Source: Empower Labs
On March 23, the controversial Stability AI CEO Emad Mostaque announced his resignation and will devote himself to the decentralized AI cause in the future. This news caused a great shock in the market.
On the one hand, Stability AI has carried the banner of open source AI alone and made important contributions to open source AI; on the other hand, it has been constantly accused of taking the achievements of other research teams on its own head. The exposure of CEO Emad's false academic qualifications and frequent bragging on social media have further damaged its reputation. Once with the support of shining investors and huge financing, Stability.AI has long been in financial difficulties and is on the verge of death.
After resigning, Emad revealed that he still holds a majority of Stability.AI's shares, enough to control the board of directors. He was not forced to resign, but believed that in the field of AI, the high concentration of power is harmful to everyone, so he chose to resign to promote changes in Stability.AI. Because Emad has an ignominious history of bragging, most people think that things are not that simple. But I would rather talk about decentralized AI, which Emad is going to devote himself to, than Stability.AI.
A few weeks ago, I participated in a discussion with Emad on decentralized AI. After that, I sorted out many of his past speeches on this topic and roughly sorted out his perspective.
Who controls the model controls the mind
If human actions are driven by an operating system, AI is quickly becoming an external core component of this operating system. Because humans have gradually adapted to outsourcing the burden of thinking to AI, this technology has become part of the way we think. However, the convenience and empowerment brought by AI also come with significant risks - whoever controls the AI model controls the world's thoughts to a certain extent.
If the public lacks understanding of the working principles and default settings of these intelligent tools, our decisions and opinions may be quietly affected. The controllers of the AI model can subtly guide people's choices, opinions and behaviors by setting specific default options. As the infrastructure of the next era, if AI is controlled by only a few commercial groups, the consequences may be disastrous. This is what Emad emphasized, the importance and urgency of decentralized AI.
Every country needs its own model
When OpenAI puts a lot of effort into the Super
Alignment project, who is responsible for the alignment of OpenAI itself with every country, every industry, and every culture in the world?
No one.
OpenAI's Super Alignment has made great efforts in basic security and common human ethics, but in the face of the diversity of different countries and cultures, is such an effort enough? Different nations and cultures often have values that are very different from those of Silicon Valley elites. Can these diverse values be fairly reflected in the AI model? When the next generation of students in countries like Kenya begin to use Silicon Valley AI to learn on a large scale, will their unique national cultural characteristics gradually disappear?
The answer is not optimistic. Therefore, Emad believes that every country, every industry, and every culture should have an AI model that represents its own characteristics. These models should be deeply rooted in the local area, fully absorbing and reflecting the collective wisdom of the country, industry and culture. This concept should be familiar to everyone, because two months ago NVIDIA also talked about the concept of sovereign AI on various occasions, which is essentially the same thing. However, Emad started talking about this on various occasions a year or two ago, much earlier than NVIDIA.
Most countries in the world simply do not have the ability to create their own AI models, and this is exactly the market Emad is targeting. He hopes to support the AI models created by each country, nation and industry by creating an underlying stack. On the stack, he hopes to achieve the development of the model in a decentralized collective collaborative way.
Emad once said that he might start/incubate a series of companies, each of which will have different professionals focusing on different key areas, such as education, medical care, finance, and of course AI models for different countries. As a practice of decentralized AI, these companies play more of a role as initiators. By providing basic models and standardized frameworks, community talents are introduced to participate in contributions. If a large number of outstanding talents in a country can be attracted to participate in contributions, these collective wisdom will eventually converge into an excellent national model.
The core is data
Using a simple metaphor, the recipe of AI model is algorithm and data, and then use some computing power to mix them together. The more data there is, the more computing power is needed to stir these data. At present, most teams in the market are pursuing better model algorithms, getting more data, and then matching them with greater computing power. But practice has proved that if the data quality is high, less data can also achieve excellent results. In other words, people are using computing power to clean up low-quality data.
This constitutes an advantage of the decentralized AI system that Emad advocates. He believes that if a structure can be established to guide the participation of a country's outstanding talents, a high-quality national data set can be assembled, and at the same time, these data can be verifiable, ownership is clear, and incentive models around data can be designed accordingly.
In this way, we can collect data that was not accessible in the past. These data are not only of higher quality, but also more authentic and fair representation of the public's voice and needs.
Small model cluster vs. single large model
In the field of AI, scaling laws have almost become an iron law, and we cannot avoid it whether we choose to pursue it or not.
Obviously, it is not realistic to organize resources in a decentralized way and then achieve general artificial intelligence (AGI) through scaling laws in the short term. For quite a long time, community-oriented AI models will also find it difficult to compete with giants like OpenAI for the crown of the most powerful model.
However, pursuing AGI and creating widely applicable AI are two different things. With the continuous advancement of technology, community-driven small and medium-sized models are rapidly improving their capabilities. It is expected that within one or two years, small and medium-sized models will be sufficient for most daily tasks. Maybe it is not the strongest, but it is practical enough and the cost is low enough to open up a wide range of application scenarios. Just like most of our online shopping does not need to use SF Express for the next day, the mixed use of models will gradually become mainstream in the future.
This brings about a very important change. When models driven by collective intelligence are widely used, the potential risks brought by a single large model controlled by a single institution are greatly reduced. If the data of a large model is contaminated, these models based on collective intelligence can also easily play the role of calibrator and make necessary corrections. This is not just about practicality and cost savings, but also a game between collective intelligence and AI God.
From a technical perspective, small models are not entirely a disadvantage. Their small size makes it easier for them to conduct additional training for vertical fields. In these fields, although their comprehensive capabilities may not be as good as those of large models, they can perform well as expert-level tools. And a cluster of small models composed of many expert models will not necessarily lose in the competition with a single large model.
More importantly, small models can effectively promote the decentralization of deployment. When talking about decentralization, we not only refer to the decentralization of model construction and data sources, but also the decentralization of governance and the decentralization of deployment. If open source models can be easily deployed on personal laptops or even mobile phones, this will constitute AI equality. Even if the centralized service provider shuts down the service, users can still rely on local AI to continue operating. Allowing people to use AI widely without restrictions is also an important goal of decentralized AI.
AI + Web3 – Scammer or the future
There is no doubt that the decentralized AI plan that Emad is actively promoting is closely related to crypto technology. He has stated that he wants to design a Web3 protocol to integrate and realize his ideas. This is because several key elements currently lacking in the AI field - data verifiability, data ownership, large-scale coordination and incentive mechanisms, and collective governance capabilities - are exactly the areas where Web3 technology excels.
Here I want to focus on governance. Because there has never been a technology as powerful as today's artificial intelligence, and it is about to have a wide and profound impact on every corner of the world. Who should decide the future direction of this technology? Who can effectively control it? Governing artificial intelligence by the board of directors of a few companies such as OpenAI is definitely not the most effective solution. Simply setting some hard shackles on AI models by regulators may not be an effective way to deal with the challenges. Collective governance may be the real solution.
In the Web3 space, experiments in collective governance are flourishing, covering multiple levels such as data governance, application governance, network governance, and organizational governance. Although most attempts are still in the exploratory stage and have experienced many failures, this is the forefront of human governance development.
In the past five years, people in the crypto space, especially around decentralized autonomous organizations (DAOs), have tried almost all governance models in human history. The innovative structure adopted by OpenAI, where a non-profit foundation controls a for-profit company, has long been widely practiced in DAOs. In my opinion, people in the Web3 world have played a governance speedrun game in the past few years. People have re-enacted thousands of years of human governance in just a few years.
One of the most common criticisms is that most Web3 governance is just a copy of the governance model that humans have had in the past and added an on-chain vote. But history has shown us that when there is such a fast growth rate and a high talent density, something completely new will soon evolve.
A less appropriate example is Internet advertising. I remember the early years when the Internet was just emerging. When you opened a news website, a huge full-screen advertisement would suddenly appear and then slowly fade away, and the webpage was covered with dense advertising blocks, which became the iconic memory of the early Internet. Because people at that time did not know what was a more effective way to promote on the Internet, they simply moved the advertising methods on traditional media to the Internet. However, with the advancement of the scene and the deepening of people's understanding of Internet technology and culture, an efficient promotion model that had never been seen in human history soon evolved, and traditional advertising was quickly eliminated on the Internet.
In my opinion, the governance of technology will also take a similar path to Internet advertising. Blockchain technology has brought greatly enhanced coordination and governance capabilities, and collective governance solutions that have never been seen in human history will grow from it. I am full of confidence in this.
Written in the end
It is difficult to draw a conclusion when predicting the future of the AI era and examining Emad's entire conception. Obviously, his plan faces huge challenges on many levels. In addition, Emad often exaggerates in the past, and it is not easy to distinguish which of his words can be taken seriously and which are just casual remarks.
However, exploring the power structure of AI is an early, extremely complex and important topic. Although Emad and other colleagues who pursue decentralized AI may not be close to the final answer, their thoughts and attempts deserve sufficient respect and attention. Although these explorations are difficult, they are brave attempts to create the future. These efforts, no matter what the results, will become a chapter in the epic history.
Maybe one day, the world will thank Emad.
Huang Bo
JinseFinance
Xu Lin
Edmund
Joy
Coinlive 
Others
Bitcoinist