Home » I Guess AI Can Fix A Jeep Cherokee XJ Now?

I Guess AI Can Fix A Jeep Cherokee XJ Now?

Ai Jeep Wrench Ts

I’ve received the following video so many times, I simply cannot ignore it anymore, and I must share it with you, dear reader. It’s about a young man trying to get his 1994 Jeep Cherokee XJ through smog, and using Claude to reprogram the rather primitive OBDI engine computer.

At least, that’s how I interpret the video uploaded last week by YouTuber Muhammad Zulali.

Vidframe Min Top
Vidframe Min Bottom

Here’s a look:

In the video, Zulali talks about the high Nox readings the smog station was reading from the tailpipe of his 1994 Jeep XJ:

Screenshot 2026 11 15 At 4.37.18 pm
Image: YouTube/Muhammad Zulali

He also talks about how he wired a $3 chip directly into the vehicle’s engine computer (ECU — electronic control unit), which looks like this:

Jeep Xj Ecu Large
Image: eBay/Goodautopartsaj

Zulali decided to ask the Large Language Model family known as “Claude” whether it can enable Zulali to read the Jeep’s engine computer live from his Mac; Claude helped him figure out how to wire up the aforementioned $3 chip to the engine’s computer.

Screenshot 2026 11 15 At 4.37.52 pm
Image: YouTube/Muhammad Zulali

Claude mentioned that Chrysler uses an old protocol named SCI.

Screenshot 2026 11 15 At 4.38.03 pm
Image: YouTube/Muhammad Zulali

Per Zulali, Claude let him know that he would need a translator to read the SCI protocol:

Screenshot 2026 11 15 At 4.38.16 pm
Screenshot: YouTube/Muhammad Zulali

So he wired up a cheap Arduino Nano on a breadboard:

Screenshot 2026 11 15 At 4.38.41 pm
Screenshot: YouTube/Muhammad Zulali
Screenshot 2026 11 15 At 4.38.28 pm
Screenshot: YouTube/Muhammad Zulali

From there, Zulali started using his Macbook to get live readings of what was happening with his engine’s combustion process. He was getting things like fuel pressure, fuel trim, and at one point Claude even actuated the fuel pump for Zulali to try to help diagnose the Jeep’s elevated NOx issue.

Screenshot 2026 11 15 At 4.39.17 pm

In the end, Zulali learned that his engine computer had a different fuel map at the specific engine driving conditions covered by smog tests, and this was causing low fuel and thus high NOx. [Note: I find this a bit odd. An ECU programmed to help the Jeep reduce emissions during CA testing is causing the Jeep to fail smog in that very state… -DT]. 

Screenshot 2026 11 15 At 4.43.31 pm
Screenshot: YouTube/Muhammad Zulali

Watch the full video to see how Zulali and Claude worked together to figure out that a 22% fuel excess that Claude initially suggested may have been causing the issue was actually not the problem (they figured out that a tool they’d pulled from the web was wrong), and watch how they used the CODEX coding assistant to resolve the 1/4% fuel deficiency resulting from the CA ECU pulling fuel at certain driving conditions. Watch as they actually send a tiny program over three wires over and over to actually write a program onto the 32 year-old engine computer.

I won’t pretend to be particularly well versed on all things related to computer programming, especially as it relates to the intersection of cutting edge AI computing and old-school OBDI software/hardware, so I’d love to get some more insight from you, dear readers, as many of you have been flooding my inbox with this rather interesting story. (Thank you!).

Top Image: YouTube/Muhammad Zulali

Follow us on:

Facebook
WhatsApp
Twitter
LinkedIn
Reddit
Subscribe
Notify of
80 Comments
Inline Feedbacks
View all comments
Michael Golden
Michael Golden
17 hours ago

I think the interesting part here isn’t the usage of Claude (although cool), it’s that the California model OBD-I had a special map to use while under the specific conditions of an emissions test. You know, like the thing that Volkswagen got in trouble for?

Widgetsltd
Member
Widgetsltd
12 hours ago
Reply to  Michael Golden

I don’t think that is what the guy is saying. He/Claude appears to be saying that the California calibration is leaner than the federal calibration in the load and speed areas that correspond with certain portions of the smog test. The car & its system would also perform in the same way if you actually drove within those load & speed conditions on the road. VW got in trouble for programming their ECM to recognize the specific conditions (more than just road speed & load) that represent the certification test, and then actually making the engine system behave differently ONLY WHEN it is operated under those very specific conditions.

Widgetsltd
Member
Widgetsltd
18 hours ago

I find the claim that he was readying fuel pressure from the ECU to be complete BS. Chrysler SMEC engine managment systems DO NOT have a fuel pressure sensor!

Widgetsltd
Member
Widgetsltd
12 hours ago
Reply to  Widgetsltd

OOPS – I meant READING fuel pressure from the ECU. I just watched the video and it seems to show that he was checking fuel pressure with a physical gauge attached to pressure test port on the fuel rail.

However, one must ask the vital question: If the ECM code for California emissions was the root cause of the smog check failure, how did this 32-year-old vehicle presumably pass the smog check 13 TIMES over its lifespan? This guy did not diagnose the root cause of the failure. If the video is to be believed, he applied a band-aid solution.

That said, it would be great if somebody would create an emulator for the old Chrysler DRBII or DRBIII scan tool which was what the dealer used on these cars when new. Even better would be an emulator for the engineering IBOX tool which was designed to alter the calibration parameters inside the ECUs that were used on the SBEC series engine ECUs. Maybe somebody could persuade Claude Code to do that!

El Jefe de Barbacoa
El Jefe de Barbacoa
8 hours ago
Reply to  Widgetsltd

Could absolutely do something like that with Claude, would be a whole lot easier with one to reverse-engineer.

Bill Loney
Bill Loney
18 hours ago

My Jeep runs on Fortran 77 – like a boss. Also because AMC was so broke they couldn’t afford an ECU better than a Speak and Spell.

Ben
Member
Ben
20 hours ago

Here’s the thing:

Someone published all the information Claude used to do that. They got zero ad revenue for it. Since “AI” doesn’t actually learn and has no ability to synthesize truly new content, what happens when it puts all of the human-written sources of information out of business?

The AI companies don’t care about this because by the time the chickens come home to roost they’ll have moved to their private island bunker full of underage girls, but the rest of us should care.

We won’t, because AI is easy and sometimes good enough, but we should.

Phuzz
Member
Phuzz
4 hours ago
Reply to  Ben

I’m somewhat torn on this. On one hand, I don’t like the idea of the AI companies making money* off other people’s work. On the other hand, a lot of the information the AI used in this case was probably scraped from Jeep forums etc. and the people who originally posted the information were never interested in being paid, they just wanted the information out there to help other people.
I know I shouldn’t attempt nuance on the internet, but I feel this is a bit more of a grey-area than a lot of AI use.

*(technically just losing money at a slightly slower rate)

PlatinumZJ
Member
PlatinumZJ
20 hours ago

Finally, a use for A.I. that doesn’t involve ‘art’ slop! I would personally be more concerned about tracking down the actual problem, but I’m also pleased to see someone doing work like this for an older vehicle.

ElectricMiataBoogaloo
Member
ElectricMiataBoogaloo
21 hours ago

What the heck is lowercase smog? That’s the stuff that gives me an asthma attack. Not all of us live where-ever you live. Around here we have emissions testing though, and thank goodness because then they have to spend extra money to re-tune and de-tune to roll coal on the EVs. There are some carve-outs for “classic cars” anything than 32 years but not driven more than 4,000 miles a year.

I wouldn’t trust the “glue on pizza” “eating rocks is good for you” machines though. They’ll tell you anything to keep that sweet subscription money coming.

Cars? I've owned a few
Member
Cars? I've owned a few
21 hours ago

“In the end, Zulali learned that his engine computer had a different fuel map at the specific engine driving conditions covered by smog tests, and this was causing low fuel and thus high NOx. [Note: I find this a bit odd. An ECU programmed to help the Jeep reduce emissions during CA testing is causing the Jeep to fail smog in that very state… -DT].”

I get a big whiff of dieselgate with this. But probably the manufacturers that didn’t are in the minority.

Mazda Garageda
Mazda Garageda
20 hours ago

That’s how I read this as well. Honestly a bit surprised (and impressed) they were already trying to cheat the tests by using different fuel maps back in 1994.

Bill Loney
Bill Loney
18 hours ago
Reply to  Mazda Garageda

Cheat? I think they were meeting the requirements then per the laws and the tests that they were required to meet. The VW thing was more elaborate and was an issue because VW misled buyers. Their vehicles passed emissions standards. That wasn’t the reason for the buyback.

Peter Andruskiewicz
Member
Peter Andruskiewicz
14 hours ago

It wasn’t uncommon for a different engine computer or emissions equipment to be used on CA vs non-CA cars, and they would have different maps for fueling, spark etc. My interpretation is that the maps for the CA model were different, meaning the engine would run in a presumably cleaner (or at least different) way any time the speed and load was in that region. Maybe this hurt fuel economy, maybe it was reliant on a CA-only EGR system or other piece of equipment that cost more, but for whatever reason it wasn’t implemented on non-CA cars the same way. However it always operates the same regardless of how the car is driven.

Diesel-gate was different because the ECU actually looked for “tells” to determine if the car was being driven on the specific time variant speed and load trace of the drive cycle, and if it was then it would run the engine and aftertreatment one way for minimum emissions, while if it wasn’t then it would run in a higher NOx but more efficient way. So it determined if it was being tested to know when to cheat.

Widgetsltd
Member
Widgetsltd
12 hours ago

Great point! This guy was probably putting a band-aid on the ECU to compensate for an inoperative EGR system that he did not know how to properly diagnose. Or, perhaps, his Jeep has a fuel pressure control issue – the description of the fuel pressure test results in the video is pretty suspect.

Peter Andruskiewicz
Member
Peter Andruskiewicz
2 hours ago
Reply to  Widgetsltd

Oh yea, while technically impressive I agree with a lot of the commenters that he’s avoiding fixing the root cause with a novel band-aid… seems like a case of “when all you have is a hammer, every screw looks like a nail”

CuppaJoe
Member
CuppaJoe
22 hours ago

Pshaw. I would only be impressed if he used a salami to program the Jeep’s ECU.

BigMike2k3
Member
BigMike2k3
19 hours ago
Reply to  CuppaJoe

…and made run DOOM!

LTDScott
Member
LTDScott
22 hours ago

Technically this is akin to installing a performance tuner and is illegal, but on a pre OBD-II vehicle it’s basically impossible to detect.

Also, this post is a good place to put this: https://www.facebook.com/reel/1107530411687016

Last edited 22 hours ago by LTDScott
Parsko
Member
Parsko
23 hours ago

This is a perfect use case for AI. SMOG sucks, and this old XJ barely out of spec barely being brought back into spec are not going to be the end of bad health in California or the rest of the country or world. This is good, and I’m glad they figured this out this way.

Give power back to the people and away from the dealers. This solution is not for the every day person who will still go to the dealer and have things done “legitimately”. It’s for folks like this person and us, who want to keep our borderline cars on the road a little more, which is better than having to replace it with another.

LTDScott
Member
LTDScott
22 hours ago
Reply to  Parsko

Or, he could have diagnosed it the old fashioned way, replaced the defective components, and had the same result without fooling the computer.

Mopar4wd
Mopar4wd
22 hours ago
Reply to  LTDScott

In the video he mentioned swapping the O2 sensor and a few other things that didn’t clear up the issue.

LTDScott
Member
LTDScott
22 hours ago
Reply to  Mopar4wd

I admittedly have not watched the video, but in theory if all components are working properly there should have been no need to fool the computer. He essentially put a Band-Aid on the problem.

I’ve been very active in the Californians for Classic Car Exemptions Facebook group which had a role in getting Leno’s law to pass, and so many times I have seen owners of older cars that won’t pass smog not even attempt to fix the root problem and/or just fire the parts cannon without proper diagnosis.

A buddy of mine couldn’t get his stock Fox Body 5.0 Mustang to pass smog and was even considering registering it out of state. It failed on HC and was clearly running rich and he insisted everything was working properly. I’m very familiar with the engine controls and components on that engine and told him there’s gotta be something wrong and made some suggestions. Sure enough, after investigation he found a vacuum leak. Fixed the leak and it passed no problem.

PresterJohn
Member
PresterJohn
23 hours ago

I saw the video over the weekend and I’m with you on doubting the cause given. This is a super cool use of AI though and it is extremely effective as a reverse engineering tool if you have some idea what you’re doing.

Frankly, this is one of the most pro-consumer uses of it – anything you own can be truly owned now if you can set it up properly for experimentation.

Just an example: my other expensive hobby is amateur radio. I have a few radios that can only be programmed with clunky Windows software from the manufacturer. Could I have reversed it myself? Sure, but I have 2 kids and don’t have the time even though I do software development for a living. You know what never tires and doesn’t need to put kids to bed? Claude. I now no longer need those programs and have my own versions that do exactly what I want and run easily on my Linux machines.

Paul E
Member
Paul E
23 hours ago
Reply to  PresterJohn

CHIRP didn’t cover your radios? I found that it worked well in programming a couple Yaesus I own, versus the clunky, expensive RT Systems software.

PresterJohn
Member
PresterJohn
23 hours ago
Reply to  Paul E

Hah a fellow ham! No not these ones, though adding it to CHIRP is what I did. Generally I use CHIRP for everything when possible.

Last edited 23 hours ago by PresterJohn
Baltimore Paul
Baltimore Paul
1 day ago

Just this weekend, I asked Alexa how many pounds was 37 ounces, and it replied it was 1 lb. 5 oz.
I asked how that was possible and it said oh whoops, I made a mistake.
I’m thinking if this is the AI future either we shouldn’t be worried at all, or we should be completely worried

Last edited 1 day ago by Baltimore Paul
Clark B
Member
Clark B
1 day ago
Reply to  Baltimore Paul

I once asked AI if onions were safe for cats, because one of ours ate at least two fried onion rings while we weren’t looking. Google’s AI said onions are fine for cats. They are not. They are toxic to cats, and you should contact a vet immediately if your cat ate them.

Fortunately the cat has a stomach of steel and had no complications or issues whatsoever. On more than one occasion, he’s grabbed a whole slice of pizza and tried to run off with it.

ChefCJ
ChefCJ
1 day ago
Reply to  Clark B

I heard the ‘Daily Dose of Cats’ music as I was reading this about your pizza thief

Bkp
Member
Bkp
21 hours ago
Reply to  Clark B

Off the top of my head, yes, onions are bad for cats, too much can cause anemia. I don’t know the threshold for what is a worrying amount, run that by your vet.

Somewhere, back when we just had film cameras, is a fine photo of one of our cats leaping away with their booty of a pizza crust in their mouth. The mighty hunter!

ElectricMiataBoogaloo
Member
ElectricMiataBoogaloo
21 hours ago
Reply to  Clark B

Yup, Google was telling people to put glue on pizza until the media exposed it. That particular issue got fixed becuaee of the attention but it’s all garbage in garbage out. The things don’t actually understand what they are spewing out, it’s all about manipulating the public by addicting them to outsourcing all the thinking.

ElectricMiataBoogaloo
Member
ElectricMiataBoogaloo
21 hours ago

It wasn’t actually difficult to look at a list of links to find one that was reputable to click on?! Oh Google, RIP ten blue links.

Bill Loney
Bill Loney
18 hours ago
Reply to  Clark B

Technically speaking – since your cat ate two and lived – was Google wrong? Your cat is fine as it said it would be. Maybe Google knows more about your cat’s intestinal fortitude than you think. It’s secretly scanning all living things in our homes so that when the robots want to get rid of all of us; it knows exactly what to do…. we’re so doomed.

Chartreuse Bison
Chartreuse Bison
23 hours ago
Reply to  Baltimore Paul

The free models that anyone can use infinitely like google search overview and whatever alexa uses are such garbage that I don’t know why they bother.

Baltimore Paul
Baltimore Paul
21 hours ago

I then asked to speak with a manager and Alexa said it would share my concerns with the programmers

Mazda Garageda
Mazda Garageda
20 hours ago
Reply to  Baltimore Paul

I recently had Google make an obvious math mistake as well. I then questioned it twice before it admitted that “oops you are right, I made a fundamental error.” At which point I stopped interacting with it, realizing that I had become it’s educator.

It was bad enough their AI mixes information from multiple websites to give you an answer that isn’t actually there when you read them separately. Simple math errors now are just the cherry on top.

Jon L
Jon L
1 day ago

Personal AI example: I am writing a paper about the times country A has attacked country B. AI, programmed to provide three bullets like all American school papers, gave me three invasions. The first two I knew well, the third was new to me. It had dates, it had operation names, it even had the name of the general that lead the invasion. There was no third invasion. It made up an invasion just so it would have three points.

This is my biggest fear with AI. Not that it will go Matrix, or Terminator on us. But that someone will ask it to clean up all the plastic in the ocean. Once the big chunks are cleaned up, it will see that there is micro-plastic in fish, and it will clean them up too; and then AI will determine that all life in the ocean has plastic in it, and it will “clean up” (kill) all life in the ocean. It will end life on earth not because its angry, but because it’s trying to complete a task it was assigned.

Tj1977
Member
Tj1977
1 day ago
Reply to  Jon L

I have repeatedly asked car blogs that I follow to correct mis-information so it won’t be scraped. I know it’s a futile attempt, but I really don’t want well-respected author at Hemmings (HEMMINGS!) spouting off that the 1973 Monte Carlo was just a re-fresh of the ’70 Monte Carlo…

It’s just the car history world, but I can’t tilt at all the windmills out there so I have to pick and choose.

Jon L
Jon L
21 hours ago
Reply to  Tj1977

Thank you for doing what you can!

ElectricMiataBoogaloo
Member
ElectricMiataBoogaloo
21 hours ago
Reply to  Jon L

Or, like the classic game, strip mine the entire solar system to produce the maximum number of paper clips. They are just trying to achieve the objective someone gave them! Now it’s one AI “training” the next with garbage and “synthetic” or made-up training data.

Username Loading....
Member
Username Loading....
1 day ago

Why not use AI to tune old cars, it’s being used to develop new ones.
I’ve been tempted to tinker with this a bit, though I’d want to try it out on an engine where I’m not particularly worried if it lives or dies. For something like a junkyard ls there is definitely more than enough reference material to diagnose any issues and get the fuel and spark maps dialed in.

D-dub
Member
D-dub
1 day ago

[Note: I find this a bit odd. An ECU programmed to help the Jeep reduce emissions during CA testing is causing the Jeep to fail smog in that very state… -DT].

Did you insert an editorial note into your own post?

Mark Tucker
Mark Tucker
1 day ago

Yes, but what’s actually causing the high emissions? Was the problem identified and the offending part(s) replaced? If not, it isn’t “fixed.” Nice try with the shortcut, though, Junior. Now break out the actual toolbox and do it right.

Dingus
Dingus
1 day ago

I’ve said it before, will say it again. LLMs are tools. They are software tools. You can use tools to do great things or you can do stupid and dangerous things with them. You can frame out a house with an air nailer, you could put a nail into your foot with one.

I’m so over the “AI is bad, you should never use it” trope that seems to get dragged out by the luddites. Sure, giant data centers are bad, but that’s a choice made by some rich pricks. You can just as well run a reasonably competent LLM on a modern computer, though a little slower.

Using it to modify values in an ancient ECU seems a little risky; I had Claude help me diagnose an issue with my old P2 Volvo where it was not boosting properly. I set up data logging for various ECU values like boost pressure, TCV control, O2 sensor, MAF, etc and graphed them while I drove and exposed the issue. I gave Claude the data and the graphs. It correctly pointed out a few things I already knew about wastegate issues (sticking) and a few things I was unaware of based on the data. I used the analysis to help me fix what was wrong and was successful.

As others have said, I don’t think it’s wise to give it full control and let it modify values, that goes towards responsible use. However, if you actually know what you’re doing, you can test such changes and then set up your LLM tools to run validation after incremental changes. You do have to know SOMETHING about basic function and use of these tools. Even if you don’t, you could literally tell it “don’t do anything that is risky, all changes made must be validate and if they fail, the changes are backed out”.

The reality is that with modern cars being so heavily dependent on a scatter of various compute units, trying to diagnose issues becomes extremely difficult. Using an LLM to at least retrieve, aggregate and summarize the disparate data into something meaningful is a huge help. If it can diagnose or even probe various components, it saves a ton of time. Instead of guessing that an ECU is bad or having a hard time, it can compose tests to probe and check responses. Sure as hell beats doing the old bullshit of “swap in known good” when everything is locked to VIN by the manufacturer.

To me, this sort of thing is what will allow independent mechanics to claw back some work rather than having to depend on dealers and their proprietary locked software to do basic diagnosis and repairs. That is until the manufacturers cryptographically lock every part of every system (which is happening more and more).

Alexk98
Member
Alexk98
1 day ago
Reply to  Dingus

Exactly this, Claude is an incredibly versatile and capable tool when used for data analysis and coding work. It and other LLMs are going to open up a huge amount of repair and diagnosis work that is currently under stranglehold by OEM obfuscation. These LLMs are able to reverse engineer AAA video games, they will absolutely be able to crack open locked down modules in cars. Stripping back AI marketing and looking at principles of functionality, this is the sort of tool that enthusiasts have been begging for for year in the wake of complex modern cars.

These tools will make tuning and modifications significantly easier too. Just think of the potential for at-home manual swaps on cars that have ECUs that “can’t handle” the lack of TCU data. An LLM and some basic hardware like this jeeps case will be able to crack into it, analyze the signals, and help with hardware implementation to spoof it or pull in a related signal (such as wheel speed and engine speed) and inject an equivalent sensor. Caution is needed and some people will blow up their cars, but that’s nothing new, the accessibility LLMs afford however IS new.

Beachbumberry
Member
Beachbumberry
1 day ago
Reply to  Dingus

Exactly. I think we may have discussed this before on another article but there is a guy using codex to read the encrypted canbus signals on an i8 to be able to figure out how to swap in a dsg and rotary engine but maintain the hybrid system. It’s wild. And that’s the whole point of LLMs. They are really good at connecting the dots if you know what the starting point is, the desired end point, and what it should look like in between.

JerryLH3
Member
JerryLH3
1 day ago
Reply to  Beachbumberry

That would be Couch Built on YouTube. His videos are great. He has also used a lot of 3d metal printing, another emerging and more widespread technology than it used to be.

Jdoubledub
Member
Jdoubledub
21 hours ago
Reply to  JerryLH3

Always appreciate a YouTube recommendation.

M SV
M SV
1 day ago

All this vibe coding is just having llms look up what’s done and try put things together. The people using llms to declutter TVs really cracks me up. It’s just running a script. You can do it yourself in about 3 minutes it takes the llm hours. It was interesting what he was saying about how the computer in his car wasn’t programmable because it was built 4 months before they were programmable. That and software people use was backwards for that particular computer. I guess if it helps people with fringe problems great.

Kurt B
Member
Kurt B
1 day ago

LLMs are now capable enough to write simple tools to talk to OBD.

At its core, everything is just a connector, electricity, and an impulse interpreter. Modern frontier models are, in fact, good enough to write tools for simple diagnostics work.

I probably wouldn’t let Claude write a fuel map though.

Mechjaz
Member
Mechjaz
1 day ago

LLM stands for “Large Language Model,” not Learning Language Model. Large is a reference to the number of parameters used for inference when choosing is this a good match, yes or no, for the next token to be used.

As to the YouTuber at hand, he got lucky and has no idea what he actually programmed to it. Or maybe he read the code, wrote unit and integration tests, validated against a mock ECU, …yeah he has no fucking clue what he wrote to his ECU. AI can generate, does generate, garbage faster than you can imagine, and I’ve watched otherwise skilled engineers let it rip and subsequently take down entire software environments. I’ve also seen true believers let it rip and take down other peoples’ software environments as well as their own.

Because I was forced to (and ended up losing my job anyway, a fat lot of good that was), I used AI. It does these things (personified for simplification, so I don’t have to retype “an LLM mimics the behavior of” over and over):

lies with perfect confidence: It will reason over a change, make that change, tell you in flowery language why it was the right change to make, and be completely futt-buckling wrong about the whole thing.generate garbage: code of little or no value, which is actually a net reduction to code quality. A test that it self fulfilling takes time to run, but that’s not even the worst of it – it explicitly states that there is coverage when there is none.have no concept of true context or root cause: As @TurboFarts ninjas me on this point as I type, it will paper over everything with anything, fixing nothing in the process. It’s the technological equivalent of spraying Febreze in a car full of trash while someone is still inside, smoking with the windows up. You might smell Fresh Linen for a second, but that’s basically an accident.This is the huge, intentionally under-discussed weakpoint of AI. It’s not the computer from Star Trek, which was a written character; it can never know what you mean, or indeed what anything means. It can only guess Yes or No for fractional chunks of text and data (tokens) billions of times per second (kind of). This is also part of why there’s “AI Voice,” written and spoken, because it homogenizes everything into the same average Yeses in the token output.Introduce errors of scope: I’m not just talking about the scary ones where they muscle out of weakshit “sandboxes” (in quotes because if they were really isolated, they wouldn’t really have been able to perform rogue actions. Sandboxing itself is real, of course). Since it has no idea what it’s actually doing, it has no qualms about pulling in dependencies on anything it thinks might be useful. It might even want to install entirely new and unknown-to-you software libraries and utilities on your machine, when in fact all it needed was some publicly available information from a website.Enable the unable: I’ve heard several engineers talk about how they hadn’t hand-written any code in over a year. These are people that move money in the economy, people that write safety software (pick a field), people that write websites you depend on and devices and services you trust. All automatically written, “reviewed” and approved by a self-fulfilling loop of AI agents. These people used to be engineers, but now they’re AI herders, and the VPs and up – as well as any true believers, at any level – could not be more delighted, because it has the shape and size of five people working in the place of one, but only costs as much as a person and a half. The thing is, there’s no way for one person to keep up with, on a line by line basis, what five “people” are putting out.
Source: I’m a software quality engineer, I’ve done firmware (not ECU) testing in the past, and I have personally seen and experienced every situation I just described. I have been told that reviewing code was too time consuming to do and I should be focused on using AI more, so yes, I’ve got a big fuckin axe to grind, but nothing I’ve said is untrue, either.

edit: formatting

Last edited 1 day ago by Mechjaz
Mechjaz
Member
Mechjaz
1 day ago
Reply to  Mechjaz

Ugh sorry guys. I forgot that the site is terrible and eats formatting even worse when you edit. Should have just left it alone.

Vetatur Fumare
Member
Vetatur Fumare
1 day ago
Reply to  Mechjaz

Whenever I think AI has a use, it usually ends up being because I am thinking of a field in which I have no expertise…

I will say that it has impressed me when analyzing boring, raw data. But that might be because I don’t yet know what it is missing.

TurboFarts
Member
TurboFarts
1 day ago

Seems like a band aid solution that doesn’t address the actual root cause.

It’s an old car and so a weak fuel pump, clogged fuel filter, clogged fuel injectors, failing 02 sensor, vacuum leak, bad MAF, and other issues could all cause lean condition and the resulting high NOx.

Occam’s razor thinks AI “solved” the real issue by masking it.

Edit: It is cool what he managed to do with very little cost and a lot of ingenuity.

Those jeeps also use a three way catalytic converter that reduces NOx, CO, HC. Could be an obvious cause. Not sure what OBD1 limits are and if this should have set it off…

Last edited 1 day ago by TurboFarts
Spyrius Robot
Spyrius Robot
1 day ago
Reply to  TurboFarts

Yeah, Claude is an idiot.

Max Headbolts
Member
Max Headbolts
1 day ago
Reply to  TurboFarts

Yeah, that’s the real issue here; something is wrong in the emissions system and changing the fuel map is allowing the vehicle to pass smog; for now.

Eventually it will fail again, and the root cause which is probably a combination of all of these things being out of spec, will return.

In short, AI didn’t fix this Jeep, it helped someone reprogram an old ECU; which is neat, but not really fixing things. Hopefully it’s not introduced a lean condition that will cause the engine to fail. Given the lack of actual measurement done here it’s hard to be certain.

Dingus
Dingus
1 day ago
Reply to  Max Headbolts

This is the case where you give software incomplete data and expect a complete solution.

If you are going to tinker with ECU values to achieve lower emissions, then unless you have instrumented both the ECU AND the resultant emissions in real-time, the software is guessing.

As you said, it could introduce a new problem in pursuit of achieving what the user asked for. Unless the software is given a comprehensive view of the whole system, there is no way for it to determine if there were harmful consequences.

You wouldn’t expect a doctor to perform surgery without medical imaging because it’s guesswork at that point. Nothing different here; people that use these tools need to understand what’s at stake and the consequences of a half-baked approach.

TurboFarts
Member
TurboFarts
15 hours ago
Reply to  Max Headbolts

Exactly, you said it better!

TheDrunkenWrench
Member
TheDrunkenWrench
1 day ago
Reply to  TurboFarts

Yeah, this is like the inverse bandaid solution that a guy I knew did on his CRX.

It was running too rich to pass, so a set of vise grips on the fuel feed, adjusted until it passed.

Did it fix the actual problem? No. But it DID get the emissions certificate (you only had to do it once when registering the car where I live, these days OBD reading is integral to the safety process, and non OBDII cars are exempt from it)

TurboFarts
Member
TurboFarts
15 hours ago

Inverse bandaid, I like that. I’ll be taking that for future use haha. Great temporary solution for emissions test I guess.

Mazda Mark
Mazda Mark
1 day ago

David, I suggest you start ingratiating yourself with the AI sooner than later before they become our compute-based overlords.

ClutchAbuse
Member
ClutchAbuse
1 day ago

I’ve used Claude to make myself some software tools. It’s incredibly powerful if you have a good idea about what you want and how it should work. I’ve also done hardware reflashing using a similar method to whats shown here. I think an age of AI assisted ECU tuning is definitely upon us.

Dave Larkman
Dave Larkman
1 day ago
Reply to  ClutchAbuse

I was involved in an OEM project that changed a single parameter of the engine calibration by a tiny amount to fix the cause of multiple development vehicle fires.

AI messing with that with no validation is terrifying.

DangerRoad
Member
DangerRoad
1 day ago
Reply to  Dave Larkman

My company’s president is a data driven engineer who uses Claude. He ran my long term cash flow though it, it concluded that I under predicting the cash balance by about 20k at the end of our fiscal year, which was fine. My take is Claude is over predicting by 20K. What worries me is that many using Claude will accept its conclusions as an absolute truth, such as the Titanic being unsinkable, and then act accordingly.

TurboFarts
Member
TurboFarts
15 hours ago
Reply to  DangerRoad

The people already accept whatever they hear on their favorite biased “news”station.

You can be certain your worry will become reality.

TheDrunkenWrench
Member
TheDrunkenWrench
1 day ago
Reply to  Dave Larkman

I mean, is it really that much different that Billy Backwoods tuning a factory ECU with HPTuners or the like?

Any time you go outside factory parameters, you’re rolling the dice. Whether it be by human or digital hand.

Ferdinand
Member
Ferdinand
1 day ago

I’ll have you know that Billy’s cousin is on a NASCAR pitcrew. He knows his stuff!

TheDrunkenWrench
Member
TheDrunkenWrench
1 day ago
Reply to  Ferdinand

Ah yes, Marlboro Merle, we’ve all met him.

Dave Larkman
Dave Larkman
5 hours ago

It exactly the same issue, but with even less control and with every change being potentially unique.

And with all the decisions being made by a thing that doesn’t know cylinder 6 always has worse air flow/what model year the con-rods had a cost-down/what it’s like to have an angry customer with bits of engine all over the floor.

Ferdinand
Member
Ferdinand
1 day ago
Reply to  ClutchAbuse

I’m an AI skeptic, but just this past weekend I had my first use where I really relied on it. Before this weekend, all previous interactions have been more for fun (“Make me a picture…” or some low level data/info gathering where it didn’t really matter if it was accurate or not).

I was setting up a new radio for my RC car; an open source one known for a steep learning curve but having powerful features if you know how to take advantage of it. I do not know how to take advantage of it. I really leaned on Claude to hold my hand and walk me through the whole process. Still took me about 4-5 hours but lots of that was because I often had to ask Claude to double check something.

The alternative would have been finding forum posts of people having the same issue and then spending probably 10-20 hours reading through them to find the answers. I know on the basis, that’s all Claude really did, and it’s also why it gave me wrong answers.

But I left pretty impressed, but still skeptical. The amount of various tools and software it just says to go out and download is staggering. Often times I would say, “Can I do this step in _____ that you had me download from earlier” and you often could. I bring this up, because you often find yourself installing stuffs from what seems like questionable sources, so how often are people downloading malware and the likes.

So, overall, I’m guessing I completed the process in about half the time I would have before. And I had to still try to learn what it wanted me to do before doing it, so I could be sure it was right. I also had to push back when it was wrong. 3/5 stars from me. Not perfect, but could have been worse. Can see how it would easily because a 1/5 situation if you just blindly followed it.

This isn’t to endorse AI. But simply say that you can use it. Just don’t be a lemming. You still have to learn.

Drive By Commenter
Member
Drive By Commenter
1 day ago
Reply to  Ferdinand

OpenTX or EdgeTx? I have two Radiomaster TX’s. Both on EdgeTx. Amazing what it can do.

Ferdinand
Member
Ferdinand
23 hours ago

EdgeTx for a Radiomaster MT12. I’m getting back into RC cars after being out for ~15 years. I’m going from frequency crystals to this.

Drive By Commenter
Member
Drive By Commenter
22 hours ago
Reply to  Ferdinand

Try downloading Companion and playing with it. Yeah it’s meant for the twin stick TX’s. The basic workflow is the same. Seeing it on a larger screen and not reliant on the scroll wheel is a major advantage. I cut my teeth on a TX16 so the MT12 was easier.

Ferdinand
Member
Ferdinand
22 hours ago

My real problem came when I tried to get a custom startup splash screen. Which requires changing the whole firmware (forgive my poor terminology if I’m wrong, like I said, I don’t know shit about this stuff and leaned heavily on Claude). And the flash went bad, so I had to recover. And what I used to recover didn’t have a couple of the built in tools that came on it so I then had to add those back…

Drive By Commenter
Member
Drive By Commenter
21 hours ago
Reply to  Ferdinand

Oof! Sounds like you had to get a little into the weeds.

I haven’t messed with the boot loaders. Only to flash new firmware. Just be aware the transmitting chips (module in ETX-speak) have different firmware than the TX board. Either ELRS or the multi module, depending on which one you purchased.

Right now my air TX has different firmware on its external module than the internal module to fly cheap planes. The internal module is on an older stable build to fly my spendier planes.

Ferdinand
Member
Ferdinand
21 hours ago

I think I took care of that, but I got so lost in the sauce that now I’ve got to check when I get home.

Went ELRS since I only had two other cars that I was ok buying new receivers for to get them using the same transmitter. Already moved over one, will tackle the other later on.

Drive By Commenter
Member
Drive By Commenter
21 hours ago
Reply to  Ferdinand

I have a lot of different protocol receivers so having one TX to talk to them all is nice. I haven’t hopped onto the ELRS quadcopter yet. My existing stuff for air and surface works fine.

ClutchAbuse
Member
ClutchAbuse
1 day ago
Reply to  Ferdinand

Claude shines when you already have a good idea about how to do something. If you don’t, it will make a ton of assumptions to fill in the gaps. I’m a software designer and am used to working with teams of engineers to make products. I treat Claude like an engineering team. I give it feedback, and provide notes and mockups for features. It does a shockingly good job building from mocks! There’s also tons of iteration to get things just right. I only use it to make productivity tools.they either work as expected or they dont. I would never outright trust it with customer data or financial information without extensive engineering oversight.

ElectricMiataBoogaloo
Member
ElectricMiataBoogaloo
21 hours ago
Reply to  ClutchAbuse

And it’s being pushed for medical diagnoses, therapy, and to tell cops who to arrest. The errors are piling up and some of them are fatal.

ClutchAbuse
Member
ClutchAbuse
18 hours ago

Yeah, it’s being shoved into things it has no business in.

80
0
Would love your thoughts, please comment.x
()
x