Building Models for One Hundred Countries: The Ambition of Stability.AI's Controversial CEO
TechFlow Selected TechFlow Selected
Building Models for One Hundred Countries: The Ambition of Stability.AI's Controversial CEO
Every country, every industry, and every culture should have AI models that represent their unique characteristics.
Navigating Web3 tides with focused insightsBy Wang Chao
On March 23, Emad Mostaque, the controversial CEO of Stability AI, announced his resignation, stating he will now fully dedicate himself to decentralized AI. This news sent significant shockwaves through the market.
Stability.AI once stood alone as a champion of open-source AI, making vital contributions to the field. Yet it has also faced persistent criticism for taking undue credit for work by other research teams. Meanwhile, revelations about CEO Emad’s falsified academic credentials and his frequent exaggerations on social media further damaged his reputation. Despite once enjoying high-profile investors and massive funding rounds, Stability.AI has long been mired in financial distress, teetering on the edge of collapse.
After stepping down, Emad revealed he still holds a majority stake in Stability.AI—enough to control the board. He insists he was not forced out, but rather chose to resign because in the field of AI, extreme concentration of power is harmful to everyone. By stepping aside, he hopes to drive change within Stability.AI. However, given Emad’s history of embellishment, most people suspect there’s more to the story. Still, beyond Stability.AI itself, I’m more interested in exploring the decentralized AI path that Emad now intends to pursue.
A few weeks ago, I participated in a discussion on decentralized AI with Emad. Since then, I’ve reviewed many of his past statements on the topic, piecing together a rough outline of his perspective.
Who controls the model, controls thought
If human actions are driven by an operating system, AI is rapidly becoming the external core component of that system. As humans increasingly offload cognitive tasks to AI, this technology becomes part of how we think. While AI brings convenience and empowerment, it also carries significant risks—whoever controls the AI models gains substantial influence over global thought.
If the public lacks understanding of how these intelligent tools function and their default configurations, our decisions and opinions may be subtly shaped without our awareness. The controllers of AI models can nudge choices, beliefs, and behaviors by setting specific defaults. As AI becomes foundational infrastructure for the next era, its control by only a few corporate entities could lead to disastrous consequences. This is precisely why Emad emphasizes the importance and urgency of decentralized AI.
Every country needs its own model
While OpenAI devotes immense effort to its SuperAlignment project, who ensures alignment between OpenAI itself and every nation, industry, and culture around the world?
No one.
OpenAI’s Super Alignment initiative makes strong efforts toward fundamental safety and shared human ethics, but is that sufficient when facing the vast diversity of nations and cultures? Different ethnic groups and societies often hold values vastly different from those of Silicon Valley elites. Can these diverse values be fairly represented in AI models? When students in countries like Kenya begin using Silicon Valley-developed AI at scale for learning, might their unique national cultural identity gradually fade away?
The outlook is far from optimistic. Therefore, Emad argues that every country, every industry, every culture should have its own AI model reflecting its distinct characteristics. These models should be deeply rooted locally, absorbing and embodying the collective wisdom of each nation, sector, and culture. This concept isn’t entirely new—just two months ago, NVIDIA began promoting “sovereign AI” across various forums. Fundamentally, it's the same idea. But Emad had already been advocating this vision publicly for over a year, well ahead of NVIDIA.
Most countries today lack the capability to build their own AI models—and this gap is exactly the market Emad is targeting. He aims to create a foundational stack enabling each nation, community, and industry to develop their own AI models. On top of this stack, he envisions decentralized, collaborative development.
Emad has suggested launching or incubating a series of companies, each staffed with domain experts focused on key areas such as education, healthcare, finance, and country-specific AI models. In the context of decentralized AI, these companies would primarily act as catalysts—providing base models and standardized frameworks to invite community participation. If enough top-tier talent from a given country contributes, their collective intelligence could coalesce into a powerful national model.
Data is the core
To put it simply, the recipe for AI models consists of algorithms and data, mixed together with computing power. The more data, the greater the computational resources required to process it. Most teams today focus on improving model algorithms, acquiring more data, and scaling up compute. But practice has shown that high-quality data can achieve excellent results even with smaller volumes. In other words, people are using brute-force computation to compensate for low-quality data.
This insight highlights a key advantage of the decentralized AI system Emad advocates. He believes that by creating a structure that engages top talent within a country, it’s possible to assemble high-quality national datasets—datasets that are verifiable, clearly owned, and capable of supporting incentive mechanisms centered on data contribution.
Through this approach, we can gather data previously inaccessible—data that is not only higher in quality, but also more authentically and fairly representative of public voices and needs.
Small model clusters vs. single large model
In AI, scaling laws have almost become an ironclad rule—an unavoidable reality whether or not we actively pursue them.
Clearly, achieving artificial general intelligence (AGI) via decentralized resource organization isn't feasible in the short term. For the foreseeable future, community-driven AI models will struggle to compete with giants like OpenAI in the race for the most powerful monolithic model.
However, pursuing AGI and building widely applicable AI are two different goals. With continuous technological progress, community-driven small and medium-sized models are rapidly improving. Within one or two years, they may already suffice for most everyday tasks. They may not be the strongest, but they’ll be practical enough and cost-effective enough to unlock vast application scenarios. Much like how most online shopping doesn’t require overnight express delivery, hybrid model usage will likely become mainstream.
This brings about a crucial shift: when collectively developed models become widespread, the risks associated with a single entity controlling a dominant model are significantly reduced. If a large model’s data becomes corrupted, these community-based models can serve as easy-to-use calibration tools, providing necessary corrections. This isn’t just about utility and cost savings—it’s a battle between collective intelligence and AI godhood.
From a technical standpoint, small models aren’t inherently inferior. Their smaller size makes them easier to fine-tune for vertical domains. While their overall capabilities may lag behind larger models, they can excel as expert-level tools within specialized fields. A cluster of such expert models may very well outperform a single large model in real-world competition.
Even more importantly, smaller models enable truly decentralized deployment. Decentralization isn’t just about distributed model development and data sourcing—it also encompasses governance and deployment. If open-source models can be easily deployed on personal laptops or even smartphones, this enables AI democratization. Even if centralized service providers shut down, users can continue operating with local AI. Enabling unrestricted, widespread access to AI is a core goal of decentralized AI.
AI + Web3 – Scam or the Future?
Unquestionably, Emad’s push for decentralized AI is closely tied to cryptographic technologies. He has stated his intention to design a Web3 protocol to integrate and realize his vision. This is because AI currently lacks several critical components—data verifiability, clear data ownership, large-scale coordination and incentive mechanisms, and collective governance capabilities—precisely the strengths of Web3 technology.
Here, I want to particularly emphasize governance. Never before has a technology been as powerful and poised to so deeply and broadly impact every corner of the world as today’s AI. Who should decide its future direction? Who can effectively control it? Governance by a handful of corporate boards like OpenAI’s is clearly not the optimal solution. Nor is rigid regulatory oversight necessarily effective. Collective governance may be the true answer.
In the Web3 space, experiments in collective governance are flourishing, spanning data governance, application governance, network governance, and organizational governance. While most attempts remain exploratory and have suffered repeated failures, this is the frontier of human governance evolution.
Over the past five years, especially around decentralized autonomous organizations (DAOs), the crypto world has nearly tried every known form of human governance. The innovative structure used by OpenAI—where a nonprofit foundation controls a for-profit company—has long been standard practice in DAOs. In my view, people in the Web3 world have essentially played a "speedrun" of governance history, reenacting millennia of human governance evolution in just a few years.
A common critique is that most Web3 governance merely replicates existing models with on-chain voting tacked on. But history shows us that with rapid iteration and high talent density, entirely new systems quickly emerge.
An imperfect analogy is internet advertising. In the early days of the web, opening a news site meant being greeted by a slow-fading fullscreen ad, surrounded by dense blocks of ads—a defining memory of early internet experience. Back then, no one knew what effective online advertising looked like, so traditional media ads were simply copied onto websites. But as contexts evolved and understanding of internet technology and culture deepened, entirely new, highly efficient advertising models emerged, quickly rendering traditional formats obsolete online.
I believe AI governance will follow a similar trajectory. Blockchain technology offers unprecedented coordination and governance capabilities. From this foundation, humanity will develop collective governance models never seen before. I am confident in this evolution.
Final Thoughts
Predicting the future of the AI era and evaluating Emad’s entire vision today is difficult. Clearly, his plan faces enormous challenges at multiple levels. Combined with Emad’s history of exaggeration, it’s hard to distinguish which parts of his vision are genuine and which are mere speculation.
Yet, discussing AI’s power structure is an early, highly complex, and profoundly important issue. Emad and others pursuing decentralized AI—their thinking, their experiments—though likely far from final answers, deserve serious attention and respect. These endeavors, though arduous, represent courageous steps toward shaping the future. Regardless of outcome, these efforts will become a chapter in history’s grand narrative.
Perhaps one day, the world will thank people like Emad.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
