Tencent's large model app Yuanbao is online, and we used it to 'challenge” GPT-4o

Tencent Yuanbao VS GPT-4o, who is better?

Recently, Tencent has changed from its past slowness and suddenly "rolled up":

On May 14th, Tencent fully open sourced the Hunyuan Wensheng graph model;

On May 17th, Tencent released a one-stop AI intelligence The three-dimensional creation and distribution platform "Tencent Yuanqi";

On May 30, the App "Tencent Yuanbao" based on the Hunyuan large model was officially launched and is currently available for download in the app store.

Tencent Yuanbao is an efficient information integration tool based on Hunyuan large model and search engine driven. It has a simple interface design. It can search for real-time information, summarize and translate uploaded multi-format documents, and can also practice speaking through voice dialogue.

Behind this upgrade of Tencent Yuanbao’s product capabilities is the continuous iteration of Tencent’s Hunyuan underlying model.

According to reports, since its debut in September 2023, the parameter scale of Tencent’s Hunyuan large model has been upgraded from hundreds of billions to one trillion, the pre-training corpus has been upgraded from one trillion to 7 trillion tokens, and it has been the first to be upgraded to multi-expert Model structure (MoE), the overall performance is improved by more than 50% compared to the Dense version.

We immediately got the qualification to experience Tencent Yuanbao, so we took it out for a "walk" today.

Tencent Yuanbao "Heads Up" GPT-4o

Compared with the Hunyuan applet version in the previous testing stage, Tencent Yuanbao provides core capabilities such as AI search, AI summary, and AI writing for work efficiency scenarios.

Without comparison, there is no right to speak. We decided to let Tencent Yuanbao compete with GPT-4o from across the ocean.

Round 1: AI Search

Nowadays, AI search is popular.

Both the "King of Search" Google and the new top OpenAI are making a fuss about AI search. Even Perplexity AI, which was founded less than 2 years ago, has become Huang Renxun's "favorite". Nvidia founder Jensen Huang has publicly stated that his favorite AI tool is Perplexity.

Tencent Yuanbao naturally launched this function.

You must know that WeChat public accounts generate a large number of high-quality and in-depth articles every day. Backed by this big tree, Tencent’s AI search function has unique advantages.

We enter "Is it suitable to buy gold now?" in the input box below "Tencent Yuanbao"

(Warm reminder: Be cautious when investing)

Tencent Yuanbao's answer:

GPT-4o's answer:

First of all, in terms of response speed, the two can be said to be comparable. In a few seconds, Tencent Yuanbao referenced 9 pieces of information and gave answers.

Secondly, in terms of answer content, Tencent Yuanbao seems to be better. GPT-4o only gives a few factors that need to be considered when buying gold, while Tencent Yuanbao's answers cover gold price trend predictions, investment risks and investment strategies, and each item is quoted, which avoids the big model's "talking about the truth". "Shortcomings.

In addition, Tencent really used good steel on the blade this time. In addition to recommending relevant public account articles, Tencent Yuanbao also launched a quick broadcast function on the homepage to summarize the latest and most important information, and each piece of information will provide a link to Tencent News.

In this round, Tencent Yuanbao wins!

Round2: Document summary

This function is designed to help users obtain and process document information efficiently. Tencent Yuanbao can process documents in multiple formats, such as PDF, Word, Excel, etc.

Tencent just released its first quarter financial report for 2024 a few days ago, and we downloaded a copy.

This 48-page, 32,000-word financial report not only contains a lot of numbers, but also uses traditional Chinese characters, which makes people’s brains hurt when they read it. This is where AI comes in handy.

We uploaded this financial report to Tencent Yuanbao and GPT-4o respectively, allowing them to analyze Tencent's revenue in the first quarter of the year.

Tencent Yuanbao:

GPT-4o:

Overall, their answers are very clear. Comparing the two, GPT-4o has richer answers. In addition to financial performance and main business performance, GPT-4o also provides operational data, financial status, etc.

Accuracy of financial statements is important. We took this pile of data and proofread it one by one, and sure enough we found the bug.

Tencent’s other income in the first quarter should be 2.06 billion yuan, but Tencent Yuanbao was written as 20.6 billion yuan. The decimal point was wrong during the conversion process.

And GPT-4o’s data is all correct.

In this round, GPT-4o wins!

Round3: Web page summary

This is a function that automatically extracts key information from web pages. When users browse a large amount of information, it can quickly grasp the key points of web page content, thereby saving reading time and improving efficiency.

Last week, this site published an article titled "Li Feifei personally wrote: Large models do not have the ability to perceive subjective feelings, even with billions of parameters." We gave the article link to Tencent Yuanbao and GPT-4o respectively, and asked them to summarize the link content.

Tencent Yuanbao:

GPT-4o:

After receiving the task, Tencent Yuanbao "immersed himself in summarizing", starting from Li Feifei's views, spatial intelligence and AI, the controversy of AI sensory ability, erroneous reasoning of AI sensory ability, AI The differences with human intelligence and future prospects are summarized in 6 aspects.

I have to say, it summarizes it quite well.

However, GPT-4o temporarily lost the link, "I cannot directly access the specific content of the link provided", and asked us to provide article description or key points, but GPT-4o was too lazy to explain clearly.

In this part, Tencent Yuanbao wins!

Round4: AI drawing

Multimodality is also a key subject of inspection.

Let’s take a look at the drawing skills of these two AIs.

We entered the same prompt word: Please help me draw a picture of a cute cartoon girl in a dress holding a white kitten, full body, yellow background, Keith Haring style doodles, clear illustration, bold lines and solid colors, simple details, minimalism, yellow background.

GPT-4o is "on strike" because due to content policy restrictions, images related to Keith Haring's style cannot be generated.

After we deleted the "Keith Haring" keyword, GPT-4o started to work:

Tencent Yuanbao is "easy to talk to" and directly publishes the picture:

Tencent Yuanbao is not discounted. Of course, the response is more pleasing to ordinary users, but this may also involve copyright issues.

76 smart phones were launched in one go, focusing on practicality and fun.

The trend of smart phones has also hit Tencent Yuanbao.

In the "Discovery" column at the top of the interface, Tencent Yuanbao has launched a total of 76 smart agents covering five categories: work, entertainment, efficiency, learning, and role. Visual inspection shows that most of them are created and published by users or developers themselves.

Among them, efficiency agents include PPT masters, work report wizards, logo design experts, promotional draft generators, recruitment masters, etc., focusing on practicality.

The lifestyle and entertainment category focuses on "fun", such as movie recommendations, Duke Zhou's Interpretation of Dreams, and the same popular game "It's Over!" "I am surrounded by beauties"...

In addition, in the face of dazzling intelligent agents, Tencent Yuanbao has also produced a first selection list, including creative paintings, ever-changing AI avatars, spoken language training, creative posts, and super translators. These 5 intelligent agents were selected.

Creative Stickers

There is a niche category on Xiaohongshu that is very popular, that is, cute pet stickers, and "Creative Stickers" is aimed at this demand.

Users only need to enter text or upload images, and then select a style.

We uploaded a picture of a scribbled puppy, and the final sticker effect is as follows:

You can also enter prompt words to generate stickers. Prompt word: Little girl eating ice cream, cute style.

Various AI avatars

This feature allows users to use AI technology to generate personalized avatars, and can also be integrated with QQ QR codes to add personalized elements to users’ QQ accounts.

This function also provides 12 styles such as Barbie, Dopamine, Retro Flowers, and White-Collar Elite. We chose the "Retro Hong Kong Comic" style and uploaded a photo of Swift.

It is worth noting that the uploaded pictures must have clear facial features and a resolution of more than 500. Avoid blurry pictures, facial occlusions, small heads, or photos of multiple people.

The generated effect is as follows:

Although the generated avatar is not on par with Swift, the style of painting is quite nice.

Due to the simple operation, we couldn’t stop playing.

This is Sophie Marceau in Barbie style:

Little Plum in retro floral style:

Fool-level operation, hand-rolling an intelligent body in minutes

Tencent Yuanbao is also online " The "Create Intelligent Agent" function completely lowers the production threshold.

Users only need to click "Create Agent" and then follow the prompts to enter the name, character settings, introduction, opening remarks, preset instructions, select the sound, and upload the logo.

For example, the "Crazy Literature in Moments" generator we created can be completed in minutes.

We asked it to post a copy of "Life is wrong, every sentence makes sense." The agent spit out 8 sentences at once, such as "Life is like playing a game. No matter how hard you try, there is always a level that you can't pass." However, we still love this game, because it is difficult to pass, and this is life. "

Hey, the logic is really self-consistent.

However, Tencent Yuanbao’s customized intelligent agent is still too “serious”. Many sentences are indeed reasonable, but not crooked or meaningful enough.

If you are too lazy to do it, you can also let AI do it for you. For example, we only enter the name "Ancient People Also Emo", click the "AI Generate" magic wand, and AI will complete the rest of the work in a few seconds. We just need to adjust the details.

The above is the detailed content of Tencent's large model app Yuanbao is online, and we used it to 'challenge” GPT-4o. For more information, please follow other related articles on the PHP Chinese website!