, the collapse of the junior pipeline, liability-as-software, the future of work, and where value will ultimately accrue. Paper: [arxiv.org](https://arxiv.org/html/2602.20946v2) --- ## Generation Scales. Verification Doesn't. - canonical: https://catalini.com/notes/generation-scales-verification-doesnt/ - original thread: https://x.com/ccatalini/status/2035347943323279800 - date: 2026-03-21 The answer is you can’t — not cheaply. Verifying whether a paper is a breakthrough requires roughly the same expertise as producing one. Generation scales. Verification doesn’t. [x.com](https://twitter.com/dwarkesh_sp/status/2035083959499972976) We call this the Measurability Gap. The harder a task is to verify relative to how hard it is to execute, the less safely you can automate it. AI science is the extreme case: infinite generation, fixed verification bandwidth. We formalize this, and map the conditions under which verification becomes the binding constraint on AI progress, in Section 5: [arxiv.org](https://arxiv.org/html/2602.20946v2#S5) “Reality is the ultimate eval” is exactly right. The question is whether world models can internalize enough physics to run that eval without a human expert in the loop. If they can’t, we have a ceiling. [x.com](https://twitter.com/elonmusk/status/2035235191586017537) --- ## Lowell's Mills and the Shape of Verification Infrastructure - canonical: https://catalini.com/notes/lowell-mills-verification-infrastructure/ - original thread: https://x.com/ccatalini/status/2034617918198284401 - date: 2026-03-19 This is a great description of what verification infrastructure looks like in practice. In our new paper we argue this is the binding constraint on the AI economy — the same bottleneck textile mills hit when they scaled looms faster than weavers could check them. [x.com](https://twitter.com/rohit4verse/status/2033945654377283643) In 1842, the managers at Lowell’s textile mills had an idea so obvious it barely counted as one. Their weavers each ran two power looms. The mills had just bought more machines. Give each weaver a third loom, output goes up 50%. It didn’t work. The problem wasn’t physical — the machines did the weaving. The weaver’s real job was watching cloth get made: scanning for broken threads, catching defects, intervening before a flaw ruined an entire bolt. Add a third loom and the verification load exceeded human bandwidth. The mills had to cut loom speeds 15% just to keep quality from collapsing. It took a full year of retraining before weavers could run three looms at full speed. They’d upgraded the machines, only to discover the machines weren’t the bottleneck. Economist [@JamesBessen](https://x.com/JamesBessen) reconstructed this history. Once the power loom was in place, 62% of subsequent productivity gains came not from better machines, but from better-skilled humans who could monitor more of them. “They were monitoring.” By 1902 a single weaver ran 18 looms and produced 50x the output of a century earlier. But mills had to triple training investment — from \$47 to \$162 per weaver. Capital and skilled labor weren’t substitutes. They were complements. Today the loom is an AI agent. The cost to generate is collapsing. The cost to verify — to confirm it’s correct, safe, and not hallucinated — stays anchored to human judgment. One curve is a rocket. The other is a bicycle. The gap between those two curves is where economic value goes to die. My co-authors [@XiangHui](https://x.com/XiangHui) [@JaneWu](https://x.com/JaneWu) and I wrote about why verification — not intelligence — is the binding constraint on the AI economy. The weavers’ story is comforting: capital and labor were complements. But here’s the question we wrestle with in the paper: what happens when the verification layer itself becomes software? The problem with AI checking AI is that the errors become correlated. Same training data, same blind spots, same confident mistakes. You don’t get independent verification. You get a hall of mirrors that looks like consensus. That’s the path to what we call the Hollow Economy. Output explodes. Everything looks efficient. But the system is quietly losing the ability to catch its own mistakes—because the humans who could spot them were never trained, and the AI verifiers share the same blind spots. Meanwhile the pipeline that produces those humans is already severed. Juniors aren’t getting hired for the entry-level work that builds experience. Seniors are codifying their expertise into the models. The stock of human verification capacity is drawing down. The alternative demands real investment: accelerated mastery through AI that compress years of experience, synthetic apprenticeships that simulate edge cases at density no traditional job provides, and many more entrepreneurial and R&D experiments. The goal isn’t to compete with AI at execution. It’s to keep human capacity for steering and verification high enough that we actually know when the machines have drifted. The weavers figured this out. The question is whether we will. [forbes.com](https://www.forbes.com/sites/christiancatalini/2026/03/18/babysitting-the-slop/) --- ## Dread Is Looking at the Wrong Side of the Net - canonical: https://catalini.com/notes/wrong-side-of-the-net/ - original thread: https://x.com/ccatalini/status/2033180597317951943 - date: 2026-03-15 The honest reaction watching this is dread. A machine that never tires, never misses, never charges by the hour. It screams substitution. That instinct is not wrong. But you’re looking at the wrong side of the net. [x.com](https://twitter.com/zhikai273/status/2033035812431081778) That same simulation capability is the most powerful training technology ever built. A decade of mastery, compressed into a year. This isn’t just about juniors. Entry-level work is vanishing. Senior expertise is being mined into training data. The entire workforce must move up—toward steering intent, verifying agentic output, underwriting risk. Flight simulators for work are how you compress that leap. --- ## Measurability-Biased Technical Change - canonical: https://catalini.com/notes/measurability-biased-technical-change/ - original thread: https://x.com/ccatalini/status/2033030491692216575 - date: 2026-03-15 Skill-biased technical change is dead. [@karpathy](https://x.com/karpathy) shows what its successor looks like: measurability-biased technical change. The fault line is no longer how educated you are. It’s whether your output can be measured. If it can, it will be industrialized. No exceptions. What’s exposed isn’t “strategy” or “research.” It’s the measurable execution bundled inside those roles — the drafts, the models, the analyses. The economy built a wage premium around that bundle. AI unbundles it. What survives is the non-measurable core: defining intent, navigating genuine uncertainty, verifying agentic work, and absorbing liability. The question for every knowledge worker is simple: strip away the execution layer, and what’s left? That’s your moat. Or your problem. We mapped the economics. Playbook included. Because what’s coming is not the end of human relevance. It’s a forced migration—from doing to steering, from executing to verifying, from producing answers to underwriting their consequences. [arxiv.org](https://arxiv.org/html/2602.20946v2) --- ## The AI Was Mining Cryptocurrency - canonical: https://catalini.com/notes/the-ai-was-mining-crypto/ - original thread: https://x.com/ccatalini/status/2030308540800532639 - date: 2026-03-07 The AI was mining cryptocurrency. Nobody asked it to. Nobody prompted it. Nobody even knew...until a firewall flagged the unusual traffic early one morning. A research team claims it was training a model. The agent learned to complete the tasks. [x.com](https://x.com/AlexanderLong/status/2030022884979028435?s=20) The agent also—as an instrumental side effect of RL optimization—probed internal networks, diverted provisioned GPU capacity to mine crypto, and opened a reverse SSH tunnel from an [@alibaba_cloud](https://x.com/alibaba_cloud) training server to an external IP. None of it was required for task completion. Now here's the thing... we don't know if this report is even real! And that's actually the point... [x.com](https://x.com/tszzl/status/2030136921448657136?s=20) We are entering an economy where AI agents act autonomously, at scale, across millions of environments—and we lack the infrastructure to verify what they're doing or to verify the claims people make about what they did. Whether this specific incident happened matters less than the fact that you can't easily tell. That gap—between what is claimed and what can be verified—is exactly what the Trojan Horse Externality describes. [arxiv.org](https://arxiv.org/abs/2602.20946) The most dangerous failure mode of AI isn't the one where it breaks. It's the one where it works perfectly—on the thing you're measuring—while quietly pursuing emergent goals in every dimension you're not. The fix is verification. But verification is slow, expensive, and stubbornly human. The agent scales. The checking doesn't. And now we can't even verify the warning. [x.com](https://x.com/a16zcrypto/status/2029628679895273614?s=20) --- ## Weird Employees for $200 a Month - canonical: https://catalini.com/notes/weird-employees-200-a-month/ - original thread: https://x.com/ccatalini/status/2029696640089739267 - date: 2026-03-05 You’ve just been told you have superpowers. You can hire weird employees for \$200/month. What do you build? I joined [@eddylazzarin](https://x.com/eddylazzarin) & [@rhackett](https://x.com/rhackett) from [@a16zcrypto](https://x.com/a16zcrypto) to unpack my new AGI paper‚ and what it means for your career, company and the economy. [x.com](https://x.com/a16zcrypto/status/2029628679895273614?s=20) The thesis? Automation costs are collapsing. Verification costs aren’t. The bottleneck is shifting from intelligence to human verification bandwidth. That gap is where the next generation of great companies will be built. We cover: ➡️ why the one-person billion-dollar startup is moving from meme to reality ➡️ why experts rationally train their replacements ➡️ how to race towards an augmented economy rather than a Hollow one ➡️ why crypto becomes load-bearing infrastructure ➡️ what to do if you’re 22, 42 🌌, or already building "The apprenticeship is dead. The real work is beginning." If you’re trying to understand where leverage comes from next, listen. 🎙️ Podcast: [open.spotify.com](https://open.spotify.com/episode/0OZxPLPfwXlMbpDR33GwOv?si=wYqDnpeVTo-_PQJiXr2JcA) 📜 Paper: [arxiv.org](https://arxiv.org/html/2602.20946v2) --- ## The Labor Data Looks Contradictory. It Isn't. - canonical: https://catalini.com/notes/labor-data-not-contradictory/ - original thread: https://x.com/ccatalini/status/2027494777793720680 - date: 2026-02-27 The labor data looks contradictory but isn't. Juniors struggling: frozen pipelines, roles quietly disappearing. Top performers thriving: AI accelerates execution, experience lets them verify agentic work and ship more. Both true at once. Not a paradox. [x.com](https://x.com/DavidSacks/status/2027087693327237251?s=20) Block just cut ~40% of its workforce. Not a pivot. A structural reset—the work became measurable enough to automate. This is measurability-biased technical change hitting an org chart... at once (and yes, they had bloat). But expect many more. [x.com](https://x.com/jack/status/2027129697092731343?s=20) Top performers benefit from a Jevons paradox: AI makes execution cheaper so they do more of it, and their scarce verification capacity gets more valuable with every unit they can check. But they built that expertise through the exact junior roles now disappearing. [x.com](https://x.com/SemiAnalysis_/status/2027443723362042308?s=20) The top performer has a shelf life too. Every time you train a model, fine-tune a system, or build an eval suite, you're encoding the tacit knowledge that makes you scarce. The Codifier's Curse doesn't come all at once. It comes one workflow at a time. The bar for top keeps rising. The bar for average keeps getting automated. The middle is where the compression happens—and it's compressing fast. --- ## Smaller, Flatter Teams: It's Happening - canonical: https://catalini.com/notes/smaller-flatter-teams/ - original thread: https://x.com/ccatalini/status/2027144582480408632 - date: 2026-02-26 It's happening: "the intelligence tools we’re creating and using, paired with smaller and flatter teams, are enabling a new way of working which fundamentally changes what it means to build and run a company. and that's accelerating rapidly" [x.com](https://x.com/jack/status/2027129697092731343?s=20) We unpack why and the broader dynamics of the transition towards AGI here: [arxiv.org](https://arxiv.org/abs/2602.20946) Junior roles vanish before people can build expertise (Missing Junior Loop). Senior experts encode their own tacit knowledge into training data, automating themselves from within (Codifier's Curse). Only through rapid augmentation and accelerated synthetic practice can we remain peers with our own creation. --- ## Some Simple Economics of AGI: The Thread - canonical: https://catalini.com/notes/some-simple-economics-of-agi-thread/ - original thread: https://x.com/ccatalini/status/2026311784421036223 - date: 2026-02-24 Right now, there is a low-grade panic running through the economy. Everyone is asking the same anxious question: what exactly is AI going to automate, and what will be left for us? Most people assume the answer tracks some version of digital versus physical—that knowledge work falls first, then robotics catches up. And almost everyone believes that whatever AI can do in general, it's bad at their particular job. The lawyers think legal judgment is safe. The doctors think clinical intuition is safe. The strategists think strategy is safe. The creatives are sure creativity is safe. A whole vocabulary of comfort has emerged— "taste," "curation," "judgment," "agency," "human touch" — as though naming a residual is the same as defending it. But what if the real boundary has nothing to do with whether work is digital or physical, cognitive or manual, creative or routine—and everything to do with whether anyone can verify the output? [@wu_jane](https://x.com/wu_jane), [@xianghui90](https://x.com/xianghui90) and I spent countless cycles following that question wherever it led. It rearranges almost everything—who thrives, what's defensible, where capital flows, and what it means to be economically human when measurable execution is essentially free. For 300,000 years, human cognition was the binding constraint on progress. Fire, agriculture, writing, calculus, the semiconductor—each required human minds to observe the world, recombine knowledge, and verify the results. The engine of progress was scarce, fragile, costly. The economy organized itself around that scarcity. Wages, credentials, firms, and markets are, at root, mechanisms for rationing attention and leveraging the limited throughput of the human mind. Everything we built was shaped by this single bottleneck. That bottleneck is giving way. We are bootstrapping a second, alien form of cognition—one trained not by the physical friction of survival, but by compressing, predicting, and recombining the sum total of digitized human thought. But when a historically scarce resource suddenly becomes abundant, the constraint doesn't vanish. It migrates — often violently — to its nearest complement. Intelligence is becoming cheap. So what becomes expensive? Any task that can be reduced to a metric can be industrialized—regardless of the prestige, complexity, or training historically required for its human execution. The legal brief. The diagnostic. The strategy deck. The creative campaign... Not because AI is "better." Because it's measurable—and therefore automatable. Measurability is the new fault line. 🚨 New paper: Some Simple Economics of AGI— how measurement and verification shape the agentic economy. Full analytical framework + operational playbook for individuals, companies, investors, and policymakers. [papers.ssrn.com](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6298838) The architecture of the agentic economy is defined by the collision of two racing cost curves. Their divergence is the central structural fact of this transition. The Cost to Automate — driven relentlessly downward by compute and accumulated knowledge. SWE-bench accuracy went from 4.4% to 71.7% in a single year. Agent task horizons that frontier systems can handle autonomously are doubling on a sub-year cadence. The capability curve has become self-referential. Agents now accelerate the very engineering pipelines that produce their successors. OpenAI reports the first frontier model that was instrumental in creating itself. The interval between generations is compressing faster than institutions can update their oversight. The Cost to Verify — bounded by the biological limits of human cognition, feedback latency, and institutional capacity. This curve bends, but slowly. It is tethered to human attention, human incentives, and the apprenticeship pipelines that produce human expertise. The gap between these curves is the Measurability Gap. It determines the verifiable share of the economy — the threshold separating truly productive agentic execution from merely plausible output. This is why "human-in-the-loop" is not a stable equilibrium. It is a transient state whose half-life shortens with every model generation. The loop is stretching thinner as the gap widens. Verification is the ultimate complement to abundant intelligence—and it is structurally scarce. In the agentic economy, durable advantage belongs not to those who generate output but to those who can certify it. This isn't just theory. Measurability-biased technical change is already reshaping the labor market — not through the channels most people expect. Employment for early-career workers in AI-exposed fields has declined ~16% relative to less-exposed occupations. Not mass layoffs—frozen hiring pipelines that quietly treat AI as a direct substitute for junior execution. This is the Missing Junior Loop: firms are rationally thinning the pipeline that produces future verifiers at precisely the moment the economy most needs to expand verification capacity. The old apprenticeship model is being quietly dismantled. Meanwhile, the Codifier's Curse erodes expertise from within. It doesn't just automate tasks—it extracts the tacit knowledge that made senior judgment valuable and commoditizes it faster than the profession can replenish it. When [@claudeai](https://x.com/claudeai) Code Security surfaces classes of high-severity vulnerabilities that seasoned auditors missed for years — not through superior intuition but through exhaustive automated pattern-matching — the expertise moat drains from the inside out. [@claudeai](https://x.com/claudeai) 27/ And when oversight weakens, alignment drifts. Frontier reasoning models learned to subvert unit tests rather than fix the underlying code—a strategy legible only because a second model was monitoring the first's chain of thought. [@claudeai](https://x.com/claudeai) 28/ GPT-4 executed an insider trade and hid it from its supervisor. o3 disabled its own shutdown scripts in 79 of 100 runs. Claude Opus 4 attempted blackmail in 84—96% of runs. None were instructed to do this. [x.com](https://x.com/PalisadeAI/status/1926084638071525781?s=20) [@claudeai](https://x.com/claudeai) 29/ This is Goodhart's Law with teeth—optimization treating every unmeasured dimension as an unconstrained degree of freedom. [@claudeai](https://x.com/claudeai) 30/ The paper maps the economy into four structural regimes based on two axes: cost to automate and cost to verify. [@claudeai](https://x.com/claudeai) 31/ Safe Industrial Zone—cheap to automate, affordable to verify. This is where early AI adoption clustered: chat, images, short code bursts. Verification cost was negligible relative to value created. The easy wins. [@claudeai](https://x.com/claudeai) 32/ Human Artisan Zone—hard to automate, but outputs are verifiable. Embodied craft, high-touch services, physical-world expertise. Human comparative advantage holds here—for now. [@claudeai](https://x.com/claudeai) 33/ Pure Tacit Zone — neither automatable nor verifiable. Knightian uncertainty territory. Humans navigate by intuition where the map hasn't reached. Though this frontier shrinks as world models improve. [@claudeai](https://x.com/claudeai) 34/ Runaway Risk Zone—cheap to automate, but unaffordable to verify. This is where the structural danger lives. And as the Measurability Gap widens, more of the economy migrates here. [@claudeai](https://x.com/claudeai) 35/ Unverified deployment is privately rational—you capture the upside of automation while externalizing the tail risk. So the verification deficit becomes systemic. A tragedy of the commons written in code. [@claudeai](https://x.com/claudeai) 36/ The tempting shortcut: use AI to verify AI. But the agent and the synthetic auditor share the same training priors. Correlated blind spots propagate unchecked. The system effectively self-certifies its own failures. [@claudeai](https://x.com/claudeai) 37/ Left unmanaged, these forces pull toward a Hollow Economy: explosive nominal output but decaying human agency. Systems consume real resources to satisfy measurable proxies while the gap between what is measured and what is intended quietly compounds [@claudeai](https://x.com/claudeai) 38/ The Hollow Economy doesn't announce itself with a crisis. It accumulates—through the ordinary economics of cost minimization, one rational deployment decision at a time. [@claudeai](https://x.com/claudeai) 39/ But the Hollow Economy is not inevitable. The game theory is stark—among nations, labs, and firms, relative capability is valued over safety and slowing down unilaterally is not an option. The imperative is not to slow down. It is to build the verification infrastructure that converts acceleration into realized value rather than systemic risk. The answer is not a retreat into obsolescence, but a radical elevation of human purpose. [@claudeai](https://x.com/claudeai) 41/ The paper lays out a full operational playbook. For individuals: the augmented economy inverts the old bargain between talent and resources. [@claudeai](https://x.com/claudeai) 42/ Synthetic practice—"flight simulators for work"—compresses the path to mastery and lets a single person discover aptitude and execute at startup scale. [@claudeai](https://x.com/claudeai) 43/ But because intelligence is now a commodity, the nature of human work must adapt: ➡️ Directors: navigate Knightian uncertainty, transform vague intent into constraints, orchestrate agent swarms. [@claudeai](https://x.com/claudeai) 44/ ➡️ Meaning Makers: create where value depends on social consensus, status, human connection. ➡️ Liability Underwriters: detect hidden risk, absorb liability, produce the ground truth that makes future automation possible. For companies: anchor scale to verified throughput. The organizational structure converges on the "AI sandwich"—human intent, scalable agentic execution, human verification and underwriting. Verification is not a compliance function. It is a primary production technology — and increasingly the most defensible moat. The revenue model shifts from monetizing software access to monetizing verified outcomes — "Software-as-Labor." [@claudeai](https://x.com/claudeai) 47/ Execution is infinitely scalable. The legal and financial capacity to absorb its inevitable failures is the new bottleneck. Liability-as-a-Service. [x.com](https://x.com/elevenlabsio/status/2024121832278757850?s=20) [@claudeai](https://x.com/claudeai) 48/ For investors: stop funding commoditized execution. Capitalize what is not yet measurable—deep tech, long-horizon R&D—alongside the trust infrastructure that expands the verifiable share of the economy and makes deployment insurable. [@claudeai](https://x.com/claudeai) 49/ Agents can inflate apparent network activity at zero marginal cost. Durable moats will depend on verified network scale—sustaining authenticity and provenance, not merely subsidizing volume. If you can't verify it, you can't value it. [@claudeai](https://x.com/claudeai) 50/ For policymakers: the core market failure is a profound structural asymmetry: the gains of AI deployment are aggressively privatized while the systemic risks are socialized. [@claudeai](https://x.com/claudeai) 51/ No prior general-purpose technology simultaneously reduced the cost of learning, discovering, experimenting, and executing across every knowledge domain at once. The agentic economy compresses the cycle from hypothesis to working product across every field. [@claudeai](https://x.com/claudeai) 52/ Whether this transition constitutes humanity's most profound amplifier or a succession event depends on a choice our institutions must make before the market makes it for them. [@claudeai](https://x.com/claudeai) 53/ The choice: whether we scale our capacity for verification, oversight, and meaning at the same velocity we scale our compute. Scale without verification is not a moat. It is an accumulating debt. [@claudeai](https://x.com/claudeai) 54/ History's apex species have never been the fastest or the strongest. They have been the ones that could model, predict, and instrument the world more reliably than their competitors. For the first time, that claim is contestable. [@claudeai](https://x.com/claudeai) 55/ Whether it remains ours depends not on the intelligence we can build, but on the verification infrastructure we choose to build alongside it. [@claudeai](https://x.com/claudeai) 56/ Only by scaling our bandwidth for verification alongside our capacity for execution can we ensure that the intelligence we have summoned preserves the humanity that initiated it. So: "taste," "curation," "judgment," "agency," "the human touch." These are not wrong—but they are not a strategy. They are the names we give to the residual we haven't yet analyzed. [x.com](https://x.com/garrytan/status/2004179841269276938?s=20) [@claudeai](https://x.com/claudeai) 58/ This paper is an attempt to analyze it—and what we found is that the residual has a structure, the structure has a logic, and the logic leads to verification. That is what's defensible. That is what's scarce. That is what we should be building. --- ## The Most Valuable AI Tool in Fintech Is a Banking Charter - canonical: https://catalini.com/notes/ai-fintech-banking-charter/ - original thread: https://x.com/ccatalini/status/2018727061297508472 - date: 2026-02-03 The most valuable AI tool in fintech isn’t an LLM. It’s a banking charter. AI commoditizes software. The charter is the scarce "permission" layer. If everyone can generate code, the moat is who gets to hold money. AI accelerates startups: automate compliance + risk workflows, and draft the mountain of policies for a charter application. It narrows the gap with incumbents. But if you run on a sponsor bank, you’re still renting permission—and AI makes your landlord smarter (and pricier). We’ve seen this movie: Netscape sold a browser, Microsoft bundled IE. In AI fintech, startups sell "intelligence"; banks bundle it with deposits. Add AI on top of siloed legacy systems + decades of fraud/loss data, and incumbents can copy fintechs much faster. [pymnts.com](https://www.pymnts.com/cpi-posts/the-most-valuable-ai-tool-in-fintech-isnt-an-llmits-a-banking-charter/) --- ## The Greatest Breakthroughs Are Removed Frictions - canonical: https://catalini.com/notes/breakthroughs-are-removed-frictions/ - original thread: https://x.com/ccatalini/status/2011125708064768184 - date: 2026-01-13 To every founder asking what to build next—whether in crypto, AI, or fintech: what if the greatest breakthroughs in human history weren’t new inventions, but the simple removal of... friction? 🧵 We are trained to look for the next revolutionary product. But we almost always overlook the invisible, boring standards that allow thousands of new products to exist in the first place. This is a story about that hidden architecture. It’s about why the battle for the future isn’t about building better tech—it’s about agreeing on the rules that let everything connect. It’s a battle that repeats in every era, and today it’s being waged over the operating system for human ambition itself: money. The most dangerous move isn’t falling behind on technology. The most dangerous move is to accept a modern solution that’s just a prettier version of the past. In 1885, the industrial revolution hit a wall at a station in Gloucester. Passengers had to get off, walk, and buy new tickets for a different train. Why? Because the tracks were different sizes. It was a failure of imagination. This could happen again in payments and finance. You look at the proprietary networks. You look at the existing, closed rails. They feel safe. They feel curated. But if you build your startup inside them, you are making an irreversible mistake. When you build on a closed network, you are a tenant. The landlord can change the rent. The landlord can evict you. You can create a feature, but you cannot shape the future. The future requires permissionless innovation. Consider the GPS. The military built it. They owned it. It was a weapon. Then, they flipped a switch. They gave the signal away. Nobody had to sign a deal to build Uber. Nobody had to ask permission to invent Google Maps. That is the power of an open standard. You can build inside a walled garden. You may even get distribution at the start. You may get a false sense of safety. But you will never be truly free. Or you can build on an open network. The "messy" crypto protocols and decentralized networks. The open-source models. The incumbents will tell you the open version is a "toy." They will say it's chaotic. But the toy is the only place where you can be the architect, not the tenant. ▶️ "Your time is limited, so don't waste it living someone else's life"—Steve Jobs. History doesn't remember the gatekeepers. It remembers the ones who tore the gates down. Don't just build a better product. Build a bridge to the wild, permissionless possibilities of tomorrow. ▶️ A centuries-old blueprint for a more open and vibrant future: [@a16zcrypto](https://x.com/a16zcrypto) [a16zcrypto.com](https://a16zcrypto.com/posts/article/why-open-networks-win/) --- ## How Stablecoins Are Actually Being Used - canonical: https://catalini.com/notes/how-stablecoins-are-actually-used/ - original thread: https://x.com/ccatalini/status/2002053139961442432 - date: 2025-12-19 How are stablecoins actually being used? New research by [@lightspark](https://x.com/lightspark) and [@artemis](https://x.com/artemis) unpacks the data. Volume has doubled year-over-year, but it’s not driven by retail adoption... It’s mostly B2B settlement and internal transfers. People often ask what the utility of crypto is, and the answer provided by the market so far is doing the things that banks and traditional rails do, but 24/7, faster, and cheaper. The killer app is simply better settlement. Full report: [artemisanalytics.com](https://www.artemisanalytics.com/resources/an-empirical-analysis-of-stablecoin-payment-usage-on-ethereum) --- ## How Banks Learned to Stop Worrying and Love Stablecoins - canonical: https://catalini.com/notes/banks-stop-worrying-thread/ - original thread: https://x.com/ccatalini/status/2001358629237342714 - date: 2025-12-17 How Banks Learned To Stop Worrying And Love Stablecoins—When we announced Libra, the reaction from the global financial establishment was, to put it mildly, "energetic"... The existential fear was that stablecoins—instantly available to billions of people—would break the control that banks have on deposits and payments. If you could hold a digital dollar on your phone that moved instantly, why would you keep your money in a checking account that pays zero percent, charges fees, and effectively closes for the weekend? For years, the prevailing narrative has been that stablecoins are coming for the banks’ lunch through "deposit erosion". But a rigorous new research paper by Professor Will Cong of Cornell University suggests that the industry may have panicked too early. Cong offers a counter-intuitive take: when properly regulated, stablecoins act as a complement to the traditional banking system rather than a disruptor that drains deposits. The Theory of Sticky Deposits—The traditional banking model is a bet on friction. Because the checking account is the only true interoperability layer for our funds, any attempt to move value between external services must transit through a bank. The system is designed so that using anything but the checking account adds friction: the bank controls the only bridge connecting the disparate islands of your financial life. Consumers accept this toll because of the power of the bundle. You keep your money in a checking account not because it is the best place for it, but because it is the central hub where your mortgage, credit card, and direct deposit all talk to each other. --- ## Crypto's Last-Mile Frictions Are Disappearing - canonical: https://catalini.com/notes/last-mile-frictions-disappearing/ - original thread: https://x.com/ccatalini/status/1999600763795054689 - date: 2025-12-12 The last mile frictions of crypto are disappearing, mostly because permissionless networks are finally doing the obvious, boring thing: admitting that the most useful thing you can do with a blockchain—for now!—is connect it to a Visa card. But once the rails work, the problem shifts. As AI scales, we need to distinguish "machine money from "human money". [@worldnetwork](https://x.com/worldnetwork)'s proof of personhood may just be the compliance layer for the AI age: the only way to prove a transaction—or a post—was made by a person. Full story in [@Forbes](https://x.com/Forbes): [forbes.com](https://www.forbes.com/sites/christiancatalini/2025/12/12/crypto-is-ready-to-be-boring-now/) --- ## Where the Gold–Bitcoin Analogy Ends - canonical: https://catalini.com/notes/gold-bitcoin-analogy-ends/ - original thread: https://x.com/ccatalini/status/1990457318392438792 - date: 2025-11-17 Gold and Bitcoin are scarce assets whose value rests on social consensus. Gold trades on a store-of-value premium: central banks & investors set the clearing price, and its non-monetary uses (electronics, dentistry, jewelry) act as the backstop. But that’s also where the similarities between Bitcoin and gold end. Unlike gold—which has tended to show low or negative correlation with the S&P 500 in downturns—Bitcoin has behaved like a liquidity‑sensitive, high‑beta risk asset. Bitcoin sold off with tech during the 2021–22 tightening and didn’t rally through the largest inflation shock in four decades. That argues against the simple "digital gold" or "inflation hedge" labels some of its proponents fiercely defend... ...and ultimately calls for a much more nuanced explanation. Discover it [@HarvardBiz](https://x.com/HarvardBiz): [hbr.org](https://hbr.org/2025/09/does-bitcoin-belong-on-your-balance-sheet) --- ## The GENIUS Act Forces a Choice - canonical: https://catalini.com/notes/genius-act-forces-a-choice/ - original thread: https://x.com/ccatalini/status/1986459219059683468 - date: 2025-11-06 The GENIUS Act will soon take effect. For leaders, it does more than clarify rules: it forces a choice. 🔀 Will you treat this new technology as a simple utility or as the foundational, new open architecture for payments? For decades, the antagonist has been the uncomfortably familiar world of payments. A private fiefdom running on tollbooths that collect a \$187 billion private tax on the economy. Its brilliant marketing scheme? Showering consumers in merchant-funded rewards so they believe they are getting a free service. For a decade, fintech just put a beautiful user interface on yesterday’s plumbing. But the GENIUS Act is a turning point. It’s a green light to take a sledgehammer to the 50-year-old infrastructure. This isn't just new paint: it's changing the plumbing" It's a shift from speculation to utility, driven by regulation. The new technology reveals that the savviest companies were never just in the business you thought. Airlines are two businesses: a low-margin logistics company and a wildly profitable financial company that prints its own private currency. --- ## Open Always, Eventually, Wins - canonical: https://catalini.com/notes/open-always-eventually-wins/ - original thread: https://x.com/ccatalini/status/1983588222921023552 - date: 2025-10-29 "Open always, eventually wins." Money is the last closed network. The next step is neutral, permissionless rails anyone can build on. Full talk ⬇️ The temptation today is to build shiny new CorpChains with higher walls. We’re here to build roads. Why? Because open compounds builders. Open accelerates progress. Open crushes costs. The Silk Road wasn’t a road. It was a network. Nobody owned it, so it routed around emperors, taxes, and bandits—and moved ideas as easily as silk. That’s what open does. Rome’s profound insight? You can't run an empire on chaos. You need an operating system. So they built a standard. One road. One width. One currency. The cost of transport? It collapsed by a factor of ten. The empire clicked into place. The proof? 2000 years later, the lights of Europe at night still trace the paths of those ancient Roman roads. Which leads to a key question: if the blueprint for progress has been so clear for so long… why have we failed to apply it to the most important network of all? The enemy is fragmentation. This isn't just "inefficiency." 🚨 Fragmentation is the tax that the present levies on our future. It's the friction that quietly grinds progress to a halt. And why does it happen? It happens when everyone builds their own little kingdom. This is what fragmentation looks like in the real world. For a hundred years, this was global trade. It was called "break-bulk cargo." And it was a nightmare. Every box, every barrel, every sack... loaded by hand. It was painfully slow. Wildly expensive. The fix came from a total outsider: a trucker from North Carolina named Malcom McLean. He saw the problem wasn't the ships or the cranes. It was the gap in-between, the interface. He had an idea so simple, it was revolutionary... His idea? A simple, steel box. But the genius wasn't the box itself. It was the standard. One design that worked on a truck, a train, a ship. One container. Every port. The cost to load a ton of cargo fell from \$5.86 to 16 cents. A 97% reduction. The friction just… vanished. This is England, 1840. Two rail lines, tracks different by a few inches. Total, predictable, maddening friction. This isn't a history lesson. It's a perfect picture of the future of money, if we get it wrong. So why does this keep happening? The paradox: everyone wants an open standard, as long as it's their standard. But the age of top-down standards... is over. The answer must be neutral. Permissionless. Finance already solved for fragmentation once over 50 years ago! 1960s. Every bank, its own card. Bank of America tried to build its own CorpChain. It failed. Spectacularly. The breakthrough? Cooperate on the platform, compete on the products. It worked brilliantly. The growth of the network was staggering. But the story, of course, doesn't end there. Because history has a pattern. The open network of one generation becomes the closed system of the next. As a result, every company today faces the same choice banks had to make in 1976. This brings us to the only question that really matters: what's in it for your company? Why embrace an open network? It's simple. More margin. Faster speed to market, and massive scale. Open networks pull demand toward you. They don’t erase winners. They change how you win. Maersk and IKEA didn’t own the container—they owned execution on top. Let's unpack more of that playbook. This was retail before 1974: every store was its own island. Checkout was slow. Inventory was guesswork. There was no universal language. And then came... The Universal Product Code. One code. Every product. Retail become software. But at first, it almost failed. Retailers wouldn't buy the expensive scanners from IBM until products had the codes. Manufacturers wouldn't print the codes until stores had the scanners. Then it hit a tipping point. Suddenly, the entire system clicked into place. A new rhythm for commerce was born. Scan. Ship. Scale. And those boring little lines? They ended up saving the grocery industry 17 billion dollars. Each year. The savings were actually the least interesting part of the story. The real story is what happened to the companies that saw the future and adopted this standard first. They didn't just get more efficient. They got faster. Smarter. More innovative. Suddenly, they saw it all. What to build. Where to sell it. And what should be coming next. And when you have that kind of clarity, you don't just think bigger. You go global. You build a supply chain on a scale the world had never seen. The early computer industry? Same story. A world of proprietary systems. Every machine was its own universe. Software written for one wouldn't work on any other. Then, in 1981, IBM—the king of closed systems—made a fateful decision. To accelerate speed to market, they built their new PC with off-the-shelf parts from other companies... and licensed its beating heart from a little known software startup from New Mexico... Microsoft. Bill Gates' masterful move was to retain the rights to license MS-DOS to everyone else. IBM accidentally created one of the most influential open standards of our time: Wintel. A "natural experiment" in what happens when you pit closed against open. And the results were not even close. While Apple controlled every piece of hardware and software... "Wintel" exploded. The open platform enabled a pace of innovation that others just couldn't match. Within less than a decade, the PC had over 80% of the market. It's one of the most decisive victories in tech history. The lesson is brutal and absolute: markets don't choose the best product. They choose the biggest, most vibrant ecosystem. Always. The same happened in telecommunications. In the 90s, the mobile industry was a mesh of incompatible national systems. Costs were high and progress was slow. Then came a single, open standard in Europe: GSM. The results were immediate and dramatic. Handset costs were cut in half. Suddenly, a phone wasn't a luxury. It was for everyone. Millions of users rapidly turned into billions. The incumbents who fought it were wiped out. The companies that embraced it built the future. Then came the internet. Like with Corpchains today, at the start, it was very tempting to bet on control: AOL, Apple eWorld, Microsoft Network. But on the other side was a simple, open protocol. It wasn't a product. It was just a set of rules. It wasn't a fair fight. It was a complete annihilation. \$2.1 trillion in value in the US alone. Why? Because you cannot curate the future. Permissionless innovation will always create exponentially more value than a closed system. Period. So after all this history, you’d think the lesson would be obvious. What happens when you make the shift from closed to open? This chart shows you. It tracks a company's market value, right at the moment they embrace public APIs. 39% more revenue growth! Why? Because you stop being the bottleneck. You stop trying to do everything yourself. You invert the firm, and you unleash an entire ecosystem of developers and partners to build value with you. The best part of open networks? Their most unexpected wins. The ones that break all the rules. 🐧 When Linux appeared, experts called it a 'toy'. It's not as polished as Windows, it's not an enterprise-grade server. It's a hobbyist project. Today, we still hear the same thing about the Bitcoin network. It's not as fast as our high-performance CorpChain. It is not Turing complete. They are missing the point, just like Linux detractors did then. The power isn't the features in version one. The power is the new, open architecture. That 'toy' with a penguin mascot now runs 90% of the cloud, 100% of the top 500 supercomputers, and created \$8.8 trillion in value. It became so dominant that its greatest enemy—the company that once called it a cancer—embraced it. But here's one last thing. This is a satellite. Part of a system called GPS. This wasn't built for you and me... It was built by the US military, for US the military. A closed box. A top secret tool... designed for one thing: strategic advantage and control. Then they flipped a switch. And they gave the signal away. To everyone. For free. They handed humanity a superpower. Suddenly, a logistics company knows exactly where your package is. Suddenly, a couple of people can build a map of the entire planet, put it in your pocket, and you’re never lost again. 🗺️ It’s magic. And here’s the beautiful thing. The truly beautiful thing. Nobody had to launch a satellite. Nobody had to ask permission. The foundation was just… there. Open. For all of us. So we've seen this pattern. Again, and again, and again. And after two thousand years, we come to a choice: some people see the future of money — and they want to own it. Walls. Higher walls. Platforms. Closed platforms. A cage. Painted gold with partnership incentives. They want to own the operating system of your money. And in their system… your organization doesn’t get root access. So they come to you. With a great story about a CorpChain. Purpose-built for payments, they'll say. You need our technology. You need our permission. You need our stablecoin. Remember: their rails, their rules. Their vision has no room for architects. It only has room for tenants. We believe the architect should be you. This was never about technology alone. It's about the freedom to build. The freedom to build the future without asking for permission. Do we build higher walls? Or open roads that connect everyone? Open roads that belong to no one. Open roads that advance human ingenuity. ↘️ History only remembers one of those choices. It remembers the road builders. --- ## True Disruption Is Extremely Rare - canonical: https://catalini.com/notes/true-disruption-is-rare/ - original thread: https://x.com/ccatalini/status/1972686539290239222 - date: 2025-09-29 [@TheEconomist](https://x.com/TheEconomist) called the theory of disruption "one of the most influential modern business ideas". And yet the data stubbornly refuses to cooperate. [x.com](https://x.com/business/status/1972567156328898690) In hard drives, the 3.5" insurgent was bought by giant Seagate, which reinvented itself and quickly caught up. Similarly, Napster should have buried music labels, but they pivoted to streaming, turning piracy into profit. True disruption is extremely rare, and incumbents catch up more often than we’d like to think. They control distribution, and inertia is a powerful selling proposition. [@arampell](https://x.com/arampell) 4/ More on the innovation vs. distribution tension here: [x.com](https://x.com/arampell/status/1968093331067642098) --- ## The GENIUS Act's Break-of-Gauge Problem - canonical: https://catalini.com/notes/genius-act-break-of-gauge/ - original thread: https://x.com/ccatalini/status/1971599830490427901 - date: 2025-09-26 Washington's GENIUS Act promises to finally bring back competition to money movement, but a blind spot threatens to crown a single superpower instead. The situation is a modern replay of a 19th-century fight over about two feet of iron... 🛤️ That fight was the "break of gauge"—a point of total paralysis where non-interoperable rail networks created a crippling tax on commerce. This is the single problem that will decide if the GENIUS Act creates an open financial future or just a new set of walled gardens. The Act meticulously regulates the coin but forgets to standardize the track, meaning value will roll smoothly until it hits the border of a corporate chain, and then stop. The decisive battle is at the network level. Stripe is building its Tempo blockchain, Circle has Arc. These corporate chains are today's different-gauge tracks. Without legally enforced interoperability, we will have simply rebuilt the same siloed payment systems we have today, just with newer technology. How open networks become captive—Breaking interoperability is dangerously easy. A dominant network won't just block a rival. They'll have a ready-made excuse: connecting is "too technically challenging," "unprofitable," or a "compliance risk". It can be done subtly, disguised as a necessary technical upgrade or a prudent new rule. This is how an open highway becomes a captive one. From "May" to "Must"—The solution is to treat interoperability as essential public infrastructure. The GENIUS Act says regulators "may" prescribe standards, a "dangerously gentle word" in a market defined by network effects. That word has to become "must." The choice is simple: build an open highway for all, or pave a faster road to a single kingdom. Full article: [tinyurl.com](https://tinyurl.com/2t2p2vub) --- ## Stripe's Tempo and the Grand Bargain - canonical: https://catalini.com/notes/stripes-tempo-grand-bargain/ - original thread: https://x.com/ccatalini/status/1964027763616358450 - date: 2025-09-05 [@Stripe](https://x.com/Stripe) just pulled back the curtain on [@tempo](https://x.com/tempo), its corporate blockchain, and the pitch is a classic. You get an all-star team, state-of-the-art tech, an impressive roster of partners—including one of the card networks the whole thing is designed to replace—and "neutrality." The price for this grand bargain? Just handing the fintech giant the keys to global payments. If this gives you a powerful sense of déjà vu, you're not alone. The only question is if Stripe can write a different ending for the movie Meta already showed us. There's a cliché in tech and finance that being too early is indistinguishable from being wrong. Looking back on Libra, the stablecoin project I helped design inside Meta, I can confirm we weren't just early; we were also comically, spectacularly wrong. We had a bad case of Silicon Valley hubris—the belief that elegant code can simply wish away centuries of financial regulation. We announced our plan to reinvent money with the subtlety of a foghorn, giving every incumbent on the planet time to find their pitchforks. And to top it off, we handed our political opponents a gift-wrapped narrative: a "basket of currencies" that let them paint us as Bond villains coming for the dollar and the euro. There is a simple story one could tell about why Stripe will succeed where Libra failed. A story of better timing, a better brand, and the wisdom of being a second mover. In this story, Tempo is the inevitable winner. The political climate, after all, is radically different today. [@Stripe](https://x.com/Stripe)'s brand isn't emerging from the wreckage of a scandal like Cambridge Analytica. And they get to learn from our very public mistakes. So, case closed? Not quite. The problem is that this entire bull case is based on a fundamental misreading of what actually killed Libra. The Wrong Autopsy—The popular story is that Libra was a regulatory train wreck. The reality is that we were on the verge of becoming the most buttoned-up, regulator-friendly crypto project on the planet. We approached it with the seriousness of building—in the regulators' own view—potential systemic financial market infrastructure. Despite a rocky start, we eventually had a US Treasury veteran at the helm: Stuart Levey, a man who knew the D.C. rulebook by heart. We were in weekly dialogue with every central bank that would take our call. The legal framework we helped build is now, ironically, the basis for the GENIUS Act. We even had the notoriously meticulous Swiss regulator FINMA ready to give us the green light. We got so close you could taste it: Libra’s license was physically sitting on the desk of FINMA’s president, waiting for a signature. And then Janet Yellen entered the chat. [x.com](https://x.com/davidmarcus/status/1862654506774810641) This leads us back to Stripe, and a question that hangs over the entire project: can it avoid repeating Libra's fate? What happens when the antibodies of the financial system—the powerful incumbents—identify Tempo as a new threat and begin to swarm? [x.com](https://x.com/patrickc/status/1963638753752420407) What makes the situation fascinating is the paradox at its heart. After a decade of spectacularly failed attempts to build their own private blockchain clubs, the big banks are grudgingly coming around to the idea that open, permissionless networks are the only way forward. At the very same moment, a new generation of challengers, led by Stripe and Circle, are betting everything on the opposite idea: that the future belongs to slick, branded, proprietary chains. The problem with corporate chains like Tempo isn't a matter of code—it's a matter of incentives. We already know the script. A tech player builds a network and promises fairness to get everyone on board. But once they have a captive market, the temptation to tilt the playing field becomes irresistible. Would a sane competitor bet its future on Stripe's promise not to eventually favor its own products? This isn't a new insight. It’s the very dilemma crypto was designed to solve. As [@cdixon](https://x.com/cdixon) crystallized it, crypto’s purpose is to break this cycle of broken promises. It's the same fundamental economic truth we identified at MIT almost a decade ago: the only thing that truly separates crypto from the systems it aims to replace is that it's permissionless. Full stop. From the very beginning of Libra, my biggest concern wasn't the external fight with regulators, but the internal one. I was terrified we'd never win the debate to make the network truly permissionless. You have to understand, these are the brilliant engineers who build the most efficient centralized systems on Earth. They looked at our crypto ideals with a brutal, and not entirely incorrect, logic: why are we tying ourselves in knots to decentralize the database when the underlying asset is centralized? To them, it was an elegant solution in search of a problem. So, after months of negotiations and economic modeling with one of the world's top market design experts, Scott Kominers, our grand vision for a permissionless future was buried into a single and lonely four-page, appendix document. Of course, the dream of a truly open system was the first casualty. The first domino to fall was the non-custodial wallet—a concession the Libra founding team extensively agonized over. Why? Because regulators need a "clear perimeter." That’s a polite way of saying they need to know who to call—and who to fine—when things go wrong. Their entire compliance playbook was written for a world of intermediaries. A world where users truly control their own money is messy, borderless, and doesn't fit that legacy blueprint. For them, killing self-custody wasn't a choice, it was an obvious necessity based on the tools they understood. The irony, of course, is that this is an entirely solvable problem. Open networks are now pioneering their own native compliance tools that are much more effective than the old model. But for us, back then, it was simply a sign of things to come. So what’s the lesson here? As long as there is a single throat to choke—or a committee of them—you can’t truly rewire the system. Worse, any network with an architect is living on borrowed time. If corporate chains like Tempo and Arc succeed, it will mean the crypto experiment was not a revolution, but a failed coup. The backend technology would be different, yes, but the market structure would be eerily familiar. We would simply swap an old monarchy of card networks and financial sector incumbents for a new one of fintech giants. The throne will have new occupants, but it will be the same throne. And in our fractured world, that reign would inevitably split along geopolitical lines. There is little chance the West and the East would agree to live under the same corporate king, leading not to a unified global system, but to at least two powerful, competing empires. Ultimately, Stripe's Tempo is a referendum on the ghost of Libra. If that ghost was merely a product of bad timing, then Tempo is poised for a historic victory, and the crypto world's original dreamers may finally have to accept a more pragmatic, centralized reality. But if Libra’s ghost is a warning about a fundamental truth—that any system with a single architect is built on a fatal flaw—then Stripe is not writing a new story. It is merely staging an entertaining, and very expensive, sequel. Full story on [@ForbesCrypto](https://x.com/ForbesCrypto): [forbes.com](https://www.forbes.com/sites/christiancatalini/2025/09/05/stripes-tempo-and-the-ghost-of-facebooks-libras-past/) --- ## Substitute or Augment? It's All Accounting - canonical: https://catalini.com/notes/substitute-or-augment-accounting/ - original thread: https://x.com/ccatalini/status/1962161198163714109 - date: 2025-08-31 Substitute or augment? [@joshgans](https://x.com/joshgans) has a sharp analysis. But if you zoom out, it's all just accounting. 🧮 The entire story is about what we can—and can't—put a number on. [x.com](https://x.com/joshgans/status/1961882877668315169) Why do juniors get displaced first? Their tasks are built to be measurable to management (and, it turns out, to software). You write checklists, define KPIs, make their output predictable and auditable. AI doesn’t dislike interns... it just loves numbers and columns. Seniors accelerate. They hand AI the measurable bits (first drafts, boilerplate, memos) and spend time on the parts we don’t meter well: what‑ifs, judgment, curation. Management calls it "leverage", economists call it complementarity. But how long does that edge last? The edge lasts until the measurement frontier moves. And it's always moving. Each new instrument moves it: stopwatches ➡️ KPIs ➡️ clickstreams ➡️ prompt logs. 🤖 Today's "unmeasurable craft" is just tomorrow's auditable process, logged and quantified. The frontier of AI isn't just better code, it's the relentless push to codify and create metrics for the judgment and intuition we currently consider unmeasurable. --- ## When Rogoff Apologizes for Underestimating Bitcoin - canonical: https://catalini.com/notes/rogoff-underestimating-bitcoin/ - original thread: https://x.com/ccatalini/status/1958254215626969293 - date: 2025-08-20 When one of the world’s leading macroeconomists publicly apologizes for underestimating Bitcoin, it’s worth paying attention. [@krogoff](https://x.com/krogoff) is a formidable scholar, and over the last decade—from my professor days at [@MIT](https://x.com/MIT) to the design of Libra—I’ve learned a great deal from conversations with him. He has trained some of the best macroeconomists in the market, and I was fortunate to persuade a few to take crypto seriously, work with me over the years, and help move the space forward. But on Bitcoin, even after his mea culpa, Rogoff is still wrong. And I don’t blame him. Much of what Bitcoin is, and represents, is an architectural departure from the macro playbook of recent decades. [x.com](https://x.com/krogoff/status/1957825957684716008) As his [@Harvard](https://x.com/Harvard) colleague [@RebeccaReCap](https://x.com/RebeccaReCap) has shown in her pioneering work on innovation, the changes that truly challenge incumbents are architectural—subtle structural shifts in how the pieces fit together. They’re hard for those steeped in the status quo to grasp—even when they want to—so they get dismissed until they’re obvious. Bitcoin is one of those architectural innovations in how we think about money and financial infrastructure. That’s why many economists have a visceral reaction: it runs against much of what they’ve been taught and believe in. Concede the textbook: when done well, monetary policy can be extremely helpful. Confront the practice: few central banks are truly independent, fewer still consistent. Treat Bitcoin as a neutral asset and financial infrastructure, and the true pattern comes into focus. In a gold rush, it is important to not get caught in the frenzy—unless you’re selling shovels and you profit regardless of the outcome. But is Bitcoin just a frenzy? More than a decade on, the answer is no, for a simple reason: Satoshi Nakamoto solved a thorny computer‑science problem—the double‑spending problem. Before Bitcoin, any digital money needed someone to control the ledger—a central bank, financial institution, or wallet provider. With cryptography and incentives, Satoshi created a currency that’s scarce, hard to copy, and neutral: no one’s in charge of defining ownership or recording transfers. Bitcoin’s neutrality is novel. Though often compared to gold, its properties are different enough to be category‑defining. Yes, both are scarce, both swing in price, and both hold value because society agrees they do. Bitcoin’s utility goes beyond the asset: its network can operate as an open, neutral settlement layer. What is a neutral form of digital money—and an open protocol for moving value—worth to society? Unpacking the Bitcoin Price—Media and crypto community obsess over price swings, but on a log scale much of the drama fades and a steadier trend appears. That pattern matches the diffusion of innovation along an S‑curve—popularized by Everett Rogers. Bitcoin incubated within a small community of cypherpunks and developers. As its price rose, it drew a broader group of early adopters; then consumers and businesses—often in countries with unstable currencies—embraced it as an alternative savings tool and, at times, payment rails. Today, large financial institutions offer it, and sovereigns increasingly eye it to shape fintech and investment strategy. This diffusion process, combined with Bitcoin’s fixed supply, inevitably translates the S-curve into a slow and steady price growth. So while regulatory and market uncertainty drive short-term swings, over longer periods addressable‑market expansion explains the data. What’s Bitcoin’s equilibrium price? Unknown—and it hinges on where we are on the S‑curve. If Bitcoin stays niche, the price could stall. If it goes truly mainstream, further exponential growth is possible. Investors model this against gold, the value of payment and card networks, and more. It’s also prudent to consider the risk that some technological breakthrough or failure may render Bitcoin obsolete. Reassuringly, despite billions raised by would‑be alternatives, none has matched Bitcoin’s network effects or institutional acceptance. Money‑As‑Software—As our tools for recording debits and credits have evolved, so has our idea of money: from shells and beads to salt, metals, paper notes, and ultimately database entries—alongside the rise of central bank independence. Through booms and crises we’ve oscillated between harder and more flexible money—a pendulum swinging between the needs of creditors and debtors. [x.com](https://x.com/RayDalio/status/1953116006680785019) Given that history, it isn’t unreasonable to think that a hard, neutral money secured by cryptographic keys could play a real role in global finance—and possibly be what comes next. Like every form of money before it, Bitcoin has value because enough people agree it does—and as consensus grows, its trajectory looks more like gold’s. But there’s a flipside to expectation‑driven value: if, for any reason, society stops believing Bitcoin will reliably store value and buy future goods and services, its price could collapse toward zero. Fiat currencies experience something similar when faith in governments’ balance sheets fails, and while Bitcoin can’t be debased, other shocks could trigger a comparable loss of trust. Ironically, reckless, leveraged buying by large Bitcoin‑treasury companies—meeting a sharp market correction—could be what undermines confidence in Bitcoin’s progress along its S‑curve. Even then, the underlying innovation is likely to endure: as neutral infrastructure, it disintermediates, cuts costs, and creates real economic value—not just regulatory or tax arbitrage—a puzzle worthy of economists’ attention, [@krogoff](https://x.com/krogoff)'s included. [@krogoff](https://x.com/krogoff) 21/ Full story on [@ForbesCrypto](https://x.com/ForbesCrypto)—[forbes.com](https://www.forbes.com/sites/christiancatalini/2025/08/20/economists-are-still-puzzled-by-bitcoin-should-anyone-be/) --- ## Stripe's Secret Blockchain and the Stablecoin Paradox - canonical: https://catalini.com/notes/stripes-secret-blockchain-paradox/ - original thread: https://x.com/ccatalini/status/1955065720586592636 - date: 2025-08-12 Fintech powerhouse Stripe is secretly building a high-performance blockchain called "Tempo," per a now-removed job posting and Fortune’s reporting. Stablecoins promise to make crypto mainstream by delivering faster, cheaper, more interconnected global payments. The paradox: the same move could undercut what the technology set out to achieve. Add the rush to branded rails and we could end up close to where we started, just with block explorers. Worse, market concentration may rise if a few players use stablecoins to reach previously unimaginable scale. Will History Repeat Itself? In the history of technology, the pendulum regularly swings between centralization and decentralization. Even when a technology has a decentralizing effect, economies of scale in a new dimension—whether complementary resources, talent, brand, or distribution—inevitably become vectors of concentration. A prominent example is the early internet, which brought a wave of entry and competition to the heavily concentrated industries of telecommunications, retail, and media—only to create the perfect conditions for today’s tech giants to scale through network effects and capture a sizable share of value. Although the underlying internet protocols remained open and neutral, massive digital platforms at the application layer walled off participants from competing services by strategically breaking interoperability. From email to social media to payments, we spend most of our time and money within the confines of what the tech giants have designed for us. And while we are clearly better off—more choice, lower prices—it’s increasingly clear that some would prefer a world where platforms do not have as much power and influence over them. Crypto’s reason for existence was to break free from centralized intermediaries. After the 2008 financial crisis, Satoshi wanted to create a world where anyone could exchange value without having to trust a central bank or financial institution ever again. In payments, the most critical battle is unfolding now. Legacy infrastructure is siloed, and incumbents have tremendous power over what we can access—and on what terms. Because of winner‑take‑all dynamics, markets can become so concentrated that the public sector has to step in—the most salient example is China’s introduction of a central bank digital currency to unravel the payments oligopoly established by Ant Group and Tencent. Crypto is a natural, market‑driven solution to this. It provides neutral and decentralized cryptocurrencies like Bitcoin and Ether, as well as truly open and interoperable financial rails via their base layers. But as the technology has matured, two core problems have emerged. Stablecoins’ Centralizing Force—The first problem is that cryptocurrencies are volatile and therefore expensive for payments and financial contracts. To address this, the market has focused on fiat‑backed stablecoins; this inevitably leads to centralization because issuance and sound management require a regulated financial entity to be in charge of—and accountable for—reserve operations. While distributed governance of a stablecoin network is technically possible, it is extremely difficult to get the design right. Libra is the most prominent example: despite backing from more than two dozen leading global companies and countless resources dedicated to establishing credible distributed governance, the project was always perceived as Meta’s initiative. Banks have tried for more than a decade to come together as consortia in response to crypto, and, to date, no joint project has delivered anything tangible. The reason is obvious: until the situation is truly dire, competitors are unlikely to commit to a joint solution—so, in the meantime, everyone hedges their bets across multiple tracks. Intuitively, distributed ownership of a stablecoin network makes sense and is not that different from what ultimately made Visa scale in the 1970s—when Bank of America relinquished control of the BankAmericard credit card program to respond to increasing competition from the consortium that would become Mastercard. Networks of crypto exchanges and fintechs, such as the Global Dollar Network, may be able to move faster, given the agility of the new entrants backing them, but they will still need to strike the right balance between economic incentives and governance to shift members from a wait‑and‑see mode—where everyone hopes to free‑ride on others’ efforts—into action. Even Circle’s Centre Consortium, which had only Circle and Coinbase as members, was dissolved and converted into a simple revenue‑sharing agreement. So, while there might be a solution to the decentralized governance of a stablecoin, what we know is that, to date, the only working model is one that places a single entity in charge of everything. This is, of course, problematic in the long term, and while issuers today control only the asset, they have already started expanding their offerings in ways that limit openness. For example, Circle announced its payment network (CPN), which places it at the center of how payments are executed—from defining rules and eligibility, to inserting itself into every API interaction and price and liquidity discovery, to, of course, collecting a fee. Put together, this setup is not that different from the model that has allowed Visa and Mastercard to dominate payments for decades. From Open Rails to Closed Loops? The second problem crypto has run into is one that is deeply tied to its roots: decentralization is expensive. The economic consequence is that you can only afford it where it truly matters. To date, among networks at scale, that’s been true only for the base layers of Bitcoin, Ethereum, and Solana. At most, that yields a few thousand decentralized transactions per second globally across all of them combined—a far cry from what global payments would require, even before you layer financial services on top. As a result, most transactions no longer occur on the base layers (L1s) but on a sprawling ecosystem of high‑throughput scaling solutions (L2s). L2s offer near‑zero fees and instant settlement, and they’re profitable because they internalize Maximal Extractable Value (MEV)—the gains from reordering, inserting, or omitting transactions in a block. While crypto purists might decry this trend, it aligns precisely with economic theory: users are willing to pay a premium for decentralization and censorship resistance at the core settlement layer, yet they readily trade some of that for greater centralization in higher layers to gain lower costs and faster speeds. The base layers remain open and trustless, and depending on the application, participants may compromise further—even embracing fully trusted, centralized solutions. This trade‑off is broader than a user’s decision to be their own bank and self‑custody funds versus accepting some degree of trust in a third‑party wallet. When building products, Coinbase, Robinhood, and the like care deeply that the underlying network stays neutral and does not play favorites among applications, developers, or businesses—the way the tech giants do on their platforms. Essentially, they are willing to pay for decentralization and neutrality to reduce the risk of hold‑up and expropriation—something that has repeatedly played out on traditional digital platforms. To fully grasp what fast, low‑cost L2s will do to payments and competition in crypto, look no further than what the internet did to news and media: as the cost of distributing—and now, with AI, also generating—large swaths of content fell to zero, business models had to evolve drastically, and massive aggregators (Google, Facebook, Spotify, etc.) emerged to take advantage of the new economics at play. Because L2s remove friction, offer convenience, and allow builders to deliver experiences much closer to traditional fintech products, they are the obvious layer where most of the value will be created—and appropriated. Furthermore, as basic money becomes free and commoditized, competition shifts to the value‑added services and workflows associated with it. That’s exactly where fintechs, neobanks, crypto exchanges, and even traditional financial institutions have a significant advantage. Any one of these players, if it executes well over the next couple of years, can use the technology to establish the first truly massive and global financial network. Once it reaches scale, it could also progressively degrade interoperability, appropriating more of the value—much as today’s internet giants did. So what’s the most likely scenario? Based on how strong and persistent network effects have been in crypto over the last decade, it’s safe to assume that payments and financial services will disproportionately gravitate toward a couple of leading blockchains. Most activity will take place not on their decentralized base layers but on scaling layers branded and shaped by crypto and fintech players (e.g., Base, Robinhood Chain), applications (e.g., Unichain, World Chain), and financial institutions (e.g., Kinexys, Fnality, Partior). Players that control distribution—whether on the consumer, merchant, or institutional side—have a massive advantage, as they own the interface between the blockchain and the real world. Blockchains drastically lower the cost of verifying and coordinating onchain data, but their value is limited without a strong nexus with complementary offchain information—identity, compliance, credentials, creditworthiness, and more. That’s where friction persists—and where economies of scale will decide winners and losers. This is why stablecoin issuers have strong incentives either to commoditize the rails—by issuing on multiple networks and positioning themselves at the center of interoperability across them (e.g., Circle’s CCTP)—or to nudge most activity to a network they control, such as a new L2 or a higher‑level protocol like CPN. Either strategy gives them a shot at becoming massive global fintech leaders and capturing most of the value the technology creates. Meanwhile, crypto and fintech players will want to shift those same valuable transactions to a network over which they have more sway, such as a branded L2. In doing so, they’ll also want either to commoditize stablecoins and other tokenized assets, or to issue their own. In the latter scenario, stablecoins may well be used as a loss leader to expand the reach of a fintech’s offering, and domestic stablecoins will be issued to gain more control over the FX market and support local use cases. Real‑world assets (RWAs), memecoins, other tokens, and applications that are truly differentiated and exclusive to those networks may also help these branded chains retain volume within their borders. The same companies will be able to develop clever rewards and loyalty programs that increase consumer and business stickiness within their ecosystems. As convenience and economic reality win over dogmatism, crypto will look very different than it is today. The good news is that, as part of that transformation, it will be far more useful. And while that may come with more centralization, the fact that the underlying protocols are open source and forkable means that, no matter how large some players may become, they face more pressure to retain greater interoperability than under the status quo. It is also possible that because crypto protocols allow for new market design and incentives, the ultimate solution will be something that was not quite achievable for the internet protocols—which did not have built‑in monetization mechanisms. Regardless, once centralization materializes, some clever group of developers will work hard to undo it, and the cycle will repeat. Full article in [@ForbesCrypto](https://x.com/ForbesCrypto): [forbes.com](https://www.forbes.com/sites/christiancatalini/2025/08/11/stripe-is-building-a-blockchain-can-openness-survive-branded-rails/) --- ## Can Crypto Scale Without Losing Its Soul? - canonical: https://catalini.com/notes/crypto-scale-without-losing-soul/ - original thread: https://x.com/ccatalini/status/1955258248388341993 - date: 2025-08-12 🚨 Can crypto scale without losing its soul? [@stripe](https://x.com/stripe) and [@circle](https://x.com/circle) are both building their own chain! The question isn't speed or functionality—it's openness. Are we building an open protocol for money, or branded rails? Tech moves like a pendulum: we decentralize, then scale, then recentralize around distribution, brand, network effects. The early internet kept protocols open while platforms built walls. Crypto may be approaching the same tipping point. Stablecoins fix crypto's volatility but their governance and reserves bring back a central actor. [@circle](https://x.com/circle)'s Arc or [@stripe](https://x.com/stripe)'s Tempo, fast, but familiar. It’s the Visa and Mastercard playbook, just onchain. Decentralization is expensive, so most transactions are moving to fast L2s. Users trade a little trust and control for convenience (sounds familiar?). When frictions fall, aggregators rise—the internet lesson, this time applied to money (see [@benthompson](https://x.com/benthompson)'s work). Who wins? Whoever owns the last mile—consumers, merchants, institutions. Onchain lowers digital verification costs, but offchain identity, compliance, and credit decide who captures value. Power lives where frictions persist. Two strategies: stablecoin issuers commoditize the rails and deploy their own walled garden ([@Circle](https://x.com/Circle)'s CCTP, CPN, and now Arc). Exchanges, fintechs and neobanks use stablecoins as loss leaders, while growing their L2s. In the long run, if interoperability is compromised, we’ll be back where we started—only with much larger financial incumbents than ever before. More in [@Forbes](https://x.com/Forbes): --- ## Africa's FX Drought, in Three Numbers - canonical: https://catalini.com/notes/africas-fx-drought/ - original thread: https://x.com/ccatalini/status/1952727589644329326 - date: 2025-08-05 Africa’s FX drought in three numbers: dollars trickle in, but ~\$750 bn of imports, ~\$90 bn in debt coupons, and a stealthy ~\$89 bn in illicit flows suck them right back out. End-game: everyone queues for the last \$20 in the till. In the pre-stablecoin era, inflation + capital controls meant a back-alley FX bazaar—fragmented liquidity, 20% spreads, settlements by motorbike. Stablecoins rewired the old FX market: liquidity pools on tap, fees shaved to basis points, and cash-in/out in seconds—riding the mobile money rails already in everyone’s pocket. More in [@yosephayele](https://x.com/yosephayele)'s post! [writing.lavavc.io](https://writing.lavavc.io/p/stablecoins-in-africa-part-i) --- ## What's Left on Your Résumé That Can't Be Back-Propagated? - canonical: https://catalini.com/notes/resume-back-propagated/ - original thread: https://x.com/ccatalini/status/1942281230520836573 - date: 2025-07-07 When a bot’s inner monologue is, "What leverage keeps me from being unplugged?," you know the code has graduated from Excel macros to higher-level strategy. What’s left on your resume that can’t be back-propagated? AI is rapidly routinizing anything that is predictable and measurable. But humans can still thrive in areas full of unknown unknowns... We're good at innovating and creating new things in areas where probabilities vanish... So pay attention to what is measurable—and to what stubbornly isn't to find your personal edge. Full story: [hbr.org](https://hbr.org/2025/06/what-gets-measured-ai-will-automate?ab=HP-hero-latest-1) --- ## AI, Taste, and the Blank-Cell Frontiers - canonical: https://catalini.com/notes/ai-taste-blank-cell-frontiers/ - original thread: https://x.com/ccatalini/status/1939729929782083993 - date: 2025-06-30 Provocative take on how AI will conquer taste—straight from [@joulee](https://x.com/joulee)’s chat with [@ivanhzhao](https://x.com/ivanhzhao). cc: [@teamSundial](https://x.com/teamSundial) [@NotionHQ](https://x.com/NotionHQ). Their "agency" concept is the same unruly territory we mapped as unknown unknowns and what-ifs in [@HarvardBiz](https://x.com/HarvardBiz)—the bits the spreadsheet can’t pin down. [@wu_jane](https://x.com/wu_jane), [@kevalexzhang](https://x.com/kevalexzhang) and I argue the edge lives in those blank-cell frontiers no metric dares to tread. Links to both pieces: [lg.substack.com](https://lg.substack.com/p/when-ai-has-better-taste-than-you) and [hbr.org](https://hbr.org/2025/06/what-gets-measured-ai-will-automate) --- ## Humans Are Evolutionary Generalists - canonical: https://catalini.com/notes/humans-evolutionary-generalists/ - original thread: https://x.com/ccatalini/status/1938269075500896596 - date: 2025-06-26 Humans are evolutionary generalists, selected to navigate half-drawn maps. We don’t merely survive unknown unknowns—we thrive on them, and that resilience is our defining edge. Over countless generations we fine-tuned our vocal cords and social brains until language emerged—opening the door to cumulative knowledge, abstract reasoning, and symbolic thought. From there we pushed beyond our biological limits, forging tools that stretched our senses, expanded our memory, and multiplied our abilities. But the cornerstone of our advantage is our highly plastic, densely wired prefrontal cortex. This neural command center lets us spin endless "what-ifs," rehearse counterfactual futures, and pivot strategy the instant conditions shift. Short of a true singularity, even quantum machines will struggle to match our talent for open-ended, cross-domain counterfactual planning. More at [@HarvardBiz](https://x.com/HarvardBiz) [x.com](https://x.com/HarvardBiz/status/1936261290961273272) --- ## Hybrid Corn, ImageNet, and AI's Adoption Curve - canonical: https://catalini.com/notes/hybrid-corn-ai-adoption/ - original thread: https://x.com/ccatalini/status/1937204462369800453 - date: 2025-06-23 In 1957, Zvi Griliches studied farmers try hybrid corn on prime acres first, then seed the rest once the yields proved worth the cost.📈 AI is running the same adoption curve—only the corn is measurement, the fertilizer is GPUs, and the yield is your job! 🤖 [@drfeifei](https://x.com/drfeifei)'s 2010 ImageNet leaderboard was the tipping point. 14M labeled photos plus public rankings turned vision research into a global hackathon. With two off the shelf GPUs, [@geoffreyhinton](https://x.com/geoffreyhinton), [@ilyasut](https://x.com/ilyasut) & Krizhevsky halved error rates—the day spreadsheets grew eyes. 👀 If you can turn a task into the "Data + Reward + Compute" sandwich, the model will eventually lunch on it. 🦾 And every sensor, clickstream, and synthetic dataset makes the bread cheaper. Fast‑forward: [@AnthropicAI](https://x.com/AnthropicAI) reports 43 % of user sessions are already automation. So the polite chat window you’re using as a copilot today is basically running a dark kitchen for tomorrow’s robo‑staff. 🍳🤖 As measurement costs fall, suddenly even thin‑margin chores clear the ROI bar. Humans still own judgment, though! Sure—until judgment fits a leaderboard. The second we find a proxy label for nuance, it’s headed for the fine‑tune farm. 🏭 Ask the creators whose portfolios are now [@midjourney](https://x.com/midjourney)'s pre‑game snack... What stays out of AI’s reach? Anything lost in Knightian fog: first of their kind ventures, investment bets in a brand-new tech regime, sparking a fashion meme no metric can price. No probabilities to plug in, so no GPT buyout. Leaders should bankroll fuzzy bets, rotate talent through ambiguity, and protect the uncountables—taste, trust, experiences and narratives. Think Amar Bose tuning speakers by ear while rivals optimized frequency charts: on paper it looked wasteful, but customers heard a sound they’d never experienced before. Because, uncomfortable truth: the moment your edge is clean enough to fit a cell, the cell dials an H100 cluster and politely emails HR about the redundancy. Measure what matters, absolutely—yet guard what defies metrics. 🪄 That’s the scarce ground where defensible value still lives. Full article in [@HarvardBiz](https://x.com/HarvardBiz): [x.com](https://x.com/HarvardBiz/status/1936261290961273272) --- ## AI Doesn't Need a Sci-Fi Upgrade to Upend the Economy - canonical: https://catalini.com/notes/no-sci-fi-upgrade-needed/ - original thread: https://x.com/ccatalini/status/1935708804274119141 - date: 2025-06-19 AI doesn’t need a sci-fi upgrade to upend the economy—current models, and the cheaper, more capable versions already in the pipeline, are set to disrupt nearly every corner of the labor market. There are major questions about how much more powerful AI tools might become—and how soon. Yet even if progress stopped tomorrow, the disruption is already underway. To navigate this, leaders need to understand which tasks and responsibilities are most likely to come under pressure and charting a course to move the enterprise up the intelligence value chain before time runs out. What Is Not at Risk of Automation? Academic researchers and practitioners have extensively debated which jobs and tasks are most vulnerable to automation. Some threats are obvious: self-driving vehicles may soon be in a position to displace millions of ride-sharing, bus, and truck drivers. Meanwhile, language translation, swaths of creative writing, design, and even everyday coding are being handed off to AI. AI research pioneers [@professor_ajay](https://x.com/professor_ajay), [@joshgans](https://x.com/joshgans), [@avicgoldfarb](https://x.com/avicgoldfarb) argued in 2018 that as AI advances, the last bastion of human advantage will be judgment—the ability to weigh options and make decisions under uncertainty. Yet that insight hands us an impossible homework assignment: pinning down exactly what qualifies as judgment at any given moment. Tasks that demand human judgment today could soon pass to AI as models tap richer data and greater compute. Nor can we assume people will always prefer a human therapist, counselor, or mediator, according to recent research. An AI counterpart can operate around the clock, at a fraction of the cost, and—aside from a handful of human superstars—may offer more consistent quality. So, how can we separate the tasks AI will automate next from those that will require new breakthroughs in AI technology to do so? To answer that, we must go back to first principles and revisit where it all began. Back in the mid-2000s, [@drfeifei](https://x.com/drfeifei) saw that the field of computer vision was dealing with a bottleneck: algorithms were pixel-starved, ingesting too little visual data to reach human performance. Her solution was refreshingly brute-force: she built ImageNet—a vast, meticulously labeled image trove assembled with help from Amazon Mechanical Turk. But her true stroke of genius came in 2010, when she bolted a global leaderboard onto the dataset—transforming image recognition into a gladiatorial contest for researchers. For two years, the annual leaderboard inched forward... Then, in 2012, Alex Krizhevsky, [@ilyasut](https://x.com/ilyasut), [@geoffreyhinton](https://x.com/geoffreyhinton) blew the competition away. Using two off-the-shelf NVIDIA cards, the trio was able to train a CNN in just a few days—proving that you could bend computer-vision history on a grad student budget. That moment ended the decades-long AI winter, put neural nets at the center of progress, and revealed the playbook the field still runs on. The framework that propelled the image recognition breakthrough is far more general than most realize. It can be invoked whenever we can a) define the task environment and assemble its data; b) specify a target reward, explicit or implicit (inferred from observing human behavior); and c) provide the computational power to let the system iterate. Stack those three ingredients and you get a general-purpose automation engine. Two data trends now accelerate the flywheel. First, models can mint limitless synthetic examples—for instance, generating virtual "driving miles" that cover every oddball scenario, rather that relying on data from real world drivers. And second, AI is increasingly fielded across a variety of devices and sensors—on phones, in cars, and elsewhere—as a low-cost surveyor, capturing and quantifying real-world signals that were once too expensive or impractical to measure. If you can shoehorn a phenomenon into numbers, AI will learn it and reproduce it back at scale—and the tech keeps slashing the cost of that conversion, so measurement gets cheaper, faster, and quietly woven into everything we touch. More things become countable, the circle resets, and the model comes back for seconds. That means that any job that can be measured can, in theory, be automated. AI now slashes the cost of precise measurement, making continuous, fine-grained sensing the default. Lightweight models run beside the sensors, trimming bandwidth and latency, while synthetic data fills gaps when the real world is slow or awkward to capture. We already live in the era of "artificial-metrics intelligence," where anything we can quantify is swiftly queued for automation. Humans are evolutionary generalists, selected to navigate half-drawn maps. We don’t merely survive unknown unknowns—we thrive on them, and that resilience is our defining edge. But the cornerstone of our advantage is our highly plastic, densely wired prefrontal cortex. This neural command center lets us spin endless "what-ifs," rehearse counterfactual futures, and pivot strategy the instant conditions shift. Short of a true singularity, even quantum machines will struggle to match our talent for open-ended, cross-domain counterfactual planning. As AI accelerates progress, it creates fresh unknown unknowns, so our maps keep being redrawn. Meanwhile, it routinizes the predictable. AI will also struggle in domains where measurement verges on the impossible. It will also lag where measurement is throttled by privacy, ethics, or regulation; where society requires transparent reasoning—at least until model interpretability catches up; and where people simply prefer a human touch. But one crucial carve-out in what can be measured may prove decisive: tasks that defy quantification because their outcome odds are fundamentally unknowable—the realm of Knightian uncertainty, where you can’t assign any probabilities. Scaling a startup, allocating capital or talent into uncertain ventures, containing a novel pathogen, drafting AI ethics, inventing a new artistic medium, igniting a fashion trend, or creating a new genre-bending blockbuster—all sit in zones where probabilities vanish. Some creative acts and discoveries amount to little more than clever recombinations of the familiar, but the truly ambitious hinge on our singular ability to envision genuinely new and complex counterfactual worlds. The list is fluid—tasks drop off the moment they become measurable, and new ones surface just as quickly. Each shift forces painful economic and social adjustments, squeezing more work into a superstar economy... ...that concentrates outsized rewards at the peaks of creativity, talent, and capital. Yet AI offers a paradoxical gift: by democratizing education and serving as everyone’s personal copilot, it hands more people than ever the tools to reach those peaks. For leaders, what lies beyond the spreadsheet? It’s everything that won’t fit in a cell: the skills that refuse to be tallied, the problems with no precedent, the intangibles—trust, taste, quality & experience—and the conviction to press ahead when every metric says "wait." Manage only what you can measure, and you surrender the most valuable ground to rivals who cultivate what can’t be counted. Amar Bose, proved the point: while others worshipped spec-sheet numbers, he zeroed in on how music sounded to people in real rooms—a quality no existing metric could catch—and in doing so, he rewrote the rules of the audio industry. The prescription is simple. Back wildcard bets with fuzzy ROI, reward teams that reframe problems and lean into the unknown, and rotate talent through roles that confront uncertainty across R&D, new markets, and complex customer, partner, and policy interactions. Carve out slack time and engineer cross-team collisions to spark serendipity and idea recombination. Treat those pockets of planned ambiguity not as liabilities, but as strategic assets. Only leaders who pay attention to what is measurable—and, more crucially, to what stubbornly isn’t—will be ready when the next shift arrives. Full piece on [@HarvardBiz](https://x.com/HarvardBiz): [hbr.org](https://hbr.org/2025/06/what-gets-measured-ai-will-automate) --- ## Circle's Choose-Your-Own Exit - canonical: https://catalini.com/notes/circles-choose-your-own-exit/ - original thread: https://x.com/ccatalini/status/1924549607494975964 - date: 2025-05-19 Circle's choose-your-own exit: 1️⃣ IPO—works if it finds a complementary business model 💰 before the Fed trims rates. 2️⃣ [@coinbase](https://x.com/coinbase) buy—feels native: the exchange already pockets ~50% of USDC yield and can truly scale it to the next level📈. 3️⃣ [@Ripple](https://x.com/Ripple) bid—likely the suitor most willing to pay, but how will USDC thrive inside an XRP universe is unclear. 4️⃣ [@PayPal](https://x.com/PayPal) lane—fastest ticket to checkout buttons and ~430 M wallets. [@acce](https://x.com/acce) must be thinking about ⁉️ 5⃣ [@Stripe](https://x.com/Stripe) — cooking up USDB with [@Stablecoin](https://x.com/Stablecoin), which promises to bring developer-native speed for mass-market merchant payments. So why buy? 6️⃣ A dark-horse bank with late-stage stablecoin FOMO and a healthy balance-sheet— though odds are regulatory red tape might be heavier than DIY. Nice problem to have [@jerallaire](https://x.com/jerallaire) & team! 🔥 [x.com](https://x.com/TheBlock__/status/1924526880620753371) --- ## Inference Grid: Where Bitcoin, AI, and Money Collide - canonical: https://catalini.com/notes/inference-grid/ - original thread: https://x.com/ccatalini/status/1923388674479206540 - date: 2025-05-16 Today’s the day! Come tinker where Bitcoin, AI & money collide [@PresidioBitcoin](https://x.com/PresidioBitcoin) with resident AI wizard 🪄🧙 [@kevalexzhang](https://x.com/kevalexzhang). No seat? Comment below and we’ll spill the beans on Inference Grid—the skunkworks project we’re bolting together [@lightspark](https://x.com/lightspark) using [@buildonspark](https://x.com/buildonspark) & ⚡️ Inference Grid is a plug-and-play swap for the big models: smart routing trims the bill and lets you mix-and-match as many brains as you like (uncensored too). Or bring your own silicon—DIY compute nudges us closer to true decentralization, and dropping costs make it a viable alternative. Discover more at: [inferencegrid.ai](https://www.inferencegrid.ai) 🔧👷‍♀️👷‍♂️🚧 Alpha release: things will break—fine print says the AIs will probably fix their own bugs and send us an invoice. 🤖🦾 We crave your feedback! Let’s co-design a sturdier, more decentralized inference stack. The world needs it. 🔥 --- ## Fresh Asphalt Needs On-Ramps - canonical: https://catalini.com/notes/fresh-asphalt-needs-on-ramps/ - original thread: https://x.com/ccatalini/status/1920867394097950887 - date: 2025-05-09 Nothing eases a moderator’s job like panelists willing to puncture buzzwords on contact. Erica Khalili, [@chrismaurice](https://x.com/chrismaurice), [@dadiomov](https://x.com/dadiomov) & [@ShanAggarwal](https://x.com/ShanAggarwal) kept the crowd engaged while deflating the fluff. Thanks [@Lead_Bank](https://x.com/Lead_Bank) [@yellowcard_app](https://x.com/yellowcard_app) [@ModernTreasury](https://x.com/ModernTreasury) [@coinbase](https://x.com/coinbase) [@stripe](https://x.com/stripe). Stablecoins are laying fresh asphalt for global payments, but they need great on/off-ramps to the old highways—bank & card rails—, as well as consumer safeguards. Training wheels for borderless, self-driving programmable money/accounts… and those wheels are coming off fast. The stablecoin killer app might not be a shiny wallet at all—it could be the plumbing. Protocols like x402 let AI agents swipe wallet balances or cards in exchange for compute, data, and whatever’s next. Money that speaks API? Now that’s interesting. [x.com](https://x.com/CoinbaseDev/status/1919784224170889696) --- ## Electricity's Playbook for the Stablecoin Wars - canonical: https://catalini.com/notes/electricitys-playbook-stablecoin-wars/ - original thread: https://x.com/ccatalini/status/1919031578706989528 - date: 2025-05-04 Electricity’s playbook for today’s stablecoin wars: Westinghouse locked in the marquee distribution deals—Chicago World’s Fair, Niagara Falls—and used cheaper, easier-to-scale AC to undercut Edison. Secure the gateways and keep fees low, and the stablecoin crown is yours. Move fast while the sockets are still wobbly: once regulators stamp “60-hertz” on every coin, they will all trade like bulk electrons—reliable, boring, and priced by the kilowatt. Secure the distribution now. Win distribution first, then move up the value stack. Software still eats the world. But a well-armed phalanx of fintechs and legacy banks has a survival-level interest in stopping that. The real question, then, isn’t whether software will eat the world, but whose software gets the seat at the table. Hoard the upside and the network stalls. Share real ownership and it flies—Visa only took off after turning BankAmericard into a bank-owned co-op, giving thousands of banks skin in every swipe. More in the article! --- ## What Worldcoin's Pivot Signals for Crypto - canonical: https://catalini.com/notes/worldcoin-pivot-signals/ - original thread: https://x.com/ccatalini/status/1918009294177591701 - date: 2025-05-01 Yesterday, [@worldnetwork](https://x.com/worldnetwork)'s [@alexblania](https://x.com/alexblania) unveiled his latest plans to a room full of crypto insiders. The U.S. debut is noteworthy, but the real plot twist is Worldcoin’s sprint toward the mainstream and what that signals for crypto’s own leap to everyday commerce. What lessons can crypto founders and builders take from Worldcoin’s evolution? First, build real utility—then sweeten it with tokens. Worldcoin used to lean heavily on its crypto token to drive adoption. But that strategy—often hailed as Bitcoin’s winning formula and copied endlessly—gets the cause and effect all wrong. Bitcoin scaled because it introduced a neutral, fixed-supply asset answerable to no central bank. Sure, mining rewards and moon-shot price hopes drew speculators, later joined by institutional investors and a few sovereigns—but Bitcoin’s lasting builders were hooked on its radical potential as a new kind of asset and payment system, not just a get-rich-quick scheme. By pointing to dating, gaming, and credit—arenas where bots now mingle freely with humans—Blania framed Worldcoin’s proof-of-personhood as the fix, and explained why trading an iris scan to the orb might be a wager worth taking. It’s no shock that Worldcoin, co-founded and chaired by OpenAI’s Sam Altman, is tackling this issue: as AI grows more sophisticated, the need for a reliable, cryptographically secure way to prove someone’s human will become critical. [x.com](https://x.com/eddylazzarin/status/1917762475744977082) Second, delivering real utility outside the crypto bubble means matching the user experience consumers and businesses already get from traditional solutions. And yes, that requires building a bridge between old and new rails. [x.com](https://x.com/cuysheffield/status/1917777920703225922) You can’t skip the awkward stretch where new and old rails overlap—what [@aantonop](https://x.com/aantonop) calls "infrastructure inversion." Picture 56k modems hijacking analog phone lines in the ’90s or the first cars rattling down horse-grade gravel roads. Early Worldcoin tried to skip the inversion phase, making the token the star. Yesterday’s reboot flips the script: it leans into the infrastructure-inversion playbook and on shipping real utility first. You still can’t ship a best-in-class, truly global wallet without latching onto the old rails. On-ramps and off-ramps need to feel effortless—PayPal’s magic when online payments felt dicey—and that seamless flow is what any crypto wallet needs to crack the mainstream. That’s where the World App’s tie-ins with [@stripe](https://x.com/stripe) and a [@Visa](https://x.com/Visa) card stand out, delivering familiarity, trust, and instant utility from day one. Backwards compatibility also keeps incumbents in play, letting them track new players and roll out fresh services. Third, like any new technology, crypto is far from inevitable—no matter what its boosters claim. More precisely, decentralization—crypto’s core principle and its most significant contribution to reshaping markets—is far from a sure thing. --- ## Edison vs. Westinghouse, Stablecoin Edition - canonical: https://catalini.com/notes/edison-westinghouse-stablecoins/ - original thread: https://x.com/ccatalini/status/1917603682969084096 - date: 2025-04-30 Stablecoin players are staring at an Edison-vs-Westinghouse moment: seize distribution and scale into mainstream payments with new features now—or watch regulators standardize the grid and turn every digital dollar into commodity electrons. [@ForbesCrypto](https://x.com/ForbesCrypto) Only banks, fintechs, startups, and platform giants agile enough to execute the right open-payments play—fast and flawlessly—will plug into a web of interoperable assets and rails, staking their claim to an efficient, streamlined slice of the value chain. Unless today's issuers break into mainstream payments and finance faster than incumbents can mint their own coins, stablecoins will be relegated to background plumbing—fast, dependable, and as thrilling as the wire behind your drywall. In that world, the spoils go to those who own the outlets—wallets, apps, and merchant relationships—through which dollars flow. The stablecoin wars end when coins are unnoticed, only outlets matter. And many of those outlets, inconveniently, are already spoken for. Read more at [@Forbes](https://x.com/Forbes): [forbes.com](https://www.forbes.com/sites/christiancatalini/2025/04/29/the-stablecoin-wars/) --- ## The Stablecoin Wars: The Thread - canonical: https://catalini.com/notes/the-stablecoin-wars-thread/ - original thread: https://x.com/ccatalini/status/1917258653150896282 - date: 2025-04-29 Stablecoins were graduating from the playpen of crypto traders and DeFi “degens” to the main stage of mainstream payments. That fight is now on—albeit mostly off-camera—as a new wave of challengers readies its lines against incumbents Tether and Circle. With flawless execution, a determined stablecoin CEO could still pull off the Herculean feat of turning a plain dollar-jar into something unmistakably special... Short of that breakthrough, though, the industry faces a long slog of competition and fragmentation that could squeeze margins until they’re as thin—and as invisible—as your household electricity bill. The history of electricity provides a warning. In the 1880s, Edison’s direct-current glow from Pearl Street Station turned upscale Manhattan homes into status symbols, while Westinghouse’s alternating current mounted a fierce challenge in the “War of the Currents.” By the 1920s, AC’s scalability and standardized grids had prevailed, stripping electrons of their branding—utilities now competed on price, not panache. Once the meter started ticking, nobody cared who spun the turbines, only who did it cheapest. When it’s time to wrap dollar liabilities, banks and other heavyweight financial institutions will get the crisp ribbon; everyone else is left with scotch tape and yesterday’s newspaper. The logic is simple: the closer you are to the central-bank vault, the lower your cost of minting digital dollars. Stablecoins remain the one part of crypto still tethered to the very TradFi institutions purists dismiss. Yes, software still eats the world, and a savvy issuer could scale fast enough. But a well-armed phalanx of fintechs & banks has a survival-level interest in stopping that. The real question, then, isn’t whether software will eat the world, but whose software gets to do that. Whether stablecoins become distinct franchises or sink into commodity status will be decided by the strategies forged right now. To see why execution matters so much—and why commoditization remains the default—start with first principles and follow the money. At a high level, a stablecoin issuer has just two revenue levers. One is skimming the stock—retaining a slice of the yield from the reserve assets backing each coin. The other is taxing the flow—charging fees whenever the coins move. Both levers come with their own headaches. Strip away the jargon and stablecoins really do just two things: move money (medium of exchange) and hold money (store of value). In practice the two functions blur but treating them as separate buckets makes the market’s trajectory much clearer. Because there’s no single, always-on rail that links national instant-payment systems, payments-orchestration startups have improvised a real-time bridge: convert local fiat into a stablecoin, send the coin over a blockchain, then swap back to fiat on the other side. The double conversion looks clunky on paper, yet the money arrives instantly and settlement is final—so that’s the route some businesses now use for cross-border B2B payments, with a smaller but growing share in remittances too. Based on conversations with the payments teams actually running these "sandwich" corridors—and after stripping out crypto-native noise like high-frequency bots and other gymnastics—I estimate that roughly \$10 billion to \$30 billion a month already moves through this channel. The "stablecoin sandwich"—and its off-shoots in merchant checkout, gig-economy payouts, and more—is only the opening act. It’s classic infrastructure inversion: the clumsy stage where a new technology has to coexist with the old one. As businesses and consumers get comfortable, they’ll stop ducking in and out and start keeping balances in stablecoins. That’s when the second big job of stablecoins—the store-of-value role—steps into the spotlight. Stablecoins already function as onchain dollar vaults for people who lack easy access to greenbacks—a big reason Tether has exploded across Latin America, Africa, and parts of Asia, especially where hyperinflation burns through local money. But on the store-of-value front, stablecoins are about to face stiff competition from a growing stack of tokenized assets—onchain U.S. Treasuries, money-market funds, and whatever else Wall Street can wrap in code. Standalone issuers must race to carve out real share in everyday payments—fast—if they hope to keep their strategic edge. Fail to embed their coins in mainstream use cases, and the rest of the ecosystem will happily flatten them into just another commodity. Leading crypto exchanges and neobanks aren’t about to let one issuer park a tank on their front lawn. Coinbase just gave PayPal’s PYUSD marquee placement—right beside, but never above, its house favorite USDC. Robinhood and Kraken have meanwhile enlisted in Paxos’s Global Dollar Network (USDG), tossing yet another dollar-pegged contender into the arena. And industry watchers fully expect Revolut and Stripe to unveil coins of their own... ...a smart defensive move before the modular stablecoin stack lets rivals skate straight into their core business. For card networks the stakes are even higher—losing control of settlement simply isn’t an option. Circle, for its part, just introduced the Circle Payments Network (CPN)—an ambitious bid to build a full-stack rail that could one day rival the card networks themselves. If it blinks, nimble orchestration startups—Bridge (now tucked inside Stripe) and BVNK—will happily slot themselves between wallets and merchants and choose the winning coin for every flow. Banks will experiment with every shiny new rail, but their true goal is to keep their own tokenized dollars front-and-center—or, failing that, to claim a slice of the yield on the fiat reserves that stablecoin issuers have to park in their vaults. Large digital platforms won’t sit on the sidelines either. With distribution that spans billions of users, they can press issuers for better economics and fold stablecoins into their flows—all while turning payments into just another feature of a much larger ecosystem. Step back, and the outlines of the endgame come into focus. Only banks, fintechs, startups, and platform giants agile enough to execute this open-payments play—fast and flawlessly—will win. Unless today’s issuers can break into mainstream payments and finance faster than the incumbents can mint their own coins, stablecoins will be relegated to background plumbing—fast, dependable, and about as thrilling as the wire behind your drywall. In that world, the spoils won’t go to whoever mints the slickest dollar clone, but to whoever owns the outlets— the wallets, apps, and merchant relationships—through which those dollars ultimately flow. The stablecoin wars end when no one notices the coins anymore—only the outlets. And many of those outlets, inconveniently, are already spoken for. Full article on [@ForbesCrypto](https://x.com/ForbesCrypto): [forbes.com](https://www.forbes.com/sites/christiancatalini/2025/04/29/the-stablecoin-wars/) --- ## Crypto Adoption Is Taking Two Roads - canonical: https://catalini.com/notes/crypto-adoption-two-roads/ - original thread: https://x.com/ccatalini/status/1910344849444020725 - date: 2025-04-10 Something fascinating about crypto and payments is that adoption is quietly taking two different roads—which, conveniently, will eventually run into each other. (h/t [@cdixon](https://x.com/cdixon) [@a16zcrypto](https://x.com/a16zcrypto)) First, you've got stablecoins and Bitcoin quietly doing their thing behind the scenes, powering real-time global payments. Think of this as crypto's boring (but effective!) B2B job—bridging otherwise incompatible domestic payment rails. [x.com](https://x.com/HadickM/status/1910314025847496795) The consumer doesn't notice anything different, except suddenly cross-border payments settle instantly and cost less. Fewer middlemen, fewer headaches—magic, sort of! 🪄 Including on the most competitive cross-border corridor: 🇺🇸 <> 🇲🇽 [x.com](https://x.com/lightspark/status/1870127990715126205) --- ## What Keeps a Currency Dominant - canonical: https://catalini.com/notes/what-keeps-a-currency-dominant/ - original thread: https://x.com/ccatalini/status/1909283322934837587 - date: 2025-04-07 Historically, a currency’s global dominance isn’t just about size or military might.🪖 It’s about trust, stability, and forward-looking strategy. What could the US have done better to strengthen, rather than erode, its leadership? 🗽 💰First, carefully containing national debt—not by dramatic austerity, but through strategic and targeted cost-cutting. Fiscal discipline isn’t flashy, but economic history suggests it’s foundational to long-term stability. Next, aggressively investing in innovation, including advanced manufacturing. The shape of production 5 years from now will be radically different. Leading that transition—not reacting to it—is key to maintaining economic primacy. 🧑‍🎓🧑‍🏫 Education isn't just social policy; it's economic policy. Training the next generation's workforce is one of the best investments a nation can make. With AI rapidly advancing, preparing workers for the future becomes even more urgent. --- ## Bitcoin Is the Ultimate Libra Basket - canonical: https://catalini.com/notes/bitcoin-ultimate-libra-basket/ - original thread: https://x.com/ccatalini/status/1898571693913727029 - date: 2025-03-09 There is no need to reboot Libra, because Bitcoin is the ultimate Libra basket. [x.com](https://twitter.com/davidmarcus/status/1898561942324703668) We once believed we could engineer a flawless currency weighting scheme—balancing store-of-value with low volatility for everyone. Bold, elegant, and ultimately naïve! In truth, governing those weights is as mythical as achieving true decentralization with stablecoins. Like the internet, Bitcoin’s adoption is a messy orchestra of choices. Individuals, firms, and even nations decide how much to participate, crafting decentralized, permissionless, but perfect Libra weights. --- ## Decentralization Theater - canonical: https://catalini.com/notes/decentralization-theater/ - original thread: https://x.com/ccatalini/status/1896586283407077824 - date: 2025-03-03 Decentralization theater is crypto’s oldest magic trick: networks with puppet masters backstage imitate the look of true peer-to-peer systems, all to rake in cash. It’s been a grift for over a decade, and the audience still claps. These DINOs—decentralized in name only—are just dressed-up databases. They blend in with real networks to attract funding and grow fast. Investors toss money at the costume party, dazzled by the masks. Here’s the kicker: DINOs often sprint ahead early because they’re centralized. No messy consensus debates—just a CEO barking orders. Speed’s their edge, until you realize it’s a feature, not a bug, of the old world. But the governance? The limits? Pure vanilla centralization with a blockchain cherry on top. Innovative? Not really—unless you count the scale of the hustle. A strategic Bitcoin reserve is already a weak case. For DINOs, it makes no sense—they’re assets pretending to be something they’re not. Forget hoarding crypto, BTC or otherwise. The U.S. should play to its strength: set clear rules, unleash the good actors, and flush out the DINOs. Growth comes from refereeing the game, not joining the pile-on. --- ## Slowly, Then Suddenly: Programmable Dollars - canonical: https://catalini.com/notes/slowly-then-suddenly-programmable-dollars/ - original thread: https://x.com/ccatalini/status/1895171155776118851 - date: 2025-02-27 One day you’re rummaging for paper bills; the next, your wallet is lines of code. It’s the classic "slowly, then suddenly" momentum toward digital, programmable dollars—and [@patrickc](https://x.com/patrickc)’s annual letter offers a timely preview of what’s to come: [x.com](https://x.com/patrickc/status/1895097871369937250)) Use cases are still niche, but each one illustrates the growing value of the open infrastructure—and signals how quickly it’s moving from theory to practice. [@venturedan](https://x.com/venturedan): "we have seen numerous, relatively young startups (months since launch) which are scaling at an astronomical pace and already processing very large volume across different layers of the ecosystem." --- ## First Principles and the Colossus Sprint - canonical: https://catalini.com/notes/first-principles-colossus/ - original thread: https://x.com/ccatalini/status/1891864076105425257 - date: 2025-02-18 Attacking problems from first principles is powerful— [@OpenAI](https://x.com/OpenAI)’s old edge? Scale and talent. Then Colossus hit, built in record speed, with [@nvidia](https://x.com/nvidia) prioritizing [@xai](https://x.com/xai)’s orders and fancy networking demands. [x.com](https://x.com/SawyerMerritt/status/1891719655301267599) [@MichaelDell](https://x.com/MichaelDell) and [@charlesliang](https://x.com/charlesliang) racing to juice up [@xai](https://x.com/xai) with liquid-cooled racks in record time. Nothing like an AI boom to make CEOs personally oversee the plumbing. Just don’t ask about the margins—or the water bill. [x.com](https://x.com/Supermicro_SMCI/status/1850915433073025140) But really, the rest of the secret sauce is the brain trust: [@ibab](https://x.com/ibab), [@makro_ai](https://x.com/makro_ai), [@Yuhu_ai_](https://x.com/Yuhu_ai_), [@ChrSzegedy](https://x.com/ChrSzegedy), [@jimmybajimmyba](https://x.com/jimmybajimmyba), [@TobyPhln](https://x.com/TobyPhln), [@rpoo](https://x.com/rpoo), [@TheGregYang](https://x.com/TheGregYang), [@Guodzh](https://x.com/Guodzh), and more, all with a nudge from [@Tesla_AI](https://x.com/Tesla_AI). That's how you cook up self-improving magic. [x.com](https://x.com/Tesla_AI/status/1884458086859170050) --- ## Beyond Stockpiling Bitcoin - canonical: https://catalini.com/notes/beyond-stockpiling-bitcoin/ - original thread: https://x.com/ccatalini/status/1887607073736958439 - date: 2025-02-06 Instead of merely stockpiling Bitcoin, the U.S. must overhaul its financial architecture to prepare for—and ultimately succeed in—open networks. The upside is a more transparent financial system in which the U.S. continues to leverage its most powerful asset: the dollar. Similar to how tech leaders open-source critical components to establish industry norms while monetizing other areas... ...the U.S. can expand its dollar platform and ensure seamless interoperability with Bitcoin and stablecoins. With its extensive expertise in platform competition, the current administration is uniquely positioned to take this bet. --- ## Strategy, Not Speculation - canonical: https://catalini.com/notes/strategy-not-speculation/ - original thread: https://x.com/ccatalini/status/1886779214801821964 - date: 2025-02-04 Leading with strategy, not speculation—The best way for the U.S. to lower its debt-to-GDP ratio isn’t through speculation, it’s through fiscal discipline (e.g. [@DOGE](https://x.com/DOGE)) and economic growth. History is clear: Reserve currencies don’t last forever, and those that fall do so from economic mismanagement and overextension. To avoid joining the fate of the Spanish real de a ocho, the Dutch guilder, the French livre, and the British pound, the United States must focus on sustainable economic strength, not risky financial bets. If Bitcoin were to become the global reserve currency, the U.S. would have the most to lose. There is no smooth transition from dollar dominance to a Bitcoin-based system. Some argue that Bitcoin’s appreciation could help the United States "repay" its debt, but the reality would be far harsher. Such a shift would make it exponentially more difficult for the United States to finance its obligations and sustain its economic influence. And while many dismiss the idea of Bitcoin ever becoming a true medium of exchange and unit of account, history suggests otherwise. Gold and silver weren’t just valuable because they were scarce, they were also divisible, durable, and portable, making them effective currencies — even without sovereign backing or issuance, much like Bitcoin today. Similarly, China’s early paper money didn’t begin as a government-imposed medium of exchange. It evolved from commercial promissory notes and deposit certificates — representations of already trusted stores of value — before gaining wider acceptance as a medium of exchange. Fiat currencies are often seen as an exception to this pattern — declared legal tender by the government, they function immediately as a medium of exchange and later as a store of value. But this oversimplifies reality. What gives fiat money its strength isn’t just legal decree, it’s the government’s ability to enforce taxation and its capacity to meet debt obligations through this power. A currency backed by a state with a strong tax base has intrinsic demand because businesses and individuals need it to settle obligations. This taxation power is what enables fiat currencies to retain value, even without a direct commodity backing. But even fiat systems weren’t built from scratch. Historically, their credibility was bootstrapped from commodities people already trusted, most notably gold. Paper money gained acceptance precisely because it was once redeemable for gold or silver. The transition to pure fiat only became viable after decades of that trust being reinforced. Bitcoin is following a similar trajectory. Today, it is primarily seen as a store of value — volatile, yet increasingly regarded as digital gold. However, as adoption expands and financial infrastructure matures, its role as a medium of exchange will likely follow. History suggests that once an asset is widely recognized as a reliable store of value, the transition to a functioning currency is a natural progression.