The day before yesterday, Zhipu held DevDay and released GLMs.
Two months earlier, OpenAI held DevDay and released GPTs.
The internet.
Congratulations to Zhipu.
Whether it is GPTs or GLMs, the most common discussion among builders is still: “Can this make money?” and “How do you make money with it?”
So can GPTs/GLMs help us, the creators, make money?
My answer may disappoint most people: yes, but most people cannot.
Why not? I will slowly explain below. This article will not talk about lofty ideas or deep technology. It starts from the most vulgar angle, money, and looks back at what I saw, heard, and felt in 2023 as an AI product manager, to talk about AI monetization and application landing.
First, let’s look at a set of data. As one of the largest third-party GPTs stores, BeBeGPTs has collected more than 100,000 GPTs records, while the official number is already 3 million-plus. Let us simply pull some data.
This data was pulled directly from the BeBeGPTs database after cleaning. Unprocessed data is not included. The data is grouped and sorted by actual GPTs conversation count. Even though we cannot collect every GPT, and even though a small amount of wrong data exists during collection, the result is still quite obvious.
Why Is Everyone Building GPTs?
Many people know me as the person behind BeBeGPTs, but I actually started building LLM applications quite early, only they were B2B products. When our project was first proposed, I brought up the concept of GPTs. In fact, many things we built were relatively early: RAG, Team, FakeJsonMode, Tools, Agent.
I have been doing product work for ten years, from B2B to government to consumer products. I have gone through many projects, some big successes, some small successes, and some that did not work. After moving into AI product work, I always felt I was very suited for it. I understood technology, product, business, and had experience. But I still ran into problems. I could not find the landing point. The ideas were great, but users were all watching from the sidelines.
The leap in AI technology came too fast for most people.
When I passionately described this thing to others, more people were confused or skeptical. No way around it. This battle is probably a long war, and right now it has only just begun.
Later, I joined all kinds of groups to meet all kinds of people, chatting with them about business, technology, and nonsense. I even worked for free as a product manager for mirror-site developers, and did free consulting for some company bosses.
I summarized that state as: GAP.
People who understand technology do not understand business. Most people familiar with business have no concept of the technology. After all, there are very few people willing to be the first to eat the crab.
Not to mention customers. Even our technical leader at the time argued with me several times because I wanted to find someone to write LangChain. “You are a product person, do not interfere with technology.” So we ended up building many wheels in Java and wasted a lot of time.
In Silicon Star’s recent interview with Zhipu CEO Zhang Peng, the same view appeared.
Why is it so hard now to find a so-called Killer App? It seems like all product managers have no clue.
Personally, I think the contradiction is that traditional internet product managers who were good at creating hit products cannot understand the newest technology. The people who truly understand the technology cannot understand so-called user needs and product needs.
Silicon Star
Do more in vertical domains. If you do not do it, you do not know what customers are thinking, and you do not know how much user value your technology ultimately helps them improve.
This is actually a process of looking for Best Practices. We hope to walk this path through, then look back, settle these things down, and empower the ecosystem.
Silicon Star
After all that setup, let us return to GPTs.
As practitioners, we are constantly watching industry dynamics and all kinds of data.
From a practitioner’s point of view, it may feel like we are surrounded by AI information every day, and everyone around us is talking about AI. But if we look at society as a whole, this group is still very, very small. Most people have not even used the real OpenAI product, let alone learned how to use it correctly.
AIGC is indeed not hard to access. But if most users face a large model directly, the most common conversations may be “hello,” “who are you,” and “tell me a joke.” Then it is over.
Why does Character.ai have such high data? From the essence, the difference is that Character.ai very clearly creates a concrete role. I am who I am, and I do what I do. Do not make users think. Most product people have probably heard that sentence.
So who should do this? Model companies? Application platforms? Neither can really do it.
It has to be done by the pioneers, the creators of GPTs.
Platforms need scenarios to feed back into products. They also need pioneers to get rich first and drive the rest. The people who get rich first also need to receive benefits, while the platform provides capabilities and tools. That is a win-win.
Eventually, a coexisting ecosystem forms.
So How Do You Make Money with GPTs/GLMs?
When OpenAI first released GPTs, people online kept saying this was the App Store era. If we must compare it that way, think carefully: the GPTs we create are our products. So where is the product’s core competitiveness and moat?
Is it the prompt?
Absolutely not.
Model capability and basic tools are provided by the platform. Prompting is not difficult. It is hard to defend, everyone will eventually catch up, and later prompts may be simplified or weakened.
Personally, I think the real core competitiveness and moat of GPTs lies in data and services, meaning customized Tools.
In the first half of the year, a closed-door Sequoia meeting mentioned a point: traditional SaaS is dead. How did it die? It died here. Traditional SaaS capabilities will be fragmented and integrated into all kinds of GPTs, and OpenAI becomes the real traffic entrance.
Let me give a real example.
There are many homogeneous products in the GPTs store now. Simple ones created with prompts can be copied in minutes. If you can write it, so can I.
There is a GPT called WebPilot. The author developed his own search interface, provided search service, and connected it to GPTs. If you want to imitate it, either you also develop a search service, or you directly buy the API interface provided by WebPilot.
There is also a Xiaohongshu writing expert. The author collected a lot of Xiaohongshu data and rules, then packaged them into GPTs through RAG. If you want to imitate it, you first need to obtain those data.
At this point, everyone should roughly have an answer in mind. You may be able to make a wave of money, dividend money, but this is certainly not a long-term way to make money.
What Will a Developer Incentive Program Look Like?
This is only a bold guess. We have done similar things ourselves, so I have thought seriously about this problem many times, but only in the context of our own GPTs product. It does not mean big companies will do the same.
And different stages will definitely have different strategies.
Poe’s plan is user acquisition. OpenAI may adopt the same strategy. But personally, I think this strategy has low conversion and a high threshold.
If payment is based on usage, both OpenAI and Poe currently use subscription models. Paying from your own pocket cannot last long either.
There is actually another strategy: split the basic service from the GPTs value-added part, give pricing power to GPTs authors, and let authors and the platform share each piece of GPTs-generated revenue.
In that case, the platform’s benefit only increases, while whether GPTs authors can make money depends on whether they can build truly valuable products or provide valuable services.
Finally, Agent Application Development
This section is about application development, not GPTs.
There are many ways to make money. In Chen Caimao’s group, I also met many friends who made money through AI.
Key merchants, writing books, selling courses, taking projects, building mirror sites, and so on.
But if we really look at it from a product perspective, the direction I choose is still Agent.
The author of WebPilot mentioned above recently developed a long-form writing Agent and provides an external interface. The effect is stunning. As someone in the same field, I studied it for a long time and still could not help praising it.
I have always been a loyal supporter of Agents. I have looked at most open-source frameworks on the market and studied the implementation of many closed-source products. Early on, I naively wanted to build a general framework and use it across different scenarios. Later, when I actually started doing it, I found that this idea was still naive at the current stage.
Years of product experience have also told me that knowing the business is the right path.
After you understand the business, combining AI capability to truly polish the product becomes the moat. This is also the direction I have always believed AI application developers should stick with.
So keep going, friends. I look forward to AGI arriving soon.
Original
This article was first published on the WeChat Official Account “白苏Elliot”: https://mp.weixin.qq.com/s/hYY0rxFmCMbIW-XyWOYCFA