While tesla autopilot has had plenty of updates over the past few years, the current version definitely isn't ready for full self-driving. Quite yet, however, tesla is planning for a massive software update coming in the next two to four months, which elon musk believes will be the key to full self-driving in 2020.. In this video, i'm gon na unveil the evidence pointing towards the new features in tesla's upcoming autopilot, update and explain why all these features make elon musk, confident in full self driving by the end of 2020.. I guess i remain confident that we will have the functionality for the basic functionality for level 5 autonomy complete this year, if you're new to the channel.

Please consider subscribing for more content like this and let's get right into it with the current version of the tesla autopilot tesla won't be able to achieve full self-driving in a short period of time. There are significant foundational issues that make it difficult to improve the safety of the autopilot on dred karpathy. The senior director of artificial intelligence at tesla explained how tesla's new update will focus more on the artificial intelligence aspect rather than the explicit decision making. Typically, when the tesla autopilot sees a stop sign, it stops at the sign.

However, this isn't going to work out because there's always certain outliers that make it impossible to achieve level 5 autonomy. Instead of this simplistic approach, tesla is utilizing its data set to monitor actual human reactions and copy those reactions for each scenario. This change in tesla's approach is called software 2.0, which is slowly taking over software 1.0 in tesla's upcoming software update, the company will be done or almost done with a software 2.0 completely getting rid of the previous foundation of autopilot in the following clip. Andres carpathi explains this rewrite using a term called neural network.

Nero networks refers to the ability to understand what an object or sign is, despite the different angles and lighting and degradations in our brain, we all have the ability to spot what a number is. Despite the differences in handwriting or pixel locations, this is what the neural network is trained. To do. Um, so basically, what's happening in the team is when i joined, we had small neural networks, doing some detections, and then these were stitched up and firmware.

Instead of the software 1.0, the c plus plus code and so on, and basically the neural network stack has been taking on more and more of the functionality. So i've shown this example or this visualization a few times in many of my previous talks, where this is sort of the autopilot software stack on the bottom. You have implicit stack, primarily vision and on the output we have steering and acceleration and roughly you can think of two code bases hidden inside the software stack. We have what i refer to as a software 1.0 code, which is good old fashioned c, plus plus explicitly designed engineered by person.

Person writes the code, and then you have what i call software 2.0 code, where the the code is an outcome of an optimization. It's a it's the compilation, it's a compiler, takes your data set and creates neural network code that runs in the car, and so it's roughly speaking, neural net. The reason i refer to it as software 2.0 is that you can take functionality from 1.0 codebase and put it into 2.0 codebase, so this boundary is fluid and you can actually engulf more and more software 1.0 stack. So over time.

Since i joined about two and a half years ago, the neural nets have expanded in how much of software 1.0 and they've taken over, and so the trend is upwards. I've shown you sort of how the software 2.0 stack is expanding through the code. It's eating. Some of the you know occupancy tracker-like techniques.

The logical conclusion of that, of course, is there's still still a perception system that creates explicit predictions and those explicit predictions are one shown in the ice, the instrument cluster and two. They are used for a policy, and this policy is still in the land of 1.0, where for us and for many others as well, where you have explicit planner that takes all the predictions and drives on that. The problem with that is, your explicit predictions will are basically doomed to never be good enough and never be complete, and so writing these planners is extremely difficult. Error prone lots of hyper parameters very tricky, and so the logical conclusion of this, of course, is that we'd like to train neural networks that actually do a lot of this planning inside the network.

In particular, we have huge data sets of people driving cars, so when people drive cars and they steer the wheel, they're, actually data labeling for you they're, showing you how to drive through any any intersection or any other kind of place, um and uh. The other thing i'll say is like most of what i've described so far relies primarily on self-supervi on supervised learning, with the exception of depth, where we massage and create a massive data set. But but we've seen recently with the use of, for example, contrastive losses and things like that. There's a lot of progress today in self-supervised learning and so we'd like to leverage some of that in the team.

Here's an example of just some videos: um. We have probably the most interesting largest data set of videos. These are coming from cameras, and so we've been really looking to some of these self surprise techniques, to speed up the process and to speed up the rate at which we learn from very little supervised data. Android certainly talks at a pace faster than normal and there's a point that andrij makes that's extremely important and must be emphasized.

Essentially, because tesla has a lot of data with people driving cars. The company can replicate a user's wheel movements when, on autopilot, for example, if there's a worn out, stop sign that tesla's cameras, don't seem to be catching onto autopilot, will be able to use the facts that humans have stopped at that area in the past and replicate It in real life, according to elon musk. This is the kind of revolutionary update that's coming in two to four months with the expansion of software 2.0 in tesla's, autopilot teslas are becoming better at spotting 3d objects. These improvements will allow tesla to better understand the environment and react to it correctly.

For example, the tesla autopilot is now able to drive through green lights, knowingly, previously teslas wouldn't react to green lights, and when it did, autopilot was just following the lead of a car. In front of it or staying in its lane, if a tesla on autopilot was the first car in its lane and the light turned green, the driver would have to hit the accelerator. However, with the improvement of tesla's neural networks, autopilot will now be able to drive through green lights, while knowing it is going through a green light. As andres carpathy explained in this presentation, software 2.0 is beginning to eat up software 1.0.

What he means by this is that him and his team are beginning to solve each small issue using software 2.0 in a presentation in china, elon musk explained how collectively these relatively rare situations become a larger problem. The there are many small problems um and then there's the the challenge of solving all those whole problems and then putting the whole system together and just and just keep addressing the long tail of problems. So you'll find that you're able to handle the vast majority of of situations, but then there'll be something very odd, and then you have to have the system figure out and train to deal with these very odd situations. Um - and this is why you could you - you need a kind of a real world situation.

Nothing is more complex and weird than the real world. Any simulation we create is necessarily a subset of the complexity of the real world. Essentially, what we will see in the upcoming autopilot update is a foundational rewrite, more reliance on neural networks and human reactions, and a lot of small issues being fixed. This upcoming update makes elon musk, confident that tesla will have level 5 autonomy, at least in california.

The tesla autopilot does work reasonably well in china. It does not work quite as well in china as it is in the u.s, because still most of our engineering is in the us, and so that tends to be the local loop of optimization. So autopilot tends to work the best in california because that's where the engineers are and then it once once it works in california, then we then extend it to the rest of the world. Most people are skeptical of achieving level 5 autonomy by the end of 2020.

With myself included, but we do have to realize that we don't know how the developments are going for tesla's upcoming autopilot update according to elon musk, a lot of functionality will happen all at once. When we transition to the new software stack most likely, it will be releasable in two to four months. Then it's a question of what functionality is proven, safe enough to enable for owners the the thing to appreciate for level five autonomy is really what level of safety is acceptable for the public streets, um relative to human safety and then uh so is it? Is it enough to be twice as safe as humans like? I do not think that the regulators will accept equivalent safety to humans, so the question is: will it be twice as safe as the requirement three times as safe five times as safe ten times as safe? So you can think of really level five autonomy as kind of like a march of nines. Like do you have 99.99 uh safety, 99.99999 percent? How many nights do you want of when what is the acceptable level and then what amount of data is required to convince regulators that it is sufficiently safe? Um? Those are the actual uh in-depth questions.

I think to be asking about level 5 autonomy that it will happen is a certainty. Let me know whether or not you think tesla will be able to achieve level 5 autonomy in the upcoming autopilot update in the comment section below. If you enjoyed this video, please hit the like button and subscribe and i'll see you in the next one.