REAX SN92 与 Jev 同攻「AI 决策经济」：26ms 本地决策

Andy ττ @bittingthembits原推发布于 2026.10.16

在 X 上打开原推

🚨 Most AI work isn't what many think. It's making decisions. The AI Decision Economy @reaxlabs SN92! Billing or support? Cheap model or expensive model? Approve, confirm or escalate? Millions of tiny choices a day. Using a giant model for each one is like hiring a lawyer to sort your mail. That's the problem REAX SN92 is going after. Jev, from TypeSafe AI, targets the same market: fast decisions for software instead of long answers for people. REAX’s approach combines open-weight alternatives with a network paid to serve and improve them. What it does: Decision Models, the company behind it, serves small, fast models that read a request and return a decision, with a confidence score, instead of generating text. • A model router that picks which AI should handle each request • About 26ms to decide when run locally, about 93ms through the hosted API • Open source under the MIT license • An engine its founder measured at about 2.6x faster than the reference server Why it's interesting msg, a European IT firm with 10,000+ employees, 34 countries and €1.5B in revenue, is evaluating the router internally. Not a contract. But real enterprise testing it on real work. Where Bittensor comes in to this: SN92 miners serve these models on big GPUs, small GPUs and even CPUs. And starting October 16, emissions get split four ways: • 50% for serving inference • 20% for making the engine faster • 30% across two specialist model competitions So miners don't just run the product. They get paid to improve it. The serving network went live on mainnet October 14. It's days old. Customers pay in euros and miners earn subnet tokens. There's no published buyback or revenue link to SN92 yet. It's largely one founder's work so far but @airesearch12 also runs JevBench, so he knows the competitive field. And that's the point The opportunity isn't one person building everything. It's one person designing a system that pays many others to make decisions cheaper, faster and better. That's what a subnet is for. 🧠 Sources: @decisionmodels_ docs, @reaxlabs engine repo $TAO / DYOR.

这条是引用着另一条写的：x.com/airesearch12/status/2110410684873814482