AI-driven automation: Transforming industries with smarter technology

During the World Government Summit, Robin Li, Founder of Baidu, highlighted the rapid progress of AI innovation, especially in the large language model inference costs. He emphasized the China's distinctive challenges in AI development, where computing constraints drive cost-efficient innovations. Also he highlighted Tesla's autonomous strategy with Baidu's Apollo Go, which runs robotaxis without drivers from the beginning. He emphasized that robotaxis are 10 times more secure than human drivers but cautioned that accidents and public opinion would affect progress. Gradual  implementation is essential for the long-term success of AI-driven transportation. Below are the key insights from his address.

Innovation cannot be planned when and where it comes. It is crucial to foster an environment that is conducive to innovation. According to Moore's law, every 18 months, the performance doubles or the price halves. However, today the large language models, the inference cost can be reduced by more than 90% over 12 months. Hence, this is much faster than the computer revolution experienced over the past few decades, and also, it's a very large area. There are all kinds of directions that can be innovative. For example, in the US, the open AIs and NSRFs all raise to build or to develop the most powerful large language models. But in China, because of the constraints of compute, there is a need of innovation to reduce the cost, both on the inference side and the training side.

And fortunately, there has been significant progress in the past year. The Baidu started from a search engine background. It's naturally very close to a large language model. In March 2023, the ErnieBot was launched, which was an answer to ChatGPT, which was launched three months after ChatGPT. That's also probably the first ChatGPT-like application ever launched by a public company, and Google launched BARD later, and Google renamed it to Gemini.

The AI Technology Stack and Market Dynamics

AI application is more important. But in the technology stack, it's like a pyramid, and most money is made on the silicon level, the GPUs, and other kinds of basic chips. Then on top of that is the cloud infrastructure, the hyperscalers or cloud service providers, and then on top of that is the language model, and lastly the application layer is at the very top, but the applications haven't made much money yet. Today, AI native applications, ARR, or Annual Recurrent Revenue, are probably like several 10 million or max several hundred million dollars. But at the GPU level, it is about tens of billions of dollars. Most of the innovation, apart from AI or in the IT industry, in the past few hundred years, relates to cost reduction. If costs can be reduced by a certain amount or certain percentage, then that means productivity increases by that kind of percentage. So that's pretty much the nature of innovation. It's happening at a rate much faster than before.

The Evolution of Autonomous Vehicles

For companies in China, there is a certain type of constraint. For example, the robo-taxi car. Even if people have a car, that robo-taxi car cost is $100,000, so there are chances to make money by operating that car in a driverless fashion. But in China, the ride-hailing price is much lower. So it's crucial to come up with a technology that costs much less in order to make sense to have a driverless operation.

There are a number of technological approaches for the self-driving thing. Tesla has their own approach. They just don't want to use any sensor other than the camera. And they would like to achieve full self-driving from the ADAS, or advanced computer-aided driving, until it can reach a fully driverless stage. But the robot taxis called Apollo Go and Waymo, there is a different approach. China likes to start from certain geographic areas. And from day one, it's operated without any driver. However, both approaches have their merit. The road condition is actually quite complicated and challenging. Sometimes people are cut off by a car coming from nowhere. There are motorcycles weaving through the traffic and sometimes people have to overtake a bus in front them.

The Future of Robotaxis and Risk Management

Technically, deploying technologies in some cities takes about two weeks. But in order to comply with the local regulations, a typical deployment cycle is about six months. Most cities require having a safety driver to drive around the road. One or two months, people can get rid of the driver but cannot the charge; then people can start to charge, and expand the areas from a certain dedicated area to the whole city. That process typically takes about six months.

In a regulatory environment, technology is already ahead. Technology is improving rapidly. Everyone is rushing ahead, and competition is also very fierce. And there are also risks associated with this kind of innovation. Although China has proven that robotaxis are much safer than human drivers, and it's at least 10 times safer.

Every year, more than a billion people get killed in traffic accidents. But with Robotaxi, that death rate can be reduced significantly. Also, in the insurance claims from the operating record, the insurance claims are only like 1/14th of a regular taxi or a regular driver's car. But still, this is a new industry, a new area. People's tolerance for accidents is very low. So if some serious accident happens, it may drag down your technology progress. It requires stopping the operation for a certain amount of time, finding out what happened, and making all the adjustments to restart the operation. So Gradual  implementation is essential. Although Robotaxi has been operating for like two or three years on a relatively large scale, it hasn't had any serious accidents.