Secretly opened a company with a few agents, and let the AI ​​do all the work of his colleagues.

From last year to this year, the large model industry has launched a comprehensive competition, and many large models have reached the "strongest" level in terms of parameters and dimensions.

However, for most ordinary users, it is actually quite difficult to intuitively feel the substantial impact of AI on work and life from the rankings and parameters.

Those cool magics at the press conference, such as generating PPT in one sentence, multi-modal AI that can recognize multiple media in real time, and voice interaction like in science fiction movies, can they become part of our productivity? That is why AI should be shown to ordinary people. Human background.

At today's Zhipu AI Open Day, CEO Zhang Peng also pointed out that imagination for the future can only have truly revolutionary significance if it is turned into realistic productivity.

▲Zhang Peng, CEO of Zhipu AI

At Zhipu AI Open Day, we saw a completely different style from other large AI model release events.

With the help of "out-of-the-box" AI agents, children as young as 1 years old have held art exhibitions, journalists have transformed into psychological counselors, and started second careers.

There is also a child as young as 9 years old. By designing his own story framework and constantly asking questions, he created the first domestic children's and AI co-authored book in just one month – "AI Boy – Mars Survival Challenge" .

In addition, we also saw the first digital employee of Zhipu Qingyan – AI Luo Yonghao with job number 001. Not only can he bring goods live and bring the "excavator"-shaped house to new heights, he is also open to the public at the Qingyan Intelligent Center at any time, which means that we can talk to him at any time.

Speaking of digital "employees", I had a bold idea. Can I also use WorkFlow – "Qingliu" generated by the collaboration of Zhipu AI agents to form an AI content company?

Use AI agents to form a content creative company

Just do it. Since I want to "start a business", I will start from my old line of work.

1. Content Creation

Recently, Nvidia CEO Jensen Huang introduced the company's new products and plans in a speech. We first used AI technology to search for relevant content. It must be said that the summary obtained by AI search is very comprehensive.

However, as the "founder and CEO" of this artificial intelligence company, I obviously will not be satisfied. So I called in a few AI agents designed to write popular headlines and created some catchy headlines.

We can then fine-tune these titles to suit our specific needs.

Guided by the title, the overall quality of the article is quite good.

That's not all. In order to ensure the accuracy and logic of the article content, we also hired a "proofreading specialist" to carefully review the article.

They will help us check multiple aspects of the article, including typos, accuracy of information sources, clarity of presentation, and logical rigor, and can also make targeted revision suggestions.

After the quality of the article is guaranteed, the agent can also provide us with more extended questions about the event to help us understand the relevant background and details in depth.

I have to say that this way of directly typing "@" in the dialog box to call up the agent often makes me feel as if I can find colleagues with different expertise anytime and anywhere in the company group. No need to explain the previous article, this "colleague" who was @ You can give me feedback immediately based on the "messages in the group", which is very convenient and natural.

2. Generate PPT of event planning plan

Not long ago, the inG Game Art Festival hosted by Ai Faner received a good response. I also wanted to try to see if I could use the intelligent workflow to complete the entire event planning.

The results were so good that I convened several different AI campaign planning agents and asked them to work together to provide optimization recommendations.

Not only does this reduce the likelihood of errors, but it also stimulates more creative inspiration.

After generating the PPT, I called the CogView AI drawing agent to generate the required pictures for each page in one sentence.

It is worth mentioning that I did not specify the requirements for each page, but it can still generate corresponding pictures for each chapter based on the previously generated PPT content.

Even if I only have a vague idea of ​​what I want, I can slowly adjust it to meet expectations and meet various tricky requirements.

3. Customize the agent and “find foreign aid”

Of course, we can also bring more functions by connecting external tools.

In the "Agent Center" of Zhipu, we can see a variety of intelligent agents, some of which can help you practice speaking, and some of which can bring happiness when you are bored.

Click "Create Agent" to automatically generate the configuration of the agent through AI without inputting too much content.

The system automatically generated avatars, ability descriptions, opening remarks and preset questions for us. Model capabilities have networking, painting, and coding checked by default.

Click on the "Tool Market" and we can see many tools used by other agents, which can be connected to Evernote, NetEase Youdao, WPS, Feishu Calendar and other tools.

Although I created a "Breaking News Editor", after adding the three tools under the "Amap" column, I can now have the ability to query local food.

I also published this "Breaking News Editor" with the hidden "Food Search" function to the "Agent Center". Friends who are interested may wish to try it to see if it can use this "Breaking News Editor" as a Use mapping software.

Become a super individual in the AI ​​era

As we have experienced before, the collaboration of agents may even be stronger than the next generation basic model, but the basic model also determines the upper limit of group intelligence.

In fact, the model GLM-4 behind the Zhipu Qingyan agent is very different from GPT in many aspects.

The latest GLM-4-9B model launched by Zhipu AI today can be said to be a step forward in all aspects. Its general performance is 40% higher than the previous generation, surpassing the Llama 3 8b model. It supports up to two million words of text and has function calling capabilities. Compared with the previous generation, it has improved by 40% and fully supports AllTools calling capabilities.

In the previous article, we showed you how to use an instruction to let the agent understand the intention on its own and automatically understand and plan complex instructions. In addition to the function we used to let CogView understand the context and generate graphs, there are also code models CodeGeeX, multi-mode The state-of-the-art large model GLM-4V and many other tools realize the seamless connection between intelligent agents and daily life through external APIs.

Today, there are more than 300,000 agents active on Qingyan APP for us to use. This way of using @ to connect different agents together is Qingyan flow, which is the process we experienced earlier. More and more, we are moving towards the GLM large model on the application side.

Smaller, faster, cheaper – and probably closer to AGI

Nathan Baschez, founder of Lex, an AI writing tool startup, believes that,

In the AI ​​era, a new type of startup company is emerging, characterized by: smaller, faster, cheaper and weirder (smaller, faster, cheaper, weirder).

In fact, in today's wave of AI entrepreneurship, we have already seen star companies such as Midjourney and Magnific AI with a very small number of people. When the workflow of AI agents penetrates into more industries, this situation will become more and more common. .

Industry giants such as Sam Altman and Huang Jen-Hsun also predict that AGI may arrive within a few years. The so-called human-level AI does not only mean that the cognitive level and understanding ability are the same as or even stronger than humans, but also the ability to Feedback and adjust planning like humans do.

By letting AI agents play different roles, such as creators and reviewers, they can iteratively improve based on each other's feedback. This multi-agent system with a clear division of labor can significantly improve work efficiency and quality.

At the same time, integrated tools expand the functional boundaries of the system, enhance flexibility and adaptability, and enable it to continue to evolve with technological development to meet the needs of more scenarios.

As the workflow progresses, complex tasks are broken down and assigned to specialized agents. This meticulous process planning and task division allows each agent to focus on what it is good at, reducing the chance of errors and improving overall collaboration efficiency.

The AI ​​agents in the workflow can not only undertake tasks independently, but also complement each other and cooperate when necessary. At the same time, human intervention can fine-tune the work of the agents to ensure the accuracy of the entire system operation. and efficiency.

It is also these characteristics that bring greater efficiency and overall optimization to the workflow system composed of AI agents, making it the "future look" of AI in the eyes of many people, and also making it closer and closer to the current AI era. The existence of "Jarvis".

There is a well-known simulation hypothesis in philosophy: all of our existence may be a simulated reality, such as a computer simulation. Simulations can contain conscious minds that may or may not be aware that they are living in a simulation.

▲ The famous science fiction movie "The Matrix" is a story centered around such a virtual world.

What we have seen in the workflow is that a group of AIs can deduce the collaborative process of humans, which may have verified this hypothesis to some extent. Will this be the beginning of using AI to create silicon-based civilization?

Perhaps the intelligent workflows we are experiencing now, viewed from the parallel time and space of AGI, are still in a very early stage just like the social organization of ancient Babylon for thousands of years. But despite this, it is the only way to AGI.

Zhang Peng, CEO of Zhipu AI, believes that there is still a big gap between the steepness of the technology curve and allowing more people to enjoy the powerful productivity of AI. How to use the powerful large model capabilities to truly help people in their work, study and life? In fact, there is still a lot of work to be done.

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo