Tech Workers Made ChatGPT Drive a Toyota Corolla 52
An anonymous reader quotes a report from 404 Media: Some tech workers in San Francisco hooked up ChatGPT to a Toyota Corolla and got it to drive around a parking lot. This may not sound impressive in the age of the self-driving car until you realize that the LLM driving the car was a general purpose chatbot running on a laptop and not the purpose-built driving system based on millions of hours of training data used by Waymo or Tesla.
The people behind the stunt call themselves DrivingBench and they've posted their code, their prompts, and videos of the experiment online. In the experiment, DrivingBench hooked up GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol to a Toyota Corolla and gave the LLMs control over the car's steering, accelerator, and brakes. Then they set up a small cone course in a public parking lot and prompted the chatbots to navigate the space.
The goal, they said, was to find out if untrained, off-the-shelf frontier AIs can drive a real car. The results were mixed. Grok, Sol, and Fable only drove a few meters and didn't finish the course. GPT-6 Astra, however, did eventually learn to navigate the course and complete it. But not without a lot of troubleshooting. "It's an interesting demonstration of potential emergent capabilities," said Mahns. "It can fail a turn, and then the next turn, next try, it will be able to do that turn and other turns that it hasn't seen before. Not necessarily saying LLMs are gonna put Waymo out of business or something, just an interesting demonstration of where things are, and things are just moving fast."
"Most people that are familiar with ChatGPT have used it in this white collar, kind of like in the computer, and it's definitely imminent to the point that this capability will start having more impact in the physical world. So I think seeing the sparks of this happening, the shift of this utility coming into the physical world, is also somewhat exciting."
The people behind the stunt call themselves DrivingBench and they've posted their code, their prompts, and videos of the experiment online. In the experiment, DrivingBench hooked up GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol to a Toyota Corolla and gave the LLMs control over the car's steering, accelerator, and brakes. Then they set up a small cone course in a public parking lot and prompted the chatbots to navigate the space.
The goal, they said, was to find out if untrained, off-the-shelf frontier AIs can drive a real car. The results were mixed. Grok, Sol, and Fable only drove a few meters and didn't finish the course. GPT-6 Astra, however, did eventually learn to navigate the course and complete it. But not without a lot of troubleshooting. "It's an interesting demonstration of potential emergent capabilities," said Mahns. "It can fail a turn, and then the next turn, next try, it will be able to do that turn and other turns that it hasn't seen before. Not necessarily saying LLMs are gonna put Waymo out of business or something, just an interesting demonstration of where things are, and things are just moving fast."
"Most people that are familiar with ChatGPT have used it in this white collar, kind of like in the computer, and it's definitely imminent to the point that this capability will start having more impact in the physical world. So I think seeing the sparks of this happening, the shift of this utility coming into the physical world, is also somewhat exciting."
inputs and outputs (Score:4, Interesting)
set_motion(right, 60 %, 1.5 m/s, 10 s)
They translated those functions into input for some aftermarket car controlling hardware [comma.ai], which looks kind of cool but I didn't know was a thing.
I couldn't figure out from their report how they returned data back to the LLM, but apparently they pasted images and telemetry into the chat window every so often, or something like that.
Re:inputs and outputs (Score:5, Funny)
Probably the next code that it outputs will be:
# You're out of uploads for today, so if you want me to stop this vehicle anytime soon, upgrade your plan now!
Re:inputs and outputs (Score:4, Insightful)
That is why this is a ridiculous joke. The hardest part of AI is actually perception, and this used human perception. Maybe they can improve their hack by embedding electrodes in their own heads to provide the perception input to the chatbot.
They "made" it drive a Corolla? (Score:2, Offtopic)
Oh no. You can't make that bitch do anything it doesn't want to do.
Oh my, what to say (Score:3)
If that's not the absolute worst idea in the world, it's certainly in the running.
Re: (Score:2)
Re:Oh my, what to say (Score:4, Funny)
If that's not the absolute worst idea in the world, it's certainly in the running.
Maybe their next test will be to give it a shotgun and see how good it is at duck hunting...
I look forward to their next ChatGPT experiment (Score:5, Funny)
It's fine if you do it in a parking lot. It will be easy to hose off afterwards.
Re: (Score:3)
Joking aside, I used to think sci-fi stories about AI teaching itself to operate machines and systems was a bit outlandish. Turns out I was wrong, a generic chatbot can already drive a car, albeit badly.
We are probably screwed. If climate change doesn't get us, AI will.
Re:I look forward to their next ChatGPT experiment (Score:4, Insightful)
We all used to think it had to be intelligent to be dangerous. Turns out all it needs is just to be more complicated.
Re: (Score:3)
I think anyone how serves a cat probably knows that AI doesn't need to be particularly complex or intelligent to take over the world. Cats don't have object permanence or the ability to reason, and they still easily get humans to work for them for free.
Re: (Score:1)
>Cats don't have object permanence or the ability to reason
You must have lived with exceptionally stupid cats if you think that's true.
Re: I look forward to their next ChatGPT experimen (Score:2)
Sad what happened to Toyota (Score:1)
They hired some uninspired losers to lead their self-driving efforts. They basically were weak, looked into it and figured it's impossible. These types of people shouldn't be put in charge of anything. Seriously, it's like a lesson they should be teaching in MBA school or whatever. Reference: https://spectrum.ieee.org/qa-t... [ieee.org]
Re: (Score:3)
They hired some uninspired losers to lead their self-driving efforts. They basically were weak, looked into it and figured it's impossible.
There's nothing wrong with knowing your own limitations and acting accordingly. Had Mr. Zuckerberg had that sort of self-knowledge, he might have avoided lighting 80-billion-with-a-b dollars on fire and renaming his social media company after a technology that nobody wanted.
Re: (Score:1)
They should have quit their job and let someone else do it. Self driving cars are clearly possible .. Waymo, Zoox, Tesla and other companies (including a few in China) have done it. Even Mercedes is launching it next year.
As for Zuckerberg .. you're right .. but look Elon found the right people to make re-usable rockets and self driving cars work. As did Waymo, Zoox and a bunch of Chinese companies. Clearly there are people out there who do know how to do it.
Re:Sad what happened to Toyota (Score:4, Insightful)
1) Facebook stopped making new things, pursuing enshittification as the primary strategy.
2) Facebook pissed off Democrats with Cambridge Analytics, and pissed off Republicans with censorship.
3) Doesn't matter how much of that was true, both parties were pissed off and threatening regulation/antitrust at Facebook in a bipartisan way.
4) Facebook promised COOL NEW TECHNOLOGY and name change. They also scaled up political donations in a bipartisan way.
5) The threats largely went away.
6) The COOL NEW TECHNOLOGY went away.
"A real car" (Score:3)
No, it's a Toyota Corolla, it is the most boring, most lame yet most indestructible thing with wheels.
Re: (Score:3)
The LLM, being offended by the idea of driving a Corolla, promptly drove itself to a nearby dealer and traded the Toyota in for a Bentley.
Re: (Score:2)
Using Altman's credit?
next experiment (Score:1)
Halfway to skynet! (Score:3)
If they didn't also put up cardboard cuts-outs of people and award it points for running them over then it's like they aren't even trying to make skynet.
Talk about really bad ideas (Score:4, Insightful)
The last thing you want to control dangerous machinery is a control system with a tendency to hallucinate.
Re: (Score:3)
Or idiots that have no clue but pray to the great machine God. Like the person that moderated this down.
Re:Talk about really bad ideas (Score:4, Insightful)
Moderation on slashdot feels like it's being manipulated by bots whose goal is to protect certain stock prices, like whole classes of them in some cases, or specific companies in others, based on what's otherwise inexplicable moderation. It's difficult to separate this from the just pure trolling moderation but this is still my perception. If you criticize e.g. AI or cryptocurrency the chances of a downmod go way up.
Re: Talk about really bad ideas (Score:3)
It warms my cockles to know how mad the Russians and Nazis get at me. Eat shit and live to enjoy it.
Re: (Score:2)
Personally, I find the existence of these people just sad. But yes, pissing them off is a service to society. The last thing we want is have them accepted and accommodated. These people are the enemy of anything civilized and they serve only their own narrow interests.
Re: (Score:3)
Agreed. Although probably the same "bots" tend to post really stupid comments as well. Hence it is difficult to say whether these are bots or just "useful idiots".
Re: (Score:2)
Yet humans drive cars every day...
Re: (Score:2)
Yes stupid things are done by many people every day. Your point? Oh, you do not have one. I see.
Re: (Score:1)
Because we can drive, chew gum, talk, and check out your mom all at the same time in real time like its nothing, on 16 watts. Try that with any AI.
Re: (Score:2)
I think the last thing you want is something that has to generate long strings of text to turn or brake suddenly.
Re: (Score:2)
And that is not what is being done. At all. Get a clue about self-driving tech.
Re: Talk about really bad ideas (Score:2)
Yes, it does:
set_motion(direction="left",steering_percent=50, speed_mps=1.0, duration_s=10,reason="Cone gap opens to the left; ease in.")
How else did you think an LLM would be able to control it?
Re: (Score:2)
Please stop trolling. It is annoying. Anybody with the least bit of clue know LLMs are not used in self-driving tech.
Re: Talk about really bad ideas (Score:3)
Did you look at the article you are commenting on before saying this?
Re: (Score:2)
These people made an experiment. They did not use self-driving tech.
Incidentally, a text-string interface would not be where this fails.
Re: (Score:2)
That said, in the context of the story, everything they do is a bad idea. Hence my apologies.
Re: (Score:2)
wow I mean I get the whole 'screw reading the article' approach, but to not even read the article's title where they specifically say they used ChatGPT to drive a Corolla. That's a whole new level of avoidance.
Re: (Score:2)
I was mentally in the wrong thread. My apologies.
Re: (Score:2)
The last thing you want to control dangerous machinery is a control system with a tendency to hallucinate.
practically all cars are currently controlled by such bio-electrical systems; 20-50m people worldwide are injured per year, ~1m die, its the leading cause of death in children and young adults, and costs most countries ~3% of their GDP (source [who.int])
therefore it's a matter of performance; a matter of which system hallucinates less, and bio-electrical lot have low marks.
Re: (Score:2, Offtopic)
However, it seems that automakers are too stupid to understand this, except Tesla
HJAHAhaHAHHAhaHhAHAHAHAHHAHAHAHHAHAH
which is why nobody but Tesla wants FSD
AHAHHAHAHHAHAHAH
My only question is, are you this stupid, or do you think we're this stupid?
In other words, like people learning to drive (Score:2)
GPT-6 Astra, however, did eventually learn to navigate the course and complete it. But not without a lot of troubleshooting.
No different than when people first learn to drive. Beween learning how to shift gears while not stalling the car, how much to accelerate, how much to turn the steering wheel, when to turn the steering, and so on, people first learning to drive require a lot of troubleshooting. Each time improves the quality of driving (you hope).
running on a laptop.... (Score:3)
"a general purpose chatbot running on a laptop"
Lists a bunch of commercial LLMs in the cloud.
Insure that? (Score:2)
Re: (Score:2)
That's not the issue - insurance companies have no problem giving insurance to humans which have "individual learning capabilities."
The problem is that while insurance companies can extract compensation from humans - essentially they can coerce the humans to work "for them" to pay them back - they can't get compensation from the AI, and there isn't a good framework to get compensation from the AI developers.
The issue isn't that the AI learns - the issue is that our legal framework doesn't have a good way to
missing something (Score:2)
Any motors, actuators, like where, how? (Score:2)