<?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 Pratyush Shivam on Medium]]></title>
        <description><![CDATA[Stories by Pratyush Shivam on Medium]]></description>
        <link>https://medium.com/@pratyushshivam?source=rss-8829ca1e29da------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Y_JvvGkO66TWu4GPidWvdw.jpeg</url>
            <title>Stories by Pratyush Shivam on Medium</title>
            <link>https://medium.com/@pratyushshivam?source=rss-8829ca1e29da------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 22:13:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pratyushshivam/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[Mount Everest in a browser tab: the real mountain, the real route, the real air]]></title>
            <link>https://medium.com/@pratyushshivam/mount-everest-in-a-browser-tab-the-real-mountain-the-real-route-the-real-air-c981424a1683?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/c981424a1683</guid>
            <category><![CDATA[mount-everest]]></category>
            <category><![CDATA[webgl]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 00:00:55 GMT</pubDate>
            <atom:updated>2026-11-17T13:40:16.124Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ITSdzBpu-OV4YMPq.jpg" /></figure><p>Base Camp to 8,848.86 m by the South Col. Six weeks of expedition in twelve minutes. Every gauge on screen is physics, pinned to arterial blood drawn from climbers at 8,400 m. Every frame of this film is a real climb, played by the game’s own test climber.</p><h3>The engineering</h3><ul><li>Terrain from space. TanDEM-X radar satellite data (Copernicus GLO-30): a 1,024 × 1,024 grid at one arc-second over 28 × 32 km, plus a 512 × 512 horizon so Lhotse, Makalu and Cho Oyu stand exactly where they stand. Radar rounds off knife-edges and reads the summit 125 m low, so the ridge is lifted to four surveyed heights, tapered to zero within 400 m. Disclosed in the README.</li><li>Detail below the data. One sample every 30 m can’t hold an icefall tower. Within 190 m of the climber, procedural seracs, crevasses and wind-packed snow are generated and blended onto the real surface.</li><li>The sky. NOAA solar equations for the exact date and Nepal clock time, shadows cast by the real relief. Summit night: every star to magnitude 5.6, rotated to Everest’s latitude and the hour.</li><li>The air. West’s Himalayan model atmosphere: 253 mmHg on the summit, matching the 1981 measurement. The textbook atmosphere says 236. It’s wrong up there.</li><li>The blood. Alveolar gas equation + Severinghaus haemoglobin curve with the Bohr shift, fitted to the 2007 NEJM samples. Model blood oxygen at 8,400 m: within 2.5 mmHg of measured.</li><li>Bottled oxygen. Regulator flow expanded by the gas laws into every breath. 2 l/min lowers the mountain by about 2 km. It never takes you to sea level.</li><li>Seeded weather. Jet stream on the summit through April, short calm windows in late May. Same seed, same season, every time.</li><li>40 tests against published measurements, plus whole simulated expeditions: the patient climber summits and comes home. The one who ignores the turnaround time doesn’t.</li></ul><p>No server. No tracking. No recordings: every sound is synthesised in the browser.</p><p>Climb it: <a href="https://pratyushshivam.github.io/thin-air/">pratyushshivam.github.io/thin-air/</a></p><p>Source: <a href="https://github.com/pratyushshivam/thin-air">github.com/pratyushshivam/thin-air</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/mount-everest-in-a-browser-tab/"><em>https://www.pratyushshivam.com</em></a><em> on November 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c981424a1683" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[One night in the Amazon rainforest in a browser tab: real forest, real animals, real sound]]></title>
            <link>https://medium.com/@pratyushshivam/one-night-in-the-amazon-rainforest-in-a-browser-tab-real-forest-real-animals-real-sound-6e611388f6f6?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e611388f6f6</guid>
            <category><![CDATA[amazon-rainforest]]></category>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[webgl]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 00:00:39 GMT</pubDate>
            <atom:updated>2026-11-17T13:52:54.949Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9MjdsIFRQLhH_JWt.jpg" /></figure><p>Understory is the floor of Yasuní, Ecuador, from 17:45 to 06:25, with one torch. Every frame of this film is the game running.</p><h3>The engineering</h3><ul><li>The forest. 702 trees per hectare and 10.3% stilt palms, from the 25-hectare Yasuní plot, with heights from Feldpausch’s fitted curve for western Amazonia. About 2,400 trunks, 5,400 plants and 560 animals, instanced: one draw call per species and per chunk.</li><li>The light. One torch, held at your eye, so every shadow hides behind what casts it. No shadow maps. Mist in the beam, dawn shafts, tone mapping and grain are one full-screen pass with 14 ray-march steps. If the frame rate drops, resolution goes first.</li><li>The eyes. Eyeshine is your own torch coming back: 1/d² on the way out, 1/d² on the way back, scaled by pupil area and the reflecting layer behind the eye. No reflecting layer, no shine. Computed per eye by the same function the tests check.</li><li>The sound. 25 real field recordings from xeno-canto, each played from where its animal sits through an HRTF panner. −6 dB per doubling of distance, plus a forest absorption curve that takes the high notes first, so a howler carries across the whole site and a katydid fades within tens of metres.</li><li>The ecology. 58 species, each with a record from Tiputini or Yasuní, at its real height and only in its own hours. Placement, activity windows and dawn order are unit-tested: 52 tests.</li><li>Deterministic. Every night is seeded. Night 7 is the same trees, the same animals and rain at the same minute, for everyone.</li></ul><p>The README lists what is not real right next to what is.</p><p>No server. No tracking.</p><p>Walk in: <a href="https://pratyushshivam.github.io/understory/">pratyushshivam.github.io/understory/</a></p><p>Source: <a href="https://github.com/pratyushshivam/understory">github.com/pratyushshivam/understory</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/one-night-in-the-amazon-in-a-browser-tab/"><em>https://www.pratyushshivam.com</em></a><em> on November 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e611388f6f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From space, Mount Everest is 125 metres too short]]></title>
            <link>https://medium.com/@pratyushshivam/from-space-mount-everest-is-125-metres-too-short-08ef2b733432?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/08ef2b733432</guid>
            <category><![CDATA[mount-everest]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 00:00:36 GMT</pubDate>
            <atom:updated>2026-11-17T14:03:12.577Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*u6vuk9QHW4uMT7Ut.jpg" /></figure><p>I sampled Copernicus GLO-30, the global elevation model built from the TanDEM-X radar satellites, along the South Col route. It puts the summit at 8,724 m. The highest pixel anywhere near the top reads 8,737 m. The surveyed height is 8,848.86 m.</p><p>And it isn’t one bad pixel. The whole ridge comes out low:</p><ul><li>South Col: −59 m</li><li>Balcony: −122 m</li><li>South Summit: −152 m</li><li>Summit: −125 m</li></ul><p>Lhotse, next door, loses 111 m.</p><p>Why: every pixel is a 30 m square. Average a knife-edge ridge over 30 metres and you get a rounded hill. Side-looking radar also struggles on faces this steep.</p><p>I found this building Thin Air, my Everest climb in the browser. The mountain came out short, so the ridge is raised back to four surveyed heights, fading to nothing within 400 m, and the README says exactly where.</p><p>The lesson travels well beyond mountains: anything sharper than your resolution gets smoothed away. Before you trust a number, ask what the pixel can see.</p><p>Climb it: <a href="https://pratyushshivam.github.io/thin-air/">pratyushshivam.github.io/thin-air/</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/from-space-everest-is-125-metres-too-short/"><em>https://www.pratyushshivam.com</em></a><em> on November 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=08ef2b733432" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[No engineering beats this number: the speed of light sets your latency]]></title>
            <link>https://medium.com/@pratyushshivam/no-engineering-beats-this-number-the-speed-of-light-sets-your-latency-9cba9f3b1615?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/9cba9f3b1615</guid>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[networking]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 00:00:46 GMT</pubDate>
            <atom:updated>2026-11-17T13:11:03.077Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5HmlCwcljWalh_Q6.jpg" /></figure><p>Frankfurt to Sydney and back can never take less than 161 milliseconds. Not with better code, not with a faster database, not with more servers in the same place.</p><p>That is physics. Light in optical fibre travels at about 204,000 km a second, 68% of its speed in vacuum, and a request has to go out and the answer has to come back. The chart is the best case: a cable running in a perfectly straight line, with nothing else in the way. Real networks are slower, because cables follow coastlines and every router adds time.</p><p>A few things this changes once you see it:</p><ul><li>100 ms of round trip buys about 10,210 km of fibre. One data centre can serve a little over a quarter of the planet well. Nobody can serve the far side under 196 ms.</li><li>“Make it faster” sometimes means “move it closer”. The fix for a slow API in Singapore is often not in the code at all.</li><li>Every extra round trip multiplies the distance. A page that makes five calls in a row from Frankfurt to Tokyo has spent half a second on physics before any work is done.</li><li>It is why systems built for many places need many copies of the data, and why keeping those copies in agreement is one of the hard problems in distributed systems.</li></ul><p>I built a small game around this limit. You get a globe and a budget, cities ask for answers in under 100 ms, and every latency in it is computed from the same physics as this chart. Nothing is made up, and the far side of the planet stays out of reach no matter what you build.</p><p>Play Latency: <a href="https://pratyushshivam.github.io/latency/">pratyushshivam.github.io/latency/</a></p><p>What is the longest round trip your users live with?</p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/speed-of-light-latency-limit/"><em>https://www.pratyushshivam.com</em></a><em> on November 16, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9cba9f3b1615" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI found 81 Firefox flaws. How fast you ship the fix is now your security]]></title>
            <link>https://medium.com/@pratyushshivam/ai-found-81-firefox-flaws-how-fast-you-ship-the-fix-is-now-your-security-8b93d440533b?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b93d440533b</guid>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 00:00:20 GMT</pubDate>
            <atom:updated>2026-11-17T14:19:08.701Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LYjb63FfUWYH8Cf1.jpg" /></figure><p><strong>81.</strong> That is how many previously unknown, high-severity flaws in Firefox OpenAI says its new security model, GPT-6 Cyber, found before attackers did. Mozilla has started shipping the fixes. The same model reportedly produced working exploits for 948 Linux kernel bugs.</p><p>The numbers are OpenAI’s own, and the model is not public. But the direction is clear, and it changes how engineering teams should think about security.</p><p>For years, the scarce thing in security was finding the bug. It took rare people and a lot of time. That is now becoming cheap, for defenders and for attackers alike.</p><p>So the scarce thing moves. It becomes how fast a team can go from “this is broken” to “the fix is live”:</p><ul><li>Can you trust your tests enough to ship a patch today, not next sprint?</li><li>Do you know every service that runs the vulnerable code?</li><li>Can one person roll out a fix without a release ceremony?</li><li>Can you roll it back in minutes if the fix breaks something?</li></ul><p>If AI can read your code faster than your team can, your release process is now part of your security.</p><p>The teams that do well will not be the ones with the most scanners. They will be the ones that can ship a fix the same day.</p><p>How fast could your team ship a security fix today?</p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/ai-found-81-firefox-flaws/"><em>https://www.pratyushshivam.com</em></a><em> on November 16, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b93d440533b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The entire observable universe in a browser tab, true to scale, in 2.95 MB]]></title>
            <link>https://medium.com/@pratyushshivam/the-entire-observable-universe-in-a-browser-tab-true-to-scale-in-2-95-mb-225c0a5f3339?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/225c0a5f3339</guid>
            <category><![CDATA[webgl]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[astronomy]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 00:00:08 GMT</pubDate>
            <atom:updated>2026-11-17T13:37:18.987Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/0*-R8JptnuICJOe0sp.jpg" /></figure><p>Earth to 46.1 billion light-years in one continuous shot: 16,301 asteroids, 109,400 real stars and 156,940 real galaxies, every one where it actually is. 60 fps on a laptop.</p><h3>The hard part was the engineering</h3><ul><li>Precision. A GPU has 32-bit floats, about 7 digits. Earth to the edge of the universe is 20 orders of magnitude. So every object stores its position relative to its parent in float64, the camera has no world position at all, and values are narrowed to float32 only at the last step. A naive renderer loses a full kilometre at Earth’s orbit; this keeps it to a micrometre.</li><li>Depth. No depth buffer spans 1,000 km to 1⁰²⁶ m. Stars and galaxies skip it entirely; each planet gets its own range fitted around its sphere. Zero flicker.</li><li>Compression. 8 bytes per star: octahedral direction in 32 bits, log distance in 16, brightness and colour in 8 each. 109,400 stars in 875 kB, sorted brightest first so phones get the familiar sky early.</li><li>Memory. Catalogues stream in and out by scale, textures step up from 512 to 4,096 px and are released when a planet shrinks. Heap stays at 15–29 MB the whole way. 1 to 4 textures and 3 to 15 draw calls per frame.</li><li>Zero frames drawn when nothing moves.</li><li>Tested against the real sky: the 2020 Jupiter-Saturn conjunction and the 2024 solar eclipse.</li></ul><p>Fly it: <a href="https://pratyushshivam.github.io/lookback/">pratyushshivam.github.io/lookback/</a></p><p>Source: <a href="https://github.com/pratyushshivam/lookback">github.com/pratyushshivam/lookback</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/the-entire-universe-in-a-browser-tab/"><em>https://www.pratyushshivam.com</em></a><em> on November 16, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=225c0a5f3339" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I folded a paper plane out of physics]]></title>
            <link>https://medium.com/@pratyushshivam/i-folded-a-paper-plane-out-of-physics-b792e046746f?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/b792e046746f</guid>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 00:00:49 GMT</pubDate>
            <atom:updated>2026-11-17T13:30:05.758Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/0*hdWJ_teBt_sQHGRW.jpg" /></figure><p>Paper Plane is a game in your browser. A top-floor window, a long hand-drawn world of rooftops, a park, a river and hills. Pull back, aim, let go. After that, the air decides.</p><p>Nothing you see in the flight is a drawn curve.</p><h3>What is under the hood</h3><ul><li>The plane is a point with a pitch angle, and 240 times a second three forces push it and one turns it. Lift and drag both come from the angle of attack, the angle between the wing and the air moving past it, and both grow with the square of the speed.</li><li>The stall is real. Past about 13 degrees the model blends from a lifting wing to a flat plate: lift falls away, drag jumps, and the plane drops until it has the speed to fly again. Throw it at the ceiling and it lands at your feet. That is the model, not a rule.</li><li>Pitch has its own equation. The nose is heavy, so the plane keeps turning to meet the air at about 5.5 degrees. Holding the screen raises that by about 4 degrees: the only control in the air. It trims at 6.5 m/s and glides 7.4 metres forward for every metre down.</li><li>The swoop is not scripted. Throw it hard and level and it climbs, slows, noses over, speeds up and climbs again. That trade of height for speed is called a phugoid, and it falls out of the equations on its own.</li><li>The air moves. Wind that is slower near the ground and gusts over time, warm air rising off sunny roofs, cold air sinking over the river, lift where the wind meets a slope, and a cushion of air near the ground. Every force uses the plane’s speed through the air, not over the ground.</li><li>One sky a day. The world and its weather come from a seed, and the seed is the date, so everyone flies the same sky and a distance means something. The step is fixed and nothing in flight is random: the same throw into the same sky lands in the same place every time. That is also why the aiming arc is the first 0.6 seconds of the real flight, not a guess.</li><li>Skill is measurable. A look-ahead autopilot flies each day against a hands-off throw. On the first six skies, letting go reaches 128 to 417 metres; careful holding reaches 205 to 527.</li><li>Every line is drawn by the same wobbly pen, sound is made on the spot with WebAudio, and the whole thing is about 28 kB of script. 24 tests. No account, no server, no tracking.</li></ul><p>Everything in the film is real footage from the page, or drawn from the game’s own functions. The throw is a real mouse drag. The 524.9 metre flight crosses 7 columns of warm air, and the holds in it were chosen by that same look-ahead autopilot.</p><p>Play it: <a href="https://pratyushshivam.github.io/paper-plane/">pratyushshivam.github.io/paper-plane/</a></p><p>Source: <a href="https://github.com/pratyushshivam/paper-plane">github.com/pratyushshivam/paper-plane</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/paper-plane-physics/"><em>https://www.pratyushshivam.com</em></a><em> on November 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b792e046746f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I built a black hole, then flew into it]]></title>
            <link>https://medium.com/@pratyushshivam/i-built-a-black-hole-then-flew-into-it-a97a092028a4?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/a97a092028a4</guid>
            <category><![CDATA[black-holes]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[webgl]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 00:00:24 GMT</pubDate>
            <atom:updated>2026-11-17T13:35:31.240Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/0*zmfgA3otTehKZA6r.jpg" /></figure><p>Event Horizon is a page where you fly a ship around a real black hole. Your clock and Earth’s clock start together. They do not stay together.</p><p>Nothing you see is a video or a painting. For every pixel, every frame, a ray of light is traced backwards through curved spacetime on your graphics card, live, in the browser.</p><h3>What is under the hood</h3><ul><li>Light follows the null geodesic equation of the Schwarzschild metric, d²u/dφ² = −u + (3/2)u², integrated per pixel with fourth-order Runge-Kutta in at most 100 steps. Stepping in angle instead of distance is what makes it fast enough to run live. A ray is either captured (the black shadow), escapes (you see the sky, bent), or crosses the disk, two or three times near the photon sphere, which draws the thin photon ring.</li><li>The disk glows by real physics: the Page &amp; Thorne (1974) thin-disk formula, brightest at 4.78 Schwarzschild radii, each ring a blackbody, shifted by gravity and Doppler together. The side coming towards you is brighter and bluer, by g⁴.</li><li>The ship flies a real timelike geodesic. Newton has r where general relativity has r − 3GM/c²; that one term makes orbits precess and turns circular orbits unstable below 3 radii.</li><li>Two clocks, added up along the path you actually fly. At the station Earth runs 1.07× faster than you, at the innermost stable orbit √2×, and at the photon sphere without limit. Fall in, and from Earth you are never seen to cross.</li><li>Three real black holes at their published masses: Sagittarius A* (4.3 million Suns), M87* (6.5 billion) and Cygnus X-1 (21 Suns). From mass and distance alone, the page computes Sagittarius A*’s shadow at 53.3 microarcseconds; the Event Horizon Telescope measured 51.8 ± 2.3.</li><li>The sky is real too: NASA’s Milky Way and 41,486 stars from the HYG catalogue, each at its catalogued position and colour.</li><li>45 tests against results known in advance. About 30 kB of code. 60 frames a second on a laptop. No account, no server, no tracking.</li></ul><p>Everything in the film is real footage from the page, or drawn from the page’s own equations.</p><p>Fly it: <a href="https://pratyushshivam.github.io/event-horizon/">pratyushshivam.github.io/event-horizon/</a></p><p>Source: <a href="https://github.com/pratyushshivam/event-horizon">github.com/pratyushshivam/event-horizon</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/black-hole-ray-tracing-in-the-browser/"><em>https://www.pratyushshivam.com</em></a><em> on November 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a97a092028a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A game a five-year-old can play and an engineer can read]]></title>
            <link>https://medium.com/@pratyushshivam/a-game-a-five-year-old-can-play-and-an-engineer-can-read-24f5030228ce?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/24f5030228ce</guid>
            <category><![CDATA[typescript]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 00:00:19 GMT</pubDate>
            <atom:updated>2026-11-17T13:33:36.815Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/0*eAMuz_9HLHvMDrlB.jpg" /></figure><p>Junction: cars, vans and trucks drive into a yard from all four sides. With one finger you draw each one a path to the garage of its colour. Don’t let them bump. You have three hearts. It starts calm.</p><h3>What is under the hood</h3><ul><li>The whole game is a pure, deterministic simulation in TypeScript: no page, no clock, no Math.random. A round is three things, a seed, the shape of the field and the lines you draw, stepped at a fixed 60 steps a second. Same seed, same lines, same round, every time.</li><li>Today’s round is seeded from the date. Every arrival takes exactly five numbers from a small seeded generator (Mulberry32): which edge, where along it, which vehicle, which garage, and the gap until the next. So everyone gets the same cars in the same order and can compare scores.</li><li>A line is just a list of points, thinned to one every 1.2 units. End it near the right garage and it snaps and locks. The vehicle walks the list at its own speed, and its body turns a little slower than its direction, which is what makes the steering look smooth.</li><li>Bumps are measured, not guessed. Every vehicle is a capsule, a line down its middle with half its width as the radius, and two touch when the distance between their lines is less than their two radii. Fifteen times a second the engine looks ahead along every line, six looks 0.3 s apart up to 1.8 s, and puts a “!” between any pair about to meet.</li><li>The difficulty is one formula: the gap between arrivals is max(1.4, 5.8 × 0.972ⁿ) seconds. Vans join at the third arrival, trucks at the tenth, pairs from the fifteenth. New garages open at 5, 13 and 24 delivered.</li><li>The yard is the same 16,000 square units on every screen, tall on a phone and wide on a laptop, so the traffic is just as dense either way.</li><li>Sound is made on the spot with WebAudio, one note per garage from a five-note scale, so any mix of deliveries sounds fine together. No audio files, no images.</li><li>Every garage has a colour and a shape, so the game never depends on telling two colours apart.</li><li>About 45 kB in total. 21 tests. No account, no server, no tracking.</li></ul><p>Everything in the film is a real round, played with real pointer events, or drawn from the game’s own code.</p><p>Play it: <a href="https://pratyushshivam.github.io/junction/">pratyushshivam.github.io/junction/</a></p><p>Source: <a href="https://github.com/pratyushshivam/junction">github.com/pratyushshivam/junction</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/junction-deterministic-game/"><em>https://www.pratyushshivam.com</em></a><em> on November 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=24f5030228ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I taught fifty cars to drive. Nobody showed them how]]></title>
            <link>https://medium.com/@pratyushshivam/i-taught-fifty-cars-to-drive-nobody-showed-them-how-fc333eb508ac?source=rss-8829ca1e29da------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc333eb508ac</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Pratyush Shivam]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 00:00:59 GMT</pubDate>
            <atom:updated>2026-11-17T17:55:11.647Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*avFAdRNW_Hu1OaUt.jpg" /></figure><p>Rookie is a page where you draw any race track with your finger, and fifty AI cars learn to drive it in front of you, from nothing, in under a minute. Then you race the one you trained.</p><p>No training data. No dataset. No backprop. No ML library. Everything is written from scratch and runs on your own device.</p><h3>What is under the hood</h3><ul><li>Each car sees the world through 7 distance sensors fanned out from 75° left to 75° right, plus its own speed. Eight numbers between 0 and 1, sixty times a second. It never knows where it is on the track or which way round it goes.</li><li>Its brain is a feed-forward neural network: 8 inputs, 6 hidden neurons, 2 outputs (steering and pedal), tanh on every neuron. 68 weights in total. Those 68 numbers are the entire driver.</li><li>It learns by neuroevolution, a genetic algorithm over the network’s weights. Rank all fifty by how far they got. The best four go through untouched, so the best can never get worse. Every other car is bred: parents picked as the best of three from the better half, weights crossed over from one parent or the other, then about one in five nudged at random. If progress stalls, the mutations grow 30% per generation and six complete newcomers join.</li><li>Generation 1 is a pile-up. In the run in this film, the first clean lap came in generation 5. Nobody drew the racing line you see wearing into the road. It is simply where the survivors kept driving.</li></ul><h3>The numbers</h3><ul><li>200 random tracks nobody had seen: all 200 learned to a clean lap. Median generation 6.</li><li>One physics step for all fifty cars takes about 0.03 ms on a laptop, around 600× faster than real time.</li><li>Deterministic: same seed, same evolution, weight for weight.</li><li>The track and the trained driver fit after the # of a link, so you can send someone your rookie to race.</li><li>The whole page is about 58 kB. No account, no server, no tracking.</li></ul><p>Everything in the film is real footage from the page or drawn from the real numbers in the code.</p><p>Try it: <a href="https://pratyushshivam.github.io/rookie/">pratyushshivam.github.io/rookie/</a></p><p>Source: <a href="https://github.com/pratyushshivam/rookie">github.com/pratyushshivam/rookie</a></p><p>I’m Pratyush Shivam, CTO of Konfidio. More of what I build and write is at pratyushshivam.com.</p><p><em>Originally published at </em><a href="https://www.pratyushshivam.com/writing/neuroevolution-fifty-cars-learn-to-drive/"><em>https://www.pratyushshivam.com</em></a><em> on November 14, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc333eb508ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>