Jev is incredible for many purposes. But not for everything that it seems capable of doing.
As you keep exploring those Jev ideas, be mindful to verify see whether it's actually good for the task. Obviously, with evals.
I stayed up til 2am and spent $1,000 benchmarking Jev Router so you don't have to.
Performance on DeepSWE was roughly the same as GPT-6 Astra on low. It costs slightly more, and it took almost 5x longer to run.
Me: Why do we need this new file for x?
Agent: I conflated the need for x with the need for a new file. We can omit it.
Me: Why should we omit the file?
Agent: I suggested removing it only to omit a file. That was an overcorrection. The file should stay in the plan.
Me: Why
Not sure when we'll achieve this again, so I'm sharing it proudly:
This year we're managing to be in the US for the US National Day, in Peru for the Peruvian National Day, and in Germany for the German National Day.
