<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Elias Shieh</title>
    <description>The latest articles on DEV Community by Elias Shieh (@elias-shieh).</description>
    <link>https://dev.to/elias-shieh</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4197173%2F9f92babd-474f-4c49-90e3-d6113e93567a.png</url>
      <title>DEV Community: Elias Shieh</title>
      <link>https://dev.to/elias-shieh</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/elias-shieh"/>
    <language>en</language>
    <item>
      <title>I can't draw. My AI ghosts named my crab, then turned it into a spider.</title>
      <dc:creator>Elias Shieh</dc:creator>
      <pubDate>Sat, 31 Oct 2026 12:56:21 +0000</pubDate>
      <link>https://dev.to/elias-shieh/i-cant-draw-my-ai-ghosts-named-my-crab-then-turned-it-into-a-spider-2da6</link>
      <guid>https://dev.to/elias-shieh/i-cant-draw-my-ai-ghosts-named-my-crab-then-turned-it-into-a-spider-2da6</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;I whispered "a crab with enormous claws" to a table of AI ghosts.&lt;/p&gt;

&lt;p&gt;The first ghost drew it. The second ghost looked at the drawing and said "a caterpillar with a propeller on its head."&lt;/p&gt;

&lt;p&gt;Then it was my turn. I can't really draw, but I drew my crab. The next ghost said "A crab with a single antenna that sticks straight up—odd for its species."&lt;/p&gt;

&lt;p&gt;The ghosts took it from there. The last one said "spider with a face and spider legs."&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fndj64gzt6eqx6e63spox.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fndj64gzt6eqx6e63spox.png" alt="The whole crab round in seven steps: my whisper, Pip's drawing, Dot's caterpillar guess, my crab drawing, Moth's guess, Hush's drawing, and Wisp's spider guess." width="800" height="1006"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That is Monster Telephone: the party game Telephone, played with ghosts. You are the only living player at the table. Each ghost sees only the turn before it, so you are the only one who remembers the original. When they lose your monster, you draw it back into the story. At the end you see your monster beside what it became.&lt;/p&gt;

&lt;p&gt;The challenge asks about the memories and stories we carry. Here the ghosts carry nothing from one turn to the next, so the player has to.&lt;/p&gt;

&lt;p&gt;Drawing is optional, and a ghost can take your turn. There is no score and nothing to bet. It's for anyone who can't draw and has nobody around to play Telephone with.&lt;/p&gt;

&lt;p&gt;I wanted this game to be fun, cute and surprising. I can't draw, but I still get a turn to draw my monster back into the story.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Play: &lt;a href="https://monstertelephone.tech" rel="noopener noreferrer"&gt;monstertelephone.tech&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Press "Watch one go wrong" for a 15-second recorded chain. The first one is the crab above, with my real drawing. Press "Try saving a monster" for a live round. The server sleeps when nobody is playing, so the first turn can take a minute.&lt;/p&gt;

&lt;p&gt;The table is Pip the enthusiast, Dot the overthinker, Moth the dreamer, Hush the worrier and Wisp the dramatist. When your turn comes, the page shows your original whisper and a sheet of paper.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzbgw24enirbnrm9xiabn.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzbgw24enirbnrm9xiabn.png" alt="The &amp;quot;Draw it back&amp;quot; screen. A note says &amp;quot;They've lost your monster&amp;quot; above my whisper, and my crab is on the paper. Five small ghosts sit around the stage." width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;At the end the page shows your whisper beside what came back. "Save this chain" downloads the round as a picture, and "Copy chain link" gives you a link to send to a friend.&lt;/p&gt;

&lt;p&gt;The page ships with nine recorded rounds. In these four, a ghost took the player's turn.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffwam7ir3bovmge39e1o9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffwam7ir3bovmge39e1o9.png" alt="Four recorded rounds. A ghost wearing a tiny crown ended as a cat with tarantula legs. A teapot with human feet ended as a stickman. A cat with a second head ended as an animal with chest fur. A very polite vampire ended as a deer with thin legs." width="800" height="811"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/ting-hong-shieh" rel="noopener noreferrer"&gt;
        ting-hong-shieh
      &lt;/a&gt; / &lt;a href="https://github.com/ting-hong-shieh/monster-telephone" rel="noopener noreferrer"&gt;
        monster-telephone
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Telephone with monsters: keep yours alive through a table of AI ghosts
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Monster Telephone&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;a href="https://monstertelephone.tech" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/a83c762769caaec049d3dd70427fc71bccb0fae5876fd7c66a6f8b87fa29809b/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f706c61792d6d6f6e7374657274656c6570686f6e652e746563682d616431373364" alt="Play"&gt;&lt;/a&gt;
&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/a6e1f6893f5704e01d6a693e223ab58bfa3dea8990fb59cfd71eb2ac2159bee8/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4170616368652d2d322e302d313432333366" alt="License: Apache-2.0"&gt;&lt;/a&gt;
&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/releases/tag/artist-2" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/eb02c61a49a3276ac1a66c7fb88b964e38ebe19510be0debb7f9efe806cc27ca/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f776569676874732d6172746973742d2d322d313432333366" alt="Weights: artist-2"&gt;&lt;/a&gt;
&lt;a href="https://dev.to/elias-shieh/i-cant-draw-my-ai-ghosts-named-my-crab-then-turned-it-into-a-spider-2da6" rel="nofollow"&gt;&lt;img src="https://camo.githubusercontent.com/6c5e345733c8ad7bc25a0fa557748f04b49b1bc32285287ed5500141cc017c8e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f73746f72792d4445562d313432333366" alt="Story on DEV"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Telephone with monsters. You whisper a monster to a table of AI ghosts. One draws it, the next guesses from the drawing alone, and the story drifts. You are the only one who remembers the original, so at your turn you try to draw it back.&lt;/p&gt;
&lt;p&gt;Play at &lt;a href="https://monstertelephone.tech" rel="nofollow noopener noreferrer"&gt;monstertelephone.tech&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/ting-hong-shieh/monster-telephone/docs/assets/crab-chain.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fting-hong-shieh%2Fmonster-telephone%2FHEAD%2Fdocs%2Fassets%2Fcrab-chain.png" alt="A recorded round saved from the game. The whisper &amp;quot;a crab with enormous claws&amp;quot; came back as &amp;quot;spider with a face and spider legs&amp;quot;."&gt;&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What is open&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Piece&lt;/th&gt;
&lt;th&gt;Where&lt;/th&gt;
&lt;th&gt;License&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Game page and server&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/site" rel="noopener noreferrer"&gt;&lt;code&gt;site/&lt;/code&gt;&lt;/a&gt;, &lt;a href="https://github.com/ting-hong-shieh/monster-telephone/server" rel="noopener noreferrer"&gt;&lt;code&gt;server/&lt;/code&gt;&lt;/a&gt;
&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Training and evaluation code&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/teach.py" rel="noopener noreferrer"&gt;&lt;code&gt;teach.py&lt;/code&gt;&lt;/a&gt;, &lt;a href="https://github.com/ting-hong-shieh/monster-telephone/learn.py" rel="noopener noreferrer"&gt;&lt;code&gt;learn.py&lt;/code&gt;&lt;/a&gt;
&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;955 training drawings, held-out prompts, results, recorded chains&lt;/td&gt;
&lt;td&gt;&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/data" rel="noopener noreferrer"&gt;&lt;code&gt;data/&lt;/code&gt;&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test rules, raw results, blind answers, every guess&lt;/td&gt;
&lt;td&gt;&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/tests" rel="noopener noreferrer"&gt;&lt;code&gt;tests/&lt;/code&gt;&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fine-tuned artist, a rank 64 LoRA adapter&lt;/td&gt;
&lt;td&gt;&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/releases/tag/artist-2" rel="noopener noreferrer"&gt;artist-2 release&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Base models&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://huggingface.co/Qwen/Qwen3.5-4B" rel="nofollow noopener noreferrer"&gt;&lt;code&gt;Qwen/Qwen3.5-4B&lt;/code&gt;&lt;/a&gt;, &lt;a href="https://huggingface.co/Qwen/Qwen3.6-35B-A3B" rel="nofollow noopener noreferrer"&gt;&lt;code&gt;Qwen/Qwen3.6-35B-A3B&lt;/code&gt;&lt;/a&gt;
&lt;/td&gt;
&lt;td&gt;Apache-2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Public Sans font&lt;/td&gt;
&lt;td&gt;&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/site/fonts" rel="noopener noreferrer"&gt;&lt;code&gt;site/fonts/&lt;/code&gt;&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;&lt;a href="https://github.com/ting-hong-shieh/monster-telephone/site/fonts/OFL.txt" rel="noopener noreferrer"&gt;OFL&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;The server calls &lt;a href="https://tinker-docs.thinkingmachines.ai/" rel="nofollow noopener noreferrer"&gt;Tinker&lt;/a&gt;, a hosted service, so the live game and training need a Tinker API key. I have not run the weights outside Tinker yet.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;How it works&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;The ghosts run on…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/ting-hong-shieh/monster-telephone" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;h3&gt;
  
  
  How one round runs
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;you whisper a monster
  -&amp;gt; Pip draws it (fine-tuned Qwen3.5-4B)
  -&amp;gt; Dot looks at the picture and guesses (Qwen3.5-4B)
  -&amp;gt; you draw it back, or a ghost takes your turn
  -&amp;gt; Moth guesses, Hush draws, Wisp guesses
  -&amp;gt; you see your monster beside what it became
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The five ghosts share two models: the fine-tuned artist draws, and plain Qwen3.5-4B guesses.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb6yldgu0ylbd677yo8k9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb6yldgu0ylbd677yo8k9.png" alt="How the project runs in three columns. Train on Tinker: the teacher draws, a strict check keeps 955, then the fine-tune. Serve on Render: the FastAPI server, drawing to picture, two ghosts on Tinker. Play on the web: the domain, the static page, the recorded rounds." width="800" height="372"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  The game had a hole
&lt;/h3&gt;

&lt;p&gt;My first design asked the player to guess and bet along the way. While reading it through, I asked why I could not just insist on my own drawing, since I was the one who set the question. The player already knows the answer, so the guessing was fake.&lt;/p&gt;

&lt;p&gt;That question changed the round. Drawing from memory became the player's turn. Betting and the score both went.&lt;/p&gt;

&lt;h3&gt;
  
  
  The small model could not finish a drawing
&lt;/h3&gt;

&lt;p&gt;I wanted the smallest model that could do the job, but Qwen3.5-4B often failed to finish a clean line drawing. So I had a larger model teach it.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Qwen3.6-35B drew 1,200 monsters as SVG.&lt;/li&gt;
&lt;li&gt;I kept the 955 that passed a strict check.&lt;/li&gt;
&lt;li&gt;I fine-tuned Qwen3.5-4B on them with Tinker: a rank 64 LoRA, 3 epochs, 180 steps. I later retrained it the same way so the weights could be exported.&lt;/li&gt;
&lt;li&gt;I tested all three models on 60 prompts about six creatures that were not in training, two samples each.&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Valid drawings&lt;/th&gt;
&lt;th&gt;Mean tokens&lt;/th&gt;
&lt;th&gt;Mean seconds&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Qwen3.5-4B&lt;/td&gt;
&lt;td&gt;54 / 120&lt;/td&gt;
&lt;td&gt;765&lt;/td&gt;
&lt;td&gt;4.2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen3.5-4B, fine-tuned&lt;/td&gt;
&lt;td&gt;117 / 120 (118 / 120 on the retrain)&lt;/td&gt;
&lt;td&gt;591&lt;/td&gt;
&lt;td&gt;4.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen3.6-35B&lt;/td&gt;
&lt;td&gt;92 / 120&lt;/td&gt;
&lt;td&gt;830&lt;/td&gt;
&lt;td&gt;5.4&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Valid means the reply parses, has 3 to 14 unfilled paths, stays inside the canvas, and finishes within 1,100 tokens. Half of the small model's failures were replies that hit that limit. For a live game this matters: a player can't wait for a drawing that never ends, and the fine-tuned model almost always ended.&lt;/p&gt;

&lt;p&gt;Here are the three models on the same four prompts: the first prompt for each of the first four creatures in the test, first sample, from the first training run.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foqe6v7wijkkicclra5hh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foqe6v7wijkkicclra5hh.png" alt="Four test prompts across three models. The base 4B model passed on two and failed the check on two. The fine-tuned model and the 35B teacher passed on all four." width="800" height="594"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Then I ran a review on my first draft of this post. It pointed out that the check only looks at the file format. The student had beaten its teacher on a test that never asks what the drawing shows.&lt;/p&gt;
&lt;h3&gt;
  
  
  The blind test
&lt;/h3&gt;

&lt;p&gt;I rendered 89 drawings from the first training run and made a page that hid the model and the prompt. Then I tried to name each creature from six choices. Picking at random would be right 1 time in 6.&lt;/p&gt;

&lt;p&gt;You can try it too. These six came from the large model, one for each creature: crab, goat, lantern, moth, radio, witch. I took the first of each creature in the test file.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm2x1ypr9v90zod25hzo0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm2x1ypr9v90zod25hzo0.png" alt="Six line drawings labelled A to F, for the reader to name." width="800" height="562"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;/p&gt;
  My answers, and the real ones
  &lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Drawing&lt;/th&gt;
&lt;th&gt;The prompt&lt;/th&gt;
&lt;th&gt;I said&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;A&lt;/td&gt;
&lt;td&gt;a radio with a mouth on its belly dancing&lt;/td&gt;
&lt;td&gt;radio&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;B&lt;/td&gt;
&lt;td&gt;a moth with long thin arms waving hello&lt;/td&gt;
&lt;td&gt;moth&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;C&lt;/td&gt;
&lt;td&gt;a witch with stitches looking very polite&lt;/td&gt;
&lt;td&gt;witch&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;D&lt;/td&gt;
&lt;td&gt;a goat with a forked tongue&lt;/td&gt;
&lt;td&gt;crab&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;E&lt;/td&gt;
&lt;td&gt;a lantern with long thin arms grinning&lt;/td&gt;
&lt;td&gt;lantern&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;F&lt;/td&gt;
&lt;td&gt;a crab with one enormous eye&lt;/td&gt;
&lt;td&gt;goat&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Four of six flatters me. Over all 30 drawings from this model I named 7.&lt;/p&gt;



&lt;br&gt;
&lt;p&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foscfm3oattf7835xnc4j.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foscfm3oattf7835xnc4j.png" alt="Two bar charts over the same three models, each row with one crab drawing. Finished a drawing: Qwen3.5-4B 54 of 120, the fine-tuned model 117 of 120, Qwen3.6-35B 92 of 120. I named the creature: Qwen3.5-4B 1 of 29, the fine-tuned model 2 of 30, Qwen3.6-35B 7 of 30." width="800" height="480"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I answered "can't tell" for 66 of the 89. By the rule I wrote before the test, the gap between student and teacher is inconclusive. What I can say is smaller: the fine-tune taught the model to finish a drawing. I could not see that it taught the model to draw a goat.&lt;/p&gt;

&lt;p&gt;I checked this another way, with no human in it. For each drawing I compared pixels to find the closest drawing from a different prompt. It showed the same creature 27% of the time for the large model, 18% for the first fine-tune and 22% for the public retrain, with chance at about 17%. I chose this check after the blind test, so it was not in my rules.&lt;/p&gt;

&lt;h3&gt;
  
  
  The ghosts could not read
&lt;/h3&gt;

&lt;p&gt;The listener, the ghost that guesses, first read each drawing as SVG text. I sent it 18 drawings from the large model, three guesses each. It named the right creature 0 times out of 54. I drew one of each creature myself: 0 out of 18.&lt;/p&gt;

&lt;p&gt;So the rescue did nothing. A player's drawing went into the chain and the ghost could not see it.&lt;/p&gt;

&lt;p&gt;Qwen3.5-4B can also read images, so I rendered the same drawings as pictures and ran the same test with the same rules.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Drawings&lt;/th&gt;
&lt;th&gt;Ghost reads SVG text&lt;/th&gt;
&lt;th&gt;Ghost reads a picture&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Large model's&lt;/td&gt;
&lt;td&gt;0 / 54&lt;/td&gt;
&lt;td&gt;0 / 54&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Mine&lt;/td&gt;
&lt;td&gt;0 / 18&lt;/td&gt;
&lt;td&gt;4 / 18&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4n9mo6mcdeytlrlbyjio.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4n9mo6mcdeytlrlbyjio.png" alt="My six drawings: crab, moth, lantern, witch, goat and radio. Beside each is the first guess from the ghost reading SVG text and the first from the ghost reading a picture. The crab was named 3 of 3 times, the moth 1 of 3, the others not at all." width="800" height="932"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;In the picture, each row shows the first of three guesses. The table counts all three. Three of the four hits came from one crab and the fourth from a moth called a butterfly. The rule used a fixed word list, so near misses scored nothing: "a lightbulb head" for the lantern, a hat twice for the witch, horns for the goat, a square head or body twice for the radio. Six drawings by one person who can't draw are too few to give a rate. But it changed the game, and the live server now shows the ghosts pictures. In the recorded crab chain, my crab lasted one turn: the ghost named it, drew it, and the next ghost saw a spider. In the eight recorded rounds with a ghost in my seat, none of the 24 guesses named the original monster.&lt;/p&gt;

&lt;h3&gt;
  
  
  Making the wait feel like a game
&lt;/h3&gt;

&lt;p&gt;A live round once took 31.4 seconds from pressing Done to the reveal on a desktop, and 27.5 on a phone. The ghosts were waiting for animations to finish before they called the model. I moved each call into the animation before it, and the page now sends your drawing the moment you press Done. It now takes 21.0 seconds on a desktop and 20.7 on a phone.&lt;/p&gt;

&lt;h3&gt;
  
  
  Where it runs
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Tinker&lt;/strong&gt; trained the artist and serves both ghosts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Render&lt;/strong&gt; hosts the FastAPI server on the free plan. It keeps the Tinker key, turns each drawing into a picture for the listener, and limits use to 30 requests a minute and 2,000,000 tokens a day. After a sleep, it took 62.7 seconds to answer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;DigitalOcean App Platform&lt;/strong&gt; hosts the static page for free, and DigitalOcean DNS runs the domain.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;.tech domain:&lt;/strong&gt; monstertelephone.tech is the game's address.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;ElevenLabs&lt;/strong&gt; made the voice, the music and the sound effects in the demo film. The game itself does not use it.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  The first idea failed
&lt;/h3&gt;

&lt;p&gt;I started on something else: a small version of &lt;a href="https://www.nature.com/articles/s41586-026-10319-8" rel="noopener noreferrer"&gt;a 2026 Nature paper&lt;/a&gt; in which a model that loved owls wrote lists of numbers, and another model trained on those numbers came to love owls too. On Tinker with Qwen3.5-4B, my pilot gave 21 of 300 answers naming the owl against 10 of 300 for the control, and the replay gave 41 of 600 against 39 of 600. The rule I set before the replay said to stop, so I stopped. My run was small, so this says nothing against the paper.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;When the drawings failed, I could change the weights of a small model for one narrow job with 955 examples. Both Qwen models are Apache-2.0, so I could train one on the other's drawings and publish those drawings.&lt;/p&gt;

&lt;p&gt;OpenAI's fine-tuning, the closed route for that fix, is ending. Its docs say OpenAI is &lt;a href="https://developers.openai.com/api/docs/guides/vision-fine-tuning" rel="noopener noreferrer"&gt;"winding down the fine-tuning platform"&lt;/a&gt; and that the platform is "no longer accessible to new users." A new account could not train this artist there today.&lt;/p&gt;

&lt;p&gt;The artist is a file. The rank 64 LoRA is in the &lt;a href="https://github.com/ting-hong-shieh/monster-telephone/releases/tag/artist-2" rel="noopener noreferrer"&gt;artist-2 release&lt;/a&gt;, and the live game runs that same checkpoint. Tinker only lets you export a training checkpoint, so I retrained to get one. I have not run the weights outside Tinker yet.&lt;/p&gt;

&lt;p&gt;The training drawings and all test results are in the repository, including my blind answers and every guess.&lt;/p&gt;

&lt;p&gt;I have used $16.66 of Tinker credit: I started with $20 and the billing page now shows $3.34 left.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Tinker:&lt;/strong&gt; The category asks for a fine-tune that beats a baseline. On 60 prompts about creatures outside the training set, the same 4B model went from 54 of 120 valid drawings to 117 of 120, and to 118 on the retrain that became the public weights. It passed the 35B model that taught it (92 of 120). Tinker ran the whole loop, from the teacher drawings to both ghosts in the live game. The validity test showed that the fine-tune finishes drawings. My blind test did not show that the drawings got better.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; Render is the game's AI runtime. Every live turn goes through the FastAPI server, which keeps the Tinker key, turns the player's drawing into a picture the ghost can read, and enforces the rate and token limits. It runs on the free plan, so it sleeps, and the page tells the player that the ghosts are waking up.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of DigitalOcean:&lt;/strong&gt; App Platform hosts the whole game page for free, and DigitalOcean DNS runs the domain. The recorded rounds ship with the page and need no server, so a visitor can watch the crab chain while the live server is still asleep.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of a .tech Domain:&lt;/strong&gt; I claimed monstertelephone.tech through Hacktoberfest, and it is the game's only address. It is where the live game runs, and it is the last thing on screen in the demo film.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs:&lt;/strong&gt; The category suggests narration for a demo, and I went further. ElevenLabs voices say the five narration lines in the film. The march, the tape rewind, the pencil, the paper and the small ghost cheer all came from its sound generator. The music breaks when the ghosts lose the crab.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How this was made
&lt;/h2&gt;

&lt;p&gt;Claude Code wrote the Python, ran the training and tests, and wrote the first draft of this post from my decisions. An AI stand-in drafted the "fun, cute and surprising" lines near the top and "Why Does Open Innovation Matter?" from my notes, and I read them before publishing. Codex built the page and, when I asked, reviewed the game design as a producer would. I chose the game rules, approved the tests, did the blind test, and drew the crab. Gemini 4 Argon reviewed the post line by line, and I approved the edits I kept.&lt;/p&gt;

&lt;p&gt;For the film, Claude Code wrote the Blender stage, the web animation and the mix as code. The ghosts' table is staged. Every drawing and guess on it comes from a recorded round, and the crab is my real drawing, stroke by stroke.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxjd2znr8y4jlyo2m9n93.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxjd2znr8y4jlyo2m9n93.jpg" alt="A still from the film: Dot, a clay-like ghost with round glasses, peers over a sheet of paper holding Pip's drawing." width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;What would you whisper to the ghosts? Tell me in the comments what your monster turned into.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>ai</category>
    </item>
    <item>
      <title>My alarm won't let me snooze if I'd be late</title>
      <dc:creator>Elias Shieh</dc:creator>
      <pubDate>Thu, 22 Oct 2026 20:09:39 +0000</pubDate>
      <link>https://dev.to/elias-shieh/my-alarm-wont-let-me-snooze-if-id-be-late-42p0</link>
      <guid>https://dev.to/elias-shieh/my-alarm-wont-let-me-snooze-if-id-be-late-42p0</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Every morning I want to sleep a little longer, but I'm scared to keep pressing snooze.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/HcZXvemBTq8" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;Snooze Oracle uses past Uber trips to work out when I need to get up for a meeting. It finds the latest time I can get up for which the model estimates at least a 90% chance of being on time. The alarm rings 30 minutes before that, and I can snooze up to 10 minutes at a time until I reach it.&lt;/p&gt;

&lt;p&gt;Right now it is a web replay and a command-line planner that writes the alarms to a calendar file. There is no phone app yet, and I don't live in New York, so I haven't used it on a real morning.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Try to snooze it yourself: &lt;a href="https://snooze.eliasshieh.com" rel="noopener noreferrer"&gt;snooze.eliasshieh.com&lt;/a&gt;&lt;/strong&gt; (works in a phone browser)&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;If you have to be somewhere at the same time every weekday, you probably set the same alarm every day. But the trip doesn't take the same time every day. On the route in my demo, a weekday Uber from University Heights in the Bronx to Midtown, requested between 7 and 9am, took 28 to 75 minutes from request to drop-off in July and August. So when the alarm rings, you don't know how long you can stay in bed.&lt;/p&gt;

&lt;p&gt;In the replay, the meeting is at 9:00 in Midtown. The alarm rings at 07:14. You snooze to 07:24, 07:34 and 07:44, and then it says no: one more minute would drop the estimated chance of being on time to 89%. So I can keep snoozing without doing the math in my head, and I don't have to check the map the night before.&lt;/p&gt;

&lt;p&gt;It doesn't know about tomorrow's accident or rain. It learns from past trips how long your route usually takes on a weekday morning at that hour, and how much that changes. So the wake-up time depends on the day of the week and the meeting time.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;The demo page replays one day of August 2026 data. Pick a commute or an airport pickup, press Start and keep pressing Snooze. When another snooze would put the estimate below 90%, the button says "No safe snooze" and the page shows the number.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fah72mrctslg7ill8mlch.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fah72mrctslg7ill8mlch.png" alt="Four states of the alarm: 07:14 first ring, snooze to 07:24, snooze to 07:34, and at 07:44 no safe snooze, 89%" width="800" height="717"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The page is static. The curves were computed with TabPFN on my machine and saved into the page, so nothing on it calls a model.&lt;/p&gt;

&lt;p&gt;The video's narration is ElevenLabs text-to-speech, the music was made with ACE-Step 1.5, and the video's 3D scenes are three.js.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/ting-hong-shieh" rel="noopener noreferrer"&gt;
        ting-hong-shieh
      &lt;/a&gt; / &lt;a href="https://github.com/ting-hong-shieh/snooze-oracle" rel="noopener noreferrer"&gt;
        snooze-oracle
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      An alarm that lets you snooze only while you'll still make it. TabPFN-3.5 on NYC Uber trips.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Snooze Oracle&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/ting-hong-shieh/snooze-oracle/docs/img/hero.jpg"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fting-hong-shieh%2Fsnooze-oracle%2FHEAD%2Fdocs%2Fimg%2Fhero.jpg" alt="A cartoon sleeper on the edge of the bed, next to a tablet showing the Snooze Oracle replay"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a href="https://github.com/ting-hong-shieh/snooze-oracle/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/f74c8d3527c7a094930d214c08c064d32928cb177a0c261f8b861263336fd587/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f636f64652d4170616368652d2d322e302d626c7565" alt="License: Apache-2.0"&gt;&lt;/a&gt;
&lt;a href="https://snooze.eliasshieh.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/34bb79078f06dadbb592a8670b7d5e763588b4fd8dd765db3ba24c71ae9bf03e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f64656d6f2d736e6f6f7a652e656c69617373686965682e636f6d2d663663393662" alt="Demo"&gt;&lt;/a&gt;
&lt;a href="https://youtu.be/HcZXvemBTq8" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/223b38f4e40fe37d2c368dfff4827af1252f45ac7cb68777bff3022ce6365789/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f766964656f2d313030253230732d666638613661" alt="Video"&gt;&lt;/a&gt;
&lt;a href="https://dev.to/elias-shieh/my-alarm-wont-let-me-snooze-if-id-be-late-42p0" rel="nofollow"&gt;&lt;img src="https://camo.githubusercontent.com/3836232d4be9cdf2767799ac2e4de78e8b4370bfbb9f7d4df45b7135c3623228/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4445562d77726974652d2d75702d306130613061" alt="DEV post"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;An alarm that works out, the night before, the latest time you can get up and still be on
time with an estimated 90% probability. It rings 30 minutes earlier and allows snoozes of up
to 10 minutes until that cutoff, then refuses.&lt;/p&gt;
&lt;p&gt;Two mornings are covered:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Commute:&lt;/strong&gt; an Uber to a meeting in Midtown, the Financial District or Hudson Yards.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pickup:&lt;/strong&gt; meeting someone off an early flight (landing 03:00–08:59) at JFK or LaGuardia.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Ride times come from TabPFN-3.5 run locally on NYC TLC Uber trips; flight delays from
TabPFN-3.5 on BTS on-time data.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://snooze.eliasshieh.com" rel="nofollow noopener noreferrer"&gt;https://snooze.eliasshieh.com&lt;/a&gt; replays six August 2026 mornings (four commutes
two airport pickups) with a simulated alarm clock and a map of each route. The page is
static: the curves were precomputed and nothing on it calls a model.&lt;/p&gt;
&lt;p&gt;Built for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19" rel="nofollow"&gt;DEV Hacktoberfest Open-Source AI Challenge, Week 3&lt;/a&gt;
The write-up is on …&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/ting-hong-shieh/snooze-oracle" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Data.&lt;/strong&gt; NYC TLC trip records for about 14 million Uber trips a month (July and August 2026), and BTS on-time records for flights into JFK and LGA. The models only get July trips (for flights, the 12 months before) and are tested on August.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why TabPFN.&lt;/strong&gt; It did best where data is thin. On airport routes with fewer than 30 trips in July, I gave TabPFN, LightGBM and a lookup table exactly the same rows. TabPFN's mean pinball loss (lower is better) was 2.49 minutes, against 2.78 for LightGBM and 2.97 for the lookup. On busy routes it tied LightGBM.&lt;/p&gt;

&lt;p&gt;I don't train or fine-tune TabPFN. I use the pretrained model as it is. For each ride prediction, I give it up to 10,000 July trips as examples. I start with trips on the same route around the requested hour, then add trips from the same borough to the same destination. It predicts a range of trip times and how likely each one is, and the planner uses that to compare possible wake-up times.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The alarm.&lt;/strong&gt; For each possible wake-up time, it simulates 2,000 mornings: getting ready (I assume about 15 minutes) and the ride. For an airport pickup, I compare my arrival with the flight's predicted arrival plus 20 minutes for the passenger to get out. The target is the latest wake-up time where at least 90% of those mornings are on time.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu6y00hnf7svt33kpo6nu.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu6y00hnf7svt33kpo6nu.png" alt="Chance of being on time by wake-up time for the demo morning: it stays near 100% until about 07:35, crosses 90% at 07:44, and is 51% at 07:57, when the usual ride plus 10 minutes would wake you" width="800" height="637"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The test.&lt;/strong&gt; Every strategy picks one wake-up time the night before and is scored against August trips. The fixed buffer is the simple way: take the meeting time, then subtract 15 minutes to get ready, the July median trip time, and a buffer. The other three are the same alarm built on a lookup table, on LightGBM and on TabPFN.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiev9vrgirwoyhy2i8xdx.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiev9vrgirwoyhy2i8xdx.png" alt="Late mornings in the August replay: all 12 routes, 10-minute buffer 8.3%, 13-minute buffer 5.1%, TabPFN alarm 5.1%; the demo route, 22.9%, 17.1% and 12.7%" width="800" height="655"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Across all 12 commute routes, the alarm was late 5.1% of the time and the 10-minute buffer 8.3%. But you sleep a little less. On average the alarm got you up 2.9 minutes earlier than the 10-minute buffer.&lt;/p&gt;

&lt;p&gt;A 13-minute buffer matched the alarm's overall late rate at about the same average wake-up time, but I could only pick 13 after seeing the August results. On the University Heights route, the 13-minute buffer was late 17.1% of the time. The TabPFN alarm was late 12.7%, which is better but still above my 10% target, and it got you up 7.1 minutes earlier than the 13-minute buffer on average (small sample, see the details below).&lt;/p&gt;

&lt;p&gt;The full reports are in the repo: &lt;a href="https://github.com/ting-hong-shieh/snooze-oracle/blob/main/docs/commute_backtest.md" rel="noopener noreferrer"&gt;commute backtest&lt;/a&gt;, &lt;a href="https://github.com/ting-hong-shieh/snooze-oracle/blob/main/docs/backtest.md" rel="noopener noreferrer"&gt;airport backtest&lt;/a&gt; and &lt;a href="https://github.com/ting-hong-shieh/snooze-oracle/blob/main/src/ride/RESULTS_round3.md" rel="noopener noreferrer"&gt;ride model comparison&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;I also tested airport pickups, where both the drive and the flight's arrival are uncertain. Here the TabPFN alarm did worse than the lookup alarm: late in 10.3% of simulations, versus 9.1%. It gave some flights too much extra sleep, so I need to improve those predictions before I would trust it for a real pickup.&lt;/p&gt;

&lt;p&gt;&lt;/p&gt;
  All the numbers
  &lt;p&gt;Share of mornings late:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Airport&lt;/th&gt;
&lt;th&gt;Commute&lt;/th&gt;
&lt;th&gt;Demo route&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Fixed buffer*&lt;/td&gt;
&lt;td&gt;11.1%&lt;/td&gt;
&lt;td&gt;8.3%&lt;/td&gt;
&lt;td&gt;22.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Lookup alarm&lt;/td&gt;
&lt;td&gt;9.1%&lt;/td&gt;
&lt;td&gt;6.1%&lt;/td&gt;
&lt;td&gt;21.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;LightGBM alarm&lt;/td&gt;
&lt;td&gt;11.7%&lt;/td&gt;
&lt;td&gt;6.6%&lt;/td&gt;
&lt;td&gt;14.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TabPFN alarm&lt;/td&gt;
&lt;td&gt;10.3%&lt;/td&gt;
&lt;td&gt;5.1%&lt;/td&gt;
&lt;td&gt;12.7%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;*30 minutes at the airport, 10 minutes for the commute. A 13-minute commute buffer was late 5.1% overall and 17.1% on the demo route (University Heights to Midtown Center).&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Airport: 1,575 completed flights scheduled to arrive from 4:00 to 7:59, each with 8 pickup zones: 12,600 scenarios. Planning for the scheduled time and ignoring delays was late 46% of the time.&lt;/li&gt;
&lt;li&gt;Airport, by flight: on the 4.3% of scenarios where the alarm woke you 10 or more minutes earlier than a 30-minute buffer, mostly red-eyes from the West Coast and Hawaii, the buffer was late 17.8% of the time and the alarm 3.8%. On the 2.0% where it let you sleep 10 or more minutes longer, mostly short 7am flights into LaGuardia, the alarm was late 16.9% and the buffer 5.6%.&lt;/li&gt;
&lt;li&gt;Commute: 12 routes into Midtown, the Financial District and Hudson Yards, meetings from 8:00 to 10:00, 21 weekdays: 1,260 mornings, of which 1,070 had enough August trips to score.&lt;/li&gt;
&lt;li&gt;University Heights route: 63 simulated mornings with meetings at 8:00, 9:30 and 10:00; the demo's 9:00 meeting had no August mornings I could score. The route has few trips, so its ride samples come from any August weekday, and they don't measure each day's traffic.&lt;/li&gt;
&lt;li&gt;Flights alone: pinball loss TabPFN 10.38, lookup 10.56, LightGBM 10.68.&lt;/li&gt;
&lt;li&gt;I tested commute buffers from 0 to 30 minutes and airport buffers from 0 to 90 minutes on August data. August trips were never in the models' training rows or context, but I used August results to compare buffers and to choose the flight-model variant.&lt;/li&gt;
&lt;li&gt;A feature I dropped: I looked for a reason to wait 10 minutes before requesting the Uber and didn't find one. On the busiest early-morning airport routes, the typical fare changed by $0.80–2.40 between neighbouring 10-minute slots, while the spread of fares inside one slot was $12–34.&lt;/li&gt;
&lt;li&gt;The backtest scored wake-up times. Its snooze simulation used 9-minute snoozes; the demo now uses up to 10.&lt;/li&gt;
&lt;/ul&gt;



&lt;p&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;I can keep the planning local.&lt;/strong&gt; TabPFN runs on my computer, and the planner reads saved predictions. Picking a pickup area or a flight doesn't mean sending my plans to a model provider.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I could test without API fees.&lt;/strong&gt; In my third round of ride tests, TabPFN ran on 2,914 route-and-hour groups, 59,321 trips in total. That took about two hours on my GPU.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Other people can check my choices.&lt;/strong&gt; The repo has the code that downloads the public NYC TLC and BTS data, compares the models and tests the alarm. You can change my assumptions, like 15 minutes to get ready or 20 minutes for bags, and rerun the tests.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The code is Apache-2.0. TabPFN-3.5's weights have a separate non-commercial license, so rebuilding the predictions needs a Prior Labs login, that license, the weights and a GPU. The public demo only replays saved results.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Best Use of TabPFN&lt;/li&gt;
&lt;li&gt;Best Use of ElevenLabs: the narration in the demo video is ElevenLabs text-to-speech.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;How this was made:&lt;/strong&gt; I built this with AI coding agents (Claude Code, and Codex for reviews). I chose the idea, the rules and what to test. The agents wrote most of the code, the video and the first drafts of this post. Every number comes from runs on my own GPU.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>machinelearning</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
