“Science and Technology Innovation Board Daily”, May 31 (Reporter Zhu Ling) Until the end of May, the hot trend in the AI application market remained unabated. On the 30th, the AI assistant app “Tencent Yuanbao” based on the mixed-element model was launched, marking that BAT has finally gathered in the field of AI consumer C-side applications.

According to reports, since its debut in September 2023, the parameter scale of Tencent's mixed-element model has been upgraded from 100 billion to trillion yuan, the pre-training corpus has been upgraded from trillion yuan to 7 trillion tokens, and the first to a multi-expert model structure (MoE). The overall performance has increased by more than 50% compared to the Dense version.

Yuanbao targets work efficiency scenarios and provides AI search, AI summarization, and AI writing capabilities; for everyday life scenarios, it also has richer gameplay, and provides various special AI applications such as oral sparring, super translators, and versatile AI avatars. At the same time, gameplay such as creating personal agents has also been added.

The “Science and Technology Innovation Board Daily” reporter made a powerful PK between Tencent Yuanbao and OpenAI's latest GPT-4O and Byte's personal assistant Doubao apps.

▍ AI efficiency tool test: Tencent Yuanbao's ability to grab information and read links is outstanding

According to the data, when people currently use products related to big models, more than 65% of the demand is focused on work/learning efficiency scenarios. In response to the three core requirements of the efficiency scenario, which are information acquisition, processing, and production, Tencent Yuanbao has all carried out commercialized exploration.

First, AI search ability competition.

When answering the question “What are the recent major events in the field of AI models in the world”, Tencent Yuanbao and GPT-4o all used a classification method to organize the answers. However, Tencent Yuanbao provided 24 highly time-sensitive references, most of which were published within the past week, making it easy for users to quickly trace their origin and further reading.

According to reports, backed by the strong support of the Tencent ecosystem, Tencent Yuanbao has effectively integrated resources from various platforms such as WeChat Search and Sogou Search, surpassing the traditional search model. The reporter clicked on the reference link in the answer and verified that the content mainly comes from high-quality resources within the Tencent ecosystem, such as WeChat accounts, as well as authoritative information sources on the Internet.

Although GPT-4o also classified information, it only provided 6 reference materials, far less than ingots, and included data from the beginning of the year, and the timeliness of the information was poor. As a result, Tencent Yuanbao has stronger AI search capabilities, can provide users with more accurate, comprehensive and timely information, and effectively improve content generation results.

Second, AI summarizes the ability competition.

Judging from the input method, Yuanbao can upload up to 10 documents in various formats such as PDF, Word, and txt, and can parse multiple WeChat public account links and URLs at once, and supports 256K native window contexts. Although GPT-4O can also summarize link content, it does not support generating domestic link summaries.

The reporter submitted links to four WeChat articles. Yuanbao analyzed the content of each article, not only accurately distinguishing the main topics of each article, but also revealed in detail the logical connections between the articles, showing the ability to integrate complex information.

Yuanbao also showed keen product details. The reporter uploaded the “Stanford University: 2024 Artificial Intelligence Index Report” file. Yuanbao first analyzed the document's recognition size and word count, and also carefully filled in the default prompts. This is a feature that GPT-4o does not have. It is worth mentioning that even for 400,000-word documents, ingots can be parsed within a few seconds, which is faster than GPT-4o.

However, when comparing GPT-4o and Yuanbao's analytical answers to the document, the reporter observed that GPT-4O performed better in providing depth and reliability of information. GPT-4o's response is more detailed and systematic. It not only provides point-by-point answers under each topic, but also introduces specific data as support, making the arguments more persuasive. In contrast, in Yuanbao's response, the opinions were not detailed enough, and there was also a lack of data information.

In addition, the reporter also prepared economic, medical, logical reasoning, and riddle topics to compare the accuracy and speed with which Tencent ingots and GPT-4o answered questions.

The reporter observed that although the accuracy rate of Yuanbao and GPT-4o is the same, and the accuracy rate is 75%, the two have different answer styles. GPT-4o presents answers in a simple, direct, and structured manner, using mathematical formulas to clearly show the calculation process and quickly communicate results; while Tencent Yuanbao focuses on the guidance and logic of problem solving ideas and provides detailed steps and analysis, but may be slightly inferior in terms of efficiency and intuitiveness.

Finally, Wenshengtu's ability competition.

The images generated by Tencent Yuanbao and GPT-4o based on the ancient poem “Little Lotus Came Sharp, Dragonflies Rise to the Head” all include key elements in the poem, such as lotus flowers and dragonflies, and more accurately capture and convey the mood in the poem. The reporter discovered that the picture of the yuanbao shows the vivid color characteristics of modern photography, while the GPT-4o image is closer to classical painting style, emphasizes soft colors and emotional expression, and is more in line with the ancient charm contained in ancient poetry.

▍ AI application testing: upgrading the fun and practicality of Tencent Ingot in everyday situations

In addition to meeting the need for efficiency, Tencent Yuanbao's “Discover” section has launched special applications in various daily life scenarios, such as versatile AI avatars, speaking sparring, super translators, and AI agents, all of which are free to use.

Big models such as Doubao, Wenxin Yiyan, and Kimi are currently in the first camp in the country. Will the comeback of the mixed-element big model launched in September 2023 be a surprise? I'm afraid it still depends on strength.

First, oral sparring service test.

The reporter discovered that Tencent Yuanbao scored users' grammar and pronunciation by simulating real 1V1 conversation scenarios. More like an exclusive private tutor, users can get personalized speaking guidance and improvement suggestions by clicking “how to optimize”. It is more suitable for learning users seeking detailed grammar and expression improvements, such as changing “what's” to “who's” to optimize sentence grammatical structure and adding “and why?” Make the conversation more detailed.

Yuanbao oral sparring conversation results

In contrast, Doubao uses virtual cartoon foreign teachers to practice conversation. The interface is simple and fun, and highly interactive. It can provide detailed information and background knowledge, making the conversation content natural and close to actual life. The downside is that it does not clearly indicate users' oral language improvement opinions.

Second, the Super Translator function test.

In terms of input methods, compared to Doubao, which only supports the three input methods of file, voice, and text, Tencent Yuanbao is quite powerful. It not only supports five input methods for files, voice, text, images, and links, but can also recognize 15 mainstream languages.

The reporter tested the documents of an English paper and found that Yuanbao's super translator function not only efficiently summarizes key points in the final text, but also provides full-text translation services, which is more suitable for high-demand translation tasks such as academic research and professional documents. In addition, Yuanbao specially designed an immersive reading mode, which further guarantees the user's reading experience and makes the translated content more clear and easy to read.

Doubao's translation results were poor compared to Yuanbao. The answers were redundant and the subject was not refined enough. At the same time, the translation speed was not as good as expected, and there were even obvious delays during the test, which affected the consistency of the user experience.

Third, the versatile AI avatar function test.

The Tencent Ingot feature provides 12 unique styles, including Barbie, dopamine, retro flowers, and white-collar elites. Users can choose different styles to try out according to their personal preferences.

The “Science and Technology Innovation Board Daily” reporter noticed that compared to the vertical AI camera track app, Tencent Yuanbao has restrictions on users uploading selfies and only allows them to upload one image, while Miaoyu Camera allows users to upload selfies with multiple lights, multiple backgrounds, multiple perspectives, and multiple expressions. Furthermore, Ingot's AI avatar feature does not include gameplay such as clay filters and TuSen videos, which have recently been popular among users.

The reporter's test found that although the AI avatar generated by the ingot was different in style, it did not meet the expected level of detail and looked relatively rough. Despite providing different style options, these avatars are slightly bland in terms of personalization, lack unique identifiable elements, tend to have uniform facial expressions, and lack vivid changes. Furthermore, the background design appears simple and highly repetitive, and lacks rich and diverse detailed processing.

ByteDoubao's avatar creation function is located in the drawing section of the discovery page. It also uses the concept of multi-style generation, using a Wensheng map instead of uploading a photo. After entering the keyword “Wong Kar-Wai-style avatar”, the reporter generated four works. Although these works try to capture a unique literary and artistic atmosphere, similar to the problems that have arisen in Tencent Yuanbao, the character's avatar needs to be enhanced in terms of vividness of detail, diversity of expressions, and complexity of the background. Also, there were deviations in the generated results that did not match keywords such as “hair length” and “location.”

Finally, AI agent testing.

Tencent Yuanbao has launched the function of an AI agent, giving character settings. The agent can either let the AI play a specific role to chat with you, or it can be an expert who is good at completing specific tasks. Users only need to click “Create Agent”, then follow the prompts to enter the name, character settings, introduction, opening line, and preset instructions, select the sound, and upload the logo. Or let AI automatically generate intelligence related information and replicate your own sounds.

Doubao's intelligent function is similar to Wen Xinyan, and also allows the creation of exclusive sounds, while at the same time being more diverse in terms of voice selection than ingots, including automatic recommendations, female voices, male voices, characters, and accents.

Overall, judging from efficiency scenario tools, Tencent Yuanbao is good at quickly grabbing information and efficiently parsing links, and has obvious advantages in terms of processing speed and multi-format input support. More importantly, by deeply integrating the massive data resources of public accounts, it can provide more time-efficient and comprehensive search results than GPT-4O, making it an AI assistant product with powerful search functions and easy to use.

Judging from everyday scene tools, Tencent Yuanbao's ability in speaking sparring and document translation is slightly superior to ByteDoubao; what they all have in common is that the intelligent functions of the two are very similar, and both have obvious room for improvement in the fineness and innovation of AI images.

The big model application market is still developing rapidly. As more players of AI products at home and abroad “participate”, the consumer market will usher in more intelligent and efficient products and services, and the competition for big model apps may enter a more intense new stage in the future.

According to the Changjiang Securities Research Report, it is recommended to continue to pay attention to the commercialization of AI in various fields such as advertising, e-commerce, film and television, games, and education.