INTERVIEW: AI Safety is Far More Important in Car than a Data Center AI: Lars Reger

At a time when artificial intelligence is shaping the conversation around next-generation mobility, Lars Reger, Global CTO of NXP Semiconductors, believes that for a “rolling robot” like a car, being absolutely trustworthy, safe, and secure is far more important, and that only after these fundamentals are ensured does intelligence come into play. In an interview, he explains that the shift underway is not about making vehicles more intelligent in isolation, but about ensuring that this intelligence operates within a framework of safety, security and trust. In a detailed interaction with Autocar Professional, Reger explains why cars must now be seen as “rolling robots”, why the industry is moving from cloud-heavy AI to edge computing, and why cybersecurity will define the next phase of automotive evolution. Edited Excerpts:

We are not hearing as much about AI in cars today. Is something still evolving at the chip level?

There is a big difference between having tools like ChatGPT in the cloud and building a robot, a trusted machine, a rolling robot in that case, a car. When you ask something to a cloud system, it goes into a huge data centre, gets processed there, and comes back. But when you move into a physical system like a vehicle, everything changes.

For the robot, it becomes much more important to be absolutely trustworthy, safe and secure, because the moment it does something you do not want, it becomes dangerous. Energy efficiency then becomes the next layer, and only after that does intelligence come into the picture. As I said, for the robot, it is much more important to be absolutely trustworthy… only then is it about intelligence.

We are now shifting from these large data centre-based models into what I would call physical AI, robotics building. For that, you need smaller models, but you need to be able to trust them completely.

You describe cars as ‘rolling robots’. How does AI actually work in such systems?

There is a big misunderstanding in how people think about intelligence in autonomous driving. It is not like you take your son at 18, hand him the keys and say, you have watched driving for years, now go and drive. That does not work. We send people to driving school.

First, they learn the rules, then they learn how to apply them. That is not creativity, that is rule-based and deterministic. You have fixed rules. In Germany, you drive on the right side. In India, you drive on the left. At a red light, you stop. If there is a speed limit, you follow it.

Only inside these boundaries is there any room for flexibility. So what we are doing is boxing in the brain into a functional, safe and secure environment. That makes the robot. This is very different from asking an AI system to generate content or create something new. In a car, unpredictability is simply not acceptable.

There is a strong narrative that data is the new oil. How do you view that?

There is this nice saying that data is the new oil. It is not. Oil does not rot, oil does not become irrelevant. Data does. Every day, your eyes capture enormous amounts of raw data, but your brain processes only a very small, meaningful portion of it. The same principle applies here. Having all the data of an entire country does not help. What matters is having the right and relevant data.

That is where the science and the art lies. If you train systems with the wrong data, they will behave in the wrong way. And in automotive, that can become dangerous very quickly. You can create biased systems, systems that respond to the wrong signals, or systems that misinterpret situations. So the focus has to be on selecting the right data, training correctly and then validating those models before deploying them.

How do you see the balance between cloud and edge computing evolving?

A lot of what we are doing will move to the edge. We are already running AI accelerators with billions of parameters together with automotive microprocessors at very low power consumption. We are talking about around 7 watts. That is less energy than sending a question to a cloud system, processing it there and getting the answer back.

Now imagine billions of devices constantly sending data to the cloud. The energy required just for data transfer becomes enormous. That is something we simply cannot afford. 

For many use cases, whether it is managing your car, your home systems or other devices, you can run these models locally, activate them when needed and avoid continuous connectivity. The cloud will still be there, especially for training and larger-scale processing, but a lot of real-time intelligence will sit at the edge.

With the move towards software-defined vehicles, does adding AI increase complexity?

Actually, it is the opposite. These software-defined architectures are a cleanup of functions. Earlier, you had many domain-based systems, powertrain, infotainment, connectivity, driver assistance, each with its own control unit. That led to a very fragmented system. Now we are moving towards central compute devices connected through zonal architectures. These central systems act like the spine or the cerebellum of the car.

You reduce the number of control units, simplify the software environment and create a more flexible system. With the right middleware, you can move functions between central and peripheral systems easily. It is similar to running the same operating system on different devices. You can shift applications back and forth without redesigning everything.

And where does AI fit into this new architecture?

You do not necessarily need separate systems for AI. You can integrate these capabilities into existing architectures as co-processors, as add-ons. That allows you to enhance the system without increasing complexity. You also get advantages in terms of data security because you can process data locally instead of sending everything to the cloud. At the same time, response times improve and systems become more reliable.

Cybersecurity is still not widely discussed in India’s automotive ecosystem. Should OEMs and suppliers act now?

There is no chance to build a connected device without cybersecurity. If you have an unconnected system, fine, but the moment it is connected, you need protection. The good news is that we already have many of the building blocks. The same technologies used in banking cards and biometric passports are being used in automotive systems.

Every device needs a secure identity, like a passport. A car can authenticate itself, communicate securely and prove what it is allowed to do. You need both hardware and software. Hardware gives you the anchor, software enables the system. One without the other is not enough.

Cars are increasingly becoming extensions of smartphones. How do you see this comparison?

I am always careful with that comparison. A car is not a smartphone on wheels. The smartphone is becoming a car without wheels. The risk levels are very different. With a smartphone, you are dealing with financial loss or reputational damage. With a car, you are dealing with physical harm. We think in what I call damage classes. You do not want to overrun a person.

You do not want your bank account hacked. You do not want your reputation damaged. That is why “trust your device” becomes so important. This applies not just to cars, but to all connected systems, whether it is a medical device, a home system or an industrial machine. If your fridge orders too much milk, you fix it. If your thermostat overheats your home, you adjust it. But if your car behaves unpredictably, that becomes a serious problem.

Your message for Indian OEMs and suppliers?

You cannot talk enough about cybersecurity. We are already working closely with Indian OEMs, including two-wheeler manufacturers, and I see a strong need for everyone to address this. At the end of the day, “you cannot sell a product that is not cyber secure.” If vehicles fail, if they get hacked or behave unpredictably, the trust is gone. And once trust is gone, it is very difficult to bring it back.