Dev recreates OpenAI's Astra-style live canvas demo: voice-driven 20-minute game build, no reloads
Beneficial-Cow-7408 · reddit · 2026-11-04
Inspired by OpenAI's GPT-6 "Astra" launch demo (talk to a canvas, watch a rocket appear and evolve in real time), a developer rebuilt the experience with a single canvas and direct API calls — no Blender, no MCP setup — as an open project called LiveLoop.
The demo runs voice-driven (originally OpenAI Realtime WebRTC, switched to ElevenLabs V4 Turbo after a 33-million-credit grant), stays live while changes happen, with no reload or restart, recorded in one take: roughly 20 minutes from a yellow circle to playable gameplay, cut to four minutes.
Testing GPT-6.1 Luna/Terra/Sol, he found all could handle initial builds, but Luna was inconsistent at live patching — about half the time the canvas didn't visibly update until a refresh or state change. He's still determining whether that's a model limitation or a bug in his implementation.
The bigger question he's chasing: not whether AI can generate a rocket, but whether you can keep editing one persistent, evolving project — correcting, branching, keeping versions — without rebuilding from scratch. LiveLoop offers 3 free anonymous builds (Claude Haiku 5.5) plus 500 credits and 5 more builds on signup.
More from coding & agent
- 3 open-weight LLMs auto-screen 15,034 records: sensitivity 1.00, specificity 0.98 — maier_ak · 2026-11-04
- Sub-4B model agent benchmark: Qwen3.5-4B hits 51.4% strict pass, thinking mode adds up to 20 points — KienShen · 2026-11-04
- External deterministic state layer proposed to fix AI coding session amnesia, with 3 papers surveyed — ApprehensiveSkin7975 · 2026-11-04
- Axion 1.0: ByteDance-Backed Open Agent Foundation Model With 3-Stage Recipe — Yuqi Jia · 2026-11-04
- Caper Pre-Executes Shell Commands in Sandboxes for Safer Agent Approvals — SingaporeManagementUniversity · 2026-11-04
- Claude Code lead: AI shifts team bottlenecks upstream to PMs and downstream to verification — Flomerboy · 2026-11-04