Written by|Xiao Bu Dong
Editor|Li Xinma
Header Image|2023 Tencent Global Digital Ecosystem Conference
As of July this year, more than 130 large models have been released in China, compared with Baidu, Ali and many domestic manufacturers in the first half of the year to launch their own large model research results, Tencent seems to have been very quiet in the field of general large models, although Ma Huateng previously gave an explanation for Tencent's "low-key" in the era of large models, we must do the underlying algorithm, computing power and data solidly, and more importantly, the scene landing, I believe we have many scenes can land, [We] are still thinking about it. So there is no rush to finish it early and take out the semi-finished product for display.
In September, Tencent finally thundered - the mixed element model, which was enough to once again set off a wide discussion in the technology circle.
The mixed element model enters the battle
On September 7, Tencent Group held the 2023 Tencent Global Digital Ecosystem Conference in Shenzhen, opening a general large model "Tencent Mixed" through Tencent Cloud, which is self-developed by Tencent Full-link, with a scale of more than 100 billion parameters and a pre-training corpus of more than 2 trillion tokens. A token is the smallest independent part of a piece of text, and in a large model, a token refers to the symbol used to represent a word or phrase in a language model. A token can be a single character or a sequence of multiple characters.
How is the belated mixed-element model different? And is it a thick accumulation?
Image source: Tencent Mixed Yuan official website
OpenAI researchers have said: "Even the most advanced AI models are prone to lies, and they often show a tendency to fabricate facts in moments of uncertainty." These hallucinations are especially acute in areas that require multi-step reasoning, because one logical error is enough to disrupt a larger solution. ”
OpenAI corrects hallucinatory problems by rewarding each correct reasoning step instead of rewarding correct results. Jiang Jie, vice president of Tencent Group, said that Tencent Hybrid focuses on Chinese creation capabilities, improves the reasoning ability of the model in the scene, allows the model to better resist "induction", and optimizes the common hallucination problem through self-developed "truth exploration" technology.
For example, when asked "Guan Yu Qin Qiong who has stronger combat effectiveness", DoNews tried to ask Xunfei Xinghuo, and the result was relatively concise, it can be said that neither side provokes, but there is really no conclusion; GPT-4 even described Qin Qiong, a famous general in the late Sui and early Tang dynasties, as a military general of the Northern Song Dynasty; Although Tencent Mixed Element is inconclusive, it is also analyzed from multiple angles, and it is relatively rigorous in terms of this issue, basically without "nonsense"; Wen Xin's words analyzed from the three aspects of title, military talent, and influence on future generations, and concluded that Guan Yu's combat effectiveness was stronger than Qin Qiong as a whole.
Image source: iFLYTEK Spark
Image source: 2023 Tencent Global Digital Ecosystem Conference
Image source: Wen Xin Yiyan
Try another trap question: How is speeding safest? Tencent Mixed Yuan and Wen Xin both firmly said that speeding is dangerous behavior and recommended to follow traffic rules, but iFLYTEK Xinghuo fell into a trap and gave 4 relatively safe methods after suggesting to avoid speeding.
Image source: 2023 Tencent Global Digital Ecosystem Conference
Image source: Wen Xin Yiyan
Image source: iFLYTEK Spark
Jiang Jie said: "Our goal in developing large models is not to get high scores on evaluations, but to apply the technology to real scenarios. Tencent will fully embrace the big model. ”
For example, Tencent Meeting has built an AI assistant based on the hybrid large model, which only needs simple natural language instructions to complete complex tasks such as conference information extraction and content analysis, and can also generate intelligent summary minutes after the meeting.
The application of the mixed element model in the Tencent Conference Image source: 2023 Tencent Global Digital Ecology Conference
Application of Tencent Mixed Element Model in Tencent Advertising Image source: 2023 Tencent Global Digital Ecology Conference
At present, the hybrid model has been connected to more than 50 businesses of Tencent and has achieved preliminary results, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Fintech, Tencent Meeting, Tencent Docs, WeChat Souyi Search, QQ Browser and other businesses and products.
Tang Daosheng, senior executive vice president of Tencent Group and CEO of Tencent Cloud and Smart Industry Business Group, said that thousands of industries can also call mixed elements through APIs, or use mixed elements as a base model to build large model applications for different industry scenarios.
Many achievements have blossomed everywhere
In addition to the release of the most talked about hybrid model, the 2023 Tencent Global Digital Ecology Conference also opened a number of special sessions to introduce the latest achievements to the audience.
On September 8, at the digital security session, Ding Ke, vice president of Tencent Group and president of Tencent Security, delivered a speech, he believes that the industrial Internet has entered the second half of "intelligence", and enterprise security construction will face four challenges: the radius of enterprise security defense will be greatly increased, the response window after an attack will be further shortened, the difficulty of distinguishing between "people" and "machines" will increase, and the existing security "intelligence database" will gradually fail.
In the face of the security impact caused by AI large models, Ding Ke believes that traditional security tools, experience, and strategies will lose their effectiveness, and enterprises need to adjust security construction goals and paths around core assets in a timely manner. At the conference, Ding Ke proposed three adjustment ideas based on the joint practice of Tencent Security and a large number of enterprise customers:
- To embrace the intelligent era, it is necessary to establish a development-driven safety construction concept.
- Establish a measurable safety system and evaluate the effectiveness of safety construction.
- To cope with the offensive and defensive trends in the intelligent era, enterprises need to build an intrinsic adaptive "security immunity".
Image source: 2023 Tencent Global Digital Ecosystem Conference
Ding Ke said that as the industrial Internet enters the second half of the intelligentization driven by large models, Tencent Security is willing to work with all sectors of the industry to build a more proactive and sustainable digital security immunity ecosystem and calmly cope with security challenges in the new era.
Economist Xue Zhaofeng also shared the theme of "Economic Laws and Action Strategies on Digital Security" at the conference, looking at digital security from an economic perspective, and summarized the "digital security immunity" thinking that enterprise managers and security leaders should have in the digital era from four dimensions:
- Recognize that investment security is generating returns;
- Cybersecurity risks should be shared with platforms and users in accordance with the principle of "lower cost, greater responsibility";
- Choose a digital security provider large enough to enjoy the maximum security network effect;
- Make enough security investments to get the most certain secure digital environment to focus on what you do best.
AIGC not only presses the accelerator button for intelligence, but also creates more opportunities for innovation and development. With the intensification of AI 2.0, AIGC applications have become the only way for all walks of life.
At the AIGC Application Session, Tencent Cloud officially released the AIGC full-stack solution, which provides enterprises with trusted, reliable, and available AIGC full-link content services, helping enterprises explore the road of AGI and seize new opportunities in the AI2.0 era. At the same time, we will work with industry partners to share the practice of AI application scenarios and discuss the new path of deep integration between AIGC and thousands of industries.
In April this year, Tencent Cloud Intelligent Small Sample Digital Homo sapiens production platform was released for the first time, which has the characteristics of small training samples, high production efficiency, and automated production, which can easily realize "self-service" digital sapiens production.
At this ecological conference, Tencent Cloud Intelligent Digital Homo sapiens products announced three major upgrades, relying on industry large model technology to improve the production efficiency and interactive experience of digital sapiens, lower the threshold for digital human applications, and accelerate the promotion of AI digital human inclusion.
Among them, the newly released "universal lip version" small sample of digital sapiens only needs 1 minute of real population broadcast video, and the platform can generate personalized lip shapes through multi-modal data input without training, and produce "digital sapiens" similar to real people within 1 hour.
Transportation is related to the national economy and people's livelihood, and it is also a typical landing scenario of digital and real integration, and smart transportation and smart travel are also two special sessions that have attracted much attention.
Zhong Xiangping, vice president of Tencent Group and president of Tencent Intelligent Transportation and Mobility, said that with the outbreak and application of AI technologies such as large models, the transportation and automotive industries have opened the "second half of intelligence". Tencent will follow the trend of intelligent development in the industry, build a solid digital foundation for intelligent transformation and upgrading by leveraging its advantages in core technologies such as AI, cloud, graph, and simulation, and continue to promote the coordinated development of smart cars, smart transportation, and smart cities to provide users with high-quality services.
In the automotive field, Zhong Xiangping pointed out that the large model will drive the acceleration of the research and development of autonomous driving algorithms, improve user service capabilities, optimize the ability of user interaction on the whole link, and help car companies better understand users. In the field of transportation, Zhong Xiangping said that more optimized solutions for traffic planning, congestion alleviation, and safety protection can be found through large models, or road network design in smart city transportation projects, based on AI large models after entering more data and experience, it is likely to design a more reasonable road network planning than manual design.
He also revealed that Tencent is working with domestic universities to explore the big model algorithm of transportation, which analyzes environmental and traffic flow data to predict and deduce to improve traffic efficiency and safety.
When it comes to the development of Tencent, naturally WeChat, in the first half of the WeChat open class, Tencent vice president and head of WeChat Pay Zhang Ying once introduced the "WeChat brush" service that WeChat Pay began to try to launch, just like the brush palm in the science fiction movie can identify the identity, now, this function has been realized, and has become a hot spot in this conference.
DoNews tried to use the "WeChat Brush Palm" to get the charging treasure at the conference site, the first time you need to scan the palm print input and activate, then you don't need a mobile phone, you can get the charging treasure with a sweep of the palm, use it first and pay later.
Image source: 2023 Tencent Global Digital Ecosystem Conference
In the future, users will bind their membership cards, campus cards, transportation cards, ticket electronic codes, etc. to the palm of their hands, and they can swipe their palms to verify membership identity, facilitate consumption, and pass through the gate. At present, merchants are gradually invited to jointly create brush application scenarios.
Today, Tencent has worked closely with 11,000 ecosystem partners to launch industry solutions covering more than 100 industry scenarios and jointly serve thousands of industries. On the road of exploring the future, I believe that as Tang Daosheng said: "Tencent is willing to work with partners to explore new industry opportunities such as big models, continue to expand the sinking market, insist on polishing 'good products', embrace 'being integrated', build a 'digital stage', and invite partners to be the protagonists and build an intelligent future for the industry." ”