Tesla FSD v12: The Quest for General Driving Intelligence
07:31 Share

Hey, it's Dave. So today I want to talk about Tesla's big shift in their approach to FSD. But more specifically, I want to explain two concepts that I think will help understand why Tesla had to change their approach with FSD, and more importantly, what it will allow them to do with their new FSD approach. Here are the two concepts. Number one, Narrow Driving Intelligence, and I'll call it NDI for short. And number two will be General Driving Intelligence, or GDI for short.

So let me explain these two ideas and I think it'll help unlock what Tesla has been doing but also their direction for the future with full self-driving.

So first, let's try to define what narrow driving intelligence is. Narrow driving intelligence is an approach to autonomous driving where you try to solve each edge case as a case-by-case problem. So let's say you face an extremely difficult edge case in driving with narrow driving intelligence. The vehicle is then programmed with a specific solution to try to handle that edge case. So

So for every challenge or every problem, you're creating a separate algorithm or a separate solution to try to fix that problem.

However, this narrow driving intelligence approach has some major limitations. And the biggest limitation is that it doesn't really improve its overall ability to handle new edge cases that it's never seen before. In other words, its overall driving intelligence or its general intelligence is not really improving. It's just able to handle more and more specific edge cases because these edge cases are being solved through specific algorithms. One analogy is the game of whack-a-mole.

Every time an edge case pops up, what Tesla's been doing in this old approach of narrow driving intelligence is they would try to whack that edge case by coming up with a specific solution. However, this was a never-ending cycle. You solve one problem or one edge case and another one pops up. So imagine you're solving thousands of these edge cases only to find out that there's another thousand edge cases that you need to solve.

It's almost like a never-ending game of whack-a-mole. No matter how many edge cases you've solved, you're always going to have more that you need to solve. The big weakness of this narrow driving intelligence approach is that the AI in this approach doesn't learn from its past experiences in a way to extrapolate that learning into new novel situations.

In other words, narrow driving intelligence is about this big collection of distinct non-generalizing algorithms that are trying to solve individual use cases but are somewhat isolated in their functions. Now let's dive into the second concept which is general driving intelligence or GDI. Now what GDI or general driving intelligence aims to do is not just take a look at each individual edge case and come up with a solution that only works in an edge case, but

but rather it tries to develop a generalized driving intelligence. So an intelligence that is able to drive in various situations, but also able to handle new and novel edge cases that is never seen before. And this is because general driving intelligence is able to take what it knows with driving and extrapolate it across various situations and domains. So how does a generalized or general driving intelligence approach work?

First off, we need to get rid of all of the code-based algorithms and rules that the software has. The narrow driving intelligence approach has been to create various rules for each different edge case and then to continue to build on those rules as you handle more and more edge cases. The problem is FSD just relies on those rules and doesn't really ever develop a general driving intelligence that it can apply to new and novel and complex situations it's never seen before.

So number one is you get rid of all of your rules and all of the specific individualized algorithms for these specific edge cases. But then you're left with an AI that is not able to handle all of these edge cases and complex situations. So what do you do? So now without the crutch of rule-based code, you create a massive neural net and you don't give it specific rules for specific situations. Rather, you feed this massive neural net millions and millions of videos of driving in various situations.

And the neural net analyzes all of this video and learns from each video how driving is done, but it learns in a more holistic and comprehensive way. It's kind of like someone understanding the fundamental principles of physics rather than just memorizing individual equations.

In other words, by feeding it millions and millions of videos, you're teaching the neural net to recognize patterns and to learn the essence of driving, similar to how a human would learn. The result is this neural net becomes a general driving intelligence AI, and is able to drive in situations not just that it's seen before in videos, but the AI is able to take its learnings and use them to intelligently navigate through situations and edge cases that's never seen before.

This is largely because it's not relying on thousands of lines of hard-coded rules. Rather, through billions of parameters or viewing angles, the neural net gains general driving intelligence from its training data. Here's one analogy of how to look at it.

it so let's say you solve 1000 edge cases with narrow driving intelligence and then you discover another thousand edge cases you're gonna have to spend a lot of energy solving right those edge cases and then there's another thousand edge cases and another thousand and it goes on and on and on on however with general driving intelligence you solve a thousand edge cases through data you feed into the neural net not through individualized algorithms or rules and

Then you discover, let's say there's more edge cases, but the neural net is automatically able to handle, let's say over half of them because it has general driving intelligence. With the remaining edge cases, you feed it more training data to solve those edge cases and you're building more and more general driving intelligence in your AI so that that can be applied to the next novel edge cases.

So what ends up happening is that the ability of your neural net to handle new edge cases that's never seen before increases over time so that the number of edge cases actually go down. And the result is that you have a truly intelligent driving AI that is able to generalize its intelligence across various driving situations. I think that's the holy grail of autonomous driving is to see an AI that's able to handle extremely complex edge cases that's never seen before

and to handle them like a human would, meaning taking all of the driving knowledge and experience that they have and applying it in this new situation, right, to successfully and safely drive through it. So if Tesla's able to successfully navigate this transition from narrow driving intelligence to general driving intelligence, it's extremely promising in my opinion. It potentially unlocks a whole new paradigm in autonomous driving and potentially unlocks a whole new world of abilities for FSD in the future as well.

Now, I'm not going to predict exactly when FSD is going to become unsupervised and when exactly robotaxis are going to roll out. That's not the point of this video. Rather, I wanted to share from a bigger picture point of view, the evolution of Tesla's approach toward FSD and why I think their new version 12, which uses an end-to-end neural network approach, is significant and probably more significant than most people realize. It's laying the foundation for a true general driving intelligence AI.

And probably when we look back, we'll see that this version 12 was the big turning point for FSD that allowed Tesla to accomplish what most people thought was not going to happen. All right, I hope this has been helpful. We'll see you guys in my next video. Thanks.