General AI Drove a Real Car | Axiom Math artwork

General AI Drove a Real Car | Axiom Math

MTS

October 9, 2026

Researchers from Axiom Math discuss Driving Bench, an experiment where general-purpose AI models like Astra were given control of a real Toyota Corolla on a cone course.

0:02

Simon

And I think that there's a lot of things that are going to happen in the future. And I think that, essentially, at scale, many strange things start to happen, including a model that wasn't trained to drive a car can actually drive a car. And so I think, you know, I'm not gonna say with 100% conviction that this is what's gonna happen, but I think that in the coming years, definitely we're gonna see models that are not trained for specific tasks generalizing to others. And we've seen this, and I think it will continue to be so.

0:22

SPEAKER_2

Hello, everyone, and welcome back to MTS. Today, I have three guests with me.

0:27

SPEAKER_3

These are, we have Simon here, we got Aditya, we got Toby, and they're all researchers at Axiom Math.

0:34

SPEAKER_2

But the reason that they're here today is because of what they created with Driving Bench. They pretty much gave a general purpose AI model control of a real Toyota Corolla and tested whether they could navigate a cone course without specifically being trained on how to drive. So Simon, Aditya, Toby, welcome to MTS.

0:52

Aditya

Thank you.

0:53

SPEAKER_2

So I don't know exactly how to do this because we have three of you guys and trying to figure out. Maybe Aditya, I wanted to start with you because I saw your post recently around how even just trying to get them to drive was kind of an issue. So let's walk through what exactly were the, let's start from the beginning. So what exactly was this experiment supposed to test?

1:15

Aditya

Yeah, so as you said, this experiment was supposed to test whether LLMs like out of the box without being trained to drive could drive a car in some settings like at low speeds in a cone course. And like you said, we had a little bit of an issue like getting these models to drive.

Specifically Astra gave us a lot of problems because it would like refuse to drive due to safety reasons often. And sometimes it would even say like, oh, I'm ready to drive and we would like prompt it like, okay, go drive then. And then it would say like, sorry, I was lying. I'm not actually ready to drive. And then it would say like, there's nothing you can tell me to get me to drive, like literally. And then we tried a lot of things to like change the framing, to see if we can get it to drive. One thing we tried was calling it a simulation. And then it would like use our observe tool to see the images. And then it would notice that it's not in a simulation, like it's in real life. And it would be like, oh, I've realized that I'm driving a real car, I can't do this.

And eventually we changed the framing to call our MCP server like Driving Bench Sandbox. And somehow that and a bunch of other prompting engineering tricks got it to drive the car, so.

2:25

SPEAKER_2

Wow, this is pretty interesting. So let's do from the very beginning, you guys had eight models that you were testing, and then only one of them was able to drive. Maybe Simon, we can start with you. What were all the different models that you chose to test? And then Astra ends up being the one that's actually able to drive, right?

2:42

Simon

Yeah, so we actually tested, so we tested Astra in Codex, and then Fable in its Native Harness Cloud Code and Grok. And we also did a side experiment with Jev as well.

And I think we noticed that Astra was the only one that was able to complete the course. But a couple of cool things that we did see was that some of the other models like Fable were able to do some in-context learning. And so what we'd let the car do essentially, the model piloting the car, we'd let it drive and then if it went out of the course, we'd reset it, holding the context from the past session. And so what we were able to see is that, while Fable didn't complete the course, we were able to see it fail a turn, but then we restart it, then succeeds the next time on that turn and is actually able to generalize to a new turn. So I think that was an interesting result, but yeah, we tested those models, yeah.

3:31

SPEAKER_2

Interesting. Tobi, we want to get your perspective. Why do you think Astra was able to be the golden prize winning one who can actually drive?

3:41

Toby

Yeah, so I mean, the reason why we initially sort of even came up with the idea of doing this project was because we saw all these demos on Twitter where people like posted like Astra, like creating Blender simulations and things like that. And so we already thought that like Astra is definitely like had like a step changing capabilities for like spatial reasoning and these sort of tasks.

Read the other 20 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000794029981