I'd love if you could just give 30 seconds on what Mobileye is, what Mentee is, and a little bit on Double AI that everyone can go back and look at later on.
Okay, so Mobileye is a 4,000 people company. It's traded on NASDAQ. I think about 20% of worldwide vehicles today on the planet have a Mobileye technology. So Mobileye sells chips with software for driving assist, a front-facing camera that performs all the safety requirements, systems with multiple cameras working on autonomous driving, about to deploy together with the Volkswagen robot taxis across US and Europe later this year. So we're all full stack from driving assist to more premium cars, hands-off systems up to robot taxis, and we're fully integrated silicon, software, electronics, everything. Mentee is the company I founded four years ago to build the humanoid robots. Two weeks ago, Mobileye acquired Mentee. So it's now part of the Mobileye group, taking the full scope of physical AI or embodied AI where AI is on a platform that interacts with the real world, that's cars and humanoid robots. Double AI is a company I founded two years ago. It's still new but it's working on building super intelligence. In a sentence, AI today is focused on productivity and GDP growth. AI in the future I believe would be working on overcoming the bottlenecks of humanity, and the biggest bottleneck of humanity are experts. People that have incredible capability of solving problems in very specific domains, whether it's science, mathematics, medicine, physics, writing code, optimizing code, all areas which experts are needed. And those experts are scarce, those experts are rare, very expensive. And if you look at the progress of humanity, those Newtons of the world, Einsteins of the world, they are responsible for taking humanity's envelope and pushing it forward. And if AI can do that, if a single computer can do that, imagine what 1 million computers can do. So this is what Double AI is working on. So more than 30 seconds, this is why I told you don't ask me that question.
I kept saying I had 15 questions prepared and he looked at me and said I'm going to talk forever on three questions. So thank you for letting me fit that in. Maybe starting on the theme of today, because it really is today, you, robot taxis, and you and I have been talking a bit about how those systems have evolved over time. And what do you think, Amnon, really differentiates the systems of today and the embedded AI versus even what we saw 5 years ago?
So what has changed is the level of learnable versus non-learnable components of a system and the level of integration. So if we look at the decade before, starting from 2013, the engine of choice was called convolutional networks. So you build a system in which the camera images would be processed to find patterns in the image that was using learnable data-driven technologies with an engine called convolutional net. And then everything else was using algorithms or non-learnable components like lifting to 3D, doing ego motion, putting together different viewpoints, tracking objects over time, and then also doing the driving policy, the decision making, which was based on classic reinforcement learning called MDP, Markov decision processes. So that was a technology up till 2021, 2022. Today the technology, let's call them the transformers technology, is based on language. What does that mean? Everything can be described by language. If you look at what are the basic atoms of language, they are called tokens. These are words. So everything can be tokenized, whether it's an image, whether it's text, whether it's whatever you can think of, if it can be described by language, it can be tokenized. So now imagine a network which receives tokens as input and outputs tokens. So basically everything can be learnable if you have data. You first describe it in language, then you build the tokens, you feed it into the network, and the network also outputs tokens. So if you have data to train the network, everything is learnable. So what this has created is that the number of non-learnable components in such a system is being drastically reduced, at the extreme even to zero. Imagine this network receiving pixels as input and output the vehicle commands. Right? This is an extreme case but it's possible, and including the driving policy, including looking at different viewpoints. So all camera inputs are fed directly and simultaneously into this network rather than each camera separately. So this is kind of the big transformation that happened in the technology. Now it has also a downside that people need to work with, and this is why companies who are building these systems are designing an architecture and not just simplistic looking at the network with input and output, because the higher the level of integration, the less explainability the system has. So now imagine you are involved in an accident. You need to explain whether to the regulator or whether to the court of law why the machine decided what it decided, presumably decided something wrong which led to an accident. You cannot say it was a black box. Second, you need to guarantee certain things. For example, you need to guarantee that you'll not violate traffic laws. You need to guarantee that you'll do all what you can in order to prevent an accident. And so this is what the regulator is expecting from you. Now if everything is one big black box, you cannot guarantee anything. And we know from language models they can hallucinate. They can make, you know, it could be very, very impressive but 5% of the time they can be very, very stupid. So how do you handle this? And this is a challenge for those companies who are building systems based on modern AI, is how to harness modern AI but at the same time guarantee certain safety elements both in terms of explainability and in terms of guaranteeing critical elements of driving, which is not violating traffic laws for example.
I think it was interesting in the round table we sat in yesterday, just to the next question. We did ask the room how many people have tried a robot taxi, and credit to the UAE, really at the forefront of its deployment, but I think about 40% of the room, maybe of the 40 odd people, 15, 16 had really been in a robot taxi. So we do have early deployment, let's say, some are more advanced than others. What are some of the lessons that we've learned in early deployment you think that we can apply as we start to rapidly scale this part of mobility?
Well, I would say first that the existing state of technology is very, very impressive. That means in small scale, few thousands of vehicles, Waymo has proven that they're very safe. Their data shows that about 10 times safer than humans, which is a very, very impressive bar to pass. Right now there are several challenges in scaling and making a big business. One is there is a back office. So you have people supervising the robot taxis. So they get feed from the cameras and they're supervising. What exactly they're doing is a closely guarded secret. In order to build scale, you somehow need to reduce the number of people that are supervising. So for example, those people cannot handle safety because then you need one teleoperator per one vehicle and very fast communications, and what happens if the communication falls and so forth. So they need only to give high level guidance. So you'll want to build technology that will enable you to gradually reduce the number of teleoperators until you reach to zero. That means there's no back office. I think the community is still far away from that. I think the most advanced, of course, is way more, but it's still making the number of teleoperators to zero is quite far. At the CES I presented the Mobileye thinking in which we use visual language models in two paths. One is slow and one is fast. The fast is about safety. So it will never hit, never violate a traffic law and never hit anyone. And the slow part is using a powerful VLM that can be in the cloud and it will analyze complicated scenes like a policeman is approaching you or something happened, situations which you refer to a teleoperator to give you guidance, you'd like the AI to do it. And I believe that those VLMs over time, big enough, trained enough, would be able to take the role of teleoperators and in that way allow for scale. So I think the game today is simply scale. And I'll also mention that it's not that there are many robot taxis out there. There's only one company that has proven it can reach a safety level. We have to be honest about that. There's only one company on the planet that has done it, which is Waymo. All the rest, including ourselves, are very confident we'll be able to do it, but there's only one company who has done it so far. So it's still, I think, there is a long journey to go through, and I believe that the opportunities are now very, very big.
