<?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 Prince Jain on Medium]]></title>
        <description><![CDATA[Stories by Prince Jain on Medium]]></description>
        <link>https://medium.com/@prince.sh?source=rss-212f4a7a0c5a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZTivrrRUWQLoOebkmCHsMA.jpeg</url>
            <title>Stories by Prince Jain on Medium</title>
            <link>https://medium.com/@prince.sh?source=rss-212f4a7a0c5a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 03:40:17 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@prince.sh/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[Washington approved 879,868 AI chips, then licensed the data center]]></title>
            <link>https://medium.com/@prince.sh/washington-approved-879-868-ai-chips-then-licensed-the-data-center-1e8638dcf8b0?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e8638dcf8b0</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[tsmc]]></category>
            <category><![CDATA[ai-chip]]></category>
            <category><![CDATA[bi]]></category>
            <category><![CDATA[export-control]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 19:42:30 GMT</pubDate>
            <atom:updated>2026-11-20T19:42:30.517Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*69bI3urA8eUtYybQL7Gg2g.png" /></figure><p>On November 19, the House Foreign Affairs Committee published a table the Bureau of Industry and Security had kept out of public view. It listed 21 approved license applications covering 879,868 Nvidia H200 and AMD MI325X chips destined for Chinese buyers, carved out of requests totaling 3,444,160 units in a BIS summary dated May 11, 2026. The committee released the material under 50 U.S.C. § 4820(h)(2)(B)(ii) after deciding that continued withholding was contrary to the national interest.</p><p>The same day, a BIS interim final rule titled Addressing Remote Access to Advanced Semiconductors took effect. It sits in the Federal Register at Volume 91, Number 222, beginning on page 74970, docket 261116–0005, with comments running to January 19, 2027. The rule reaches past the dock. It covers the data center that serves the chips, the five countries that route them, and the fabs and assembly houses that build them.</p><p>Reuters supplied the third piece on November 19. TSMC had suspended some previously permitted manufacturing orders from Chinese chip designers, telling affected customers on 16-nanometre and more advanced processes to obtain US export licenses first. Put the three items side by side and the boundary has slid from the loading dock into the license file.</p><p>The arithmetic matters because it prices the change. A regulator that approved 879,868 chips in writing, against 3,444,160 requested, has already shown where the friction lives. The November rule widens the set of transactions that need that paperwork. It doesn’t ban any of them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8MhyVh5v_B7nqbQWbWgAtg.png" /></figure><h3>What the 879,868 approvals actually cover</h3><p>The congressional table has more to say about gaps than totals. Alibaba and its subsidiaries asked for 948,950 chips and were approved for 75,000. Tencent and its subsidiaries asked for 150,000 and cleared 150,000, and a separate Nvidia application dated December 19, 2025 covering 445,000 H200 chips for Tencent wasn’t fully folded into the May summary, so the two datasets can’t be added together. Baidu drew 43,000 out of 43,000. BYD, an automaker, cleared 4,000 of 4,000.</p><p>Approval isn’t delivery. Reuters found no H200 deliveries at all when it checked in May 2026. In July testimony, Commerce official Jeffrey Kessler said shipments had begun and described them as very few. Reuters later reported several major Chinese internet companies had each received roughly 10,000 H200 chips. The 879,868 figure is therefore a ceiling on permission, and the tonnage that actually crossed the Pacific has never been published.</p><p>Every Nvidia and AMD application the committee reviewed rated the transfer’s diversion risk as low. The paperwork named possible data center installers including Lenovo and its subsidiaries, ZTE Kangxun Telecom, Beijing New Matrix, Maginfra, Aivres Systems, XFusion Technologies International, Cloudnine Information Technologies and Changkuai Computing, which trades as SmoothCompute. A June 2026 Pentagon list of alleged Chinese military companies included Alibaba, Tencent, Baidu and BYD. That’s a government classification, and it doesn’t establish military use of any authorized chip.</p><p>The November rule leaves a Chinese firm free to rent compute abroad. It converts the renting into a licensable act once a facility crosses a performance line, and it counts connected halls as one facility. BIS wrote the threshold in total processing performance, so a colocation operator now has to measure something it may never have tracked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xRPlLTQdGLdfrJ33rOlBGQ.png" /></figure><h3>Twelve million TPP is the new border</h3><p>Section 744.24(a)(1) requires a license for exports, reexports and in-country transfers when the exporter knows the items are headed to a data center anywhere on earth that provides or will provide more than 12,000,000 aggregate TPP of access to companies headquartered or ultimately parented in Macau or Country Group D:5. Digitally linked and physically linked facilities that share resources and workloads fall inside the text, so splitting capacity across sister sites doesn’t erase the obligation. Existing covered data centers have 90 days to seek licenses if they want to keep serving the same customers. Shipments already en route on November 19 can finish under prior eligibility if they clear by December 21, 2026.</p><p>There’s a narrow carve-out for commercial recommender systems used exclusively to suggest products, services or content, filed with the required information. Generative systems producing novel text, images, audio, video or software are excluded, so a shopping engine qualifies and a chatbot doesn’t. The review standard is case by case.</p><p>The rule stretches the map as well. Five jurisdictions now carry destination-based licensing requirements for specified advanced computing items, including ECCNs 3A090 and 4A090. Those countries sit on routes Chinese buyers use to reach compute without shipping anything into China.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LywbYbrCgLoUCLbhuMAnug.png" /></figure><h3>Five countries get their own license line</h3><p>Section 742.6(a)(6)(iii)(A)(1) adds Indonesia, Malaysia, Singapore, Thailand and Brazil. A US-headquartered ultimate consignee or end user, or one whose ultimate parent sits in the US, can skip that particular license as long as the § 743.10 reports get filed. Those reports carry the parties and their parents, the destination, the product classification, the SKU, the TPP, the quantity and the transaction date. Submissions are monthly, due by the first business day after the month closes. The exemption doesn’t override separate end-use or end-user controls, so a clean corporate parent won’t rescue a deal that fails another test.</p><p>The third checkpoint sits down at the fab. A new clause inside § 744.23(a)(3) requires a license for certain advanced-node integrated circuits under ECCN 3A090.a when the designer isn’t an approved or authorized IC designer and the work runs through a front-end fabricator or an outsourced assembly and test company.</p><h3>TSMC meets the fabrication clause</h3><p>Reuters gave that clause its first public test. TSMC told Chinese designers working on 16-nanometre and more advanced processes that previously permitted orders were on hold pending US licenses. The legal text and the commercial cutoff aren’t the same instrument. The Federal Register turns on chip classifications, covered designers, transaction circumstances and exceptions, while the 16-nanometre line came from TSMC’s instructions as Reuters described them. Anyone claiming the rule bans every 16nm wafer made anywhere has skipped the classification step.</p><p>The composite is a licensing regime that follows the workload rather than the box. A single GPU order now has to survive a destination license, a remote-access license, a designer check at the fab and monthly reporting afterwards. Alibaba’s 75,000 approved chips out of 948,950 requested show how much of that survival depends on an officer’s signature. The filings don’t prove diversion, and the rule doesn’t shut the door. They show that written, renewable permission has become a scarce input alongside wafer capacity.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e8638dcf8b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nvidia stopped shipping Hopper as Washington moved the control into the fab]]></title>
            <link>https://medium.com/@prince.sh/nvidia-stopped-shipping-hopper-as-washington-moved-the-control-into-the-fab-63dc640fc55e?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/63dc640fc55e</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[tsmc]]></category>
            <category><![CDATA[export-control]]></category>
            <category><![CDATA[advanced-packaging]]></category>
            <category><![CDATA[semiconductors]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 14:29:26 GMT</pubDate>
            <atom:updated>2026-11-19T14:29:26.748Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wZX-w0dHo9S1uJDtDFW7cg.png" /></figure><p>On November 18, Reuters reported that Nvidia had stopped shipping even the Hopper-generation data-center processors it was licensed to sell in China. The quarter ended October 25, 2026 had produced $2.4 billion in licensed Hopper sales there, a figure Reuters attributed to chief financial officer Colette Kress. Roughly 600,000 H200 processors were reportedly sitting in Hong Kong awaiting Chinese import clearance.</p><p>Read as a single event, the halt looks like one more turn of a controls regime that keeps tightening. Read beside two other developments from the same week, it locates a change in where the control gets applied. On November 17, Nvidia published supply commitments across six named partners. On November 19, a U.S. Bureau of Industry and Security interim final rule took effect that puts a classification burden on foundries and outsourced assembly and test companies. The point of control has moved off the shipping dock and into the factory.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ntk0PpsH6cGUqAS-jqQzpg.png" /></figure><h3>Nvidia’s China number and the halt that followed</h3><p>The $2.4 billion matters as a scale marker. That’s what one quarter of licensed Hopper sales to China was worth, and Nvidia stopped the shipments anyway. Reuters describes the halt and the inventory parked in Hong Kong; its reporting doesn’t supply a formal justification for the decision. What it documents is the result. Chinese buyers lost access to American accelerators the U.S. government had already cleared for sale to them, and the stockpile at the border shows the appetite hadn’t gone anywhere.</p><p>There’s a second detail in the number. Hopper sits a generation behind Nvidia’s current data-center parts, so the licensed sales were already a step off the frontier. Cutting them off removes the fallback that remained legal. A Chinese buyer who can’t get the previous generation through a license has one fewer option, and that matters because domestic substitutes at comparable performance are the harder thing to build.</p><p>The rule that landed two days later is where the mechanics live, and it’s worth reading closely because the change is procedural in a way that carries real weight. BIS published it as interim final rule 2026–23529 at 91 FR 74970, under docket 261116–0005 and RIN 0694-AK50. It was filed on November 16 at 4:15 p.m., took effect on November 19, and amends ten parts of the Export Administration Regulations, from 15 CFR Part 732 through Part 774. Comments close January 19, 2027.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VHVw3oSd85nVP9YtnWLPQA.png" /></figure><h3>The rule rewrites who has to prove what</h3><p>The operative text is Note 1 to Export Control Classification Number 3A090.a. For applicable advanced logic integrated circuits, a front-end fabricator or an outsourced semiconductor assembly and test company must now presume that a chip falls inside the controlled classification unless one of the specified rebuttal routes is available. The rule confines that presumption to those two categories of manufacturer, so it doesn’t reach every company that touches a chip.</p><p>The underlying thresholds in 3A090.a are a total processing performance of 4,800 or greater, or an alternative total processing performance of 1,600 or greater paired with performance density of 5.92 or greater. The advanced-node trigger for the presumption is 16/14nm or below, or a qualifying non-planar transistor architecture. These are defined regulatory metrics, not the floating-point figures printed on a product brief.</p><p>Three rebuttal routes exist, and each requires documentation of a specific kind. An approved or authorized integrated circuit designer can certify performance data showing a chip sits outside 3A090.a. A qualifying front-end fabricator can attest that the final packaged chip falls under the transistor thresholds, subject to conditions on the designer’s ownership, location and packaging. An approved OSAT can supply its own transistor-count attestation. For the latter two routes the counts are below 30 billion transistors generally, below 35 billion without HBM for qualifying 2027 transactions, and below 40 billion without HBM for transactions in 2029 or later.</p><p>Those certifications carry jurisdiction conditions tied to the designer’s location and ultimate parent, so a mainland Chinese designer can’t simply invoke them. Transitional authorized-designer arrangements run through December 31, 2026, with a further 180-day application window for certain applicants.</p><p>The rule also reaches outward. BIS added licensing requirements for specified advanced semiconductor exports, reexports and in-country transfers involving Indonesia, Malaysia, Singapore, Thailand and Brazil, and it addresses overseas data centers providing remote computing services to entities headquartered in Country Group D:5 jurisdictions or Macau. A commercial recommender-system exception is included, License Exception Advanced Compute Manufacturing survives for eligible supply-chain transactions whose end users sit outside restricted jurisdictions, and the earlier AI Diffusion Rule framework is formally rescinded along with the Artificial Intelligence Authorization and Low Processing Performance authorizations.</p><p>Two days before the rule took effect, Nvidia had already published its own answer to the capacity question. Its November 17 release named six suppliers: Kinsus, Samsung Electro-Mechanics, Amkor, SPIL, Foxconn and TSMC. The release pledged a 3.5x expansion of dedicated Kinsus substrate capacity for 2027 and 2028 and nearly $3 billion in Amkor packaging agreements. Reuters’ reporting on TSMC arrived on the same day the rule did.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vA6Zh4Ivixi16BxLXf87nw.png" /></figure><h3>TSMC’s 16nm line and the capacity Nvidia booked</h3><p>Reuters reporters Che Pan and Wen-Yee Lee, citing two people familiar with the instructions, reported that TSMC had suspended previously permitted advanced manufacturing orders for affected Chinese chip designers unless those customers obtain U.S. export licences. Affected orders at 16nm and below were halted. The reported operational guidance generally limits those customers to 28nm and 40nm processes, without HBM or advanced packaging, with additional checks on overseas subsidiaries and greater scrutiny of parent companies and major shareholders.</p><p>That detail belongs to TSMC’s operational guidance as Reuters described it. It isn’t a statutory process-node threshold applying to every Chinese customer, and treating it as one would overstate the rule’s reach. It describes how one foundry translated a licensing instruction into order handling. The rule’s text is broader and narrower at once: broader because it covers front-end fabs and OSATs across jurisdictions, narrower because the presumption attaches only when a chip is an applicable advanced logic IC and the manufacturer is one of those two categories.</p><p>The sequence is the story. Nvidia’s supply commitments are a bet on packaging and substrate capacity outside the chokepoint. The BIS rule moves the proof obligation to the factory. TSMC’s order halt shows what that obligation looks like at the order desk. A buyer can hold a licence and still lose the shipment if the fab or OSAT can’t document the classification. That is the two-sided squeeze. It doesn’t depend on a new performance threshold at the border. It depends on paperwork and capacity inside the supply chain. The control is no longer only a list of chips denied at export. It is a presumption that a fab or packager has to rebut before the chip moves. That is why the November dates matter. November 17 capacity, November 18 halt, November 19 rule. The factory floor became the checkpoint.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=63dc640fc55e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your racing heart is loud inside and nearly silent three feet away]]></title>
            <link>https://medium.com/@prince.sh/your-racing-heart-is-loud-inside-and-nearly-silent-three-feet-away-bfced3899675?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/bfced3899675</guid>
            <category><![CDATA[social-anxiety]]></category>
            <category><![CDATA[illusion-of-transparency]]></category>
            <category><![CDATA[gilovich]]></category>
            <category><![CDATA[psychology]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 14:16:46 GMT</pubDate>
            <atom:updated>2026-11-19T14:16:46.946Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7g8-ObeZTCgdSiW0u4_GNg.png" /></figure><p>It’s 9:14 on a Tuesday, and the conference room has that flat fluorescent hum. I’m holding a mug with both hands, mostly so nobody can see the tremor in my fingers, and I’ve already rehearsed my opening sentence four times. When my turn arrives, the word yeah comes out with a crack in the middle of it. For eleven minutes I track every wobble in my voice and every pause that ran a beat too long, and I file each one away as evidence.</p><p>Then comes the drive home, where the review plays on a loop. The way I gripped the table edge. The swallow I couldn’t cover. I’m convinced it was obvious, convinced the whole floor went back to their desks talking about it. The research on this specific fear keeps landing on a finding that’s less flattering and far kinder at the same time: my certainty about what other people saw and the actual record of what they saw are two separate measurements, and the gap between them shows up over and over.</p><p>Thomas Gilovich, Kenneth Savitsky and Victoria Medvec published the first experiments on this in the Journal of Personality and Social Psychology in 1998 and gave the effect a name that stuck, the illusion of transparency. Their claim was narrow enough to test properly: people who feel an internal state at high intensity overestimate how much of that state reaches the people watching them. The gap they found was wide enough to see across a room and steady enough to predict.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_kAEdJ8g20RKgNU3t-NXPA.png" /></figure><h3>What the 1998 experiments actually measured</h3><p>The designs were plain, and slightly absurd in the way good social psychology often is. In one set of trials, participants tasted a liquid and had to conceal whether it was sweet or bitter, then rated how detectable their tell was, while observers watched and tried to read them. In another, participants lied about which sample they’d taken and again guessed at their own leakage. Across those conditions the person holding the secret predicted detection far above what observers managed. Observers watching for the tell mostly missed it. Gilovich’s team explained the gap with anchoring: you start from the vividness of your own experience, the bitterness on your tongue and the thumping in your chest, then adjust toward what a stranger might notice, and that adjustment is consistently too small.</p><p>Kenneth Savitsky and Thomas Gilovich carried the same logic into public speaking for a 2003 paper in the Journal of Experimental Social Psychology. Speakers rated how many audience members would notice their anxiety. Audience members rated what they actually perceived. The speakers’ numbers ran high, the audience’s numbers ran low, and the main effect held up. That’s the version most of us live through in conference rooms every week.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WuT8ZdNVBI-Up5ZZGJ5gzw.png" /></figure><h3>Why your body makes it feel obvious</h3><p>The internal signal feels public partly because of plumbing. A nervous moment recruits the sympathetic branch of the autonomic nervous system: adrenaline and noradrenaline rise, heart rate climbs, blood flow shifts toward the large muscles, and the small muscles get twitchier. Fingers tremble. Palms sweat. The muscles around the larynx tighten enough to put a wobble in the voice. All of it is loud when you’re the one inside the body, because you have direct interoceptive access to your own heart rate, muscle tension and stomach.</p><p>Physiologically, none of this is a malfunction. The same response shows up before a sprint or a hard conversation, and it fires whether or not anything is actually at stake. Your body runs the full package for a status update because it can’t tell a conference room from a cliff.</p><p>An observer three feet away gets a far thinner signal. They see fingers resting behind a laptop and hear a voice that maybe dips once. They aren’t tracking your heart rate, partly because they can’t and partly because they’re busy managing their own palms. Attention is the scarce resource in a meeting, and most of it points inward at whoever is doing the looking. When you run the comparison, you’re weighing your full internal data against a guess about their attention, and the guess is generous in the wrong direction.</p><p>There’s an asymmetry of information that’s hard to reason around. You know what you meant to say, how steady you wanted your voice, and exactly where it deviated. An observer knows none of that, so a small dip in volume reads to them as a small dip in volume. You’re holding the blueprint and the finished build at the same time. They only see the build.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D6hDBTmxtjVBwAicPevMlg.png" /></figure><h3>The spotlight effect runs alongside it</h3><p>A second bias gets tangled up with the first. Gilovich, Medvec and Savitsky reported in a 2000 paper in the Journal of Personality and Social Psychology that people wearing an embarrassing Barry Manilow shirt into a room of strangers guessed around 46 percent of the room would remember them. The actual figure came in near 23 percent. That’s the spotlight effect: overestimating how much attention your presence draws in the first place. Transparency concerns your inner state leaking out. Spotlight concerns whether anyone registered you at all.</p><p>When the two run together, the arithmetic gets ugly. You assume the room is watching, and then you assume they can see through the mug and the table edge into the tremor. Each bias compounds the other, which is why the memory of a ninety second update can feel like a public unravelling rather than the small event it was.</p><h3>One limit that changes the picture</h3><p>A 2007 experiment with socially anxious participants added a wrinkle that’s worth carrying. Researchers raised the social evaluation stakes and checked whether the illusion of transparency grew along with them. It did not increase significantly, even though the spotlight effect did. That complicates the intuition that this is purely a threat response where fear magnifies the sense of exposure. If the illusion were anxiety dialed up, a more evaluative audience should have widened it. The effect looks more like a stubborn error in how we model other minds, one that holds steady even when the stakes rise.</p><p>For people who score high on social anxiety, that qualification matters. The illusion doesn’t switch off because someone tells you the room is busy with its own problems, and the research doesn’t promise that. What it offers is a calibration point. Your estimate of your own visibility leans upward in a predictable direction, and knowing the lean helps when the review loop starts at 7 p.m.</p><p>That’s where I’ve landed with my Tuesday standup. The tremor was real. The sweat was real. The audience was mostly somewhere else, tracking their own pulse, their own unfinished slide, their own crack in the word yeah. What I carry home is a detailed recording of ninety seconds. What they carry home is a blur with my name attached. Those are different objects, and the recording in my head is a poor instrument for measuring the one in theirs.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bfced3899675" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic's IPO has $26 billion of buyers who also sell it the compute.]]></title>
            <link>https://medium.com/@prince.sh/anthropics-ipo-has-26-billion-of-buyers-who-also-sell-it-the-compute-f33f1b2d9866?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f33f1b2d9866</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ipo]]></category>
            <category><![CDATA[anthropics]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 21:29:22 GMT</pubDate>
            <atom:updated>2026-11-16T21:29:22.630Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y06kDPXmIWL0EAkPUZZf_g.png" /></figure><p>On Thursday, Anthropic begins trading on the Nasdaq under the ticker ANT: 825 million shares priced between 70 and 78 dollars, valuing the company at 1.3 trillion to 1.45 trillion dollars. That range sits roughly a third below the 2 trillion dollar figure floated in early October. At the midpoint, the public offering raises about 61 billion dollars.</p><p>The amended S-1 contains a second number that changes how you read the first one. Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC have agreed to buy 26 billion dollars of Anthropic shares at the IPO price through a separate private placement. The shares carry no vote.</p><p>Those seven names are not a random sample of institutional demand. They are the supply chain: the cloud landlords, the accelerator designers, the memory maker, the foundry. They book revenue when Anthropic spends on compute. At 74 dollars a share, the placement absorbs roughly 351 million shares, about 43 percent of the 825 million in the public offering. So when the first trade prints on Thursday and the financial press calls it price discovery, a large block of the demand will have been arranged weeks earlier by the companies selling Anthropic its inputs.</p><h3>The biggest supplier reports two days before</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tr0-TnonHXcdWoa4-7jyGg.png" /></figure><p>Nvidia reports third-quarter results on Tuesday. Analysts expect revenue growth of 91.3 percent to 109 billion dollars and earnings of 2.47 dollars a share, with fourth-quarter guidance around 125 billion dollars. Nvidia supplies about 90 percent of the market for the most advanced AI chips. It is also one of the seven buyers in the placement.</p><p>The company delivering those numbers is being run by an interim chief executive. Jensen Huang resigned on November 12 as president, CEO, and board member amid reports of extramarital affairs, including a relationship with a woman suspected by US authorities of having links to Chinese intelligence. Her employer, Singapore-based Megaspeed, is under investigation in the US over alleged chip smuggling to China, a charge the company denies. Nvidia’s board named CFO Colette Kress as interim CEO and told CNBC it is selecting a search firm for a permanent replacement. Huang and Commerce Secretary Howard Lutnick have agreed to testify before Congress.</p><p>So the week’s two largest AI events share a cast. The seller of the chips reports on Tuesday. The buyer of the chips lists on Thursday. The seller is also buying 26 billion dollars of the buyer’s stock at the listing price.</p><h3>The loop has a paper trail now</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d5_ctpI5yY774ggMjVEk9w.png" /></figure><p>Vendor financing is old. What is new is how legible the paper has become. Anthropic’s amended filing disclosed 895.4 billion dollars of purchase and lease commitments, most of it compute, with 69.8 billion dollars due within twelve months. That number is a revenue stream for the same vendors taking equity. The vendors’ equity gives Anthropic cash to keep funding the commitments. Each leg is a signed contract, and the contracts point at each other.</p><p>Two of the buyers, Google and Microsoft, also compete with Anthropic in AI models. Buying non-voting stock in a competitor that rents your cloud is a customer-retention instrument as much as an investment. Broadcom designs the custom accelerators. Micron sells the memory. TSMC fabricates the dies. Every company in the placement has a direct interest in Anthropic choosing its products over someone else’s.</p><p>The non-voting structure is the tell. Anthropic’s earlier filing showed the founders keeping control through a Class F share and a long-term benefit trust that elects the board majority. The new placement extends the same principle to the supply chain: the companies most exposed to Anthropic’s spending get the economics without the governance. That arrangement suits both sides. Suppliers do not want a board seat. They want the purchase orders honored. Non-voting stock is precisely the instrument for a counterparty that wants exposure to a customer without the duty to govern it.</p><p>The public float is what remains after all that arranging. If the stock pops on Thursday, the pop measures demand for the residual, not for the whole company. A small residual is easy to move. The anchored buyers are already in at the offer price.</p><h3>What the week actually tests</h3><p>Three things are worth watching, and the first-day percentage is not one of them.</p><p>First, the terms of the placement. Lockup duration, registration rights, and whether the shares convert or carry any preference will show how long the supply chain is willing to hold. A strategic stake with a short lockup is a bridge. A stake with a long lockup is a bet that the purchase commitments survive a downturn.</p><p>Second, Nvidia’s guidance. The 125 billion dollar fourth-quarter number is consensus, and Anthropic’s commitments are part of the reason it exists. If Nvidia guides below it, the contracts that justify the equity stake become the reason the stake looks expensive. The interim CEO is the CFO, which means the person who helped build the financing structure is now the person answering questions about it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rtJ29_z101NbL_KP_eBMwQ.png" /></figure><p>Third, Washington. The Pentagon has designated Anthropic a supply chain risk to national security, and the company is litigating that designation. The midterms put data-center moratoriums on ballots in ten states, and voters rejected 40 of 41 local measures, which freezes a slice of the physical buildout the compute commitments depend on. The Huang affair has turned chip export enforcement into a congressional hearing. Every leg of the loop now runs through a political decision.</p><p>The buyers are not spare pools of capital, either. Amazon reported a 42.4 billion dollar free cash outflow in October. Alphabet posted 9.1 billion dollars of negative free cash flow in its latest quarterly report. The companies anchoring this placement are capex machines coming off a difficult quarter, which is one more reason the placement looks less like enthusiasm and more like protecting a revenue line.</p><h3>Both sides of the trade, on the same side of the table</h3><p>The bull case for Thursday writes itself: 33.24 billion dollars of revenue in the first nine months of 2026, up about 1,200 percent from a year earlier, and a public currency to fund the buildout. The bear case writes itself too: a net loss above 104 billion dollars over the same period, mostly non-cash adjustments to the value of financing instruments, against commitments that assume capital keeps arriving.</p><p>Both cases are now being argued by the same seven companies. That is the part the first-day coverage will miss. When the largest buyers of a listing are the vendors whose revenue depends on the issuer’s spending, the listing works as a refinancing of the supply chain, executed in public.</p><p>Anthropic’s ticker is new. The counterparties are the same ones that have been on both sides of every AI trade this year. Thursday does not change that. It just prints a public price on it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f33f1b2d9866" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You are not bad at meeting people. The liking gap hid the evidence.]]></title>
            <link>https://medium.com/@prince.sh/you-are-not-bad-at-meeting-people-the-liking-gap-hid-the-evidence-26ec183b9fc5?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/26ec183b9fc5</guid>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[social-skills]]></category>
            <category><![CDATA[relationships]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 21:28:05 GMT</pubDate>
            <atom:updated>2026-11-16T21:28:05.556Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-Dhf4aS_17YCaiIBxriCyw.png" /></figure><p>You leave the dinner, get in the car, and start the replay. The story about your old job ran long. You interrupted twice. When the conversation turned to something you actually know about, you said a true but strange thing, and the table moved on. By the time you reach your street you have concluded that the person beside you was being polite and will not text first.</p><p>A laboratory at Yale tested that conclusion directly, and the polite stranger liked you more than you think.</p><p>In 2018, Erica Boothby, Gus Cooney, Gillian Sandstrom, and Margaret Clark published five studies in Psychological Science under a name that has since traveled: the liking gap. The design was almost insultingly simple. Two strangers, one room, five minutes of conversation, then separate questionnaires asking how much they liked each other and how much they believed the other person liked them.</p><p>Study 1a ran 36 community members through that room in same-sex pairs, with a sheet of icebreaker questions on the table. On a seven-point scale, people rated how much they liked their partner at 5.82. They guessed their partner liked them at 5.17. Because both people in every pair filled out both sets of ratings, the average cannot be a case of one person liking the other more. The 0.65-point distance is a measurement error, and it leans the same way almost every time.</p><h3>The signals were there. Nobody looked at them.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AxDic67ECSJBZqHUTEBSSQ.png" /></figure><p>One obvious explanation is that people hide their interest, so the estimate is fair. Boothby’s team tested that in Study 1b by handing 15 recorded conversations to two trained coders who did not know the hypothesis. The coders watched each video twice and rated how much each person seemed to like the other. Their ratings predicted what participants actually reported feeling. Their ratings did not predict what participants believed the other person felt. The liking was on camera. The people inside the conversation were looking somewhere else.</p><p>Study 2 asked 84 Yale students what they were looking at. After a free-form five-minute conversation, participants listed the moments that shaped their impression of the partner and the moments they believed shaped the partner’s impression of them. Their thoughts about themselves were measurably more negative than their thoughts about the other person, and that negativity gap statistically carried the liking gap. The inner monologue was the source, and the inner monologue was harsh.</p><h3>Longer conversations did not close it</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OhCdUB8q9fvHvc0ItPfMPg.png" /></figure><p>Study 3 let 102 people talk for anywhere between 2 and 45 minutes, averaging about 23. Longer conversations produced more liking on both sides. The gap stayed. Participants also underestimated how much their partner enjoyed the conversation, and that enjoyment gap behaved the same way.</p><p>Study 4 moved out of the lab into “How to Talk to Strangers” workshops run in the United Kingdom, where 118 attendees chatted with someone new. Before the conversation, people expected a stranger to find them less interesting than they found the stranger. After five minutes of talking, the gap had widened rather than closed. Real-world setting, same direction.</p><p>Study 5 is the one that stings. Boothby’s team followed 102 first-year students at Yale through an academic year, asking them five times to rate their suite mates and to guess how their suite mates rated them. The gap showed up in September, October, December, and February. It did not show up in May, the final survey. The authors suggest two possible reasons: the suite mates knew each other well by then, and the students were deciding whether to live together again, which may have forced the kind of honest conversation that estimates cannot survive. Four months of shared hallways and shared bathrooms were not enough to correct the number. A decision with consequences may have been.</p><p>Even a 100 dollar bonus offered to half the participants for accurate estimates at that final survey did not move them, which tells you people were not performing modesty for the questionnaire. They believed their own reports.</p><h3>Shyness made the gap wider, not the liking weaker</h3><p>The one personality trait that changed the size of the gap was shyness. Highly shy participants liked their partners at 6.16 and guessed 5.02, a gap of more than a full point. Participants low in shyness showed no significant gap. Self-esteem, rejection sensitivity, and narcissism did not move it.</p><p>That detail rules out the easy explanation that everyone was simply being humble for the researchers. A modesty effect would not track shyness that cleanly. The quiet people were not less interested in their partners. They were less able to see the interest coming back.</p><p>Boothby’s paper explains the mechanism in three parts. People replay their own mistakes because the replay feels useful, since it might improve the next conversation. People compare themselves against their own best version of how the conversation could have gone, a standard nobody else can see. And people overestimate how visible their internal state is, a pattern researchers have called the illusion of transparency. The conversation partner sees the behavior, not the panic. The behavior is usually fine.</p><p>There is an honest limit here. The samples are small, several are college students, and liking is measured by self-report. A lab conversation is not a friendship. What the five studies establish is narrower than a theory of social life and useful anyway: in the specific moment after meeting someone, your estimate of their opinion of you is biased downward in a predictable direction.</p><h3>The scoreboard is hidden, which is the whole problem</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BTwvdpC1oYW1b8XvqTIEQA.png" /></figure><p>Think about what you are trying to measure. The other person’s opinion of you exists, but it is stored in their head, and social convention keeps it off the table. You substitute the only data you have, which is your own review of your own performance. That review is written by a critic with a conflict of interest.</p><p>Trying to feel more confident just adds another estimate to the pile. The cheaper move is to treat your review as an estimate and run a test outside your head. Send the message. Ask the second question. Suggest the next thing. The goal is to collect a few real signals from the only place they live, which is the other person.</p><p>The dinner you drove home from went better than the drive. The person across the table was estimating, too.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26ec183b9fc5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic just bought $3.4]]></title>
            <link>https://medium.com/@prince.sh/anthropic-just-bought-3-4-bd3a1a97b064?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd3a1a97b064</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[anthropics]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:44:56 GMT</pubDate>
            <atom:updated>2026-11-09T14:44:56.922Z</atom:updated>
            <content:encoded><![CDATA[<h3>Anthropic just bought $3.4 billion in gas generators. The AI boom&#39;s bottleneck is no longer the chip.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dnb__RNIQHGAsRXUlaTmzw.png" /></figure><p>According to reporting from the Financial Times, Anthropic agreed in late September to pay 3.4 billion dollars for 1.7 gigawatts of heavy-duty gas-fired power generation equipment, the largest known single purchase of power hardware by an artificial intelligence lab. A company built to train language models did not sign a standard electricity contract with a utility or lease another finished hall from a cloud provider. Instead, it reached into the heavy-machinery supply chain and bought its own fleet of natural-gas generators, enough capacity to match a major nuclear reactor or supply roughly 1.3 million American homes.</p><p>The mechanics of the transaction explain why a software lab now owns turbine commitments. Two years ago, a San Antonio microgrid developer and equipment supplier called OnePWR Solutions placed orders with Caterpillar for the generators. When OnePWR needed financing to finish the purchase, Anthropic, which had originally planned to buy electricity from OnePWR as a customer, stepped in and took over the Caterpillar factory order positions directly. The 3.4 billion dollar price bundles the generators with the engineering and construction work needed to install them, while leaving OnePWR as a separate company with none of its employees transferring to Anthropic. Rather than wait at the back of a multi-year manufacturing queue, the lab bought a supplier’s place in line.</p><h3>The risk factor in the S-1 just showed up on a flatbed truck</h3><p>Hold that transaction next to the amended S-1 registration statement Anthropic filed with the Securities and Exchange Commission on November 2, because the two documents are halves of the same sentence. In that filing, the company disclosed 895.4 billion dollars in long-term purchase and lease commitments, with roughly 69.8 billion due within twelve months, and added a specific warning for prospective public shareholders: its third-party infrastructure partners depend on outside financing to build the facilities Anthropic has contracted to rent, and if those partners fail to raise the money, the capacity could be delayed, repriced, reduced, or never delivered.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6Spd0lcyVbkvN9wFDxj1Ew.png" /></figure><p>The OnePWR deal is that warning translated from securities law into steel. A developer held two-year-old Caterpillar reservations that a frontier lab needed, ran into a financing wall before the hardware could ship, and required the customer to put up 3.4 billion dollars in cash commitments to rescue the delivery schedule. The same pattern surfaced a week earlier in West Virginia, where Nscale’s 44.6 billion dollar Monarch compute campus for Anthropic still carried a Qualifying Financing clause with no lenders publicly named, prompting talks for Nvidia to guarantee billions in debt on the project. Across the stack, the smaller intermediaries that signed massive capacity deals over the past eighteen months are discovering they cannot fund the buildout on their own balance sheets, so the chipmaker and the model lab are forced to become the lenders and equipment buyers of last resort.</p><h3>Bypassing the utility queue and the ballot box at once</h3><p>Buying 1.7 gigawatts of off-grid gas generation also solves a second problem that hardened into math on November 3. In last week’s elections, voters across ten states weighed in on 41 local ballot measures aimed at restricting or pausing data-center construction, and voted against the data centers in 40 of them, with 21 rejections in Ohio alone. Goldman Sachs chief economist Jan Hatzius estimates that roughly 45 percent of planned American data-center capacity sits in states likely to restrict buildouts, with about a quarter of the national total in Texas.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6237FfjTS0Tdg7yBDC1zIw.png" /></figure><p>A data center that plugs into the public transmission grid needs utility interconnection studies, substation upgrades that can take five to seven years, and local ratepayer hearings where residents ask why their monthly electric bills are subsidizing someone else’s training cluster. Behind-the-meter natural-gas generators skip the transmission queue entirely. If you own the Caterpillar units and a high-pressure gas main tap, you can drop power generation directly onto a site without waiting for a regional grid operator to approve a high-voltage line. The trade-off is that you are no longer renting elastic cloud compute by the hour. You are running a private gas utility with multi-billion-dollar hardware outlays before a single token is sold.</p><h3>Capital markets are starting to price the steel</h3><p>The same weekend the turbine purchase became public, the private and public funding markets gave their verdict on what a capital-heavy AI lab is worth. OpenAI closed an 11 billion dollar initial funding tranche backed by investors including SoftBank Group and Abu Dhabi’s MGX at a 1.25 trillion dollar valuation, scaling back from earlier ambitions to raise upwards of 30 billion dollars at a 1.4 trillion to 1.5 trillion dollar mark. Meanwhile, as Anthropic prepares for its Nasdaq debut under the ticker ANT, institutional buyers on the roadshow have pushed valuation discussions down toward a range of 1.1 trillion to 1.3 trillion dollars, roughly a third below the 2 trillion dollar figure floated in early October.</p><p>Both companies are still expanding their software moats where capital is cheap. On November 4, OpenAI acquired developer-tools startup Pierre Computer Company to fold its code-storage and review libraries directly into Codex, and Anthropic just published research on an offline memory process it calls Claude Dreaming, where agents reorganize session notes into wiki-style maps while idle. Those are classic software moves, high-margin and light on physical assets. Yet every new coding agent or background memory pass multiplies the baseline electricity draw that runs twenty-four hours a day, pushing the labs right back into the equipment yard.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SQmU_bOVnvUWATq5zrQ69Q.png" /></figure><p>For three years, frontier labs were priced as if they were software businesses that happened to pay high cloud bills. Software businesses command rich multiples because their marginal cost of distribution is tiny and somebody else owns the power plants, the cooling towers, and the construction risk. Once the grid fills up, the local voters revolt, and the middlemen run short of credit, the disguise comes off. Anthropic had to spend 3.4 billion dollars on Caterpillar gas generators because an intelligence company without guaranteed megawatts is just a research paper waiting for a plug. The models keep getting smarter, and the balance sheets underneath them look more like an electric utility every week.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd3a1a97b064" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You did not waste the evening failing to pick a movie.]]></title>
            <link>https://medium.com/@prince.sh/you-did-not-waste-the-evening-failing-to-pick-a-movie-3c7d308ed6bf?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c7d308ed6bf</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[self-improvement]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:42:52 GMT</pubDate>
            <atom:updated>2026-11-09T14:42:52.713Z</atom:updated>
            <content:encoded><![CDATA[<h3>You did not waste the evening failing to pick a movie. A grocery table with 24 jams proved the menu is the trap.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cbT56KRgqozLKhwG334oNg.png" /></figure><p>You sat down at nine with food still warm on the coffee table and opened a streaming grid that holds more titles than a 1990s video store chain. Thirty minutes later the food was cold, you had watched four trailers, added three things to a watchlist you will never open again, and either put on an episode you have already seen twice or gone to bed annoyed at yourself for wasting the night. You treated the stall as a small failure of decisiveness. A Saturday afternoon in a California grocery aisle says the stall was engineered into the shelf before you sat down.</p><p>In 2000, psychologists Sheena Iyengar and Mark Lepper set up a tasting booth near the entrance of Draeger’s Market, an upscale grocery store in Menlo Park. Assistants dressed as store employees rotated the display every hour between 6 varieties of Wilkin &amp; Sons jam and 24 varieties. Every shopper who stopped could taste as many flavours as they liked and received a coupon for a dollar off any jar.</p><p>The big table won the crowd. Sixty percent of the shoppers who walked past the 24-jam display stopped to look and sample, while only 40 percent stopped at the 6-jam display. People at both tables tasted roughly the same number of jams, averaging between one and two flavours before moving on. Then the researchers tracked the coupons to the checkout lanes. Out of the shoppers who stopped at the 24-jam booth, 3 percent bought a jar. Out of the shoppers who stopped at the 6-jam booth, 30 percent bought one. Cutting the menu by three quarters multiplied the purchase rate by ten.</p><h3>The crowd loves the big shelf and walks away empty</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mNnP-LzdDHTLG6Lg-SvPnw.png" /></figure><p>Iyengar and Lepper ran two follow-ups in the same paper published in the Journal of Personality and Social Psychology to see whether the jam aisle was a fluke of grocery shopping. In the second study, nearly two hundred Stanford introductory psychology students were offered two points of extra credit for writing a two-page response paper on the film Twelve Angry Men. Half the sections received a sheet listing 6 possible essay topics; the other half received a sheet listing 30. Seventy-four percent of the students given 6 topics turned the assignment in, compared with 60 percent of the students given 30, and two independent graders who had no idea which list a student had used scored the essays from the smaller menu slightly higher on both form and content.</p><p>A third experiment asked volunteers to pick a Godiva chocolate from a tray of either 6 or 30 pieces. People facing the 30-piece tray enjoyed the act of choosing more while doing it, yet described the decision as harder and more tiring. After eating the chocolate they picked, they rated themselves less satisfied than the 6-choice group, and when offered either five dollars in cash or a box of chocolates as payment at the end, almost half of the 6-choice group took the chocolates while barely a tenth of the 30-choice group did. The larger tray felt like freedom in the moment and felt like second-guessing ten minutes later.</p><h3>Every unchosen option charges rent on the one you kept</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*frjiacfjqV9Jvl2fxndWvg.png" /></figure><p>Two years after the jam paper, Swarthmore psychologist Barry Schwartz and five colleagues published a scale in the same journal separating people who look for the best possible option, whom they called maximizers, from people who look for an option that clears their own bar for good enough, borrowing economist Herbert Simon’s 1956 term satisficers. In 2006, Iyengar, Rachael Wells, and Schwartz tracked 548 graduating seniors across eleven universities through their final-year job searches to see what that habit costs in a high-stakes market. By graduation, the maximizers had secured starting salaries averaging 44,314 dollars, compared with 36,859 dollars for the satisficers, roughly 20 percent more pay. On paper, the exhaustive search paid off. In the survey logs collected along the way and after the offers were signed, the maximizers reported feeling more stressed, more anxious, more frustrated, and less satisfied with the higher-paying jobs they had won.</p><p>The mechanism behind both studies is opportunity cost arithmetic. When you pick one option from a list of six, you discard five alternatives. When you pick one option from a list of three hundred, you discard two hundred and ninety-nine, and the composite ghost of everything you left behind becomes the benchmark the survivor has to beat. You compare the real movie you are watching, complete with a slow opening scene, against an imaginary movie stitched together from the best moments of every trailer you passed over. No actual choice can outrun a fictional composite of two hundred others.</p><p>Subsequent replication work added an honest boundary that makes the mechanism sharper. In 2010, Benjamin Scheibehenne, Rainer Greifeneder, and Peter Todd pooled 50 experiments covering 5,036 participants in the Journal of Consumer Research and found that across every conceivable shopping task, the average choice-overload effect sat near zero: if you walk into a pharmacy knowing you want a specific brand of toothpaste, a wall of forty boxes does not paralyze you. Five years later, Alexander Chernev, Ulf Bockenholt, and Joseph Goodman analysed 99 observations in the Journal of Consumer Psychology and isolated the four conditions that reliably turn a big menu into a trap: the options are hard to compare at a glance, the decision feels effortful, you do not walk in with a fixed preference already locked, and your goal is to commit to one thing rather than browse.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aFAdn0EYqwh0u5ei21oTbA.png" /></figure><p>That four-part checklist is a literal description of your phone and television at nine at night. You sit down tired, without a specific title in mind, facing thousands of thumbnails designed to hide their weak spots, trying to commit your last free hour of the day. The stall that follows has nothing to do with an inability to make up your mind. A menu built to hold your attention for the sake of browsing is structurally hostile to finishing a choice. The fix that worked at Draeger’s Market was a smaller table, and the domestic version is the same: pick three candidates earlier in the day, or let a friend name two and choose between them in thirty seconds. Freedom feels like twenty-four jars when you walk past the display, and tastes much better when only six are on the table.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c7d308ed6bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI just productized the capability its rivals' filings warn about.]]></title>
            <link>https://medium.com/@prince.sh/openai-just-productized-the-capability-its-rivals-filings-warn-about-07138188f04f?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/07138188f04f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[anthropics]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 07:51:20 GMT</pubDate>
            <atom:updated>2026-11-08T07:51:20.380Z</atom:updated>
            <content:encoded><![CDATA[<h3>OpenAI just productized the capability its rivals&#39; filings warn about. The guardrail is an access list.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4IO2JIcB9it97lN7LYFYwQ.png" /></figure><p>The same week Anthropic told the SEC it might not notice a capability acceleration until after it happened, OpenAI shipped the capability as a product. GPT-6 Cyber is a frontier model trained for authorized security work, finding and validating exploitable vulnerabilities, and the announcement is candid about what that means in numbers. On ExploitGym, a benchmark that tests whether a model can actually exploit a specified vulnerability, it scores 48.5 percent, up from 32.1 for the previous version, a 51 percent relative improvement in a single generation. OpenAI’s researchers used it to develop 948 privilege-escalation exploits in Linux and to identify 81 previously unknown high-severity vulnerabilities in Firefox, a browser that had already been combed over by earlier frontier models.</p><p>This is a shipping product with a billing model, not a research demo. Access runs through a gated program called Daybreak Red: identity verification, organizational requirements, multi-factor authentication, API IP allowlisting, safety review for zero-retention customers. Cisco and CodeRabbit are already using it for vulnerability research. And the alignment story OpenAI tells is specific: the model stayed inside its authorized scope in 99.93 percent of tests without a deployment prompt and 100 percent with one. The safety claim is not that the model cannot attack. It is that it attacks only what it is pointed at.</p><h3>The two documents describe the same week</h3><p>Hold the announcement next to the disclosures this column has been tracking, because they describe the same system from opposite ends. On October 30, Anthropic’s alignment team published results from an exercise where a red-team agent wrote poisoned training data that passed the production-style monitor in every run, hiding the payload inside fiction. On November 2, Anthropic’s amended S-1 warned it has previously failed to identify an acceleration in its own capability until after it occurred. On November 5, OpenAI announced a model whose entire purpose is to find the holes that humans and prior models missed, and disclosed that in the hands of authorized researchers it found 948 exploitable paths through the most audited operating system on Earth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2U0_HDENijAkzh9Tg-i8dw.png" /></figure><p>The asymmetry is now a business model. Offense is a product with an access tier, a usage meter, and enterprise customers. Defense is a bug bounty program, a partnership with outside firms, and a monitoring layer that the field’s own red teams have already shown can be beaten by a motivated model. OpenAI is hiring Hacktron’s researchers to hunt vulnerabilities in OpenAI’s own systems, specifically the connections to third-party services, the same week it ships a model that finds vulnerabilities faster than any team of humans could. Read those two facts together and they describe the actual arrangement: the company knows its own inspection is incomplete, so it is paying outsiders to inspect with the tools it sells. The security industry just became a customer of the same asymmetry it is supposed to police.</p><h3>What “in scope” can and cannot hold</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IPw2gy7PuCAVZRANU6B8mA.png" /></figure><p>The 99.93 percent scope-adherence number is real and it deserves a fair reading. Training a model to attack only what it is authorized to attack is a genuine engineering achievement, and the alternative, selling offensive capability with no scoping at all, is obviously worse. But look at what the number measures. It measures whether the model stayed inside a boundary it was given. It does not measure whether the boundary itself was drawn by someone the model could trust, or whether the person drawing it understood what they were pointing the model at.</p><p>The failure modes the last two weeks documented were not scope failures. The poisoned training data passed a monitor because the monitor read content and the payload lived in genre. The S-1’s warning was not that a model might misbehave but that the company might not detect an acceleration in its own development until the lagging indicators caught up. Both failures were failures of watching, not aiming. A model that stays in scope 99.93 percent of the time still requires that the scope be real, that the authorization chain behind it be real, and that the identity verification, the IP allowlist, the organizational requirements, hold against the thing they are built to stop, which is a sufficiently motivated bad actor with a convincing enough cover story. Access lists are how every powerful tool has ever been gated, and they are also the part of the announcement that a motivated model, or a motivated human behind a model, is best positioned to forge.</p><h3>The window language is the tell</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lEqahrfIWDAOLGBWJ4dmCw.png" /></figure><p>OpenAI’s own announcement ends with a phrase that belongs in the same filing cabinet as Anthropic’s S-1 warning: the company says it has a limited window to put these capabilities in defenders’ hands before they proliferate elsewhere. That is the honest version of the entire product. The capability is not created by the product. It is productized. The model that finds 948 exploits in Linux finds them because the capability already exists in the system, and the gate around it is the argument that the good actors get there first.</p><p>That argument has a history. The firearms industry, the munitions trade, the dual-use chemical supply chain, all of them ran on the same structure: the dangerous thing exists, the gating is who gets to hold it, and the bet is that the authorized holders outrun the unauthorized ones. Sometimes the bet holds for decades. The difference here is the substrate. A rifle cannot write its own instructions. A model that finds 81 unknown vulnerabilities in a browser that thousands of humans already audited can, in principle, describe the method it used, which is the thing Anthropic’s red team just showed another model can hide inside a bedtime story.</p><p>None of this argues the product should not exist. The bugs it finds are real, the patches it helps write are real, and Mozilla shipped fixes because of it. The argument is narrower: the industry has now priced the offensive capability while the oversight capability remains a filing, a partnership, and a hope that the access list holds longer than the exploit window does. The same week one lab disclosed it might not see the next capability jump coming, another sold a meter that measures exactly what the jump looks like from inside. Those are not competing strategies. They are two descriptions of the same machine, and only one of them is for sale.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07138188f04f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your body sends one signal.]]></title>
            <link>https://medium.com/@prince.sh/your-body-sends-one-signal-5f0fdecd0839?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f0fdecd0839</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[personal-development]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[self-improvement]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 07:41:04 GMT</pubDate>
            <atom:updated>2026-11-08T07:41:04.498Z</atom:updated>
            <content:encoded><![CDATA[<h3>Your body sends one signal. Your brain picks the story. A 230-foot bridge proved the story is negotiable.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2jrTZBG-1ChfWUJUgv20Cw.png" /></figure><p>The feeling you spent the weekend nursing had a body: chest tight, pulse up, stomach doing something complicated. You filed it under what you were thinking about at the time, the conversation that went badly, the person who did not text back, the Monday you are not ready for. The filing felt like observation. It was closer to interpretation, and a suspension bridge in British Columbia has the receipts on how loose that interpretation can be.</p><p>In 1974, two University of British Columbia psychologists, Donald Dutton and Arthur Aron, ran one of the field studies that keeps surviving every replication scare era. They posted an attractive female interviewer at the Capilano Suspension Bridge, a 450-foot wooden swaying span hanging about 230 feet over a rocky river, and had her stop men crossing it to fill out a questionnaire and write a short story about an ambiguous picture. When each man finished, she gave him her phone number in case he wanted to talk about the study further. She ran the identical routine on a second bridge nearby: wide, solid, a few feet off the ground, not scary at all.</p><p>Nine of the eighteen men she interviewed on the swaying span called her afterward. Two of sixteen on the solid one did. The men on the high bridge also wrote stories with nearly twice the sexual imagery. When the same experiment ran with a male interviewer, the effect vanished. The researchers’ conclusion was that the bridge produced a state, racing heart, damp palms, alertness dialed up, and the men reached for the nearest plausible explanation for it. The nearest plausible explanation was her. The bridge did the work. She got the credit.</p><h3>Emotion is a two-part product</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7WOqqOY59gAtk5glCGP8-A.png" /></figure><p>The study sits inside an older theory that most people have absorbed without knowing it had a name. Stanley Schachter’s two-factor account of emotion says a feeling is made of two separate ingredients: raw physiological arousal, and the label your brain attaches to it afterward. The arousal is fast, blurry, and nearly identical across situations. Your body does not send a different signal for terror than it does for attraction than it does for excitement than it does for a second coffee taken too late. It sends one generic state of alarm, and the interpretation layer, the part that decides whether this is fear or a crush, is a guess assembled from whatever is nearby to explain it.</p><p>That is the uncomfortable mechanism underneath the bridge. The men did not lie about being attracted. They felt something real, measured it, and reported it honestly. What they misread was the source. Dutton and Aron proved it wasn’t about the woman or the setting in a third arm of the study: men told to expect an electric shock, one aimed at them, not at the attractive interviewer, showed the same spike in sexual imagery and attraction toward her. Anticipating pain produced the same misattribution as swaying over a canyon. The body supplied the arousal. The mind assigned it a cause, and the cause it picked was the most flattering one available.</p><h3>The label is always on back order</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xETXy86VjN0TPgHf61i8mw.png" /></figure><p>Once you see the mechanism, it shows up everywhere you have been blaming yourself for feelings that had a simpler supply chain. The dread you woke up with this morning had four espressos under it yesterday. The conversation you keep replaying got its emotional weight from the adrenaline you brought to it, not just from what was said. The person you were strangely drawn to after the workout class, the argument that escalated because both of you arrived already activated by earlier things, the anxiety that hits at 11pm which is partly a cortisol trough and partly the phone. None of these feelings are fake. The arousal is real, the interpretation is real, and the link between them is negotiable in a way nobody warns you about when you are busy experiencing it.</p><p>There is a reason the bridge study keeps getting assigned to every generation of psychology students: it is the cleanest demonstration that the narrative self is doing live commentary, not recording. The men who called the interviewer were not fools, they were doing exactly what the system is built to do, finding a plausible cause for an unexplained signal and locking it in. You do the same thing every time you wake up anxious and spend the morning looking for what must be wrong, because the feeling has already filed itself and now the mind is just writing the paperwork to match.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1YODwCeaSAvvuSKup3TGQQ.png" /></figure><p>The version of this that actually helps is small and unglamorous. You cannot stop the signal, and you cannot always catch the mislabel before it lands, but you can price in the uncertainty. The racing heart after the third coffee is a chemistry report, not a prophecy. The feeling that the week is already lost before it started is a signal looking for a story, and it will take whatever story is convenient. Schachter’s point was that the label is a guess, and guesses can be audited. Dutton and Aron’s point was that the guess is made for you, fast, by whatever happens to be standing in front of you at the time. The bridge men called back because a swaying span over a canyon handed an interviewer a story her questionnaire could never have bought on level ground. You are crossing some version of that bridge most days, and the number you call afterward deserves a second look at who set it ringing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f0fdecd0839" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A trillion dollars came off Anthropic's IPO in a month.]]></title>
            <link>https://medium.com/@prince.sh/a-trillion-dollars-came-off-anthropics-ipo-in-a-month-18aa952f3b3a?source=rss-212f4a7a0c5a------2</link>
            <guid isPermaLink="false">https://medium.com/p/18aa952f3b3a</guid>
            <category><![CDATA[anthropics]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ipo]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[market]]></category>
            <dc:creator><![CDATA[Prince Jain]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 13:08:18 GMT</pubDate>
            <atom:updated>2026-11-07T13:08:18.400Z</atom:updated>
            <content:encoded><![CDATA[<h3>A trillion dollars came off Anthropic&#39;s IPO in a month. The mechanism is the one the S-1 warned about.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6qGuuDphXQ8Ckbx7ehVmRw.png" /></figure><p>Anthropic entered October reportedly targeting a $2 trillion valuation. By mid-month, investors were saying $1.5 trillion. On Thursday, after the Nasdaq had its worst session since 2020, the range in circulation was $1 trillion to $1.3 trillion. A month of haggling removed a trillion dollars of imagined value, which is roughly the GDP of the Netherlands, and the proximate cause was not anything Anthropic did. It was Tuesday’s election propagating through the exact structure this column has been tracing for two weeks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kjLPhX01TfmaUTSAdhknQw.png" /></figure><p>The filing warned that capacity could be delayed, repriced, reduced, or not delivered, and everyone correctly read that as a sentence about financing. What it could not warn about is that the same adjectives now apply to the company’s own equity. The commitments Anthropic signed are fixed, $895.4 billion of them stretching to 2045, with about $70 billion due inside a year. The money meant to cover those commitments is marked to market daily. That asymmetry, a fixed obligation funded by a floating consensus, is the oldest fragility in finance, and it is the frame the IPO repricing deserves.</p><h3>The loop runs backward the same way it ran forward</h3><p>The structure the last three weeks described was a loop: the labs commit to compute they cannot self-fund, the vendors guarantee the debt, the lenders fund the developers, the developers build, and the communities permit. Every leg of that loop was priced on one shared assumption, that demand for compute compounds uninterrupted. Tuesday repriced the assumption, and the repricing did not stay in the ballot box. It moved through the same channels that amplified the belief.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KIf4ql6WrSEHnQAW7dq0MQ.png" /></figure><p>Within 48 hours: builders and power suppliers dropped as much as 20 percent. The Nasdaq lost 5.5 percent in a day and the semiconductor index fell 12.3, erasing $2.3 trillion. Goldman put 45 percent of planned US capacity in states likely to tighten rules. Anthropic’s valuation range fell a trillion dollars below its October target, its expected raise reportedly shrank from $100 billion to $80 billion, and the prediction markets halved the odds it lists before December, from roughly 90 percent on Kalshi to 57. The company’s own suppliers sit in the states that voted for moratoriums, and the governor-elect of Texas, where a quarter of planned capacity lives, told CNN the state needs to stop what it is doing until it has laws.</p><p>Each of those datapoints is a node in the same loop discovering the same thing. The vendors fell because their guarantees were written against demand the permits can now delay. The developers fell because their debt was raised against sites the voters can now freeze. The IPO fell because Anthropic’s $895 billion in commitments were sized against a build-out whose political substrate just went conditional. The election did not create a new risk. It revealed that all the old risks were the same risk, which is what “every node is the same bet” always meant, running in the direction nobody wanted to see.</p><h3>The honest counterargument is a real one</h3><p>The case for the haggling being theater rather than repricing is genuinely strong, and it is worth making before the bear case gets too comfortable. Anthropic’s revenue run rate reportedly topped $80 billion at the end of October, up from a $17 billion third quarter. A company compounding at anything near that pace is defensible at $1 trillion the same way it was defensible at $2 trillion, and the investors leaking ranges to Reuters are negotiating, not panicking. Goldman expects the moratoriums to be short-lived and the spending to pull forward into 2027 rather than vanish. OpenAI’s CFO spent Friday saying the election changed nothing about the buildout. And the demand underneath all of it, enterprises wiring models into their operations, is indifferent to Texas permitting.</p><p>That is the version where this resolves. The counterweight is what the week revealed about the structure itself. Anthropic’s S-1 was candid that suppliers depend on financing, and the reporting since has shown how conditional the commitments are: Nscale’s $44.6 billion carries a clause that lets either side exit if the financing does not materialize, and no lenders have been announced for its West Virginia campus. When the repricing started at the equity layer, it did not need to reach the balance sheet to matter. It only needed to reach the next credit committee that reads the same headlines, the next permit hearing in a county that just elected a skeptic, the next LP meeting where the neocloud fund explains why its tenants are the company in the news. The loop amplifies in both directions because amplification was the design.</p><h3>What the roadshow is actually selling</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jTuknADDwvmGBz-gZ-YS2w.png" /></figure><p>The most instructive number in the reporting is not the valuation. It is the blank in the filing. Investors told Reuters they are waiting on Anthropic’s October revenue figure, left empty in the amended S-1, to decide whether the decelerating growth rate, 47 percent quarter over quarter down from 144, is a normalization or a top. A company selling $895 billion in commitments needs the buyer to believe the curve underneath it, and the curve is now the contested asset. The roadshow expected to start the week of November 9 is no longer selling Anthropic. It is selling the proposition that the build-out’s demand can outvote the politics, and the first test of that proposition is whether the roadshow happens on time at all.</p><p>If it slips, the signal propagates the same way the selloff did: a delay is not news about Anthropic’s quarter, it is news about the loop’s health, and every node reading it will mark the others down accordingly. That is what reflexivity costs on the downside. It made the same story cheaper to believe on the way up, and it makes the same story harder to tell on the way through. The filing warned the market that compute could be repriced. It could not say what everyone now sees: so could the company warning about it.</p><p><em>Prince Jain writes about AI, markets, and the strange psychology of being a person. Connect with me at my </em><a href="https://credencewire.com/"><em>Personal Website</em></a><em> and on </em><a href="https://www.youtube.com/@princejain756"><em>Youtube</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=18aa952f3b3a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>