ZihangZ

5 ms·

1.
by ZihangZ 5mo ago
Small caution from using agents here: the useful chart is the one generated from code, tests, or traces, not the one the model draws from its own explanation. I've had models produce very reasonable Mermaid diagrams that matched the in
2.
by ZihangZ 5mo ago
+1 to the CI/isolation point. That is the part that makes these setups work for me too: make the failure cheap to reproduce, make stderr visible, make the agent rerun the same command after the patch. A lot of bad agent behavior is re
3.
by ZihangZ 5mo ago
Yeah, LAN is the line for me. USB-side sshd is a weird dev leftover; LAN means it’s now in the home threat model.
4.
by ZihangZ 5mo ago
Built a ROS 2 cycling helmet. IMU + GPS, but we didn't fuse them, on purpose. Speed is just haversine between GPS fixes. IMU only does turn detection and crash/fall. No EKF. Under bridges or urban canyons I'd rather have spee
5.
by ZihangZ 5mo ago
Yeah, this is pretty common once a device has any real DSP in it. There's usually some stripped-down Linux on an ARM SoC underneath, and the vendor BSP just happens to ship with sshd on. Not necessarily malice, more like nobody on the
6.
by ZihangZ 5mo ago
I don't think SSH vs OpenTofu is the core issue here. For agents, declarative plans are still valuable because they are reviewable. The interesting question is whether exe.dev changes the primitive: resource pools for many isolated VM-
7.
by ZihangZ 5mo ago
This matches my experience. When I write the boring glue code myself, I get a map of the project in my head. When I let an agent write too much of the structure, the code may work, but a week later every small change starts with "where
8.
by ZihangZ 5mo ago
This matches what I've seen too — the hallucination gets much worse when the loop has no external verifier. "Does this board work?" has no ground truth inside the model, so it defaults to optimistic narration. What OP is doin