Opus 5.5 Ad Pipelines, Remote Orca, Jev Training — AI Daily Sep 26

Opus 5.5 dominated the day on two fronts: one-shotting full video ads and orchestrating other models. Builders chained Suno, ElevenLabs, Qwen Image 2.1, LTX 2.5, and Kie APIs — or generated React/SVG/Remotion 'puppet' animations directly in Claude Code — with end-to-end ad generation in ~5 minutes on a local 5090. @realcrischico shared a Mega folder with his full process plus an AI Board of Advisors persona pack. On the orchestration side, Opus 5.5 was praised for spawning Astra sub-panes, calling Kimi for research, and killing scope creep, with several members maxing multiple $200 Anthropic subs to keep it running. Infra chatter split between Remotion Lambda (~10¢ per 3-min render in ~60s, winning a head-to-head vs Hyperframes on style) and remote Orca setups for traveling builders — Tailscale latency tests, local VPS+SIM tricks, and a codex bridge letting GPT Live message a parent OMP orchestrator. Model routing was a recurring theme: Luna dropped from 120-150 TPS to 35-40, Baseten held at 33k tokens/sec, and Omniroute was recommended as a Codex account-rotation wrapper despite bugs on long calls. Members also compared reasoning-effort levels, finding xhigh over-engineers ('a business plan for a breakfast restaurant when you asked for pancakes'). @tounano ran a deep thread on encoder vs decoder architectures and is training embeddable Jev-style small models on RunPod community pods (~$1.50/hr), targeting iPhone and Raspberry Pi. @leewardbound floated calorie/nutrient classification as a killer use case. Meta's new Muse assistant got hype as a potential mass-market winner, and Anthropic hit legal turbulence after being designated a Pentagon supply-chain risk.