Horizon Yu Kai: ChatGPT "crazy", why has automatic driving not come?

Yu Kai, founder and CEO of Horizon

In November 2022, OpenAI announced the latest large-language pre-training model, ChatGPT, which opened the development race for large models. As one of the important scenarios of AI landing, autonomous driving will have further development in this wave, and it has also caused many people to discuss: why is ChatGPT "soaring", but automatic driving is delayed?

On April 1, at the China Electric Vehicle 100 High-level Forum, Yu Kai, founder & CEO of Horizon, expressed his views on this. In his view, the key is that the challenges facing ChatGPT and autonomous driving are very different.

First of all, ChatGPT may replace or assist some white-collar jobs in the future, but the job itself has a relatively high error tolerance, such as ghostwriting, even if it is not perfect, but people can modify it on the basis of what he wrote. But automatic driving is different, especially unmanned driving, human life is at stake, and the error tolerance rate can only be zero.

Secondly, in terms of computing, ChatGPT is computing in the cloud, which can achieve sufficient power supply and a very good water cooling system. But the car side relies on electromagnetism, and the challenge of heat dissipation is very large, which means that automatic driving cannot use such a large model and such a large calculation.

Under various challenges, perhaps autonomous driving, especially unmanned driving, cannot make as amazing progress as ChatGPT. Yu Kai is even pessimistic that even L3 will not be realized in ten years.

From the perspective of the entire autonomous driving industry, especially L4 autonomous driving, there is indeed a pessimistic and depressed atmosphere. Whether it is Google, Waymo, or the commercialization of Cruise, layoffs, or the closure of Argo, which Ford and Volkswagen invested, the industry is returning to calm, returning to the essence of business and user value.

What is user value? Yu Kai cited a survey data: 87% of users really need the feeling of ease during driving, hoping to eliminate tension and fatigue. In other words, users do not currently need to be truly driverless.

"From the beginning of advanced assisted driving, it is already creating value for users." Yu Kai continued.

However, it is also assisted driving, but the needs in Japan and abroad are also very different. In Japan and Europe, assisted driving is mainly L1 and L2, that is, driven by safety regulations, including safety functions such as AEB, ACC, lane keeping, and emergency automatic braking. "But in China, it's completely different, security is the passing line, more driven by user experience and user value."

Therefore, Yu Kai believes that do not be too anxious about automatic driving. By 2025, what is really necessary is to make the high-speed NOA, loop NOA, closed-road autonomous driving experience silky smooth under a reasonable cost performance, and the price cannot be too expensive. At the same time, there must be considerable investment to really make the NOA available in more complex urban areas.

Although the industry generally believes that 2022 is the first year of L2 + (high-speed NOA) mass production, in Yu Kai's view, there will be at least three years of good progress in technology research and development.

Especially in terms of algorithms, Yu Kai said, "We must constantly optimize our software algorithms for a given computing power, use more data, and constantly approach the upper limit of user experience." ”

Based on the inspiration of ChatGPT, Yu Kai said that he will continue to use larger data, larger models, and unsupervised to learn the common sense of human driving. "For example, given the current traffic environment, given a navigation map, given a driver's entire driving behavior history, how to predict his next driving action? Learning can actually be obtained from a large number of unsupervised behaviors that do not require labeling. Building such a large-language model of autoregressive driving is what we need to do next. ”

When the automotive field continues to approach the field of autonomous driving, many car companies have begun to focus on autonomous driving chips. In addition to GAC's investment in Yixing Intelligence, the Volkswagen Group and Great Wall Motor have also focused on the field of autonomous driving chips through their own methods.

In this regard, Yu Kai said that every car manufacturer should carefully consider the matter of creating their own self-driving chip. The first is the huge amount of capital and cost investment in research and development; The second is how to stay competitive.

"Self-research" and "other mountain stones" are a strategic choice, and Yu Kai's advice to car manufacturers is that "if your sales are expected to be less than 1 million vehicles, the efficiency of this fund is not high." ”

For the future development of autonomous driving, Yu Kai also expressed some "anti-consensus" views.

He believes that in ten years, even Level 3 autonomous driving will not be realized. However, on dedicated roads, autonomous driving can be realized on autonomous driving dedicated roads with vehicle-cloud collaboration. However, there is a premise, that is, no one is allowed to drive in this environment, and all cars must be automatically driven on special roads.

At the same time, he also believes that with the development of this artificial intelligence, a machine hegemony will be formed. On the surface, human beings feel very satisfied with the development of algorithms now, but in fact, they give up their own thinking more. "The machine is a black box, how did it develop? It's out of control for us. ”

In his view, it is necessary to promote artificial intelligence computing that is distributed, localized, protects user privacy, and computing is transparent to humans. For example, the future vehicle will actually become a distributed energy storage center, hundreds of millions of cars equipped with thousands of terabytes of computing power chips, how to achieve its corresponding functions when parking, which is the world's largest computing resource pool.

"In the future, it will definitely be computing other than on-board computing to provide infinite possibilities." I think the computing revolution has just begun. Yu Kai finally said.

(This article was first published on Titanium Media App, author|Han Jingxian, editor|Zhang Min)