Software is staring to eat the auto industry. For the first time, the most advanced cars on the road today now have the ability to see and understand what’s happening in the world around them. Although elusive, the goal is to make the AI’s understanding of the world so rich that it can control the car on its own very reliably without a human driver. Skeptics and cynics argue that this technology is “useless” or even “dangerous” because it’s not ready for driverless operation today. Here, I want to advocate the opposing position: That this technology is not only extremely useful to have in a car, but that it is urgent that we start building it into every new vehicle manufactured in the United States today.
Why should cars employ “AI” to understand what’s happening around them?
- To Provide Active Safety Features for Human Drivers — The same perception system that allows an autonomous car to see what’s happening around it can also be very useful when a human is driving the car. To drive itself, your car needs to know where the lane lines are, and it needs to know how to check if it’s clear to make a lane change. These building blocks let automakers create features like Lane Departure Avoidance, which can detect when drivers are drifting out of their lane with an oncoming car about to hit them and apply corrective action or warn the driver to avert a collision. Even skeptics agree these features are a great thing and prevent thousands of collisions a day. At a time when fatalities and injuries are at record highs, NHTSA should move to mandate not only Automatic Emergency Braking but a comprehensive set of advanced AI-powered active safety features that can be updated over the air on all new vehicles. That software update part is important, because consumers tend to turn off active safety features in most cars due to their imperfections. If the features are being actively improved, they’re more likely to be used.
- To Make Insurance Cheaper and Incentivize Safer Driving — Five states now have Tesla Insurance, and four of those states will actually adjust your monthly insurance rate based on your driving safety score. What is the safety score? Take the same perception system you use to make the car drive itself and use it to grade a human’s driving, and that’s the safety score. With Tesla Insurance, safe drivers pay less while those taking extra risk chip in more to help cover the extra damage caused. The result is not only cheaper insurance, especially for groups like young males that are traditionally labeled risky drivers, but an actual reduction in crashes. That’s right: Drivers who tried to get a higher driver safety score to save money on their car insurance actually got in less crashes. That AI system may not be good enough to drive a car on its own just yet, but it’s good enough to judge human driving and pay drivers for making safe choices. These are powerful incentives, and like the active safety features even critics have to love insurance that pays drivers to drive safer. Without the AI system developed for self-driving, it would not be possible to watch the driver and judge whether they are driving safely or not.
- The Visualization — Every car with an advanced perception system usually has some kind of visualization to help humans understand what the perception system is able to see. In a Tesla for example, every time users glance down to check their speed they get a full 360 degree representation of their surroundings. Is there someone walking up behind the car? Before they switch gears into reverse, they’ll be able to see it. The visualization is like an extra set of eyes for every driver, and will greatly increase drivers’ awareness of their surroundings. Greater awareness ultimately means less accidents.
- The Convenience — It is extremely convenient to have a car that can try and drive you anywhere you ask, even at this early stage where human supervisors will need to take back control often. Not everyone will agree that it’s for them today, but many people will love it and many more will fall in love as the software improves. Alarmists argue that advanced ADAS that can sometimes complete entire drives on its own is dangerous, as users may become complacent and fail to take over when needed to prevent a collision. For the most part, these concerns are entirely unfounded. All of the most advanced systems on the market now monitor driver attention to ensure that human drivers are focused on the road and ready to take over at a moment’s notice. More importantly, what detractors miss is that ADAS technology has been on the roads for a decade now. There are already millions using much more primitive versions of the technology, and it hasn’t been the end of the world. For example, Tesla’s free basic Autopilot and most competitors’ ADAS systems do not read traffic lights and will try and blow through every red light at full speed unless manually disengaged. Despite the fact that people use these features on city streets every day, they manage to always disengage when needed and are adept at working around the limitations of these systems, which they become very familiar with. How could a system that can actually handle traffic lights and will stop on its own 99.999999% of the time without user action be more dangerous than a system that never handled them at all, stopping on its own without user takeover 0% of the time? Not only is FSD Beta much more convenient, the fact that it handles so much more automatically makes it much safer than more primitive systems that previously required the driver to takeover often for common road scenarios. The software tends to make much safer and more conservative choices than a human driver would have made.
- We must start building out the fleet — The question is not when we get the first reliable driverless car, but how soon we can build 10 million of them. One driverless car is useless. It may be perfect for whoever is lucky enough to ride in it, but you can’t build a ubiquitous ride network that challenges car ownership with just one car. Even if someone like Waymo solved autonomy today, how long would it take them to ramp production and get 10 million cars on the road? How many more years? We can’t afford to wait until the software is perfected. We need to start working hard towards that 10 million number now, standardizing the hardware and repeatedly and forcefully iterating on the software. That way, as the software becomes highly reliable there will be a large base of devices ready to capitalize on it. For this reason, it would make sense for Tesla to focus on optimizing for maximum units produced. Doing so also yields an additional benefit: Lots and lots of testing. There is really only one way to prove an autonomous car works and is safe: You have to try it, in the real world. When your lane centering technology has been tested across billions of miles, day and night, rain or shine, that is testing and learning that provides stronger building blocks for an ultimately driverless car. When driverless cars do finally launch at scale, they will benefit from years and years of testing under human supervision.
Those are just a handful of reasons why I think it’s so important to start putting this technology into every new car today. I want to start writing some more about this technology because I think it’s incredibly powerful and important, and I think it’s a major breakthrough in software. So much of what I see discussed in public is just misconceptions, silliness, and confusion. It doesn’t help that TSLAQ, the anti-Autopilot cult and the media are eager to stoke public fears to sell papers and puts. However, I’m confident that as people start to understand the technology better, their attitudes will change. This is not about driverless cars, but rather a set of great benefits we can take advantage of today.
People are so distracted thinking about driverless cars that they’re missing the transformation happening before their eyes. This isn’t about some autonomous future, cars are starting to gain amazing AI abilities today. This is the foundation of a massive new computing platform that will take center stage in our lives: When AI does the driving, humans are suddenly freed to do whatever they like, opening the Pandora’s box of in car software opportunities.
I believe the companies that ultimately survive in the driverless race will be the ones that can find a way to monetize benefits of the technology that users can enjoy today. If your company is aiming straight for driverless as your MVP, you’re going to need constant infusions of capital from starry-eyed VCs and private equity firms at eye-popping valuations. They may have been happy to invest in companies burning billions of a dollars a year thus far, but as the tide goes out we may find that many autonomy players are naked. But hey, I’m sure the Fed will make sure to not to freeze up equity markets on them just when they’re needed most.
4 thoughts on “Why We Need AI in Every New Vehicle Yesterday”
I know that, you are software engineer.
However, Do you have deep knowledge about algorithms and software systems behind this FSD’s every software parts?
A lot of my Computer science professors, researchers on machine learning, AI, computer says, Tesla FSD beta may not be as advanced as Waymo or Cruise. It has nothing to do with its performance and and safety. It is related to many deep calculations, deep algorithms behind those self driving system.
This is what those profs and researchers said to me, as they understand more about those than general public.
To the contrary, the algorithms are far more advanced as they’re operating on just computer vision rather than a more concrete spatial awareness from LIDAR localization.
This was fire! Really well thought out and argued, it wouldn’t surprise me if this post will save lives. That said, we can argue to the cows come home, but the statistics will rule the roost. I’m pretty sure they will decide this issue in favor of your arguments, as the Tesla numbers clearly already shows. When you mix the potent motivator of (insurance) money into it, I think this will be game over. It’s just a matter of how fast, where your best contribution through this writing was found. I agree with you 100% that speed is of the essence, any more delays and Luddite action will almost be criminal (as I think Sandy Munro has already pointed out).
Great thoughts as usual.
But you made an (unimportant) mistake in the penultimate paragraph:
“When AI does the driving, humans are suddenly freed to do whatever they like, opening the Pandora’s box of in car software opportunities.”
Pandora’s box was classically filled with war and pestilence; unwanted stuff. Do not open! Maybe you mean Cornucopia, the ever-refilling horn of wealth and all things good? 🙂