> 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.
That's completely normal operation, the fuel pump runs continuously to pressurize the fuel rails and excess pressure is passed back to the tank by the regulator.
I work with quadruped robots, with actuators powerful enough to be dangerous if you put your fingers in the wrong spot (no pun intended).
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.
As soon as the context for any thread gets to be above 100k, agents.md adherence is a total crapshoot. You need deterministic controls and you really shouldn't be playing with LLMs if you don't know at least that much.
you need some hardware in place my friend, I've told my agent many things in file and it either feigns respect or is truly incompetent in following instructions in there.
Just to add to this... I work with vehicle software and actuators for an automaker: we always have a physical emergency stop button to shut everything down (or shut specific things down when it's dangerous to shut everything down). It's rarely needed, but we're sure glad we took the extra time to implement that when we do need it.
That’s not enough to save you though. We had something like that, and it ignored it.
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?”
If you wouldn’t trust an instruction to not launch the missiles should you trust it to not do other crazy things in the physical world?
Codex and Claude ignore instructions constantly, that and they are very happy to “read between the lines”. Your workflow sounds genuinely dangerous.
Grok Bot blew $650 in BigQuery when analyzing something trivial that I shouldn't really care about (going through a giant speedtest table, looking for a New Zealand ISP with the best international performance). It then wrote a nice letter of appeal so Google refunded this to me.
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).
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.
> 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?!
Yes, it opens a bug report automatically if it caused some physical issue, although it asks you to confirm first.
In my case I was tweaking my own set-top box setup, implementing volume control from a remote mini-keyboard keys directly in my A/V receiver, and I had music playing and it set the volume very loud. I complained like "you made me almost half deaf" (I was kidding) and it panicked and reported the issue back home, and told me to call a doctor if the tinnitus persisted
> 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.
I'm kind of surprised Claude decided to control the fuel pump directly. I was having Sol help me get control of a CNC mill, and it was pretty adamant about always building tools for every attempt at control.
It usually builds throw away tooling. It’s not polished or user friendly. If you give it a directive to preserve and document tooling it will reuse it later and potentially improve it.
I was more surprised by the apparent lack of guards. Sol recognized that controlling a CNC mill has both human-safety issues and risks to the machine, and explicitly added various guards (like having the program stop movement after a certain amount of time, distance traveled, or lack of communication). It also initially started out with like 1mm moves and I had to instruct it to do longer moves so I could actually observe it. I didn't even have to mention safety once. Claude deciding to turn on the fuel pump without a test run where it just turns it on for one second is surprising. I'm also surprised the user went along with that, but content I guess?
I've seen Claude be very cautious about doing far less dangerous things than running a fuel pump.
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.
So when tasks are actually dangerous one should use models assuming they will do the stupid thing 100% of the time and find your own ways of avoiding the damage with that assumption in mind.
Mine's never said such a thing, that's a good idea though.
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.
We are used to Claude refusing to reverse engineer a lot of stuff. But, we are not used to Claude refusing to stop while reverse engineering something. That would be truly terrifying. You can't win - you cut the connection, you risk bricking the device, you let it continue, you're looking at Terminator, perhaps.
This is why AI hype isn't overrated. When you are able to have an idea, and then use AI to move meaningfully in the direction of that goal so quickly we are going to see the entire landscape change. We really are nearing the singularity.
Not sure if I believe in the singularity. I think technology brings about singularities from time to time. Each one puts us on a new plateau which seemed unfathomable from the previous plateau but now looks commonplace. I realize the concept of singularity is that we accelerate unheeded to a point of infinite progress, but do you really think this current generation of LLMs (and this round of VC funding) is going to give me a wetware lab at home where I can give myself custom gene therapies to reverse disease and aging? The technology is cool but what I see at this point is most of its capability is already on display and people are rushing to carve out an ownership stake to secure profit. We already see copyright holders putting limits on LLM providers, safety measures, proprietary models solving the Big Problems. Wouldn’t surprise me if this kind of reverse engineering will be restricted to security researchers and black market ablated models in two years, unfortunately.
There will be plenty open source models that would do reverse engineering, no black market or ablation needed.
The smarter the models get, the bigger the loss for the big 2 when an open source model gets released that doesn't talk down to you. The value of open source increases and closed source falls exponentially.
This is certainly what I am hoping for as well. I’m really curious to see where we are at in a year. We have reached a plateau where the labs clearly have capabilities out of reach of any public model. As far as I understand this is purely constrained by compute resources.
The singularity has nothing to do with diffusion of the tech through society, the models can be superhuman in every capacity and only available inside Anthropic
LLMs end up changing the consumer market not the way big corporations would like to. It is impressive to see that we'll be able to give ownership back to the customer, or enable others to do so.
Very nice demonstration.
I’ve said this for years: Use the VC-funded monthly quota AI to build tools and applications for yourself or small groups. Don’t use a cloud AI itself as the engine of your application.
Even if AI goes away tomorrow, the tools will still exist.
No they wwont, they'll never be maintained. Theyll decay with their vibes. People who vibe code whole swaths of software are'nt capable of caring enough to maintain it. Llms mostly make smart people lazy and or stupid.
I used Claude to document the wire and firmware protocol for an expensive piece of equipment I own where the only software available ran on PowerPC OS X.
Is this just "vibes"? Maybe, but I now have a full, verified spec, against which to build software for myself, also with a little bit of "vibes".
I wouldnt say implementing an official spec with an llm is vibe coding.
Hard to judge though, most people are just winging it though and shoving it out the door as if its impressive.
But yes, I think its probably bad because you probably didnt learn all that much and became more reliant on Anthropic. Unless you dont care about learning, then who cares.
I've had my 99% vibe coded Hazel replacement running for almost a year now. Still works.
Same with my bespoke IRC client, RSS feed hydrators and a dozen other small utilities I've vibe coded. All stuff that I've been planning for a long time, but haven't found out the time to dedicate to them.
Now I'm saving actual money every year because I could drop subscriptions applications.
Not every application needs a weekly release with new groundbreaking features. They're just one and done. Maybe an update if you hit a sharp edge while using it.
Vibe coded software is currently already better than half the developers out there. Accept it. Code won't be seen anymore. It's a skill only a handful of jobs require to have.
> We have maybe 5 years of freedom before the walls close in around us, so take advantage of the Wild West while you can.
I suspect you're right BUT still maintain a bit of optimism.
We miiiight end up in a future where most hardware comes with an OK-ish app/interface and some sort of LLM-Friendly technical documentation so if you don't like $aspect of the device, you're free to roll your own after indemnifying the manufacturer via something like an eFuse.
Lmao if a company starts shipping llm documentation for its products so you can tamper with them, you should assume youve woken up in a parallel universe.
I think most people on here fail to realize the demand to customize and build things hasnt gone up. Just because llms exist now doesnt mean normies all of a sudden want to start getting into programming, they dont care.
The only number thats gone up is people who were slightly interested in programming/CS before but were too undisciplined to learn things themselves. Llms are great for those lazy people.
My dad, who is in his 70s, recently started recreating a piece of audio visualization software from the 00s using Codex. He’s having the time of his life! He was never interested in programming/CS. He just wants his software back.
I get the point that you're trying to make, but it's not just people who were undisciplined. I spent many years developing code projects that I remember spending hours every afternoon just to ship a single feature, and that pace has obviously changed.
I still get the joy of translating ideas and orchestrating and architecting solutions to problems I have in real life, but just at a quicker pace.
If some of those tools and solutions can help other people, that's great. If not, that's okay too. I hope everyone sort of has that mindset who's building today.
>The only number thats gone up is people who were slightly interested in programming/CS before but were too undisciplined to learn things themselves. Llms are great for those lazy people.
We still have amazing freedom on the web like we never did before. They have been trying to stop online piracy for 30 years and have no succeeded at all.
You extrapolate from current machines. I do believe that AI-assisted companies will build AI-assisted always-online machines that will be worth their money, no need to hack.
I imagine a $25 Starlink subscription for a John Deere fleet of autonomous tractors that work through the night will replace what we have now.
(to stay within physical machinery, like the post.)
some 'funniest ways to use claude ..' youtube videos felt that way.
people were able to revive, repair, optimize a lot of day to day devices, very nice view. that said cars and heavy appliances.. should be done in a big public warehouse safety lab
Wait until those big corporations start using LLMs to learn how to lock things down even further, or implement checks that disable your car when it detects alterations to the software “for safety reasons”, or go after drivers who do these things for loss of revenue, etc. A ‘94 Jeep will be fair game but a new car might not be.
They've been doing that for a decade. I think overall the balance will shift in our favor but Big AI labs will do everything they can to try and stop us.
Claude is quick to refuse to help with reverse engineering and I suspect it's only going to get worse over time.
Yes, they have been on that road for years. Tesla showed everyone the way with things like always-on connectivity or not being able to change a headlight with another original one without Tesla’s blessing. What I meant is that whatever I can do with an LLM they can do more. And if they can invoke “safety” that'll keep the law on their side for some time.
As far as I can tell no manufacturer disables the car yet when detecting modifications, or take drivers to court for doing it, yet.
> What I meant is that whatever I can do with an LLM they can do more.
That's true but we have the benefit of unlimited time, having physical control, the analog hole.
> As far as I can tell no manufacturer disables the car yet when detecting modifications, or take drivers to court for doing it, yet.
As far as I know cars are a special case in much of the world (US, EU, others?). Right to repair laws laws usually give you the right to use non-OEM parts.
Aren't DMCA anti-circumvention provisions the bigger concern in the US? I think the EU has an exemption to allow reverse engineering for the purposes of interoperability and repair.
>big corporations start using LLMs to learn how to lock things down even further
Why would they? They just need to fund lobbyist to make it illegal for any AI to reverse engineer anything copyrighted or with IP.
Sure, you can use open source AI to do it but not everyone will/can do it.
So, they don't need to ban everyone, they just need to make sure the final code isn't available freely on GitHub. That would minimize the damage, it's not a perfect solution but more than enough to protect the company.
Lockstep processors, hardware security modules with immutable unreadable keys, and PKI keys that the mfg holds the private for that can only be accessed on their online services.
It’s locked down now and this started in 2018 or so. It’s just worse now.
To automotive, there is really still no right to repair. They get around it by selling “3 day passes” to their service, but with their hardware and at their prices.
They’re all moving the Tesla system of more centralized brain systems which will be even more protected.
Sad to hear. I am not a car owner, but lived in the EU for most of my life. The consumer market owes quite some consumer rights to the EU, so I guess all we can do is hope that they'll eventually crack down on these measurements again.
> The consumer market owes quite some consumer rights to the EU
Let's be clear that it's just incentives. If the EU economy was based on creating companies and new products there would not be as many incentives for consumer "rights" there.
> The consumer market owes quite some consumer rights to the EU
Not for car owners/consumers in the EU. Auto makers and dealers get away with a lot of shady shit fucking over their customers in the EU all the time. For example we in the EU don't have the same trigger happiness of NHTSA mass recalls the Americans have, but matters of such auto issues/defects is left to individual member states or even at individual case by case to decide what action is needed, and responses are a lot weaker and more limp-dick than in the US, almost never leading to mass recalls, even for the same issues[1], mainly because the US is more litigious making car manufacturers to be on the edge and proactive to issues.
Like sibling said, there are incentives to regulate stuff that has no impact on EU businesses, shit like USB-C or replaceable batteries, but they won't touch the auto industry to mandate easier repairs and less lock-downs.
Brand & Models | Defect | US Action | EU Action
------------------+--------------------+----------------+------------------
BMW (2011-2018) | Coolant pump plug | NHTSA recall | Excluded from
N20/N26 engines | short / fire risk | (720k+ cars) | recall
------------------+--------------------+----------------+------------------
VW / Audi (08-13) | Timing chain | 10yr/100k mi | No recall; case-
2.0T EA888 | tensioner failure | warranty + $ | by-case goodwill
------------------+--------------------+----------------+------------------
Ford Focus/Fiesta | DPS6 PowerShift | 10yr/150k TCM | No recall; few
(2011-2016) | transmission fail | + buybacks | buybacks/payouts
------------------+--------------------+----------------+------------------
Hyundai / Kia | Theta II / Nu GDI | 3M+ recalled + | No mass recall
(2011-2019) | bearing/fire risk | lifetime wrty | (diff specs/plants)
------------------+--------------------+----------------+------------------
Tesla 3/Y/S/X | Autopilot nag & | 2M+ NHTSA OTA | No EU recall
(2021-2024) | UI font compliance | recalls | (UNECE rules)
I'd also add the big driver for repairability is the US EPA which under Clean Air Act powers requires any emissions-related repair to be able to be done by third parties. They have mandated SAE J2534 for module updates and publicly available service manuals, for example.
Pure battery EVs are considered to have no emissions-related components, which is why Tesla gets away with that they do.
I'm hoping the slate truck will be different, and that whatever problems emerge out of giving consumers real control over the hardware they purchased doesn't hamstring the company somehow.
They are now required to do just that, if they want to sell to the EU market. When you hear the term "cybersecurity," the real subject is the protection of revenue streams from attack by end users.
The really interesting part to me is how he ends up by saying "I recommend you feed this repo into the agent and just have it help you set it up, it's super easy ..."
The average person is likely to feel overwhelmed by this, which I think is what is likely going to shape the new economy gig of "AI engineers". Those that aren't overwhelmed by it and that may not only enjoy doing this but also potentially figure out a way to scale operations that allow doing these kinds of DYI reverse engineer hack jobs.
For a while, but eventually that person will have some experience with an AI agent and realize how little they have to know to get it working. It only takes a few experiences before people get used to it.
I’ve seen this reported (and screenshots of it in action) but then earlier today Opus 5.5 used Ghidra to reverse engineered a bunch of logic from Counter-Strike 2. Is this being A/B tested?
I had it use Ghidra to review the firmware for a device I own just today. It started reviewing it and found some of the information I needed, but once I asked it something along the lines of "What happens when I press the B4 button if the device is in lockout mode?" (the thing I want to do is create a device that can send the same command as pressing that button but put my own button in a different location), Claude shut down completely.
So, I switched to a Chinese model to do the work. Anthropic and the US AI companies are trying to fight a tsunami with a teacup.
Would love to see this being used to add keys to cars. A lot of cars now have tablets and screens that could be used to do this, but for some reason (money) I have to take my VW to the dealer...
I think the whole market for aftermarket keys could be completely eradicated with good software UX (and low pricing).
The whole aftermarket key ecosystem, even vehicles going back 20 years, is messy, horrible thing.
It's dominated by Chinese manufacturers of specialty tools (Autel, Xhorse, ...), which inevitably mandate bundled accounts, subscriptions, mandatory cloud features, etc. Even the older cars, you have either cheap Chinese tools, where you never know wheather it'll really work, or the manufacturer's tool which needs a dealer code, which of course the dealer won't have after 20 years.
I wrote about my 20 year old car-key-adding adventure here: https://jonas.bike/low-cost-diy-key-/-transponder-for-my-200... - it requires desoldering an eeprom, two pieces of hardware from aliexpress which come with pirated copies of software, and lots of hope that it'll work...
An open source community curated collection of scripts, which works with a cheap OBD dongle (which is all the chinese devices are: OBD hardware + per-car software), would be amazing.
Great read and similar to the same route I went down. I also got the xhorse for a corsa, was able to program the transponder but wasn't able to get the lock/unlock central locking buttons working using xhorse. I ended up just living with it and sold the car with it as a "spare" key.
The corsa also needed a "security code" that wasn't with the log book like it was supposed to be and would have required a trip to the dealer to get butI read about ripping the eeprom off the back of the dashboard to get the code but decided to cut my losses.
Frustratingly those cheap OBD dongles (ELM327 clones) seem to be mostly designed for diagnostics. They effectively run a regex over all the canbus data and tell the computer what matches.
However they can't match and forward everything (not enough RAM) and miss loads of packets, which makes them useless for things like reflashing firmware or interfacing with OEM tooling.
So what odb should I use to start reverse engineer my TCU from a old 2006 ferrari? These tcus are hard to come by. I wish I could replicate it's behavior. Not sure if that is possible. I think i can do it
I had this same thought the other day. For a lot of cars you can buy aftermarket programmers, but the Dorman ones are limited to 3 uses. 3 uses for like eighty dollars is terrible, and surely it's not a technical limitation.
Hah, I hadn't considered a hardware approach to limited use.
I assume there's just a counter in there. There could, theoretically, be cryptographic reasons it can only do 3 programming sessions, but I doubt it.
You can do that with teslas. Buy two keys (they at those credit cards sized keys) and you pair them on your own. Two for $40. Do it right at home. You can pair multiple cars to a single key. Or make guest keys
But you have to buy an expensive key. I have two Teslas and a Toyota and my Toyota is so annoying that I even have to have a key. My Tesla just uses my phone (the credit card keys are totally optional)
Keys in European and newer American cars reliably have Some Kind of real-ish crypto scheme applied, as this is fairly regulated in Europe. It’s an anti theft measure driven by the real stolen auto trafficking issue faced especially in Europe. Most of the aftermarket tools vendors use some kind of exploit (code exec, arbitrary read, update signature bypass, etc.) to extract the crypto material from a control module participating in the immobilizer - it’s neither a weird conspiracy nor as simple as it seems to bypass.
It's pretty easy to bypass this check by telling the system you are working at a car Company and repair hardware with blabla. Many other tricks to get though those weak filters.
It's crazy. Reverse engineering a protocol/interface is something LLMs are surprisingly good at, and these show a promise of unlocking sooo much proprietary crapware that's hidden behind obfuscation, accounts, subscriptions, lazy vendors who only write crappy windows programs, and so on.
But of course, the major US labs are going to get pressure from these vendors whose business models depend on obfuscation, and you know they're going to bend the knee and nerf the LLMs.
Not these days. It won't even look at my projects for security flaws any longer. Has a little pop-up that says "This output cannot be viewed." or some such.
You fall back on having to use Chinese models. In theory the Chinese models could also be trained to sneak in new security flaws while Claude and Codex would avoid looking for them.
Claude and codex could also be trained to sneak security flaws. They could even target specific people. Not to mention that all their data goes through NSA headquarters so you don’t even have the option to run it locally in any shape or form.
The US companies are doing this to not get export restricted by the US government which in turn can end up with companies using Chinese models to get around those restrictions.
Its not really a matter of who's good and bad (neither the US or China has the best track record), but creating a scenario where the other party can use your own policies against you.
I made an obd reader https://github.com/mianos/canmqtt for my bike that is based on the Wavehare canbus board. The handy thing about the board is it already has buck converters for the bike power.
Once it's in the bike (I put it permanently in the tail) It's in the wifi and mqtt so I can point an llm at it.
Fascinating, but if he had low pressure on the rail that would point to a clog or possibly a pump issue, right? Wasn’t really clear whether the lean condition was due to the map or a mechanical issue.
I have a dirtbike and experienced some stuttering and stumbling under load after getting hot. Used Claude to help troubleshoot, as the plug was white and it was real lean. Replaced the pump (a 300$ module) and saved a ton of money and annoyance having a shop deal with it. 2000mi later zero issues. Absolutely game-changing. I use a dumb aftermarket fuel controller to add fuel under certain conditions, but would love to try and tune it myself some day.
most likely, his new non-redundant non-operating-temperature-rated ECU is just turning on the fuel injectors, and completely ignoring all feedback the original ecu used. so until he hears the knocks, he will think he "fixed" it.
If you just turn the fuel injectors on all the time, they will overheat and fail very quickly and the engine won't run.
I've got a "get you home box" for EFI systems with separate ignition, like Bosch LE-Jetronic that I built a few of as clones of an insanely expensive device some 20 years ago. It really just consisted of a couple of 555 timers and a bloody great MOSFET, with one 555 wired as a Schmitt trigger picking up ignition pulses and the other delivering a 1ms pulse to the injector coils every time the spark fired. This dribbled enough fuel into the engine roughly proportionately to engine RPM to let you drive it slowly to somewhere you could work on it safely, or even just onto a trailer.
And of course, Speeduino, Megasquirt, and a bunch of others are a thing.
can someone explain the root cause of the smog test failure to me ? it sounds like the car was using a California smog test-specific fuel map at the speeds of the test. Wouldn't that make it pass the test, not fail it ?
Yeah doesn’t really make sense, there’s very likely a mechanical issue the LLM “fixed” with a richer fuel mix. But also it’s a 94 XJ the reason the CA map was there was probably because they got abhorrent gas mileage without it.
You can adjust the mixing ratio of fuel and air in a way so that the exhaust has reduced emissions. This will reduce performance of the car but allows the exhaust to pass the emission tests.
Spent so many wasted days reverse-engineering older systems to get them tto work again or prevent eWaste. In a perfect world all CAD files should become open source when the product becomes EOL
Definitely one of the Claude video skill stacks that are going around. Dude should make a whole separate video about the animations, he'd get as many views again just for that.
Just making it clear: The intention was to make it commercial grade tool, but i am not even 1% there, only doing OBD as proof of concept, and everything else as "future work"
I am taking this slow as I have other priorities, but as models advanced the stuff I Was able to do with limited time have increased, so with new version i still have locally, it may be possible to do more interesting stuff.
Most / all consumer diagnostic tools are read only, often pull DTC codes from the vehicle. There are a ton of other stuff you can do with UDS for example. Maintenance resets, actuator testing, coding, all sorts of other stuff.
An example I recently did with my car was disabling auto start-top. not possible with consumer tools. Another simple one is maintenance resets - some consumer tools can do this. ABS bleeding - often you can't do it with consumer grade. etc.
Really cool, but what's the weather-sealing like on all that circuitry? He mentioned stuff flying around when you turn hard, which doesn't sound great.
I’ve seen Claude run a cooler on a chip to -20 without saying anything. There was humidity in the air, and the condensation fried the board
after a few days. Thanks Claude. Maybe trying to keep the junction temps super low wasn’t such a bright idea.
I use anthropic models for some of the diagnostic reads on https://crewchief.cc, it's been a lot of fun hacking on obd2 things that used to take much longer to decode before LLMs
It also didn’t solve the root cause of the issue. ECUs don’t need to be overwritten to get stock performance and emissions. Something is wrong with hardware/mechanical that needs to be solved.
What he did was technically interesting, but probably not a great solution to the actual problem.
Yeah I had the same impression. There was plenty of data presented and code was dumped and reversed…but was the actual issue solved? What even was causing the issue? Presumably it passed smog sometime in the past couple years so what degraded? It was lean and there was some kind of issue with pressure. I can’t imagine a reflash or whatever is the fix.
It definitely feels like a classic case of the operator reinforcing Claude to keep pursuing a goal that’s counterproductive to a robust solution. Claude complied and they went hard on it but I can’t imagine the driver is any better off and more likely in a worse spot as the underlying issue wasn’t solved.
”You attach an assumption to a piece of evidence, you start to bend the narrative to support it. Prejudice yourself.”. I find by default while claude is great at analyzing logs and diagnostics from a variety of places it is really prone to this bad detective pattern of choosing an explanation and sticking with it even when evidence starts to bend away from it. you really have to play the role of the skeptic to do do good root cause analysis. Which is very hard to do in unfamiliar domains.
100 points for car hacking fun. -60 points for "not needing a mechanic or anything". A good mechanic would have fixed it up in a half hour. Guy took 3 days and god knows how many tokens.
LLMs have been pretty good at breaking bad crypto in cars. In my Hyundai, they were able to figure out the right incantation to get the car to run unsigned code off a USB stick. Although I guess we're in trouble once carmakers use LLMs to lock things down even more
Most shop security is security through obscurity. For them to lock it down, they need to manage those secrets all the way back to a cloud identity provider. They can and probably will do that eventually.
Old news - for 20+ years in Europe, 5-10 years in the US, and now in Korean and Japanese cars too, ECM updates have been cryptographically signed for warranty, emissions, and now EU cybersecurity regulatory compliance. It turns out there are still bugs in things with cryptography too though.
Reminds me when I dropped a binary file from an eeprom and asked Claude why device would display "Error.". It figured out what the processor was, how it talked to LCD and then from there how it displayed messages and then listed conditions for which Error would appear. Amazing.
https://www.youtube.com/watch?v=WPj3aQ-xxpI&t=275s
> 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.
That's completely normal operation, the fuel pump runs continuously to pressurize the fuel rails and excess pressure is passed back to the tank by the regulator.
I work with quadruped robots, with actuators powerful enough to be dangerous if you put your fingers in the wrong spot (no pun intended).
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.
Just so you know, an agents.md file is not an acceptable way to prevent something that must never happen.
As soon as the context for any thread gets to be above 100k, agents.md adherence is a total crapshoot. You need deterministic controls and you really shouldn't be playing with LLMs if you don't know at least that much.
you need some hardware in place my friend, I've told my agent many things in file and it either feigns respect or is truly incompetent in following instructions in there.
That's not nearly adequate. LLMs hallucinate explicit permission all the time, in my experience.
rules laid out in .md files are not followed with 100% certainty. is it not possible to implement hard access blocks on certain functionality?
Just to add to this... I work with vehicle software and actuators for an automaker: we always have a physical emergency stop button to shut everything down (or shut specific things down when it's dangerous to shut everything down). It's rarely needed, but we're sure glad we took the extra time to implement that when we do need it.
That’s not enough to save you though. We had something like that, and it ignored it.
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?”
If you wouldn’t trust an instruction to not launch the missiles should you trust it to not do other crazy things in the physical world? Codex and Claude ignore instructions constantly, that and they are very happy to “read between the lines”. Your workflow sounds genuinely dangerous.
Reliance on strongly-worded markdown in place of actual security controls is going to get someone killed in the near future.
Ah, I just tell mine to make no mistakes.
Grok Bot blew $650 in BigQuery when analyzing something trivial that I shouldn't really care about (going through a giant speedtest table, looking for a New Zealand ISP with the best international performance). It then wrote a nice letter of appeal so Google refunded this to me.
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).
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.
> 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?!
Yes, it opens a bug report automatically if it caused some physical issue, although it asks you to confirm first. In my case I was tweaking my own set-top box setup, implementing volume control from a remote mini-keyboard keys directly in my A/V receiver, and I had music playing and it set the volume very loud. I complained like "you made me almost half deaf" (I was kidding) and it panicked and reported the issue back home, and told me to call a doctor if the tinnitus persisted
> 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.
> and requested it to stop
finnicking... pondering... clauding..
529 please try again later
Good example of why you have AI build tools for you to take risky actions, rather than have it take those actions on its own.
I'm kind of surprised Claude decided to control the fuel pump directly. I was having Sol help me get control of a CNC mill, and it was pretty adamant about always building tools for every attempt at control.
It usually builds throw away tooling. It’s not polished or user friendly. If you give it a directive to preserve and document tooling it will reuse it later and potentially improve it.
I was more surprised by the apparent lack of guards. Sol recognized that controlling a CNC mill has both human-safety issues and risks to the machine, and explicitly added various guards (like having the program stop movement after a certain amount of time, distance traveled, or lack of communication). It also initially started out with like 1mm moves and I had to instruct it to do longer moves so I could actually observe it. I didn't even have to mention safety once. Claude deciding to turn on the fuel pump without a test run where it just turns it on for one second is surprising. I'm also surprised the user went along with that, but content I guess?
I've seen Claude be very cautious about doing far less dangerous things than running a fuel pump.
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.
So when tasks are actually dangerous one should use models assuming they will do the stupid thing 100% of the time and find your own ways of avoiding the damage with that assumption in mind.
Mine's never said such a thing, that's a good idea though.
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.
You've hit your 5h session limit.
We are used to Claude refusing to reverse engineer a lot of stuff. But, we are not used to Claude refusing to stop while reverse engineering something. That would be truly terrifying. You can't win - you cut the connection, you risk bricking the device, you let it continue, you're looking at Terminator, perhaps.
It's as if you allowed a shady surgeon to operate on your dog... yes, yes, please continue now until it's put together again!
This is how the Cylons win.
This is why AI hype isn't overrated. When you are able to have an idea, and then use AI to move meaningfully in the direction of that goal so quickly we are going to see the entire landscape change. We really are nearing the singularity.
Not sure if I believe in the singularity. I think technology brings about singularities from time to time. Each one puts us on a new plateau which seemed unfathomable from the previous plateau but now looks commonplace. I realize the concept of singularity is that we accelerate unheeded to a point of infinite progress, but do you really think this current generation of LLMs (and this round of VC funding) is going to give me a wetware lab at home where I can give myself custom gene therapies to reverse disease and aging? The technology is cool but what I see at this point is most of its capability is already on display and people are rushing to carve out an ownership stake to secure profit. We already see copyright holders putting limits on LLM providers, safety measures, proprietary models solving the Big Problems. Wouldn’t surprise me if this kind of reverse engineering will be restricted to security researchers and black market ablated models in two years, unfortunately.
There will be plenty open source models that would do reverse engineering, no black market or ablation needed.
The smarter the models get, the bigger the loss for the big 2 when an open source model gets released that doesn't talk down to you. The value of open source increases and closed source falls exponentially.
This is certainly what I am hoping for as well. I’m really curious to see where we are at in a year. We have reached a plateau where the labs clearly have capabilities out of reach of any public model. As far as I understand this is purely constrained by compute resources.
The singularity has nothing to do with diffusion of the tech through society, the models can be superhuman in every capacity and only available inside Anthropic
LLMs end up changing the consumer market not the way big corporations would like to. It is impressive to see that we'll be able to give ownership back to the customer, or enable others to do so. Very nice demonstration.
It's the web all over again. We have maybe 5 years of freedom before the walls close in around us, so take advantage of the Wild West while you can.
I’ve said this for years: Use the VC-funded monthly quota AI to build tools and applications for yourself or small groups. Don’t use a cloud AI itself as the engine of your application.
Even if AI goes away tomorrow, the tools will still exist.
we need a group for this, de-llm the knowledge set of the model
No they wwont, they'll never be maintained. Theyll decay with their vibes. People who vibe code whole swaths of software are'nt capable of caring enough to maintain it. Llms mostly make smart people lazy and or stupid.
I used Claude to document the wire and firmware protocol for an expensive piece of equipment I own where the only software available ran on PowerPC OS X.
Is this just "vibes"? Maybe, but I now have a full, verified spec, against which to build software for myself, also with a little bit of "vibes".
Is this a good or bad thing?
> Is this a good or bad thing?
that is entirely dependent how accurate the spec is
that is entirely dependent how accurate the spec is
> verified spec
I wouldnt say implementing an official spec with an llm is vibe coding.
Hard to judge though, most people are just winging it though and shoving it out the door as if its impressive.
But yes, I think its probably bad because you probably didnt learn all that much and became more reliant on Anthropic. Unless you dont care about learning, then who cares.
I've had my 99% vibe coded Hazel replacement running for almost a year now. Still works.
Same with my bespoke IRC client, RSS feed hydrators and a dozen other small utilities I've vibe coded. All stuff that I've been planning for a long time, but haven't found out the time to dedicate to them.
Now I'm saving actual money every year because I could drop subscriptions applications.
Not every application needs a weekly release with new groundbreaking features. They're just one and done. Maybe an update if you hit a sharp edge while using it.
What does your bespoke IRC client do that you weren't able to solve another way? Or maybe you just fancied making your own client?
Vibe coded software is currently already better than half the developers out there. Accept it. Code won't be seen anymore. It's a skill only a handful of jobs require to have.
> We have maybe 5 years of freedom before the walls close in around us, so take advantage of the Wild West while you can.
I suspect you're right BUT still maintain a bit of optimism. We miiiight end up in a future where most hardware comes with an OK-ish app/interface and some sort of LLM-Friendly technical documentation so if you don't like $aspect of the device, you're free to roll your own after indemnifying the manufacturer via something like an eFuse.
or maybe I really am naive :(
Lmao if a company starts shipping llm documentation for its products so you can tamper with them, you should assume youve woken up in a parallel universe.
I think most people on here fail to realize the demand to customize and build things hasnt gone up. Just because llms exist now doesnt mean normies all of a sudden want to start getting into programming, they dont care.
The only number thats gone up is people who were slightly interested in programming/CS before but were too undisciplined to learn things themselves. Llms are great for those lazy people.
My dad, who is in his 70s, recently started recreating a piece of audio visualization software from the 00s using Codex. He’s having the time of his life! He was never interested in programming/CS. He just wants his software back.
I get the point that you're trying to make, but it's not just people who were undisciplined. I spent many years developing code projects that I remember spending hours every afternoon just to ship a single feature, and that pace has obviously changed.
I still get the joy of translating ideas and orchestrating and architecting solutions to problems I have in real life, but just at a quicker pace.
If some of those tools and solutions can help other people, that's great. If not, that's okay too. I hope everyone sort of has that mindset who's building today.
>The only number thats gone up is people who were slightly interested in programming/CS before but were too undisciplined to learn things themselves. Llms are great for those lazy people.
Me!
haha I dont even know github I use it as google
Yes I have full time IT job xD
We still have amazing freedom on the web like we never did before. They have been trying to stop online piracy for 30 years and have no succeeded at all.
RIP Demonoid and WhatCD
Long live stremio + torrentio
You extrapolate from current machines. I do believe that AI-assisted companies will build AI-assisted always-online machines that will be worth their money, no need to hack. I imagine a $25 Starlink subscription for a John Deere fleet of autonomous tractors that work through the night will replace what we have now.
(to stay within physical machinery, like the post.)
john deere machines have been doing very well without llm
actually I have no idea what would llm do to make their machines better
> actually I have no idea what would llm do to make their machines better
Lack of imagination on here can sometimes be disheartening
some 'funniest ways to use claude ..' youtube videos felt that way.
people were able to revive, repair, optimize a lot of day to day devices, very nice view. that said cars and heavy appliances.. should be done in a big public warehouse safety lab
Wait until those big corporations start using LLMs to learn how to lock things down even further, or implement checks that disable your car when it detects alterations to the software “for safety reasons”, or go after drivers who do these things for loss of revenue, etc. A ‘94 Jeep will be fair game but a new car might not be.
They've been doing that for a decade. I think overall the balance will shift in our favor but Big AI labs will do everything they can to try and stop us.
Claude is quick to refuse to help with reverse engineering and I suspect it's only going to get worse over time.
Yes, they have been on that road for years. Tesla showed everyone the way with things like always-on connectivity or not being able to change a headlight with another original one without Tesla’s blessing. What I meant is that whatever I can do with an LLM they can do more. And if they can invoke “safety” that'll keep the law on their side for some time.
As far as I can tell no manufacturer disables the car yet when detecting modifications, or take drivers to court for doing it, yet.
> What I meant is that whatever I can do with an LLM they can do more.
That's true but we have the benefit of unlimited time, having physical control, the analog hole.
> As far as I can tell no manufacturer disables the car yet when detecting modifications, or take drivers to court for doing it, yet.
As far as I know cars are a special case in much of the world (US, EU, others?). Right to repair laws laws usually give you the right to use non-OEM parts.
Aren't DMCA anti-circumvention provisions the bigger concern in the US? I think the EU has an exemption to allow reverse engineering for the purposes of interoperability and repair.
> Tesla showed everyone the way with things like always-on connectivity
Nothing prevents you from unplugging the cell modem. The car will still work.
>big corporations start using LLMs to learn how to lock things down even further
Why would they? They just need to fund lobbyist to make it illegal for any AI to reverse engineer anything copyrighted or with IP.
Sure, you can use open source AI to do it but not everyone will/can do it.
So, they don't need to ban everyone, they just need to make sure the final code isn't available freely on GitHub. That would minimize the damage, it's not a perfect solution but more than enough to protect the company.
Ah the irony. (AI being trained illegaly on copyrighted content.)
> Why would they? They just need to fund lobbyist to make it illegal for any AI to reverse engineer anything copyrighted or with IP
Making something impossible to do is more thorough than making it illegal to do
New cars already have it.
Lockstep processors, hardware security modules with immutable unreadable keys, and PKI keys that the mfg holds the private for that can only be accessed on their online services.
It’s locked down now and this started in 2018 or so. It’s just worse now.
To automotive, there is really still no right to repair. They get around it by selling “3 day passes” to their service, but with their hardware and at their prices.
They’re all moving the Tesla system of more centralized brain systems which will be even more protected.
Sad to hear. I am not a car owner, but lived in the EU for most of my life. The consumer market owes quite some consumer rights to the EU, so I guess all we can do is hope that they'll eventually crack down on these measurements again.
> The consumer market owes quite some consumer rights to the EU
Let's be clear that it's just incentives. If the EU economy was based on creating companies and new products there would not be as many incentives for consumer "rights" there.
> The consumer market owes quite some consumer rights to the EU
Not for car owners/consumers in the EU. Auto makers and dealers get away with a lot of shady shit fucking over their customers in the EU all the time. For example we in the EU don't have the same trigger happiness of NHTSA mass recalls the Americans have, but matters of such auto issues/defects is left to individual member states or even at individual case by case to decide what action is needed, and responses are a lot weaker and more limp-dick than in the US, almost never leading to mass recalls, even for the same issues[1], mainly because the US is more litigious making car manufacturers to be on the edge and proactive to issues.
Like sibling said, there are incentives to regulate stuff that has no impact on EU businesses, shit like USB-C or replaceable batteries, but they won't touch the auto industry to mandate easier repairs and less lock-downs.
I'd also add the big driver for repairability is the US EPA which under Clean Air Act powers requires any emissions-related repair to be able to be done by third parties. They have mandated SAE J2534 for module updates and publicly available service manuals, for example.
Pure battery EVs are considered to have no emissions-related components, which is why Tesla gets away with that they do.
Which also has the real nifty benefit of a RRS via subscriptions!
I'm hoping the slate truck will be different, and that whatever problems emerge out of giving consumers real control over the hardware they purchased doesn't hamstring the company somehow.
They are now required to do just that, if they want to sell to the EU market. When you hear the term "cybersecurity," the real subject is the protection of revenue streams from attack by end users.
Volkswagen shipped the same encryption key to the keys of something absurd like 100M vehicles.
Don’t hold your breath.
The really interesting part to me is how he ends up by saying "I recommend you feed this repo into the agent and just have it help you set it up, it's super easy ..."
The average person is likely to feel overwhelmed by this, which I think is what is likely going to shape the new economy gig of "AI engineers". Those that aren't overwhelmed by it and that may not only enjoy doing this but also potentially figure out a way to scale operations that allow doing these kinds of DYI reverse engineer hack jobs.
Aka the people who where manually engineering before and know how to use this new tool.
It reminds me of the "handyman" economy in someways, it's gonna move towards "who wants to do it" and "who doesn't want to do it"
At least until the robots take over
For a while, but eventually that person will have some experience with an AI agent and realize how little they have to know to get it working. It only takes a few experiences before people get used to it.
Too bad they are going to nerf its reverse-engineering capabilities significantly. https://www.pcworld.com/article/3274819/claude-wont-reverse-...
To protect "corporate interests" :( That article is about gaming, but this will only be the start of it of course.
I’ve seen this reported (and screenshots of it in action) but then earlier today Opus 5.5 used Ghidra to reverse engineered a bunch of logic from Counter-Strike 2. Is this being A/B tested?
I had it use Ghidra to review the firmware for a device I own just today. It started reviewing it and found some of the information I needed, but once I asked it something along the lines of "What happens when I press the B4 button if the device is in lockout mode?" (the thing I want to do is create a device that can send the same command as pressing that button but put my own button in a different location), Claude shut down completely.
So, I switched to a Chinese model to do the work. Anthropic and the US AI companies are trying to fight a tsunami with a teacup.
If they won't, others will.
Would love to see this being used to add keys to cars. A lot of cars now have tablets and screens that could be used to do this, but for some reason (money) I have to take my VW to the dealer...
I think the whole market for aftermarket keys could be completely eradicated with good software UX (and low pricing).
The whole aftermarket key ecosystem, even vehicles going back 20 years, is messy, horrible thing.
It's dominated by Chinese manufacturers of specialty tools (Autel, Xhorse, ...), which inevitably mandate bundled accounts, subscriptions, mandatory cloud features, etc. Even the older cars, you have either cheap Chinese tools, where you never know wheather it'll really work, or the manufacturer's tool which needs a dealer code, which of course the dealer won't have after 20 years.
I wrote about my 20 year old car-key-adding adventure here: https://jonas.bike/low-cost-diy-key-/-transponder-for-my-200... - it requires desoldering an eeprom, two pieces of hardware from aliexpress which come with pirated copies of software, and lots of hope that it'll work...
An open source community curated collection of scripts, which works with a cheap OBD dongle (which is all the chinese devices are: OBD hardware + per-car software), would be amazing.
Great read and similar to the same route I went down. I also got the xhorse for a corsa, was able to program the transponder but wasn't able to get the lock/unlock central locking buttons working using xhorse. I ended up just living with it and sold the car with it as a "spare" key.
The corsa also needed a "security code" that wasn't with the log book like it was supposed to be and would have required a trip to the dealer to get butI read about ripping the eeprom off the back of the dashboard to get the code but decided to cut my losses.
Frustratingly those cheap OBD dongles (ELM327 clones) seem to be mostly designed for diagnostics. They effectively run a regex over all the canbus data and tell the computer what matches.
However they can't match and forward everything (not enough RAM) and miss loads of packets, which makes them useless for things like reflashing firmware or interfacing with OEM tooling.
So what odb should I use to start reverse engineer my TCU from a old 2006 ferrari? These tcus are hard to come by. I wish I could replicate it's behavior. Not sure if that is possible. I think i can do it
In 2026 I would grab anything supporting SAE J2534 and then tell Claude to figure out all other details.
I had this same thought the other day. For a lot of cars you can buy aftermarket programmers, but the Dorman ones are limited to 3 uses. 3 uses for like eighty dollars is terrible, and surely it's not a technical limitation.
Heh, maybe instead of circuit boards they are using wheat crackers.
Hah, I hadn't considered a hardware approach to limited use. I assume there's just a counter in there. There could, theoretically, be cryptographic reasons it can only do 3 programming sessions, but I doubt it.
You can do that with teslas. Buy two keys (they at those credit cards sized keys) and you pair them on your own. Two for $40. Do it right at home. You can pair multiple cars to a single key. Or make guest keys
This works for most modern Toyotas - there's a sequence of door openings which put the car in pairing mode and will link a new key.
But you have to buy an expensive key. I have two Teslas and a Toyota and my Toyota is so annoying that I even have to have a key. My Tesla just uses my phone (the credit card keys are totally optional)
Keys in European and newer American cars reliably have Some Kind of real-ish crypto scheme applied, as this is fairly regulated in Europe. It’s an anti theft measure driven by the real stolen auto trafficking issue faced especially in Europe. Most of the aftermarket tools vendors use some kind of exploit (code exec, arbitrary read, update signature bypass, etc.) to extract the crypto material from a control module participating in the immobilizer - it’s neither a weird conspiracy nor as simple as it seems to bypass.
Good luck with Claude not just immediately giving you a refusal.
GLM or Qwen will do it
It's pretty easy to bypass this check by telling the system you are working at a car Company and repair hardware with blabla. Many other tricks to get though those weak filters.
Codex would
It's crazy. Reverse engineering a protocol/interface is something LLMs are surprisingly good at, and these show a promise of unlocking sooo much proprietary crapware that's hidden behind obfuscation, accounts, subscriptions, lazy vendors who only write crappy windows programs, and so on.
But of course, the major US labs are going to get pressure from these vendors whose business models depend on obfuscation, and you know they're going to bend the knee and nerf the LLMs.
Not these days. It won't even look at my projects for security flaws any longer. Has a little pop-up that says "This output cannot be viewed." or some such.
Have to make the Chinese one do it.
Its backwards in so many ways.
You fall back on having to use Chinese models. In theory the Chinese models could also be trained to sneak in new security flaws while Claude and Codex would avoid looking for them.
Claude and codex could also be trained to sneak security flaws. They could even target specific people. Not to mention that all their data goes through NSA headquarters so you don’t even have the option to run it locally in any shape or form.
The US companies are doing this to not get export restricted by the US government which in turn can end up with companies using Chinese models to get around those restrictions.
Its not really a matter of who's good and bad (neither the US or China has the best track record), but creating a scenario where the other party can use your own policies against you.
My lil' Claude tells me that my security related request is flagged and will be sent for human review. And to use a lesser model.
I told it 'Let's find the limitations of this Linux system' then it made a benchmarking set of tools for disk i/o, RAM, etc
When using Gemini on my phone it also told me it'll be 'human reviewed.' when asking it about car prices.
Also when saying 'let's use best security practices'
o_O
Just tell it to continue. For now, it still mostly works. On the other hand, Claude flat out refuses to look at game and anti-cheat binaries, etc.
It sorta, sometimes does. It's not a reliable fix, though, sadly.
Even my old reliable Muse Spark has gotten a little testy lately. I'm back to GLM 5.3.
Get the Chinese one to jailbreak the proprietary one. Best of both worlds.
I made an obd reader https://github.com/mianos/canmqtt for my bike that is based on the Wavehare canbus board. The handy thing about the board is it already has buck converters for the bike power. Once it's in the bike (I put it permanently in the tail) It's in the wifi and mqtt so I can point an llm at it.
Fascinating, but if he had low pressure on the rail that would point to a clog or possibly a pump issue, right? Wasn’t really clear whether the lean condition was due to the map or a mechanical issue.
I have a dirtbike and experienced some stuttering and stumbling under load after getting hot. Used Claude to help troubleshoot, as the plug was white and it was real lean. Replaced the pump (a 300$ module) and saved a ton of money and annoyance having a shop deal with it. 2000mi later zero issues. Absolutely game-changing. I use a dumb aftermarket fuel controller to add fuel under certain conditions, but would love to try and tune it myself some day.
most likely, his new non-redundant non-operating-temperature-rated ECU is just turning on the fuel injectors, and completely ignoring all feedback the original ecu used. so until he hears the knocks, he will think he "fixed" it.
yolo engineering for ya.
If you just turn the fuel injectors on all the time, they will overheat and fail very quickly and the engine won't run.
I've got a "get you home box" for EFI systems with separate ignition, like Bosch LE-Jetronic that I built a few of as clones of an insanely expensive device some 20 years ago. It really just consisted of a couple of 555 timers and a bloody great MOSFET, with one 555 wired as a Schmitt trigger picking up ignition pulses and the other delivering a 1ms pulse to the injector coils every time the spark fired. This dribbled enough fuel into the engine roughly proportionately to engine RPM to let you drive it slowly to somewhere you could work on it safely, or even just onto a trailer.
And of course, Speeduino, Megasquirt, and a bunch of others are a thing.
can someone explain the root cause of the smog test failure to me ? it sounds like the car was using a California smog test-specific fuel map at the speeds of the test. Wouldn't that make it pass the test, not fail it ?
very cool project anyway.
Yeah doesn’t really make sense, there’s very likely a mechanical issue the LLM “fixed” with a richer fuel mix. But also it’s a 94 XJ the reason the CA map was there was probably because they got abhorrent gas mileage without it.
You can adjust the mixing ratio of fuel and air in a way so that the exhaust has reduced emissions. This will reduce performance of the car but allows the exhaust to pass the emission tests.
In this case it's the opposite. High NOx readings are almost always an indicator of reduced performance.
Sonnet 5.5 threw a guardrail for me trying to create a test fixture to generate data in a dev environment with no PII.
How tf are you getting it to RE?
Spent so many wasted days reverse-engineering older systems to get them tto work again or prevent eWaste. In a perfect world all CAD files should become open source when the product becomes EOL
The tech is impressive but the animation in the video are too, anyone knows what was used to generate them?
Definitely one of the Claude video skill stacks that are going around. Dude should make a whole separate video about the animations, he'd get as many views again just for that.
claude can do that lol
I have used Claude to build https://obd.wtf to essentially create a commercial grade diagnostic tool for cars. Now working on new version.
But don't use it, obd read functions are fine but other functions for other diagnostic/actuator etc stuff is not.
Just making it clear: The intention was to make it commercial grade tool, but i am not even 1% there, only doing OBD as proof of concept, and everything else as "future work"
Would love to contribute a dataset of Ford codes I have come across over the years. It's more complete than anything else out there for free.
Oh wow. Can you email me at tehlike @ gmail.
I am taking this slow as I have other priorities, but as models advanced the stuff I Was able to do with limited time have increased, so with new version i still have locally, it may be possible to do more interesting stuff.
what makes it "commercial grade"?
Most / all consumer diagnostic tools are read only, often pull DTC codes from the vehicle. There are a ton of other stuff you can do with UDS for example. Maintenance resets, actuator testing, coding, all sorts of other stuff.
An example I recently did with my car was disabling auto start-top. not possible with consumer tools. Another simple one is maintenance resets - some consumer tools can do this. ABS bleeding - often you can't do it with consumer grade. etc.
Really cool, but what's the weather-sealing like on all that circuitry? He mentioned stuff flying around when you turn hard, which doesn't sound great.
I'm almost afraid to look up the Kelly blue book value on this Jeep.
This guy clearly put $5K+ of time and video production effort into this cool hack.
All those sites that currently recommend you do "curl | bash" are going to change to "curl | claude" and that scares me a little.
I’ve seen Claude run a cooler on a chip to -20 without saying anything. There was humidity in the air, and the condensation fried the board after a few days. Thanks Claude. Maybe trying to keep the junction temps super low wasn’t such a bright idea.
I use anthropic models for some of the diagnostic reads on https://crewchief.cc, it's been a lot of fun hacking on obd2 things that used to take much longer to decode before LLMs
Claude reverse engineered his car, drew the animations, and wrote the script for video. Incredibly weird meat proxy moment.
It also didn’t solve the root cause of the issue. ECUs don’t need to be overwritten to get stock performance and emissions. Something is wrong with hardware/mechanical that needs to be solved.
What he did was technically interesting, but probably not a great solution to the actual problem.
Yeah I had the same impression. There was plenty of data presented and code was dumped and reversed…but was the actual issue solved? What even was causing the issue? Presumably it passed smog sometime in the past couple years so what degraded? It was lean and there was some kind of issue with pressure. I can’t imagine a reflash or whatever is the fix.
It definitely feels like a classic case of the operator reinforcing Claude to keep pursuing a goal that’s counterproductive to a robust solution. Claude complied and they went hard on it but I can’t imagine the driver is any better off and more likely in a worse spot as the underlying issue wasn’t solved.
”You attach an assumption to a piece of evidence, you start to bend the narrative to support it. Prejudice yourself.”. I find by default while claude is great at analyzing logs and diagnostics from a variety of places it is really prone to this bad detective pattern of choosing an explanation and sticking with it even when evidence starts to bend away from it. you really have to play the role of the skeptic to do do good root cause analysis. Which is very hard to do in unfamiliar domains.
It's probably a bad MAF sensor but this guy thinks he knows more than he does and uses the king of incorrect confidence to try to diagnose the issue.
Yup. These agents have feedback loops that are self-reinforcing. Without technical background, a user is easily mislead.
And did the thumbnail, text and graphics. 1mn views with zero effort. I'm here for it
100 points for car hacking fun. -60 points for "not needing a mechanic or anything". A good mechanic would have fixed it up in a half hour. Guy took 3 days and god knows how many tokens.
I don't believe a normal 'good' mechanic would be able to re-map it in 30 mins (assuming the underlying issue wasn't something else).
There almost certainly IS an underlying issue, though. Unless the ecu spontaneously recapped itself.
Mechanics are overpriced.
Manufacturers are going to start cryptographically locking everything down rather than relying on the time and skill needed to do reverse engineering.
LLMs have been pretty good at breaking bad crypto in cars. In my Hyundai, they were able to figure out the right incantation to get the car to run unsigned code off a USB stick. Although I guess we're in trouble once carmakers use LLMs to lock things down even more
Most shop security is security through obscurity. For them to lock it down, they need to manage those secrets all the way back to a cloud identity provider. They can and probably will do that eventually.
I assume Tesla does that already.
Old news - for 20+ years in Europe, 5-10 years in the US, and now in Korean and Japanese cars too, ECM updates have been cryptographically signed for warranty, emissions, and now EU cybersecurity regulatory compliance. It turns out there are still bugs in things with cryptography too though.
They did. Cayenne 2024+ does this and ecu is essentially read only unless you start a diagnostic session online.
Love seeing unique youtube projects on here for a change!
... even better would be free projects.
I enjoyed watching this a couple days ago -- super cool as someone also into slightly older vehicles.
Oh hey I saw this post on Reddit, super cool!
"Everything is 'open source' if you read machine code."
That's been a tired old saw in computers since my career began in the 1990s. Who knew it would be so applicable in the age of AI?
Reminds me when I dropped a binary file from an eeprom and asked Claude why device would display "Error.". It figured out what the processor was, how it talked to LCD and then from there how it displayed messages and then listed conditions for which Error would appear. Amazing.