Decentralized AI: An Alternative Path for Generative AI

Centralized AI Vs. Decentralized AI

Centralized AI (CeAI) is the prevalent approach where the user (client) interacts with a service provider, such as ChatGPT (server). CeAI has many severe drawbacks that necessitate alternatives to fix those problems. Decentralized AI (DeAI) offers an alternative future. DeAI like CeAI isn’t one thing but a stack of different technologies. I’ve tried to outline some of them below.

CeAI is plagued by three major problems: poor control of data, lack of transparency, and dependence on a few companies (or a single company)

Poor control of data

CeAI typically rely on a central authority to manage and control user data. This means that user data may be vulnerable to security breaches, data leaks, and unauthorized access. Many users and organizations are understandably skeptical about entrusting personal information to ChatGPT. Furthermore, in the unlikely event of OpenAI's sudden shutdown, all of your conversations would be lost. A more likely scenario could be OpenAI's decision to de-platform you, as we have seen with other major tech companies, resulting in the loss of your data.

Lack of transparency

CeAI lack transparency due to its closed source nature, which can make it difficult for users to understand how their data is being used and by whom. Until recently, OpenAI was using your data to improve their model and currently have a 30 day data retention policy. In essence, model providers have an incentive to make their models better to remain competitive. Will they use your data to improve their service without your consent? Well, behavioural patterns show AI companies follow the philosophy of “it's easier to ask forgiveness than it is to get permission”.

The lack of transparency in CeAI also creates problems in putting checks and balances on the bias inherent in ML models. With no visibility into the data the model was trained on and its internal structure, it becomes difficult to mitigate the biases. Decisions on how biased the model will be ultimately rest with the central authority. Since these models will be used in many real-world applications (e.g., hiring), this is likely to result in unfair treatment of certain individuals or groups and perpetuate the biases.

Dependance on a few companies

In contrast to the early days of the internet, where services were built on open protocols, today's AI apps are built on top of APIs controlled by a few companies. This puts AI apps built on centralized companies in a risky position. The AI companies can update their APIs, increase their prices, or change their terms on a whim.

The concentration of power with a few companies also creates a significant risk, especially when there are conflicting interests with the user's objectives. In my opinion, the concentration of AI is a fast route to a totalitarian surveillance state/capitalism, which is a future that should concern us all.

Intention

While the open-source movement in AI is a step in the right direction, I believe there are approaches that go even further. For the past two years, we have been researching decentralized AI, which offers a solution to the problems of centralized AI. The decentralized AI stack comprises various technologies, and the level of decentralization can be selected according to the specific use case. I’ll present our vision of AI which aims to have greater user control, transparency, collaboration and ultimately more utility.

Algovera’s Vision

User Controlled Data

One of the insights we gained from our research is the significant benefits of separating data from the app. This means that the data uploaded to the platform is separate from the platform itself. Our solution to this is to use decentralized storage networks to host encrypted data. Since the users themselves control the data, they can decide who has access to it and how it can be used. The platform serves as a means to effectively query and write to decentralized storage networks.

Additionally, we are exploring the local-first paradigm to provide greater control of your data. With this approach, the data never leaves your local computer to query a server, and the program runs on your laptop. As hardware continues to become more powerful and models get smaller, it will soon be possible to run programs similar to ChatGPT on your laptop efficiently. Data privacy concerns becoming less important since the data never leaves your local computer.

Secure Collaborative Spaces

While running programs locally is useful, collaboration and data sharing between users can provide even more benefits. To support collaboration, a hybrid approach can be taken. This involves users storing and processing their data locally while also having the option to share and collaborate on data with others through cloud-based servers or decentralized storage networks. Similar to Git and Github, we intend to build a user-friendly ways to manage data collaboration.

Collaboration can occur between users or between users and developers. Initiatives like OpenMined and Zama are building tools that allow data scientists to perform machine learning while protecting privacy. These tools provide data scientists with access to data to solve problems while ensuring that users retain control over their data and privacy. The challenge lies in finding ways to enable users to control and contribute their data. Various approaches, such as marketplaces and rewards, are being tested to enable data scientists to access data. However, after researching this topic, we have found that users value the utility that can be derived from their data more than the monetary benefits. At Algovera, we prioritize giving users utility from their data while supporting secondary functionality for data scientists to use data with the users' permission.

Medium for Collective & Autonomous Intelligence

We need to ensure the protection of our own data and collaborate with others in secure spaces before introducing autonomy. AutoGPT, which aims to transform GPT-4 into an autonomous application, has recently gained a lot of interest. However, any attempt to achieve autonomy without decentralised infrastructure is not truly autonomous. I agree with Gene Kogan's proposal of the ventriloquist rule: if an application is a puppet of human control, then it's not truly autonomous. Therefore, to facilitate truly autonomous AI agents, we need to support decentralised infrastructure because autonomy is an emergent phenomenon of decentralisation.

Autonomous AI agents require a way to interact and transact value with humans. However, our current financial and banking system is unsuitable for this form of transaction. Fortunately, we have autonomous system of value transaction native to the internet - cryptocurrencies. Cryptocurrencies are the perfect match to enable transactions between humans and AI. What's even more exciting is that AI agents +/- humans can work together like a swarm/hive mind to achieve higher complexity goals and coordination.

All of this is possible, but we must build systems that don't compromise our privacy or human agency.

We are building Algovera so that AI assistants runs locally yet give the flexibility to to collaborate with others in ways that don’t comprise our privacy. This lays the foundation to build and support truly autonomous AI agents that interact with humans through decentralized infrastructure.