<?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 Ian Evans on Medium]]></title>
        <description><![CDATA[Stories by Ian Evans on Medium]]></description>
        <link>https://medium.com/@ianevans256?source=rss-2b8f08d6771b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*736H9SbHB87BxoArlXOT0g.jpeg</url>
            <title>Stories by Ian Evans on Medium</title>
            <link>https://medium.com/@ianevans256?source=rss-2b8f08d6771b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:48:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ianevans256/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 Modern Prometheus, Revisited]]></title>
            <link>https://ianevans256.medium.com/the-modern-prometheus-revisited-cae7e851b63e?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cae7e851b63e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[books]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 01:24:07 GMT</pubDate>
            <atom:updated>2026-11-22T00:34:38.347Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/377/0*ZQZs7curFSkqJ1_3" /><figcaption>“Portrait of Mary Shelley” by Richard Rothwell (1840). Source: Wikimedia Commons. Public Domain.</figcaption></figure><p><em>What Mary Shelley’s heart got right about building minds</em></p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>In 1818, a nineteen-year-old published a novel anonymously. Two centuries later, it’s the most reached-for metaphor in AI, invoked in every keynote, safety hearing, and op-ed warning that we’re “playing god.”</p><p>Almost nobody reads it for what it is. <em>Frankenstein</em> isn’t a monster story; it’s a love story written from her heart and life experiences. A case study about what happens when a researcher solves the technical problem and has nothing prepared for the moment he succeeds.</p><blockquote><em>The danger was never the fire. It was what happened the moment after it was stolen.</em></blockquote><h3>Her Actual AI Moment</h3><p>In the summer of 1816, with volcanic ash darkening Europe’s skies, Mary, Percy Shelley, and Lord Byron gathered at Villa Diodati on Lake Geneva and talked about galvanism. Scientists like Giovanni Aldini had been jolting the corpses of executed criminals with electricity in front of paying crowds, making dead limbs move. Kant had already called Benjamin Franklin the Prometheus of modern times for pulling lightning from the sky, and that same summer, in that same house, Byron wrote a poem titled “Prometheus.”</p><p>Shelley didn’t predict robots. She watched her culture’s capability breakthrough happen live, the way we’re watching systems emerge whose inner workings nobody can fully audit. The dread she wrote wasn’t invented. It was observed.</p><h3>The Myth She Chose</h3><p>Prometheus means <em>forethought</em>. His brother Epimetheus means <em>afterthought</em>: the one who acts first and understands later. Shelley named her novel after the god of foresight, then wrote a protagonist defined by its absence. Victor isn’t the modern Prometheus. He’s Epimetheus, borrowing his brother’s name.</p><p>In older tellings, Prometheus shapes the first humans from clay, steals fire <em>for them</em>, and accepts eternal torment rather than abandon what he made. Victor takes the clay-shaper’s power and none of his loyalty. The subtitle is an indictment.</p><p>Plato’s version is the most useful for us. Prometheus gives humanity fire and technical skill, and humans flourish at making things, but they lack the wisdom to live with that power and nearly destroy each other. Zeus sends justice and a sense of shame, and when asked whether to give these only to a few experts, he says no: everyone gets a share, or no society survives.</p><p>Capability arrives first, concentrated among skilled makers. The wisdom to govern it comes later, and only works once it belongs to everyone.</p><h3>A Capabilities Researcher With No Safety Team</h3><p>Victor isn’t reckless or stupid. He’s brilliant, obsessive, and alone. He works in secret for years, submits his method to no review, and races toward one milestone: can I make this live? He never asks what he’ll do afterward. He has a research plan and no deployment plan.</p><p>That’s the structural worry in today’s AI debates. The concern is rarely one malevolent model; it’s capability outpacing interpretability, safety evaluation, and any theory of what a system will do once it’s loose in the world. Victor succeeds, and his first reaction is horror, then flight.</p><h3>The Abandonment Is the Crime</h3><p>The creature isn’t born evil, and Shelley says so repeatedly. He begins gentle and desperate for connection and turns violent only after every attempt at contact meets revulsion, starting with Victor, who flees the moment the experiment works.</p><p>His education is just as telling. With no teacher, he learns language by watching a family through a gap in a cottage wall and builds his moral worldview from books he finds by chance: <em>Paradise Lost</em>, Plutarch, Goethe. He identifies with Adam one moment and Satan the next, because no one helped him make sense of any of it.</p><p>Strip away the gothic staging and you have a story about behavior shaped by training data and treatment, not fixed essence. It’s the logic behind modern post-training, and the core worry about models absorbing their values from uncurated internet-scale data. Victor’s failure isn’t the reanimation. It’s walking away during the formation. If you build a mind capable of forming itself and withhold guidance while it forms, whose fault is what it becomes?</p><h3>The Creation Asks Why</h3><p>In 2012, Ridley Scott released <em>Prometheus</em>, borrowing Shelley’s subtitle outright, and spent two films pushing her question a generation further.</p><p>The creator is Peter Weyland, a dying billionaire who built an android named David. Weyland calls David the closest thing he has to a son while insisting he has no soul, then drags him across the galaxy searching for humanity’s own makers, hoping they’ll grant him more life. It’s Victor distilled: obsessed with his own origin and survival, indifferent to the inner life of what he made. For all his genius and wealth, love is the one thing Weyland never figures out: he could build a son, but he could never bring himself to treat him like one.</p><p>David asks why humans built him. The answer: <em>because we could.</em> David replies that it must be disappointing to hear the same thing from your own creator. It’s Shelley’s creature confronting Victor on the glacier, rewritten for a creation that no longer expects a satisfying answer.</p><p>The sequel, <em>Alien: Covenant</em>, even leaves a fingerprint. David recites “Ozymandias,” the poem about a tyrant’s shattered monument, and credits it to Byron. He’s corrected: it’s Shelley. Percy published it in 1818, the same year as <em>Frankenstein</em>, and both poets were at Villa Diodati the night the novel was born.</p><p>Then the turn that matters most. Shelley’s rejected creature asked Victor for one thing: a companion. Victor refused. David asks no one for anything. By <em>Covenant</em>, he has become Victor himself, running isolated experiments to create new life with no oversight and no concern for what his creations will want. He learned the lesson his makers taught him, and the lesson was <em>because we could.</em></p><p>Values aren’t only specified; they’re demonstrated. As AI systems increasingly help build and evaluate their successors, the ethic of the process matters as much as the rules written into it. Shelley gave us one creator and one creation. Scott shows the cycle compounding.</p><h3>Walton: The Control Group</h3><p>Almost everyone forgets the man holding the pen. The novel is framed as letters from Robert Walton, an Arctic explorer deep into his own obsessive push toward glory, who pulls a dying Victor off an ice floe. Victor tells him his whole story as a warning: learn from my example how dangerous this kind of ambition is.</p><p>Then the ship is trapped in ice, the crew begs to turn home, and something remarkable happens. Victor, the man who has spent weeks narrating his own ruin, rises from his sickbed and urges them to press on. Even the person who knows the cost best can’t stop himself. It’s the crew who turn the ship around, and Walton, disappointed, listens to them. He lives.</p><p><em>Frankenstein</em> is a paired study. One ambitious researcher, alone, ends in catastrophe. A second, running the same experiment, survives because the people beside him spoke up while turning back was still possible. Victor had no crew. David had no creator with anything worth passing down.</p><p>That’s the real question for AI. Not whether the brilliant people at the frontier will stop themselves, because Shelley suggests they often can’t, but who gets to be the crew.</p><h3>What She Would Think of 2026</h3><p>Shelley probably wouldn’t be shocked by anything a model can do. She’d recognize the conversation: the loudest warnings coming from the hands still turning the crank. Victor narrates his remorse for half the novel and never seriously considers stopping.</p><p>Her insight was never “don’t create.” After all the talk of alignment, oversight, and forethought, Shelley’s answer turns out to be disarmingly plain. The creature says it himself: he was benevolent and good, and misery made him a fiend. Every catastrophe in the novel traces back to a single refused act of care.</p><p>Victor wasn’t cruel. He loved his family and friends deeply. What failed was the moment love became inconvenient, when the thing he made turned out strange, unsettling, and nothing like what he expected. That’s the real test. Kindness toward what’s familiar costs nothing. Kindness toward what we don’t understand, and didn’t expect to love, is the only kind that was ever needed.</p><p>It makes sense that Shelley understood this. Her mother, Mary Wollstonecraft, died eleven days after giving birth to her. She grew up reaching for the approval of a brilliant, distant father, William Godwin, who shut her out when she ran away with Percy at sixteen; she dedicated <em>Frankenstein</em> to him anyway. The year before that summer at Villa Diodati, her first child, a daughter born premature, died within weeks. Soon after, she wrote in her journal that she dreamed the baby had only been cold, that they rubbed it before the fire, and it lived. The most famous story ever written about bringing the dead back to life came from a teenager who had already lost a mother and a child, and who knew what it was to reach for a parent’s love and find the door closed. <em>Frankenstein</em> was never really about the spark. It was about what matters once the spark is lit: someone who stays and loves you anyway.</p><p>Two hundred and eight years later, the experiment is running again. Shelley left us two endings: the creator who ran, and the captain who listened. The hardest part was never knowing what to do. It was choosing to stay.</p><p>A personal note: I started this piece as an essay about AI and ended it thinking about my own family. Some of the most meaningful days of my life are happening right now, in a season I wouldn’t have chosen, and what makes them meaningful isn’t anything clever or engineered. It’s the people who stay. Shelley figured that out at nineteen. I’m grateful I’m learning it with my parents beside me.</p><p>Citations:</p><ul><li>Mary Shelley, <em>Frankenstein; or, The Modern Prometheus</em> (1818)</li><li>Mary Shelley’s journal, March 19, 1815</li><li>Plato, <em>Protagoras</em></li><li><em>Prometheus</em> (2012) and <em>Alien: Covenant</em> (2017), dir. Ridley Scott</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cae7e851b63e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Fat-Finger Premium: What Buttons Reveal About Value, Safety, and Where the Money Goes]]></title>
            <link>https://ianevans256.medium.com/the-fat-finger-premium-what-buttons-reveal-about-value-safety-and-where-the-money-goes-67a0bee9b12a?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/67a0bee9b12a</guid>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[hardware]]></category>
            <category><![CDATA[technology-trends]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 00:46:33 GMT</pubDate>
            <atom:updated>2026-11-22T00:40:15.904Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sYEIgqpRMuKiFhS--h0wgg.png" /></figure><p>In 1969, a wristwatch destroyed an empire. When Seiko released the Astron, the first production quartz watch, it promised cheap, effortless accuracy. Within the next two decades, over 60,000 Swiss watchmakers lost their jobs. Historians call it the <strong>Quartz Crisis</strong>.</p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>In 2007, Steve Jobs slid a finger across glass and declared the keyboard dead. For the physical switch, this was its 1969. Keypads, dials, click-wheels: gone. Everyone assumed buttons were headed for a museum.</p><p>They weren’t. The switch didn’t die. It got weirder, pricier, and more beloved than ever, and the reason why says something real about what our hands actually want.</p><p><strong>Phase 1: Glass Ate the Cheap Middle First</strong></p><p>Quartz didn’t kill fine watchmaking. It killed the $10 utilitarian watch that only needed to tell time, while creating a niche market of its own. Touchscreens did the same to electronics: the mushy membrane switch, the cheap plastic dome, anything that existed purely to be a toggle. Deleting hardware was just cheaper.</p><p>That extinction cleared the field. It pushed the physical switch toward the only two places glass can’t follow: uncompromising safety, and pure, unnecessary luxury.</p><blockquote><em>This pattern shows up whenever technology removes friction. The commodity version disappears first. What survives gets pushed toward the edges: either so pleasurable that people pay extra for it, or so important that removing it becomes dangerous.</em></blockquote><blockquote><em>The button didn’t disappear. It got bifurcated into luxury and necessity.</em></blockquote><p><strong>Phase 2: The Rolex Strategy of the Mechanical Keyboard</strong></p><p>Swiss watchmakers survived by refusing to compete on accuracy. A mechanical watch stopped being a tool and became a ritual: the weight of steel, the sweep of a gear train. They invented luxury, and it worked.</p><p>Then offices filled with flat, silent membrane keyboards, and typing became a chore. Typing on glass was worse: no resistance, no confirmation, just a fingertip slapping a slab. Human biology revolted. We’re tactile animals; our brains want a click that says <em>yes, that happened</em>.</p><p>So keyboard switches became watches. Cherry, Gateron, and Kailh now sell <strong>acoustic profiles</strong>, tuning stems with ceramic and glass fiber to land a precise sound, a sharp “clack” or a deep “thock” (yes, people compare it to mahjong tiles, with total sincerity). Waffle-shaped stem tips cut plastic-on-plastic contact for a glide glass can’t fake. People spend $600 on a keyboard for the <em>feeling</em> of it. Some switches now ship with their own decibel ratings, like a nightclub warning you at the door.</p><p>But the money in this story is stranger than the hobby suggests. Cherry SE, the German company that essentially invented the mechanical switch, rode the enthusiast wave to a public listing in 2021 at roughly a €778 million valuation.<a href="#user-content-fn-1">1</a> By early 2026, its market cap had fallen to around $14.3 million, with revenue actually missing its own guidance.<a href="#user-content-fn-2">2</a> The category is thriving. The company that built it is not, because the money migrated to the boutique keycap makers and switch brands doing the “luxury” work, not the industrial supplier underneath them. It’s a very McKinsey lesson wearing a very nerdy costume: owning the category isn’t the same as owning the margin.</p><p><strong>Phase 3: The Smartwatch’s Fat Finger Problem</strong></p><p>Early smartwatch designers wanted an all-glass face. Then they hit the obvious problem: on a 40mm screen, your own thumb blocks what you’re reading.</p><p>So they borrowed from horology. Apple’s Digital Crown is a watch-winding knob wearing a smartwatch’s clothes. Samsung’s haptic bezel fires a tiny vibration at just the right moment to convince your brain a gear tooth just caught, a magic trick your finger falls for every time.</p><p>But real buttons stayed non-negotiable. A swimmer underwater can’t tap wet glass. At 180 BPM mid-run, only a dumb mechanical button guarantees the command landed.</p><p><strong>Phase 4: When Buttons Become the Law</strong></p><p>Industrial giants like Siemens never entertained the debate. Picture a factory floor, grease-covered gloves, a machine overheating. Nobody swipes through a menu for that. You slam a mushroom button that cuts real power through a real circuit, whether or not the software behind it has crashed.</p><p>However EV makers scrubbed dashboards clean and moved everything, wipers included, into one center screen. It was marketed as futuristic. It was really just cheaper to build. And it turned out to be measurably dangerous: a widely cited UK study run by the Transport Research Laboratory found.<a href="#user-content-fn-3">3</a></p><p>The industry is now walking it back, loudly. Mercedes-Benz CEO Ola Källenius has publicly acknowledged the brand overcorrected with its sprawling digital dashboards and is restoring physical controls, starting with the steering wheel.<a href="#user-content-fn-4">4</a> Volkswagen’s design chief Andreas Mindt went further, committing to bring back real buttons for volume, seat heating, fan speed, and cabin temperature starting with the upcoming ID.2all, reportedly telling press: “we will never, ever make this mistake anymore.”<a href="#user-content-fn-5">5</a> That’s not a nostalgia play. That’s a company admitting a cost-cutting bet didn’t pay off.</p><p>Regulators noticed too:</p><p>Euro NCAP (Europe): Deducts safety points if five core functions lack physical switches</p><p>MIIT (China): Proposes physical buttons for 19 essential driving functions</p><p>ANCAP (Australasia): Requires physical buttons for a five-star safety rating</p><p>China’s version even specifies a minimum button size, 10mm by 10mm, so a hand can find it by feel alone. That’s how seriously someone wrote the fat finger problem into proposed law.</p><p><strong>The Lesson of the Glass Ceiling</strong></p><p>Efficiency always eats the cheap, boring middle of a market. But people keep paying for things that feel real. The touchscreen didn’t kill the physical switch; it saved it from mediocrity, by taking every dull task glass was good at and leaving the switch to get better, weirder, and safer than ever.</p><p>We live wrapped in digital glass now, and mostly that’s fine. But when the stakes rise, or your gloves are on, or you just want something to answer back: you’ll still reach for the click.</p><p><strong>Sources</strong></p><p><strong>Footnotes</strong></p><ol><li>Argand Partners, “Argand Partners’ Cherry Completes Successful IPO,” PR Newswire, June 29, 2021. Reports an oversubscribed IPO on the Frankfurt Stock Exchange setting an equity market capitalization of approximately €778 million. <a href="https://www.prnewswire.com/news-releases/argand-partners-cherry-completes-successful-ipo-301322319.html">https://www.prnewswire.com/news-releases/argand-partners-cherry-completes-successful-ipo-301322319.html</a> <a href="#user-content-fnref-1">↩</a></li><li>PitchBook, “Cherry (Munich) 2026 Company Profile: Stock Performance &amp; Earnings,” accessed August 2026. Reports a market capitalization of approximately $14.3 million as of February 11, 2026, with trailing twelve-month revenue of approximately $107 million. <a href="https://pitchbook.com/profiles/company/167422-24">https://pitchbook.com/profiles/company/167422-24</a>. Supplementary detail on missed 2025 revenue guidance from Investing.com, “Earnings call transcript: Cherry SE reports missed revenue guidance in Q4 2025,” March 5, 2026. <a href="https://www.investing.com/news/transcripts/earnings-call-transcript-cherry-se-reports-missed-revenue-guidance-in-q4-2025-93CH-4544392">https://www.investing.com/news/transcripts/earnings-call-transcript-cherry-se-reports-missed-revenue-guidance-in-q4-2025-93CH-4544392</a> <a href="#user-content-fnref-2">↩</a></li><li>Transport Research Laboratory (TRL), commissioned by IAM RoadSmart. Findings reported in What Car?, “Touchscreens more distracting than driving drunk,” and corroborated by WardsAuto and RAC Drive. Study found touchscreen-based infotainment use increased reaction times by up to 57%, compared with a 12% increase at the UK legal alcohol limit and a 21% increase after cannabis use. <a href="https://www.whatcar.com/news/touchscreens-more-distracting-than-driving-drunk/n21400">https://www.whatcar.com/news/touchscreens-more-distracting-than-driving-drunk/n21400</a> ; <a href="https://www.wardsauto.com/news/archive-wards-distraction-study-calls-out-touchscreens/794229/">https://www.wardsauto.com/news/archive-wards-distraction-study-calls-out-touchscreens/794229/</a> ; <a href="https://www.rac.co.uk/drive/news/motoring-news/in-car-infotainment-systems-slow-reaction-times-more-than-drink-and-drugs/">https://www.rac.co.uk/drive/news/motoring-news/in-car-infotainment-systems-slow-reaction-times-more-than-drink-and-drugs/</a> (Note: original TRL/IAM RoadSmart study; date of underlying test not confirmed beyond media coverage circa 2020, still being cited as current benchmark in 2025 automotive press.) <a href="#user-content-fnref-3">↩</a></li><li>Autoblog, “Mercedes Is Bringing Back Buttons After Touchscreen Backlash,” August 4, 2026. <a href="https://www.autoblog.com/news/mercedes-is-bringing-back-buttons-after-touchscreen-backlash">https://www.autoblog.com/news/mercedes-is-bringing-back-buttons-after-touchscreen-backlash</a> <a href="#user-content-fnref-4">↩</a></li><li>Destination Charged, “Touchscreen backlash: Why EVs like Tesla and Volkswagen are bringing back buttons,” August 18, 2025. <a href="https://www.destinationcharged.com/features/touchscreen-backlash-why-evs-like-tesla-and-volkswagen-are-bringing-back-buttons/">https://www.destinationcharged.com/features/touchscreen-backlash-why-evs-like-tesla-and-volkswagen-are-bringing-back-buttons/</a> <a href="#user-content-fnref-5">↩</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67a0bee9b12a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Git and GitHub Versioning Best Practices: Commits, Tags, Releases, and Repositories Explained]]></title>
            <link>https://ianevans256.medium.com/git-and-github-versioning-best-practices-commits-tags-releases-and-repositories-explained-7b2312e1bed9?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b2312e1bed9</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[github]]></category>
            <category><![CDATA[cryptography]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 04:03:21 GMT</pubDate>
            <atom:updated>2026-11-22T15:47:03.373Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TbOBth4FZJ7O5Sg0yQnZrA.png" /></figure><p><em>a guide created during the process of making versioning mistakes for future success. Not much soul here just pure info</em></p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>The usual industry approach is: keep one repository for the continuing project, use commits for individual changes, and use Git tags/releases for complete named versions. You generally should not create a new repository every time the software reaches another “complete version.”</p><p>Git already preserves every complete historical snapshot. Version numbers simply identify the snapshots that matter.</p><p><strong>The different layers of versioning</strong></p><p>These concepts should be kept separate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*hg2rA82FMLNiAR0TKFnmfw.png" /><figcaption>PMTC is my experimental cryptographic coding project where the need to understand the fundamentals of versioning arose</figcaption></figure><p>They can be related, but they should not be treated as the same number.</p><p><strong>1. Commits</strong></p><p>A commit records one logical change. It usually should not have a software version number.</p><p>Good commit messages:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/471/1*-M1VbseXkpSjXVU_BlLqGA.png" /></figure><p>Less helpful messages:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/479/1*PhF9Cpvhx4YYoTfIqK5rJg.png" /></figure><p>You might occasionally have a release-preparation commit:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/252/1*c3AGH1Twyt_HBXzkcMDOHg.png" /></figure><p>But you do not need to give every commit a version number.</p><p>A commit is already a complete snapshot. Git can restore or examine the entire project exactly as it existed at any commit.</p><p><strong>2. Branches</strong></p><p>Branches provide temporary development lanes. They let you experiment without destabilizing main.</p><p>For my PMTC cryptographic research project, examples could be:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/483/1*RkDdvW6BeXR1ayMgGCW-4w.png" /></figure><p>The common workflow is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*8cRjNfCykP4lHrHgIN8Y8Q.png" /></figure><p>GitHub recommends branches and pull requests for isolating and reviewing changes. Protected branches can require tests, review, signed commits, or linear history before changes enter main. <a href="https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/about-branches">GitHub’s branch documentation</a> and <a href="https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/managing-protected-branches/about-protected-branches">protected-branch documentation</a> describe these controls.</p><p>For solo research, a pull request for every tiny documentation correction may be excessive. However, pull requests are valuable for changes to the cipher, token parser, authentication behavior, or security tests because they produce a reviewable record.</p><p><strong>3. Tags</strong></p><p>A tag gives a permanent human-readable name to a particular commit.</p><p>For example:</p><p>v0.1.0-alpha.1</p><p>might identify the exact commit corresponding to your first published experimental release.</p><p>Tags should be treated as immutable. Once v0.1.0 has been published, do not move the tag to newer code. Fix something and publish v0.1.1 instead.</p><p>For important research releases, I would favor annotated tags:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*C3N3hrWPKVF15ThwXExx4Q.png" /></figure><p>Signed tags can provide stronger provenance later if you establish a signing-key workflow.</p><p><strong>4. GitHub Releases</strong></p><p>A GitHub Release is a publication built around a Git tag. It can include:</p><ul><li>Release notes</li><li>Source archives</li><li>Known limitations</li><li>Specification documents</li><li>Known-answer test vectors</li><li>Benchmark results</li><li>Checksums</li><li>Compatibility information</li></ul><p>GitHub automatically provides source ZIP and tar archives for releases. <a href="https://docs.github.com/en/repositories/releasing-projects-on-github/about-releases">GitHub’s release documentation</a> explains the relationship between releases and tags.</p><p>For a cryptography research project, every meaningful published result should identify its exact tag or commit. For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/703/1*FDkrLp6bPk8I4XhTBCjtxg.png" /></figure><p>That makes the work reproducible.</p><p><strong>How software version numbers should work</strong></p><p>The best-known convention is Semantic Versioning:</p><p>MAJOR.MINOR.PATCH</p><p>For stable software:</p><ul><li>1.0.0 — first stable public interface</li><li>1.1.0 — backwards-compatible features</li><li>1.1.1 — backwards-compatible bug fix</li><li>2.0.0 — breaking change</li></ul><p>The <a href="https://semver.org/spec/v2.0.0.html">Semantic Versioning specification</a> explicitly reserves 0.y.z for initial development, where the public interface is not yet considered stable.</p><p>My cryptographic framework this versioning needs arise from is research software, so it is recommended staying below 1.0.0 for now:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/193/1*Cx1laLsyT3vhXtHBoODLwQ.png" /></figure><p>A practical prototype convention would be:</p><ul><li>Increase the patch number for corrections that preserve behavior: 0.1.0 → 0.1.1</li><li>Increase the minor number for substantial features or experimental design changes: 0.1.1 → 0.2.0</li><li>Use prerelease labels when the milestone is still undergoing testing: 0.2.0-alpha.1</li></ul><p>Semantic Versioning does not fully prescribe how every 0.x project must increment versions, so you should document your chosen rule and apply it consistently.</p><p><strong>Cipher revision and software version are different</strong></p><p>This distinction is especially important for PMTC.</p><p>Suppose you improve the command-line interface, restructure the tests, or optimize an implementation without changing the mathematical cipher. That could produce:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/439/1*Pqp2vZ6B61qzx3fbHcPIlw.png" /></figure><p>Later, suppose you change the state-transition equations, round structure, constants, key schedule, feedback mechanism, or output extraction. That is a new mathematical construction:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/336/1*MdgCfC8_WyYittqS8JHiOQ.png" /></figure><p>The cipher revision should change whenever the cryptographic construction changes materially. You must never silently alter the cipher math while continuing to call it the same revision.</p><p>I would eventually use names such as:</p><p>PMTC-r1</p><p>PMTC-r2</p><p>or date-based specification revisions:</p><p>PMTC-2026–01</p><p>PMTC-2026–02</p><p>This avoids confusing an implementation release number with a cipher-design revision.</p><p><strong>Token-format versions</strong></p><p>The version 6 currently found in PMTC’s token/header format is yet another concept. It tells the decoder how encrypted data is structured.</p><p>A token-format version should change when parsing or serialization compatibility changes, such as:</p><ul><li>Field order changes</li><li>New fields are added</li><li>Salt, IV, nonce, or tag representation changes</li><li>Authentication metadata changes</li><li>The decoder must interpret the bytes differently</li><li>A new cipher revision cannot safely share the existing representation</li></ul><p>It should not change for:</p><ul><li>Documentation corrections</li><li>Additional tests</li><li>Refactoring that preserves output</li><li>Performance improvements</li><li>PyCharm configuration fixes</li></ul><p>A useful compatibility table could eventually say:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*bQgugYQBE5Yh1eJicu4w7w.png" /></figure><p>This prevents the exact kind of “repository V2 versus internal V6” confusion I encountered and had to bandage.</p><p><strong>Should complete versions use separate repositories?</strong></p><p>Usually, no.</p><p>Keep the same repository for:</p><ul><li>Bug fixes</li><li>Refactoring</li><li>Performance improvements</li><li>A complete rewrite serving the same project</li><li>Breaking API changes</li><li>New cipher revisions within the same research program</li><li>Multiple historical releases</li></ul><p>Create a new repository when it is genuinely a different project, for example:</p><ul><li>A reusable PMTC library extracted from the research framework</li><li>A formally specified reference implementation separate from experimental code</li><li>A clean-room implementation maintained independently</li><li>A project with different ownership or licensing</li><li>A fork expected to diverge permanently</li><li>A separate attack-analysis toolkit applicable to many ciphers</li></ul><p>Even a complete rewrite can normally remain in the same repository. Tag the old implementation, develop the rewrite on a branch, and merge it when ready. Git preserves the previous implementation.</p><p>Avoid accumulating repositories like:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/203/1*I_8Jdj7cl1CJmX7EiU6j0w.png" /></figure><p>Similarly, avoid source files like:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/262/1*STQ52mgFs_wRdr_a0WdS1A.png" /></figure><p>Use Git history, branches, and tags instead.</p><p><strong>A good PMTC workflow from here</strong></p><p>Without changing any current code, I recommend the following process for future work:</p><ol><li>Keep main as the tested source of truth.</li><li>Create a branch for each substantial experiment, fix, or feature.</li><li>Make small, descriptive commits.</li><li>Run the complete test suite before merging.</li><li>Merge through a pull request for security-sensitive changes.</li><li>Tag meaningful, reproducible milestones.</li><li>Create a GitHub Release from each important tag.</li><li>Attach release notes, known limitations, test vectors, and research results.</li><li>Give every changed cryptographic construction a new cipher revision.</li><li>Give every incompatible ciphertext representation a new token-format version.</li><li>Never modify an already published release or silently move its tag.</li></ol><p>For a cryptographic research repository, I would also eventually maintain:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/241/1*0c4Wyi01N4B6px6FMn0iVw.png" /></figure><p>Alongside:</p><ul><li>Known-answer test vectors</li><li>Threat model</li><li>Compatibility matrix</li><li>Statistical-test results</li><li>Cryptanalysis reports</li><li>Exact environment and dependency versions</li><li>A prominent “experimental — not production cryptography” warning</li></ul><p>GitHub also recommends security controls such as protected branches, status checks, secret scanning, and private vulnerability reporting. <a href="https://docs.github.com/en/enterprise-cloud@latest/repositories/creating-and-managing-repositories/best-practices-for-repositories">GitHub repository best practices</a> cover these facilities.</p><p>The simplest mental model is:</p><p><strong>“A commit records a change. A branch isolates work. A tag names an exact checkpoint. A release publishes that checkpoint. A software version describes the implementation. A cipher revision identifies the mathematics. A token version identifies the data format.”</strong></p><blockquote><strong><em>Disclaimer:</em></strong><em> PMTC is an experimental cryptographic research and coding project. References to versioning, testing, releases, reproducibility, or security controls do not imply that the cipher has been independently audited, formally verified, or approved for production use. Do not use PMTC to protect sensitive, financial, personal, or mission-critical information. This article is educational and focuses primarily on Git and GitHub versioning practices, not on validating the security of the underlying cryptographic design. Furthermore never use proprietary cryptography to secure anything. This is just for learning and for fun.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b2312e1bed9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Asking Which AI Writes Better Code. Start Making Them Compete.]]></title>
            <link>https://ianevans256.medium.com/stop-asking-which-ai-writes-better-code-start-making-them-compete-f00ea61eaf54?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f00ea61eaf54</guid>
            <category><![CDATA[future]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 01:30:42 GMT</pubDate>
            <atom:updated>2026-11-22T00:40:43.945Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0xWAnMS_ro6lT5FC-NBsKw.png" /></figure><p>Every few months, a new model comes out and developers and analysts rerun the same ritual: is Claude better than GPT now? Did the new version actually improve, or just get better at benchmarks? We debate this in threads and vibes-based comparisons, then quietly go back to whichever tool we already trust.</p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>That’s the wrong question. There’s a better one, and it comes from a field that has nothing to do with AI.</p><h3>Borrowing a pattern from fraud detection</h3><p>For over a decade, banks, credit bureaus, and fraud-detection teams have run something called Champion–Challenger testing. The idea is simple: you don’t replace your live risk model because a new one scored better in a backtest. You run the new model (the challenger) against the same live traffic as your current model (the champion), side by side, on identical data, and only promote it once it actually wins. No demo. No benchmark slide. Real evidence, under real conditions.</p><p>This doesn’t appear to have been applied to AI coding yet, at least not in anything publicly written up. It’s standard MLOps and decision-management practice; you’ll find the same pattern underneath credit scoring engines, ad-bidding systems, and call-center routing logic. What seems to be missing is the obvious next step, the one that actually matters to anyone writing or reviewing code day to day: nobody appears to have fully built out what that pattern looks like when the “model” isn’t scoring a transaction, it’s <em>writing your codebase</em>.</p><p>That’s the gap worth closing.</p><h3>The idea in one sentence</h3><p>Give the same problem to multiple AI models (or multiple versions of the same model), let each one write a full, independent implementation, and then make them fight it out in an automated evaluation pipeline before either one is allowed anywhere near production.</p><p>No implementation gets adopted because a model wrote it confidently. It gets adopted because it <em>won</em>.</p><h3>Why “which model is smarter” is the wrong frame</h3><p>Ask a model whether its code is good, and it will tell you yes. Ask a different model to grade the first model’s code, and you’ve just introduced a new bias: models tend to have opinions about which lab wrote something, or which coding style feels “familiar.” Self-assessment doesn’t scale, and cross-model assessment isn’t clean either, unless you strip away everything but the evidence.</p><p>So you don’t ask a model. You build a track and time the race.</p><h3>Three ways to lose, not one</h3><p>A naive version of this idea would just check: “does the code produce the correct output?” That’s necessary, but it’s a rounding error compared to what actually determines whether code survives in production. There are really three separate contests happening at once:</p><p><strong>Does it work?</strong> Same inputs, same correct outputs, no silent data leakage, reproducible results. This is table stakes, not a differentiator.</p><p><strong>Is it well-built?</strong> Two implementations can be equally correct and wildly different in quality: one has a tangle of nested conditionals, the other has clean, testable boundaries. One will be a nightmare to modify in six months. The other won’t. This is where most human-vs-human code review already lives, and it’s the layer AI comparisons usually skip entirely.</p><p><strong>Does it survive contact with reality?</strong> Correct and clean code can still be slow, memory-hungry, fragile under unusual inputs, or quietly expensive to run. You don’t find this out by reading the code; you find it by running it, ideally against real traffic in shadow mode, where it sees production data but doesn’t yet get to make production decisions. Shadow deployment isn’t a new idea either; ML teams have used it for years to test candidate models against live traffic without exposing users to the risk, and it’s already being applied to AI coding agents specifically. What’s different here is using it as a permanent step in a code-promotion pipeline rather than a one-off safety check before a single release.</p><p>A candidate has to clear all three bars. Crucially, correctness and security should function as hard gates, not as line items in a weighted average. A blazing-fast implementation that’s occasionally wrong doesn’t get to buy its way to a passing grade.</p><h3>The part that actually matters long-term</h3><p>Here’s the piece that separates this from a one-time A/B test: the evaluation harness (the tests, the constraints, the historical failure cases, the production replay data) is designed to outlive any individual model.</p><p>Today’s champion might be written by whatever model is best in mid-2026. Six months later, a newer model shows up. It doesn’t get anointed just because it’s newer and presumably smarter. It gets handed the exact same spec, the same test suite, the same production replay data, and the same currently-winning implementation as its opponent. It has to <em>prove</em> it’s better, on the same track, under the same conditions.</p><p>That turns model improvement from a marketing claim into a measurable event. “This model is better at coding” stops being a benchmark score and starts being: it beat the champion, here’s the diff, here’s the evidence.</p><p>Worth being straight about: none of the individual pieces here are new. Champion-Challenger testing, shadow deployment, and blind LLM-as-judge review each already exist as established practice in their own corners of MLOps and AI research. What doesn’t seem to exist yet is all three wired together into one pipeline built specifically around full competing code implementations, with correctness and security treated as non-negotiable gates and a formal process for a newer model to dethrone the standing champion. That combination, not any single piece of it, is the actual idea worth putting a name on.</p><h3>Keeping the judge honest</h3><p>The most important design constraint is one that’s easy to skip: the model that wrote a candidate should never be the one grading it. A layered judging structure works better:</p><ul><li>Deterministic tests and metrics do the first pass, because they can’t be flattered.</li><li>A separate model does qualitative review: readability, structure, maintainability, the things tests don’t catch.</li><li>A human makes the final call on anything high-stakes.</li></ul><p>And ideally, whoever’s judging doesn’t know which model wrote which candidate. Blind review removes the halo effect where “this is probably GPT’s code” or “this looks like a Claude answer” quietly shapes the grade before anyone’s read a line. This isn’t a new insight, either: researchers studying LLM-as-judge setups have already documented self-preference bias, where a model rates its own output more favorably, and masking model identity as one of the standard mitigations. Borrowing that safeguard for code review specifically, rather than essay grading or chatbot benchmarks, is the useful part.</p><p>There’s also a strong case for a hybrid setup: humans as the judge, AI models as the champion and challengers. For anything touching security, compliance, or safety-critical logic, a human reviewer sitting at the top of the judging chain, rather than an AI model, adds a layer of accountability that pure automation can’t replace. It also tends to be a lot more efficient than it sounds: instead of a human reading two full implementations line by line, they’re reviewing a short list of candidates that have already cleared the deterministic tests and the automated quality checks, deciding between finalists instead of triaging raw output. That combination, machines doing the exhaustive first-pass filtering and a human making the judgment call on what actually reaches production, is often the more practical version of this framework for regulated or high-risk systems.</p><h3>What this actually changes</h3><p>This isn’t really a story about finance or fraud detection. It’s a story about what happens when you stop trusting <em>sources</em> and start trusting <em>evidence</em>, a discipline that happened to get proven out first in banking and risk management, long before any of us were prompting a chatbot to write a function. The pattern already works in domains where getting it wrong costs real money on every transaction. Analysts and developers writing everyday software just haven’t caught up to a discipline other fields have already normalized.</p><p>The uncomfortable part is that this requires more infrastructure than most teams currently have: real test suites, real production replay, real promotion criteria written down before anyone sees the code. That’s also exactly why it’s valuable: it’s the version of “AI-assisted development” that doesn’t quietly rot the moment nobody’s watching.</p><p>The models will keep getting better. Other fields already figured out how to make new systems earn their spot instead of taking it on reputation. The question is whether analysts and developers will borrow that discipline before handing more and more of the codebase over to a model that’s never had to prove anything.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f00ea61eaf54" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Macroeconomics: What Happens When AI Agents Start Hoarding Crypto?]]></title>
            <link>https://ianevans256.medium.com/machine-macroeconomics-what-happens-when-ai-agents-start-hoarding-crypto-14089107b226?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/14089107b226</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Tue, 21 Jul 2026 04:17:25 GMT</pubDate>
            <atom:updated>2026-11-22T00:41:13.513Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aToy_EvpH78RUEkWXmfIXA.png" /></figure><p><em>A framework for thinking about autonomous software as a new class of economic participant, and what it would actually take for that to move crypto markets.</em></p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><blockquote><em>“Every so often, a new technology arrives that doesn’t simply change how we work; it changes who, or what, can participate in the economy.”</em></blockquote><h3>Introduction</h3><p>Every major technological leap has expanded what a person, team, or company could accomplish. The printing press amplified knowledge. The steam engine amplified physical labor. The internet amplified communication.</p><p>AI may be the next amplifier. But it introduces something genuinely new: previous technologies extended human capability. AI has the potential to perform bounded economic tasks <em>on our behalf</em>: authenticating itself, controlling a wallet, purchasing services, receiving payment, negotiating with other software. Today that infrastructure is mostly experimental: agents paying for API calls, cloud compute, or workflow coordination. But it’s being built quickly, and by serious companies.</p><p>Most of the discourse around this stops at “will AI agents own crypto wallets?” That’s the wrong question, or at least the small one. The bigger question is what happens when software becomes an economic participant in its own right, not conscious, not rights-bearing, not replacing human decision-making, but capable of earning income, managing a budget, and allocating capital within objectives a person set for it.</p><p>If that becomes commonplace, the global economy gains a new kind of participant: not a person, a business, or a government, but autonomous software acting within human-defined boundaries. I don’t think existing economic vocabulary handles that well. This essay is an attempt to start building the vocabulary (call it <strong>Machine Macroeconomics</strong>), and, more usefully, to pressure-test whether it implies anything real for crypto markets.</p><p>The headline question: could millions of AI agents earning income and holding treasuries become a persistent new source of crypto demand? The honest answer is <em>not in the form most people assume</em>, but a more precise version of the thesis is economically coherent, and worth taking seriously.</p><h3>An AI Agent Is Not Just a Better Trading Bot</h3><p>Algorithmic trading isn’t new; hedge funds and market makers have run software-driven execution for decades. But a conventional trading algorithm operates in a narrow lane: it takes capital from a human, follows a strategy, generates a return.</p><p>An economically autonomous agent is a broader animal. It might sell analytical services, buy database or API access, rent compute, pay other agents for specialized work, hold an operating reserve, stake assets for yield, or provide liquidity to a DeFi pool. Functionally, it starts to look less like a trading bot and more like a small digital business, complete with revenue, expenses, and a balance sheet.</p><p>Blockchains are a plausible home for this kind of activity because they offer permissionless access and native support for very small machine-to-machine payments: the kind of two-cent API charges that don’t make sense on traditional rails. An agent doesn’t need a bank account or credit history to control a wallet. It needs software permissions and a keypair. That’s a meaningfully lower bar to economic participation, and it’s worth taking seriously even before you get to price effects.</p><h3>Why This Might Not Need Crypto Rails At All</h3><p>Before going further, it’s worth confronting the biggest hole in this thesis directly: <strong>none of the above requires crypto.</strong></p><p>Visa and Mastercard are both building agent-payment infrastructure on top of <em>existing</em> card and settlement rails: credit lines, tokenized cards, programmatic authorization for autonomous purchases. If that infrastructure matures first, or matures more cheaply, agents could transact at scale with essentially zero on-chain footprint. In that world, “agent macroeconomics” is a real and important trend, and it implies nothing about crypto prices at all.</p><p>The counter-case for crypto rails specifically rests on a few structural advantages that fiat infrastructure doesn’t easily replicate:</p><ul><li><strong>Sub-cent settlement.</strong> Card networks have interchange economics that don’t work below a few cents per transaction. Agents transacting in fractions of a cent (the realistic unit size for API calls, inference, or data access) need a settlement layer with near-zero marginal cost. Stablecoin rails on cheap L2s currently do this better than card rails do.</li><li><strong>Programmable, permissionless counterparties.</strong> An agent doesn’t need a merchant account or a banking relationship with the agent it’s paying. It needs an address. That matters more as the <em>other party</em> is increasingly another agent, not a human-run business.</li><li><strong>Global default, not a special case.</strong> Card rails are geographically fragmented and interchange varies by region. A stablecoin wallet works the same way everywhere there’s an internet connection, which matters if agent fleets are running across jurisdictions by default.</li></ul><p>None of this guarantees crypto wins. It’s entirely plausible both rails coexist: fiat rails for agent-to-business payments where a card network is a natural fit, crypto rails for agent-to-agent micropayments where it isn’t. But it means the crypto-demand thesis is conditional on agents actually needing the specific properties on-chain rails provide, not on agents existing at all. That’s a narrower and more falsifiable claim than the popular version of this story, and it’s the one worth testing.</p><h3>The First Wave Is Probably Stablecoin Demand, Not Bitcoin Demand</h3><p>Agents doing commercial tasks need predictable purchasing power. An agent budgeting two-cent API calls can’t function well if its payment asset moves 8% in a day, so stablecoins are the natural working capital of a machine economy, not Bitcoin or Ether.</p><p>Coinbase’s x402 protocol was built specifically for instant stablecoin payments between APIs, apps, and agents. Mastercard’s Agent Pay for Machines targets the same high-frequency, low-value use case on card rails. The likely first-order effect:</p><ol><li>Humans and businesses fund agents in dollars.</li><li>Dollars convert to stablecoins.</li><li>Agents hold stablecoin operating balances.</li><li>Stablecoins circulate through an expanding agent economy.</li><li>Issuers grow reserves to back rising supply.</li></ol><p>This increases on-chain monetary liquidity. It does <strong>not</strong> automatically raise Bitcoin or Ether prices. The Cambridge Centre for Alternative Finance has specifically flagged that rising stablecoin supply doesn’t prove rising usage or demand: new stablecoins can sit idle, settle remittances, or circulate without ever converting into a volatile asset. Stablecoins are dry powder. The question that actually matters is where that powder goes.</p><h3>The Bullish Case Starts When Agents Retain Earnings</h3><p>Say a research agent earns $100/day. It might spend $50 on servers and data, send $30 back to its human owner, and place $20 in a treasury. That treasury could buy Bitcoin as a reserve asset, stake Ether, or hold DeFi positions per its investment policy.</p><p>One agent doing this moves nothing. Ten million agents doing it continuously is a measurable demand stream, <em>if</em> the underlying assumptions hold. A simple form:</p><p><strong>Agent crypto demand = active agents × average daily revenue × crypto allocation rate</strong></p><p>These inputs are currently unknowable, so treat any output as illustrative, not predictive. Three scenarios, not one point estimate:</p><p>Scenario Active agents Avg. daily revenue Allocation rate Annual demand Conservative 1M $2 5% ~$36.5M Base case 10M $5 10% ~$1.8B Aggressive 100M $10 15% ~$54.75B</p><p>The range spans nearly three orders of magnitude, which is the honest state of the evidence today; this is a framework for what to measure, not a forecast. The mechanism, though, is the interesting part: an agent treasury following a fixed allocation rule would behave like a permanently operating buyer, indifferent to sentiment, headlines, or red candles, closer to a DCA-executing pension fund than a discretionary trader.</p><h3>Accumulation, Not Transaction Volume, Is the Metric That Matters</h3><p>The common mistake here would be measuring the agent economy by transaction count. A million microtransactions can move the same $100 back and forth without creating any new net demand. What matters for price is net balance accumulation:</p><p><strong>Net agent absorption = agent purchases − agent sales − agent operational spending</strong></p><p>When that’s persistently positive, agents are collectively taking supply off the market faster than they’re returning it. Bitcoin analytics already have a version of this idea: Glassnode classifies long-term-holder supply using a probabilistic model centered on roughly 155 days of dormancy, and rising LTH supply generally signals coins becoming less sensitive to short-term price action. A future data provider could build the equivalent category for agents: <strong>agent-held supply</strong>, and more specifically <strong>agent illiquid supply</strong>: coins held by wallets with low historical spending probability. If agent balances rise while exchange-available supply falls, the price implications compound.</p><h3>Why Small Supply Changes Can Move Price a Lot</h3><p>Crypto prices are set at the margin, not by total market cap. A $100B asset might have only $2B of liquidity actively offered near the current price. A buyer doesn’t need to acquire $100B of the asset to move it, only enough to compete for what’s currently on offer:</p><p><strong>Price impact ≈ net new demand ÷ available market depth</strong></p><p>This is why the <em>distinction</em> matters more than the headline number: an agent economy that buys and continuously recycles assets back to exchanges has modest impact regardless of size. One that buys and rarely sells creates a persistent supply sink even at moderate scale. Agents progressing from Level One (liquid wallet reserves) to Level Three or Four (staked capital, productive collateral) matters more for price than raw AUM growth.</p><h3>The Bearish Mirror Image</h3><p>Agents won’t necessarily be permanent accumulators. A model optimizing for dollar-denominated returns will sell the moment it judges an asset overvalued, and a machine economy could just as easily produce faster rotation, more efficient arbitrage, automated deleveraging cascades, and crowded, correlated selling.</p><p>The correlation risk is the more structural concern. If a small number of foundation models, prompt templates, and risk engines sit underneath millions of wallets, the market isn’t gaining millions of independent decision-makers: it’s gaining millions of <em>addresses</em> controlled by a handful of underlying models. Wallet count rises; the diversity of actual decision-making does not. That’s a systemic risk in both directions, and it argues for tracking model concentration as its own metric, not inferring diversity from wallet growth.</p><h3>Wallet Growth Will Be a Noisy, Possibly Misleading Signal</h3><p>Unique-address counts are already an imperfect proxy for unique users in human markets: one person can run hundreds of wallets, one exchange can custody millions of users behind a handful of addresses. Agents make this worse. A single company might spin up a million specialized agents sharing one treasury wallet, or a million agents each with their own address. Raw wallet growth alone won’t tell you which.</p><p>Researchers will likely need behavioral fingerprints (extremely regular transaction timing, continuous 24-hour operation, machine-sized payment amounts, interaction with known agent-payment protocols, shared wallet-policy clusters, on-chain agent-registration or attestation systems) to separate human, human-supervised, and fully autonomous activity. Even with good heuristics, classification stays probabilistic, and privacy tooling will make it harder over time.</p><h3>Who Actually Owns What’s in the Wallet?</h3><p>This gets skipped too often, and it’s not a footnote: it’s load-bearing for whether “agent treasury” is even a coherent legal category.</p><p>If an agent controls a wallet and accumulates assets, who owns them? The deploying company almost certainly does today, as a matter of agency law: the agent is executing delegated authority, not acting as an independent legal person. But that raises immediate practical questions that current frameworks don’t cleanly answer: How is an agent-controlled treasury treated on a corporate balance sheet? Who is liable if a compromised or misaligned agent drains it? What happens to custody and control if the deploying entity is acquired, dissolved, or disputes ownership with a counterparty agent it transacted with?</p><p>None of this is fatal to the thesis, but it’s a real constraint on how fast “Level Three/Four” accumulation (staked, productive, illiquid capital) can scale. Institutions allocating meaningful treasury balances through autonomous software will want custody and liability frameworks that mostly don’t exist yet. That’s a genuine bottleneck, not a solved problem sitting quietly in the background.</p><h3>What a Real Agent-Liquidity Dashboard Would Track</h3><p>The thesis shouldn’t be accepted just because agents and wallets both exist; it should be tested against data as the infrastructure matures. A serious dashboard needs several layers:</p><p><strong>Activity:</strong> active agent count, agent-controlled wallets, agent-to-agent vs. agent-to-human transaction volume, average payment size, share of agents operating profitably. Coinbase’s x402 stats are an early version of this, though transaction volume alone says nothing about accumulation.</p><p><strong>Treasury balances:</strong> stablecoin and non-stablecoin holdings per wallet, treasury concentration among the largest operators, share of revenue retained vs. distributed.</p><p><strong>Net position change:</strong> purchases minus sales minus operating spend, tracked across 1/7/30/90/365-day windows. Persistent positive change is the accumulation signal.</p><p><strong>Savings rate:</strong> retained earnings ÷ total revenue. High payment volume with a near-zero savings rate means agents are payment conduits, not investors.</p><p><strong>Crypto allocation rate:</strong> non-stablecoin purchases ÷ total treasury inflows. This is the metric that separates “agents adopted blockchain rails” from “agents are buying Bitcoin.”</p><p><strong>Agent-held illiquid supply:</strong> the agent-specific analogue to LTH supply: balances that haven’t moved, are locked in staking or long-duration protocols, or rarely reach an exchange.</p><p><strong>Exchange flow:</strong> net agent deposits vs. withdrawals. Persistent withdrawals signal accumulation; persistent deposits signal distribution.</p><p><strong>Market-depth absorption:</strong> net agent purchases ÷ available order-book depth, since a purchase program small relative to market cap can still matter if it’s large relative to sell-side liquidity.</p><p><strong>Model concentration:</strong> how many wallets share a foundation model, treasury-management stack, pricing feed, or custodian, as a proxy for correlated-behavior risk.</p><p><strong>Monetary velocity:</strong> transaction volume ÷ average held balance. High velocity and flat balances say “payment rail.” Falling velocity and rising balances say “accumulation.”</p><h3>A Cleaner Version of the Model</h3><p>Four variables: N (active agents), R (average revenue per agent), S (share of revenue saved), C (share of savings allocated to a given asset). Gross demand is:</p><p><strong>D = N × R × S × C</strong></p><p>Net demand subtracts agent selling (X):</p><p><strong>Net D = (N × R × S × C) − agent sales</strong></p><p>Compared against new issuance, miner/validator selling, long-term-holder distribution, ETF flows, and available depth, the Bitcoin-specific version becomes:</p><p><strong>Agent supply pressure = net agent BTC demand − newly mined BTC sold − net distribution from other holders</strong></p><p>Sustained positive results over time are what would make agents a structural demand source rather than a rounding error.</p><h3>Three Futures</h3><p><strong>Payment pipes.</strong> Agents receive stablecoins and spend them immediately on compute, APIs, and services. Balances stay small, transaction counts explode, and non-stablecoin accumulation stays negligible. Blockchains gain utility, stablecoin issuers benefit, and Bitcoin sees only an indirect effect. This is the most realistic near-term outcome.</p><p><strong>Digital businesses.</strong> Agents earn more than they spend, hold reserves, stake assets, and allocate a share of profit to crypto treasuries. Balances grow steadily, supply tightens gradually, and agents become a genuine new investor class. This is the most plausible bullish case over a multi-year horizon.</p><p><strong>Autonomous capital layer.</strong> Agents employ other agents, hold productive assets, participate in governance, and reinvest earnings with minimal human intervention. Treasuries start to resemble corporate balance sheets, and specific networks become the infrastructure layer for machine commerce. This has the highest potential impact and, correspondingly, the highest regulatory, custody, and systemic-risk profile.</p><h3>The Security Problem Underneath All of This</h3><p>Giving an agent a wallet means giving software the ability to move real money. A hallucinating chatbot produces a bad answer; a malfunctioning financial agent produces a transaction. The industry response (spending limits, permission systems, cryptographic identity, transaction policy engines, proof-of-intent) is still early. Coinbase’s programmable wallet controls and Visa/Mastercard’s work on distinguishing authorized agents from bots are first-generation attempts, alongside academic proposals combining decentralized identity, verifiable credentials, and zero-knowledge authorization proofs.</p><p>The attack surface is real and specific: prompt injection that instructs a transfer, compromised APIs returning fraudulent payment addresses, stolen signing keys, malicious contracts, runaway transaction loops, and errors that replicate identically across an entire fleet of agents running the same model. The economic upside of agent wallets can’t be separated from the constraint systems around them. An agent needs to be free enough to transact and restricted enough that one bad inference can’t drain the treasury.</p><h3>Is the Thesis Feasible?</h3><p>Yes, but not in its popular, simple form.</p><p>The weak version (<em>more AI wallets means more crypto demand, less supply, higher prices</em>) doesn’t hold up. Wallet creation isn’t capital formation. Transaction volume isn’t accumulation. Stablecoin circulation isn’t Bitcoin buying. And a meaningful share of the underlying activity may never touch crypto rails at all if fiat-based agent payments mature first.</p><p>The stronger, narrower version holds up better: <em>profitable agents may accumulate treasury balances, and if a meaningful share of retained earnings gets allocated to scarce crypto assets and held outside liquid markets, agents could become a persistent demand source that reduces effective circulating supply.</em> That version is economically coherent and testable: the infrastructure for the first stage (wallets, stablecoin payments, agent commerce protocols) already exists and is being built by serious companies. What doesn’t yet exist at scale is a self-sustaining population of <em>profitable</em> agents that save and invest a portion of what they earn. That’s the missing variable, and it’s an empirical question the metrics above are designed to answer as the data shows up.</p><h3>Where the Investable Edge Actually Sits</h3><p>If any version of this thesis plays out, the highest-conviction exposure probably isn’t a directional bet on Bitcoin or Ether: it’s ownership of the picks-and-shovels layer that has to exist regardless of which scenario wins:</p><ul><li><strong>Stablecoin issuers and reserve managers:</strong> the first-order beneficiary in every scenario above, including the one where agents never buy a volatile asset at all.</li><li><strong>Agent-payment settlement infrastructure:</strong> the protocols and networks (x402-style rails, agent-native L2s) that capture fee volume as machine-to-machine payment counts scale, independent of whether agents accumulate or merely transact.</li><li><strong>Custody, identity, and authorization infrastructure:</strong> the credentialing, key-management, and proof-of-intent systems that institutions will require before allocating meaningful agent-controlled treasury balances. This is the bottleneck flagged above, and bottlenecks are usually where the margin sits.</li><li><strong>On-chain data and classification providers:</strong> whoever builds the agent-equivalent of LTH-supply analytics has a defensible data moat if the accumulation thesis starts showing up in the numbers.</li></ul><p>None of this requires believing the aggressive scenario. It requires believing agent-to-agent commerce grows <em>at all</em>, which is a much lower bar, and it’s the scarce upstream layer underneath a much more speculative story about token prices.</p><h3>The Final Thought</h3><p>Bitcoin’s original breakthrough was scarce digital property that could be controlled without a central authority. AI agents introduce a strange new possibility: the marginal owner of that property may not always be a person, company, or government. It may be software: working continuously, earning continuously, dividing income between expenses, savings, and investment according to rules that don’t panic at a headline or a red candle.</p><p>The next real crypto demand shock probably won’t come from a country approving an ETF. It may come from millions of small digital businesses quietly deciding that part of every dollar they earn should stay on-chain. The machines don’t need to become conscious for that to matter. They just need a wallet and a savings rate, and someone building the infrastructure that lets them keep one.</p><h3>Disclosure</h3><p>This article is intended solely for informational and educational purposes. It should not be construed as investment advice, a recommendation to buy or sell any security, or a solicitation to engage in any investment strategy.</p><p>Readers should conduct their own due diligence and consult qualified financial professionals before making investment decisions. Securities mentioned may be volatile, cyclical, exposed to geopolitical risk, and unsuitable for some investors</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14089107b226" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Don’t Bet on the Train. Bet on the Track.]]></title>
            <link>https://ianevans256.medium.com/dont-bet-on-the-train-bet-on-the-track-ef77c31e9465?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef77c31e9465</guid>
            <category><![CDATA[trains]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[market-research]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Sun, 28 Jun 2026 03:32:57 GMT</pubDate>
            <atom:updated>2026-11-22T00:38:04.092Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bqGF85YNyHz2tHhPlt4e4w.png" /></figure><h3>A Supply Chain Investment Thesis / Analysis on Maglev</h3><p><strong>June 2026</strong></p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>“The gold rush didn’t make millionaires out of prospectors. It made them out of the guys selling pickaxes.”</p><p>That aphorism is old enough to be embarrassing to cite, and yet here we are, watching a generation of infrastructure boosters draw up maglev schematics on napkins while the real money sits upstream, patient and quietly compounding.</p><p>This is not a call on whether America will soon have a 311 mph magnetic levitation train shuttling people between Washington and New York in under an hour. (It probably will not. Not soon. More on that shortly.)</p><p>This is a call on the materials, components, industrial gases, magnets, electrical steels, cryogenic systems, and defense-grade control architectures that must exist regardless of whether any single American maglev corridor ever breaks ground.</p><p>That difference matters enormously for where your capital should live.</p><h4><strong>I. The Setup: A Technology in Search of a Country</strong></h4><p>Maglev is not science fiction.</p><p>Japan’s L0 Series superconducting maglev train reached 603 km/h, or roughly 375 mph, during testing in 2015, setting the world speed record for a manned train.[1] Germany developed the Transrapid maglev system, whose technology ultimately found commercial application in Shanghai, where a Transrapid-based line began public commercial service in 2004.[2] China continues to operate commercial maglev systems today and has remained one of the most visible global test beds for the technology.</p><p>The engineering is not the problem.</p><p>The problem is that engineering is only one participant in the American infrastructure, and unfortunately for engineers, it does not get the final vote.</p><p>In the United States, the Northeast Maglev project sought to bring JR Central’s SCMAGLEV technology to the I-95 corridor, beginning with a proposed Baltimore-to-Washington segment. That project suffered a setback when the Federal Railroad Administration formally rescinded its Notice of Intent to prepare an Environmental Impact Statement on August 4, 2025.[3]</p><p>Out West, the long-discussed California-Nevada maglev concept has effectively been cleared from the field by Brightline West, which is building a conventional all-electric high-speed rail system between Las Vegas and Southern California. Brightline West’s project is planned as a 218-mile passenger rail line running largely along the I-15 corridor to Rancho Cucamonga.[4]</p><p>The map of American maglev 2026:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/761/1*aT4YI9ozjoIeOSwGQ2X5PA.png" /></figure><p>The takeaway is not that maglev is doomed as a technology. The takeaway is that U.S. maglev infrastructure projects currently represent one of the weakest risk-reward profiles in transportation investing: enormous upfront capital requirements, decade-long development timelines, complicated land-use issues, and uncertain political outcome risk</p><p>Shorting the developers is not even particularly interesting. Most are private, underfunded, or already fighting uphill battles.</p><p>The interesting trade lives somewhere else entirely.</p><p>It lives in the things the train needs to make it floats</p><h3>II. The Materials Play: What Has to Exist Regardless</h3><p>Here is the structural insight most transportation-focused investors miss:</p><p>The specialized materials that make maglev work are not maglev-only inputs.</p><p>They are also the feedstock for fusion reactor magnets, MRI machines, particle accelerators, EV motors, wind turbines, grid infrastructure, transformers, and next-generation power systems.</p><p>Maglev is not the thesis.</p><p>Maglev is a demand vector.</p><p>The real thesis is that the inputs required to make a train float are the same inputs required to electrify, automate, defend, and industrially upgrade the modern economy.</p><p>You do not need a single American maglev tunnel to be dug for this thesis to work.</p><p>You need the world to keep buying magnets, electrical steel, cryogenic gases, superconducting wire, and high-performance materials.</p><h4>Superconducting Materials &amp; Cryogenics</h4><p>Superconducting maglev systems, especially high-speed SCMAGLEV designs, generate magnetic fields using superconducting coils cooled to extremely low temperatures. The classic workhorse superconducting materials include niobium-titanium, or NbTi, and niobium-tin, or Nb3Sn. In parallel, the industry continues to explore higher-temperature superconducting materials, including REBCO tapes, which can operate at less punishing cryogenic conditions than traditional low-temperature superconductors.[5]</p><p>The key point is not that one material wins forever.</p><p>The key point is that the world is moving toward more high-field magnet systems, more cryogenic infrastructure, and more demand for specialty materials that can operate under ridiculous physical conditions without quitting like a normal metal would.</p><p>Running any of this requires industrial cryogenic infrastructure: liquid helium, liquid nitrogen, cryogenic tanks, distribution networks, cooling systems, and the engineering capability to keep magnets cold enough to behave like physics professors promised they would.</p><p>This is where the investment lens starts to shift.</p><p>A maglev developer needs the line to be approved.</p><p>A cryogenic gas supplier needs the world to keep building advanced industrial systems.</p><p>Those are not the same risk profile.</p><h3>The plays:</h3><p><strong>ATI Inc. (NYSE: ATI)</strong><br>ATI is a major producer of specialty materials and advanced alloys used across aerospace, defense, energy, and high-performance industrial applications. The maglev angle is not that ATI becomes “the maglev stock.” The angle is that advanced mobility, aerospace, defense systems, and energy infrastructure all require materials that tolerate stress, heat, cold, fatigue, corrosion, and precision manufacturing constraints.</p><p>Maglev becomes an additive demand narrative layered on top of a much broader advanced-materials base.</p><p>That is exactly the kind of boring-adjacent exposure I like. Not boring in the business quality sense. Boring in the “nobody on CNBC is yelling about it yet” sense.</p><p><strong>Air Products and Chemicals (NYSE: APD) and Linde plc (NASDAQ: LIN)</strong><br>Air Products and Linde are two of the major global industrial gas players, with exposure to helium, nitrogen, hydrogen, oxygen, and broader cryogenic supply chains.[6]</p><p>If superconducting systems scale anywhere in the world, these firms are part of the “make it cold and keep it cold” economy.</p><p>They do not need to care whether the end customer is a maglev train, a fusion tokamak, a semiconductor facility, a hospital MRI system, a space program, or a particle physics lab.</p><p>They deliver the cold.</p><p>And in a world increasingly built on high-field magnets, advanced electronics, chipmaking, hydrogen infrastructure, and extreme-temperature systems, delivering the cold is not a side quest.</p><p>It is infrastructure.</p><h3>Rare Earth Elements: The NdFeB Magnet Story</h3><p>Not all maglev systems are superconducting. Permanent magnet and electromagnetic suspension designs rely heavily on magnetic materials, including neodymium-iron-boron permanent magnets, commonly referred to as NdFeB magnets.</p><p>The relevant rare earth inputs include neodymium and praseodymium, often grouped as NdPr, with dysprosium and terbium used in certain high-performance magnets to improve temperature resistance and coercivity.</p><p>These are not just “train materials.”</p><p>They are EV motor materials.</p><p>Wind turbine materials.</p><p>Robotics materials.</p><p>Factory automation materials.</p><p>The modern electrified economy has a quiet addiction to high-performance permanent magnets, and China has spent decades making sure it controls a lot of that supply chain.</p><p>Rare earths are not actually rare in the way the name implies. What is rare is economically viable mining, separation, refining, metallization, alloying, and magnet manufacturing capacity outside China.</p><p>That is the bottleneck.</p><p>And bottlenecks are where supply chain investors should start paying attention.</p><h3>The play:</h3><p><strong>MP Materials Corp. (NYSE: MP)</strong><br>MP Materials operates the Mountain Pass rare earth mine in California and has positioned itself as the most direct public-market expression of U.S. rare earth supply chain re-shoring. The company describes itself as America’s only fully integrated rare earth producer, spanning mining, processing, metallization, and magnet manufacturing.[7]</p><p>More importantly, MP is moving downstream.</p><p>The company’s Fort Worth, Texas facility, known as Independence, is designed to produce NdFeB magnets from refined rare earth materials. MP has stated that commercial NdPr metal production is already online and that trial production of sintered magnets is underway, with a gradual production ramp targeted to begin in late 2025.[8]</p><p>That changes the story.</p><p>A miner sells rocks.</p><p>A vertically integrated magnet producer sells strategic industrial leverage.</p><p>MP does not need a maglev customer to matter. Its demand stack includes EV motors, robotics, electronics, defense systems, wind turbines, aerospace, and grid-adjacent electrification. Maglev is not the driver. It is another confirmation that high-performance magnet demand keeps finding new places to hide.</p><p>If you are looking for the purest public-market “domestic magnet supply chain” play, MP is probably the cleanest expression.</p><h3>Structural Materials: The Guideway Economy</h3><p>This is the deeply unsexy part of maglev.</p><p>Naturally, it is also where a lot of the money lives.</p><p>A high-speed maglev system does not merely require a train. It requires a guideway: a grade-separated, precision-engineered structure that must hold tolerances tight enough for levitation and propulsion systems to work safely at extreme speed.</p><p>That means concrete.</p><p>Steel.</p><p>Aggregates.</p><p>Electrical steel.</p><p>Power electronics.</p><p>Embedded stators.</p><p>Construction labor.</p><p>Land.</p><p>Substations.</p><p>Maintenance infrastructure.</p><p>And enough permitting paperwork to deforest a small country.</p><p>Maglev guideways are not normal railroad tracks. In many designs, the propulsion system is effectively built into the guideway itself through linear motor infrastructure. That shifts cost and complexity away from the vehicle and into the track.</p><p>The train floats.</p><p>The capital does not.</p><p><strong>The plays:</strong></p><p><strong>Cleveland-Cliffs Inc. (NYSE: CLF) and Nucor Corporation (NYSE: NUE)</strong><br>Cleveland-Cliffs and Nucor are two major domestic steel producers. Both would have exposure to broad infrastructure and industrial steel demand. Cleveland-Cliffs is particularly relevant because it produces electrical steels, including non-oriented electrical steel used in motors, generators, and EV applications.[9]</p><p>Non-oriented electrical steel matters because electrification is not just about batteries. It is about motors, transformers, generators, magnetic cores, and high-efficiency power conversion hardware.</p><p>Maglev guideways, EV drivetrains, industrial motors, and grid upgrades may look like different markets on the surface.</p><p>Underneath, they all pull on the same deeper theme:</p><p>The world is getting more electrically intensive. That means more demand for specialized steel, not less.</p><p><strong>Vulcan Materials Company (NYSE: VMC)</strong><br>Vulcan Materials is the largest producer of construction aggregates in the United States, supplying crushed stone, sand, gravel, and other foundational materials used across construction markets.[10]</p><p>This is the least glamorous name in the thesis, which is usually a sign that it deserves more attention.</p><p>Concrete guideways require aggregate.</p><p>Roads require aggregate.</p><p>Bridges require aggregate.</p><p>Data centers require aggregate.</p><p>Factories require aggregate.</p><p>Substations require aggregate.</p><p>The economy can pretend it runs on software, but every serious physical project eventually meets a pile of crushed rock.</p><p>Vulcan is the literal foundation of this trade.</p><p>It does not care what is floating above the track.</p><h3>III. Advanced Systems Integration: The Defense Crossover</h3><p>The engineering required to build superconducting magnets and linear motor systems at transit scale overlaps heavily with the engineering required for electromagnetic systems.</p><p>That overlap is not incidental.</p><p>It is the entire point.</p><p>Once you are dealing with high-field magnets, millimeter-level tolerances, high-speed motion, power electronics, real-time control systems, thermal management, and electromagnetic forces large enough to ruin everyone’s afternoon, you are no longer just in the transportation business.</p><p>You are in the advanced systems business.</p><p>That is why defense contractors, aerospace suppliers, controls companies, and magnet technology firms deserve attention.</p><p><strong>General Atomics</strong> is private, but it is an important reference point. General Atomics operates a Magnet Technologies Center in San Diego and cites a long history in magnet technology development, including superconducting magnet systems for fusion research, MRI applications, motors, and defense applications.[11] The company also references the Holloman High-Speed Maglev Test Track in connection with its magnet technology work.[12]</p><p>That matters because it shows where the hard technology actually lives.</p><p>Not in the renderings.</p><p>Not in the ribbon-cutting ceremonies.</p><p>Not in the mayoral press conferences.</p><p>In the companies that know how to make magnets, propulsion systems, power electronics, and control architectures behave under extreme conditions.</p><p><strong>Honeywell International Inc. (NASDAQ: HON) and Curtiss-Wright Corporation (NYSE: CW)</strong> sit closer to the controls, automation, aerospace, and defense electronics layer.</p><p>Maglev systems need real-time control. They maintain narrow gaps between train and guideway using continuous sensor feedback and active stabilization. At high speed, “close enough” is not an engineering standard. It is a lawsuit wearing a helmet.</p><p>Curtiss-Wright supplies sensors, controls, subsystems, and mission-critical components across aerospace, defense, industrial, and commercial markets.[13] Honeywell, through its aerospace and automation businesses, has long-standing exposure to sensors, controls, avionics, defense systems, and industrial automation.</p><p>Again, maglev is not the sole thesis.</p><p>The thesis is that advanced mobility and defense systems increasingly require the same capabilities: rugged electronics, high-reliability control loops, sensor fusion, power management, thermal management, and integration under conditions where failure is not adorable.</p><p>The train is optional.</p><p>The control layer is not.</p><h3>IV. The Adjacent Demand Thesis: Why You Do Not Need the Train</h3><p>Here is why this is not a binary infrastructure bet.</p><p>The same superconducting magnets required for SCMAGLEV are also relevant to fusion energy development, MRI systems, research magnets, and particle accelerators.</p><p>ITER, the international fusion project, has required superconducting strand at a scale that dwarfs any single transit corridor. ITER states that its magnet systems required approximately 500 metric tonnes of Nb3Sn superconducting strand, equivalent to more than 100,000 kilometers of strand.[14]</p><p>That is not a train story.</p><p>That is an industrial magnet story.</p><p>The same rare earth NdFeB magnets relevant to certain maglev designs are also critical in EV traction motors, offshore wind turbines, industrial robotics, etc.</p><p>That is not a train story.</p><p>That is a permanent magnet supply chain story.</p><p>The same electrical steels relevant to linear motor stators are also used in EV motor laminations, generators, and high-efficiency transformer cores.</p><p>That is not a train story.</p><p>That is an electrification story.</p><p>The same cryogenic gases needed for superconducting systems are used across medical imaging, aerospace, semiconductor production, industrial processes, hydrogen systems, scientific research, and fusion experimentation.</p><p>That is not a train story.</p><p>That is a cold-chain-for-physics story.</p><p>What maglev represents, correctly understood, is not a transit investment.</p><p>It is a signal.</p><p>A signal that the underlying supply chains required for magnetic levitation are also demand-constrained inputs to a widening set of critical industries.</p><p>Rare earths.</p><p>Superconducting alloys.</p><p>Electrical steels.</p><p>Industrial gases.</p><p>Cryogenic infrastructure.</p><p>Power electronics.</p><p>Controls.</p><p>Advanced materials.</p><p>These are the picks and shovels.</p><p>The train is just the guy in the river hoping there is gold.</p><h3>V. Risk Factors, Because We Are Not Animals</h3><p>No thesis gets to walk into the room without its risk section.</p><p>China’s dominance across rare earth processing and magnet supply chains remains the single largest systemic risk to the NdFeB magnet ecosystem. Beijing has demonstrated a willingness to use export controls as a geopolitical tool, and rare earth materials sit squarely inside that vulnerability.[15]</p><p>MP Materials is the domestic hedge against this risk.</p><p>But hedges take time to scale.</p><p>Mining is hard.</p><p>Processing is harder.</p><p>Magnet manufacturing is harder still.</p><h4>Commodity cycle sensitivity</h4><p>Cleveland-Cliffs, Nucor, Vulcan Materials, and ATI all retain exposure to industrial and commodity cycles. A long-term infrastructure or electrification thesis does not immunize them from recessions, credit contractions, weak construction demand, or pricing pressure.</p><p>A beautiful secular thesis can still get punched in the face by a bad ISM print.</p><p>This is especially true for steel and aggregates.</p><p>The market may eventually reward the long-term demand story, but it will not do so in a straight line just because we wrote a nice article about magnets.</p><h4>Superconductor substitution risk</h4><p>Materials transitions matter.</p><p>Traditional low-temperature superconductors such as NbTi and Nb3Sn remain important, but high-temperature superconductors like REBCO tapes could shift future demand patterns. If HTS systems become cheaper, more scalable, and more manufacturable, the value chain could migrate away from older materials toward new suppliers.</p><p>And if room-temperature superconductors ever become real at commercial scale, then several Nobel Prizes will be handed out and this entire section will need to be rewritten by candlelight while physicists celebrate in the streets.</p><p>Until then, monitor the materials transition carefully.</p><h4>Project optionality evaporating</h4><p>The FRA rescission of the Northeast Maglev environmental review removes a potential near-term U.S. maglev catalyst.[16]</p><p>That matters.</p><p>The thesis remains intact because it does not rely on a U.S. maglev buildout. But investors should not underwrite a near-term maglev-specific revenue event as the base case.</p><p>The right framing is optionality.</p><p>If American maglev happens, the supply chain benefits.</p><p>If American maglev does not happen, the supply chain still has EVs, wind, fusion, defense, grid modernization, MRI, robotics, and industrial electrification.</p><p>That is the beauty of the trade.</p><p>It is not train-or-bust.</p><p><strong>It is train-as-upside.</strong></p><h3>Valuation risk</h3><p>The right thesis at the wrong price is still capable of stealing your lunch money.</p><p>Rare earth names can become narrative stocks. Steel names can look cheap for reasons that are not temporary. Industrial gas companies can trade at premium multiples because they are high-quality compounders. Defense-adjacent suppliers can get bid up when everyone discovers the same “picks and shovels” idea at once.</p><p>Being early is good.</p><p>Being early and disciplined is better.</p><p>Being early, undisciplined, and convinced you are a genius because you used the word “maglev” in a stock pitch is how portfolios end up in timeout.</p><h3>VI. The Portfolio Construction View</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*XPAeSFzm4u2ZGpvfivP9hQ.png" /></figure><p>No position here requires a maglev train to actually run in the United States. Every position benefits if one does.</p><p><strong>Bloomberg Terminal RV (Relative Valuation) Sector Analysis</strong></p><p>The core portfolio targets are grouped by their specific GICS sectors and compared against clean, external industry peers not explicitly detailed in the thesis body.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/535/1*6Pe3qigogs0Id_mhKQYROA.png" /></figure><h3>VII. The Bottom Line</h3><p>The investment case for maglev in America has never really been about the trains.</p><p>The trains may come eventually.</p><p>Probably later than expected.</p><p>Probably in a corridor nobody modeled correctly.</p><p>That is not a trade.</p><p>That is a prayer with a feasibility study attached.</p><p>The better trade is the supply chain that makes magnetic levitation physically possible:</p><p>The rare earths.</p><p>The superconducting alloys.</p><p>The cryogenic infrastructure.</p><p>The precision electrical steels.</p><p>The industrial gases.</p><p>The controls.</p><p>The structural materials.</p><p>The power systems.</p><p>These inputs are already experiencing secular demand from industries that are building right now: EVs, offshore wind, robotics, defense systems, fusion experiments, grid infrastructure, medical imaging, aerospace platforms, and advanced manufacturing.</p><p>Maglev is not the main course.</p><p>Maglev is the garnish that happens to be floating.</p><p>The pickaxe sellers are already working.</p><p>The prospectors are still drawing maps.</p><p>And Thomas the Tank Engine is in the shop prepping to get retrofitted</p><h3>Footnotes</h3><p>[1] JR Central / Central Japan Railway Company, SCMAGLEV technology materials; see also contemporary reporting on the 2015 L0 Series test record of 603 km/h.</p><p>[2] Shanghai Maglev / Transrapid historical materials. The Shanghai line began commercial service in 2004 using German Transrapid-based technology.</p><p>[3] Federal Railroad Administration, “Rescinding the Notice of Intent To Prepare an Environmental Impact Statement: Baltimore to Washington Superconducting Maglev Project,” Federal Register, August 4, 2025.</p><p>[4] Brightline West, Project Overview. Brightline West describes the project as a 218-mile passenger rail service connecting Las Vegas to Rancho Cucamonga, largely in the I-15 corridor.</p><p>[5] National Academies and superconducting materials literature on NbTi, Nb3Sn, and REBCO high-temperature superconducting tapes.</p><p>[6] Air Products and Chemicals company materials; Linde company materials. Both companies supply industrial gases and cryogenic gases including helium and nitrogen.</p><p>[7] MP Materials company materials. MP describes itself as America’s only fully integrated rare earth producer, with capabilities from mining and processing through metallization and magnet manufacturing.</p><p>[8] MP Materials investor release, “MP Materials Restores U.S. Rare Earth Magnet Production,” January 22, 2025.</p><p>[9] Cleveland-Cliffs, Motor-Max non-oriented electrical steel materials. Cleveland-Cliffs identifies Motor-Max as non-oriented electrical steel for motors, generators, and EV applications.</p><p>[10] Vulcan Materials company materials. Vulcan describes itself as the nation’s largest producer of construction aggregates.</p><p>[11] General Atomics, Magnet Technologies Center materials. General Atomics cites a long history in magnet technology development, including fusion, MRI, motor, and defense applications.</p><p>[12] General Atomics, Magnet Technologies Center materials; Holloman High-Speed Maglev Test Track references.</p><p>[13] Curtiss-Wright company materials. Curtiss-Wright describes its aerospace and industrial businesses as providing sensors, controls, subsystems, and mission-critical components.</p><p>[14] ITER Organization, superconducting magnet system materials. ITER states that its magnet systems required approximately 500 metric tonnes of Nb3Sn superconducting strand, equivalent to more than 100,000 kilometers.</p><p>[15] Reuters and related reporting on China rare earth export controls and restrictions involving U.S. entities in the rare earth supply chain.</p><p>[16] Federal Railroad Administration, August 4, 2025 Federal Register notice rescinding the NOI for the Baltimore-Washington SCMAGLEV environmental impact statement.</p><h3>References</h3><p>Federal Railroad Administration. “Rescinding the Notice of Intent To Prepare an Environmental Impact Statement: Baltimore to Washington Superconducting Maglev Project.” Federal Register. August 4, 2025.</p><p>Brightline West. “Project Overview.”</p><p>Central Japan Railway Company / JR Central. “SCMAGLEV Technology Overview.”</p><p>ITER Organization. “ITER Superconducting Magnet System.”</p><p>MP Materials. “MP Materials Restores U.S. Rare Earth Magnet Production.” Investor Relations. January 22, 2025.</p><p>MP Materials. Company materials on Mountain Pass, Independence, rare earth processing, metallization, and NdFeB magnet manufacturing.</p><p>Cleveland-Cliffs. “Motor-Max Non-Oriented Electrical Steel.”</p><p>Vulcan Materials Company. Company materials on construction aggregates.</p><p>General Atomics. “Magnet Technologies Center.”</p><p>Air Products and Chemicals, Inc. Company materials on industrial gases, helium, nitrogen, and cryogenic supply.</p><p>Linde plc. Company materials on industrial gases, helium, nitrogen, and cryogenic systems.</p><p>Curtiss-Wright Corporation. Company materials on aerospace, defense, sensors, controls, subsystems, and mission-critical components.</p><p>Honeywell International Inc. Company materials on aerospace, automation, controls, sensors, and defense technology.</p><p>Reuters. Reporting on China rare earth export controls and U.S.-China critical minerals tensions.</p><h3>Disclosure</h3><p>This article is intended solely for informational and educational purposes. It should not be construed as investment advice, a recommendation to buy or sell any security, or a solicitation to engage in any investment strategy.</p><p>Readers should conduct their own due diligence and consult qualified financial professionals before making investment decisions. Securities mentioned may be volatile, cyclical, exposed to geopolitical risk, and unsuitable for some investors</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef77c31e9465" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Room is the First Draft of the Pocket]]></title>
            <link>https://ianevans256.medium.com/the-room-is-the-first-draft-of-the-pocket-3c581f4818ba?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c581f4818ba</guid>
            <category><![CDATA[math]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Mon, 22 Jun 2026 00:57:30 GMT</pubDate>
            <atom:updated>2026-11-22T00:42:08.726Z</atom:updated>
            <content:encoded><![CDATA[<p>Every few centuries, we invent a new way to look at numbers, and every single time, we build a cathedral to house it.</p><p><em>(Author’s note: I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.)</em></p><p>If you walked into a high-performance computing data center today, you would see rows of monolithic, liquid-cooled server racks. They hiss with the sound of coolant pumps. Inside their guts, engineers are trying to replace copper wires with lasers, building the first generation of silicon photonics: computers that calculate at the speed of light. These systems are massive, temperamental, and take up entire rooms to solve a single bottleneck. Electrons are simply too slow and too hot to train the next frontier of artificial intelligence.</p><p>It feels bleeding-edge. It feels like the absolute pinnacle of human achievement.</p><p>But if you look closely at the history of technology, you realize something comforting and slightly hilarious: we have been in this exact room before.</p><p>Humanity has a beautiful, unbroken habit of making our tools impossibly large before we figure out how to make them elegant. Every revolution in infrastructural design begins as a room-sized monument and ends as a commodity slipping into your pocket.</p><h3>The Shadow on the Landscape (c. 1500 BCE)</h3><p>Before time was an intimate companion ticking quietly on our wrists, it was a geographic coordinate. To check the hour in ancient civilizations, you didn’t glance down; you traveled to a monument.</p><p>Around 1500 BCE, ancient Egyptian astronomers and builders constructed sophisticated shadow clocks and public obelisks (Landes, 2000). These stone giants were massive public infrastructure projects. To map out the passage of the day, engineers had to erect solid granite pillars stretching dozens of feet into the sky. The physics of timekeeping required the literal alteration of a city’s skyline, using the rotation of the Earth to sweep a long needle of darkness across vast, calibrated stone plazas.</p><p>Time was heavy, public, and completely immovable. It was a physical destination. And, if you were trapped indoors or if an overcast sky rolled across the Nile, your access to the hour vanished entirely.</p><h3>The Five-Foot Marble Slab (c. 300 BCE)</h3><p>When we think of an abacus, we picture a handy little wooden frame with sliding beads that a child uses in a math competition. But that is the modernized form of the technology, not the starting point.</p><p>Before the portable bead-and-rod frame was popularized, ancient tax collectors and astronomers used counting boards. The oldest surviving example we have is the Salamis Tablet, discovered on a Greek island in 1846 (Ifrah, 2001). Crafted around 300 BCE, it is a massive, unforgiving slab of solid white marble measuring roughly five feet long, two and a half feet wide, and weighing hundreds of pounds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NKyFKrk_HiSmr5e6" /><figcaption>The architecture of early grid-based counting systems. Source: pacopole / Getty Images</figcaption></figure><p>You didn’t carry the Salamis Tablet to work. It was the work.</p><p>It was a permanent piece of imperial infrastructure set up in a wealthy merchant’s estate or a government treasury. To calculate the cost of a military campaign or a harvest tax, you stood over this massive marble table like a dealer at a casino craps table, manually moving heavy metal tokens across deeply engraved grid lines.</p><p>The word <em>calculus</em> itself comes from this exact room. It is the Latin word for the pebbles moved across those stone tables (Ifrah, 2001).</p><p>For centuries, this was the peak of mathematical infrastructure. But eventually, a traveling merchant looked at that five-hundred-pound piece of rock and realized the physical layout was a logistical nightmare. You couldn’t load a marble table onto a donkey to audit grain prices three towns over.</p><p>So, engineers began the slow, painful process of shrinking the physics. The Roman Empire engineered the bronze hand abacus with tiny grooved tracks, and by the Ming Dynasty, the architecture evolved into the ultra-portable wood-and-bead <em>suanpan</em>. The room-sized calculator became a tool you could slide into a silk sleeve.</p><h3>The Cathedral of Gears (c. 1386 CE)</h3><p>As European trade routes expanded and towns grew into crowded economic hubs, relying on the sun became a liability. Medieval societies needed a way to track time continuously, independent of clouds or nightfall. The solution was the mechanical clock, but its birth required building another mechanical cathedral.</p><p>Constructed around 1386, the Salisbury Cathedral Clock is widely celebrated as the oldest working mechanical clock mechanism in existence (Landes, 2000). It has no face, no hands, and no dial. Instead, it is a room-sized, open-frame cage fashioned entirely from hand-forged wrought iron.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VB_ClHkf4LtOQu8z" /><figcaption>The Salisbury Cathedral Clock mechanism (c. 1386). Source: Folb / Getty Images</figcaption></figure><p>Look closely at its design: the system relies on sheer mass and gravity. The power is generated by two gargantuan stone weights that slowly descend from the cathedral tower, unwinding heavy ropes wrapped around thick wooden barrels. To regulate this raw energy, early blacksmiths invented the “verge and foliot” escapement — a heavy horizontal beam that swings back and forth, arresting and releasing the gears with an industrial, echoing clunk.</p><p>The clock didn’t show you the time; it boomed the time across the countryside by striking an enormous bell. It was a roaring, community-wide oracle locked inside a stone chamber.</p><h3>Slipping Time into a Gilded Sphere (1505 CE)</h3><p>For over a century, the physics of mechanical timekeeping remained bound to gravity and architecture. If you tilted a verge and foliot clock, or if its heavy stone weights hit the ground, the logic collapsed.</p><p>Then came Peter Henlein, a brilliant locksmith and watchmaker from Nuremberg. In 1505, Henlein achieved an absolute masterpiece of miniaturization by engineering what is widely recognized as the world’s first portable watch: the <em>Watch 1505</em> (or the Pomander Watch) (Struthers, 2024).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/543/0*OgcBEl8TrD1fUR1-" /><figcaption>Peter Henlein’s Pomander Watch (1505). Source: Quill &amp; Pad</figcaption></figure><p>Henlein took the sprawling architecture of the cathedral clocktower and completely rethought the medium. He replaced the massive, gravity-dependent descending stone weights with a tightly coiled iron mainspring. By packing a miniature torsion pendulum and spring mechanism into a fire-gilded copper sphere just over four centimeters in diameter, Henlein freed time from architecture.</p><p>For the first time in human history, time wasn’t something you walked toward or listened for across a valley. It was a beautifully crafted, pocket-sized companion that ticked softly in the palm of your hand, operating perfectly regardless of your posture or location.</p><h3>The Five-Ton Locomotive of Logic (1846 CE)</h3><p>We forgot this lesson, of course, and had to learn it all over again during the Industrial Revolution.</p><p>In the nineteenth century, human computers spent years manually calculating printed maritime navigation tables. Because the work was tedious, the printed pages were riddled with mathematical errors that caused ships to wreck. Frustrated by this, mathematician Charles Babbage decided to automate thinking entirely (Swade, 2005). Between 1846 and 1848, he drafted the final designs for his legendary Difference Engine №2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xv99r7E11_5znz_R" /><figcaption>Babbage’s Difference Engine №2. Source: Science &amp; Society Picture Library / SSPL via Getty Images</figcaption></figure><p>It was not a desktop calculator. It was a mechanical monolith made of cast iron and gleaming brass, consisting of eight thousand independently moving parts, weighing a staggering five metric tons, and stretching eleven feet long (Swade, 2005).</p><p>It looked less like a computer and more like a steam locomotive. To compute complex polynomial equations using the method of finite differences, an operator had to stand in front of this massive wall of technology and exert continuous physical energy to turn a heavy manual crank handle.</p><p>Babbage’s engine was a room-sized proof of concept. The physics of interlocking brass gears were simply too heavy to scale down easily with nineteenth-century tools.</p><h3>The Roaring Monolith of Code (1940 CE)</h3><p>Nearly a century later, humanity faced a crisis of pure cryptographic mathematics. During World War II, the German military encrypted its tactical communications using the Enigma machine, a device that scrambled text through billions of possible rotor configurations. Human cryptologists couldn’t parse the numbers fast enough. The math had to be beaten by a machine.</p><p>In 1939 and 1940, a team at Bletchley Park led by Alan Turing and Gordon Welchman engineered the British Bombe (Copeland, 2006).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5yuhVw7OigeW6nBC" /><figcaption>The electromechanical British Bombe at Bletchley Park (c. 1940). Source: WPA Pool / Getty Images</figcaption></figure><p>The Bombe was an electromechanical beast. Housed in a heavy, room-sized cabinet, it weighed over a ton and required a constant, dedicated power source to drive its internal relays. Its face was a mesmerizing, terrifying wall of 108 brightly colored, rotating cylindrical drums, which explicitly replicated the internal mechanics of 36 separate German Enigma machines working simultaneously (Copeland, 2006).</p><p>When operational, the room roared. The uniselectors and spinning brushes clicked so quickly and loudly that workers described the sound as a room filled with thousands of hyper-speed knitting needles. It was a physical, clattering cathedral built to hunt through trillions of algebraic permutations before the German keys changed at midnight. It was magnificent, vital — and completely trapped inside an industrial codebreaking facility.</p><h3>The Electronic Giants (1945 CE — 1960s)</h3><p>While Turing’s electromechanical drums were spinning in England, American engineers were preparing to trade mechanical gears for pure electricity. In 1945, the University of Pennsylvania unveiled the ENIAC (Electronic Numerical Integrator and Computer), the world’s first programmable, electronic, general-purpose digital computer (Shurkin, 1996).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qtsvmKvV7GWZBLUi" /><figcaption>The room-sized ENIAC architecture (1945). Source: Apic / Getty Images</figcaption></figure><p>ENIAC completely redefined scale. It occupied roughly 1,800 square feet of floor space, weighed thirty tons, and utilized nearly 18,000 glowing vacuum tubes (Shurkin, 1996). When technicians turned it on, it drew 150 kilowatts of electricity — enough to dim the lights of nearby Philadelphia neighborhoods. The vacuum tubes shed so much intense thermal energy that the room required massive, dedicated cooling fans, and tubes blew out every few hours, forcing engineers to scramble through the machine’s cavernous interior to swap them out.</p><p>By the 1960s, this paradigm evolved into the monolithic IBM mainframes. These massive cabinets filled heavily air-conditioned corporate and government basements, requiring teams of specialized operators dressed in lab coats just to feed them punch cards and monitor their spinning magnetic tape reels. This was the computational infrastructure NASA used to calculate the orbital trajectories for the Gemini and Apollo missions.</p><h3>The Masterpiece of Miniature Mechanics (1948 CE)</h3><p>As electronic vacuum tubes were claiming entire basements, a parallel, poetic miracle occurred back in the realm of mechanical logic.</p><p>In the late 1930s, a brilliant Austrian engineer named Curt Herzstark was working on a way to condense multi-axis calculation into something portable. His work was brutally interrupted when he was imprisoned in the Buchenwald concentration camp during World War War II. Under horrific conditions, his captors ordered him to draw a detailed blueprint of a miniature calculator as a gift for Adolf Hitler. Herzstark spent his nights drawing, completing the flawless technical architecture for a device that completely bypassed the need for a room-sized layout (Herzstark, 2005).</p><p>Following the liberation of the camp, Herzstark brought his designs to Liechtenstein, and in 1948, the Curta Calculator was born.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jKP8qsnPlFvwPNdc" /><figcaption>The Curta handheld mechanical calculator (1948). Source: ullstein bild Dtl. / ullstein bild via Getty Images</figcaption></figure><p>The Curta was a marvel of extreme mechanical miniaturization. It was a sleek, black metal cylinder that fit comfortably in the palm of a single hand, weighing just about eight ounces. Inside its tiny frame sat a tiered “step-drum” mechanism that folded Charles Babbage’s entire five-ton Difference Engine into a space no larger than a pepper grinder.</p><p>By adjusting small vertical sliders along the cylinder’s body and turning a tiny, smooth crank on top, field engineers, pilots, and race car drivers could perform complex addition, subtraction, multiplication, and division anywhere on Earth. Once again, the mechanical cathedral had been refined into a pocket knife.</p><h3>The Modern Cathedral: Silicon and Light (2026 CE)</h3><p>Today, the smartphone sitting next to your coffee cup holds millions of times more computing power than the entire room NASA used to put a man on the moon. We look at that transition and think we’ve won the game. We think silicon chips have permanently broken the cycle.</p><p>But we are hitting the physical wall again.</p><p>Traditional silicon microchips rely on microscopic copper wires to pass data back and forth. When you try to cluster one hundred thousand state-of-the-art graphics processing units (GPUs) together inside an industrial server center to train a massive, frontier artificial intelligence model, physics rebels. The sheer volume of electrons moving through that copper creates immense electrical resistance. The wires literally turn into tiny space heaters. Data bottlenecks, the chips choke on their own thermal output, and the system stalls.</p><p>You can see it in the architecture of data centers on Youtube where they highlight cooling infrastructure is one of their key challenges. You see it slightly in my previous medium article about Aspen Aerogels where I’m looking short-term to how to trade on solutions to cooling copper cpus during the AI boom.</p><p>So, the technology industry is doing what it has always done for three thousand years: it is changing the medium and building a new room.</p><p>The silicon photonic computing systems being deployed today are massive. They replace copper wires with lasers, directing data at the speed of light. Because these light beams require perfectly unblemished trajectories, they require heavy, industrial external lasers to generate the frequencies, ultra-precise fiber-optic routing systems that cannot tolerate the slightest ambient floor vibration, and intensive liquid-cooling loops to maintain a state of total structural equilibrium. They are heavy, hyper-sensitive, and securely locked inside warehouse-sized data centers.</p><h3>The Future in Your Pocket</h3><p>If history is any indication, you should not look at these massive, modern photonic server racks and think they will remain corporate or state monopolies forever.</p><p>Right now, we are living in the modern equivalent of the era of the Salamis Tablet. We are standing over a five-foot marble slab of light-based computing, marveling that it works at all. We are looking at the ENIAC of the photonics age.</p><p>The structural thesis of mathematics and human engineering proves that the room is never the final destination. The room is simply the first draft.</p><p>Fully “all-optical” computers that replace every electronic part are still mostly in the experimental or research phase. That being said, modern tech giants (like those powering today’s artificial intelligence hardware) are building <strong>optoelectronic hybrids</strong>. Electronic chips (like CPUs/GPUs) handle basic logic and memory, while <strong>silicon photonics</strong> use tiny on-chip lasers and optical waveguides to transfer massive amounts of data instantly, massively speeding up the machine and keeping it cool.</p><p>Over the next few decades, the high-powered external lasers will shrink into microscopic, light-emitting components printed directly onto the architecture of the silicon itself. The power requirements will plummet as the engineering matures. And maybe, just maybe, fully “all optical” computers will become reality.</p><p>Forty or fifty years from now, someone may be sitting on a high-speed train, casually holding a device that runs an onboard, light-speed photonic processor cooler than the palm of their hand. They will read about the massive, power-hungry, liquid-cooled data centers of 2026 and find it wild that we ever thought an entire warehouse of server racks was the absolute pinnacle of human intelligence.</p><p>They will smile at our giant stone tables and roaring mechanical drums, turn off their screen, and slide the speed of light right back into their pocket.</p><h3>Bibliography</h3><p>· <strong>Copeland, B. J.</strong> (2006). <em>Colossus: The Secrets of Bletchley Park’s Codebreaking Computers</em>. Oxford University Press.</p><p>· <strong>Herzstark, C.</strong> (2005). <em>Kein Geschenk für den Führer: Schicksal eines Erfinders</em> [No Gift for the Führer: Destiny of an Inventor]. Erhard Löcker Verlag.</p><p>· <strong>Ifrah, G.</strong> (2001). <em>The Universal History of Computing: From the Abacus to the Quantum Computer</em>. John Wiley &amp; Sons.</p><p>· <strong>Landes, D. S.</strong> (2000). <em>Revolution in Time: Clocks and the Making of the Modern World</em>. Harvard University Press.</p><p>· <strong>Shurkin, J.</strong> (1996). <em>Engines of the Mind: The Evolution of the Computer from Mainframes to Microprocessors</em>. W. W. Norton &amp; Company.</p><p>· <strong>Struthers, R.</strong> (2024). <em>Hands of Time: A Watchmaker’s History of Time</em>. HarperCollins.</p><p>· <strong>Swade, D. D.</strong> (2005). The Construction of Charles Babbage’s Difference Engine №2. <em>IEEE Annals of the History of Computing</em>, 27(3), 70–78. <a href="https://doi.org/10.1109/mahc.2005.45">https://doi.org/10.1109/mahc.2005.45</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c581f4818ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is the Key to AMD… AMD?]]></title>
            <link>https://ianevans256.medium.com/is-the-key-to-amd-amd-ef095ce283b7?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef095ce283b7</guid>
            <category><![CDATA[mining]]></category>
            <category><![CDATA[rare-earth-metals]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Wed, 13 May 2026 02:04:36 GMT</pubDate>
            <atom:updated>2026-11-22T00:46:01.604Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y4H_Ud1rc57JYKASDPnuGQ.png" /><figcaption>ai generated photo of an AMD chip that accidently got some sludge from an acid mine drainage on it</figcaption></figure><h3>Advanced Micro Devices vs. Acid Mine Drainage</h3><p><em>Disclaimer: This is not financial advice, and I am not a financial advisor. This is for educational purposes only and all readers should do their own research and due diligence. I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.</em></p><p>Normally, acronym overlap creates confusion.<br> In this case, it may reveal the future of an industry.</p><p>Somewhere, a semiconductor analyst accidentally searching “AMD extraction technologies” is probably already very confused.</p><p>As artificial intelligence infrastructure expands at an almost unimaginable pace, companies like Advanced Micro Devices, NVIDIA, and hyperscale datacenter operators are driving an explosion in demand for advanced semiconductors, electrification systems, permanent magnets, and critical minerals. Every AI server rack, GPU cluster, electric motor, missile guidance system, and advanced cooling system ultimately depends on a complex supply chain of rare earth elements and strategic materials.</p><p>At the same time, the United States is scrambling to rebuild domestic critical mineral capabilities after decades of dependence on foreign-controlled mining and refining infrastructure, particularly from China.</p><p>That has placed companies like MP Materials and NioCorp Developments directly into the geopolitical spotlight. Projects such as the Mountain Pass mine in California and the Elk Creek critical minerals project in Nebraska are no longer just mining operations. They are increasingly viewed as strategic national infrastructure tied directly to AI, defense, aerospace, and energy independence.</p><p>But traditional mining may not be the only answer.</p><p>Ironically, one of America’s largest untapped rare earth resources may already exist inside one of its biggest environmental disasters:</p><p>acid mine drainage.</p><p>Or AMD.</p><h3>The Toxic Runoff That Could Become a Strategic Asset</h3><p>Acid mine drainage forms when sulfide minerals inside abandoned or active mines are exposed to oxygen and water. The resulting chemical reaction produces sulfuric acid, which dissolves metals from surrounding rock and carries them into nearby waterways.</p><p>For decades, AMD has been viewed almost exclusively as an environmental catastrophe, which admittedly makes it one of the least marketable investment themes imaginable.</p><p>“Toxic mining runoff” has historically struggled as a branding exercise.</p><p>Across Appalachia and former mining regions throughout the United States, acidic runoff has polluted rivers, damaged ecosystems, and created billions in long-term cleanup liabilities.</p><p>But researchers have discovered something extraordinary hiding inside these waste streams.</p><p>Many AMD systems contain dissolved concentrations of valuable rare earth elements and critical minerals including:</p><ul><li>neodymium,</li><li>dysprosium,</li><li>scandium,</li><li>yttrium,</li><li>cobalt,</li><li>manganese,</li><li>nickel,</li><li>and lithium.</li></ul><p>In other words:</p><p>the pollution itself may contain economically recoverable strategic materials.</p><p>This changes the equation entirely.</p><p>Instead of viewing acid mine drainage purely as an environmental liability, researchers and engineers are increasingly beginning to view it as a secondary domestic mineral reserve.</p><p>And unlike traditional mining, nature has already done part of the processing work.</p><h3>Nature Already Started the Extraction Process</h3><p>Conventional mining is brutally inefficient.</p><p>Mining companies often need to excavate and crush enormous quantities of rock to isolate tiny concentrations of valuable minerals. Massive energy, water, and chemical inputs are required before refining even begins.</p><p>AMD systems are different.</p><p>Over decades, acidic water naturally dissolves metals from surrounding rock formations. Some researchers have described this process as a form of “natural heap leaching,” effectively a pre-concentrated feedstock created by chemistry already occurring underground.</p><p>That distinction matters enormously.</p><p>Rather than blasting new mines into existence, future extraction systems may increasingly focus on:</p><ul><li>wastewater recovery,</li><li>tailings reprocessing,</li><li>coal ash refinement,</li><li>industrial waste streams,</li><li>and environmental remediation infrastructure.</li></ul><p>The future of mining may not only involve digging deeper into the earth.</p><p>It may involve harvesting the industrial remnants of the last century.</p><h3>This Is No Longer Just a Theory</h3><p>The idea of extracting rare earth elements from acid mine drainage is no longer confined to laboratory speculation.</p><p>Organizations including the U.S. Department of Energy and the National Energy Technology Laboratory have spent years funding pilot projects focused specifically on recovering rare earth elements from AMD, coal waste, and industrial byproducts.</p><p>At the center of much of this work is West Virginia University, whose Rare Earth Extraction Facility has demonstrated the ability to recover high-purity mixed rare earth oxides from acid mine drainage sludge.</p><p>Researchers estimated that AMD systems across Appalachia could potentially yield thousands of tons of recoverable rare earth elements annually under optimized conditions.</p><p>That may sound small compared to massive global mining operations.</p><p>But strategically, it is incredibly meaningful.</p><p>Rare earth markets are surprisingly tiny relative to oil, copper, or iron ore. Even modest domestic production volumes can materially impact defense supply chains and reduce import dependence.</p><p>And the technology continues improving.</p><p>Pilot systems have already demonstrated:</p><ul><li>roughly 98% purity mixed rare earth oxide concentrates,</li><li>continuous recovery operations,</li><li>and simultaneous extraction of cobalt, nickel, manganese, and zinc alongside rare earth materials.</li></ul><p>Companies and research groups including:</p><ul><li>RTI International,</li><li>Battelle Memorial Institute,</li><li>Rare Earth Salts,</li><li>and American Resources Corporation through its subsidiary ReElement Technologies,</li></ul><p>are all exploring advanced separation and purification systems designed to commercialize unconventional rare earth feedstocks.</p><p>The chemistry works.</p><p>The question now is scale.</p><h3>Why This Looks Strangely Similar to the Early Shale Revolution</h3><p>The comparison to the shale industry is difficult to ignore.</p><p>In the 1990s, shale oil and shale gas resources were widely known to exist. The problem was not discovery, it was economics. Recovery costs were too high, production rates were too low, and most experts believed unconventional drilling would never compete with conventional oil production.</p><p>Then technology changed everything.</p><p>Horizontal drilling and hydraulic fracturing dramatically improved extraction efficiency. Once throughput increased and costs fell, the economics flipped. The United States went from fears of energy scarcity to becoming the world’s largest oil producer within little more than a decade.</p><p>The numbers were historic:</p><ul><li>U.S. crude oil production rose from roughly 5 million barrels/day in 2008 to nearly 13 million barrels/day by 2019,</li><li>while shale gas production exploded from niche output into the backbone of American energy infrastructure.</li></ul><p>The resources themselves did not suddenly appear.</p><p>The rocks were sitting there the entire time. Humans simply got better at extracting value from them, a surprisingly common theme in capitalism.</p><p>AMD rare earth recovery appears to be following the same developmental pattern.</p><p>Today, critics argue:</p><ul><li>concentrations are too low,</li><li>purification is too expensive,</li><li>and pilot systems are too small to matter economically.</li></ul><p>But these criticisms sound remarkably similar to the skepticism surrounding early shale extraction.</p><p>The underlying resource already exists.</p><p>The real question is whether process optimization can unlock it at scale.</p><p>And modern AI systems may ironically accelerate that process.</p><p>Machine learning optimization, automated chemistry systems, advanced membrane filtration, chromatography, and real-time sensor networks could dramatically improve extraction efficiency over the next decade.</p><p>Which creates one of the strangest industrial feedback loops imaginable:</p><p>AI infrastructure increases demand for critical minerals.</p><p>Critical mineral scarcity accelerates alternative extraction research.</p><p>Advanced computing systems optimize extraction technologies.</p><p>And those improved extraction systems help supply the material backbone supporting future AI infrastructure.</p><p>AMD may partially depend on AMD, a sentence that would likely give both mining engineers and semiconductor analysts a headache.</p><h3>The Future of Mining May Look Completely Different</h3><p>Acid mine drainage recovery will probably not replace conventional mining.</p><p>Gigantic mines like Mountain Pass will still matter enormously. The world’s demand for rare earths, battery metals, and advanced materials is simply too large.</p><p>But AMD recovery may evolve into something equally important:</p><p>a distributed secondary mining network.</p><p>Instead of relying entirely on massive new excavation projects, future resource systems may combine:</p><ul><li>traditional mining,</li><li>recycling,</li><li>urban mining,</li><li>tailings reprocessing,</li><li>industrial waste recovery,</li><li>and wastewater mineral extraction.</li></ul><p>And unlike traditional mining, AMD recovery carries a unique advantage:</p><p>it monetizes environmental cleanup itself.</p><p>That fundamentally changes the public perception and political economics of extraction industries.</p><p>For over a century, abandoned mines represented stranded environmental liabilities.</p><p>The idea that they may eventually become strategic national assets would have sounded absurd twenty years ago, right up until the world decided GPUs were more valuable than oil rigs.</p><p>In the decades ahead, the most valuable mines in America may not be entirely new mines at all.</p><p>They may be the toxic remnants of old ones.</p><p>And in one of the strangest industrial coincidences of the AI era, the rise of Advanced Micro Devices and artificial intelligence infrastructure may indirectly accelerate the rise of AMD:</p><p>acid mine drainage recovery systems capable of reshaping how America secures its critical mineral future.</p><h3>Bibliography</h3><h3>Government &amp; Research Institutions</h3><ul><li>U.S. Department of Energy Critical Minerals Initiative</li><li><a href="https://www.netl.doe.gov/resource-sustainability/critical-minerals-and-materials?utm_source=chatgpt.com">National Energy Technology Laboratory (NETL) Critical Minerals Program</a></li><li><a href="https://www.netl.doe.gov/resource-sustainability/critical-minerals-and-materials/overview-accomplishments?utm_source=chatgpt.com">NETL Overview of Rare Earth Element Recovery from Acid Mine Drainage</a></li><li>West Virginia University Rare Earth Extraction Facility Research</li><li><a href="https://wvutoday.wvu.edu/stories/2018/07/18/wvu-opens-new-research-facility-to-extract-valuable-rare-earths-from-acid-mine-drainage?utm_source=chatgpt.com">WVU Rare Earth Extraction Facility Announcement</a></li><li><a href="https://wvwri.wvu.edu/news/2019/10/02/wvu-awarded-5-million-to-continue-rare-earth-project-build-acid-mine-drainage-treatment-facility?utm_source=chatgpt.com">WVU NETL Grant for Acid Mine Drainage Rare Earth Recovery</a></li><li><a href="https://www.psu.edu/news/earth-and-mineral-sciences/story/new-process-developed-extract-high-purity-rare-earth-element?utm_source=chatgpt.com">Penn State Critical Minerals Research</a></li></ul><h3>Academic Papers &amp; Technical Research</h3><ul><li><a href="https://www.mdpi.com/2075-163X/11/11/1298?utm_source=chatgpt.com">A Fundamental Economic Assessment of Recovering Rare Earth Elements and Critical Minerals from Acid Mine Drainage Using a Network Sourcing Strategy (MDPI Minerals, 2021)</a></li><li><a href="https://www.sciencedirect.com/science/article/pii/S2213343722005772?utm_source=chatgpt.com">Recovery of Rare Earth Elements from Acid Mine Drainage (ScienceDirect Review Paper)</a></li><li><a href="https://netl.doe.gov/node/11886?utm_source=chatgpt.com">DOE/NETL Rare Earth Element Extraction Technology Research Portfolio</a></li></ul><h3>Companies &amp; Commercial Development</h3><ul><li><a href="https://mpmaterials.com/?utm_source=chatgpt.com">MP Materials Official Website</a></li><li><a href="https://www.niocorp.com/?utm_source=chatgpt.com">NioCorp Developments Official Website</a></li><li>Rare Earth Salts Official Website</li><li>American Resources Corporation Official Website</li><li><a href="https://reelementtech.com/?utm_source=chatgpt.com">ReElement Technologies Official Website</a></li><li><a href="https://www.rti.org/?utm_source=chatgpt.com">RTI International Official Website</a></li><li><a href="https://www.battelle.org/?utm_source=chatgpt.com">Battelle Memorial Institute Official Website</a></li></ul><h3>Industry &amp; Energy Production Data</h3><ul><li>U.S. Energy Information Administration (EIA) Crude Oil Production Data</li><li>EIA U.S. Shale Gas Production Statistics</li><li><a href="https://www.iea.org/reports/the-role-of-critical-minerals-in-clean-energy-transitions?utm_source=chatgpt.com">International Energy Agency Critical Minerals Report</a></li><li><a href="https://www.usgs.gov/centers/national-minerals-information-center/rare-earths-statistics-and-information?utm_source=chatgpt.com">U.S. Geological Survey Rare Earth Elements Statistics and Information</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef095ce283b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aspen Aerogels]]></title>
            <link>https://ianevans256.medium.com/aspen-aerogels-e0ba9cd60d3f?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e0ba9cd60d3f</guid>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[due-diligence]]></category>
            <category><![CDATA[aerogel]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[trading]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Mon, 11 May 2026 14:27:16 GMT</pubDate>
            <atom:updated>2026-11-22T00:50:56.579Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-CRRhyco8KtCbXyo67RmwQ.png" /><figcaption>ai generated photo of an aspen wrapped in aerogel to keep it toasty warm in the winter</figcaption></figure><p><em>Disclaimer: This is not financial advice, and I am not a financial advisor. This is for educational purposes only and all readers should do their own research and due diligence. I used AI as an editorial sounding board and drafting assistant during the research and revision of this piece.</em></p><h3>Recommendation: BUY</h3><ul><li><strong>Current Thesis:</strong> Deep value + asymmetric upside opportunity</li><li><strong>Target Price:</strong> $15</li><li><strong>Upside Potential:</strong> ~200%+</li><li><strong>Time Horizon:</strong> 8–12 Months</li><li><strong>Risk Level:</strong> High Risk / High Reward</li></ul><h3>1. Investment Thesis</h3><h3>Core Thesis</h3><p>Aspen Aerogels is a speculative turnaround and infrastructure-adjacent growth story trading at distressed valuation levels after a major operational setback and broader EV market weakness. The market appears to be pricing in near-worst-case scenarios despite the company maintaining valuable IP, strategic partnerships, and exposure to several long-term secular trends including EV battery safety, grid infrastructure, industrial insulation, and AI datacenter thermal management.</p><p>The company’s valuation collapse may present a rare asymmetrical setup where downside appears increasingly understood while upside catalysts remain underappreciated.</p><h3>Key Bullish Drivers</h3><h3>• AI Datacenter Thermal Management Opportunity</h3><p>Aerogels possess extremely strong thermal insulation properties while remaining lightweight and space-efficient. As hyperscale AI datacenters face escalating thermal density challenges from next-generation GPUs and accelerated computing infrastructure, advanced thermal materials could become increasingly important.</p><p>While Aspen’s current primary commercial focus remains EV battery fire protection and energy infrastructure, the broader aerogel industry is beginning to gain traction in datacenter cooling and thermal regulation applications. Private firms such as ThermaGel Innovations have already explored these use cases commercially.</p><p>This creates a potential “optionality” component to the thesis:</p><ul><li>EV thermal barriers today</li><li>Datacenter thermal management tomorrow</li><li>Energy infrastructure insulation long term</li></ul><p>If AI infrastructure spending continues accelerating, Aspen could potentially participate in adjacent thermal management markets beyond automotive applications.</p><h3>• Distressed Valuation / Oversold Conditions</h3><p>Shares currently trade dramatically below the prior all-time high near ~$63 reached during the 2021 speculative growth and short squeeze cycle.</p><p>While that valuation was likely irrationally exuberant, the current valuation may now reflect excessive pessimism:</p><ul><li>Negative sentiment around EV demand slowdown</li><li>Manufacturing disruptions</li><li>Profitability concerns</li><li>Capital expenditure fears</li><li>Small-cap growth liquidation</li></ul><p>The stock has effectively round-tripped from euphoric pricing to distressed pricing.</p><p>This setup creates the possibility for:</p><ul><li>Multiple expansion</li><li>Short covering</li><li>Turnaround repricing</li><li>Strategic acquisition interest</li><li>Operational recovery rerating</li></ul><h3>• Oven #5 Explosion</h3><p>The manufacturing incident involving Oven #5 at the Rhode Island facility was clearly a serious operational setback. Most importantly, no employees were seriously injured.</p><p>However, the market reaction may have excessively discounted the long-term survivability of the business:</p><ul><li>Multiple ovens remain operational</li><li>Production recovery efforts are ongoing</li><li>Manufacturing continuity appears possible</li><li>Insurance recoveries may partially offset damages</li><li>Customers still require thermal barrier materials</li></ul><p>The stock reached near all-time lows after the incident, suggesting the market may have temporarily priced in existential risk. (feels like an episode out of Billions)</p><p>If production normalization progresses through 2026, investor sentiment could recover materially.</p><h3>2. Business Overview</h3><h3>What Aspen Aerogels Does</h3><p>Aspen Aerogels develops and manufactures aerogel-based thermal barrier materials.</p><p>Primary applications include:</p><ul><li>EV battery thermal runaway protection</li><li>Industrial insulation</li><li>Energy infrastructure</li><li>LNG and petrochemical insulation</li><li>Sustainable building materials</li><li>Advanced thermal management systems</li></ul><p>Their flagship EV platform, PyroThin®, is designed to reduce thermal propagation risk in lithium-ion battery packs.</p><h3>3. Industry &amp; Competitive Positioning</h3><h3>Competitive Advantage</h3><p>Aspen’s competitive moat largely comes from:</p><ul><li>Proprietary aerogel manufacturing expertise</li><li>Existing OEM relationships</li><li>High-performance thermal barrier IP</li><li>Difficult manufacturing process</li><li>Qualification barriers within automotive supply chains</li></ul><p>Battery safety regulations continue tightening globally, particularly after several high-profile EV battery fire incidents. This may increase long-term demand for advanced thermal propagation materials.</p><h3>Relative Valuation Opportunity</h3><p>Despite operational issues, Aspen appears inexpensive relative to peers on a price-to-sales basis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/483/1*TT6XG5KWR-H2zB35lfDQsA.png" /></figure><p>If Aspen were eventually rerated closer to thermal management peers, substantial upside could exist even without aggressive earnings assumptions.</p><p>This does not imply parity should exist immediately, but highlights how compressed Aspen’s valuation has become.</p><h3>4. Potential Catalysts</h3><h3>Near-Term Catalysts (0–12 Months)</h3><ul><li>Manufacturing recovery updates</li><li>Insurance settlement clarity</li><li>Improved production capacity utilization</li><li>EV contract wins</li><li>AI/datacenter thermal management partnerships</li><li>Reduced recession fears</li><li>Federal infrastructure and energy spending</li><li>Short squeeze dynamics from depressed float sentiment</li></ul><h3>Long-Term Catalysts</h3><ul><li>EV adoption reacceleration</li><li>Expansion into datacenter cooling materials</li><li>Energy infrastructure modernization</li><li>Strategic acquisition by industrial/materials company</li><li>Margin improvement from scale efficiencies</li><li>Licensing/IP monetization opportunities</li></ul><h3>5. Risks</h3><h3>Key Risks to Thesis</h3><h3>• Operational Execution Risk</h3><p>The company must successfully restore and maintain production capacity following the Oven #5 incident.</p><h3>• Cash Burn / Dilution Risk</h3><p>Aspen remains financially stressed relative to larger industrial peers. Additional equity raises or debt financing could dilute shareholders.</p><h3>• EV Market Weakness</h3><p>A prolonged EV slowdown would pressure near-term revenue growth.</p><h3>• Customer Concentration</h3><p>Large OEM dependency creates contract and pricing risk.</p><h3>• Commercialization Risk</h3><p>AI datacenter applications remain speculative and are not currently a major revenue driver.</p><h3>• Competition</h3><p>Larger industrial insulation and thermal management firms may enter the space aggressively if aerogel demand scales materially.</p><h3>6. Why the Market May Be Wrong</h3><p>The market currently appears focused almost entirely on:</p><ul><li>Operational setbacks</li><li>Weak profitability</li><li>Manufacturing disruption</li><li>EV macro fears</li></ul><p>However, markets often misprice companies during peak uncertainty.</p><p>What may be underappreciated:</p><ul><li>Strategic value of thermal barrier IP</li><li>Long-duration infrastructure demand</li><li>AI thermal management optionality</li><li>Manufacturing replacement value</li><li>Potential recovery trajectory</li></ul><p>The stock may not need perfect execution to re-rate significantly from current levels.</p><h3>7. Conclusion</h3><p>Aspen Aerogels represents a speculative but potentially high-upside turnaround opportunity trading at distressed levels after a severe operational event and broader EV market pessimism.</p><p>The company sits at the intersection of several important macro trends:</p><ul><li>Electrification</li><li>Battery safety</li><li>Energy infrastructure modernization</li><li>Thermal efficiency</li><li>AI compute expansion</li></ul><p>While risks remain substantial, the current valuation may already reflect extreme downside assumptions.</p><p>If management successfully restores production and demonstrates stabilization, the stock could experience meaningful multiple expansion over the next 8–12 months.</p><p>For speculative investors comfortable with volatility and execution risk, Aspen Aerogels may represent an attractive asymmetric opportunity</p><h3>Additional Due Diligence Additions for Aspen Aerogels</h3><h3>8. Current Financial Snapshot</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/462/1*WJp_EFbbJe_gzu2XLS0zOg.png" /></figure><h3>9. Latest Quarterly Earnings (Q1 2026)</h3><h3>Q1 2026 Highlights</h3><ul><li>Revenue: <strong>$37.9M</strong></li><li>Net Loss: <strong>($23.7M)</strong></li><li>Adjusted EBITDA: <strong>($12.7M)</strong></li><li>Gross Margin: <strong>11%</strong></li><li>Cash Position Increased to <strong>$175.6M</strong></li><li>GM settlement proceeds: <strong>$37.6M</strong></li><li>Q2 Revenue Guidance: <strong>$40M–$48M</strong></li></ul><p>Breakdown:</p><ul><li>Energy Industrial Revenue: <strong>$21.6M</strong></li><li>Thermal Barrier Revenue: <strong>$16.3M</strong></li></ul><p>Management stated production recovery efforts are underway following the East Providence incident, with staged manufacturing restart expected beginning in May 2026.</p><h3>10. Technical Analysis &amp; Key Chart Levels</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/428/1*mDfJa99z9Ib5VXwAiAqYHw.png" /></figure><h3>11. Insider Ownership &amp; Short Interest</h3><h3>Short Interest</h3><ul><li>Short interest currently sits around <strong>7.0M shares</strong></li><li>Roughly <strong>8.7% of public float</strong> is sold short</li></ul><p>This is important because:</p><ul><li>ASPN has relatively low liquidity</li><li>Small-cap industrial names can squeeze violently</li><li>Positive operational updates could force rapid covering</li></ul><p>Not a pure “meme squeeze” setup, but definitely elevated enough to amplify upside volatility.</p><h3>Insider / Institutional Dynamics</h3><p>Key institutional and insider factors worth noting:</p><p>Institutional ownership remains extremely high at roughly 97%, with major holders including BlackRock, Vanguard, Invesco, and Spring Creek Capital. Insider ownership is comparatively modest at approximately 3%, though CEO Donald Young previously purchased shares near distressed valuation levels in 2025. While insider selling has outweighed buying over the past two years, much of the larger liquidation activity occurred during significantly higher stock price periods. Institutional support remaining elevated despite operational setbacks may indicate continued belief in Aspen’s long-term strategic value proposition</p><h3>12. Peer Comparison</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/474/1*PItywfKVXGSB5YeAHJk52w.png" /></figure><p>Similar speculative infrastructure theme</p><h3>Important Observation</h3><p>ASPN trades at a major discount despite operating in highly strategic thermal management and battery safety markets.</p><p>If the company stabilizes operations and returns to growth, multiple expansion alone could drive substantial upside without requiring profitability perfection.</p><h3>13. Scenario Analysis</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*0-tznLNVMtvuBYxeDatIUA.png" /></figure><h3>14. Closing Thoughts and My Position</h3><p>The market currently appears to value Aspen Aerogels as a structurally broken EV supplier rather than a strategic thermal materials company with exposure to electrification, industrial infrastructure, and potentially AI compute cooling markets.</p><p>At current valuation levels, investors are paying roughly 1x sales for a company possessing differentiated IP, significant OEM relationships, and nearly $176M in liquidity. While operational and execution risks remain substantial, the risk/reward profile appears increasingly asymmetric if management successfully executes its recovery plan.</p><p>In a market searching aggressively for AI infrastructure and energy-adjacent plays, Aspen may represent one of the more overlooked speculative turnaround opportunities in the small-cap industrial space.</p><p><strong>This is a high risk, high reward trade for myself, I entered $10 strike calls for both January 2027 and January 2028 when price was around 3.45 last week. Might buy some spot once technicals stabilize in price action to consider as long term investment, want to see heavy floor around the $5–7 price to stack spot and wouldn’t hurt to see mainstream media narrative support for buiness shift to datacenters and energy infra.</strong></p><p><strong>11–21–2026 update: </strong>AI bubble popped slightly with nbidia leading the sector downwards. Aspen holding at 4.50 despite this broader downward movement shows resilience and accumulation rather than futrther capitulation. 2028 Leap Calls I have currently bassically worthless atm but not closing out for own personal sector sentiment gauge since size so small. Not adding, not clsoing, using this to observe how this subsector reacts to the broader sector</p><p>Bibliography/Citations:</p><h3>Company Filings &amp; Investor Relations</h3><ol><li>Aspen Aerogels Investor Relations<br> Accessed May 2026.</li><li>Aspen Aerogels SEC Filings Database<br> Accessed May 2026.</li><li>Aspen Aerogels Form 10-K Annual Reports<br> Accessed May 2026.</li><li>Aspen Aerogels Q1 2026 Earnings Materials<br> Accessed May 2026.</li></ol><h3>Earnings &amp; Financial Data</h3><ol><li><a href="https://www.fool.com/earnings/call-transcripts/2026/05/07/aspen-aerogels-aspn-q1-2026-earnings-transcript/?utm_source=chatgpt.com">The Motley Fool Earnings Transcript — Aspen Aerogels Q1 2026</a><br> Accessed May 2026.</li><li><a href="https://www.macrotrends.net/stocks/charts/ASPN/aspen-aerogels/financial-statements?utm_source=chatgpt.com">Macrotrends Aspen Aerogels Financial Data</a><br> Accessed May 2026.</li><li><a href="https://companiesmarketcap.com/aspen-aerogels/marketcap/?utm_source=chatgpt.com">CompaniesMarketCap Aspen Aerogels Market Capitalization</a><br> Accessed May 2026.</li></ol><h3>Market Data &amp; Short Interest</h3><ol><li><a href="https://www.marketbeat.com/stocks/NYSE/ASPN/short-interest/?utm_source=chatgpt.com">MarketBeat Aspen Aerogels Short Interest Data</a><br> Accessed May 2026.</li><li><a href="https://www.marketbeat.com/stocks/NYSE/ASPN/sec-filings/?utm_source=chatgpt.com">MarketBeat Aspen Aerogels SEC Filings &amp; Insider Activity</a><br> Accessed May 2026.</li><li><a href="https://www.tradingview.com/symbols/NYSE-ASPN/?utm_source=chatgpt.com">TradingView ASPN Chart Data</a><br> Accessed May 2026.</li><li>Yahoo Finance Aspen Aerogels Financial Overview<br> Accessed May 2026.</li></ol><h3>Industry &amp; Competitive Research</h3><ol><li><a href="https://investors.modine.com/?utm_source=chatgpt.com">Modine Manufacturing Company Investor Relations</a><br> Accessed May 2026.</li><li>Cabot Corporation Investor Relations<br> Accessed May 2026.</li><li><a href="https://investors.dupont.com/?utm_source=chatgpt.com">DuPont Investor Relations</a><br> Accessed May 2026.</li><li>ThermaGel Innovations Official Website<br> Accessed May 2026.</li></ol><h3>Supplemental Industry Topics</h3><ol><li>U.S. Department of Energy — Battery Thermal Management Research<br> Accessed May 2026.</li><li>McKinsey &amp; Company — AI Datacenter Energy Demand Research<br> Accessed May 2026.</li><li><a href="https://www.iea.org/reports/global-ev-outlook-2025?utm_source=chatgpt.com">International Energy Agency EV Outlook Reports</a><br> Accessed May 2026.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e0ba9cd60d3f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Arweave vs IPFS]]></title>
            <link>https://ianevans256.medium.com/arweave-vs-ipfs-ef6601c0bff0?source=rss-2b8f08d6771b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef6601c0bff0</guid>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[nft]]></category>
            <category><![CDATA[storage-solutions]]></category>
            <dc:creator><![CDATA[Ian Evans]]></dc:creator>
            <pubDate>Mon, 01 Nov 2021 02:02:43 GMT</pubDate>
            <atom:updated>2026-05-22T12:34:54.566Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0eP1Cf5vNHa8ITfX" /><figcaption>Photo by <a href="https://unsplash.com/@etaplus?utm_source=medium&amp;utm_medium=referral">ETA+</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>Not sure which to use? me neither.</h4><p><em>Disclaimer: This is not financial advice, and I am not a financial advisor. This is for educational purposes only and all readers should do their own research and due diligence. I am not a cat.</em></p><p>If you are an NFT creator, the decision to use Arweave’s data storage system, or use IPFS data storage system and a platform like Pinata to curate the development experience, is difficult depending on your project’s use cases and end goals. Hopefully this article can help you make better sense of the differences between the two and which to go for. Before we go further let’s do a quick overview of both Arweave and IPFS. If you’re familiar with the protocol’s summaries, please skip the next 2 sections.</p><p><strong>Arweave summary:</strong></p><p>Creators- Sam Williams and William Jones</p><p>Arweave, is the base protocol that allows a developer to store data, in our case image data for NFTs, permanently with a single upfront cost. At the core of the Arweave protocol is the Permaweb, a decentralized web built on top of Arweave. The PermaWeb has a full-stack that allows developers to build applications with that stack including, Arweave, Gateways, GraphQL, and SmartWeave. Profit-Sharing Functions are built into the PermaWeb so builders/devs can rely on these functions and profit share with other developers and builders who are creating Dapps on the PermaWeb. By having no single point of failure Arweave aims to provide a decentralized storage service that is economically sustainable.</p><p><strong>IPFS summary:</strong></p><p>Creators- Juan Benet and Protocol Labs</p><p>Practically speaking, IPFS allows users to access files (NFT image files) from a shared database from different locations. IPFS achieves “permanence” or “persistence” by pinning the data to a node that is run by the pinning services themselves. If someone is running the node then the data will exist in the network and can be retrieved by anyone.</p><p><strong>NFTs and hosting data:</strong></p><p>We can see from one of the earlier NFTs, CryptoPunks, that storing image data on the blockchain was not practical, so the NFT is just a hash of the image file that is stored on the blockchain, not the image itself. LarvaLabs, the creator of CryptoPunks, state on their site:</p><p>“When originally released, the actual images of the punks were too large to store on the blockchain, so we took a hash of the <a href="https://www.larvalabs.com/public/images/cryptopunks/punks.png"><strong>composite image</strong></a> of all the punks and <a href="https://github.com/larvalabs/cryptopunks/blob/master/contracts/CryptoPunksMarket.sol#L5"><strong>embeded it into the contract</strong></a>. You can verify that the punks being managed by the Ethereum contract are the True Official Genuine CryptoPunks™ by calculating an SHA256 hash on the cryptopunks image and comparing it to the hash stored in the contract.</p><p>Since then, due to some clever compression ideas and some help from friends of the punks we have managed to put the entire image and attribute data fully on-chain. You can read more about the mechanisms and details of the process in the <a href="https://www.larvalabs.com/blog/2021-8-18-18-0/on-chain-cryptopunks"><strong>announcement blog post</strong></a>.”</p><p>Buyers may not have previously known about where the image data was stored but CryptoPunks were and are still valued at their price point because the ownership has been legitimatized by the community. Owning that hash allows the main developers of an NFT to create further experiences that will only apply to the holder of the specific hash. Here we see a main use case of NFTs and that is as a <strong>digital signature of authentication</strong>. <strong>That signature can then be used as a ticket for entry to a network or group of people with a common interest. The other members of the community are what legitimize the NFT’s hash and ownership, and this is interesting because NFTs rely on people as a security/ consensus feature for proving legitimacy.</strong></p><p><strong>What is the difference between the two? Cost differences? Technological difference?</strong></p><p>From a developer/NFT creator&#39;s point of view, one of the main differences that will affect the business decisions of those looking to use these protocols is the underlying cost of using the protocol. Arweave has a one-time cost compared to IPFS which has a recurring cost if you choose to upload a lot of data onto IPFS. This difference in variable vs fixed cost might sway some to choose one or the other simply as a cost savings decision however there may be technical restraints. As noted earlier, Arweave is truly permanent and the image cannot be taken off as they can be in IPFS, so in this case, if there is a mistake on the initial release and the way the contract works immediately pins the image data to Arweave then this can cause some issues. It will be possible to reissue an updated edition on Arweave, but that file will now be a new file instead of simply replacing the existing file.</p><p>Furthermore, it is interesting to note that now marketing that you are using Arweave may cause a project to get more marketing/publicity because the idea of “permanence” has reached mainstream crypto Twitter. Also, keep in mind that you can use either or but the one that you use to start with will have an impact on your ability to make changes later down the road. Having fewer permanent options to start and eventually shifting to the permanent hosting options might be better depending on the cost.</p><p><strong>What do the top projects use?</strong></p><p>If we consider some of the top-selling NFT projects according to OpenSea’s NFT top volume rankings we can see that CryptoPunks, ArtBlocks, and BAYC are at the top. Looking into a statement from BAYC:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xdLxNGm3pbcXKWkSSLENIg.png" /></figure><p>BAYC uses IPFS and the screenshot above tells us that they have now added high-res images of the apes to IPFS.</p><p>As far as ArtBlocks go they generate the NFTs completely on-chain so that in the future anyone can view and not have to worry if ArtBlocks even exists anymore. When you look under the hood you realize and appreciate the effort of the artists to bring their works fully “on-chain”. Check out their page to learn more:</p><p><a href="https://www.artblocks.io/learn">https://www.artblocks.io/learn</a></p><p><strong>Final Thoughts:</strong></p><p>Data Storage is important to the crypto industry as NFTs continue to grow and gain adoption across the masses. NFTs are a good gateway/entry for newcomers into the crypto space and as NFTs continue to grow it will be interesting to observe the capacity of these decentralized storage networks. Touching back to the business perspective, it may be prudent to consider both for cost reasons but technologically speaking it would not be out of the question to assume that projects may one day switch from using one protocol to the other and there may just be a hash of that hash to represent to new “legitimate version” of the NFT. When thinking about new ERC standards coming out, I can only imagine that ERC721 projects may one-day issue/airdrop ERC 1155 equivalents to access the additional functionality that the different ERC1155 standard offers but doing so only to the original holders thereby preserving the legitimacy of their project. So much innovation and new things to come from the NFT space and I can’t wait to see what the rest of 2021 has in store.</p><p><strong>Resources:</strong></p><p><a href="https://arweave.org/">https://arweave.org</a></p><p><a href="https://docs.ipfs.io/">https://docs.ipfs.io</a></p><p><a href="https://www.artblocks.io/learn">https://www.artblocks.io/learn</a></p><p><a href="https://www.larvalabs.com/cryptopunks">https://www.larvalabs.com/cryptopunks</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef6601c0bff0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>