The massive model is a "pit", but the vertical model is not.
@商科星球original
Author丨Yuan Jing
Edit丨Big Rabbit
In April 2023, the news about GPT-5 caused a sensation. At that time, people had both expectations and fears for huge large models. What's more, it is believed that human history may end because of this.
But soon, practitioners discovered that there are "uneconomical" and "inapplicable" defects in a large number of large models. Immediately, vertical models began to emerge. As a result, in Q2 and Q3 of 2023, vertical models are appearing in large quantities.
In addition to the released vertical models, according to preliminary calculations by Data Science Planet (ID: digital-planet), more products will come out in the future.
The truth is getting clearer and clearer. After talking to several industry veterans, we discovered the future of the vertical model. It is not limited to providing a "cognitive dialog box" that transcends individual intelligence, but can also combine specific industries to empower thousands of industries.
Now, people are jumping out of the "pit" that OpenAI has "dig" for the industry, and are beginning to face up to the true value brought by this technology.
01
The more generic, the more awkward
In the past, people's perception of the universal model is that it is a good teacher and friend, and it knows everything. From a product point of view, it is considered a cognitive product, and the user experience is reflected in the answers given in the dialogue.
In the past, scientists focused on "how to present dialogue answers smarter". But in fact, most user needs other than cognitive functions are ignored.
An entrepreneur who focuses on marketing and acquiring customers said to Digital Planet (ID: digital-planet): "We need more multi-modal data to make user portraits more accurate, but with the increase in the number of parameters of the general large model Big, this kind of feedback results are getting more and more inaccurate.”
He's not the only practitioner complaining. More people believe that in large models such as LLaMA7B-65B, the increase in the number of parameters is of little help to some industries. A practitioner in the overseas field even thinks that "3B is enough."
For vertical industries, the increased parameters only broaden the knowledge of the large model, but have not been implemented in specific industries. In other words, the "intelligence" added by the large model is not needed in thousands of industries, and objectively causes the mismatch between supply and demand.
A practitioner said, "It (referring to a foreign large model) used to be at the level of a junior high school graduate, and now it may be a post-doctorate, but in our industry, it is enough to graduate from a high school or a college student." Therefore, under a given calculation budget , the best performance in the vertical domain is not achieved by the largest model, but can be achieved by a smaller model trained on more data.
The above is the fundamental reason why the vertical model can be so large today.
02
"I want it now"
These days, the popularity of small models and intelligent agents (AI Agents) seems to confirm the above facts. That is to say, for industry practitioners, the value of developing useful industry tools is much more than that of a seemingly "omnipotent" Copilot.
A small model practitioner who went overseas told Digital Planet (ID: digital-planet) that after their project has successfully obtained seed financing, it is expected to easily achieve 20 million in revenue this year. As far as the essence of the project is concerned, it uses open source models and public data combined with multi-modality to build overseas marketing tools. The project can bypass "intermediary" products such as Temu, TikTok, Amazon or Shopify, and reach users in Southeast Asia through EDM, instant messaging and phone/SMS.
Although, telemarketing looks earthy and annoying. But in the ToB industry, after all kinds of customer acquisition methods are fleeting, it proves a conclusion that has been proven repeatedly: the most brutal approach is often the most effective.
Back to the topic. In recent days, we have seen the reality that practitioners from all walks of life are starting to speculate on the general model. Now, more than ever, people hope to be simple, direct, and even directly generate revenue.
A practitioner in the clothing industry gave a more straightforward answer: he is neither optimistic about the intelligent assistant nor the large domestic model. The explanation given is: "The story of Copilot has been told to the end by Microsoft. The gap between domestic large models is not big. Now that there are more and more open source data sets, the advantages of domestic related products are even smaller."
At the time when the big model is being beaten to death, it seems that the market space for intelligent assistants can only be hoped by Xinchuang.
The realistic market conditions described above lead to several outcomes:
First, for domestic non-general-purpose large-scale model manufacturers, choosing to be vertical means competing with countless domestic "workshop-style" AIGC teams. How Yangchun Baixue can defeat the ants, the former's business model still needs to be considered;
Second, due to the unfriendly GPU sales environment, more and more teams are renting computing resources and launching products by training small models. Compared with the interfaces provided by domestic manufacturers, many enterprises are more willing to use GitHub open source software;
In this way, the small model that was once regarded as a "compromise" by the industry has been "resurrected" under the cost-effective thinking, and it is showing an intensifying trend. But in turn, this seems to be a slap in the face for AGI.
03
Is it a model or an income
Model or revenue, that is the question.
Some people say that AIGC seems to be hot, but it is rarely invested. It is undoubtedly a false fire in the industry and a beautiful bubble; Few people make a fortune in silence.
Objectively speaking, the above situations exist at the same time. The reason: On the one hand, for investors, "I don't invest in prestigious schools, big factories, and majors", this is a correct belief, which is difficult to change in the short term; It is rare, coupled with the habit of using GitHub, it is inevitable to follow the trend blindly.
Among the many projects that Digital Planet (ID: digital-planet) has been in contact with, some CTOs in BPs even have front-end backgrounds and senior database engineers in large factories, but we cannot question the hematopoietic ability of such teams.
"When you start from a position, you have already got the answer." An investor said to us. Silicon Valley's model burns a lot of money, and local cases seem to be able to make money day by day. We see a flourishing and distorted world. "Since the first-tier funds think we are a third-tier project, let the third-tier funds invest." The founder of a certain project said with emotion.
Is it the model, or the income? Some people choose the former, while others are more realistic.
At the moment, the "omnipotent" AGI seems to be drifting away. If the vertical model proves itself, isn't the previous industry "digging holes" for itself?
Vertical and general, who can laugh at the end?