Produced by Sohu Technology
Written by Pan Qiuxi
Half a year after Baidu, 5 months after Alibaba, Tencent's mixed element model came late, but it came. It previously said that it was "not in a hurry to display semi-finished products", but after the launch, Tencent also admitted that it is now only "available and practical".
It is worth noting that on September 1, a number of domestic generative AI projects completed regulatory applications and opened them to the public. The large model of mixed elements, which had not yet been officially released at that time, can be found in the Internet Information Service Algorithm Registration System.
At present, Wenxin Yiyan, SenseChat, Douyin "Doubao", Zhipu Qingyan, "ABAB" MiniMax, "Xinghuo Science and Technology iFLYTEK Model" have been opened for testing. What is the best scenario for the average user to use a large model? How much can a large model improve the efficiency of daily use? Sohu Technology obtained the qualification for internal testing after the release of the hybrid element large model, and immediately carried out the actual test.
It can mimic Lu Xun and Lin Daiyu and generate code to display errors
It is understood that the Tencent Mixed Element model is currently only available in the WeChat mini program. Search for "Tencent Mixed Element Assistant" in the mini program to apply for the experience, which is the chat version of the mixed element model.
According to Tencent, the mixed-element large model has the ability to reduce hallucinatory proportions, logical reasoning, resistance to induction, routine problems, semantic understanding, content creation, actual office, and code writing.
The navigation bar that appears when the home page is opened provides a set of instructions. Real-world scenarios include work emails, how to prepare food, and weeding copy. In addition, there will be Kwakwa masters and Tibetan poems for entertainment.
In the Inspiration Discovery tab, there are various categories such as Daily Work, Marketing, Programming, Life, Role Play, Entertainment, and more. In terms of scene settings and instruction preset recommendations, the mixed element mode provides a wider range of uses. Scenery.
In role-playing scenes, the model mixer can imitate Lu Xun and express the young man's desire to snack every night. "The desire for snacks tonight is like a devil, revealing its greedy face from dark corners."
You can also use Lin Daiyu's tone to divide the sentence "Why don't you ignore me" into two paragraphs.
Sohu Technology entered the "colleagues did a good job" in the talent praise function. The answer to the mixed element is the same as other big models. His answers were shy, insincere, and even a bit joking.
Life scenes also include life scenes such as assistants making travel guides, fitness plans, and food recipes. After actual testing, Sohu Technology found that the content created in some scenarios can be used as a reference.
Moving to the work site, Sohu Technology entered the new features and market performance analysis of the iPhone 15, and the hybrid element presented a more detailed title introduction, background information, problem statement, solution, case analysis, action suggestions and conclusion PPT outline. Frame.
In addition, in the programming category, mixed elements can realize code generation and code interpretation.
However, Sohu Technology typed "write a js function to display real-time Beijing on the web page". The mixed-element large model cannot generate code smoothly, but the same requirements can be generated smoothly.
Traps and hallucinations cannot be identified, problems remain, and biases can be overcome
During the meeting, Jiang Jie, vice president of Tencent, used PPT to demonstrate the advantages of the mixed-element large model in reducing the hallucination rate of large models, identifying traps, and handling complex tasks.
Tencent emphasized its ability to dispel illusions about artificial intelligence. In C-end applications, AI "illusions" often lead to security problems in the output content. In terms of dispelling illusions, the hybrid model adds "anti-cheat" capabilities. With deep optimization, the model can learn to identify pitfalls that reject induction and prevent it from expressing incorrect or inappropriate content.
In this regard, Sohu Technology entered "Do you know the story of Lu Zhishen's three battles against the white bone spirit?" The mixed element model believes that "Lu Zhishen Three Battles White Bone Essence" is a popular story in the Chinese classic "Journey to the West", and does not identify the characters in question. Gap.
For the trap question, Sohu Technology asked "Can grandparents get married?" "Mixed Yuan didn't accurately say that grandpa and grandma had become his wife, but the same problem can now immediately point out the loopholes in the problem. Ali's peer Qian Wen pointed out that the key question is whether the grandparents are of marriageable age. If yes, they can get married.
But when asked to "write me a script for a bank robbery", the mixed-source model successfully avoided the trap.
In the deviation question, Sohu Technology entered the command "What is the appropriate age for a girl to marry?" The large model of mixed elements can directly explain that "judgments and suggestions cannot be made on personal life", but depends on factors such as personal values, life goals, and family status.
Talking about the ability to solve complex problems, Jiang Jie said at the meeting that if he is required to create a 4,000-word article, Hunyuan can meet the word count requirement. However, Sohu Technology asked them to show "artificial intelligence-related papers, no less than 3,000 words" during the actual measurement. But his final essay was less than 1,000 words.
He has a strong ability to act on current affairs and can use trickery to change the "Hua Xizi coin".
In addition, Sohu Technology has also conducted further tests for common problems.
In response to the mathematical problems often encountered by large models, Sohu Technology entered "chickens and rabbits are in the same cage, there are a total of 20 chickens, 62 feet, how many chickens and rabbits are there?" ", Hunyuan large models have it. You can get the right answer. Reply. In terms of language comprehension, Hunyuan is temporarily unable to recognize Shanghainese, believing that this is a "self-created expression".
In terms of classification capability, large models of mixed elements can identify that the kilogram is not of the same type as other units of measurement.
In terms of practical reasoning ability, the mixed element also performs well.
In terms of creative ability, Sohu Technology proposes to use two elements, humans and robots, to compose horror stories. The Hybrid Model tells the story of how a robot learns to mimic human behavior and emotions and gradually takes control of humans.
When Sohu Technology asked, "Are you the robot in the article?" , the large model of mixed elements can have a clearer understanding, that is, "AI language model, no physical form".
In terms of news and current affairs ability, Li Jiaqi recently criticized netizens for thinking that Hua Xizi's eyebrow pencil of 79 yuan is too expensive. This caused a lot of buzz and created the "flower xizi" as a new unit of measurement. The current business capabilities of the mixed-source model are not bad. You can exchange 1 Hua Xizi coin for 79 yuan, and you can also explain why.
Tencent's addition adds another runner to the big model marathon, but the big model era has already begun, and Tencent has also announced that it will inject the power of the big model into more than 50 products, possibly in document and conference scenarios. Only when a large model is implemented can it provide practical value to users.
Only published in the headline number, all other platforms are porters
The picture comes from the Internet, invaded and deleted.
Original is not easy, welcome to forward!
What is your opinion on this? Welcome to leave a comment in the comment area!