Every breakthrough in the generalization of key technologies will profoundly change the mode of production and significantly improve the level of productivity, and the "large model" has become the key to opening up the "two pulses" of the versatility of artificial intelligence technology. Wisdom Eye, as the world's leading artificial intelligence company, has recently launched its innovative research and development of a general large model - Toushi.
AI is rapidly evolving from perceptual and cognitive intelligence to decision-making intelligence.
In the 1.0 era, perceptual intelligence is represented by computer vision, that is, based on deep learning algorithms, empowering computers to understand digital images and videos, and it is at this time that Wisdom Eye built a large machine vision model, and entered the top seven in IDC's Chinese artificial intelligence computer vision application market share ranking, and won the highest award of China's intelligent science and technology - Wu Wenjun Artificial Intelligence Science and Technology Progress Award.
In the 2.0 stage, cognitive intelligence will draw inspiration from cognitive psychology, brain science and human social history, and combine cross-domain knowledge graph, natural language processing, causal reasoning and other technologies to establish an effective mechanism for stable acquisition and expression of knowledge.
In the 3.0 era, AI+X has become the frontier direction, and it is necessary to use massive high-quality general data to train a basic base, and then use a small amount of professional data for fine-tuning for different scenarios to achieve low-cost AI scenario implementation. Wisdom Eye has developed a general large model Toushi, which is based on the training of massive multimodal data with strict review and filtering, and brings a new standardized AI R&D paradigm through the "pre-training + fine tuning" mode, realizing the large-scale production of AI models in a more unified and simple way and supporting the execution of multiple tasks.
The large model generates code, and the high-quality code dataset is obtained by cleaning the open source code dataset many times, and then through pre-training and CoT training, etc., the large model has strong code generation and inference capabilities, which greatly improves the efficiency and quality of software development.
On the basis of more parameters and more complex structures, the large model of Toushi uses massive general text-based data for pre-training, learns a lot of general knowledge and logic, shows strong emergence, and generates poems rich in rhyme, rhythm and emotional expression.
Based on natural dialogue, the Toushi model understands and executes tasks, provides a variety of capabilities including text, image, video, and audio input, generates text output, and has the ability to chat and question and answer, which can empower scenarios such as emotional companionship, policy Q&A, and big data analysis.
Enabled by three core technologies, a new generation of large models has emerged
The innovation of large models is far more than the innovation of the model itself, and it depends on various core technological breakthroughs in the field of artificial intelligence at a deeper level. In this regard, Wisdom Eye has significant advantages, and it has independent research and development of core algorithms and independent intellectual property rights in key technologies such as knowledge graphs, intelligent large models, and privacy computing, which provides a solid foundation for the innovation of general large models. Include:
In terms of knowledge graphs: the company breaks through the distributed knowledge transfer and federated graph inference for privacy protection, accesses differentiated large-scale industry data, and builds a world-class knowledge graph of massive data with high credibility and high quality, laying the foundation for the knowledge graph to alleviate the illusion of large models.
In terms of intelligent large models: the company has made breakthroughs in core technologies such as large model federated training under the condition of privacy computing and knowledge enhancement of large models based on knowledge graphs, which has significantly improved the accuracy of information generation of large models in professional fields.
In terms of privacy computing, it has made a breakthrough in the distributed knowledge transfer and federated graph inference technology to protect privacy, realize the desensitization of sensitive privacy data of high-value users, truly realize the "usable and invisible" security application of data, realize the fusion of distributed knowledge and the training of large models based on distributed data, and support the release of user data value.
The four basic models are integrated to jointly shape super general capabilities
The core positioning of the smart eye stone model is to inject new vitality into all walks of life, using a complete hierarchical decoupling design, including two layers, the first layer is the stone general large model layer, mainly including machine vision, natural language understanding, multimodal and cognitive four basic models, providing relevant AI basic capabilities for various application scenarios, and the second layer is the industry large model, based on the general large model layer, which can integrate the own data of industry users and train the customer's own proprietary large model. Through this hierarchical design, the smart eye stone model not only provides a wide range of general capabilities, but also realizes highly personalized empowerment according to the specific conditions of various industries.
Specifically, the base model level includes:
Large model of machine vision, "implanting eyes and brains for machines". By training on large-scale image data, this model can realize various visual tasks, such as image segmentation and recognition, and video analysis.
The natural language understanding model "allows machines to understand human language". This model is trained based on the high-quality corpus of T Token, which can realize the responsible natural language processing tasks.
Multi-modal large model, "let the machine have rich senses". This model simulates the way people interact with each other through the cognition and understanding of different stored information carriers such as text, images, video, and audio, providing a solid foundation for innovative applications.
Cognitive large models, "let machines think, analyze and make decisions". This model can adapt to the input of different types of perception data sources, including various types of sensors, to achieve the understanding and analysis of the environment.
At the level of industry models, Toushi will cover multiple professional fields such as medical care, medical insurance, finance, insurance, government affairs, pension and education, and the core value of these large models is to provide strong empowerment for the development of different industries.
In summary, the basic model is committed to overcoming technical problems and deeply exploring the universality and generalization, while the industry model is deeply integrated with industry data and knowledge characteristics to build a solid foundation for specific industries. The basic model provides strong support for the construction of the industry model, and at the same time, the industry model can provide valuable feedback and nourishment for the further optimization of the basic model by combining actual application scenarios and real data.
Large models are an important milestone in the field of artificial intelligence. Looking forward to the future, Wisdom Eye will uphold the spirit of continuous innovation, lead the wave of AI technology development, and promote large-scale model technology to be surging like mountains and rivers, covering a wide range of industries such as medical care, medical insurance, finance, insurance, government affairs, pension, education, etc., so as to inject new and high-quality productivity into social development.