正文 · AI 翻译
IT Home reported on November 14 that, according to foreign media outlet Tom's Hardware on the 13th local time, an AI enthusiast in the United States, Muhammad Zurali, used AI and simple electronic wiring to get a 32-year-old 1994 Jeep Cherokee through California's emissions test.
The old vehicle's engine computer has very limited capabilities, but after connecting it to a laptop with an Arduino Nano development board, Zurali was able not only to read vehicle data but also gradually to gain control over some parameters.
A video he posted documents the entire process of tackling the emissions problem: repeated failed tests, successfully connecting to the engine computer and reading data, and then adjusting the engine to address excessive emissions. Throughout the process, Claude provided guidance, while Codex wrote the code.
The Cherokee had previously failed California emissions tests several times because its nitrogen oxide emissions exceeded the limit at specific speeds. Initially, the vehicle failed the test at 40 km/h. Zurali then adjusted the oxygen proportion in the air-fuel mixture, but unexpectedly made matters worse: it failed both the 40 km/h and 80 km/h tests.
Zurali admitted that he had previously been tuning almost blindly, without accurate data, and that each retest cost nearly $100 (IT Home notes that this is about 668.5 yuan at the current exchange rate).
Fortunately, AI is available to help in 2026, and besides AI, all that is needed is a $3 microcontroller (about 20.1 yuan at the current exchange rate); all the electronic components together cost no more than $15 (about 100.3 yuan at the current exchange rate).
The first problem was how to get a 32-year-old engine computer to communicate with a modern laptop. Claude found that the Cherokee uses Chrysler's SCI serial communication bus, with a communication speed of about 7?800 baud and electrical signal polarity opposite to that of a standard serial port. An Arduino Nano development board could handle the signal inversion, allowing the laptop to read vehicle data. Zurali mounted the board on a breadboard and connected several wires, three of which went to the Cherokee's 6-pin diagnostic connector.
The first connection succeeded, though initially the system could only record coolant temperature and battery voltage and plot how they changed. Zurali wanted more engine data and also hoped to adjust some parameters directly to address the excessive emissions.
Because it had mistakenly referenced data from other vehicle models, the system twice incorrectly concluded that the engine's air-fuel mixture was too rich and that fuel injection was 22% too high. Zurali then had Claude collect data during test drives, separately recording engine parameters and fault codes while the vehicle traveled at 40 km/h and 80 km/h to identify the cause of the problem.
Claude was also given permission to control parameters such as fuel pressure. With what appeared to be reliable diagnostic results and a solution, Zurali originally planned to flash the AI-modified program to the engine computer so the vehicle would pass all of California's emissions tests. But the Cherokee's engine computer was simply too old: it predated by a few months the versions whose firmware could be reflashed, and did not support the operation at all.
Instead, Zurali had Codex write custom code to a nearby EEPROM memory chip and run it directly. Zurali said frankly, “Claude did most of the diagnostic work, and Codex handled almost all of the memory work. I just did the wiring and drove the car.”
The entire process also required no help from a repair technician and no reliance on proprietary diagnostic equipment from a Chrysler dealer.
Zurali also uploaded the project used to diagnose and resolve the Cherokee's emissions problem to GitHub, naming it xj-scanner. The project uses the GPLv3 open-source license and is free for anyone to use.