What is DeAI?

#ai #libertai

The use of AI to assist in day-to-day tasks has exploded in the past year. People are looking for products like ChatGPT, Claude, and Perplexity to help supplement their workflows, check their writing, or find inspiration when starting a new project. At the same time, builders are learning to fit AI into their applications to provide innovative user experiences and solve hitherto intractable problems. Recently, we’ve seen massive development and buzz around the intersection of Web3 and AI. We call this new space DeAI, or decentralized AI, and it’s exciting and groundbreaking. But what exactly is it?

To understand what DeAI is, it’s helpful to have some background on what it isn’t first. Let’s look at the aforementioned products: ChatGPT, Perplexity, and Claude. These are amazing, innovative projects in many senses. They are also what are called ‘centralized’ and ‘closed source’ projects, meaning that:

• The infrastructure that powers their features and functionality is owned or otherwise managed by a single entity or group of entities. 

• The data and software that powers these projects’ are secret.

Centralized AI products are a bit like if you could only get your favorite food (let’s say it is a panini) from a single restaurant. Let’s pretend that said restaurant was the only place in town with enough panini presses to feed everyone and was also hoarding access to the secret recipe. 

Wouldn’t it be nice if more people had access to your favorite panino? DeAI projects are like if we applied that question to AI! Notably, with DeAI:

Infrastructure runs on decentralized networks of computers that anyone can participate in to provide computation. Aleph.im, Bacalhau, and IoNet are great examples of projects that enable this.

• the data and software that powers them are developed as open-source projects. Llama and Mixtral are great examples of Open Source models that can be downloaded and run by anyone

DeAI projects effectively provide services similar to those of companies like OpenAI and Anthropic. Still, they power these services by leveraging openly available, price-competitive infrastructure and open-source models that anyone can download, modify, and deploy. It’s like you could get your favorite panino from your neighbor instead of being forced to always go to the same restaurant to satisfy your cravings. Bringing this approach to AI also benefits you, the user!

Redundancy

Sticking with our centralized panini example, what happens if the restaurant is closed (or if you get banned from the shop for espousing radical views on the redistribution of paninis)? You can’t get your panini! This situation is what we call a single point of failure. 

Centralized, closed-source products like ChatGPT also have a single point of failure: they are the only service provider of their specific combination of infrastructure and software. If, for some reason, you weren’t able to access ChatGPT from OpenAI, that would be the end of the story. OpenAI is the only service that could provide you with exactly ChatGPT, as its infrastructure is centralized and its code is secret. This is true regardless of whether OpenAI stopped existing tomorrow, if its servers went down, or if OpenAI just decided to stop providing its services to you specifically. You could try and use another product, but you’d have to find a way to research new options, migrate your personal data, and re-implement your workflows on a new platform – it would be better if we could ask someone else to provide us the equivalent service on a network where data and infrastructure is portable!

Now, think back to our decentralized-panini paradigm. We may prefer going to our neighbor’s house to source our paninis and satisfy our cravings. Now presume that said neighbor was out of town, busy, or didn’t want to make us panino today – oh no! But it’s okay! We know that our friend also can make us suitable panino at their house, which is only a little further down the road. We have what we could call here a redundant system for sourcing panino that is independent of any one fallible source.

DeAI projects are also redundant systems in the sense that, because they are built on open infrastructure and open-source software, we don’t have to rely on any one vendor for our services. If the service provider we’re using goes offline or starts denying our requests, we can move to another provider on the same network and ask them to run our freely available open-source code. Users, then, are independent of any one company for their services and can choose where they get access to AI – it’s redundant!

Privacy

DeAI projects also bring several privacy benefits to users by operating on decentralized compute networks backed by groups of independent service providers. To start understanding this, let’s bring everything back to panini. Let’s imagine there are non-trivial consequences for our privacy if we have to go to the same restaurant for all of our panini. For instance, let’s say the restaurant collects information on us, tracking when we come in, keeping a record of our orders, and watching our reactions while we eat. The restaurant may use this information to improve its panini, change staffing levels, or even sell that information to another restaurant developing a similar product. In this example, sharing information about our eating habits might not be extremely sensitive. Still, there is something about it that is troublesome – even if the outcomes are mostly innocuous in this example, we’d rather not compromise our right to privacy to access something as fundamental as panini!

Centralized AI products work like a single restaurant in this regard. As centralized purveyors of services and data, they not only exist as a single point of failure that users end up being dependent on, but they can also get enormous insight into users’ interactions with AI. Take ChatGPT, for example. When you log in to your OpenAI account, you can review your chat histories, ingest personal documents, and provide feedback on responses. For users, these look like helpful features of a well-developed chatbot product. For OpenAI, however, this is an invaluable source of personal data they can independently monetize and use to better train their closed-source products. Even as you use OpenAI products, OpenAI uses your personal data to learn your behavior and implement new products, generating value for themselves at the expense of your privacy.

DeAI products are different for several reasons. Primarily, since they operate on permissionless, decentralized computing networks, no single provider has undue insight into your interactions with AI. Instead, requests get distributed across networks as anonymous requests, providing users with increased privacy and making it harder for any single entity to associate requests with their identities. Instead of getting too familiar with any one panino maker in your neighborhood, you could strategically and anonymously go to different panino makers to help hide your preferences. In the case of ordering a panini, this might be a little extreme, but it’s great to have these options when we’re asking AI models questions about our potentially sensitive personal data!

Additionally, projects within the DeAI space aim to provide increased privacy for users by deploying innovative technologies in computing and encryption. Notably, using encrypted virtual machines and fully homomorphic encryption can make users’ interactions with AI on decentralized networks completely private by providing an encrypted environment to fulfill users’ requests and a method to completely encrypt requests themselves before providing them to AI models. These technologies enable users to interact with AI models without fear of compromising their privacy since they effectively hide the details of requests from service providers. Therefore, you can provide and ask questions about sensitive personal data to and from AI products without sharing it with service providers. If this sounds like magic, you’re probably starting to understand how unique these tools are for users! But don’t worry, all of this is just powered by an obscene amount of math!

Cost

Finally, DeAI also brings users competitive, demand-based pricing through decentralized computing networks. Traditionally, if developers or researchers wanted to source large amounts of computing for training and running new AI models, they had to either source hardware themselves or go to one of the few cloud providers that can effectively provide the cutting-edge GPUs they require, i.e., AWS, GCP, or Azure. If you are a team sufficiently large enough to raise trillions of dollars to build your supercomputers or commit to large, extensive contracts with one of the three leading cloud providers, sourcing hardware is not necessarily a pain point. But if you’re a small team or individual just wanting to provide AI services for yourself or others backed by cutting-edge chips, you face long wait lists and high prices. That means that even as computing starts to look more like a commodity in terms of demand, traditional centralized solutions impose smaller, less competitive markets that bring increased costs and deficient developer experience for teams everywhere. 

Decentralized computing networks enable computing to be bought and sold more like a commodity in open markets. They enable data centers and service providers to pool and advertise their services and prices on permissionless networks, providing an increased total supply of computing for teams everywhere and encouraging healthy competition among service providers. At the time of writing, we’re already starting to see reduced prices upwards of 50% on decentralized infrastructure networks like Akash compared to traditional cloud services providers! Users are free to set their price and requirements and match with a service provider that meets their needs, rather than a few large companies being able to dictate the terms under which teams can access computing.

When it comes to DeAI, this means that we can build AI products that allow users to pay costs that align much closer to their preferences. Rather than paying whatever monthly premium OpenAI or Anthropic imposes on users, users can choose what they want to pay (within reason) and for what services. If users have urgent AI requirements that must be fulfilled immediately, they can pay higher premiums to ensure their requests are completed. Otherwise, if they would rather wait for prices to go down, they’re welcome to delay their queries when demand decreases at a different point in their day.

So what?

This piece was a fun primer on what we mean when talking about DeAI. It was meant to convey a few of the problems innovative teams within this space are trying to solve and the benefits we can deliver to users everywhere. However, it’s worth it to take a step back and think about these issues in the broader context of the future of AI and its potential to change the world. DeAI is urgent because its purpose is to help make the development of AI technologies open, collaborative, and maximally democratic. The stakes of not building DeAI are that a technology that very well may significantly change and direct our lives is only within reach of a few large and well-funded companies and not necessarily developed in the best interests of its users. This is the paradigm under which we’ve found ourselves contending with the business models of companies like Alphabet and Meta. We know now that ownership of data and infrastructure is critical to our experience on the web. If we don’t want our data to be harvested and deployed against us to persuade us to watch stuff, buy stuff, and stay scrolling, we must try to learn from our mistakes as we build the next generation of data-driven products and AI models. We urgently need to think deeply about how we want to distribute the ownership of data and AI in a future where those things will become incredibly valuable.

DeAI is critical in helping ensure that advancements in AI align with the interests of people everywhere by making development open, infrastructure accessible, and interaction maximally private and secure. If you care about ensuring a future where these values are upheld, DeAI technologies are worth investing your attention in.

Thanks for reading!