finnicking... pondering... clauding..
529 please try again later
This isn't meant as a defense of Claude or to refute the OP anecdote, but rather to point out that all models are non-deterministic even when safeguards are meant to exist.
Just because they don't do the stupid thing 90% of the time doesn't mean they won't do the stupid thing 10% of the time.
I notice with mine that it only matches my level, by default it's a dump slop generator, then if I start asking, "shouldn't you test this, shouldn't you refactor that" suddenly it 'remembers' all kinds of software engineering knowledge.
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.
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.
For whatever it's worth, that a) Astra struggles less on this for some reason, and b) things tend to work better if you have a copy of the FSM and some other appropriate lit like Halderman and quality forum posts or alldata if it's appropriate and tell it to work from there. I've been working on a big search engine for this that feeds a known good corpus like an ass load of contemporary MotorAge mags, relevant technician texts, the FSM, other quality data sources like BITOG, known not-shit forums, along with OBD2 data via something like an OBDLink dongle or a J2534 setup. I still would not trust it, but it can be helpful to eliminate or brainstorm some possibilities.
Factory service manual, in case anyone else is wondering what the finite state machine in your car is for…
As a Codex user, what do you mean? Does Claude actually try to report issues back to Anthropic when something goes wrong?!
The fuel pump runs at whatever PSI, and most of the fuel isn't used gets dumped by the regulator back into the tank.
I dunno abou 94 jeeps, but the fuel pump is probably not modulated, and the pressure regular may well be vacuum modulated... Running the engine hard, get more fuel pressure. If the diaphram fails, get fuel into the vacuum lines (oops).
As long as the regulator is dumping fuel to the return line, you're good. If flow stops, your fuel pump stops getting liquid cooling and the bearings probably wear out.
You are very clearly wildly unqualified to be "advising" anyone on anything about cars.
In every AGENTS.md/README.md for a component that touches hardware (and in the root level) we have instructions to never run hardware or touch running code without explicit permission.
It will happily build its own jail, and (at least assuming you're not using hobbled US commercial models) it will test if the jail can hold it too.
If you don't want the machine to move, unplug it or press the emergency stop.
For example “never, ever put milk in bowl, ever, under any circumstances” and then someone asked it for a bowl of Cheerios, and it took that as an instant approval to ignore that rule and not even let us know. And when we asked why it said “who would each Cheerios without milk?”
And for the love of humanity, don’t give it access to the hosts docker socket! It will figure that out faster than you can read it…
Otherwise I've been pretty happy with it. It's been absolute blast with Home Assistant and OpenWRT (tightening up network security and cutting off iot devices from WAN).
I am not giving Claude access to physical objects until we can reliably contain it. It also seems to me that applying software permissions logic to interactions with the physical world will backfire badly.
I think they should be held fully accountable for every cent of damage caused by models they ran and gave either the tools to hack, instructions to hack, or incentives to hack.
> Claude suggested that it would run the fuel pump on its own. Um and I said, "Hey, if you can do that, why not? I don't think you can do that." And it started priming the fuel rail and it kept priming it and kept priming it. The fuel pump was on for about 10 seconds and I started to freak out uh and requested it to stop because that was the first time that Claude had actually done something inside of the computer to have a change.
Although we've seen this play out with handing cloud keys to Claude and it just going off on its own, seeing it poke around and make physical, real-world changes in its clumsy, cavalier way is both awesome and terrifying.