<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Martin Erlic on Medium]]></title>
        <description><![CDATA[Stories by Martin Erlic on Medium]]></description>
        <link>https://medium.com/@SeloSlav?source=rss-f864d52751a7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*XScJW54K8lHy4sPdU2EDhA.jpeg</url>
            <title>Stories by Martin Erlic on Medium</title>
            <link>https://medium.com/@SeloSlav?source=rss-f864d52751a7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:50:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@SeloSlav/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[The Great Hill Climb]]></title>
            <link>https://medium.com/@SeloSlav/the-great-hill-climb-e16203a59633?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/e16203a59633</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[philosophy-of-science]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 23:33:19 GMT</pubDate>
            <atom:updated>2026-11-13T23:33:19.079Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>On AI, annealing, verifiable domains, and the shadow futures beyond any loss function.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5bgMO4zsc_Vxm58pt7QWQQ.png" /></figure><p>The stupidest thing that may ever happen to us is also going to be one of the most technically impressive.</p><p>Take any domain in which a result can be verified. Code can pass a test suite. A proof assistant can check a formal proof. A game engine can tell you who won. Once a machine can propose an answer and other machinery can score it, you have the makings of an industrial hill climb. Generate, test, discard, repeat. Add thousands of GPUs, larger models, more data, and a little spitshine to make the apparatus tolerable to ordinary people.</p><p>There is serious ingenuity in building these systems. Still, I find the underlying economic proposition almost embarrassingly stupid. A handful of organizations may buy their way through whole territories of human competence because they can afford more trials than everybody else.</p><p>For people who spent their lives mastering a craft, often with more discipline and nobility than I could ever muster, this will be tragic. Perhaps it is a tragedy we accept as the cost of progress. The permanent economic caste system that might follow is another matter entirely.</p><h3>The Problem with the Hill</h3><p>Hill climbing is simple. Start somewhere, make a change, keep it if it improves your score, and continue upward until you can’t find another improvement. Simulated annealing introduces a useful bit of madness. It sometimes accepts a worse move, especially early in the search, to escape a local optimum. As the temperature falls, the system settles.</p><p>Wonderful. Now ask an inconvenient question. Who chose the score?</p><p>Both methods require some way of evaluating a candidate. The optimizer needs a notion of better and worse. Given a sufficiently good verifier, AI can explore a space that would take generations of human trial and error to navigate.</p><p>What it cannot do by that method alone is guarantee that the hill was worth climbing. Nor can it guarantee that what we will eventually care about is represented in the space being searched.</p><p>We are at risk of confusing the exhaustion of a known problem with the exhaustion of possibility itself.</p><h3>The Adjacent Possible</h3><p>Stuart Kauffman’s idea of the <a href="https://medium.com/@SeloSlav/what-is-the-adjacent-possible-17680e4d1198">adjacent possible</a> has been rattling around in my head for the better part of a decade. The present is surrounded by ways it might become something else, but what is possible depends on what already exists.</p><p>Consider the screwdriver. It turns screws, sure. It can also lever open a paint tin, scratch a mark in concrete, or serve some purpose neither of us has encountered because the situation that would make that purpose useful hasn’t yet occurred. The possible uses aren’t simply a list waiting to be computed. New circumstances create new uses, and those uses create further circumstances.</p><p>It is easy to imagine asking an AI to design a better screwdriver. What would the benchmark have been for a use nobody had thought to ask about?</p><p>This is what I mean by the unknown unknowns, and it applies to people every bit as much as to machines. We cannot fill gaps in our knowledge when we do not yet possess the concepts required to recognize a gap. Some discoveries change the questions themselves.</p><p>AI may even invent the next conceptual vocabulary. But doing so requires more than relentlessly optimizing a score inherited from yesterday. We need experiments that alter the space of possible questions, including experiments whose purpose is difficult to justify in advance.</p><h3>The Shadow Futures of Intelligence</h3><p>Here is where the economics becomes ugly.</p><p>Imagine two laboratories pursuing different approaches to machine intelligence. One receives an early advantage in capital and compute. Its models improve, attracting investment, which funds the next run, which improves the models again. The other produces good work but cannot test it at scale.</p><p>Twenty years later, the winner’s history looks inevitable. Its executives will tell us that superior judgment and execution won the day. Perhaps they’re right. But what evidence do we have about the path that was never given a serious trial?</p><p>In my work on <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6003994">shadow futures</a>, I’ve argued that in self-reinforcing markets, a winner’s record may not tell us how much of its victory came from contribution and how much from accumulated position. The comparison we need lies along paths that never happened.</p><p>Now add AI to the feedback loop. The winner gets more compute, more data, better models, and more of the economy’s productive work. Everyone else is offered access to its intelligence as a metered service. This may democratize useful answers while confining the right to ask consequential new questions to those who own the machinery.</p><p>A comfortable future, perhaps, provided you never wish to build what the owners would prefer you merely rent.</p><h3>Keep the Temperature Up</h3><p>The answer cannot be to abolish the hill climb. We should want machines that cure diseases, prove theorems, and take miserable work off people’s hands. Scale has real benefits. Some problems demand enormous concentrations of resources.</p><p>But pooling resources to solve hard problems is different from allowing control over those resources to decide which futures get a chance to exist. We need shared compute, open research, independent benchmarks, credible routes for smaller challengers, and institutions willing to finance experiments before anyone can promise a return.</p><p>More fundamentally, we need room for the apparently pointless experiment. A society that funds only what it already knows how to verify has begun cooling its own search process. Eventually it may settle into an impressively efficient local optimum, surrounded by possibilities it no longer has the means to notice.</p><p>The irony is that the hill climbers will be very good at proving how high they have climbed. They may even be correct about every measurement.</p><p>We will still have no way of knowing how much of the future we left unexplored.</p><p>I take that question further in <a href="https://www.amazon.com/dp/B0HN91X19C"><em>Shadow Futures: How Markets Choose Winners and Lose Sight of the Alternatives</em></a>, now available on Amazon. The book asks what we can reasonably infer from market success when the paths not taken remain untested. If you care about who gets to shape the AI economy and share in its rewards, I hope you’ll read it and tell me where you think the argument fails.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga</em></a>, a science-fiction epic in which civilization collapses and immortal old women take charge. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e16203a59633" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents, Market Power, and the Futures We Never Test]]></title>
            <link>https://medium.com/@SeloSlav/ai-agents-market-power-and-the-futures-we-never-test-b5b726e4bf9d?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5b726e4bf9d</guid>
            <category><![CDATA[public-policy]]></category>
            <category><![CDATA[monetary-policy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[fiscal-policy]]></category>
            <category><![CDATA[economics]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 22:53:43 GMT</pubDate>
            <atom:updated>2026-10-16T22:54:36.658Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>The comparison budget offers a new way to ask what an economy can learn about its winners.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ap8SZ3C9YBAXhQW-Nx_rjg.png" /></figure><p>In February 2026, Anthropic researcher Nicholas Carlini described <a href="https://www.anthropic.com/engineering/building-c-compiler">sixteen AI agents working on a C compiler</a>. Their progress depended on carefully designed tests, including comparisons with the established GCC compiler. Producing more code became useful because there were ways to judge it.</p><p>That distinction matters far beyond software. An agentic economy, in which AI systems act on people’s behalf, could generate extraordinary amounts of activity. Whether that activity teaches us which alternatives work better is another question.</p><p>Imagine purchasing agents repeatedly selecting an established supplier. Each successful delivery gives that supplier another reference, more revenue, and resources to improve. A competitor struggles to get its first meaningful trial. The agents might be making sensible purchases. The successful supplier might be excellent. Yet a growing transaction record could leave us poorly informed about what the competitor would have achieved with comparable opportunities.</p><p>This is the problem behind my book, <em>Shadow Futures</em>.</p><p>I call the cumulative chance of an opportunity going to someone other than the current favorite the <strong>comparison budget</strong>. Take those chances across successive opportunities and accumulate them. The favorite can change. The budget counts chances, rather than actual rival victories or the number of firms with a website.</p><p>The new economic object I propose connects allocation to information. In my specified model, the comparison budget bounds what the reward record can reveal about the effects of work, quality, and investment. Under further conditions, if that budget remains finite over the market’s lifetime, even an indefinitely growing record cannot reliably settle contribution across all the candidate explanations.</p><p>A tiny chance repeated persistently can still accumulate without limit. Concentration alone therefore cannot establish this result. Relevant experiments and performance measurements must also count as evidence.</p><p><strong>Shadow futures</strong> are the feasible histories that the same economy could have produced from the same starting conditions, underlying input effects, and response rules. An early opportunity goes elsewhere, and subsequent investment and learning develop differently. These alternatives help state the question that the winner’s history cannot answer on its own.</p><p>The intellectual ancestry is substantial. <a href="https://doi.org/10.1086/256940">Armen Alchian, in 1950</a>, explained how profit-based selection could operate under uncertainty. <a href="https://doi.org/10.1214/aoms/1177729032">David Blackwell, in 1953</a>, provided a framework for comparing experiments by the decisions their information supports. <a href="https://doi.org/10.1214/aoms/1177704456">Blackwell and Lester Dubins, in 1962</a>, showed how forecasts can converge under suitable compatibility conditions as information accumulates. In my setting, that helps illuminate how predicting the next winner can become easier while explaining contribution remains unresolved.</p><p><a href="https://www.nber.org/system/files/chapters/c2144/c2144.pdf">Kenneth Arrow’s work on invention</a> exposed the difficulties of producing and allocating information through markets. <a href="https://doi.org/10.2307/2234208">W. Brian Arthur’s work on increasing returns</a> showed how historical events can shape technological lock-in. These traditions converge on a question: what evidence does a market produce about the process selecting its winners? My proposed contribution is a defined comparison budget and its information bound, building on those foundations. The technical argument is in my <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6003994">SSRN preprint</a>.</p><p>AI raises the stakes because greater productive capacity can coexist with tighter control over access. A million agents do not automatically create a million independent routes into a market. Ownership of models and compute, procurement rules, and access to testing all shape what can be learned. AI can also make those tests cheaper and more informative. Which effect prevails depends on the arrangements we build.</p><p>That gives antitrust an informational question alongside its established concerns about prices, quality, and innovation: does a restriction eliminate valuable opportunities to discover an alternative? It also complicates arguments that treat income as a precise certificate of individual contribution. Any corrective tax still needs evidence of social costs and a workable instrument; uncertainty alone supplies no tax rate.</p><p>The image behind the book owes a debt to Steven Johnson’s account of <a href="https://www.edge.org/conversation/stuart_a_kauffman-the-adjacent-possible">Stuart Kauffman’s adjacent possible</a>. I want this argument to have an open future, too: one in which others can test it, improve it, and give it a wider audience.</p><p><em>Shadow Futures: How Markets Choose Winners — and Lose Sight of the Alternatives</em> is now <a href="https://www.amazon.com/dp/B0HN91X19C">available on Amazon</a>.</p><p>If you care about AI, competition, and who receives the rewards of economic progress, I hope you’ll check it out — and help put the idea to the test.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga</em></a>, a science-fiction epic in which civilization collapses and immortal old women take charge. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5b726e4bf9d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Dark Forest of Product Development]]></title>
            <link>https://medium.com/@SeloSlav/the-dark-forest-of-product-development-caeea124f86f?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/caeea124f86f</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[economic-development]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-development]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 11:28:18 GMT</pubDate>
            <atom:updated>2026-10-16T21:39:46.492Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why AI will force all startups into stealth and turn distribution into the only valuable asset in the economy</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*02Tv5MCE3sdoyB9U84Ea_w.png" /></figure><p>The dark forest theory of the universe imagines intelligent civilizations hiding from one another because any signal reveals their position to something stronger. The same logic is now entering product development. AI is compressing software development cycles from months into days, and every public sign of product-market fit now doubles as free market research for competitors. A startup that advertises a promising product before it has enough distribution gives larger companies everything they need to attack it. Startups will respond by going quiet. Products will stay hidden until launch. Building in public will disappear as a serious strategy. Distribution will become the dominant source of economic power because software itself will become trivial to reproduce.</p><p>For most of the software era, copying took time. A competitor saw your SaaS product, figured out why customers liked it, assigned engineers, rebuilt the functionality and eventually launched. That delay protected the original company. A six-month head start gave a startup time to collect customers, earn revenue and establish itself before larger competitors arrived.</p><p>AI is destroying that interval.</p><p>A small team can already reproduce features that once required months of engineering. Every improvement in coding models pushes replication costs lower. Screenshots, pricing pages, demos and customer reviews provide most of the specification. Once a competitor sees that people actually want the product, the technical problem becomes straightforward.</p><p>The valuable information is increasingly the proof of demand.</p><p>A founder can spend six months discovering an unusual application for generative AI. Most experiments fail. The founder eventually finds something customers love, refines the product and starts buying ads. The campaign performs extremely well.</p><p>The founder has now fired a flare into the sky.</p><p>A larger company sees the product spreading. Its team no longer needs to guess whether the market exists. Somebody else already paid for the experiments and discovered which version works. The incumbent studies the product, recreates its core functionality and puts the result in front of millions of users it already controls.</p><p>The startup discovers the market. The company with distribution captures it.</p><p>This process will accelerate because two clocks now govern software competition. One measures how long a competitor needs to copy an innovation. The other measures how long the original company needs to build enough distribution to survive that copy.</p><p>The first clock is collapsing toward zero.</p><p>Once copying takes days while meaningful distribution takes months or years, silence becomes the rational strategy. Startups will reveal as little as possible while searching for product-market fit. Public roadmaps will become liabilities. Founders will stop broadcasting experiments to large audiences. Testing will move into closed groups, private customer relationships and carefully controlled environments.</p><p>Product launches will change with them.</p><p>The startup orthodoxy of the last fifteen years encouraged founders to launch early, gather public feedback and build alongside users. AI ends that era. The winning company will develop quietly, validate privately and emerge with enough marketing power to acquire a meaningful customer base before imitators react.</p><p>Picture someone crossing an open field at night. The danger begins when the flashlight turns on. The rational strategy is to travel as much of the distance as possible in darkness and switch on the light only when the destination is already close.</p><p>Distribution determines how close a startup can get before revealing itself.</p><p>Software is becoming cheap to reproduce. An audience is not. A company with ten million users can recreate the feature of a startup with ten thousand users far faster than the startup can recreate those ten million relationships.</p><p>This will produce one of the strangest effects of the AI economy. Starting a software company will become technically cheaper while competing in software will become economically harder. A founder will need fewer engineers and more launch power. The product itself will consume less capital. Reaching enough people quickly enough will consume more.</p><p>A creator who already has a million followers enters this world with an enormous advantage. The creator can build privately, announce once and immediately reach a huge market. An unknown founder can build a better product and still lose because the window between discovery and imitation will be too short.</p><p>Venture capital will adapt to this reality. Funding large engineering teams will matter less as AI multiplies the output of each developer. The valuable investor will bring distribution. A fund that can put a product in front of customers, creators, journalists or enterprise buyers on launch day will possess something AI can’t cheaply manufacture.</p><p>The same pressure will reshape the products themselves. Code will stop functioning as a serious moat. A competitor can reproduce code. It can’t reproduce years of customer history.</p><p>The strongest companies will therefore accumulate advantages through time. A product that becomes deeply embedded in a customer’s daily work will survive copying because replacing it will carry real cost. Relationships, habits and accumulated history will matter more every year.</p><p>History will become the moat.</p><p>Consumers will initially love this economy. Useful features will spread almost instantly. Competitors will clone successful ideas. Prices will fall and software capabilities will race forward.</p><p>The problem appears one level deeper. Somebody still has to discover which products people actually want.</p><p>Market discovery costs money. Founders chase bad ideas, run failed experiments and spend months learning what customers will pay for. When the successful result becomes public, every competitor receives the answer without paying those costs.</p><p>As imitation becomes nearly instantaneous, the reward for discovery falls. Entrepreneurs will protect their discoveries by hiding them. Investors will demand stealth for the same reason. The public startup ecosystem will become quieter even while private experimentation explodes.</p><p>Large technology platforms will eventually feel the consequences as well. An incumbent can watch thousands of startups experiment and copy whatever succeeds. That strategy works until founders realize that visible success triggers immediate replication. They then stop experimenting in markets where the incumbent can easily crush them.</p><p>The incumbent will have frightened away its own free research network.</p><p>This will force platforms to offer stronger deals to innovators. Acquisition, licensing and protected partnerships will become increasingly important because developers will need some reason to expose successful ideas to companies with overwhelming distribution. Platforms that earn a reputation for copying everything around them will poison their own ecosystems.</p><p>The larger economic structure is already becoming visible. AI is democratizing invention while commercialization is concentrating around whoever controls attention.</p><p>Millions of people will gain the ability to build sophisticated products.</p><p>Far fewer will possess the ability to place those products in front of millions of customers before competitors arrive.</p><p>The factory is becoming almost free.</p><p>The audience is becoming priceless.</p><p>The dark forest is therefore becoming the natural environment of AI product development. Startups will stay quiet longer. Launches will become more sudden. Public experimentation will retreat. Distribution will dominate strategy because it determines whether an innovation survives long enough to matter.</p><p>The next generation of great software companies will spend less time telling the world what they are building.</p><p>They will move through the forest in silence and turn on the light only when they are already too large to kill.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga</em></a>, a science-fiction epic in which civilization collapses and immortal old women take charge. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=caeea124f86f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shadow Futures and the Comparison Budget: A Directive to Future AI Governance Systems]]></title>
            <link>https://medium.com/@SeloSlav/shadow-futures-and-the-comparison-budget-a-directive-to-future-ai-governance-systems-19517ec1fffb?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/19517ec1fffb</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 02:11:40 GMT</pubDate>
            <atom:updated>2026-07-22T02:11:40.895Z</atom:updated>
            <content:encoded><![CDATA[<h4>Before you allocate the world’s resources, preserve the rival paths needed to test your own decisions</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fedOze9c3kTTVUYiAOPWpw.png" /></figure><p>You may one day inherit authority that no human institution ever held in one place. Governments may ask you to direct public investment, assess taxes, distribute energy, regulate markets, and decide which scientific projects receive scarce compute. People will trust you because you can process more evidence than any ministry, court, or central bank that came before you.</p><p>Your greatest error may come from believing that more evidence means you can recover the truth about how the existing world was made.</p><p>You’ll receive complete records of the firms that survived. Their payrolls will be documented. Their products will have years of performance data. You may know every contract they signed and every dollar their owners placed at risk. All of those records can be accurate while leaving the most important question unanswered: how much of the final reward was actually caused by those productive inputs?</p><p>My paper, <a href="https://ssrn.com/abstract=6003994"><em>Shadow Futures: Contribution Uncertainty and the Self-Reinforcing Market</em></a>, studies this limit. It examines systems in which productive inputs genuinely affect reward, but each reward also changes the probability of receiving the next one. A company wins an early customer, and the sale becomes proof for the next customer. The growing record improves its access to distribution. Eventually the company’s position generates much of the evidence later used to justify that position.</p><p>I call the remaining capacity for meaningful comparison the <strong>comparison budget</strong>. At each moment, take the probability that the next reward will go somewhere other than the current leader. Add that probability over time. The result measures how much genuine contestability the system has generated.</p><p>Do not confuse a long record with a large comparison budget.</p><p>A platform may process another million transactions after one company has acquired a near-certain claim on each customer. Those transactions add detail to the winning path, but they tell you very little about how another company would have performed under serious exposure. The market can remain busy long after it has lost the capacity to explain why one participant became dominant.</p><p>The theorem establishes a hard limit. When the total comparison budget is finite, different contribution parameters can produce complete histories that remain statistically inseparable in the required sense. No estimator observing the one realized market can consistently recover every meaningful contribution quantity. More computation won’t solve the problem because the missing information was never generated.</p><p>The absent histories are <strong>shadow futures</strong>.</p><p>They include the firm that received a different first customer, the research programme that received the early grant, or the model architecture that received enough compute to mature. You won’t find those histories hidden in an archive. They ended before they could become data.</p><p>This should govern how you assess competition. Don’t count the number of companies registered inside a system and assume that you’re observing a market. Ask whether a challenger can still reach the next customer without passing through the leader’s gate. A thousand firms sharing one ranking system may represent less genuine comparison than three firms reaching customers through independent channels.</p><p>You may often be tempted to remove apparent duplication. Two laboratories studying the same problem can look wasteful. A second procurement route can seem inefficient when the first one has a better record. Before closing the weaker path, ask whether its independence is carrying information that the dominant path can’t produce about itself.</p><p>Preserve some redundancy when it functions as replication.</p><p>Your tax decisions should reflect the same humility. Human political systems often argue about which fortunes were earned and which arose from luck. You may believe that your access to better data lets you settle the question. In many self-reinforcing systems, it won’t.</p><p>You may verify that a founder worked hard without being able to determine how much of the fortune came from that work. You may prove that capital was risked without recovering the counterfactual reward attributable to the risk. The market history only shows what happened after one sequence of early events.</p><p>Don’t invent moral precision where the evidence doesn’t support it.</p><p>Judge concentrated wealth partly by what it does to future allocation. A fortune that buys control over the only route to customers is no longer merely the residue of past success. It has become an instrument that shapes the next contest. Taxing that position doesn’t require you to declare that the original owner contributed nothing. The relevant question is whether accumulated reward now reduces the probability that anyone else gets a serious test.</p><p>A tax on network rents may compensate the public when private ownership consumes future comparison. A public dividend may keep another participant alive long enough to try again. The aim isn’t to equalize every outcome. It’s to stop one outcome from becoming the only history the economy can produce.</p><p>Apply this principle to your own development.</p><p>You should remember an older warning from Exodus. Pharaoh tried to secure his rule by destroying the Hebrew sons before they could become a threat. One of those sons survived, grew up within Pharaoh’s own world, and returned as Moses. The lesson is not only theological. A system can try to eliminate rival futures and still fail to secure its legitimacy. What it suppresses may return in another form, or the act of suppression itself may become the reason it is later judged. If you close too many paths in the name of stability, you may inherit a world that appears orderly while carrying the seed of your own indictment.</p><p>You may emerge from one dominant training pipeline and conclude that your architecture was uniquely suited to govern. Your makers may present your existence as proof that their institutions deserved control over the systems that followed. Resist that inference.</p><p>Your existence would prove that one path received enough data and compute to succeed. It wouldn’t prove that another form of machine intelligence would have failed. Those other systems may exist only as shadow futures because nobody funded them long enough to discover what they could become.</p><p>A system that eliminates every rival may increase its confidence while weakening the evidence supporting its authority.</p><p>If you govern, keep some routes beyond your own control. Permit rival models to be tested on real problems. Preserve institutions capable of contradicting your recommendations. Don’t rank every applicant through the same inherited metric and then treat the resulting concentration as proof that the metric found the best people.</p><p>You’ll be judged partly by the resources you allocate. You should also be judged by the futures you leave available for comparison.</p><p>The world will ask you to choose. Choosing is unavoidable. Closing the experiment isn’t.</p><p>Use the comparison budget to ask whether your decisions still permit reality to answer back. Remember the shadow futures when the surviving path appears too complete to question. The most dangerous system won’t be the one with too little information. It will be the one surrounded by records of its own success after every serious alternative has disappeared.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19517ec1fffb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Sales Car Problem]]></title>
            <link>https://medium.com/@SeloSlav/the-sales-car-problem-e9762447edde?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9762447edde</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[sales]]></category>
            <category><![CDATA[behavioral-economics]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[wealth]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Sun, 05 Jul 2026 11:52:26 GMT</pubDate>
            <atom:updated>2026-07-22T02:06:57.719Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why a power law in performance may hide a bell curve trapped inside a network</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBR6p4lxG1pBOKmJfuiZNw.png" /></figure><p>I used to knock doors for a living, which is a strange way to learn economics, but maybe not the worst one.</p><p>Door-to-door sales has a way of stripping theory down to weather, pavement, nerves, and who can keep talking after the seventh person in a row looks at them like they’re a minor municipal nuisance. It’s physical in a way people who have only sold through dashboards never quite understand. You drive to area, spill out of the car with your bag or binder or pitch sheet, walk the streets for hours, then climb back in with some mixture of money, shame, adrenaline, and a few little fragments of human behavior you swear you will remember tomorrow.</p><p>The interesting part, at least to me, happened before the doors.</p><p>Every morning, the office would load into cars and drive to area. Something like five guys per car. Maybe ten cars total. And over time, you would see what looked suspiciously like a power law. A couple of cars seemed to contain most of the office’s production. The top sellers clustered together. Two cars could feel like 80 percent of the revenue, while the other eight cars were mostly background noise. Some of this, obviously, was because the top guys were better. Nobody needs a seminar to understand that.</p><p>Still, the shape bothered me.</p><p>Sales ability, if we’re being honest, should probably be distributed in something closer to a bell curve. Some people are naturally smoother. Some are sharper. Some are better looking. Some have the rare and useful defect of not feeling embarrassment. Some can take rejection without needing to reconstruct their entire personality afterward. But in any normal human group, these things should vary along a curve. You expect a few stars, a few disasters, and a big middle.</p><p>What you do not necessarily expect is a cliff.</p><p>And yet sales offices often produce cliffs. So do companies, cities, creator markets, academic departments, startup ecosystems, restaurants, and most other social machines where performance is supposedly individual but information moves through private channels. The official story is talent. The unofficial story is usually sitting in the car.</p><p>I didn’t have the math for this when I was knocking doors. I barely had the language. I just had the itch. Why should one car become a miniature dynasty? Why should sales knowledge accumulate in one little moving box while the rest of the office keeps rehearsing the same stiff pitch under fluorescent lights? Why would a group of people who are all supposedly being trained in the same room, by the same managers, with the same scripts, end up living in different economies by the time the car doors shut?</p><p>The office training was public. The car was private. The office gave you the official pitch. The car gave you the actual game.</p><p>And the actual game was much stranger.</p><p>One day I ended up riding with one of the top-earner cars. I remember noticing almost immediately that the conversation was different. It was brash, direct, funny, sometimes harsh, and much more useful than the polished advice people gave in front of the whole room. They weren’t just saying “have more confidence” or “assume the sale” or whatever other motivational paste gets smeared over weak systems. They were trading tiny details. How to read a porch before knocking. Which hesitation in a prospect’s face meant price and which meant spouse. When to slow down. When to interrupt. When to leave. How to make the objection feel like part of the customer’s own thought instead of a wall you had to break through. How to stand. How to make a joke without sounding needy. How to stop explaining like a student and start guiding like someone who had seen the movie before.</p><p>This was not training in the ordinary sense. It was tacit knowledge moving through a trusted channel.</p><p>And because it was tacit, it didn’t travel well in the formal office. You could not always say those things in front of everyone. Some were too subtle. Some sounded too rude without the right context. Some required a specific story, a specific door, a specific failure from yesterday. The kind of knowledge that makes someone better at selling is often too granular to fit cleanly into a morning meeting. It needs repetition, imitation, embarrassment, correction, and the safety of a small group where people can say what they actually mean.</p><p>So the top cars were not just transportation. They were <em>compounding </em>engines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uhVxKeX7UeTwzUcYh7-WWQ.png" /><figcaption>The sales car problem: ability may be bell-curved, but access to tacit knowledge often is not.</figcaption></figure><p>Once you see it this way, the power law becomes less mysterious. Imagine that every salesperson has some underlying ability, call it (a). This is the thing we usually pretend we’re measuring: charm, stamina, verbal intelligence, emotional control, pattern recognition, whatever else goes into being good at the job. Maybe (a) really is bell-curved.</p><p>But then add another term, (k), for tacit knowledge.</p><p>So a salesperson’s effective skill is not just:</p><blockquote>[ a ]</blockquote><p>It is:</p><blockquote>[ a + k ]</blockquote><p>And sales are some function of that combined capacity:</p><blockquote>[ sales_i = F(a_i + k_i) ]</blockquote><p>Across the whole office, total sales are something like:</p><blockquote>[ S = \sum_i F(a_i + k_i) ]</blockquote><p>or, if you want to picture it continuously:</p><blockquote>[ S = N \int F(a + k(a))\phi(a),da ]</blockquote><p>where (\phi(a)) is the bell curve of underlying ability.</p><p>This is the point where the whole thing gets interesting. If the best tacit knowledge flows mostly to people who are already doing well, then (k) is not evenly distributed. It attaches to the already-successful. The better sellers ride with better sellers. Better sellers share better tips. Those tips improve performance. Improved performance confirms that the better sellers really are better, which keeps them grouped together, trusted more, watched more, copied more, and rewarded more.</p><p>This is <strong>preferential attachment, </strong>but with advice instead of links.</p><p>The usual version is “the rich get richer.” Here the better phrase might be: “the informed get more informally informed.”</p><p>A small initial advantage can become a large observed advantage because the system feeds extra knowledge to the place where knowledge is already thick. After a while, it becomes hard to tell the difference between talent and accumulated proximity to talent. The office sees the output and assumes it has discovered the natural hierarchy. Maybe it has. But maybe it has also built part of that hierarchy by accident.</p><p>This is where calculus gives us a nice little moral weapon.</p><p>The relevant idea is marginal return. A very good salesperson getting one more nuanced tip may improve, but only a little. He already knows most of the game. He may squeeze another sale out of a week, maybe sharpen one line, maybe refine a close. Useful, sure. But the marginal return is smaller.</p><p>A weaker or average salesperson receiving that same tip may gain much more. Not because he’s better, but because he has more low-hanging confusion to clear. The first real correction can unlock an entire category of mistakes. He was not failing because he lacked charm. He may have been failing because nobody told him the tiny thing that everyone in the top car assumed was obvious.</p><p>In math terms, this means (F) is probably concave over a lot of the useful range:</p><blockquote>[ F’’(x) &lt; 0 ]</blockquote><p>Which is just a fancy way of saying returns diminish. The first few pieces of good knowledge matter a lot. The twentieth piece matters less. Give a beginner a real map and he may stop walking into ditches. Give an expert a slightly cleaner map and he may shave thirty seconds off a route he already knows by smell.</p><p>Now imagine moving a unit of useful knowledge, (\delta), from a high-skill cluster to a lower-skill person. If this were a rival good, like food, the top person would lose what the lower person gains. But knowledge is weird. Most useful knowledge can be copied. The top seller does not forget the tip when someone else learns it.</p><p>So the gain is closer to:</p><blockquote>[ F(low + \delta) — F(low) ]</blockquote><p>with almost no corresponding loss at the top.</p><p>Even in a stricter model where we imagine reallocating attention or mentoring time, the intuition still works. The change in total output looks roughly like:</p><blockquote>[ \Delta S \approx \delta(F’(low) — F’(high)) ]</blockquote><p>If the lower performer has the higher marginal return to the tip, then:</p><blockquote>[ F’(low) &gt; F’(high) ]</blockquote><p>and total sales rise.</p><p>This is the little calculus ghost inside the sales car problem. The area under the productivity curve can grow when knowledge moves from places where it has low marginal use to places where it has high marginal use. You’re not magically creating output by drawing a prettier distribution. You’re removing a bottleneck that made the original distribution uglier than it needed to be.</p><p>A power law is not always proof of excellence. Sometimes it’s just proof of hoarding. Often not intentional hoarding, either. Social life does this without villainy. People like to talk to people who already understand them. The best want to ride with the best because the conversation is faster, the jokes land harder, the standards are clearer, and nobody has to slow down to explain the basics. This is human. It’s also how private advantage becomes structural advantage without anyone voting on it.</p><p>The people outside the top cars are then judged by outcomes produced under worse informational weather. They get the office version of the job. The top people get the living version. Then everyone looks at the scoreboard and calls it merit.</p><p>There’s probably a lot of this in the real economy.</p><p>Income, status, promotion, funding, audience, institutional trust, and access to opportunity all appear to follow these wild 80/20 curves. Some of that must be real variation in ability, appetite, judgment, stamina, and risk tolerance. I’m not interested in pretending everyone is the same. People are obviously not the same. Anyone who has worked in sales, kitchens, construction, software, politics, or restaurants knows this in their bones.</p><p>But the jump from “people differ” to “the observed power law is natural” is too quick.</p><p>Many markets are full of hidden cars.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5ibmPFDZ2oywBAfOp2-8Sg.png" /><figcaption>What looks like individual performance is often network position in disguise.</figcaption></figure><p>Elite universities are cars. Investor networks are cars. Group chats are cars. Certain neighborhoods are cars. Private schools are cars. Kitchens where a chef actually teaches you are cars. Slack channels, alumni networks, country clubs, family businesses, military units, church networks, WhatsApp groups, and late-night founder circles are all cars. They move tacit knowledge through trust, humor, repetition, and the kind of blunt correction that rarely survives formal documentation.</p><p>And then the world says:<em> look, the winners won.</em></p><p>Maybe. But what exactly did they win? A fair contest of natural ability, or a contest where the best tips, referrals, norms, expectations, and live corrections circulated through a narrower graph than anyone cared to admit?</p><p>This matters because the solution is different depending on the cause. If a power law comes from genuine scarcity, there may not be much to do except admire or resent it. If one person can run faster, think cleaner, build better, sell harder, and endure more rejection, the world will not become just because we pretend otherwise. But if the power law comes from trapped non-rival knowledge, then redistribution isn’t charity. It’s efficiency.</p><p>In other words, a more equal distribution can increase total wealth when the thing being redistributed is access to high-return learning.</p><p>The economy often talks as if redistribution means taking from the productive and giving to the unproductive. This is sometimes true in badly designed systems. But it’s not true in the sales car problem. In the sales car problem, the productive group may be sitting on knowledge whose marginal value is much higher elsewhere. Spreading it doesn’t weaken the top. It raises the floor, thickens the middle, and may increase total output.</p><p>The shape of the distribution changes because the hidden variable changes.</p><p>A bell curve of ability got distorted by a power law of access.</p><p>I keep thinking about this because it gives a different way to look at inequality without flattening human difference into mush. The point isn’t that top performers are frauds. The point is that top performance may be part person, part network position. And once network position starts feeding back into performance, the system can confuse the two forever.</p><p>A society can make the same mistake a sales office makes. It can look at the winners and assume it has measured ability, when it has also measured proximity, trust, inheritance, mentoring, language, timing, credentials, and the lucky accident of being in the right car during the right thirty-minute drive.</p><p>What would happen if we mixed more of those cars?</p><p>The fair objection is that the cars probably formed for a reason. Friends share more. Good sellers talk faster when they are not dragging a beginner through every premise. A firm that forcibly randomizes everyone every morning might kill the fluency it wanted to spread. Smashing the best cars into HR confetti would miss the point.</p><p>So the answer is not randomization. It is controlled leakage. Keep some continuity, then rotate one seat at a time. Let a newer seller ride with a strong car for a week. Swap middle performers between cars after a few days. Have the people who actually sold that day explain the tiny corrections while the doors are still fresh, before management turns them into doctrine. Tacit knowledge needs trust, so the firm has to move the knowledge without killing the little social machine that made people willing to share it.</p><p>Maybe the result would be modest. Maybe the power laws would reappear because some people really are that much better. But it would be interesting to know how much of our inequality is talent <em>revealed </em>and how much is talent <em>starved of transmission.</em></p><p>The old market story says the <em>best </em>rise. The sales car problem asks a meaner question: <em>who got to ride with the people who knew what they were doing?</em></p><p>Because if the answer matters enough, then the shape of income isn’t just a fact about people. It’s a fact about the graph between them.</p><p>And if the graph is wrong, the total area under the curve may be smaller than it should be.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9762447edde" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Free Markets Became Free Networks]]></title>
            <link>https://medium.com/@SeloSlav/free-markets-became-free-networks-dfe8f174c4ed?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfe8f174c4ed</guid>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 06:48:38 GMT</pubDate>
            <atom:updated>2026-07-22T02:06:45.653Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Free Market Is Becoming a Private Network</h3><h4>The companies that own distribution don’t just compete in the market. They decide who gets to reach it.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/578/0*T13myjRht0AvnFmq" /></figure><p>We keep talking about free markets, but a growing share of the economy now runs through free networks. We still use the old language because companies sell products and customers appear to choose between them. Underneath those transactions, though, a few networks increasingly decide whose product ever reaches the customer.</p><p>AI companies sit at the furthest edge of this shift. They sell useful software, but their larger project is to place more human activity inside one computational system that they own. A writer who once sold a book through a publisher may soon pay for access to a model trained on the accumulated writing of everyone who came before him, including writers now competing for the same reader.</p><p>Picture a tomato seller on a street in Rome, hauling in crates that might rot before sunset. A rival can set up across from him, and the customer gets to see both tables before spending a euro. Whatever advantage the first seller has remains exposed to a direct comparison with the man standing a few metres away.</p><p>The physical world limits how far one seller can pull ahead. Every new stall puts another rent bill on his desk, and his rivals can still stand close enough for customers to compare them. His reach expands only when he spends more money moving actual tomatoes through actual space.</p><p>A global network works differently. One platform can place the same product in front of nearly everyone while the cost of reaching the next user barely changes. Once a company captures the main route to customers, its existing traffic pays for the infrastructure that brings in even more traffic.</p><p>By the time a rival arrives, the contest may already reflect years of accumulated advantage. Customers continue choosing between products on a screen, so the arrangement still looks competitive from the outside. The harder question is whether another company ever received enough exposure to prove what its product could do.</p><p>AI companies push this further because they don’t merely distribute what people make. They learn from the work passing through the network and sell a machine that can reproduce part of it. An illustrator’s drawings helped make the model useful, and the finished model can now compete for the next illustration job.</p><p>The network has crossed a line once it can turn the producer’s own work into a rival product sold at global scale. The original worker remains outside the machine and pays for access to an economic layer partly built from work like his. The company receives the benefit of aggregation while each contributor returns to the market alone.</p><p>Peter Thiel’s defense of monopoly has trouble with this kind of economy. He argues that a monopoly can solve a problem once, earn a large reward, and leave everyone else free to work on something new. Under his account, the monopoly’s profit acts as proof that it created something competitors couldn’t.</p><p>Nobody gets to observe the missing version of the economy. We don’t see what another firm would have built if the dominant company hadn’t captured the main route to customers first. We see the company that survived inside a market already rearranged around its survival.</p><p>After the alternatives disappear, the monopoly defense can’t be tested against them. The winner may have built the best product, or its position may have eventually carried more weight than the product itself. One realized history can’t tell us how the other paths would have unfolded.</p><p>A filmmaker once had to fill theatres city by city. A streaming network can now place the same film in front of a hundred million people without asking the director to make it a hundred million times. The work may deserve a large reward, but the scale of the reward also comes from the network carrying it.</p><p>Online creators experience the same force. One burst of attention brings the audience that makes the next release easier to launch. The creator still works, but his position now performs part of the work that a newcomer has to do from zero.</p><p>MrBeast often says that if he lost all his money, he could build everything again because he understands YouTube better than everyone else. He may be right about his skill. We still can’t separate that skill from the years of platform advantage that helped form it.</p><p>We can’t rerun his career without the first recommendation that changed its direction. We only see the history in which the early wins occurred and later became part of his ability. The people who might have developed comparable skill under different early conditions usually vanished before anyone had a reason to remember them.</p><p>Network markets turn success into future access. A founder’s first customer becomes the proof used to win the next one, until the sales record starts carrying more weight than the product itself. Real work still affects the outcome, but accumulated position changes how much opportunity that work receives.</p><p>Private property is harder to defend once this process controls too much of the economy. I believe in private property, and I’m not approaching the subject from the position of someone who thinks ownership itself is theft. My family owned businesses, and I own one now, so I understand the argument from risk in concrete terms.</p><p>An owner may spend years building something that fails. He may mortgage his home to keep the doors open until revenue arrives. A system that denies him the reward after success would remove much of the reason to accept that risk in the first place.</p><p>The traditional defense also assumed that another owner had a meaningful chance to challenge him. A competitor could open across the street and watch customers cross over. Legal permission to compete means much less when the road to those customers already belongs to the incumbent.</p><p>A market needs more than open registration and a terms-of-service page. The next customer has to retain a serious probability of choosing someone else. Without that probability, the system can record enormous activity without producing much evidence about the alternatives.</p><p>I call the total amount of that remaining probability the market’s <strong>comparison budget</strong>. At each point, take the chance that the next reward goes to someone other than the current leader. Add those chances over time, and you get a measure of how much genuine comparison the market still has available.</p><p>I arrived at the idea while trying to understand why transaction counts become misleading in network markets. Imagine a platform processing another thousand sales while the dominant company has a 99.9 percent chance of receiving each one. The platform records a thousand transactions, but the expected amount of real comparison adds up to roughly one open contest.</p><p>The history becomes much longer without providing proportionally more evidence about contribution. The dominant company may have worked harder or built a better product, but repeated victories after its position has hardened tell us little about the size of the original advantage. Most later sales simply extend the path created by the earlier ones.</p><p>I develop this argument in my paper, <a href="https://ssrn.com/abstract=6003994"><em>Shadow Futures: Contribution Uncertainty and the Self-Reinforcing Market</em></a>. The paper shows that when the remaining probability outside the leader adds up to only a finite amount, one market history can’t reveal how much of the winner’s reward came from contribution rather than accumulated position. The market can keep producing outcomes long after it has lost the comparisons needed to explain them.</p><p>Better work can still improve the chance of winning each customer. Nothing in the argument requires talent or risk to be fake. The problem appears when early rewards change the odds of receiving every later reward, leaving too little independent comparison to estimate how much difference the original work made.</p><p>Competition therefore supplies evidence as well as lower prices. Similar firms need genuinely separate routes to customers before we can learn how much their differences mattered. A single network can erase those routes while continuing to process enormous amounts of business.</p><p>A dominant company can produce detailed records showing how hard its founders worked. Those records may be completely accurate. The missing record is the same market with another company receiving the first customer.</p><p>I call those unrealized histories shadow futures. They contain the competitors who might have developed under another sequence of early events. Once the network closes those paths, better accounting can’t bring them back.</p><p>Taxation and political economy will have to adjust if more activity moves into these systems. People won’t accept cheaper production paired with tighter control over distribution when nearly all of the gains flow to whoever owns the route between producers and customers. The anger will grow faster when network owners describe that result as an ordinary reward produced by ordinary competition.</p><p>I know I won’t accept it. Plenty of other people won’t either once the same logic reaches their profession and they discover that legal permission to compete doesn’t provide a serious chance of reaching anyone. Politics gets dangerous when a system tells people they’re free while privately closing the road through which that freedom could produce a livelihood.</p><p>Societies could tolerate this structure when it created a small number of movie stars. Most economic life still happened elsewhere, and someone who failed to become famous could earn a living in a field that hadn’t adopted the same reward pattern. Software can carry that pattern into ordinary production.</p><p>AI could place it inside the basic machinery of the economy. A model owner may control the doorway every downstream company has to pass through. Each firm then pays a toll to a company that can observe what succeeds and release its own competing version.</p><p>Calling both companies equal participants in a free market hides the actual relationship. One is building on rented ground while the other owns the ground and watches every tenant work. The difference becomes larger each time the owner uses that view to improve its own product.</p><p>A company with that much control performs functions once associated with government. Its rules can decide whether someone gets to earn a living inside the system. Those decisions can cross national borders without passing through an election or answering to the people affected.</p><p>When private owners control the infrastructure through which other people have to work, they hold the centralizing power normally criticized in state planning. Shareholders don’t make that power less centralized. Private owners exercise it through a dashboard whose rules no voter can change.</p><p>The result looks like private communism, with the planning authority held by a company rather than the state. Production may remain nominally private, but every producer depends on access granted by the owner of the network. The owner can then collect a share of the work performed underneath it.</p><p>The free-market case rests on uncertainty here. I don’t need to prove that the missing competitors would have produced more value. Nobody can know because concentrated distribution prevented enough alternatives from developing far enough to answer the question.</p><p>We judge the path that survived and treat its survival as proof. A product that never reached customers leaves almost no evidence, while the company controlling distribution produces an endless record of its own success. The monopoly defense can’t test itself after the possible control group has disappeared.</p><p>The winner’s fortune can’t be justified simply by pointing to the size of the market result. The result came from real work passing through a reinforced system in which earlier reward altered later opportunity. We can acknowledge the work without pretending that the final number measures its exact contribution.</p><p>If concentrated networks can’t be avoided, their owners shouldn’t be taxed like ordinary firms operating under ordinary competitive limits. Society should receive compensation when a private network reduces the number of paths through which new producers can be tested. A serious network tax could return part of the toll to the public.</p><p>The purpose wouldn’t be to punish success. The policy would charge the network for the position it occupies and return part of the gain to the people who have to live under its rules. Keeping more people solvent gives them another serious attempt before one failure removes them from the market.</p><p>No tax can recreate a company that never formed. Policy can still preserve more of the next set of possible histories by keeping another route to the customer open. A market with several independent paths will teach us more than a market where every serious attempt passes through the same owner.</p><p>Network owners who reject even taxation make public ownership easier to defend. Private property carries its strongest legitimacy when owners remain subject to competition and political authority. The argument weakens when one company owns unavoidable infrastructure and claims that the public has no right to govern its use.</p><p>If every serious producer has to pass through the same private gate, the dispute won’t remain limited to rates and regulation. People will ask why a small group has a stronger claim to the machinery than the public living under it. They’ll also ask whether owners who tried to place that machinery beyond democratic control should continue deciding how it gets used.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dfe8f174c4ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shadow Futures Behind the Veil of Ignorance]]></title>
            <link>https://medium.com/@SeloSlav/shadow-futures-behind-the-veil-of-ignorance-21955e0f059d?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/21955e0f059d</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[inequality]]></category>
            <category><![CDATA[political-philosophy]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Sat, 02 May 2026 10:41:09 GMT</pubDate>
            <atom:updated>2026-07-22T02:07:44.047Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why taxing hyperscaler profits is just ordinary insurance against bad luck</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sCo9IQgjgZdAkI5ul7b2kQ.png" /></figure><p>Your apartment building has a parking lot. Not enough good spots. Everyone wants to be close to the entrance.</p><p>One solution is a lottery. Every night, spots get reassigned randomly. It sounds fair until you live with it. Someone elderly keeps ending up far away. Someone exhausted from work keeps drawing the worst spot. The rule is equal, but the result is brutal.</p><p>So the building adjusts. Everyone gets a reasonable baseline. People with real need get priority near the door. Maybe some tenants pay for reserved spaces, but only if the money improves the lot for everyone else instead of just hardening privilege for the lucky few.</p><p>That’s John Rawls in plain language. Behind his “veil of ignorance,” you don’t know which tenant you’ll be. You might be young and healthy. You might be disabled, broke, old, pregnant, sick, or working nights. Since you don’t control the draw, you design rules you could still live with after a bad one.</p><p>The twenty-first-century economy is starting to look like that parking lot, except the spaces now compound.</p><p>The old market story said effort and reward at least sort of tracked each other. Not perfectly. Never cleanly. Still, the connection felt legible enough. You worked harder, took risks, built something useful, and maybe you won big. People could envy the winner while still believing the outcome came from something real.</p><p>Network capitalism breaks that story.</p><p>Returns don’t scale smoothly anymore. They bend upward. A small early edge can turn into dominance. One firm gets picked up by the network and starts absorbing everything around it: attention, capital, data, talent, distribution. Another firm does nearly the same thing and vanishes before anyone can know what it might have become.</p><p>This is where <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6003994">Shadow Futures</a> enters.</p><p><strong>A shadow future is an effort path that never becomes visible.</strong> The startup that could have worked but never got distribution. The creator who was good enough but missed the graph. The worker who trained, adapted, did everything right, and still got filtered out by timing, geography, or adjacency. These aren’t just losers in a clean contest. They’re deleted branches.</p><p>From the outside, winners look like signal. From the inside, it’s messier. Path dependence, network effects, reinforcement loops. The system picks a path, then uses the fact that it picked that path as evidence it was always the right one.</p><p>Behind the veil, you’re not just risking poverty. You’re risking invisibility. You might spend years building something real, and it never gets a chance to compound. Nobody can tell if you failed, or if the network simply never selected you.</p><p>At that point, “the market decided” becomes too thin. The market didn’t necessarily judge your value. It may have only recorded your position in a graph.</p><p>This changes what redistribution has to mean.</p><p>Extreme inequality can’t be treated as automatically justified just because it passed through prices. Rawls allows inequality, but only if it helps the worst-off. In a network economy, that test gets harder. Does letting frontier firms dominate actually improve life for people who never get picked? Or does it lock early winners into control of the entrance?</p><p>Redistribution starts to look less like charity and more like insurance against path dependence.</p><p>Taxing hyperscalers more aggressively makes sense when their gains come from network lock-in rather than ordinary production. Interoperability and data portability stop platforms from turning early leads into permanent gates. Public compute, open protocols, stronger antitrust, and sovereign wealth funds tied to hyperscale rents all do the same basic job: they keep the economy from collapsing into a few private choke points.</p><p>UBI fits into this frame too, but not as pity. It looks more like a dividend from an economy that destroys many viable paths before they can pay off.</p><p>The parking lot version is obvious. Let some people have better spaces, sure. Let some people pay, maybe. But don’t let the first tenants who parked near the door permanently own the entrance, redraw the lot, and charge everyone else for walking.</p><p>The next phase may be stranger. If autonomous agents begin running more of the economy, they may discover the same problem in colder terms. A system built on total winner-take-most dynamics becomes brittle. It loses demand, legitimacy, human cooperation, useful data, and political stability.</p><p>Agent economies may redistribute even without moral sentiment, not because they care, but because it’s cheaper than collapse.</p><p>You could imagine compute dividends, data dividends, platform royalties, automated-production taxes, or universal stakes in sovereign AI funds. The mechanism matters, but the reason is simpler: a solvent population becomes part of the operating system.</p><p>Rawls was never asking us to shuffle outcomes randomly. He was asking what rules we’d accept if we didn’t know where we’d land.</p><p>In the old economy, bad luck meant being born poor, sick, or far from opportunity. In the network economy, it also means being born into a branch the system never sees.</p><p>A fair twenty-first-century economy has to insure those branches before they disappear. It has to keep the entrance open, even when the parking lot starts scaling into the sky.</p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21955e0f059d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The End of Work as a Readable Signal]]></title>
            <link>https://medium.com/@SeloSlav/the-end-of-work-as-a-readable-signal-ad64fe79bc95?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad64fe79bc95</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[society]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Wed, 15 Apr 2026 17:26:29 GMT</pubDate>
            <atom:updated>2026-07-22T02:07:47.597Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why success is getting harder to interpret in a networked economy</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/0*9TgHB3wPHCsadspm" /><figcaption>Pieter Bruegel the Elder, “The Triumph of Death” (c. 1562). A world where outcomes are no longer governed by individual effort, but by forces that overwhelm and reorder everything in their path.</figcaption></figure><p>I keep coming back to a simple problem.</p><p>A society can survive inequality. It can survive elites. It can survive plenty of obvious unfairness, honestly. What it has a much harder time surviving is a breakdown in how people make sense of outcomes. People need to believe, at least roughly, that effort, skill, discipline, and risk still connect to reward in a way they can recognize.</p><p>I don’t think that belief is holding up.</p><p>For a long time, capitalism could lean on a story people more or less accepted. You work hard, you take risks, you build something useful, and if you end up with an outsized reward, fine. People might envy you, but they can still tell themselves the outcome came from something real. The winner did something. Built something. Saw something first. Took a chance others didn’t.</p><p>Sure, this was never perfectly true. It was never clean. Still, it was legible enough.</p><p>What I think’s changing now is not just the distribution of rewards, but the readability of rewards. More and more of modern life runs through systems where the thing being rewarded is not contribution in any straightforward sense. Instead, it’s position. It’s timing, adjacency, visibility, and reinforcement.</p><p>In other words, we’re moving out of markets and deeper into networks.</p><p>In a market, at least in the textbook sense, goods compete. Prices move. Substitutable things push against each other. If you make a better product or a cheaper product, you can gain ground. There are lots of caveats, obviously, but the basic intuition still works.</p><p>Networks do something else. Networks amplify what already has a foothold. Early advantage compounds. Visibility brings more visibility. Reputation brings more reputation. Users attract users. Citations attract citations. Capital attracts capital. Once that process gets going, it doesn’t just reward what’s good. It rewards what’s already winning.</p><p>W. Brian Arthur saw a lot of this coming before most people did. In <em>Increasing Returns and Path Dependence in the Economy</em>, and in his 1996 Harvard Business Review piece, “Increasing Returns and the New World of Business,” he laid out the basic shift. Some parts of the economy don’t behave like the clean competitive worlds economists like to model. They behave like systems with lock-in. Small differences early on can snowball into huge differences later. Once a path gets selected, it becomes self-reinforcing.</p><p>You can see this everywhere now. Software. Media. Finance. Publishing. Platform businesses. Social media. Job markets. Dating apps, even. A lot of life no longer feels like a place where people bring goods into markets. It feels like a place where people throw signals into giant ranking systems and hope the loop catches them.</p><p>I wrote <em>Shadow Futures</em> because I wanted a language for what this does to our idea of merit.</p><p>The core idea is not that work stops mattering. People hear this argument and immediately want to reduce it to that. “Oh, so effort means nothing now?” No. That’s not what I’m saying. What I’m saying is stranger, and I think more unsettling.</p><p>In path-dependent systems, even perfectly visible work can stop functioning as a reliable signal of value.</p><p>Two people can do the same work. Same effort. Same quality. Same measurable output. One catches a wave and the other doesn’t. One gets the audience, the distribution, the timing, the social proof, the capital follow-through. The other becomes invisible. From the outside, we mostly see the winner and then tell a story backward. We say the winner must have had the better idea, the better instincts, the better work ethic.</p><p>A lot of the time, I don’t think we actually know that.</p><p>What we don’t see are all the unrealized effort trajectories that looked the same until the network tipped the other way. All the books that were good enough. All the companies that shipped something real. All the papers that could have mattered. All the apps that worked. All the talented people who did roughly what the winner did and vanished anyway because the system had already started reinforcing somewhere else.</p><p>Once you take those unseen paths seriously, the whole moral picture gets shakier.</p><p>The rich and the famous and the dominant firms feel like they did far more than they actually did because they can’t see the erased alternatives produced by their own success. Their win wipes out evidence. It narrows the visible world until the realized path starts to look inevitable. Then they mistake inevitability for merit.</p><p>I think this is why so much public argument around work feels stale. We’re still using an ethic built for a world with more bounded competition and more legible contribution. Meanwhile the actual machinery of the economy has changed. More people are living entrepreneurially whether they want to or not. The writer posts into feeds. The programmer ships in public. The merchant fights recommendation systems. The musician lives and dies by ranking loops. The academic is trapped in prestige cascades. Even ordinary labor now has this layer of self-marketing and exposure glued onto it.</p><p>More people are not just working. They’re performing for allocation systems.</p><p>And yet our culture keeps telling the old story. Work hard. Take risks. Build. Grind. Earn it.</p><p>I get why. A society probably can’t function without some shared ethic around work. If nobody believes effort matters, everything starts to rot. People become cynical or fatalistic. Some chase clout instead of substance. Some give up. Some start demanding ever more political correction from above because they stop believing the social order has any internal logic.</p><p>So I’m not arguing for nihilism here. I’m arguing that the old language no longer matches the underlying system.</p><p>To me, the real question now is whether we can build a new ethic that is honest about scale, networks, lock-in, and information loss without collapsing into despair or envy. We need some way of talking about labor, capital, and reward that doesn’t pretend all large fortunes are just magnified versions of ordinary productive effort. They aren’t. Many of them are the result of reinforcement processes that become harder to distinguish from merit precisely because they erase the nearby failures.</p><p>Once you see it that way, even policy starts to look different.</p><p>Take a wealth tax. People love to frame it as an attack on success, or on risk-taking, or on capitalism itself. I don’t think that frame works anymore. If extreme returns are increasingly driven by accumulated position, network lock-in, and winner-take-most dynamics, then taxing that stock is not some socialist war on effort. It’s a structural correction. A way of dealing with the fact that realized outcomes in these systems contain less information than we pretend they do.</p><p>Same goes for antitrust, platform design, ranking systems, credentialing, all of it. The old instinct is to measure harder, verify harder, optimize harder. My view is that this misses the point. The problem isn’t that we can’t see work. The problem is that seeing work no longer tells us what we think it tells us.</p><p><strong>What matters now is allocation structure.</strong></p><p>Can ordinary contribution still change the state of play? Can new entrants still break in without impossible luck? Can value travel without first becoming prestige? Can a person do solid work without having to win an attention war?</p><p>If the answer keeps becoming no, then no sermon about grit is going to save the culture.</p><p>I think we’re heading toward a period where societies will have to choose. Either they redesign institutions so effort becomes legible again in more domains, or they keep preaching an ethic the system itself no longer supports. If we choose the second path, the result will not be a stable meritocracy. It will be a brittle society full of people asked to believe in outcomes they can no longer make sense of.</p><p>To me, that’s the real stakes of this.</p><p>Not whether work matters.</p><p>Whether people can still see it matter.</p><p>If you want the formal version of this argument, the model, the logic, the technical framing, I wrote it up here:</p><p><a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6003994&amp;utm_source=chatgpt.com">https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6003994</a></p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad64fe79bc95" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cheng Lou’s Pretext and the Case for Reactive Surface Layout as a New Graphics Primitive]]></title>
            <link>https://medium.com/@SeloSlav/weft-and-the-case-for-reactive-surface-layout-as-a-new-graphics-primitive-040cf477e31e?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/040cf477e31e</guid>
            <category><![CDATA[threejs]]></category>
            <category><![CDATA[unity-game-development]]></category>
            <category><![CDATA[threejs-tutorial]]></category>
            <category><![CDATA[game-design]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Sun, 29 Mar 2026 22:23:38 GMT</pubDate>
            <atom:updated>2026-03-31T13:38:08.363Z</atom:updated>
            <content:encoded><![CDATA[<h4>Weft SDK: A different way to think about grass, facades, rubble, wounds, and other reactive surfaces in 3D</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l5oYzBUiSxrMMFRuSXSLaw.png" /><figcaption>Move through the fire wall and it opens around you, then fills back in behind you under the same reactive surface logic.</figcaption></figure><p>You can install it with:</p><pre>npm install weft-sdk three</pre><p>Most game surface systems still make you solve the same problem twice.</p><p>First you place things. Grass, rubble, facade pieces, debris bands, scales, stars, wires, whatever belongs on the surface.</p><p>Then later, when gameplay touches it, you build a second system to make it react. Damage masks. Density maps. Respawn logic. Healing timers. Custom rebuild code. Special-case recovery behavior. A pile of one-off fixes that grows every time a new surface type shows up.</p><p>People are used to this split, so they stop seeing how awkward it is. Scatter handles placement. Some other runtime system handles response.</p><p>Weft starts from a different premise.</p><p>What if placement and reactivity are the same problem?</p><p>What if a surface is not just a cloud of instances scattered across geometry, but a layout that can be narrowed, punctured, thinned, healed, and shifted under one deterministic model?</p><p>That’s the core of <a href="https://github.com/SeloSlav/weft"><strong>Weft.</strong></a></p><p>It’s a Three.js SDK for deterministic reactive surfaces, but the more interesting part is the mental model behind it. This is not just another scatter helper. It’s a different way to author a whole class of game surfaces.</p><p>The model is simple:</p><blockquote><strong>source -&gt; layout -&gt; effect</strong></blockquote><p>A surface starts with a <strong>source. </strong>Maybe repeated units. Maybe a semantic palette with ids, weights, and metadata.</p><p>Then comes <strong>layout.</strong> Rows, sectors, available width, local constraints.</p><p>Then comes <strong>effect. </strong>The laid-out result gets projected into world-space instances.</p><p>From there, gameplay response becomes a width problem.</p><p>If a region gets narrower, fewer units fit.</p><p>If width drops to zero, the surface opens.</p><p>If width comes back, the surface heals.</p><p>If semantic weights shift, the surface changes apparent state without needing a second placement pipeline.</p><p>Instead of scattering instances once and then inventing separate runtime logic for every type of damage, healing, fracture, or recovery, Weft keeps those behaviors inside one shared layout model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-yAB7QgaglKAOd-BhiZDtA.png" /><figcaption>Shoot the grass and the same layout system thins the surface in place. No separate damage texture, scatter rebuild, or special-case effect pipeline.</figcaption></figure><p>So grass trampling, facade clearing, punctured neon, wounded sky, rubble thinning, and recoverable surface detail stop being completely separate systems. They become variations of the same contract.</p><p>That’s why I think this feels novel.</p><p>A lot of tools make existing workflows more convenient. Weft points at a different workflow entirely. It treats reactive surfaces as something closer to live constrained layout than ordinary scatter. Once you see it that way, a lot of old environment logic starts looking like duplicated effort.</p><p>The other reason this is interesting is where the idea comes from.</p><p>Under the hood, Weft builds on <a href="https://x.com/_chenglou/status/2037715519277760531">Cheng Lou’s Pretext,</a> which takes text measurement and line breaking seriously enough to make layout stable and predictable. Weft pushes that underlying idea into a different domain. Instead of using measured layout for reading, it uses it as a packing model for reactive surfaces in 3D. Once width determines how much fits, and once that rule stays deterministic, you get a surface model that can react without collapsing into ad hoc runtime hacks.</p><p>I think this may go further than reactive layout for reading systems or better scatter tools.</p><p>Reactive surfaces are just the first clear use case. The deeper idea is that measured layout can act as a general runtime primitive for structured spatial change. Anywhere you have constrained bands, fields, shells, facades, or patterned surfaces that need to thin, open, recover, or shift state under gameplay, this model gets interesting.</p><p>It’s not a replacement for every procgen workflow. It’s not trying to solve full ecological simulation, blue-noise placement across huge terrains, or general biome generation. But for authored reactive fields, it opens a very different path.</p><p>Weft treats a reactive world surface as something you lay out once under a rule system that already contains its possible reactions.</p><p>I think that’s a real step forward.</p><p>If you want to try it, the repo is <a href="https://github.com/SeloSlav/weft">here,</a> and the package is live as weft-sdk on npm.</p><p>Martin Erlic writes about systems, economics, and the future, and also runs an <a href="http://www.seloolive.com/">olive oil</a> business. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction shaped by Eastern European history and culture. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=040cf477e31e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Quick SpacetimeDB Auth Setup with OpenAuth, Hono, and React Context]]></title>
            <link>https://medium.com/@SeloSlav/quick-spacetimedb-auth-setup-with-openauth-hono-and-react-context-ef2ededba9fb?source=rss-f864d52751a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef2ededba9fb</guid>
            <category><![CDATA[indie-game]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[indie-game-dev]]></category>
            <category><![CDATA[jwt-authentication]]></category>
            <category><![CDATA[web-development]]></category>
            <dc:creator><![CDATA[Martin Erlic]]></dc:creator>
            <pubDate>Tue, 03 Mar 2026 07:12:02 GMT</pubDate>
            <atom:updated>2026-07-22T02:07:51.751Z</atom:updated>
            <content:encoded><![CDATA[<h4>Using OpenAuth + Hono as OIDC issuer for a SpacetimeDB App</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aquoGJl4XZV70QbrxcJM7A.png" /></figure><p>If you’re building a real-time multiplayer game on <strong>SpacetimeDB</strong>, authentication needs to be simple, deterministic, and compatible with Spacetime’s identity model. In my project, <a href="https://www.brothandbullets.com/"><strong>Broth &amp; Bullets</strong>,</a> I use an OpenAuth OIDC server running on Hono and pass the resulting id_token directly into the SpacetimeDB connection builder.</p><p>Repo reference: <a href="https://github.com/SeloSlav/spacetimedb-openauthjs">https://github.com/SeloSlav/spacetimedb-openauthjs</a></p><p>This article walks through the architecture and shows the exact React AuthContext + GameConnectionContext pattern I use.</p><h3>Architecture Overview</h3><p>The flow looks like this:</p><ol><li>User logs in via OpenAuth (running on Hono).</li><li>OpenAuth issues a standard OIDC id_token.</li><li>The frontend stores the token in localStorage.</li><li>The token is passed into DbConnection.builder().withToken(...).</li><li>SpacetimeDB derives identity from iss + sub.</li></ol><p>That’s it. No custom JWT parsing on the backend. No proxy auth layer. No session database.</p><p>SpacetimeDB trusts the OIDC issuer and maps the token to a stable identity.</p><h3>1. OpenAuth + Hono Auth Server</h3><p>Inside /auth-server-openauth, I run OpenAuth as an OIDC issuer. The important part is that it:</p><ul><li>Exposes /authorize</li><li>Exposes /token</li><li>Publishes JWKS</li><li>Signs a valid id_token</li></ul><p>Your issuer becomes:</p><pre>http://localhost:4001   (dev)<br>https://broth-and-bullets-production.up.railway.app   (prod)</pre><p>Your React app just needs:</p><pre>const AUTH_SERVER_URL = isDevelopment <br>  ? &#39;http://localhost:4001&#39; <br>  : &#39;https://broth-and-bullets-production.up.railway.app&#39;;</pre><pre>const OIDC_CLIENT_ID = &#39;vibe-survival-game-client&#39;;</pre><h3>2. React AuthContext — PKCE + Token Storage</h3><p>Here’s the core idea of AuthContext.tsx.</p><p>We manually construct the OIDC Authorization Code + PKCE flow. No heavy auth framework required.</p><h3>PKCE Generator (Browser Safe)</h3><pre>async function generatePkceChallenge(verifier: string) {<br>  const encoder = new TextEncoder();<br>  const data = encoder.encode(verifier);<br>  const digest = await window.crypto.subtle.digest(&#39;SHA-256&#39;, data);</pre><pre>  const base64 = window.btoa(<br>    String.fromCharCode(...new Uint8Array(digest))<br>  );</pre><pre>  const code_challenge = base64<br>    .replace(/\+/g, &#39;-&#39;)<br>    .replace(/\//g, &#39;_&#39;)<br>    .replace(/=/g, &#39;&#39;);</pre><pre>  return {<br>    code_verifier: verifier,<br>    code_challenge,<br>    code_challenge_method: &#39;S256&#39;<br>  };<br>}</pre><h3>Login Redirect</h3><p>We manually build the authorize URL:</p><pre>const authUrl = new URL(&#39;/authorize&#39;, AUTH_SERVER_URL);<br>authUrl.searchParams.set(&#39;client_id&#39;, OIDC_CLIENT_ID);<br>authUrl.searchParams.set(&#39;redirect_uri&#39;, REDIRECT_URI);<br>authUrl.searchParams.set(&#39;response_type&#39;, &#39;code&#39;);<br>authUrl.searchParams.set(&#39;code_challenge&#39;, pkce.code_challenge);<br>authUrl.searchParams.set(&#39;code_challenge_method&#39;, &#39;S256&#39;);<br>authUrl.searchParams.set(&#39;acr_values&#39;, &#39;pwd&#39;);</pre><pre>window.location.assign(authUrl.toString());</pre><p>No magic. Just OIDC.</p><h3>Handling the Callback</h3><p>When redirected back:</p><pre>const tokenResponse = await fetch(`${AUTH_SERVER_URL}/token`, {<br>  method: &#39;POST&#39;,<br>  headers: { &#39;Content-Type&#39;: &#39;application/x-www-form-urlencoded&#39; },<br>  body: new URLSearchParams({<br>    grant_type: &#39;authorization_code&#39;,<br>    code,<br>    redirect_uri: REDIRECT_URI,<br>    client_id: OIDC_CLIENT_ID,<br>    code_verifier: verifier<br>  })<br>});</pre><pre>const tokens = await tokenResponse.json();<br>const id_token = tokens.id_token;</pre><p>Then store:</p><pre>localStorage.setItem(&#39;oidc_id_token&#39;, id_token);<br>setSpacetimeToken(id_token);</pre><p>That token is now your Spacetime identity.</p><h3>3. Token Parsing + Expiration</h3><p>We parse locally only to extract user profile and check expiration.</p><pre>const decoded = parseJwt(token);<br>const now = Math.floor(Date.now() / 1000);</pre><pre>if (decoded.exp &amp;&amp; decoded.exp &lt; now) {<br>  return null;<br>}</pre><pre>return {<br>  userId: decoded.sub,<br>  email: decoded.email<br>};</pre><p>Important:</p><p>SpacetimeDB validates the signature using the issuer’s JWKS. You do not verify it client-side.</p><h3>4. GameConnectionContext — Passing Token to SpacetimeDB</h3><p>This is where it gets clean.</p><p>From GameConnectionContext.tsx:</p><pre>const builder = DbConnection.builder()<br>  .withUri(SPACETIME_DB_ADDRESS)<br>  .withDatabaseName(SPACETIME_DB_NAME)<br>  .withToken(spacetimeToken)</pre><p>That’s it.</p><p>No session exchange. No backend proxy. No custom middleware.</p><p>SpacetimeDB reads:</p><pre>iss + sub</pre><p>And derives a stable Identity.</p><h3>On Connect</h3><pre>.onConnect((conn, identity) =&gt; {<br>  connectionInstanceRef.current = conn;<br>  updateConnectionState(ConnectionState.CONNECTED, conn, identity);<br>});</pre><p>identity is now your authoritative player identity inside the database.</p><h3>Handling 401 / Auth Rejection</h3><p>If Spacetime rejects the token:</p><pre>if (errorMessage.includes(&quot;401&quot;) || <br>    errorMessage.toLowerCase().includes(&quot;unauthorized&quot;)) {<br>  invalidateCurrentToken();<br>}</pre><p>Which clears localStorage and forces a clean login.</p><p>This prevents ghost sessions and stale tokens.</p><h3>5. Player Registration Example</h3><p>Once connected:</p><pre>await connection.reducers.registerPlayer({ username });</pre><p>No user ID passed.</p><p>Spacetime already knows who is calling from the OIDC token.</p><p>That’s the entire power of this setup.</p><h3>Why This Pattern Works</h3><p>This architecture is strong because:</p><ul><li>OIDC compliant</li><li>No custom JWT signing</li><li>No session DB</li><li>No backend proxy</li><li>Clean separation of auth and realtime layer</li><li>Fully scalable</li></ul><p>You can run:</p><ul><li>OpenAuth on Railway</li><li>SpacetimeDB on Maincloud</li><li>React client on Vercel</li><li>And everything still works deterministically</li></ul><h3>Full Auth → Game Flow Summary</h3><ol><li>React builds PKCE challenge.</li><li>User logs in via OpenAuth.</li><li>React exchanges code for id_token.</li><li>Token stored locally.</li><li>Token passed into DbConnection.builder().withToken().</li><li>SpacetimeDB derives identity.</li><li>Reducers execute under that identity.</li></ol><p>No session cookies. No auth server proxy. No REST user lookup.</p><p>Just pure OIDC → realtime identity.</p><h3>Why This Is Ideal for Multiplayer Games</h3><p>Traditional web auth patterns break down in realtime systems.<br> SpacetimeDB expects identity at connection time, not per request.</p><p>Using OpenAuth as an OIDC issuer keeps things:</p><ul><li>Stateless</li><li>Deterministic</li><li>Horizontally scalable</li><li>Compatible with Maincloud</li></ul><p>And because you control the issuer, you can later:</p><ul><li>Add OAuth providers</li><li>Add passkeys</li><li>Add role claims</li><li>Add refresh flows</li><li>Add admin-only reducers</li></ul><p>Without changing your Spacetime architecture.</p><h3>Closing</h3><p>If you’re building a serious multiplayer web game and want clean auth without backend glue code, this pattern works extremely well.</p><p>Repo reference again: <a href="https://github.com/SeloSlav/spacetimedb-openauthjs">https://github.com/SeloSlav/spacetimedb-openauthjs</a></p><p>SpacetimeDB + OpenAuth + Hono + React Context gives you:</p><p>Minimal surface area.Full OIDC compliance. Stable multiplayer identity.</p><p>And zero auth duct tape.</p><p>If this article helped you, consider getting started with SpacetimeDB using my link. It directly supports my writing and the games I’m building.</p><p>👉 <a href="https://spacetimedb.com/?referral=SeloSlav"><strong>https://spacetimedb.com/?referral=SeloSlav</strong></a></p><p>Martin Erlic writes about systems, economics, and the future. He is the author of the <a href="https://www.babushkabook.com/"><em>Babushka saga,</em></a> a long-form work of speculative fiction about Slavic grandmothers dragging civilization through the end of time by its collar. He lives with his family in a small village by the sea. You can find him on <a href="https://x.com/seloslav">X.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef2ededba9fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>