Stability AI Chronicle: Why It Can Fail in 2024

When you ask three language models – ChatGPT, Claude, and Bing – about 'when and how Stability AI started,' you will receive three different answers with three different dates and a lot of loud words about democratization, open source, human potential, and extensive work in AI. That's what characterizes Stability AI the most – inaccuracy, vagueness, and, most of the time, severe exaggeration. It's an example of how a good talker can dominate the narrative of AI models. It is also an example of how the embezzlement of VCs by buzzwords stops them from careful due diligence.

Stability AI is still in the Unicorn Family with a valuation of around $1 billion. But there are opinions that 2024 will be Stability's last year of existence. In any case, the company and its CEO are very interesting cases to study. Let’s explore the world of Emad Mostaque, the way to build a unicorn company in six days, and how it all can fail.

The Forbes correspondents did a really good job, discovering that Stability AI's journey started in 2019 by Emad Mostaque as 'an AI-powered data hub that global agencies would use to make decisions about Covid-19.' It launched with a virtual event hosted by Stanford HAI in July 2020, with AI expert Fei-Fei Li and representatives from UNESCO, WHO, and the World Bank as speakers. Emad was also a speaker, presenting ‘Knowledge Graph Architecture.’

The project didn’t take off. After that, Mostaque was all over the place: vending machine network, emotional support dog NFTs, etc. 

Mostaque certainly deserves credit for seeing the opportunity in Generative AI that, at that moment, wasn't yet a huge thing. He also deserves credit for noticing the potential behind the Latent Diffusion model and lending the computational power he had from AWS to let researchers check and improve the model’s performance.

We don’t know how he presented that help to the researchers, but they did launch an improved model in August 2022 under the name Stability Diffusion. This naming led to another appropriation by Stability AI: it simply made everyone believe that Stable Diffusion equals Stability AI and that Mostaque was the main evangelist behind the model and its open-source nature.

Which was at least an exaggeration but a very beneficial one. It led to ->

There has been a significant improvement in Stable Diffusion's financial situation. Previously, it was not exactly clear how the company was paying its bills. However, a statement provided to Forbes by lawyers representing Zehra Qureshi, Mostaque’s wife, revealed that she had been providing 'emotional and financial support' to her husband’s business since 2021.

October 2022: Stability AI raised $101 million in a seed round led by Coatue and Lightspeed Venture Partners, with participation from O’Shaughnessy Ventures LLC. This funding round catapulted Stability AI into unicorn status. Later, Mostaque would say about this round: “I did the whole round in like six days and I made sure that they were fully supportive and we maintained our independence. So there's no commercialization pressure or anything like that”.

A year later Coatue would ask Emad Mostaque to step down, pointing out that his leadership had prompted several senior managers to leave and placed the startup in a tenuous financial position. The company lost its chief operating officer, chief people officer, VP of engineering, VP of product, VP of applied machine learning, VP of comms, head of research, head of audio, and general counsel.

May 2023: Secured less than $25 million through a convertible note deal with Sound Ventures. Reports have surfaced that Stability AI faced difficulties raising funding at a $4 billion valuation. However, Stability AI issued a statement clarifying that they have not been actively fundraising, nor have they launched a formal round, and they have not experienced fundraising problems.

October 2023: Raised just under $50 million in a convertible note deal from Intel.

The company sells memberships with the main feature of allowing to use what they call the Core Models for commercial applications.

However, according to Pymnts “the company’s expenses, including bills and payroll, surpassed its revenue, leading to concerns about its financial sustainability. Stability AI was spending about $8 million a month on bills and payroll in October. The company generated $1.2 million in revenue in August and was projected to earn $3 million in November from software and services.” 

Stability AI has primarily offered its models for free, encouraging research and experimentation. The new membership model approach marks a strategic shift, aiming to standardize commercial use and facilitate enterprise deployment, while still supporting independent and research initiatives 

Whenever he can, Emad Mostaque emphasizes the continued openness of their models. He said that all membership levels, including commercial and non-commercial users, will have access to the code and weights, similar to the Meta Llama model.

But as with everything involving Mostaque, it seems that he didn’t stand by his own words. And people on Reddit are happy to catch him on that. 

In 2022, Stability AI had an eventful year, marked by both collaboration and controversy. They worked with Runway and LMU Munich to create Stable Diffusion, an innovative AI model that generates images from text descriptions. This project was a big deal because it was open-source, meaning anyone could use and contribute to it.

However, there was a bit of a disagreement with Runway. The discord between the companies arose from the release of the Stable Diffusion model, a collaborative initiative led by Patrick Esser from Runway and Robin Rombach from LMU Munich (also, Stability AI). The model's code was open-sourced under the CreativeML Open RAIL M License, thanks to a trilateral effort involving Runway, Stability AI, and LMU Munich's CompVis group. Tensions flared when Stability AI lodged a takedown request against Runway citing an IP leak with Runway ML's SD 1.5 model, an extended version of Stable Diffusion. Runway's CEO, Cris, refuted the IP breach claim in a Hugging Face discussion, appreciating Stability AI for a compute donation aiding the retraining of the original model. The takedown request was later withdrawn by Stability AI, suggesting a resolution to the disagreement.

Besides this hiccup, Stability AI made significant strides with Stable Diffusion. It became very popular, with tons of developers using it to create billions of images. Stable Diffusion made waves before ChatGPT and was one of the first to pump the generative AI hype stage. The only concerning thing about it is that Stability AI made it look like Stable Diffusion was their proprietary model. 

Nonetheless, in 2023, Stability AI continues to roll out new models and products:

Stability AI's models span across different modalities such as audio, image, video, language, and even 3D. Since the success of Stability AI is intertwined with the success of other researchers and Runway AI, we present the timeline we created for our Runway profile:

In December 2023, Stanford researchers reported that they’d identified over 1,000 child exploitation images in a widely used image database used to train AI models, including Stable Diffusion 1.5. The study on CSAM in generative ML training data, focusing on the LAION-5B dataset used in models like Stable Diffusion, reveals significant findings. It identified 3,226 entries suspected as CSAM, with many confirmed by third parties. Using various detection methods, the research uncovered 1,679 matches through PhotoDNA and an additional 495 through MD5 hash comparison. The KNN analysis based on image embeddings led to the discovery of more potential CSAM instances. This underlines the substantial presence of CSAM in widely used datasets for training generative machine learning models, reflecting a critical issue in the field of AI and machine learning.

As if replying to that, Stability AI announced a new hire: Ella Irwin, Elon Musk’s former head of trust and safety at X, has joined the company as its senior vice president of integrity.

Stability AI's mission is to "democratize AI and build a global foundation to activate humanity’s potential," a goal centered on creating open AI accessible to everyone that serves as a tool "by the people, for the people." On this platform of good intentions and wild exaggeration of self achievements, they built a substantial community, claiming to have over 200,000 members across various research hubs.

Apart from the research made by the internal Stability AI team, there are papers and frameworks created entirely by the community. According to their research page, Stability AI has published nine papers, many in partnership with top-tier universities like the University of Washington, Hebrew University of Jerusalem, and the California Institute of Technology, and research institutes including the Allen Institute for AI, LAION, and the Princeton Neuroscience Institute, among others.

The company also supports the research across the globe providing the compute resources as it’s mentioned in one of the research papers, Humans in 4D, created by the researchers from the University of California and published at ICCV 2023 but supported by the Stability AI compute grant. The team also works on open-sourcing datasets like Pick-a-Pic, a dataset of user preferences for text-to-image generation.

Stability AI presented their views in written evidence to the House of Lords Communications and Digital Select Committee inquiry about large language models, sharing the stage with other tech giants like OpenAI, Meta, Microsoft, and around a hundred other organizations. These views were also restated in the critique of the EU AI Act, arguing that its “one size fits all” approach might inadvertently smother the flames of grassroots innovation.

In their manifesto, Stability AI laid out their vision:

Emad Mostaque tries to show balance in his approach to AI regulation. He's not completely against regulation, nor does he accept it without question.

Stability AI, a once prominent AI startup that gave generative AI a significant push, is on the brink of collapse after a tumultuous 2023. The company has seen a mass exodus of top executives and has consistently been losing talent. Its major investors, Coatue and Lightspeed, distanced themselves due to disagreements with CEO Emad Mostaque. However, Emad Mostaque is still invited to many media and podcasts for interviews, continues to broadcast his views and opinions, and shies away from the reality of the situation. Efforts to raise funds at a $4 billion valuation failed. Although Intel's recent $50 million investment offers temporary relief, Stability AI's high expenditure rate overshadows this. It seems that the company is looking for a buyer, but interest is scarce.

“He’s probably the most visionary person I've ever met,” says Christian Cantrell, who left a two-decade career at Adobe to join Stability in October (he quit six months later and launched his own startup).

But sometimes the line between a visionary and a well-spoken exaggerator is blurry.

Thank you for reading, please feel free to share with your friends and colleagues. In the next couple of weeks, we are announcing our referral program 🤍