Pari mielenkiintoista kommenttia Mobileyen tulospuhelusta, joissa viitataan suoraan Teslaan tai autonomiseen ajamiseen.
We believe that SuperVision provides a validated bridge to a true eyes-off system across a wide domain, which is seen by many OEMs as a true value driver long term. But the performance requirements for eyes-off are really underappreciated by the public and also by certain OEMs who are throwing everything they have at an eyes-on system with seemingly no clear plan on how to boost mean time between failure from one safety intervention every few hours to one every hundred of thousands of hours.
Now, whether [Tesla] could introduce robotaxi using only cameras, we are kind of skeptical, but no, we don’t know what they’re going to introduce. Maybe they’ll introduce a robotaxi with additional active sensors, not only camera. We believe that eyes-off systems—rather than calling this robotaxi, let’s call it eyes-off systems because eyes off means that you can drive autonomously on selective type of roads, not necessarily on every type of road. It’s still a great value. Eyes-off system, which is our Chauffeur product line has a great value proposition. We also believe that in time, it will even overtake the Level 2-Plus in terms of volume, but we see that as something for the next decade in terms of the volume ramp. SuperVision is this decade and eyes-off would be, in terms of scaling and overtaking Level 2-Plus, we see that as something for the next decade. But we’ll be very happy to be proven wrong and to have this accelerated.
Amnon, seven months ago you posted on LinkedIn that Tesla decision to adopt an end-to-end generative AI approach to full self-driving to train neural networks was—I’m quoting—neither necessary nor sufficient for full self-driving programs. Do you still feel the same way today, Amnon?
Now in my prepared remarks, I mentioned that on the EyeQ6, we’re going to have end-to-end both perception and actuation, and that does not contradict the point that we made. The Tesla end-to-end is the sole technology. Our end-to-end is just one engine on top of multiple engines in order to create a decomposable system that is explainable, that is modifiable, that you can explain what it does to the regulatory bodies, that you can customize the driving experience for OEM. If you look at some of our competitors like Waymo, they have the same view that there’s a very, very strong reliance on the neural networks, on data-driven networks, language models, but at the end of the day, it needs to be a system that is designed to be explainable and modifiable. We’re not against end-to-end. We are against end-toend being the sole engine for the system. Back at the CES a few months ago in January. I presented Mobileye’s end-to-end perception engine, right? What I call the multi by the power of 5, how to build an end-to-end perception engine, and this is running on the EyeQ6. We have also another engine, which also includes actuation. So this is going from videos to actuation as an end-to-end, but it’s a component. It’s a subsystem of a more complex system.
Amnon, in the past, I think Mobileye has discussed something like you were hoping for 10x better miles per disengagement from SuperVision when we compare it to something like full self-driving. I think a lot of those comments were pre Version 12. Any thought on how you think SuperVision, again, as a supervised eyes-on system, the competitive statistics stacks up today?
We are targeting the current generation with EyeQ5. It’s improving all the time. We have over-the-air update every two months or so. We are close to achieving a 100-hour mean time between intervention on highways, less so in the urban but it’s more than one or two hours of mean time between intervention. On the EyeQ6 system, as I mentioned in my prepared remarks, just for the camera subsystem, it’s about 1,000 hours of mean time between intervention on highways. I don’t know what is the mean time to intervention on Tesla’s Version 12. I don’t know if anyone measured that, but these are the kinds of things that we measure in terms of KPIs on how we progress.