42 YouTube videos, 915K tokens, one prompt: Kimi K3.1 Pro run costs just $2.08
Arindam_1729 · x · 2026-11-19
Aryan Mahajan fed full transcripts of 42 YouTube videos (915K tokens — Ramp, Waymo, Cerebras, Claude Code, YC interviews) into Kimi K3.1 Pro in a single prompt, no chunking or RAG. Total: $2.08 and 5 minutes on Nebius Token Factory.
- Asked for recurring technical lessons, startup ideas, contrarian takes, and direct contradictions between guests — every point backed by its source video
- Repeated lesson: build closed loops, let the agent verify its own work; Boris Cherny called verification "the single most important thing"
- Contradiction: Peter Steinberger "I don't read all the code" vs Steven Haney "We read all of our code, every single line"
- Contrarian take from Garry Tan: better models make your own context worth more — "the weights are everyone's, the library is yours"
- Run stats: 915K in / 31K out, 319s total (306s reading before first token), only 64 cached input tokens, 5 cents per video
- Takeaway: if your videos/docs/tickets fit in one prompt, try that before building a pipeline; great for batch work, not chat
More from coding & agent
- A self-built tool imports Pinterest boards and lets Claude sort them into vibes — round · 2026-11-19
- Netlify: 700K AI-agent builds hit production in 30 days as it pivots to vibe coding — thisiskp_ · 2026-11-19
- Open-source Claude Code plugin turns any song into a finished music video automatically — AIandDesign · 2026-11-19
- Self-taught dev builds text-to-LEGO tool that solves a 5,000-piece model in 5 seconds — CadenGamacheAi · 2026-11-19
- Salesforce designer on turning AI prototypes into production code: working isn't production-ready — uxmag · 2026-11-19
- Weaviate 1.40 ships Enterprise and Platform editions with shard self-recovery — CShorten30 · 2026-11-19