The picture is from the movie "Infernal Affairs"

©Industry Quadrant Original

Author: Shancha

edit丨Qianjiang

withoutSome people dare to easily question a company that is leading the times, provided that there are no problems of its own.

butOpenAIThe problems have obviously been exposed.

In November 2023, a "palace fight" surrounding the stay or departure of OpenAI CEO Sam Altman became one of the most important news in the technology circle at the time.

A similar thing happened recently. According to The Information, John Schulman, co-founder and head of alignment at OpenAI, has jumped to OpenAI's competitor Anthropic.

While Schulman was sparking public opinion, he and Altman were involved in a palace fight last year.VortexGreg Brockman, president and co-founder of OpenAI, also revealed on X that he is on a long-term leave that will last until the end of the year.

According to netizens’ speculation, Brockman’s leave may be due to the signing of a confidentiality agreement, which means that after the leave, Brockman will most likely leave OpenAI. Prior to this, Ilya Sutskever, co-founder and chief scientist of OpenAI, had already left in May this year and founded his own company.

So far, the original 11 founders of OpenAI remain stable, with only Sam Altman and Wojciech Zaremba, who is responsible for OpenAI's language and code, remaining.

Putting these things together, it's clear that OpenAI is going through an extremely painful tear.

This turmoil may ultimately affect the question of when and how humans will achieve AGI.。

Cracks in the fortress

From the information available, the core reason for tearing up OpenAI is:Regardless of everythinglandMove forward, forward, and forward again towards commercialization?It is better to give up commercialization appropriately and take a more stable, reliable and safer path.

In fact, the differences caused by this route dispute have existed in OpenAI for a long time and have always been the fuse that caused chaos within OpenAI.

In 2020, Dario Amodei, then OpenAI's vice president of research, and Daniela, then OpenAI's vice president of security and policy, resigned together. They led OpenAI's core employees to establish a new company called Anthropic AI next door to OpenAI.

One of the important reasons why Dario and others "rebelled" against OpenAI was that they felt that OpenAI did not invest enough in AI safety or did not pay enough attention to AI safety issues. Therefore, the purpose of Anthropic, which they founded, is to develop safer and more controllable AI.

now,Anthropic hasBecome OpenAI's most important competitor.Anthropic is therefore called the rebel of OpenAI.

The same story happened again at the end of last year, when the "cautious and safe" faction represented by OpenAI's chief scientist Ilya launched a mutiny against Altman, but ultimately failed due to the collective support of OpenAI employees for Altman.

Until this year, Schulman, who just left his job, was also troubled by the issue of "AI safety".

On August 6, OpenAI co-founder John Schulman officially announced his resignation on X. The style of his farewell speech was exactly the same as that of Ilya Sutskever, who officially announced his resignation three months ago.

For example, they all expressed in their writing that "leaving OpenAI was a difficult decision." In addition, they also expressed their obsession with "AI safety alignment" research.

The whereabouts of the two people after leaving their jobs also proves this.

Ilya Sutskever founded SSI, an intelligent security company, after leaving the company to improve AI safety capabilities. He wrote on X that unlike the safety work at OpenAI, "we focus on one thing, which means we are not disturbed by management expenses or product cycles."

John Schulman, who just announced his resignation, also said that he will achieve the goal of "AI safety alignment" at Anthropic.

▲Screenshot of Schulman and Ilya’s tweets

But there is a lot of drama in this matter.

For example, Schulman, who has now jumped ship, and Brockman, who has been "marginalized," were both staunch supporters of Altman during the "palace fight" incident last November. Brockman was the most prominent among them, as he was dismissed and returned at the same time as Altman.

Judging from their stance last year, they do not seem to agree with the cautious safety represented by Ilya, and believe that OpenAI needs Altman more; but judging from their choice today, they do not seem to agree with Altman’s radical commercialization strategy.

Obviously, in thisVortexAt the center of the vortex, Schulman and Brockman are the most conflicted people.On the one hand, they hope OpenAI can accelerate commercialization, but on the other hand, they are worried that the huge potential of AI will bring unpredictable consequences.

Perhaps, at the time of the last coup, Schulman and others were still trying to find a balance between radical commercialization and cautious AI safety. But obviously, after six months of development, they have lost hope of finding this balance in OpenAI.

We have no way of knowing what has happened to OpenAI over the past six months surrounding this conflict.But what we do know is that the rift within OpenAI caused by the dispute over safety and commercialization is getting bigger and bigger.

From the outside, OpenAI is still strong. No company in the world can really compete with it. But the great dams often collapse because of ant holes, and the strong castles often begin to collapse from the inside.

The cracks at OpenAI have already appeared. Will they spread into an unstoppable avalanche?

The OpenAI paradox

In fact, OpenAI’s challenge does not lie in internal disagreements, but rather in the difficulties of its own business model, which may be the reason why Altman is determined to accelerate commercialization at all costs.

Because if it cannot achieve self-sustaining growth, OpenAI may face collapse in the near future.

In June this year, Altman disclosed revenue data to employees. According to the revenue estimate for the first half of 2024, OpenAI's revenue for the whole year of 2024 could reach about US$3.5 billion.

$3.5 billion in revenue is an astronomical figure for most companies, but it is far from enough for OpenAI.

The American technology media The Information has made an estimate that based on OpenAI's expenditure, its expenses in 2024 may be as high as US$8.5 billion. Even if its revenue of US$3.5 billion is included, its losses in 2024 will be as high as US$5 billion.

At this rate of consumption, without new capital injection, OpenAI's cash flow will be exhausted within a year.

OpenAI can certainly continue to seek support in the capital market, but this still cannot solve the problem of OpenAI's long-term survival, and its commercialization still faces challenges.

In July this year, the foreign AI research organization FutureSearch released a research report on OpenAI's revenue structure, which estimated that OpenAI's recurring revenue will reach US$3.4 billion in 2024. This is close to the data released by Altman some time ago.

According to FutureSearch's estimates, OpenAI's revenue is mainly divided into four parts.

Among them, ChatGPT Plus subscription revenue for ordinary users was approximately US$1.9 billion, accounting for 55% of the total revenue; ChatGPT Enterprise for large corporate customers contributed approximately US$744 million in revenue, accounting for 21%; ChatGPT Team for small and medium-sized enterprises and team users contributed approximately US$290 million in revenue, accounting for approximately 8%; finally, API interface services contributed approximately US$510 million in revenue, accounting for 15%.

▲Image source:FutureSearch

As you can see, subscription revenue is still OpenAI’s main source of revenue, accounting for 84% of its total revenue. FutureSearch shows that OpenAI has about 9.88 million paying users per month, of which 7.7 million are ChatGPT Plus users.

But this part of OpenAI's base is also declining because people are disenchanted with AI.byTo the challenge.

According to data reported by foreign media, the number of visits to OpenAI began to drop sharply after April this year, from 1.8 billion visits per month at its peak to less than 300 million in June. This will obviously also have an impact on the revenue of OpenAI, which mainly relies on C-end users.

▲OpenAI website visits

In addition, OpenAI's other commercialization explorations are also facing challenges.

In November 2023, OpenAI released GPTs, which is similar to the APP Store, in order to enrich the application ecosystem. After it was launched in January this year, the novelty of GPTs quickly faded, and coupled with the insecurity of the GPTs product itself and the limited value it provided, its traffic continued to decline.

In February, GPTs’ visits accounted for only 1.5% of the ChatGPT webpage. In January of this year, Microsoft launched a similar product called GPT Builder, which was shut down in June.

▲picture:One user named X said that GPTs have some flaws, such as causing hallucinations and data leakage.

Although commercialization has not met expectations, OpenAI is still increasing its investment in research and development. According to FutureSearch's calculations, as the company develops more complex models, OpenAI's expenditure is expected to continue to exceed revenue growth, and the company may need to raise tens of billions of dollars to meet costs.

OpenAI can certainly continue to obtain financing in the capital market, but between the huge amount of capital investment and the still distant break-even line, I believe everyone will hesitate for a while longer.。

Collapse begins with the loss of confidence.

In fact, it is not just OpenAI. Almost all large-model startups, while facing the hot industry atmosphere, are also shrouded in the same dark cloud - how to make money.

In June this year, Fu Sheng raised a sharp question at the DingTalk Ecosystem Conference:"No matter what, I still can't understand how big model companies make money."

This problem is like the elephant in the room, which almost everyone can see with their eyes open. But at the same time, most people choose to ignore it intentionally or unintentionally, and they either look to the bright future of achieving AGI or indulge in the current enthusiasm of the big model.

But the problem still exists.

Not long after Fu Sheng raised this question, I happened to participate in a small-scale exchange of large-scale model entrepreneurs. I found at the scene that almost all the entrepreneurs present had undisguised anxiety on their faces.

This is the straightforward answer given by an entrepreneur when asked about the reason. However, the greater pressure lies in the fact that “the money from the last round of financing will be almost spent by next year, and the new financing has not been finalized. Investors are increasingly pressing on the progress of commercialization.”

The news from outside has an even more panic-stricken feeling.

In early June, several star large-scale model startups in the United States were exposed to be in deep trouble of broken capital chain and had to seek to sell themselves, including Adept, Character.AI and Stability AI, which was a star enterprise that launched Stable Diffusion.

The key to selling out is still that large-model startups do not have a self-consistent business model, that is, either the technology cannot make money without a practical application, or, like OpenAI, the money earned is far from covering the company's R&D investment.

But this problem has obviously existed for a long time. Almost since the second half of 2023, the focus of discussions in the large model industry has shifted from developing large models with larger parameter scales and stronger performance to finding the implementation of PMF and product applications.

Investors are more sensitive to changes in the industry. Since the beginning of this year, almost all investors who are concerned about big models and have been contacted by "Industry Quadrant" have expressed that they will focus more on the application of big models.

But this trend did not last long. Starting from June, a number of large-model entrepreneurs mentioned in their exchanges with the "Industry Quadrant" that investors were no longer interested in application-layer startups."Basically, they just watch but don't invest. The first thing they ask is revenue, target customers, and growth targets."

In order to meet the requirements of the capital market and to survive, since the second half of this year, more and more large-model startups that once focused on 2C have begun to change direction and entered the 2B field.

This transformation has also brought another kind of grand occasion to the large model industry, that is, if you just attend a large model industry summit,Among every ten companies I saw, almost five were engaged in AI customer service. Among the remaining five, three were engaged in marketing content (copywriting + pictures) generation, and two were engaged in sales training.

“Although it is competitive, the B-side is easier to implement. It is easier to make money and explain to investors."A startup company manager told us with a helpless toneroad。

But there is actually another subtext behind this tone, which is that we feel that more and more entrepreneurs are beginning to be less confident that big models will change the world.

In fact, although the large models represented by GPT-4 have demonstrated amazing capabilities, at the application level, whether it is Microsoft's Bing and Copilot, or domestic companies that integrate large model capabilities such as DingTalk, Feishu,WPS and Tencent Meeting,What big models bring are essentially “micro-improvements”, and the efficiency gains brought about by these micro-improvements are also very limited. The “disruption” that people imagined has not arrived.

In this context, on the one hand, native scenarios and applications of large models are still a long way off, and on the other hand, the patience of capital is getting less and less. When the return on huge investments is nowhere in sight, confidence begins to slowly waver.

This is obviously a very dangerous signal, because the collapse of any dream begins with the collapse of confidence.。

end

whenHowever, the dream of AGI does not stop there.

To be more precise, today's pessimism is just a common sentiment when the market faces difficulties, because the group of people who truly believe in AGI and are truly leading the progress of the times have never wavered and are still fighting on the front line to realize AGI.

For example, those who left OpenAI, although John Schulman jumped to Anthropic, he will still focus on the work of AI alignment; after Ilya Stuskever left, he founded a new laboratory Superintellgence, dedicated to building safe superintelligence.

According to statistics, in the past five years, more than 30 senior executives of OpenAI have left the company and founded more well-known AI companies after leaving OpenAI, including Anthropic, OpenAI’s main competitor, and Inflection AI and Adept AI, two of the so-called “Four Little Dragons of Silicon Valley AI”.

So from this perspective, OpenAI’s turmoil also has a feeling of “when a whale dies, all things come to life”.

Since the beginning of this year, domestic large-scale model startups are still developing rapidly, especially some enterprises. For example, Dark Side of the Moon has completed two rounds of financing this year, with a financing scale of more than 3 billion yuan. In addition, Zhipu AI and Zero One Everything have also completed hundreds of millions of dollars in financing.

Faced with such a track that is still full of faith and vitality, the market should also give it more time and confidence.

But at the same time, we should also examine these crazy believers and the path to achieving AGI.。