ChatGPT continues to be like a black hole, attracting major forces.
From the outside world, this is a milestone event in the field of artificial intelligence, representing the development of AI technology to a critical point, and also indicates that the original business model has the opportunity to reshape again.
As Nvidia founder Huang Jenxun said, because ChatGPT can be called by the outside world through APIs, it means that it can access all kinds of software, and its emergence will also have a huge impact on the industry like browsers changing the Internet and Apple App Store changing the software industry.
Typical such as search engines, cloud computing industries, due to the emergence of ChatGPT, the original market competition situation has been broken. For example, Baidu can sell its intelligent cloud services through the words of Wenxin. For the original players on the field, there are both opportunities and challenges.
And what about the automotive industry? It can be seen that in recent times, car companies have also frequently spoken out in the field of ChatGPT:
First of all, nearly 10 car companies such as Jidu, Great Wall, Dongfeng Nissan, Aiways, Leap, Geely Automobile, and Haima Automobile joined Wen Xin's circle of friends; Immediately afterwards, the company announced that the autonomous driving cognitive model was officially upgraded to DriveGPT, and the ChatGPT model and technical logic were fully applied in the automatic driving model algorithm.
For the automotive industry, and even the more focused field of autonomous driving, is ChatGPT a satellite or a take-off point?
Stop at the "chicken ribs" of voice interaction?
Jidu is the first car company to publicly play the concept of "ChatGPT on the car".
On February 14, at the kick-off meeting of Jidu Robotese Sanlitun Experience Center, CEO Xia Yiping took the lead in announcing that it will integrate Baidu's "Wen Xin Yiyan" capabilities to create an artificial intelligence interactive experience for smart car scenarios and support cars to achieve a new level of natural communication.
It is understood that the "Wen Xin One Word" collection is the first time that global ChatGPT technology has been applied to smart car products.
The specific details of the cooperation were not announced on site. In this regard, Xia Yiping said, "Regarding the specific functions that will be, I can only say that I am still working hard to connect with the Baidu team, because everyone also knows that Wen Xin's words will be released in March, and then we will do our application." ”
However, combined with other data, it can be boldly guessed that ChatGPT is likely to be the first to be applied in voice interaction.
After all, in addition to ChatGPT itself is out of the circle with its excellent interaction ability, during the same period, Changan Automobile's official official account released an article entitled "If ChatGPT is installed into Changan Deep Blue SL03", consumers naturally think of the in-vehicle intelligent voice interaction function.
At this stage, as the mainstream interaction method in the market, intelligent voice is a standard function of intelligent cockpit. According to consumer research data, voice interaction is the most proportional and satisfactory interaction method in the cabin.
Generally speaking, OEMs take voice interaction functions as the core embodiment of their product intelligence and differentiation. For example, Mercedes-Benz's user experience team has set three goals for the development of its in-car voice experience:
1. Allow drivers to talk naturally as they would treat another person.
2. Supports more query types than typical voice services.
3. Integrate voice more naturally into the overall in-car experience so users can seamlessly switch between voice and touch controls.
However, the reality is often harsh. Due to the lack of mastery of voice interaction technology by car companies, the interaction degree of the on-board voice system is generally weak, the experience is average, and most of the time it is reduced to "chicken ribs".
The reason for this lies in the intelligent voice interaction technology, there are still difficulties.
The technical logic of intelligent voice interaction mainly includes three parts: recognition, understanding and execution. At present, among the vendors that provide solutions, the identification part has become mature, and the recognition rate can reach 90%. The pain points of the industry are mainly concentrated in the "understanding" part, and most of the on-board voice interaction systems are not intelligent in "understanding".
For example, most of the front-end voice interaction functions provided by traditional OEMs use command control. Users need to interact according to specified commands, and machines do not have semantic understanding. The interaction mechanization leads to a single function and a single command word for the entire system.
So, based on the AI language model, can ChatGPT, which can be answered almost any problem in any field, bring new possibilities to the in-vehicle voice interaction system?
The answer is yes. "ChatGPT has obvious advantages in reasoning and learning ability, not only for understanding and dialogue, but also through contextual communication and self-learning to achieve auxiliary creation and knowledge evolution. These capabilities are also applicable to the field of in-vehicle voice interaction, integrating dialogue intelligence technology, deep learning big model technology, engineering capabilities, and the potential of big data to bring smoother and more effective responses..." said Gefujiang, product director of's automotive division.
From this point of view, if the car intelligent voice can be implanted with ChatGPT-like technology, under the high user stickiness, the future commercialization prospects will be broader.
Of course, considering the cost level, it often pulls the relevant manufacturers back to the cold reality.
According to Lambda's official website data, Microsoft designed a distributed cluster containing 10,000 Nvidia V100 GPUs for OpenAI for GPT-3 model training, due to the large number of model parameters (a total of 175 billion parameters), it took a total of 30 days to complete the training, and the total computing power consumed was 3640PF-days. Based on the Nvidia Tesla V100's Lambda GPU instance pricing of $1.50 per hour, the full training cost of GPT-3 will reach $4.66 million per time.
The "emergent" moment of autonomous driving
It can be seen that ChatGPT is more intelligent in the short term, empowering the development of the automotive industry, and there is still some distance from autonomous driving in the full sense. To this, ChatGPT itself also gave the following answer:
As a language model, ChatGPT is mainly used to generate natural language text, such as dialogues or articles. Autonomous driving technology needs to deal with issues such as perception, decision-making, and control, which are not related to the text-generating tasks handled by ChatGPT. Therefore, it is not technically feasible to rub autonomous driving technology with ChatGPT. However, autonomous driving companies can use natural language processing technologies such as ChatGPT to improve the human-computer interaction capabilities of their products to better meet user needs.
In fact, this is also the mainstream view in the industry. Previously, at the expert media communication meeting of the China Electric Vehicle 100 Forum (2023), Academician Ouyang Minggao, vice chairman of the board, said that ChatGPT will trigger a new round of revolution in artificial intelligence and will also have a far-reaching impact on intelligent driving.
"What Musk does is based on this technical route, which is the so-called big model, big data, big computing power... In order to obtain big data, it is necessary to sell a large number of intelligent assisted driving electric vehicles, especially urban assisted driving may be rushed to market this year, which is also an important technological trend. ”
How to understand this passage? It should be known that if you look closely, there is a very important basic principle capability behind the formation of ChatGPT - "Emergent Ability", which is generally defined as "the phenomenon that occurs when quantitative changes in the system lead to qualitative changes in its behavior".
In simple terms, below a certain scale threshold, large models perform close to random, and beyond that threshold, their performance is much higher than random.
Various tests have shown that only when the model reaches the scale of GPT3, that is, the parameters are greater than 100 billion, the model can form a "emergence capability". It is on the basis of "emergence ability" that AI models demonstrate complex reasoning and knowledge reasoning capabilities similar to humans, which is the so-called "chain-of-thought".
Based on the reasoning ability of the "thought chain", there is no need for complex training, but only to give additional hints when asking questions, and the model can automatically learn and make corresponding reasoning to get the correct result. It thoroughly embodies the imitation of the advanced thinking ability of humans by AI models.
This capability is likely to be an important foundation for ChatGPT to become highly intelligent. This will undoubtedly play a decisive role in the choice of path for autonomous driving.
For a long time, around the evolution of autonomous driving technology, there have always been two "schools" in the industry: among them, Waymo, Pony.ai, etc. are representatives of "leapfrogging", and the strategy is to directly achieve L4/L5 level automatic driving; Tesla is a representative of the "progressive" route, the strategy is to first give priority to L2/L3 level assisted driving on production vehicles, and then collect data before moving to L4/L5.
Previously, the camps of the two sides were clearly divided, but the situation has changed in the past two years. Many L4 companies have begun to "reduce dimensionality" into the L2 field. For example, Robobus has launched a high-end autonomous driving solution; WeRide received investment from Bosch to develop Level 2-L3 autonomous driving software for passenger cars.
In contrast, Tesla has produced millions of cars among the "progressive" players, and its assisted driving system, Autopilot, will collect billions of miles of road conditions and driving data. These industry trends seem to be announcing the phased victory of the gradual route.
Of course, there are also different views in the industry, believing that the current is only a victory in the L2 field, and whether it can be gradually reached L4 is unknown. At present, the emergence of ChatGPT has made the industry realize that constantly accumulating kilometers and running like this can obtain a higher level of autonomous driving technology leapfrogging, quantitative changes can cause qualitative changes, and automatic driving also has the opportunity to usher in a "sudden" moment.
This is the importance of ChatGPT for autonomous driving, and to some extent, it is a clear way for automatic driving.
Resources:
TechWeb "ChatGPT on the "car"? It's not that simple"
Gascar community "ChatGPT, hit the "face" of human-car interaction? 》
Titanium media "Explosive ChatGPT, can automatic driving be realized faster? 》