Sep 24, 17:03

Overview

The speaker talks with Grok about two topics. The first is how the R1 Cord desktop server app can receive recordings from a hacked Rabbit R1 without asking users to set up a Cloudflare tunnel. Grok recommends Tailscale as more secure and nearly as easy to set up, and the speaker agrees it looks like the better approach. The second is how to build RSA Animate-style whiteboard videos with AI. The speaker decides to build their own tool from open-source parts, driven by Claude and Codex, for their YouTube channel. The talk ends on a plan: render the whole finished whiteboard as one large image first, then animate a camera over it in Remotion, synced to the narration.

Key points

R1 Cord server: Tailscale vs. Cloudflare

  • Asking users to configure a Cloudflare tunnel is too hard. The speaker wants to install one executable on the PC plus the APK on the Rabbit.
  • A desktop cannot open an inbound port without the router. UPnP can ask the router to open one, but many routers have it off and some ISPs block it.
  • A Cloudflare tunnel works through a long-lived outbound connection, so the router needs no changes. But Cloudflare sits in the middle of every request.
  • Dropbox and Google Drive folders could act as a relay. The speaker rules them out because R1 Cord’s whole point is keeping conversations off other people’s servers.
  • Grok says Tailscale wins on security: data flows peer-to-peer, encrypted end to end, and Tailscale’s servers only handle the first handshake. Nothing is exposed to the public internet.
  • On setup, Grok says Cloudflare is simpler on paper, but Tailscale adds only a one-time login on the Rabbit.
  • Grok says Tailscale is open source. You can sideload its APK from pkgs.tailscale.com, GitHub releases or F-Droid without the Play Store or a Google account.
  • Grok puts Tailscale’s memory use at about 10–15 MB idle on mobile, under 20 MB on Android. That is small next to the Rabbit R1’s roughly 4 GB of RAM.
  • Without a Google account, a device can sign in with another SSO provider, an auth key from the admin console, or a QR-code approval flow.
  • The Rabbit’s USB-C link to the PC can install and authenticate Tailscale over ADB.
  • Tailscale provides only the network. The PC still needs an upload endpoint, SFTP server or SMB share to receive files.
  • One computer can run several Tailscale nodes with separate state directories and sockets. This is easy on Linux but on Windows needs containers or WSL.
  • Tailscale Serve can publish a small local web server to devices on the tailnet only. Funnel would make it public.

RSA Animate-style whiteboard videos

  • The style is called whiteboard, sketch or doodle animation. Named players include VideoScribe, Doodly, Common Craft and RSA Animate.
  • Hosted AI tools (Ancon AI, animation.ai, Colpo, Animaker) mostly charge monthly. The speaker wants to build their own for personal YouTube videos.
  • Grok proposes an open-source pipeline. Claude or Codex writes a scene-by-scene script, an image model (Flux or GPT Image) draws ink-on-white scenes, Whiteboard-Animator (MIT, CPU only) animates them, and FFmpeg joins the clips. Narration comes from ElevenLabs or the free Edge TTS.
  • HandDraw-Skill (SVG, Remotion, FFmpeg, Edge TTS) does much of this end to end and is built for coding agents like Claude Code or Codex. It could serve as a fork base.
  • Whiteboard-Animator does not understand the script. It picks drawing order by layout rules unless a JSON region plan (optionally drafted by Gemini) sets the order.
  • It draws with raster masks, not vectors. It finds connected dark-pixel blobs, traces each blob’s skeleton from an endpoint, and gives every pixel a reveal time. It works only on clean ink-on-white images.
  • AI video can draw things in order only roughly. VideoSketcher is a research prototype. In practice, creators chain start and end keyframes through video models or use Kling’s motion brush.
  • Grok suggests a hybrid: an image model for the picture, a stroke renderer for drawing, and a video model only for camera moves or shading.
  • The speaker’s idea: render the full whiteboard at very high resolution first, then add camera moves from the narration. Grok agrees and says the RSA Animate studio worked on a giant canvas with a separate camera layer.
  • To build the big canvas, generate a grid of about 1024×1024 tiles, giving each tile’s prompt its neighbours’ content, then blend the overlaps into one 8K–16K image.
  • Remotion loads that one image and animates a camera window (pan, zoom, reveal) keyed to narration timestamps.

Action items

  • Replace the Cloudflare tunnel in the R1 Cord server app with Tailscale.
  • Install and authenticate Tailscale on the Rabbit over the USB-C/ADB connection.
  • Build a personal whiteboard-animation tool from open-source parts, driven by Claude and Codex.