Jev 打 Doom：0.114 秒决策，比 GPT 5.6 快 75 倍
TypeSafe 官方 demo 把游戏状态转成结构化文本让 Jev 每秒决策十次，同时公开了它在国际象棋上被 Fable 5.1 打爆的失败案例。
NEW AI MODEL BEAT GPT 5.6'S RESPONSE TIME BY 75X PLAYING DOOM TypeSafe own launch demo has the receipt: Jev answered a Doom decision in 0.114 seconds GPT 5.6 Terra took 8.566 seconds for the same call Not a benchmark claim - a published sid by side number The Doom bot never sees a frame Game state gets turned into structured text - health, enemy position, distance, geometry - and Jev answers the same question roughly ten times a second: what now TypeSafe's own honest line about it: a normal scripted Doom bot would still play better The point was never skill, it was speed at $7 an hour in API cost A separate developer wired the identical pattern into Subway Surfers and ran 50 game sessions at once on one screen, superhuman reaction time, under a cent total. Someone else had it building a Minecraft flag from a blueprint for about a cent in two minutes Then the honest failure, published by TypeSafe itself, not hidden: Jev V13 played 5+0 blitz chess against Fable 5.1, one API call per move Fable outplayed it - by move 29 it was already +16 > A model built to be fast at bounded decisions lost badly the moment the decision space opened up into actual strategic depth That's the real boundary worth remembering before you route anything through it: reflexes, not reasoning Great at "jump or duck, right now" Not built to out think a model that's still allowed to think full below
这条是引用着另一条写的：x.com/notegoyard/status/2103135241342693883