news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Cressey from Aofei Temple
Quantum Bit | Public Account QbitAI

Finally, the domestic large model can compete with GPT-4o in comprehensive capabilities.

In the test of 12 indicators, this modelIt surpasses GPT-4o in 11 aspects, and its overall ability is 10 percentage points higher。

Moreover, its advantages are more obvious in Chinese-specific fields such as poetry appreciation."The Hardest Chinese Benchmark" also has a major breakthrough。

However, this is not the result of a single large model manufacturer, but rather the result ofA "joint team" consisting of 16 manufacturers。

The team was initiated by 360. In addition to its own participation, other large companies including BAT also participated.Baidu, ByteDance, Tencent, Alibaba, Huawei, domestic large model "Five Little Tigers",Zhipu AI, Dark Side of the Moon, MiniMax, Baichuan Intelligence, Zero One Everything, and there are five head vertical manufacturers,SenseTime, iFlytek, TAL, Magic Cube, Mianbi Intelligence, a total of 16 manufacturers have arrived. (Note: The above ranking is in no particular order)

Now, this "joint team" has been invited into the product - inAI AssistantAmong them, all users can use it for free.

More than ten large models for you to work on

Among AI assistants, large models from 16 manufacturers have come together and can communicate directly on the same platform.Choose whichever you want。

And in the middleYou can switch models at any time, the system will remember the previous context and allow the subsequent model to continue to complete the conversation.

For example, in the following dialog window, we asked which one was bigger, Zhipu 9.11 or 9.8, and then switched the model to Xinghuo and asked directly how to compare.

From the conversation, we can see that Xinghuo, who came on later, accurately understood that the four words "how to compare" were asking about the comparison of the size of decimals.

Of course, for the same problem, you can also directlySummon another model to PK on the spot。

While the models are competing, we can also see more information or answers, which not only makes the answers more detailed but also allows for cross-verification.

For example, we raised a question about the relationship between two characters in the TV series "Yongzheng Dynasty", and the question was first asked to Zhipu.

Then we asked Wen Xin Yi Yan to answer the question to see if we could get more information. It turned out that Zhipu’s answer was correct, and Wen Xin Yi Yan gave a more detailed supplement.

More importantly, the 360 ​​AI Assistant is also extremely friendly to those who suffer from decision-making difficulties or users who are not sure which model is more suitable.

As long as you select the "AI Assistant" entity as the dialogue model, the system will judge your intention based on the content of the dialogue, and thenAutomatically match the most suitable model。

For example, when completing writing tasks, the AI ​​assistant will assign us a person who is good at copywriting.

When you encounter a programming problem, you can summon DeepSeek, which has strong coding capabilities.

Questions that mainly require logical reasoning may be handled by Zhipu.

Of course, the task classification displayed in the interface is relatively general. During actual operation, the AI ​​assistant also divides the tasks into more fine-grained categories.

In addition, when selecting a model, the AI ​​assistant will first conduct an online search.

Therefore, another benefit of using an AI assistant is that you can get the latest information without having to worry about when the model’s knowledge base is updated.

For some common tasks, the AI ​​assistant platform also preparesDedicated assistant, can better realize these functions.

In addition to being used on web pages, AI assistantsThere are two other entrances - desktop and 360 security browser。

For example, in 360 Security Browser, after installing the AI ​​Assistant plug-in, you will see a floating ball in the lower right corner when browsing the web.

With just one click, you can summon the AI ​​assistant in the right sidebar, and you can also quickly communicate with the AI ​​in the same window while writing.

In addition, when the mouse passes over the floating ball, a new button will appear above it. Clicking it will allow you to summarize the currently browsed page with one click.

You can also ask questions to the AI ​​assistant about the details on the page.

It also supports summarizing English content.

In addition, for individual words and sentences on the page, an AI assistant toolbar will appear after selecting them, which can translate and explain the selected part, or search the Internet for more relevant information.

The desktop version is implemented with 360 Safe Guard, which has similar functions to the browser, but expands the scope of word-scanning summoning from web pages to the entire system.

So what kind of technology is used behind the AI ​​assistant?

Unique "Expert Collaboration" Architecture

In fact, this method of scheduling large models according to demand is also a new technology launched by 360.CoE (Collaboration-of-Experts)。

We know that many domestic models are on par with or even surpass OpenAI in terms of individual indicators, but in terms of overall strength, the gap becomes apparent.

The idea of ​​360 isChange this "single-handed" model, build a large model "expert cluster" and form a hybrid large model, so that each can take advantage of its strengths and fight against GPT-4o as a "joint team".

As a result, the hybrid large model composed of 16 domestic large models based on the 360 ​​CoE architecture achieved an overall score of 80.49 in the test of 12 indicators, surpassing the 69.22 points of GPT-4o.

And except for the code, the other 11 indicators are better than GPT-4o.

Especially on questions with more Chinese characteristics, such as "retarded bar" and poetry appreciation, CoE's leading advantage is more obvious.

Compared with the MoE (Mixture-of-Experts) architecture, 360's CoE model isSpeed, intelligence and costIt has significant advantages at all three levels.

CoE optimizes reasoning resource allocation, improves efficiency, and reduces costs through intent recognition and task scheduling.Reasoning costs reduced by 90%。

In order to schedule the models in the CoE architecture in the most efficient way, at least two aspects of work are indispensable.

First,A comprehensive assessment of the capabilities of these models,Only in this way can we understand the areas in which each model excels and know what tasks should be assigned to the model.

To this end, 360 conducted a comprehensive test of the performance of the models in the expert database in 12 fields to understand the unique capabilities of the models.

△Except GPT-4o, the names of other models have been hidden

Another aspect isInterpretation of user intent- Understanding of task requirements is obviously an indispensable foundation for the allocation model.

Based on the technology and data accumulated in the past 10 years of search engine development, 360 has trained a dedicated model that can recognize more than 100 million intent classifications.

Making AI more inclusive

In addition to technical expertise, we can't help but ask, how did 360 organize the team and bring 15 major model manufacturers "into the hub"?

360 founder and chairman Zhou Hongyi said that the starting point for manufacturers to cooperate isLarge models require huge investments. Only when people use them can the costs be covered and the products can be continuously improved.。

360 has a large number of users and can open two star scenes: desktop and browser, which can bring a huge user group to these models.

Compared with plug-ins, these two entrances bring users very close to the capabilities of big models. However, the biggest demand for big models today is to be close to scenes and users so that users can use them.

In addition, the integrated AI assistant avoids the capability shortcomings of a single large model and can surpass GPT-4o by taking advantage of its strengths. This gives manufacturers an opportunity to allow users to use it, and more scenarios will be released in the future.

It can be said that this model of 360 AI Assistant is a good solution to improve the model level before the arrival of AGI, and it is also beneficial to increasing the penetration rate of AI.

Recently, large domestic model manufacturers have begun to reach a consensus, become more open, and provide cheaper Token APIs.

Therefore, 360's opening of the two entrances, desktop and browser, is not only to "gather talents from all over the world", but also to comply with the general trend of openness.

At a higher level, 360 alsoWe hope to make AI more accessible to more people。

Zhou Hongyi believes thatAI will not eliminate people, but it will eliminate those who do not know how to use AI.; and for those who know how to use it, AI will be a powerful tool.

But at the same time, AI itself should not be condescending, but should give anyone who is willing to learn it the opportunity to master it.

This is what 360 means by AI accessibility - allowing everyone to enjoy the capabilities brought by AI and avoid falling behind the AI ​​era.

AI Assistant Portal:
https://bot.360.com