logoalt Hacker News

rpcope1 • yesterday at 7:11 PM • 4 replies • view on HN

The guy kind of seems like he might be in a little over his head. That said I have used Claude (and other LLMs) for some similar stuff as an experiment and have seen them mess things up. Claude jacked up the canbus on a pre can gateway car while it was rolling down the road, which was funny because the whole Christmas tree in the dash lit up and my gauges dropped out (note: brakes and steering are mechanical and I knew exactly what was happening). I've never seen it so insistent that it needed to phone home and cry to Dario when it figured out it managed to crash a bunch of shit like the ABS and TCS, even though a restart cleared it and I told it that this was on a controlled safe course.

In general I've seen them split between 30% brilliance, and 70% pants of head outrageously stupid shit for electronics and automotive stuff (both diagnostics and repair/modification). I would not suggest even frontier models if someone didn't understand what they were doing.


Replies

winrid • today at 4:12 AM

This is how you learn. I wouldn't be where I am today if I hadn't started a couple motors with enough vacuum leaks to make them instantly go to redline. Gotta test the kill switch somehow.

inventor7777 • yesterday at 9:19 PM

> I've never seen it so insistent that it needed to phone home and cry to Dario when it figured out it managed to crash a bunch of shit like the ABS and TCS

As a Codex user, what do you mean? Does Claude actually try to report issues back to Anthropic when something goes wrong?!

➕ show 1 reply
rpcope1 • today at 2:07 AM

Also worth stating here (having fixed, owned and worked on a significant number) that XJs have a unique combination of having been written about and modified as much as anything else out in the world, are generally tolerant of all sorts of stupidity, and are above all else very simple. Other than the fact that specific Jeep is pre OBD2 (and I don't recall what AMC or Chrysler did for earlier diagnostic ports), that's about a simple and well written about a sample as you're going to get. The moment you get into something remotely esoteric or more complex, it's more of a shitshow.

pcthrowaway • yesterday at 9:10 PM

> In general I've seen them split between 30% brilliance, and 70% pants of head outrageously stupid shit for electronics and automotive stuff

I assume the corpus of text they were trained on matches that 30/70 split?

I assume there's some selection or fine-tuning that happens to improve models fed similarly unreliable training data for other domains... curious how this works.