用 Jev 做 agent 微决策省 277 倍成本
视频演示 agent 一晚上 3700 次小决策全用 Jev 只花 $0.16，不确定的再交给 GPT-6，准确率反而更高。
this is pure f*cking treasure AN EX-OPENAI RESEARCHER BUILT AN AI THAT CAN'T WRITE A SINGLE WORD. IT JUST MATCHED GPT-6 AS A JUDGE AT 277X LESS. most of what an agent does isn't writing. it's tiny calls. read this or skip it. keep it or drop it. is this new. is this good enough and most people pay writer prices for every single one of them the video is one night of that. pink boxes are Jev making those calls. purple is the only place a big model actually writes here's the math for one night: 312 sources x about 12 calls each = ~3,700 decisions all on GPT-6: 3,700 x $12.182 per 1,000 = $45 all on Jev: 3,700 x $0.044 per 1,000 = $0.16 send the 14 unsure ones up to GPT-6 and you're at about $0.33 that last step is from a Carnegie Mellon paper. Jev answers first, and only the ones it isn't sure about go up to GPT-6. that setup scored 93.4%. GPT-6 on everything scored 92.5% so it's cheaper and slightly more accurate using GPT-6 to decide if a link is worth opening is like paying a lawyer to sort your mail
这条是引用着另一条写的：x.com/gippp69/status/2103394701063737504