Shulou(Shulou.com)11/24 Report--

The fire of the big model is still burning.

With the iteration of technology and the development of the industry, the current attention to the large model has gradually shifted from the technological breakthrough itself to the combination with the vertical industry and commercial potential. Meta recently launched an open source commercially available version of its language model LLaMa-2, which allows startups and enterprises to build customized software based on the technology. Not long ago, Microsoft also announced that it would replace the old version of voice assistant Cortana with AI assistant Copilot constructed by ChatGPT to provide more powerful intelligent auxiliary services for Windows users. There is an industry view that after more than half a year of development, the current large model has ushered in a commercial time. However, the problems that have been criticized since its birth, such as professionalism, reliability, cost and so on, still restrict the use experience of large model technology, and limit the commercial potential of large model to a certain extent.

In the view of Teslian, which is driven by AIoT as the underlying technology, large model merchants still face the challenges of multimodal and heterogeneous data processing, and Teslian's way to break the situation is to cut into the smart city domain model and gradually build an AIoT model with multimodal capability through the way of "model + system".

"Intelligent emergence" is still limited to language tasks.

Multimodal heterogeneous data becomes the focus of AGI breakthrough

In natural science, new and unpredictable properties, characteristics, or behaviors that appear in a complex system are called "emergence". In the field of AI, with the growth of data scale and the number of model parameters, large models jump to the surface with new capabilities, which is called "intelligent emergence". "intelligent emergence" is a concept that accompanies the popularity of large models, and many people worry that AI will no longer be controlled. In the view of Hua Xiansheng, chief technical officer of Tesco, "intelligent emergence" can not solve all the problems in the field of language models, and in areas other than language models, it is more subject to multi-dimensional, multi-modal heterogeneous data.

Hua Xiansheng, Chief Technology Officer of Tesco

Shao Ling, chief scientist of Tesco and president of Tesco International, used "World Model" to explain this phenomenon. " The ultimate goal of AGI (General artificial Intelligence) is to learn and train a World Model based on its understanding of the world. The current LLMs (Big language Model, Large Language Model) is based on language training, and its understanding of the world is bound to be incomplete. Although language can be seen as a projection of human knowledge, and LLMs also has the ability of a compressed version of World Model to some extent, the defects are also significant. "

Shao Ling believes that the lack of credibility of ChatGPT, such as "serious nonsense", leads to this. " To solve this problem, on the one hand, we can make the model larger and improve the diversity and quality of data; on the other hand, we can also use domain-specific knowledge base to assist, but these can not fundamentally solve this problem. In order to achieve a more perfect World Model, it is essentially necessary to integrate language, image, video, 3D and other multi-modal data and knowledge, which is also the development trend of large models in the future. "

Shao Ling, chief scientist of Tesco and president of Tesco International.

However, at present, the generalization of large model technology in other modal data is not ideal. Take CV (computer Vision) as an example. Although CV has long been ahead of NLP (Natural language processing) in terms of technology landing, CV lags behind NLP in the battle of large models. Tesco AI R & D experts believe that this is due to the difference between the two data samples-the training of NLP is a self-generated process, and all corpus are tagged by the former corpus to predict the latter; while the images collected through the Internet are difficult to be directly used for training, and the image data itself has higher dimensions, more redundant information, and more difficulty in learning.

In the field of AIoT, the data in AIoT scene is usually multimodal and heterogeneous, including text, image, voice and sensor data. To integrate and use these different types of data effectively, we need to overcome the challenges of data representation, fusion and alignment. In the aspect of data modeling, there is no final conclusion on the form of modeling of all kinds of sensor data-whether the reference text is used as the sequence input, or the reference image is used as the matrix input, or it is a brand-new modeling method. it has also become an urgent breakthrough topic for the development of large models.

Teslian builds a big model of AIoT:

Richer perceptual dimensions and stronger ability to act

Against the above background, Hua Xiansheng, CTO of Tesco, divides the AIoT model into two types based on the form of implementation:

1. A model that introduces multimodal IoT data into the model and achieves practical application results.

The domain model of 2.AIoT technical service related domain.

He said frankly: "the bottom layer of the AIoT model that can be seen on the market is still text and vision, not the real AIoT model." Teslian's approach is to cut in from the domain model and gradually expand to the multimodal data model through communication with the system.

Earlier this year, Teslian, as a representative company in the field of metro AIoT, revealed that it was developing a multimodal large model for AIoT scenarios. Teslian's large model portfolio consists of a basic general model and five vertical domains, covering global perception, energy dual-carbon, embodied intelligence, economic brain, system interaction and other scenarios.

Hua Xiansheng interprets the value of the AIoT big model built by Teslian with "richer perceptual dimensions and stronger action ability". "Teslian has a large number of actual running AIoT devices and data, and is currently querying and analyzing AIoT data based on the large model, manipulating and managing AIoT devices, and giving the personalized experience feedback of each user in the scene to the big model for analysis and optimization, so as to build an AIoT scene agent."

Teslian Chongqing AI CITY

According to Teslian's planning, the representation of large models under its smart city business is mainly divided into three categories: large model-based super smart park system, large model-driven AGI robot, and large model-driven visual design. Take the large model-driven robot as an example, the robot defined by Teslian includes hardware robot and software robot. Like the Moss in wandering the Earth, the AGI robot driven by a large model knows everything about the park and can be everyone's "secretary". " Each AGI robot is actually a small model, but when combined with the large model behind it, it can become a powerful agent. It can be everyone's receptionist, agent, or secretary. It knows what you need and can answer your questions. Of course, it's not just the AGI robot. Driven by the AIoT model, every robot, every screen, and even everyone's mobile phone and APP can become an 'intelligent robot'. " Hua Xiansheng shares further. Driven by the large model, the future smart city will have four characteristics: man-machine coexistence, data-reality integration, intelligent evolution, green and low-carbon. These characteristics have been seen in AI CITY, the flagship product of Teslian's three intelligent cities in Chongqing, Wuhan and Deyang. With the data flywheel of "data promotion model and model change data", its AIoT model is constantly evolving, and AI CITY is gradually growing into an organic whole with vitality.

Large models are not the only path to AGI

Model + system may be the way to solve the problem

Since the birth of ChatGPT, many people in the industry think that ChatGPT has touched the edge of AGI. In Shao Ling's view, however, the big model is not the only path to AGI. "the current LLMs's understanding of the world is still very superficial," he said, citing Yann LeCun, the 2018 Turing Award winner and father of convolutional neural networks (CNN).

Yann LeCun has publicly criticized the current LLMs craze many times. He believes that the original AI model is still very useful for specific domain tasks corresponding to training data, such as translation and image recognition. However, LLMs,Yann LeCun, represented by GPT and BERT, thinks that they are essentially autoregressive text generation models, and because they only train on the text, they can only understand the world very shallowly.

In response to the goal of "achieving the Human level AI", in its position paper "A Path Towards Autonomous Machine Intelligence" published in June 2022, Yann LeCun raised three revolutionary questions:

Question 1: how do machines learn to represent the world, to predict, and to act on observation?

Question 2. How can machines reason and plan in a way that is compatible with gradient-based learning?

Question 3. How can machines learn representation perception and action planning in a hierarchical way and at different levels of abstraction and time scales?

Yann, LeCun (2022). A Path Towards Autonomous Machine Intelligence, Courant.

Interestingly, when we ask the same question ("is large model technology the only path to general artificial intelligence?") To challenge ChatGPT, ChatGPT also gives the answer: large model technology is an important progress in the field of artificial intelligence, but it is not the only path to general artificial intelligence. The realization of general artificial intelligence may need to combine a variety of technologies and methods to carry out research and development in an interdisciplinary way.

Hua Xiansheng elaborated on the connection between AI and scene, platform and system on the TEDxHuangpu stage.

In Hua Xiansheng's view, the combination of models and systems is another path to relatively general intelligence. He points out that there are three paths to AGI today: big model, new model, and "model + system". "our definition of system is broader, even including the fusion of large and small models, the integration of models and rules, and the transformation of models and rules, and so on."

"if a large model can land in the field of AIoT, it needs to be deeply coupled and connected with the system. After communicating with the system, we can realize the use of IoT-aware data and the control and scheduling of IoT equipment. Language is a way of interaction, and communicating with the system is to translate the language into instructions that the system can execute. Instructions may be to obtain information, may be analysis, control, or control based on analysis. This is currently the easiest way for us to implement a large multimodal model, and it is even a faster path to AGI that we can try. "

The robot in AI PARK has the ability of vehicle environment coordination, which can complete task scheduling in complex environment.

Although the system is a relatively abstract concept, Qian Xuesen once defined the system as "an organic whole with specific functions, which is composed of several components that interact and depend on each other." From this, we can easily see the big model ambition of Teslian, which is to give the inner vitality of the future city with the deep embedded model. With the development of artificial intelligence from the perceptual stage to the cognitive stage, agents from perception to cognition and then to execution can also be seen as a complete system. "maybe the machines of the future will have a sense of touch and smell, and machines will be able to understand what they mean and make decisions through terminal devices. This is how we understand the landing of AGI technology in cities," says Hua Xiansheng.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.