I saw a very interesting conjecture "In 2050, AI will be widely used in people's lives;In the year 2100, AI can already imitate human thinking, and it looks like a real personIn 2900, artificial intelligence was born, AI had autonomous consciousness, human beings controlled all the resources of the earth, anti-gravity vehicles were born, and human beings entered the first-class civilizationIn the year 4000, human beings mastered the resources of the solar system with the help of artificial intelligence brains and entered the second-level civilizationIn 5000 years, due to the limited life span of human beings, it was impossible to break through the solar system and explore the Milky Way, and the artificial brain fused all human consciousness, and the superbrain was born, and in the years that followed, it created an artificial wormhole and marched to the Milky Way ......”
The picture is drawn by AI.
After all, AI has become the hottest outlet of the year, and AI models such as OpenAI, Google Gemini, Wenxin Yiyan, and Ali Tongyi Qianwen have emergedWhat does this have to do with cars?It's too relevant, more and more AI large models are starting to be installed, and a new era is coming. What is an AI model?
AI large models refer to large-scale neural network models built using deep learning Xi technology, which have hundreds of millions of parameters and can be trained on large amounts of data to demonstrate strong language understanding, generation, and reasoning capabilities in a variety of tasks and domains. 
These large models are often based on a transformer architecture that uses a self-attention mechanism to process sequential data, such as text or speech. The self-attention mechanism enables the model to focus on different parts of the sequence as it processes the input sequence and to establish associations between contexts. 1.BERT, which is a Transformer-based bidirectional encoded representation model;2.GPT, which is a transformer-based generative pre-trained model;3.T5, which is a Transformer-based text-to-text conversion model. Building and training these large models requires a lot of computing resources and data, so they are often developed and maintained by large research institutes or companies. 
Calculation, data", you know why the AI large models are "squeezed" into the car!First, the computing power of the car is strong, and the maximum production car is more than 1000tops;Second, there are many sensors, which can receive all kinds of data, hundreds of sensors for a car, and more than 20 for a mobile phone. Cars are constantly moving, and road variables, building variables on both sides, pedestrian variables, and even weather variables will generate a large amount of data, and the AI model is constantly receiving data and "calculating" it, so that it is possible to become more and more intelligent. 
Seeing this, some smart friends have found that AI large models can be used in two directions, one is assisted driving, and the goal is to achieve unmanned driving;The second is the intelligent cockpit, which aims to communicate without barriers. We will not discuss how the AI model worksWhat are the specific neural nodes of a neural network and what are their functions?These are too complicated for us ordinary consumers to understand, and this article only talks about what changes it can bring about when it is carried on the carWhat kind of cars are on board?
Which cars are equipped with AI models?The most interesting thing is, of course, the ideal, the recent OTA 50 officially introduced the concept of AI, which is embodied in intelligent driving and intelligent cockpit. The Ideal AD intelligent driving algorithm has been upgraded from the original modular algorithm architecture based on multiple small models and manual rules to an end-to-end algorithm architecture based on large models. 
ad pro 3.0 integrates important capabilities such as BEV large model, MPC model** control, and spatio-temporal joint planningad max 3.0 takes this a step further and adds the occupancy network algorithm. 
The large BEV model can generate information about the environment around the vehicle by capturing images from multiple cameras, and then converting this information into a floor plan for processing and analysis by the autonomous driving system. The main advantage of MPC is that it is possible to optimize the current time slot while taking into account future time slots (future events). 
The occupancy network algorithm is a 3D reconstruction and scene understanding method based on deep learning Xi, which has strong representation Xi ability and generalization ability. Taken together, these things become a system that can "analyze and make decisions on its own." The Ideal L series model (MAX version) can handle complex scenarios such as autonomous entry and exit of complex ramps, autonomous lane changes in congested road sections, and autonomous passage through mixed pedestrian and vehicle sections in urban areas. After all, Huawei, Xiaopeng and other companies have long been able to implement relevant functions, and the technical principles are similar, that is to say, many car companies have already used AI models. 
It's just that because there is no visible carrier (such as a screen) for users to perceive, the concept of AI is not obvious. So, the chatbot appeared, of course, this is the most superficial explanation, the professional term is still xxgpt. 
According to the official statement, the Ideal internal team reconstructed the original base model of Mind GPT, giving Mind GPT a variety of capabilities such as "understanding, generating, perceiving, remembering and reasoning".
Mind GPT can accurately identify fast and complex functional commands, and create "shortcuts" that match the needs of the car ownerMind GPT can also compose poems, lyrics, and even generate paintingsota 5.After 0, the ideal car machine also brings a browser function, and the user's entertainment channels are more abundant. Judging from the current features, the ideal Mind GPT is not much different from other GPTs. However, remember the earlier mention that "building and training these large models requires a lot of computing resources and data"?Compared with giants such as Microsoft and Google, and even with Alibaba, the ideal resources may be somewhat weak.
The time-saving way to save time and effort is to directly connect to a large model, such as Mercedes-Benz, mainly through a partnership with Microsoft, using Azure OpenAI Service to integrate ChatGPT into the car, with the help of the preview version of ChatGPT in Azure OpenAI Service, developers can integrate customized AI experiences directly into their own applications. It is worth mentioning that Mercedes-Benz is a fee-based service and takes the form of a monthly subscription.
In addition, Musk's AI company XAI's first large-scale model product, Grok AI, was also officially released, in addition to linkage with social platforms (formerly Twitter), Musk also retained the possibility of Grok AI on Tesla. Domestic models, in addition to ideal, Star Era ES is equipped with iFLYTEK's Spark model;Jiyue 01 will be equipped with a word of Wenxin;Zhijie S7 is equipped with Huawei Pangu large model;The Zhiji LS6 will be equipped with Alibaba's Tongyi Qianwen ......It is foreseeable that the AI model will be the new "highlight" configuration, and more and more models will be equipped with it, and with the development of technology, it will likely trigger an industry revolution. Limitations
Remember the conjecture at the beginning?AI has been gradually improved over time, and the AI installed in the car is only in its infancy, and many functions may not satisfy us. Or borrowing an ideal example, voice control is better than before, but there are still some voice commands that it can't recognize, to put it bluntly, the current AI is still in the stage of imitating humans, and even the imitation is a bit clumsy. Therefore, at this stage, AI still needs a large amount of data for training, and finally reaches "human-like", or even surpasses human thinking, as for how long this time, no one knows. At present, there are still some problems with large AI models: computing resources are limited, and the underlying layer developed by some car companies may soon encounter bottlenecks
At the level of intelligent driving, the acquisition and annotation of data may be a difficult problem, especially for data in some rare scenarios and complex situationsAt the privacy level, the in-vehicle environment has very high requirements for security and privacy protection, so it is necessary to establish a sound security and privacy protection mechanism to fully protect model data and user privacy. Write at the end
Although the AI model has not yet been perfected, it does bring a better experience in intelligent driving, human-computer interaction, etc., I believe that with the deepening of training, it is not false that the car is getting smarter and smarter, let us look forward to AI leading the car revolution!