Lab grows neurons on chips to slash video model costs 8x vs Wan 2.2 at $0.05 per clip
PeterDiamandis · x · 2026-10-31
Justin Hackney, co-founder of Wonder Studios, visited biological computing startup TBC in San Francisco and tested its video model mischief 1. TBC trains living neurons on multi-electrode chips by converting real images and video into patterns of electrical stimulation, records the neural spikes, then decodes that activity into algorithms that improve open-source video models — adding under 0.1% to model size, with neurons staying in the lab and inference on standard silicon.
The pitch is efficiency: a human brain runs on roughly 20W vs 700W for a fully loaded H100. On cost, mischief 1 runs $0.01/sec at 720p on fal and OpenRouter — about $0.05 per 5-second 720p clip vs $0.405 for Wan 2.2 A14B at fal's list price, roughly 8x cheaper. Hackney calls this a "green model" and is using it in production on his Dream Canvas pipeline.
More from Multimodal
- 'You wouldn't think I made this with AI': another bar-raising AI video — Effective_Regular770 · 2026-10-31
- Redditor remakes LTX trailer with MiniMax: 10s 720p in 4 minutes on RTX 5090 — aurelm · 2026-10-31
- Sketching a striking three.js scene in minutes with Claude and Tripo — nptacek · 2026-10-31
- AI-generated Hogwarts-style feast video wows Reddit — Quiet_Cobra · 2026-10-31
- AI-made Spanish pop anime 'Unicornios en el Bosque' shared on Reddit — Sea_Examination_1534 · 2026-10-31
- NVIDIA releases open-weight Nemotron 3.5 Transcribe: 30B MoE ASR model covering 40 languages — AccBalanced · 2026-10-31