Pay attention, this man has connected more than 1,000 large models, allowing you to plug and switch seamlessly.

Recently, a visual AI workflow has been launched:

gives you an intuitive drag-and-drop interface. Drag, pull, and drag to arrange your own on an infinite canvas. Workflow.

As the saying goes, soldiers are expensive but quick, Qubits heard that within 48 hours of this AI Workflow going online, users had already configured personal workflows with more than 100 nodes.

# Without giving away, what I want to talk about today is Dify, the LLMOps company, and its CEO Zhang Luyu.

Zhang Luyu is also the founder of Dify.

Before joining the business, I had 11 years of experience in the Internet industry. I am engaged in product design, understand project management, and have some unique insights into SaaS.

Later, he was also responsible for product and operation management of Tencent Cloud CODING DevOps team, serving platform products for over one million developer users.

After starting his business, Dify quickly worked with him and his team to become the absolute leader in the field of "open source LLM application development platform", with 22k stars on GitHub so far.

This time, with the help of the new AI Workflow function, we specifically found Zhang Luyu and chatted with the man who helped 200,000 AI applications connect to large models.

He said:

The future of programming may be flow programming.

Drag and drop to complete workflow settings

Not long ago, AI guru Andrew Ng published a point of view that sparked widespread discussion. He asserted that the AI ??agent workflow will drive huge advances in artificial intelligence this year, perhaps even surpassing the next generation of basic models.

Andrew Ng regards workflow as an important trend and calls on all people working in AI to pay attention to it.

So Dify launched AI workflow while the iron was hot. (But in fact, Dify saw this opportunity half a year ago, we will talk about this later.)

After it was officially launched, it was Aunt Jiang’s:

The core is an intuitive drag and drop interface.

Users can build their own Workflow on an infinite canvas by connecting different nodes.

Yes, it is an "infinite canvas" - the current version allows the continuous expansion of the types of available nodes, and each node is configurable. Users can define input and output for each node to ensure the working logic of the workflow and data flows as expected.

In other words, you can test your Workflows end-to-end, or you can test each node individually to quickly locate the problem.

It is worth noting that some core nodes are supported first, including:

• LLM: Choose any mainstream large-scale language model and define its Input and Output
• Tools: Use built-in and custom tools to extend what your workflow can do.
• Intent Classifier: Let LLM automatically classify user input and perform work flow according to different categories.
• Knowledge Retrieval: Mount contextual data from existing knowledge bases to your LLMs.
• Code: Execute custom Python or Node.js code.
• If/Else Block: Define conditional logic to create branched Workflows.

In addition, AI workflow supports the import and export of DSL, which Zhang Luyu called "one of the coolest features."

Simply put, users can export their own workflows and then import them into other workspaces, allowing the Workflows to come and go freely, while also customizing them again as needed.

This feature opens a window for collaboration, sharing, and building on the work of others within the community.

And the workflow can utilize all the capabilities of the Dify platform. Users have set up a knowledge base of search solutions, first-party rich tools provided by the platform, and customization Tools, etc., can be orchestrated as one of the node capabilities.

It can be said that it is seamlessly adapted to the Dify ecosystem.

The threshold will not be lowered to an infinitely low level

After the launch of AI Workflow, Dify has two application modes.

One is Workflow, and the other is traditional Chatflow.

"The vast majority of users will use the traditional Chatbot type, which has no complex logic behind it and basically relies on the self-operation of the large model it is connected to." Zhang Luyu explained that, in contrast, Workflow will be much more powerful, but this does not mean that you cannot use Workflow to build a chatbot. It’s now clear that although “workflow” is translated as workflow, it actually represents the operating mechanism behind Dify.

The reason why Dify evolved in such a direction is that the team has its own "belief."

We believe that programming in the future may be

(work)

flow.

It is understood that Dify’s next plan is to launch RAG’s Pipeline and prompt’s collaborative paradigm.

So, which groups does Dify, which is constantly patching itself, regard as its users and potential users?

"Suppose there are 10 million developers in China, and 500,000 of them now have the ability to create and influence others. Large models have a natural attraction for them; the remaining people need time to understand And accept new things. "

People who are valuable, have retention, can be fully productized, and have rich skills in their own industry or skills - Zhang Luyu "defines" the main user group of Dify service. , and called him "a true professional developer that we think of, or this extremely creative innovator."

This group accounts for no more than 10% of the entire population.

Therefore, Zhang Luyu also confessed that what DIfy has been doing is to lower the threshold for using tools, but the threshold will not be so low as to be non-existent.

Dify will not lower the usage threshold to zero.

We also talked about the choice of when the function will be launched.

Dify launched AI Workflow not too early. Since the second half of last year, taking advantage of the popularity of the Agent concept, many startups have launched similar functions.

But Dify said that this time was a proactive choice. Before GPTs was launched, the team had some insights into the opportunities of Workflow.

We want to observe the dynamics of the model side and the product side, but also deliver complete capabilities and not fight unprepared battles. "In fact, we

deliberately slow down the pace "Comed from a big company, "but I'm a rebel"

Dify has its own rhythm, which probably has something to do with the team's assessment of its own technology.

When asked if he was worried about the fate of large manufacturers squeezing market space, the answer he got was three positive words: not afraid.

Their reasons are as follows:

First of all, the market competition does not seem so urgent and fierce. When Dify was established in May last year, the core founding team had been worried about entering the market too late and fearing that some team would get the upper hand. “But the fact is that no other team has done it.”

Secondly, big manufacturers often choose to bet because they see such an opportunity in the market, driven by supply; Dify chooses to bet because it sees real Demand drives R&D.

Not only is it not afraid of big companies, Dify is also not afraid of the power of similar startups. Zhang Luyu said that although there are many entrepreneurs in the current wave, most of them are still the wave of entrepreneurs ten years ago. The labels on this group of people are that they have practiced in Dachang. "While being labeled with experience in a large factory, the complexity of the writing process between departments in a large factory has also become an inherent thinking template for these people."

Those who have also stayed in a large factory Zhang Luyu calls himself a "big factory rebel" and flaunts his deviance -

I don't belong to the

(constrained by thinking)

category. My first job was as an engineer at a large gaming company when I was 21 years old. At that time, as long as I judged that what I was doing was valuable and the product was good, I would ignore all processes and get the product online, which is called stowaway launch.

This is illegal in many companies, but I must do it. why? If I follow the process, it may not work; but when I put it online, the value brought by the user experience is significant.

This is the same as starting a business, you need to accept risks. Entrepreneurship is risky, and you need to be cautious about the outcome. Fortunately, there are some things that make Zhang Luyu happy.

He said that the most exciting thing about the past year is that so many young people are pouring into this industry, and countless fresh graduates from top schools are willing to join it.

There is no shortage of good opportunities for this group of people. After graduation, they can easily obtain high-level positions in Fortune 500 companies or large factories, but now they are willing to spend very little money to do things related to large models.

"In the past, it was difficult for us to recruit such people. We didn't even know where to find such resumes, let alone persuade them to join."

"Are there many such people in our company?"

"Already at least 1/3? But I think this proportion is not high enough. These people should become our main force."