At WAIC 2024, there are many strong players.
Not only traditional Internet manufacturers such as Baidu, Alibaba, and Tencent,
There are also players in vertical fields such as Kuaishou, Bilibili, Liepin, and Ape Tutoring.
There are also companies focusing on the AI field, such as Baichuan, Facewall Intelligence, Minimax, and Mobvoi.
In addition, the conference attracted Amazon, Google, Qualcomm and other international manufacturers to participate in the exhibition.
Finally, traditional industry companies such as State Grid, China Unicom, Bank of Communications, and China Shipping have also entered the AI field and demonstrated their innovative achievements.
During the WAIC period, various companies made major releases and updates, including SenseTime's Ririxin 5.5, Zhipu's CodeGeex code model, Baidu's Wenxin model 4.0, Mido's government affairs model 3.0, and NetEase's robot brand Smart. The latest technologies and products displayed by major manufacturers have become the highlight of this conference.
SenseTime released Rixin 5.5
On July 5, SenseTime released the first large model with streaming native multimodal interaction capabilities in China, "Ririxin SenseNova 5.5", at WAIC 2024. Compared to the "RiRixin 5.0" released two months ago, the performance of the new version has been improved by 30%.
"RiRixin 5.5" contains a 600 billion parameter base model, using synthetic high-order thinking chain data, and the reasoning thinking ability is significantly enhanced. It excels in mathematical logic, English, and instruction following.
SenseTime's "WYSIWYG" model "5O" brings a new streaming multi-modal interactive experience. By integrating cross-modal information, RiRixin 5o can process various forms of information such as voice, text, images, and videos in real time, enabling natural and smooth AI interactions, bringing an unprecedented user experience.
The device-side model "Ririxin 5.5 Lite" has been fully upgraded, improving accuracy by 10%, inference efficiency by 15%, and first packet delay by 40%. This model is even better than GPT-4o in terms of multimodal capabilities.
In addition, SenseTime's "controllable" character video generation model Vimi is also one of the treasures of this WAIC.
智谱发布CodeGeex代码大模型
At the Industry Forum of the World Artificial Intelligence Conference on July 4, Zhang Peng, CEO of Zhipu AI, pointed out that large models bring new opportunities for human-like cognitive ability. Large models not only improve the versatility of AI, but also significantly reduce costs, making them an important tool for empowering the real economy.
On July 5th, at the GLM-New Generation Base Large Model Technology Frontier and Industrial Application Forum, Zhipu AI released the latest CodeGeeX4-ALL-9B code large model. The model performs well in code generation, code completion and interpretation, network search, etc., covering various scenarios of programming development.
CodeGeeX4-ALL-9B has performed well in multiple authoritative code capability evaluations, and is the most powerful model with less than 10 billion parameters. It has more than 1 million individual users and is available for free download in all major IDEs.
Zhipu AI's booth showcased the bigmodel.cn platform and its series of innovative achievements, attracting a large number of visitors. The Zhipu Town on the platform shows the application of large models in public affairs, medical care, finance and other industries, providing diversified intelligent services for enterprises and users.
Mido released the government affairs model 3.0
On July 4, Mido released the newly upgraded Honeycomb Government Model 3.0 at the World Artificial Intelligence Conference, showing the latest achievements in more than 20 core application scenarios such as government hotline, government intelligent Q&A, and judicial document proofreading.
Honeycomb 3.0 has been comprehensively improved in terms of R&D ecology, training data, and model architecture. The localization of the whole process, high-quality datasets, and high-performance hybrid architecture are the three highlights, which significantly improve the application efficiency in government affairs scenarios.
In terms of R&D ecosystem, Honeynest 3.0 integrates the localized software and hardware ecosystem, and uses Ascend hardware, CANN heterogeneous computing architecture, and MindSpore AI framework for the entire process from training to inference, providing customers with safe and controllable solutions.
In terms of training data, Honeynest 3.0 adds 20 billion pre-trained corpus, 100,000+ fine-tuning data and 10,000+ alignment data, improving the professionalism, accuracy and security alignment ability of model output.
In terms of model architecture, Honeynest 3.0 adopts the Llama Pro+MoE architecture, which has stronger ability to handle complex tasks.
The newly released Intelligent Public Opinion V Assistant 1.0 brings an intelligent and efficient new experience to public opinion analysis.
Baidu is free to open Wenxin Model 4.0
On June 28, at the WAVE SUMMIT Deep Learning Developer Conference 2024, Baidu CTO Wang Haifeng announced the latest data of Wenxin Yiyan and officially released Wenxin Large Model 4.0 Turbo and PaddlePaddle Framework 3.0.
Immediately, from July 5th, Baidu Wenxin Intelligent Twins Platform (AgentBuilder) will open Wenxin Large Model 4.0 for free. Developers can flexibly choose to use version 3.5 or 4.0 of the Wenxin model on the platform to make agents.
NetEase released the robot brand Smart
On July 4, NetEase Fuxi released its first robot brand "Smart" at WAIC 2024. Based on the self-developed industrial model and AOP technology ideas, the brand has launched excavation robots and loading robots, and has participated in 50 key construction projects in many provinces.
NetEase Smart Excavation Robot has functions such as automatic loading, one-key slope brushing, and one-key leveling, and can operate continuously for 10 hours in extreme environments, with a single machine efficiency of more than 80% of that of a real person. The smart loading robot realizes all-weather unmanned loading operations in the concrete batching plant, which effectively alleviates the recruitment problem in the construction machinery industry.
NetEase adheres to AI innovation and pragmatism, not only making progress in the field of engineering, but also applying AI technology to games, education, music, and other fields. The "VR Large Space" product launched by NetEase Yaotai helps the integrated development of online digital scenes and offline physical scenes.
Conclusion: Focus on practical application and innovation
At this WAIC 2024 Industry Forum, Robin Li's remarks sparked widespread discussion. He pointed out that the focus of the domestic large model should be on practical application, rather than just the college entrance examination questions. I couldn't agree more, and ultimately the market will favor apps that actually solve problems.
He also mentioned the problem of wasting computing power to do duplicate work. I have a grain of salt. Each company has its own R&D goals, and it should be up to the market to decide whether there is duplication and waste.
Robin Li believes that open source is an IQ tax, but I believe that open source and closed source complement each other, and the contribution of open source to technological progress cannot be ignored. Through open source, more people can participate in technological innovation and jointly promote the development of the industry.
WAIC 2024 showcases the blockbuster announcements of many major manufacturers, and these technological advancements and innovative applications will lead the development direction of future technologies. It is hoped that major companies can continue to focus on practical applications and bring more practical and efficient products and solutions to the market.
If you find this article helpful, please like, bookmark, retweet and share. In the meantime, please follow me for more updates and insights on artificial intelligence!