作者｜Jessica
E-mail|[email protected]
Last week, many people found that Microsoft's official website was suddenly updated with a notice that "GPT Builder is about to be discontinued". It was announced that it will end support for Copilot GPT from July 10, and will delete all created GPT on the platform along with related data within four days.
This was so unexpected, because just 3 months ago, GPT Builder was officially launched by Microsoft as part of the Copilot Pro service. Subscribers who pay $20 per month can create custom GPT chatbots for their own use or share based on their specific needs.
The official explanation for the "service suspension" at the speed of light just after its launch is: "We are continuing to evaluate scalable strategies for consumer Copilot and prioritizing the core product experience, while still working to provide opportunities for developers." To that end, we will shift the focus of GPT to commercial and enterprise use cases, and stop working on GPT for consumer-grade Copilot. ”
In short, after months of experimentation, Microsoft has now concluded that GPT Builder is not a viable product on the consumer side. While the company is fully committed to AI projects such as Copilot, it will only support them when they make real business sense.
Also in the same awkward situation are the "ontology" of Copilot GPTs, the ChatGPT version of GPTs and GPT Store with the same product logic.
With GPT-4, Sora, Voice Engine, and GPT-4o triggering rounds of discussion, OpenAI's treasure chest implosion models are frequent, and GPTs have gradually become the name that is hidden behind and rarely mentioned. But people also vaguely remember its grand appearance — the Silicon Valley winter seen as epoch-making and could change the entire AI industry.
1
Debut is the peak, but it is short-lived?
At the OpenAI Developer Conference on November 6 last year, Ultraman introduced the concept of GPTs to the world for the first time. Users don't need any programming knowledge, and they can easily customize their own personalized GPT with just natural language instructions and a knowledge base upload.
GPTs have multi-modal capabilities, support external data source extensions and API integration, and can also be published to the upcoming GPT Store to participate in community sharing and monetization. Altman confidently described the search, categorization, and recommendation mechanisms of the GPT Store, similar to the App Store, and demonstrated how to build a complete version of the "Mentor" chatbot in a matter of minutes.
Overnight, the technical barriers that stood in front of ordinary people were overturned, the term "developer" was redefined, and countless AI startups were bloodbathed. People see OpenAI's ambition to emulate Apple and Google and create a complete ecosystem of developers and users. At that time, "changing the sky", "exploding", and "subverting" had not yet become common words, and the press conference was described by Jianghu as "OpenAI's iPhone moment".
Unexpectedly, a bloody palace fight disrupted the release rhythm of GPT Store. As a result, the GPT Store, which was originally scheduled to be launched at the end of November last year, was finally opened to Plus users on January 10 this year. At the same time, the problems exposed in the use of GPTs have also come one after another.
1
With the use of a big rollover, the developers are dissatisfied, and the GPTs are cold
With its affordable technical requirements, GPTs quickly attracted a large number of AI enthusiasts from all over the world. As of the end of January this year, more than 3 million GPTs have been created by users.
However, it is precisely because of the low threshold for creation that everyone finds that the experience of custom GPT that is only called up with simple prompts is not much different from that of direct dialogue with GPT-4, and it is even a "low-profile version": it cannot adapt to the knowledge base, is easily distracted and hallucinous; Instead of following 100% of the custom instructions, "the answers you get in the ChatGPT interface are 5 times better"; It is essentially equivalent to the basic RAG platform, but it can only upload up to 10 files, and fails to provide functions such as vector database calls. And easy to implement means easy to be replaced, millions of GPTs are seriously homogeneous, good and bad, but there are very few really good ones.
In contrast, it is better to use an AI assistant that is powerful, flexible, and versatile enough to meet all kinds of complex needs. According to Sametime Web data, GPTs accounted for only 1.5% of ChatGPT's web visits in February. Since the opening of GPT Store, weekly user traffic has also leveled off.
What's more serious is the security risk exposed by some independent GPTs. Many users have said that they were able to download original knowledge files from GPTs after only a few rounds of dialogue inducements, and there is a risk of data leakage.
For example, someone easily obtained the data table of "2021 job salaries in major U.S. technology companies" uploaded by the author from GPT on the salary comparison website Levels.fyi. There are also people who directly set out the source prompts of OpenAI's 16 official GPTs. Ordinary users can copy and paste prompts, and then use a few knowledge base files to reproduce them instantly.
When it comes to the developer's side, it is also full of complaints.
On the one hand, until the OpenAI spring conference on May 13, only paid users could access the GPT Store, which greatly limited the user growth potential of GPTs. And due to the lack of user data analysis functions provided by OpenAI's backend, in a world of data-driven decision-making, creators cannot get full insight and it is difficult to attract new users.
Kirill Demochkin, founder of Hintloop, said that of the more than 36,000 custom chatbots he analyzed, only 5% of GPTs had 150 to 500 active users per day, with the vast majority only attracting one or two users per day.
On the other hand, there is the issue of the visibility of new creators in the GPT Store. Although everyone can participate, the current setting of the store makes it difficult for new creators of GPTs to gain attention. Only OpenAI's official recommended rankings can be displayed under each category. The community has shouted that OpenAI can learn from TikTok and use similar big data push or discovery methods to give new creators more opportunities to show themselves.
On top of that, the biggest problem is OpenAI's vague monetization policy. As early as the developer conference, Ultraman promised that the release of GPTs would generate actual revenue, but it has been stuck in the pie drawing stage and has not announced any payment rules. People don't get a share of the money, so they naturally lack the motivation to build high-quality apps.
By the end of March, seeing that GPTs were about to "cool down", OpenAI had to urgently decide to cooperate with a small number of American developers to issue corresponding rewards based on the use of GPTs. The goal is to "create a vibrant ecosystem where developers' creativity and influence are rewarded." As for more specific details of the reward, OpenAI did not disclose.
In stark contrast, OpenAI director Adam D'Angelo built Poe, an AI chat platform. With the same creator economy model, Poe not only took action earlier and developed a clear and comprehensive "monetization plan", but also expanded the scope of application to 23 countries and regions, including Hong Kong, China.
At Poe, creators can earn money by setting chatbot paywalls and prices per message, as well as incentivizing up to $50 per subscription by attracting new users. Poe explained how earnings are calculated, how to set up payouts, and how to get more conversions. At the beginning, it was rumored that Adam D'Angelo was furious because he was served a job by Ultraman, and who thought that Poe Road would run more and more clearly, GPTs knelt down.
Although later, in order to turn the tide, OpenAI launched the cool function of Aite GPTs directly in the dialog box, allowing 3 million chatbots to "work for you", but first find a specific GPTs conversation, next time Aite, select from the drop-down menu, and then enter instructions - obviously a ChatGPT can do things, if you have to toss the operation thread more lengthy, I am afraid that there are very few people who really use it. Last month, OpenAI announced that access was open to all free users, but that doesn't seem to be saving either.
1
为啥GPT Store成不了App Store
Carrying the ambitions of Ultraman and OpenAI, from the shocking debut to the loneliness of the fading halo. So far, at least, GPTs haven't been what they thought they would be.
OpenAI's top researchers have said that LLMs could become an operating system, similar to Apple's iOS. However, both iOS and Android, their early popular apps include video games, navigation maps, social media, browser tools, and more, all of which are distinctive, easy-to-use, and understandable. But early generative AI-powered applications were often less intuitive, and most people didn't get into the habit of relying on chatbots to create charts, plan trips, and solve math problems.
Other analysts say that the success of the App Store and Play Store also benefits from the fact that they are part of the iPhone and Android phones. Running on electronic devices that consumers can't live without, it already embraces a massive user base. On the other hand, although ChatGPT has accumulated 180 million users, which is really remarkable, it is still a novelty for more ordinary people. It is no wonder that Ultraman has not given up his plan to enter AI hardware, trying to capture this ecological highland first.
However, the collapse of GPTs and GPT Store does not affect OpenAI's earning power in the slightest. The Information recently reported that OpenAI's annualized revenue has doubled to $3.4 billion in the past six months. At the end of 2023, the figure was still $1.6 billion, compared to $1 billion in the same period last summer. Anthropic's plan is to generate more than $850 million in annualized revenue by the end of the year. Canadian rival Cohere earned just $22 million in April—OpenAI still cliff-ahead of its peers.
Of OpenAI's $3.4 billion, $3.2 billion comes from ChatGPT's subscription service and fees for software developers to access its models through APIs. Since these models are deployed in the Microsoft cloud, Microsoft takes a cut of them. OpenAI will also receive some of the proceeds from the OpenAI models that Microsoft sells to its Azure cloud customers. On an annualized basis, that earnings now represent about $200 million, or just 20 percent of the revenue Microsoft generates from the business, Altman told employees.
Now, Microsoft has cut the consumer-facing Copilot GPTs to focus on enterprise-level, which coincides with OpenAI's own business route. Since the beginning of this year, OpenAI's chief operating officer, Brad Lightcap, has publicly stated many times that the company is focusing on exploring more to B models, including a full range of customized services for enterprise-level models.
For GPTs, it is more like an experimental product in the AI era. It has eliminated the original GPT plug-in, and has also created a sensation of beautiful imagination around the world. Now, even though it is still glorious, maybe a hundred GPTs will fall, and a thousand will stand up again.
Resources:
https://wallstreetcn.com/articles/3710745
https://www.theinformation.com/articles/openais-annualized-revenue-doubles-to-3-4-billion-since-late-2023?rc=iczwj4