Make A Decentralized Internet, With AI: NANDA Is Coming

Many experts who speak at modern conferences and write papers on large language models (LLMs) will agree that we need to understand why things work in some way and how to leverage them.

Benjamin Franklin made history by putting the metallic system on the rope and thinking about why Lightning works like this.

He was a member of the philosophy circle at that time. We need something similar.

With that in mind, I was impressed by Abhishek Singh’s speech on “The Future of Chaos, Coordination and Agent AI” at an April event.

Decentralized network

Singh is one of many researchers working on the basic concept of decentralized networking and how this relates to the new technologies we have.

In his speech, he talked about the “triad dilemma” related to continuity, heterogeneity and scalability (if this is not a quantitative language, I don’t know what it is).

Speaking of Chaos Theory 2.0, he talked about the connection between decentralized networks and algorithms as the main goal of what he called the "emergence phenomenon" of AI agents.

"One way of thinking about how these two mental models fit is: the way we solve intelligence now is through a large system located in a large tech company and able to do all tasks at the same time. And another view, more from a decentralized perspective, is that many cerebellars interact. A single cerebellar is not powerful enough, but combined (they become powerful)."

When I hear this, I always mention Marvin Minsky’s “Society of the Mind” principle, partly because he is my hero, partly because I think it is crucial to the problem we are working on.

Decentralization and its challenges

Singh also pointed out the challenges of implementing a decentralized program.

Some of them involve privacy, verification and orchestration.

Others are related to how we design the network we are discussing with user experiences in the crowd or engineering.

Singh mentioned that complex models and large-scale collaboration contain inherent problems that need to be solved.

Others are related to motivation.

https://www.php.cn/link/c63858b17de9d0649b59be0c57201b9d

“Many individuals and organizations host open datasets and volunteer computing resources for altruistic reasons,” Singh wrote in the aforementioned paper. "The clear incentive structures that contribute to decentralized systems may squeeze out altruism and reduce participation by those who wish to help rather than profit. External rewards have been observed to cover intrinsic motivations over time. When we compensate people for activities that once voluntarily, they tend to lose their intrinsic interest in those activities. Therefore, designing effective incentive plans requires careful consideration of community norms, social motivation, and human psychology. The goal should be to complement existing altruism without replacing it. A hybrid approach to combining economic and reputation-based incentives may help."

Therefore, all of this must be taken into account when making these design choices.

In its own words

I admit here that the paper I got on about decentralized AI, Singh is one of the authors, is quite intensive and discusses in detail how decentralized AI works and the challenges it faces.

So I put it into Google Notebook to get some sort of human response from two invisible characters in the Generative Conversation Tool.

In the first few minutes, they reviewed some basic concepts, described a single data center as a huge vulnerability, and talked about the “severe reminder” (they used this phrase more than once) what happens when they are breached.

Then there is data ownership.

Then they had a conversation about the definition of decentralized AI (for fun, you can compare it with the definition of Singh):

“At its core, decentralized AI is about enabling different entities, companies, individuals, and even our devices to collaborate on AI development and deployment,” the female voice said. “But the key difference is that this kind of cooperation does not require a single central authority to control everything.”

"Yes," the male voice replied. "No big boss."

"That's exactly that," his companion said. "Imagine different parties with their own distributed resources, their own data, their own computing power, working together even if they don't have complete trust with each other, or may not want to hand over control to a central player. So, it's not so much a huge AI brain on some central server, it's a smaller, interconnected working network."

They continued like this for a while. It seems to me that the "people" of Notebook LM seem to be heading in a simplistic direction. So I went back to the paper itself and looked at some of the larger chunks of what it presented.

Here is a paragraph in the conclusion:

“This article illuminates the advantages, use cases, and challenges of decentralized AI. We believe that decentralized AI development can unlock previously inaccessible data and computing resources, enabling AI systems to flourish in data-sensitive areas such as healthcare. We propose a self-organizing perspective and argue that five key components are needed to enable self-organization among decentralized entities: privacy, verifiability, motivation, orchestration, and crowd user experience. This self-organizing approach addresses several limitations of the current centralized paradigm, which heavily relies on integration and trust in a few dominant entities. … We believe that decentralized AI has the potential to empower individuals, catalyze innovation, and shape a future where AI benefits society as a whole.”

You can also view various Venn diagrams showing how to solve certain problems inherent in traditional AI systems.

Decentralized vectors

Now, one thing I noticed in Singh’s speech—at the end, he mentioned an abbreviation that might be very critical to decentralized AI.

It is called NANDA, a networked proxy and decentralized AI, which is being studied by a team including Singh and my colleague Ramesh Raska at MIT.

For complete transparency, they mentioned me as a collaborator on the website in a less direct way.

But those who work on this work have front seats and can see what it will look like when we build a new decentralized internet and leverage the power of AI.

This is something we should focus on as we move forward in 2025 and we will start to see more performance in the actual capabilities of AI.

The above is the detailed content of Make A Decentralized Internet, With AI: NANDA Is Coming. For more information, please follow other related articles on the PHP Chinese website!