创作者反思 Jev 自驾与无人机 demo 的模拟环境
作者对比自己的生图成绩与 Jev 演示，指出模拟道路和障碍赛道没有真实传感器噪声与失败模式。
I GOT 10 OUT OF 10 BY REMOVING EVERYTHING THAT COULD FAIL. SO DID THEY. ten generations last week, 150 seconds of finished film, zero re-rolls. best result i have ever had by a factor of six. then i read what i actually built. no character in half of them. no events in any of them. flat overcast light with one source. a camera that never changes direction. petals for motion, because petals cannot be wrong. i did not get better. i built a room where failure was not available. five days ago i counted a directory of gpt-6 astra demos: 18 cases, 6 creator-supplied live links. twelve of the eighteen are videos of a thing rather than the thing. today jev launched, a system-1 model from a co-creator of chatgpt, and the criticism landing on it is the same shape. the self-driving demo runs in a simulator with a pre-built car on a pre-built road. the drone demo runs in a simulated obstacle course. no real hardware, no sensor noise, no safety-critical failure modes. i have not tested jev and i am not going to pretend otherwise. i am pointing at the structure, not the model. the structure is this. a simulated road has no sensor noise. a recorded demo has no try-it button. a fifteen second shot with no events has nothing to get wrong. in all three cases the impressive number is real and the conditions did the work. so when you see my 10 out of 10, apply the same test you should apply to theirs. ask what was removed. i've been reading my own results like they measured skill. turns out they measured how carefully i chose what not to attempt. 18 demos → 6 links 1 driving demo → 0 real roads 10 generations → 0 events ai generated
这条是引用着另一条写的：x.com/whoyatagarasu/status/2099448026666299752