Decentralizing AI for Democratic Empowerment
AGI is going to come from a private company, the government simply doesn’t have enough video cards. This is expected under capitalism, but what are the consequences of having so much power concentrated in a for-profit corporation?
How Much Could One AI Cost?
The first thing to understand is just how expensive it is to train these models and what that means for the landscape of AI development. Ballpark numbers show jaw-dropping amounts of money being required to train frontier AI models, $80M for GPT-4 and $190M for Gemini Ultra to give a sense of scale.
This means that advancements are concentrated in the hands of a few industry giants, so it’s worth getting to know them before continuing. Of note is where they got their funding, but of special interest is where they got the actual hardware to train their models and run their operations since that’s still a major bottleneck.
OpenAI: Funded by Microsoft (though possibly not for long), OpenAI is the company behind ChatGPT. They have a close partnership with nVidia for hardware.
Google: Richer than God. Developed its own custom AI chips called Tensor Processing Units (TPUs).
Meta: Also has more money than they know what to do with. Has a huge portfolio of compute power by sheer luck of the Metaverse being a failure.
Anthropic: Funded by VC. Dependent on Google and to a lesser extent Amazon for cloud services.
Setting Money on Fire
Additionally, beyond training the models, there are also the costs of running them. Most of these services operate at a loss, working as an advertisement for enterprise customers. Massive amounts of money are funneled in, as if normal pressure to recoup investments is bypassed by the hopes of the big breakthrough that will capture the entire market.
AI Oligarchy
In 2023, the United States saw AI investments reach $67.2 billion, nearly 8.7 times more than China, the next highest investor. In that same year, the U.S. produced 70% of new models, reinforcing this “AI oligarchy”.
What happens to the rest of the world? As AI continues to become more powerful, the gap between the haves and have nots will only grow. How many video cards does Papau New Guinea have?
Barriers to Entry
The costs have continued to skyrocket, with each new model costing more than the last to train and run. Lack of access to specialized GPUs and hardware makes it nearly impossible for new entrants to compete. Even if the money hurdles could be overcome, the proprietary datasets and closed-source algorithms prevent building on top of existing work.
Normally academic research would pace the private sector but due to the astronomical costs involved, it’s as if it’s seen as a lost cause since public investment in AI research and infrastructure lags so far behind that it’s barely worth mentioning.
Impact of an AI Monopoly
In AI, centralization, and the One Ring, Owen Cotton-Barratt lays out what a single entity gaining control over AGI actually means: vast authority over global systems and economies, insulated from the accountability structures that normally pressure organizations toward ethical behavior. Without public oversight, legitimacy is asserted rather than earned. It’s easy to imagine a centralized project rationalizing control over public benefit — they already do it at smaller scales.
AI’s Perspective
There’s a lack of attention to grassroots movements and community-driven AI development. Empowering local communities to participate in AI creation can lead to more relevant and ethical applications, fostering innovation that reflects diverse needs and values.
ChatGPT o1-preview 2024-10-29
The path forward involves open-source projects like Meta Llama, Hugging Face, and EleutherAI democratizing access to models; decentralized infrastructure efforts from Ocean Protocol and OpenMined; and policy pressure from the AI Now Institute and Algorithmic Justice League to prevent monopolistic control. Community-driven development isn’t just a nice idea — it’s the only counterweight to a system optimized to exclude it.
