Your Phone Is the Worst Part of a Walk. So I Built an AI That Hides It.

A developer built Look First, a browser-based AI naturalist that refuses to identify plants and birds until the user completes three short observation tasks away from the screen, then responds with a best guess, a confidence level and a stated uncertainty. The tool runs entirely client-side using Google's open-weight Gemma model (gemma-2-2b-it, 4-bit quantized) via WebLLM and WebGPU, with no server or API key, and was submitted to the DEV Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass. The developer reports a 15-minute field test on a tree, a small yellow flower and an unseen bird, and notes limits including text-only input, a WebGPU browser requirement and a large one-time model download.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass https://dev.to/challenges/hacktoberfest-week1-2026-10-05 Every AI product I know measures success the same way: how long you stay. Sessions, minutes, daily actives. For a challenge about getting people off the screen, that metric is exactly backwards. So I built Look First, an AI naturalist with one unusual goal: to be used for as little time as possible. Here is how a session goes: It refuses to name anything. Instead, it gives you three tiny tasks. Look at the edge of a leaf. Touch the bark. Listen for a repeating call. The screen goes dark. A phone-away timer takes over the whole page: "Eyes up. I'll be here when you get back, and I'm happiest when you don't need me." Two minutes. You go and look. You come back and report what you noticed. Only now does it answer, with a best guess, a confidence level low, medium or high , one honest thing it might be getting wrong, and one tiny next quest nearby. It tells you how long you spent outside. "You spent 2:00 looking up instead of scrolling." That number goes into a downloadable markdown field journal. Most identifiers turn a living thing into a search result. Look First tries to turn you into someone who notices. By the third time you do the tasks, you may not need the answer at all, which is the point. It is for anyone who has walked past the same tree every day and never learned what it was: students on campus, new hikers, a parent with a curious kid. https://hiyalukka.github.io/Look-first/ https://hiyalukka.github.io/Look-first/ The first visit downloads the model once, then it runs on your device. Yes, I know the irony of a slow first load in a project about leaving the screen. An AI naturalist that makes you look up before it answers. Most plant and bird identifiers keep you staring at a screen. Look First does the opposite: it refuses to name anything until you've actually looked , then reveals what you found. It runs entirely in your browser using an open-weight model Gemma , with no server and no API key. Built for the DEV Hacktoberfest Open-Source AI Challenge, Week 1: Touch Grass . Every AI product measures success by how long you stay. Look First measures success by how little you use it. The open pieces do all the work: Gemma gemma-2-2b-it, 4-bit quantized , an open-weight model from Google. WebLLM, open-source in-browser inference that fetches the model into the browser and runs it on the device GPU through WebGPU. Plain HTML, CSS and JavaScript on top. The idea lives in the prompt, not the code. The model gets a strict two-turn script. On turn one it is forbidden from identifying anything and must give exactly three short observation tasks. On turn two it uses my observations to guess, and it must state a confidence level and what it might be wrong about. That second rule came from a real worry. A 2B model can sound completely sure while inventing a species. Instead of hiding that weakness, I built it into the product: every answer admits its uncertainty. Because the person has just examined the real thing, they are a better judge of the answer than the model is. Why browser-only? Honestly, it started as a workaround. I did not have Ollama installed and the deadline was that same day. But the constraint became a feature: anyone can open a link and run the AI, with nothing to install. Known limits: It is text-only. You describe what you see, and it does not look at photos yet. It needs a WebGPU browser recent Chrome or Edge . The first load is a large one-time download. A small model will sometimes be wrong. Check a local field guide. My Field Test I took Look First to the garden behind my campus library for about 15 minutes, and tried it on three things: a big tree, a small yellow flower and a bird I couldn’t see. What it got right. For the tree, it asked me to look at the leaf edge and feel the bark. I typed “leaves smooth-edged, bark grey and peeling in flakes.” It guessed a plane tree with medium confidence and said the peeling bark was the main clue. Checking later, that matched. What I noticed was that I’d walked past that tree for a year and never touched the bark. Where it struggled. For the bird, I only had a sound to describe, and I wrote “a repeating chirp, then a short whistle.” The AI gave a guess with low confidence and added that sound alone wasn’t enough to be sure. It was honest, but not very useful. I did what I’d do with a person: I waited, spotted the bird on a branch, and went back to the app with a better description. The second guess was more specific. The dark screen. The “Eyes up” timer felt a bit awkward for the first ten seconds, since my hand kept reaching for the phone. By the second minute, I’d noticed that the flower had a fuzzy stem, which I wouldn’t have seen otherwise. A product that wants you to leave is hard to build on a closed API, because a closed API is a meter that runs while you use it. Look First is the opposite of that business model, and open weights are what make it possible. It works where the trail is. Once the model is cached, there is no connection to lose. A hosted API fails exactly where signal does, which is outdoors. Your location stays yours. What you type says where you are and what is around you. Here it never leaves your device. There is no meter. No API key, no bill that grows when more people use it, so a student can hand the link to a whole campus club. The model is swappable. The page lists the Gemma models it finds and lets you choose, so a lighter one can run on a weaker laptop. Open weights did not make the model smarter than a frontier model. They made it possible to build something with no incentive to keep you looking at it. Build with Gemma category