JonsTube Lab

Use open source vision models on yt-dlp frames to measure the glass, bottle, sip count and setting in every water video and publish the dataset as an explorable lab.

Object detection ran on every water thumbnail (not video frames, which YouTube blocks from the build box), so glass and bottle presence is measured but sip counts are not.

Cluster every thumbnail with open CLIP embeddings into a zoomable 2D map where similar looking videos sit near each other.

Uses open source MobileNetV3 image features instead of CLIP (lighter for the build machine); all thumbnails covered.

Build a radial chart per year showing the dominant colors of every video posted that year.

Render a full GitHub style calendar of every upload since 2011 with click through to each day.

Publish the full video catalog as JSON, CSV and SQLite files with a documented schema for anyone to remix.

Ship a free Vercel serverless API with endpoints for videos, channels, numbers and random picks.

Model upload pace to project when Jon Counts to 100,000 and Jon Drinks Water will hit future milestones.

Run a nightly free GitHub Action that checks every embed and flags videos that went private or vanished.

Live checker and nightly sample are real; the scheduled job is written but not running because the site deploys without a connected repo.

Analyze every title for word frequency, length trends and naming patterns across all channels.

Generate a visual profile for each channel showing length mix, posting rhythm, colors and era.

Add a chat box where a free open source model answers questions about the catalog using site metadata only.

Answers come from a rule based engine over site metadata, not a language model, so nothing is invented.

Let viewers draw a rough shape or color and find thumbnails that match using CLIP in the browser.

Matches a painted 6 by 4 color layout instead of CLIP embeddings.

Let viewers drop in a photo and find the most visually similar JonsTube videos on device.

On device color layout matching instead of CLIP.

Run open source object detection on frames so viewers can browse every video that features tape, a carrot, a phone and more.

COCO detector on thumbnails; COCO has no tape class.

Detect visual scene cuts with ffmpeg and let viewers skip between them on longer videos.

Covers the first 30 to 60 seconds of the 13 Tech Thursday videos available on the build box; YouTube blocks downloads of the rest.

Use Whisper and a small open model to write a short recap card for each spoken video.

Extractive recaps from Whisper transcripts of the first 30 to 60 seconds of 13 Tech Thursday videos; no generative model.

Index Whisper output so viewers can search for a phrase Jon actually said and jump to that moment.

Whisper tiny.en transcripts of the first 30 to 60 seconds of 13 Tech Thursday videos.

Translate every title and description into a dozen languages with open translation models and a language switcher.

Translation runs on demand in the browser (Opus MT via transformers.js) rather than precomputed for every title.

Classify videos by visual tone so viewers can browse by calm, bright, cozy or busy.

Use audio energy and scene analysis to suggest highlight moments shown as marked hotspots on long videos.

Audio energy hotspots for the same 13 video subset.

Create a full screen lean back mode with numbered channels and a remote style dial that flips between live feeling streams.

Run a deterministic 24 hour programming schedule so every visitor sees the same video at the same moment like real TV.

Show a scrolling program guide for the broadcast channel with time slots for the next 24 hours.

Play up to nine muted videos at once in a grid, with click to promote one to sound.

Watch two videos synchronized next to each other, such as water video 1 and water video 3000.

Auto pair an early video with a recent one from the same channel and show a draggable split view.

Turn idle screens into a slow rotating gallery of thumbnails and silent water clips.

Thumbnails only; no water clips, to avoid rehosting video.

Let viewers pick any past date and see the homepage exactly as the catalog stood that day.

Let viewers say how much time they have and auto build a perfect fit lineup to fill it.

Build a calm wind down sequence that fades and stops after a chosen time.

Build a game where viewers guess which of two videos has more YouTube views.

Slowly unblur a thumbnail and score viewers on how fast they name the channel.

Show a random frame and challenge viewers to guess the upload year.

Uses thumbnails instead of random frames.

Ship a Wordle style daily puzzle built from titles and numbers with a shareable result grid.

Create a playful canvas sandbox where viewers tape virtual objects together.

Turn carrot audio into a rhythm tapping game generated from real crunch timings.

Crunch timings could not be extracted (YouTube blocked audio downloads), so players record their own chart against a real video, then replay it.

Hide clues across pages that lead viewers to specific numbered videos for a final reveal.

Turn any thumbnail into a drag and drop jigsaw puzzle with selectable difficulty.

Flash thumbnails rapidly and have viewers sort them into the right channel against a clock.

Stitch existing clips into branching adventures where viewers pick what happens next.

Let viewers pick moments from many videos and play them back as a seamless embed sequence they can share by link.

Let viewers caption any thumbnail with templates and download the image locally.

Precompute short open licensed preview loops and let viewers trim and export GIFs of their favorite moments.

Frames are the thumbnail plus YouTube auto still frames; no trimmed video loops.

Assemble thousands of water thumbnails into a photo mosaic of any image a viewer uploads in the browser.

Build a board of short clips of signature sounds like crunches, sips and tape tears.

Pads play short real clips in the official embed; sip and crunch moments are not auto detected.

Create printable posters for any channel or milestone with auto layout from real thumbnails.

Generate phone and desktop wallpapers from channel art and thumbnails in any resolution.

Let viewers design retro cassette style artwork for their own lists and export it.

Use ffmpeg on the box to render an annual montage of the year's uploads posted as a JonsTube feature.

Montages for 2024 and 2025 rendered from thumbnails, not footage.

Generate fun poems built entirely from real video titles with a remix button.

Present a guided narrated tour of the channel history using open source text to speech over archive clips.

Narrated with browser speech synthesis, not a cloned voice

Write long form illustrated pages explaining how each channel began, sourced from public descriptions.

Divide the archive into named eras with a cinematic scrolling story page for each.

Build a virtual gallery room where each milestone video is a framed exhibit with a plaque.

Generate a printable PDF zine every week covering new uploads, fun facts and puzzles.

Weekly issue is generated at build time; no automatic weekly schedule

Show a rotating trivia fact computed from the real catalog each time someone arrives.

Publish scrollytelling articles like how long it would take to watch everything with live charts.

Create a wiki style reference for recurring objects, places and running bits across videos.

Give each round number milestone its own page with context and neighboring videos.

Visualize every link between channels as a network graph of shared topics and themes.

Render every video as a star in a Three.js galaxy where channels form constellations.

Build a 3D room with a TV, couch and shelves of tapes where viewers browse by picking up cassettes.

Offer a free WebXR mode that places the player on a giant screen in a virtual cinema.

WebXR theater shows posters and opens the official embed outside VR; YouTube embeds cannot render inside WebXR

Let viewers walk down a 3D corridor where each year is a room of videos.

Show every water video as a glass on endless shelves that fill as numbers grow.

Grow a virtual garden where each carrot video is a plant that blooms when watched.

Present Everyday Repairs as tools hanging on a pegboard that open episodes.

Build an alternate interface styled like a 1990s desktop with draggable windows for each channel.

Wrap games and the archive in a pixel arcade lobby with machine cabinets per channel.

Plot videos with publicly stated locations on an OpenStreetMap powered map.

No publicly stated locations found in 10,906 titles and descriptions, so the map is empty by design

Search JonsTube, JonTok and InstaJon from one box with merged results.

Let viewers export and import a single local profile file that works across all three sister sites.

JonsTube side only: JonTok and InstaJon need an import button to read the file

Show live freshness of each sister site with last sync times and new item counts.

Link a JonsTube video to its matching JonTok clip and InstaJon post in a unified view.

Matches by number; only works where all three sites share a Short

Provide copy paste widgets so anyone can embed a random video or latest upload on their own site.

Publish a living style guide documenting components shared across the three sites.

Tokens and components documented from JonsTube; sister sites have not adopted them

Auto build a page summarizing new content across all three sites each week.

Sister sites expose no watch history, so the recap covers JonsTube history plus public sister catalogs

Ship a free open source extension that adds a new tab page with a random JonsTube video.

Generate ready to publish daily post text and images for an approved bot, held as drafts for approval.

Link every video, clip and post across the three sites into one browsable graph of shared subjects and dates.

Run yt-dlp on a free schedule to pull new uploads, rebuild and deploy with zero manual steps.

Pipeline scripted and a workflow written, but no schedule is running

Detect broken thumbnails and regenerate them from frames automatically.

Healing retries alternate YouTube thumbnail sizes; no self-generated frames

Use Playwright screenshots on every deploy to catch layout breaks before going live.

Run automated test scripts that walk through key flows daily and report failures.

Journeys run on demand, not on a schedule

Track performance scores over time in a public chart on the changelog.

Split the search index into lazily loaded shards so search stays instant as the catalog grows.

Demo index for the lab; the main site search still uses the full index

Compute recommendations entirely on device so nothing personal leaves the browser.

Publish an automated audit of every cookie and request the site makes.

Use perceptual hashing to find duplicate uploads and link them together.

Show missing metadata, broken links and sync gaps on a private admin page.

Use the free Open Meteo API to tint the homepage and pick fitting videos for local weather.

Shift the layout and picks across morning, afternoon and late night automatically.

Generate themed homepage takeovers for major holidays from matching archive videos.

Celebrate channel birthdays and video anniversaries with auto built tribute pages.

Show a scrolling stock ticker style bar of the newest numbers across counting channels.

Open one hidden archive pick per day in December with a door animation.

Build a themed summer program of daily watch missions drawn from the archive.

Generate an audio only stream of calm clips with open source music generation underneath.

Procedural Web Audio lo-fi instead of an ML music model, to stay in browser and free

Serve a special late night interface with starfield visuals and quiet videos after midnight.

Let viewers write a note stored only in their browser that reopens on a future date alongside a chosen video.