mpshizzle
Well-Known Member
- Thread starter
- #1
You guys... UHF is opening an exciting door. Over the last few days I've put in about 150 miles trying it out. It's rough around the edges, for sure. But it's also showing a lot of promise. In a weird way, I'd say it feels less like they're writing code and more like taming an animal.
Based on my initial impressions, the best way I can think of to describe UHF is that it's like they loosened the reins on their LDM (Large Driving Model). It seems like they’re slowly taking more and more of the guard rails away as it gets a bit smarter. But... the funny thing about AI models is you can't really just turn behaviors on and off. There's no line of code that says "If there are no lines, hand control back to the driver". In this initial release of UHF it's been explicitly told to not work on unmarked roads. But, once or twice I've actually seen it do just that. (To be clear, I don't recommend you try this. When this did happen it did OK, but not great on these unmarked roads). Or when Rivian was giving demos of their Point to Point prototype at autonomy day it was told to not turn right on red. And for the most part, it didn't! But on some of the test drives (not mine), it took the right on red anyway.
So it's definitely a battle Rivian is having to fight here. You want to give the AI a long enough leash that it can figure out real world situations like crossing the yellow line to get around a delivery truck. But at the same time, you need it to follow rules and drive safely. Even though they have loosened the proverbial reins a bit in UHF, they clearly are taking a very cautious approach. It feels like they are keeping a lot of guard rails up until they are sure the system can safely perform that task. For example still restricting auto lane changes to the freeway, or not stopping for stop lights and stop signs.
But this is where my curiosity has been piqued. (And also where I would love some input from the community)
I'm very curious what the model is really capable of. I want to know what things the model technically IS capable of, but there's just a guard rail preventing it, and what things the model just ISN'T capable of. So let me know.. What are the scenarios you are curious about? What cases do you want me to test? Or if you have encountered an interesting situation I should investigate, let me know! I don't think I'll be able to find answers to all of it. But you can bet I'm gonna do my best to find as many answers as I can!
If you're interested in going a bit deeper on the topic, you can watch the video about this here:
Based on my initial impressions, the best way I can think of to describe UHF is that it's like they loosened the reins on their LDM (Large Driving Model). It seems like they’re slowly taking more and more of the guard rails away as it gets a bit smarter. But... the funny thing about AI models is you can't really just turn behaviors on and off. There's no line of code that says "If there are no lines, hand control back to the driver". In this initial release of UHF it's been explicitly told to not work on unmarked roads. But, once or twice I've actually seen it do just that. (To be clear, I don't recommend you try this. When this did happen it did OK, but not great on these unmarked roads). Or when Rivian was giving demos of their Point to Point prototype at autonomy day it was told to not turn right on red. And for the most part, it didn't! But on some of the test drives (not mine), it took the right on red anyway.
So it's definitely a battle Rivian is having to fight here. You want to give the AI a long enough leash that it can figure out real world situations like crossing the yellow line to get around a delivery truck. But at the same time, you need it to follow rules and drive safely. Even though they have loosened the proverbial reins a bit in UHF, they clearly are taking a very cautious approach. It feels like they are keeping a lot of guard rails up until they are sure the system can safely perform that task. For example still restricting auto lane changes to the freeway, or not stopping for stop lights and stop signs.
But this is where my curiosity has been piqued. (And also where I would love some input from the community)
I'm very curious what the model is really capable of. I want to know what things the model technically IS capable of, but there's just a guard rail preventing it, and what things the model just ISN'T capable of. So let me know.. What are the scenarios you are curious about? What cases do you want me to test? Or if you have encountered an interesting situation I should investigate, let me know! I don't think I'll be able to find answers to all of it. But you can bet I'm gonna do my best to find as many answers as I can!
If you're interested in going a bit deeper on the topic, you can watch the video about this here:
Sponsored
Last edited: