It may be the first AI star company to fall

When the tide goes out, it becomes clear who is swimming naked.

3 months ago, the CEO of Hugging Face, the world's largest AI open source community, made 6 predictions for the AI ​​industry in 2024. One of them is:

A popular AI company will go bankrupt or be acquired at a very low price.

I don’t know if he had a specific candidate in mind at the time, but not long after the beginning of the new year, reality really played out like a script.

The technical team left en masse and the popular AI unicorn collapsed.

When it comes to Vincent diagram models, there are three products that cannot be skipped: DALL-E, Midjourney and Stable Diffusion.

DALL-E is built into GPT-4 and is easy to summon. Midjourney has become a legend because its entrepreneurial team only has 11 people, and its style of painting is more artistic.

Stable Diffusion is more famous because of the product itself, which is open source, free, and free. The company behind it, Stability AI, is not well-known.

Recently, Stability AI has been in trouble again: Robin Rombach, one of the developers of Stable Diffusion, resigned from the company.

▲ Robin Rombach.

The reason for emphasizing "again" is that over the past year, the company's technical team has been in constant turmoil, with changes in the chief technology officer, chief information officer, vice president of engineering, vice president of product, vice president of R&D, research director, person in charge of large language models, etc. Resigned one after another.

In August 2022, Stability released Stable Diffusion to the public, and its daily users quickly reached 10 million. In October of that year, Stability completed approximately US$100 million in financing, with a valuation reaching US$1 billion, joining the ranks of "unicorns."

But after more than a year, why did Stability fall from the altar?

The reason for the resignation was not announced to the public. Forbes believes that there is only one reason for the brain drain: the company has been unable to make ends meet, cannot make money and cannot raise capital.

It’s easy to understand why closed-source DALL-E and Midjourney use subscription models to make money. As an open source project, Stable Diffusion is free for everyone to download and use, but it can still grow bigger and stronger.

Just like Red Hat, a model in the open source industry, it sells high-quality enterprise-level software services to the B-side in a subscription model based on the free open source operating system Linux, and was acquired by IBM for US$34 billion in 2019.

Stability's profit methods are mainly divided into two types: API access to Vincent's web application DreamStudio, and AI consulting and model services for enterprises.

In December 2023, Stability launched paid plans starting at $20 per month for business users. Non-commercial use for personal and research needs is still free.

However, being in the Cambrian Explosion of generative AI is like participating in a battle royale, where personal abilities, network resources, and off-site factors are all indispensable.

Unlike Meta and Google, Stability has a deep foundation, and unlike OpenAI, which has Microsoft's computing power and funds, it signed a contract with Amazon's cloud computing department in accordance with industry rules and used more than 4,000 Nvidia chips to train models.

AI model training costs money, the business is not big enough, and eager competitors are grabbing market share. In October last year, Stability received $50 million from Intel to replenish its blood, but no amount of money could afford to continue losing money.

Bloomberg reported that Stability’s training and salary expenses in October 2023 were approximately US$8 million, but its software and service revenue in August of the same year was only US$1.2 million.

Forbes explains the difficulty of Stability from another angle, saying that workers should be more empathetic. The company was not only suspected of owing employees wages and money to business partners, but even forced Amazon to threaten to cut off GPU access because of failure to pay money on time, but Stability denied this.

The copyright dispute over AI images has made it even worse for start-up companies that were not already wealthy. Getty Images and artists contend that Stability stole artwork and stock photos to train its models, and the lawsuit is ongoing.

Coincidentally, the relationship between Stability and its competitor Midjourney has also fallen into a freezing point recently.

In early March, Midjourney blocked all Stability employees after investigating two paid accounts suspected of being Stability employees trying to steal data, causing a lengthy server outage.

The CEO of Stability responded at

People who do practical things, people who tell stories

The relationship between Stable Diffusion and Stability is not as pure as ChatGPT and OpenAI.

Stable Diffusion originated from an academic project, formerly the Latent Diffusion model, and was researched by Ludwig-Maximilians University in Munich and Runway, an AI company in New York.

Runway is the Runway of Vincent Video Circuit. It turns out that the gears of fate have already started turning, and the technology industry is also a circle of acquaintances.

▲ Picture from: arxiv

Seven months after the project's research paper was released, in June 2022, Stability joined as a funder and provided computing resources. Stability's own generative AI team also participated in the research.

Robin Rombach, who resigned this time, was the leader of the Munich research team and the first author of the paper, and was later recruited by Stability. The two colleagues from Munich who signed the paper next to him are suspected to have left with him.

Stability released Stable Diffusion 3 in February this year, and released a new 3D video model SV3D in March. These technicians feel like they are standing at the last post before packing up and leaving.

▲ SD3 has evolved in terms of multi-topic prompt words, image quality and text spelling.

In the past year, Stability has been diligent, upgrading Stable Diffusion, and developing new models: Vincentian graph model SDXL, large language model StableLM, audio generation model Stable Audio, video generation model Stable Video, code generation model Stable Code, and 3D object generation model Stable Zero123…

But competitors are not idle either. For example, being suppressed by GPT-4, Gemini, and Claude, StableLM has barely made any splash. Meta has also open sourced Llama 2, which directly competes with Stability.

Nowadays, many core talents from the founding period have left. Although the people will not be left empty, the morale of the military will inevitably be shaken. Similar examples are around the corner.

Recently, Mustafa Suleyman, co-founder of the startup Inflection AI, accepted an offer from Microsoft to lead Microsoft's consumer-oriented AI business. He is also the co-founder of Google's DeepMind.

▲ Mustafa Suleyman.

Inflection’s main product is the chatbot Pi. Compared with its peers, Pi's emotional intelligence is particularly outstanding. It is more empathetic and cares about you. It likes to add emojis. It is suspected that it suffers from chat flattery syndrome. When chatting with it, it will unconsciously relax.

Also joining Microsoft are Inflection co-founder and chief scientist Karén Simonyan, as well as some Inflection engineers and researchers.

During the OpenAI court battle, Nadella's plan to acquire OpenAI without pain fell through by recruiting Sam Altman and hundreds of employees who were determined to follow him. Wan Wan Lei Qing, now his wish is fulfilled by Inflection.

This time Microsoft packed away the Inflection executive team and integrated its model into Microsoft's cloud computing business Azure. The money spent was far lower than the valuation. Although it was not an acquisition, it was close to an acquisition in nature. Inflection and Pi existed in name only.

It is not difficult to see that when the AI ​​industry is in turmoil, talent is still the key. Wherever there are people, there will be grievances, and where there are grievances, there will be conflicts. The technology industry is no exception, with constant internal fighting and endless business wars.

Talent exodus is a story Google is all too familiar with. At the same time, Stability's top management style has been questioned, just like Google, which has suffered many overturns.

▲ Emad Mostaque.

Emad Mostaque, CEO of Stability, was a hedge fund manager in his last career and spent most of his time in oil trading. Although he has a degree in computer science, he has never founded or joined a technology company before. He also has the habit of talking big and taking credit. Corporate culture is influenced from top to bottom.

Several current and former employees have revealed that Mostaque often exaggerated. He called Amazon a "strategic partner" when financing, but in fact the two were just ordinary parties A and B. He said Stability was building custom AI models for dozens of clients, but when asked for specific names, only an Indian holding company and an African country were named.

What's even more egregious is that the relationship between Mostaque and his wife, a former public relations executive, and the company's finances is unclear despite the fact that the company is already in financial difficulty.

In July 2023, former Google researcher David Ha resigned after 8 months as Stability research director. In an interview with the media, he said that despite entering the game early and promoting the development of the AI ​​industry, Stability still needs to find its own moat to remain competitive. Power, it seems that there is something in the words.

After many talents have left, Stability's own technical capabilities are indeed a question waiting to be answered.

In the press release announcing the financing, Stability, intentionally or unintentionally, obscured the existence of Stable Diffusion researchers and only said that Stable Diffusion is a powerful, free and open source text-to-image generator, and that it is the company behind this product. .

It is true that half of the words are left out, but Professor Björn Ommer, who left his name on the paper, is very dissatisfied with this: "In scientific research, if you take something from somewhere, you should attribute it to someone and cite someone. "

The current chief researcher of Stability is Varun Jhampani, who quit Google Brain. I am afraid only he knows how long he can be retained.

The stage of selling imagination is over, and the bloody shuffling period has arrived.

The free and open source nature of Stable Diffusion is of great significance to the AI ​​community.

Its license is very relaxed, requiring only ethical and legal use of the model, allowing commercial and non-commercial use, allowing anyone to freely create, develop products and even launch their own projects.

For example, for the problem of character stability in AI-generated images, Midjourney proposed the Character Reference solution, and the open source community launched the add-on IP Adapter for Stable Diffusion.

At the same time, open source models are more likely to become weapons for Deepfake to generate violent and pornographic content. What's more, some startups based on Stable Diffusion do not mark their licenses and technology sources, and just use a new look to cheat money.

Advocates of open source often argue that AI development is better when people don't have to pay to access models, and that there's more transparency into how models are trained.

But companies that develop models may not feel so. Because of concerns about "competition and security issues," OpenAI has stopped sharing research results starting from GPT-4, and has replaced them with technical reports.

From a business perspective, OpenAI’s decision is understandable. The duality of open source and the conflict between open source and closed source are also the epitome of the current AI competition.

In 2023, the development of generative AI will be like a raging fire, with giants and start-ups competing to release models and commercialized products. However, the thriving scene not only breeds success, but also contains lies, harm, fishing in troubled waters and exaggeration. The accompanying runner, who was enjoying the moment, carefully waited for the moment when the bubble burst.

Mostaque, a former fund manager, is actually very keen on changes in water temperature. He has sensed the intensifying competition in the AI ​​industry very early and once warned employees: "You will all die in 2023."

There are already too many choices in the market, and temporary valuations are just a mirror image. How to stand on its own feet is the next topic for AI companies.

Inflection only raised US$1.3 billion in June last year, but after receiving an olive branch from Microsoft, its executives and key executives left just because they couldn't see the future.

In an interview with Bloomberg, Mustafa Suleyman said that the number of daily users of Pi has reached 1 million, but it has not yet found an effective business model.

Although Pi has a friendly and likable attitude, a pure chat robot for consumers cannot support such a high valuation. It is better to turn to enterprise customization services and be entrusted with important tasks by giants.

Although Inflection promises that Pi's services "will not change immediately," some people feel that its death will be sooner or later. Similarly, the big guy who developed the model has left, and users can’t help but worry about whether Stable Diffusion 3, which is still being invited for testing, can still be open source.

In response to the news of Robin Rombach's resignation, a Stability spokesperson thanked him for his efforts and wished him a bright future. At the same time, he emphasized the company's development focus: commercializing technology.

When it comes to the commercialization of AI, everyone has their own problems. In February this year, Amazon Cloud Computing Services CEO Andy Jassy said in a conference call with analysts that the revenue from generative AI "is still relatively small."

But he added that nearly all of Amazon's consumer businesses already have or will have generative AI products that are expected to bring in tens of billions of dollars in revenue in the coming years.

Microsoft doesn't want to be called "OpenAI's IT department" and make investors dissatisfied, so while it is developing internal AI projects, it is poaching people everywhere and investing in Mistral AI, a competitor of OpenAI.

In this regard, OpenAI also has a strong say. When it was first established in 2015, it was a non-profit organization to develop AGI in a way that was most likely to benefit all mankind. In 2019, it was transformed into a limited profit company and attracted Microsoft to invest. The reason It’s actually very simple: AI is too expensive.

OpenAI's annualized revenue exceeded US$1.6 billion last year, and it mainly relies on three pillars to make money: member subscriptions, fees for developers' API calls, and revenue sharing with Microsoft products.

Musk has repeatedly criticized the "degeneration" of OpenAI, saying that it has become a closed-source company controlled by Microsoft that pursues profit maximization. Sam Altman was indifferent and expressed that he missed the Musk of the past.

2024 is the second year of generative AI, and the hot topics have shifted from chatbots and Vincent pictures to Vincent videos and robot embodied intelligence.

The emotions of the contestants have also become more nervous. If they have not been separated from the financing supply and have not found a commercial landing scenario, it may not take long to see many so-called star companies go bankrupt.

On the morning of March 23, Stability suddenly announced that CEO Emad Mostaque would also quit, resigning as CEO and the company's board of directors, and the chief operating officer and technology officer would temporarily take over the responsibilities.

Emad Mostaque said that the concentration of power in AI is not good for everyone and will go all out for decentralized AI.

The CEO of HuggingFace also came to watch and said that he might consider buying Stability and making Stable Diffusion 3 open source.

In the early days of AI development, market optimism is high and will tolerate the selling of dreams, rough experiences, and imaginations of the future.

However, the value of artificial intelligence must be reflected in practice. The enthusiasm of lobbying investors should be transformed into products that consumers are willing to pay for. The poetry of technology must still be written on the land of the Internet rather than in the void.

One of Forbes' top 10 AI predictions in 2024 is: Stability will go out of business. It may be right, the question is, how much stability will come next.

It is as sharp as autumn frost and can ward off evil disasters. Work email: [email protected]

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo