My company hasn't been killed by OpenAI yet

Produced by | Tiger Sniff Technology Group

Author | Wang Xin

Edit | Wang Yipeng

Header | Visual China

Every time OpenAI releases a feature, it wipes out a startup.

For example, after ChatGPT has built-in PDF processing, startups that provide PDF interactive functions have been hit hard: Alex Reibman, a data scientist who launched ChatOCR, found that 72.4% of people will reduce their use of the PDF plug-in after ChatGPT's dimensionality reduction crackdown.

On April 25, Sam Altman, in his speech at Stanford University, once again sounded the alarm for startups that are about to face the impact of GPT5:

"GPT4 is a 'bad student', and GPT5 will improve significantly. Many entrepreneurship and research projects focus on improving the shortcomings of existing AI, which is actually based on the assumption that AI technology is stagnant. However, this will lose value as more advanced models such as GPT-5, GPT-6, etc. emerge. ”

AI entrepreneurs, how to survive the wild wave of OpenAI?

The track of AI interview may bring some enlightenment:

At the beginning of 2014, after ten years of AI interviews, there were more than 30 companies that existed, and now there are only about 6 companies left in China.

Survivors of the AI 1.0 and AI 2.0 eras have discovered:

First, the truly valuable products have been verified before the advent of the era of large models, and this value does not depend on the large model, which only enhances the value on the original basis.

Second, be sure to do specialized AI in specific verticals.

This is the fundamental reason why these companies have survived to this day.

Not killed by OpenAI? Almost

The CEO of an AI interview company recalled the horror story released by GPT: "After the release of GPT, VCs collectively missed the domestic AI start-ups, and we almost died if we couldn't raise money." ”

Before the release of GPT, the company developed its own first-generation AI interview model with hundreds of millions of parameters, and he believed that "at that time, our technology was on par with the United States", but after the release of GPT, everything changed.

GPT is the new rules of this AI race. The most terrifying thing is that the NLP technology they used at that time could not find a "pick-up man" in the venture capital circle where CV algorithms are popular.

If you can't beat it, join in. As a result, some companies chose to access OpenAI's API, but it also caused a small storm.

The person in charge of the company told Tiger Sniff: "After the large model that called the OpenAI API was removed from the shelves due to data security issues, it only took me a second to find inner peace." ”

The reason is that they found that they still had to take the road of self-development in the end, so many companies chose to make self-developed large models based on the open source model Llama 2.

Now it seems that not being able to call OpenAI's API is not a bad thing for these companies.

After last November's conference, more and more developers began to feel anxious that if they continued to use OpenAI's API for application-layer development, OpenAI could eventually release a product that competed with them. In this case, the closer the application layer is to the API tool, the more dangerous it becomes. Only by finding a scene that can be commercialized can we have the last laugh.

The fortunate thing about this track is that the value of this track is not how superb the technology of the underlying model is, but that PMF (Product-Market Fit) has been done before the large model is used.

This is typical with a nail first, then a hammer. So where is the nail in the recruitment scenario? Let's look at two sets of data first.

1. BYD will recruit 31,800 fresh graduates in 2023, and they received 120,000 resumes within 24 hours of the announcement.

2. The annual recruitment budget of a labor-intensive group branch is hundreds of millions of dollars.

This leads to two problems:

1. HR cannot quickly interview and screen a large number of candidates in a short time, which leads to the fact that many companies' school recruitment will even use lottery to screen candidates from thousands of resumes, and many resumes will not even be opened by HR.

2. In the blue-collar recruitment market, such as factory assembly line workers and courier brothers, there is a huge profit and rent-seeking space for a large number of jobs recruited with the help of labor intermediaries, and it is difficult to recruit with the unified standards of a group.

To a certain extent, the AI interview has knocked down these two nails: it gives each candidate the opportunity to be interviewed by the AI interviewer once, interviewing tens of thousands of people in one day, and it also allows chain groups such as SF Express to unify standards and be more open and transparent when recruiting couriers, cutting off the profit margins of labor intermediaries.

Liang Gongjun, CEO of a blue-collar AI interview company, told Tiger Sniff that by the end of 2023, their system had interviewed 8 million people, and 10 million people are expected to be added in 2024. At present, the SF Express brothers you see have all been screened by the Haina AI interview system, and this happened before their large model was launched.

Therefore, AI interviews solve some real problems based on recruitment scenarios, but this does not rely solely on the capabilities of large models.

With the big model, you can solve more problems. Liang Gongjun found that the technology they used before was NLP and rule engine in the AI 1.0 era, when most of the AI interview questions were fixed questions, and the scoring criteria for judging candidates were also based on keyword recognition. Only blue-collar workers with lower requirements can be interviewed.

After the large model is launched, they can also roll a roll in white-collar workers and school recruitment scenarios with high interview requirements. Interview questions will be generated according to the candidate's competency and a second round of follow-up questions will be formed immediately based on the candidate's answers. The large model scores candidates based on their performance and outputs a full set of reports.

In addition, after the wave of large models has swept away, customers have become more receptive to AI interviews, which is undoubtedly a good thing for AI interviews that are essentially SaaS.

Deeper, drooping, faster

AI investor Ben Parr believes that in order to avoid becoming cannon fodder for OpenAI and other large manufacturers, AI startups must focus more on serving the customers you serve and have more vertical data.

What a subdivided vertical is the AI interview track?

Liang Gongjun told Tiger Sniff that the AI interview is the next link in the entire AI recruitment scenario, and the AI interview is divided into two different route scenarios: white-collar school recruitment and blue-gray collar recruitment. Taking blue-collar workers as an example, blue-gray collar recruitment in different industries such as couriers and factory workers have different recruitment models.

He believes that the barrier to this track is not in technology, but in the complex engineering advantages of industry experts and massive data.

To gain an engineering advantage, you need to focus on a specific vertical scenario. In this case, the choice of the main route is particularly important.

Haina AI chose to focus on the blue-gray collar route, based on the following three judgments:

First, there are 850 million blue-collar workers and 350 million white-collar workers in China, and blue-collar workers are highly mobile, of which about 300 million blue-collar workers have to find jobs three times a year, assuming that each job change interviews with three companies, it means that the total number of blue-collar workers interviewed in a year is 2.7 billion. The number of people employed in these concentrated fourth- and fifth-tier cities is the real silent majority.

Second, due to the current technical limitations, AI interviews cannot be cut into high-end talent recruitment scenarios. However, for blue-gray collar recruitment with relatively simple assessment standards such as couriers, coffee shop clerks, and supermarket employees, the technology of the AI 1.0 era is enough to penetrate these tracks, and even faster and better than human interviewers.

Third, a huge trend is gradually emerging in this huge market - chaining and head concentration.

Taking Luckin Coffee as an example, they only have 13,000 stores in August 2023, but after rapid expansion through franchises, Luckin Coffee now has nearly 17,000 stores.

These rapidly expanding giants all need a uniform set of criteria for hiring employees. And AI interviews naturally have the attribute of standardizing talents.

This has enabled AI interview companies to accumulate dedicated, niche datasets to train models in blue-collar recruitment scenarios. The recruitment database, which is inaccessible to these big companies, is a "fence" separating OpenAI from startups.

Although generative AI is in full swing, for startups: not trying to build generative AI tools for everyone, but building for verticals with special needs, is an important reason why this track has not disappeared in the long river of the AI era.

After the AI competition has entered the white heat, an obvious trend is that domestic AI manufacturers have also begun to move on this track. However, some entrepreneurs are still full of confidence in this: they believe that the unique user platform advantages of large manufacturers should cut into the AI recruitment simulation track for candidates.

An entrepreneur posted such a circle of friends: If the national team is out of the field and dominates the rivers and lakes, then AGI is just around the corner.

I asked him: Do you have a sense of crisis?

He replied that it is not possible to cover all applications with the most advanced pedestal models.

In the 90s, Microsoft also wanted to completely monopolize the PC and software market, with the vision of "putting Windows on every computer", and for this reason, it did not hesitate to shoot Netscape (the browser with the largest market share) and other up-and-comers.

But later, the era of a hundred schools of thought contended came.

The biggest problem with this track at the moment is not these potential competitors, but the overall shrinkage of the job market, which has brought about a decline in the demand for potential customers.

An employee of a large technology company revealed: "Last year, our department opened dozens of positions on the official website, but in fact we didn't plan to recruit anyone. ”

In this case, I'm afraid that even OpenAI will not be able to recover.

This track has been toppled by the wave of OpenAI, relying on PMF and enough vertical to outperform the big manufacturers, but how long it can survive is still unknown.

People who are changing and want to change the world are all in the Tiger Sniff APP