Centralized vs. Decentralized AI

Originally a thread on X/Twitter:

Imagine a world where AI makes our jobs easier, frees us from mundane tasks, and effectively democratizes superintelligence. In that same world, AI tools might whisper in your ear to recommend the job you never knew existed, the perfect partner based on complex compatibility algorithms, and even the news you “should” read. This future is part utopian and part dystopian with it being unclear what the “net impact” will be on society.

So it shouldn’t come as a surprise that a philosophical movement is brewing that believes that the development of AI technology should be slowed down or even stopped completely. But there’s another movement that believes that many of the dystopian problems that AI could create are problems of Centralization, not of the core AI technology.

The Centralized Powerhouse

Big tech companies like Microsoft, Google and Facebook hold immense power. They can afford to buy up the vast majority of GPU availability, they control massive data centers, they can pay for/assemble the largest data sets, and as a result they can churn out ever-more sophisticated AI models.

But, Centralized AI (CAI) raises serious concerns about privacy. Imagine a world where every search, every purchase, every social media post is fed into a giant AI maw, creating an unnervingly detailed profile of your life. Worse yet, these “black box” algorithms might be biased or trained to operate in ways that reflect the values of their CAI masters.

From a Government standpoint, CAI offers a tempting solution for national security and social control. Governments could leverage centralized AI for real-time crime prediction and prevention, or even streamline tax collection and social services. However, the potential for misuse is high. Imagine a world where governments can track your every move, censor information they deem undesirable, or manipulate public opinion through targeted AI campaigns. This raises serious concerns about civil liberties and democratic control.

Is Decentralization A Fix?

Proponents of Decentralized AI (DAI) believe it’s a compelling alternative. Imagine a network of independent computers working together, sharing processing power and data securely. This could pave the way for transparency in how the models work and what data is being used to train them. It could level the playing field and allow conglomerations of smaller players to innovate and flourish.

However, DAI is still in its infancy. Distributing computing power will almost definitionally be less efficient, potentially slowing down development and producing less effective model. Security risks also become a concern due to decentralized networks having unique weak points that could be attacked. And coordinating a multitude of stakeholders in a DAI future could be a bureaucratic nightmare, especially if Regulatory requirements require very specific oversight.

From a Government standpoint, the lack of control inherent in DAI is a major concern. Regulating a decentralized system would be a monumental task, potentially hindering efforts to combat terrorism, cybercrime, or tax evasion. Governments might also struggle to collect taxes or implement social programs in a DAI world where data and resources are not centralized.

It’s easy to Future Cast that CAI is the Frontrunner

Let’s be honest, corporations are profit-driven entities. CAI offers a clear path to monetization from these corporations and that will be a key driver in adoption and reinvestment.

DAI, on the other hand, has potential but also requires a paradigm shift. DAI is facing headwinds because it will require coordination of many participants and rely on the widespread adoption of complex technologies like blockchains and incentive structures/tokenomics. These technologies are currently plagued by fraud and regulatory uncertainty and until these issues are addressed, corporations are unlikely to invest heavily in a technology with such a high barrier to entry.

It’s also easy to Future Cast that DAI is likely to see experimentation happen first for consumer apps

Similar to how Duck Duck Go found product-market-fit in a customer segment that had privacy concerns, DAI could find early adoption for AI use cases that center around privacy or censorship resistance. Many crypto/blockchain teams have spun up in the space and it’s likely that the DAI movement will test and learn delivering AI applications to tech-forward consumers who are already comfortable with the crypto/blockchain space.

So while the future is unknowable, it’s clear that there’s a massive global resource allocation of people and money flowing into the AI ecosystem that’s funding every form and fashion of experiment that’s imaginable!