Introduction of ChatGPT
GPT is the abbreviation of Generative Pre-trained Transformer. Transformer is a deep learning model based entirely on the self-attentive mechanism. ChatGPT, therefore, is a generative, pre-trained algorithmic model and an AI chatbot for general users, currently mainly based on text chat, which uses advanced natural language processing (NLP) to conduct realistic conversations with humans. Currently, ChatGPT can generate articles, fictional stories, poems and even computer codes. ChatGPT can also answer questions, participate in conversations and, in some cases, provide detailed answers to very specific questions and queries. Chatbots are actually not new, they use keyword search technology and then match answers, which is common in our daily lives.
Other generative AI models can perform similar tasks for images, sounds, and videos. In addition, ChatGPT can be fine-tuned for training the process of adapting LLM to a specific task or domain by training on a smaller set of relevant data, which is the business model ChatGPT is currently expanding into, with applications in a certain niche.
Underlying Technologies
ChatGPT utilizes OpenAI's latest language model NLP (Natural Language Processing), which is based on the Large Language Model (LLM) model GPT-3 plus special techniques to fine-tune ChatGPT using supervised learning and Reinforcement Learning from Human Feedback (RLHF).
NLP, natural language processing, referring to the interaction between human language and computer, relies on the following key technologies: LSTM model and a small number of improved CNN models, RNN as a typical feature extractor; Sequence to Sequence (or encoder-decoder can also be called) + Attention as a variety of specific tasks. The LLM (Large Language Model) model, which is the model to which ChatGPT currently belongs, is a subset of artificial intelligence, and as the name implies, "large" means massive data. It has been trained on large amounts of text data to produce human-like responses to conversations or other natural language input. With algorithms comes the need for computing centers to store and process the data and computational processing, and the computing center behind ChatGPT is Microsoft's Azure cloud computing center.
ChatGPT's Positive Influence on the Automotive Industry
Driver-Vehicle Interaction: Intelligent In-Vehicle Infotainment
The emergence of ChatGPT has started to make the industry re-examine artificial intelligence, and also made the new energy vehicles (NEVs), which pursue intelligence, more aware of the lack of human-computer interaction in cars. At present, there are still many car companies whose car-machine interaction was being criticized for remaining at the level of 10 years ago. Here, we focus on voice interaction within intelligent in-vehicle infotainment systems that acquire a strong association with ChatGPT.
Voice interaction involves three main focuses: recognition, understanding and execution. According to the research from Gasgoo Auto, among the current solution providers, the technologies of recognition have become mature, with a recognization accuracy rate of over 90%. The pain point of the industry mainly focuses on the understanding part, and most of the in-vehicle voice interaction systems are not intelligent in "understanding". This is mainly reflected in two aspects: complicated operation and interaction. Firstly, most manufacturers provide voice interaction solutions through a combination of touch screen activities and voice recognition mechanisms, with different software applications of different built-in voice programs, resulting in operational inconvenience. Previously, most of the front-end voice interaction provided by traditional OEMs used command control, and users need to follow specified commands to interact, whereas AI not have the ability to understand semantics. In addition, although the recognition accuracy rate has reached a high level, drivers are, after all, individuals with independent spirits rather than a robot, and "slips of the tongue" may occur at any time. Therefore, there is a great uncertainty in voice interaction, and the lack of a system that adapts to the user's voice usage habits makes it impossible to achieve normal interaction and accomplish the goals set by drivers. As a result, accurate understanding of drivers' wording mainly involves NLP (Natural Language Processing) technologies, whose understanding of user-input speech is inextricably linked to their own scene strategies and multi-round dialogues, and directly determines the intelligence of in-vehicle voice interaction systems.
In contrast, compared with ChatGPT, the intelligence level of the current in-car voice interaction system still has room for development. ChatGPT is a typical example of one of the most basic and widespread examples of intelligent human-computer interaction. It is a computer program that responds like an intelligent entity when conversing via text or voice, and understands one or more human languages through NLP. From a technical point of view, ChatGPT is based on a large-scale pre-trained language model (GPT-3.5). With its powerful language understanding capabilities, it learns by working on large-scale data with human annotation and feedback, thus allowing the pre-trained language model to better understand human questions and give better responses.
By using ChatGPT, vehicles can interact with the driver via voice and texts and provide real-time feedback to the driver about the vehicle status, driving information and more. This enables better interaction with the vehicle and provides a better driving experience for the driver. ChatGPT can also help drivers better understand their driving behavior and offer targeted driving advice by analyzing driver behavior data, thus improving driving safety.
Autonomous Driving: Accelerated Iteration of Functional Development
The industry is divided on whether ChatGPT can be effectively applied to autonomous driving. some people believe that autonomous driving requires more graphics, image and data processing capabilities, with higher requirements for image algorithms and little relevance to natural language processing capabilities. Therefore, ChatGPT's contribution to achieve autonomous driving is currently small. However, others believe that with the improvement of the arithmetic power of in-vehicle devices and satellites, networks and other devices, AI-based intelligent driving will have stronger capabilities. The emergence of ChatGPT gives the population a glimpse of a possibility. After training, AI in achieving autonomous driving is expected to appear in a few years.
AIGC (AI Generated Content), is the technical form of AI autonomous direct production of content. From an application perspective, it is the release of ChatGPT that exploded the whole industry and made AIGC technology go viral. The launch of ChatGPT has significantly raised the expectations of AIGC technology at the application level inside and outside the industry. If ChatGPT's powerful image recognition ability and powerful deep learning algorithms are combined, it will no longer be difficult to improve the extraction speed of driving-related information. It can even complete the image annotation work, which would otherwise require a lot of labor and time. This will significantly improve the speed of data extraction for intelligent driving technologies, thus promoting the accelerated iteration of autonomous driving functional development.
In smart cockpit and smart driving software code-generating scenarios, ChatGPT, or more powerful, more professional AIGC products may replace human work. At that time, the job of human software engineers will mainly be to input instructions to the machine, and to check and optimize the code already generated by AI products.
Customer Service: Highly Efficient and Comprehensive
Leveraging ChatGPT's high level of conversational skills seems effortless when it comes to helping organizations streamline customer conversations and automate responses. Its natural language processing capabilities seem sophisticated enough to allow organizations to provide fast and effective customer assistance while maintaining quality. Answering common customer questions such as product details, cost, and availability should be a fairly easy task for ChatGPT. This can be a very valuable 24/7 support that greatly improves customer experience and allows staff to focus on more complex customer issues. Increased efficiency may be another value-add, as AI should be able to handle and process multiple customer interactions simultaneously, responding very quickly to multiple queries at once.
Streamlining the service process, on the other hand, means using ChatGPT to assist with specific tasks related to servicing vehicles, such as scheduling appointments, processing payments, and collecting information from customers. The goal of this application is to reduce manual work and increase the efficiency of the service process. In the case of vehicle repair, ChatGPT can support organizations in scheduling tasks by allowing customers to plan their own service appointments, eliminating the need for phone calls or face-to-face interactions.
In essence, in today's competitive manufacturing environment, a high-quality product is not enough to stand out. Instead, manufacturers need differentiators, and the most cost-effective way to create these differentiators is often to personalize the customer experience (CX) in the manufacturing process. Personalized marketing and sales using ChatGPT is the most obvious application area. In the automotive industry, ChatGPT might ask buyers what type of car they are looking for, their budget and desired features. Based on this data, ChatGPT can offer a selection of cars that fit the customer's requirements and provide additional information about each vehicle. This can enhance customer experience by making the buying process more efficient and fun.
Possible Risks to Consider
Heavy Reliance on Massive Training Scales and Reinforcement Learning Algorithms
Back in 2018, OpenAI released the generative pre-training Transform model GPT-1 and introduced the optimized GPT-2 and GPT-3 respectively in the following two years. The training parameter size of GPT-3 increased from 117 million in GPT1 to 175 billion. GPT-3.5, which incubated ChatGPT, has improved the training parameter size by another order of magnitude. Currently, the training scale of the GPT-4 model under development may reach 100 trillion.
Moreover, OpenAI has introduced artificial labeled data and PPO reinforcement learning algorithm for ChatGPT, which can combine the relevance model with huge parameters and human feedback to reinforce learning through feedback in interaction with humans. In other words, continuous development for ChatGPT can only be achieved via user feedback.
Security Issues for Users and Developers
Smart vehicles involve the safety of human lives, and any little mistake or hesitation may possibly result in a huge disaster. Therefore, even if ChatGPT or AIGC technology is used, it can only be applied in limited scenarios in the early stage, such as optimizing travel routes and cutting down energy consumption. Even if the related technology is mature enough, people are not yet able to sense the direction of ChatGPT's perception and learning, and it is not clear how accurate the AIGC technology is. For instance, at the level of image annotation involving intelligent driving, consequences caused by any traffic scene not being accurately identified would be unimaginable.
In addition, ChatGPT is a product that can be used by everyone. If a hacker finds vulnerability in ChatGPT, or vulnerability in a piece of code written by ChatGPT is being detected, any product that uses this software to write codes will be in danger. If vehicles are involved, privacy and security, or even life safety, are at risk. At the same time, there are many challenges at the legal and regulatory level for ChatGPT. This mainly because relevant regulations and laws are not ready and complete.