<?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 Henry Feng on Medium]]></title>
        <description><![CDATA[Stories by Henry Feng on Medium]]></description>
        <link>https://medium.com/@henryfeng_4644?source=rss-6d73b6d55519------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*-KqofK_jGJpdl6g-.</url>
            <title>Stories by Henry Feng on Medium</title>
            <link>https://medium.com/@henryfeng_4644?source=rss-6d73b6d55519------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 06:05:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@henryfeng_4644/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[ThiColor, translate the colors you see into the language of paint.]]></title>
            <link>https://medium.com/@henryfeng_4644/thicolor-translate-the-colors-you-see-into-the-language-of-paint-316236b8f5f2?source=rss-6d73b6d55519------2</link>
            <guid isPermaLink="false">https://medium.com/p/316236b8f5f2</guid>
            <dc:creator><![CDATA[Henry Feng]]></dc:creator>
            <pubDate>Sat, 20 Jun 2026 07:39:40 GMT</pubDate>
            <atom:updated>2026-06-20T07:39:40.071Z</atom:updated>
            <content:encoded><![CDATA[<h3>ThiColor, translate the colors you see into the language of paint. Everyone is racing to make AI replace us. I made one that makes us better.</h3><h3><em>How I spent two weeks teaching an AI to think like a painter, and what I learned about color, math, and the kind of AI I actually want to live with.</em></h3><p>For the past year, the loudest pitch in tech has been some version of the same sentence:</p><p>You don’t need to draw anymore. You don’t need to write anymore. You don’t need to design anymore. You don’t even need to write code anymore. The AI will do it.</p><p>I work in AI. I build AI products for a living. And I’ve spent more hours than I can count painting, badly, in the corner of my apartment.</p><p>Those two parts of me have been at war for a while.</p><p>So two weeks ago I decided to try the opposite hypothesis. <strong>What if I built an AI that didn’t replace the painter, but taught them how to see?</strong></p><p>I want to walk you through what I built, why, and what I learned along the way. The app is called <strong>ThiColor</strong>. It’s on the App Store today.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xSbL8TTw2THRXQEHBPPJWg.png" /></figure><h3>The problem I’ve been carrying around for a decade</h3><p>If you’ve ever painted seriously, you know this feeling. You look at a Monet. The purple in his water lilies. The exact pitch of a sky he saw one afternoon in 1899.</p><p>You go home, you squeeze pigment onto your palette, and you try to mix it.</p><p>You fail.</p><p>Not because your hand is wrong. Because you have no idea <strong>which tubes he reached for</strong>. Which yellow he used to break up that ultramarine. Which white he chose so the mixture wouldn’t go chalky.</p><p>Paint is also expensive. A tube of cadmium yellow can run forty dollars. A tube of cobalt blue, half that. Ten failed mixes and you’ve squeezed a hundred bucks straight into the trash.</p><p>What I’ve wanted, for years, is a tool that could look at any color in front of me and translate it into a sentence:</p><p><em>“Use these three tubes. In this ratio. That’s how you mix it.”</em></p><p>So I sat down with Claude Code and started building.</p><h3>My first version was completely wrong, and that’s where it got interesting</h3><p>The naive approach is obvious. Sample the color from a photo, get its RGB, then walk through every pigment in the library, weighted-average their RGB values, find the combination closest to the target.</p><p>I wrote it in an afternoon. It worked. The math was clean. The recipes the app suggested were elegant.</p><p>And when I actually mixed them on a real palette with real paint, the results were <strong>garbage</strong>. Muddy, dull, wrong in a way that’s hard to describe unless you’ve stared at it.</p><p>I went back to Claude and asked the question I should have asked first:</p><p><em>“Why does mixing pigments by averaging their RGB values produce the wrong color on canvas?”</em></p><p>Its answer taught me something I had been painting around for ten years without ever putting into words:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vQTX0vijLyOcUB3y9AK7RA.png" /></figure><p>Then it pointed me to a piece of color science from 1931 called the <strong>Kubelka–Munk model</strong>. It’s what the industrial paint industry has used for nearly a hundred years to formulate the color of paint in the can you buy at Home Depot. The math models how pigments actually combine in a film of binder.</p><p>I threw away my first version and rewrote the entire solver around K–M.</p><p>What the app gives you now isn’t a screen-color approximation. It’s a <strong>physically grounded paint recipe</strong> — a combination of pigments in measurable ratios that, if you actually mix them on a palette, will produce the color you pointed your camera at.</p><p>That was the first moment where it clicked. The AI didn’t replace anything I do as a painter. <strong>It taught me a piece of my own craft.</strong>Then I taught the AI to think the way a painter actually thinks</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SDcvCpJ1NSc6_RaofF24AQ.png" /></figure><p>The K–M version was beautiful and slow. Forty-eight pigments in the library, up to three pigments per recipe — that’s roughly three million combinations to evaluate, each one requiring a full color-difference calculation. On a phone it took several seconds. Unusable.</p><p>I told Claude: <em>“This has to be faster. By a lot.”</em></p><p>It said: <em>“Want to try pruning the search space?”</em></p><p>I said: <em>“How?”</em></p><p>And it said something that I had to sit with for a few seconds before I understood:</p><blockquote><em>A real painter doesn’t pick up all 48 tubes and try each one. They glance at the target color, reach for the five or ten tubes that look closest, and start mixing from there. Your algorithm can do the same thing.</em></blockquote><p>So we rewrote it. Rank the pigments by how close each single one gets to the target. Take only the closest 12. Build the two-pigment and three-pigment combinations from that pool.</p><p><strong>Three million calculations dropped to seventeen hundred. Five seconds dropped to fifty milliseconds.</strong></p><p>And the accuracy barely moved. Because the optimal mixture, almost always, lives in that small neighborhood of hue-adjacent pigments anyway. Just like a painter knows.</p><p>This is the piece of code I’m proudest of in the whole project. The AI didn’t replace the painter’s intuition with brute force. <strong>It translated the painter’s intuition into an algorithm.</strong></p><p>That, to me, is what AI should do.</p><h3>Then I got greedy. I wanted to paint with Monet’s palette.</h3><p>Once the solver worked, an idea wouldn’t leave me alone.</p><p>What if I could constrain the search to only the pigments <strong>Monet actually used</strong>? Or Vermeer. Or Sargent. Or Zorn.</p><p>Not in a metaphorical sense. Literally — the exact tubes that sat on their palette.</p><p>The idea is more than a gimmick. It’s a teaching method. You don’t just learn a color. You learn <strong>the way a specific master saw the world</strong> — the limited language of pigment they chose to translate everything they looked at into.</p><p>The problem: this data does not exist as an API. There’s no JSON file of “Monet’s palette.” Just scattered evidence across painter biographies, museum technical research, and pigment chemistry history.</p><p>So I went and dug. Painters’ letters. Conservation reports. The pigment chemistry literature on what was commercially available in Northern France in the 1880s. Then I encoded it.</p><p>Four masters, hand-curated:</p><ul><li><strong>Monet</strong> — Impressionist palette, the way he stripped earth tones out and worked almost entirely in cool blues, lemon yellow, vermilion, and lead white.</li><li><strong>Vermeer</strong> — including his famous use of natural lapis lazuli ultramarine, with a small set of earths and a lead-tin yellow that hasn’t been commercially produced in three hundred years.</li><li><strong>Sargent</strong> — the alla prima palette he used for his society portraits, viridian and rose madder and ivory black.</li><li><strong>Zorn</strong> — the legendary four-color palette: yellow ochre, vermilion, ivory black, titanium white. That’s it. Some painters spend their entire careers trying to do skin tones with just those four.</li></ul><p>I even added historical corrections. Vermeer’s lead white is toxic and no modern painter can use it, so the app substitutes titanium white. Sargent’s vermilion is replaced by cadmium red. Each substitution is annotated so you can see what was swapped and why.</p><p>The effect is what I hoped. You point your phone at the orange basketball hoop outside your office window, tap “Monet’s palette,” and the app tells you which combination of cobalt blue, ultramarine, cadmium yellow, and lead white Monet would have used to render that hoop.</p><p>You’re not just looking at a color anymore. <strong>You’re borrowing a master’s eyes.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9rlpp0X0ALPSpTwwjtpMJg.png" /></figure><h3>The second core feature: watching a master paint, step by step</h3><p>Every art student has wanted the same thing, secretly. Not to see the finished painting in a museum. To see <strong>the process</strong>. The first mark on the blank canvas. The order things happen. When the underpainting dries. When the artist switches from a brush to a palette knife.</p><p>This is the part the textbooks can’t really teach you, and the part most art classes never have time to walk you through.</p><p>So the second feature in ThiColor is called <strong>Breakdown</strong>.</p><p>You hand it a reference photo. It splits the painting into three stages — <strong>underpainting, blocking-in, final</strong> — and shows you what each stage would look like if you painted it in that style. Beneath each stage, it lists the new colors introduced and gives you a Kubelka–Munk paint recipe for each one.</p><p>The hard part isn’t generating images. The hard part is the <strong>prompt</strong>.</p><p>To make the AI actually understand the difference between, say, Impressionism and academic realism, you can’t just say “paint this in an Impressionist style.” You have to tell it that in Impressionism the underpainting doesn’t lay down hard outlines — it establishes color-temperature relationships. That the blocking-in stage has to avoid smooth gradients. That the final stage is made of broken color: small juxtaposed strokes that the eye fuses into form, no smooth surfaces, no hard edges, the figure dissolving into a field of color.</p><p>I rewrote that prompt twenty-something times with Claude before it produced steps a painter would actually recognize as Impressionist process.</p><p>Then the loop closes. The pigment recommendations from feature one get plugged into each step of feature two. So when stage two introduces a warm orange in the foreground, the app doesn’t just show you the color — it tells you to mix cadmium yellow medium with a touch of burnt umber, in this ratio, on a palette you actually own.</p><p><strong>It isn’t a reference image. It’s a screenplay you can follow with a brush.</strong></p><p>I want to share one detail because it captures, for me, why I love working this way.</p><p>Each step of a breakdown is a one-to-two-megabyte PNG. If a user opens their history and waits three seconds to see the first thumbnail, the experience falls apart.</p><p>So every time the server generates a step image, it also produces a 384-pixel JPEG thumbnail at quality 78. That’s roughly 30 to 50 kilobytes — a single packet over 4G. The client uses a small SwiftUI component called ProgressiveImage that stacks the thumbnail underneath the full image. The thumbnail appears almost instantly. When the full PNG finishes decoding, it swaps in on top, and the thumbnail layer is released so memory doesn&#39;t double.</p><p>I deliberately didn’t add a fade transition between the two layers. Because while a breakdown is still generating, the client polls the server every few seconds for status, which causes the parent view to re-render. A fade transition would make the already-loaded image flicker out and back in on every poll — making it look like the network had dropped. A hard cut looks worse in isolation, but it looks <strong>honest</strong> in motion.</p><p>That kind of detail is the thing that grinds you down when you build software alone. With an AI sitting next to you, walking through tradeoffs in plain English, you can actually polish all the way down to the floor.</p><p>This is what “vibe coding” means to me, when I strip away the hype:</p><p><strong>It’s not the AI skipping the hard parts for you. It’s the AI sitting with you while you don’t skip them.</strong></p><h3>The bugs were real, and they were stupid</h3><p>I want to be honest about the unglamorous parts.</p><p>I tried to ship version 1.5 to the App Store. Built and uploaded five times in a row. Every single upload came through to Apple as <strong>version 1.0</strong>. I lost an entire morning before Claude and I traced it: an old Info.plist had a hardcoded &lt;string&gt;1.0&lt;/string&gt; for the version string, which silently overrode the build setting. Twenty seconds to fix, two hours to find.</p><p>I created a subscription product in App Store Connect, then deleted it because I’d named it wrong. Apple’s docs don’t tell you, but the product ID is <strong>permanently burned</strong>. Once you’ve used it, you can never recreate one with the same identifier — not even after deletion. I had to rename basic.yearly to basic.yearly2 across the client enum, the StoreKit configuration, and the server&#39;s receipt-to-tier mapping. Three files, ten minutes with Claude, problem solved.</p><p>My server got auto-stopped one morning because the billing account hit a low balance. The Breakdown feature went down. I was SSH-ed in within five minutes, with Claude on the other monitor helping me write a balance-check script so it wouldn’t happen again.</p><p>Ten thousand lines of Swift. Two thousand lines of Python. Four submitted builds. A surprising number of small, real, embarrassing mistakes.</p><p>Every one of them was a real bug. And every one of them was a small lesson, made survivable because there was a smart, patient collaborator in the loop.</p><h3>What I actually believe about AI, after these two weeks</h3><p>There are too many people talking about AI right now. So many that I’ve reached the point of not wanting to talk about it at all. I just want to use it to build things.</p><p>I’ve also stopped believing the “AI replaces humans” framing, at least in the way it gets repeated in venture pitches and on tech Twitter. In two weeks of working with one, every single hour produced the opposite outcome:</p><p>It didn’t paint for me. <strong>It taught me the physics of paint.</strong><br> It didn’t think for me. <strong>It translated my intuition into a working algorithm.</strong><br> It didn’t learn to paint for me. <strong>It showed me how Monet would have rendered the basketball court outside my window.</strong></p><p>The AI helped me build this tool. And the tool, in turn, will help another group of people — painters, illustrators, students — see colors they’ve never been able to see and follow steps they’ve never been able to follow.</p><p>A person, made better because of AI. Who then uses that same AI to make more people better.</p><p>That’s the loop I think AI should be running.</p><p><strong>Not the one where it takes our hands. The one where it opens our eyes.</strong></p><h3>How to find it</h3><p>It’s called <strong>ThiColor</strong>. On the App Store, worldwide.</p><ul><li>The color picker and basic recipes are free, with a daily quota.</li><li>Master palettes and Breakdown are part of the Basic and Pro subscriptions, which mostly exist to cover the cost of running the AI image-generation service.</li></ul><p>If you paint, or if you’ve ever been frustrated by trying to mix a color from a photo, I’d love for you to try it. If you build with AI and the kind of work I described here resonates with you, I’d love to hear from you.</p><p>This is the project I’m proudest of right now. I hope it’s useful to you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=316236b8f5f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[UnaPod, AI read and audio assist, How You Handle Information Overload Might Change!]]></title>
            <link>https://medium.com/@henryfeng_4644/unapod-ai-read-and-audio-assist-how-you-handle-information-overload-might-change-46d377751076?source=rss-6d73b6d55519------2</link>
            <guid isPermaLink="false">https://medium.com/p/46d377751076</guid>
            <category><![CDATA[read]]></category>
            <category><![CDATA[audio]]></category>
            <category><![CDATA[rss]]></category>
            <category><![CDATA[pódcast]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Henry Feng]]></dc:creator>
            <pubDate>Thu, 11 Jul 2024 09:00:56 GMT</pubDate>
            <atom:updated>2024-07-11T09:02:14.271Z</atom:updated>
            <content:encoded><![CDATA[<p>In recent years, we’ve been overwhelmed with information from YouTube, Twitter, TikTok, and countless news media. This is particularly challenging for startups, researchers, investors, and knowledge-intensive professionals. Although subscribing to a news RSS feed helps stay updated, we still encounter issues such as multiple channels reporting on the same topic, insufficient time to read everything during a busy day, and a lack of prioritization in the importance of the information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*--SH8IqxViBWWdPCIDNgnA.png" /></figure><p>So we built a tool to help you manage and distill information efficiently. In Unapod, you can:</p><ul><li><strong>Choose Your Interests</strong>: Subscribe to news media RSS links or provide specific interest keywords.</li><li><strong>AI Summaries</strong>: Our AI highlights key points from your saved content.</li><li><strong>Read or Listen</strong>: Access summarized content by either reading the article or listening to the audio podcast while driving.</li></ul><p><a href="https://forms.gle/ZMwna45t5phPgTfp8">Join the waitlist</a> and transform how you manage information. Or you could get more info about this on <a href="http://unapod.com/">unapod.com</a></p><p>#Startup #FounderLife #AI #Productivity #Innovation #unapod</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46d377751076" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>