Jev + Opus 5.5 + Dots：5 分钟 1000 亿次决策的流水线
5800 个节点持续跑 Jev 决策，Opus 只做 96 次探测找热点，再由 Dots 代理处理 31 次交付。
100 BILLION DECISIONS IN 5 MINUTES. TypeSafe’s JEV + Opus 5.5 + Dots built a system where the biggest model barely has to move. JEV is the brain. 5,800 nodes running continuously. 100B decisions in 5 minutes, with each decision taking ~3ns. Opus 5.5 is the spider. It doesn't read the whole stream. It walks the brain, finds where activity is hot, and probes only those nodes. 96 probes in 5 minutes. Dots are the workers. Echo, Sprout, Pocket and Stella ingest, train, store and rank what the spider brings back. 31 deliveries from those 96 probes. So the whole pipeline becomes: 100B decisions → 96 probes → 31 deliveries → 1 screen That's the part I find interesting. Most AI systems do: input → biggest model → answer This does: cheap fast layer → finds what matters → expensive model investigates → agents act The frontier model doesn't need to see everything. It just needs to know where to look. The question I'm asking now: How rarely can your most expensive model move while still making the right decisions? https://x.com/deezzex/status/2107476721888886994/video/1
这条是引用着另一条写的：x.com/arindam_1729/status/2103468454083002537