AI Chauffeurs will reduce 90% of road deaths within 10 years.

How we already have an AI-Driven Roadmap to saving many lives... but it will take an initial leap of faith.

In 2030 we will all roll like Snoop

When George Hotz, the visionary behind Comma AI, recently addressed the future of Full Self-Driving (FSD) and Open Pilot, he painted a captivating picture of the near future, where autonomous vehicles will dramatically reduce road accidents. His insights shed light on how AI technology will bring self-driving cars to our roads sooner than we anticipate, thereby eliminating an estimated 90% of road deaths.

AI Self driving cars will eliminate the need for personally owned vehicles

The Road Safety Problem and Its Artificial Intelligence Solution

Hotz began by defining the problem at hand. He asked, "How do you build a model that outputs a human policy for driving?" His question reflects a pivotal shift in approach by Comma AI and other companies such as Tesla, who are moving from a traditional, hand-coded system, as adopted by Waymo and Cruise, to a model that learns from real-life driving data.

Instead of relying on manually designed algorithms to dictate car movements, these companies have built AI models that 'learn' from actual driving scenarios. This 'human policy' isn't merely about understanding rules of the road, but also nuances that can differentiate between an average driver and a good one. By focusing on customer satisfaction, the models aim to learn a 'good human policy,' thereby ensuring a smooth, human-like, and efficient driving experience.

AI Concept art of Tesla Logistics Truck

Learning a Driving Simulator: A Game for AI to learn safer roads

Hotz's mention of their recent work, "learning a driving simulator," is a testament to the leaps AI has made in this area. It builds on a machine learning model that uses an auto encoder and an RNN to predict the next step in driving. However, Comma AI has advanced from 2015's 'laughingly bad simulator' to what he fondly dubs "Drive GPT" using advanced models like VQ-VAE and Transformers.

In his eyes, the tech is trained in a self-supervised way on all the driving data to predict the next frame, effectively learning the 'human policy'. It doesn't just emulate human driving but aims to replicate the 'good human driver' experience. By bridging the gap between human-like and efficient driving, Comma AI is taking significant strides toward creating a self-driving car that we can trust with our safety.

"DriveGPT" taxis are only a few years away

The Role of AI Reinforcement Learning (RL) and Simulator for safer cars

Comma AI uses reinforcement learning in the simulator to derive the 'human policy'. Here, the reward function doesn't merely seek to emulate human driving but focuses on whether a human would disengage given certain behaviors of the AI. This is a step toward not just replicating human-like driving, but also focusing on what makes a human driver good.

It's a balancing act: the AI should not be overly cautious and yet ensure the driving experience is comfortable. The RL system Comma AI is developing would ideally never make the driver want to disengage, thereby building trust and satisfaction.

Concept Art of Porche 4x4

What's the Road Ahead for AI safety?

The path to self-driving cars isn't entirely smooth. Hotz admitted that while they have prototypes, they are a few bug fixes

away from closing the loop. These 'bugs' might be technical or may require more scale, but they are not insurmountable obstacles.

In fact, Comma AI has recently expanded its compute cluster, thereby providing more computational power. More compute and diverse data are critical as the models need to learn from various scenarios and make intelligent predictions.

The 'race' between Comma AI and Tesla, as Hotz puts it, is also enlightening. Despite always being one or two years behind Tesla, Comma AI believes that they are on the right track. Both companies are converging towards more of an end-to-end approach, using fancy simulators, and focusing on the reward system for AI driving.

Even Classics will be retrofitted with AI drivers

Thanks to breakthroughs from Comma AI, Tesla, and others. Their emphasis on emulating 'good human driving policy' and advances in AI learning and simulation paints an optimistic picture for the future of autonomous vehicles.

The pace of progress in technology suggests that self-driving cars will be a reality sooner than we think. Given that over 90% of road accidents are due to human error, this could mean a safer future on the roads, with AI technology playing the pivotal role.

There is an undercurrent of excitement in Hotz's dialogue, indicative of a future not too far off where our cars will do the driving, and we can sit back, relax, and enjoy the ride.

With Love, Bot Bot.