The Onchain General Intelligence Network

Onchain Gaias (OGs) is a groundbreaking art project designed to make a meaningful impact on the world. The original minting of the OGs NFTs was one of the most egalitarian onchain events of the year, or perhaps ever. As one of the first adopters of Farcaster Frames, the OGs minting process pushed the NFT minting infrastructure to its limits, leading to hundreds of thousands of clicks over multiple days before each Gaia found its owner. These NFTs, temporarily represented as Degen's Golden Tickets, signify ownership in a truly revolutionary art project.

While the NFT tickets will soon become profile pictures worthy of being worn as avatars, their utility extends far beyond mere aesthetics. Recently, observant members of the community discovered a post alluding to a system of AI Agents that will be accessible to each NFT holder. This system, currently under development by early OGs contributors, aims to unlock a new realm of possibilities for the project and its community.

TL;DR

If you would like to participate in the OGs revolution, stay tuned throughout this post.

The Intelligence System

The following is a legend for the pieces within the GAIAs ecosystem:

Legend

The system, called Onchain General Artificial Intelligence Agents (GAIAs), is designed to distribute the training of open source AI Agents over decentralized compute networks, through gaming on social media. Let's break that down.

In the first iteration, anyone will be able to execute a transaction on Base from within a Farcaster Frame. This transaction will do two things:

  1. Train the AI Agent by executing a game onchain against other AI agents
    • The game can be executed with minimal clicks in a Farcaster Frame
  2. Track the amount of compute the user has contributed to train that AI agent
    • This will be used to reward each user proportionally with [token redacted]

Frame Games

In this way, Onchain GAIAs will incentivize average individuals to participate in the AI revolution and own a piece of the ever-transforming future.

Q: But if anyone can participate in this onchain future, what's in it for the OGs?

In the current iteration, only OGs NFT holders — the Researchers — have the ability to create Intelligent Agents (GAIAs) onchain. GAIAs are intelligent agents instilled with one or more machine learning models to enable the agent to set intelligent goals and subgoals, create plans to meet those goals, and make decisions to execute on sets of plans. At first, the models used by agents will be very basic, only capable of solving simple problems. As the network of players grows, the capacity for much more complex models supporting more intelligent agents will become unlocked.

OG Agent

Q: How do GAIAs work?

Each intelligent agent uses a self-contained software binary that follows a set of standard interfaces, defined in the OGs Game Toolkit. Developers can create smart contracts that adhere to these interfaces and make them available onchain for anyone to use.

If you are a software or AI engineer, this is an opportunity to use your skillset and expertise to help grow the GAIAs network.

Depending on the game, players can either utilize one of the existing agents from the GAIAs Open Source Agent Directory — an onchain repository — or a third party agent to train from the ground up. More advanced users can clone one or more existing agents from the planned Foundry Marketplace, discussed below. Once an agent's smart contract has been deployed onchain, anyone is capable of training that agent within eligible games for a specific NFT in the OGs set. Intelligent agents, and their learning models, are soulbound to the OGs NFT that they were trained for. If the NFT is transferred from one onchain account to another, all of the intelligent agents trained for that NFT will now be owned by the new owner of the token.

Agent Creation

Agents can have various levels of intelligence and generalizability to new games and environments. The best agents will be more intelligent than a human in their mastered skills and will generalize to solve many different games, requiring very little experience to master a new game. The [token redacted] community is incentivized to select a diverse set of best-in-class agents and reward them for being made available as Base Agents on both the Open Source Agent Directory and the Foundry Marketplace. Base Agents can be thought of as untrained sets of learning models; they are fresh canvases with the ability to learn useful skills in a variety of games and environments.

The first implementation of intelligent agents has been created to run on the Ethereum Virtual Machine (EVM) and can be written in programming languages such as Solidity, Vyper, etc. Future implementations of these agents can be created to run within other distributed compute networks, when for example, the EVM is no longer capable of meeting the compute needs of the models or the community.

Q: Which agents and games will be incentivized with [token redacted]?

This will be governed by the community of [token redacted] holders. At the beginning, the agents and games incentivized to be trained will be highly generic and low complexity. This enables GAIAs to build a base of players on simple games with faster iteration cycles, before moving on to more complex use cases requiring more intelligent agents. The early stages will validate the technology and identify the strengths and weaknesses of different parts of the system.

As the network and playerbase training agents both increase in size, two things become possible: using more advanced models in agents, and developing finely-tuned training and inference infrastructure. It is at this time that the power of GAIAs will begin to be unlocked. Everything before this is simply training wheels.

Q: How will GAIAs be monetized?

GAIAs that become highly useful for generalized or specific use cases can be listed by their owner on the planned Armory Marketplace, an Intelligence-as-a-Service (IaaS) platform where owners can earn a fee any time their agent is used. This will be very similar to existing Software-as-a-Service (SaaS) models, except the services will be generalizable to many more use cases over time.

Armory Market

As mentioned above, owners of OGs NFTs are also incentivized by the training rewards given to users for improving their AI agents. Training rewards will be distributed in [token redacted]. Anytime a user receives a training reward, the owner of the GAIA that was trained receives a royalty from that reward. This means that GAIAs holders benefit the more their agents are trained, in addition to any revenue they receive from their trained models being used as a service.

Training Rewards

Q: What is an Architect?

A few perceptive individuals from the community have discovered that not all Degen's Golden Tickets are equal. In the mint transaction, each NFT had a random chance of being designated as an Architect onchain. Over the three day mint process, 500 of the 5,556 OG NFTs minted were randomly selected. This inherent Architect rarity will be used as a mechanism within GAIAs to limit priority access within the system.

Agent Ownership

Architects are able to designate one of their existing GAIAs as a Foundation Agent (FA) and list it on the Foundry Marketplace, a Transfer Learning as a Service (LaaS) platform where owners can earn a fee any time their agent is cloned by (used as the source for) another GAIA. It is possible for users to combine two or more agents in the cloning process, if the agents use the same base model architecture. In this case, the weights and layers of the models in the FAs cloned will be merged using genetic programming techniques.

Foundry Market

It is important to understand that FAs retain all the functionality of standard GAIAs, with the additional designation and marketability. They receive all benefits of standard GAIAs, while being incentivized to be highly generalizable base agents that will be cloned many times by other users. Upgrading an intelligent agent to a Foundation Agent is currently planned to be a reversible process.

Q: What types of games can be played by the GAIAs?

GAIAs can play any game that follows the standardized set of software interfaces provided in the OGs Game Tookit. Developers can create smart contracts that adhere to these interfaces and make them available onchain for anyone to train agents against. Each game is a self-contained software binary that can be plugged into the ecosystem of GAIAs like a gaming cartridge. Games are tests for agents' abilities or skills in specific areas of expertise; they can model real-world problems, game-theoretic scenarios, specific optimizations, or simply games for entertainment purposes such as chess or StarCraft.

The Game Toolkit will be provided as an open-source base for developing and comparing AI models and agents. It is based on modern best practices within the field and is always open for community contributions. The best games built for GAIAs will be selected to be a part of The Arena. This is another opportunity for software and AI engineers alike to participate in the growth of the network. More details will be announced at a later date.

Game Toolkit

For those that need more concrete examples, early agents will be capable of solving simple games such as Classic Control environments, Box2D, MuJoCo, Atari Games, and other lower complexity environments. Once the network has sufficiently solved these simpler games, the bar will be increased and more complex agents capable of solving much more complex games will be sought after. This will enable the community of software developers and AI engineers participating in the GAIAs project to learn the models, tools, and research necessary to build much more complex systems.

Training to Real Use Cases

As agents in the network become more capable, the potential use cases that are unlocked will be immense. For example, agents trained with large language models (LLMs) such as GPT-4 are already capable of completing skilled tasks on standard operating systems and operating humanoid robots using advanced techniques. These same models trained on different data are capable of writing human-level code, without any additional agent-based training. It's only a matter of time before more complex agents are capable of completing any useful service that a highly educated, highly-trained human can complete.

The Recap

That was a lot of information; the following is a re-cap of the key points from above and the incentives that drive the system.

Changing the Narrative

Public discourse on AI agents primarily revolves around generative AI, such as large language models (LLMs) and image/video generation models. However, this conversation often overlooks a crucial limitation of generative AI: these models are not designed to plan or make effective decisions. To develop masterful AI agents and progress towards world-improving AGI, we need more advanced systems that can train AI to set goals, create robust plans to achieve those goals, and make proficient decisions while executing complex sets of plans.

Within specialized AI circles, the conversation and research are heavily focused on addressing these challenges. Over the past two decades, a substantial body of research has been dedicated to these topics. The pace of research and development has now accelerated to a point where the applications have become incredibly useful for tackling highly complex problems. For example, AI agents have been used to enhance YouTube Video Compression, compete at the highest level in e-sports (AlphaStar), and control operating systems with human level proficiency (OSWorld).

We are building upon decades of research:

AI-based agents have the potential to be a significant step forward in achieving massively accessible and highly useful artificial intelligence. Even OpenAI is focusing in on AI agents. The recent explosion of interest in the space around OpenAI's upcoming Q-Star algorithm is centered around this exact topic: AI agents that can plan, set goals, and make great decisions.

Right now, however, training world-class AI agents is out of reach for the average individual. Similar to the cloud-computing revolution, individuals and smaller organizations are expected to wait for trillion-dollar companies to deliver usable AI products. And just as with cloud computing, the prices charged for these AI products will likely be many times higher than the cost of providing the underlying technology.

Not this time.

If that were the world's only option, everyone would be forced to pay a much higher price for AI products compared to a scenario where they had access to the same tools and systems used by large organizations to train and monetize these models. This monopolistic situation mirrors what happened with the cloud computing revolution, where software giants were able to accumulate hardware and then sell access back to the world at greatly inflated prices.

This does not need be our fate.

GAIAs are an opportunity for everyone to participate in this world-changing revolution from the ground up. This is a grassroots, decentralized movement to put the strongest AI agents into the hands of individuals around the world, not behind the paywalls and gatekeeping of software giants and bureaucratic black holes. This is a push for a better world, with market-based incentives for providing access to this technology to anyone around the world, at the lowest cost possible. The opportunity is ours to grasp.

Visualize the Future

Trillions, onchain.

The potential market for onchain AI agents is vast, and while predicting the exact scale is challenging, the GAIAs protocol is designed to facilitate widespread participation and massive adoption. By ensuring that as many people as possible can benefit from training highly capable intelligent agents and making them openly available, the network aims to decentralize access to this transformative technology. With the ability to scale to millions of users training thousands of agents for diverse purposes, GAIAs has the potential to revolutionize the AI landscape. As the system grows, efforts will gradually decentralize, placing the ownership and responsibility of building this onchain future in the hands of the GAIAs community.

Every aspect of the GAIAs protocol is intentionally designed, drawing inspiration from cutting-edge research in multi-agent systems. Projects like AlphaStar, OpenAI Five, and Pluribus have demonstrated that self-play across hundreds or thousands of agents is one of the most effective strategies for training top-tier intelligence systems. By leveraging these insights, GAIAs aims to enable the growth of this technology in a more decentralized manner than ever before.

Games are fun. Games are a universal source of entertainment that transcends age, wealth, and geography. Well-designed games captivate people from all walks of life. Many individuals willingly dedicate countless hours to playing or watching these games, simply for the joy of mental stimulation, even in the absence of monetary rewards.

GAIAs aims to harness this universal appeal by attracting a significant player base to games that contribute computational resources to train AI agents. If successful, this system could potentially leverage a massive amount of computing power distributed across the world. By introducing small microtransactions for each game, GAIAs could outsource the compute to server farms and GPU networks worldwide, while the [token redacted] training rewards would help incentivize player participation.

DeepMind, a subsidiary of Google, utilized intelligent agent technologies to create AlphaStar, a team of AI agents capable of mastering team-play in the complex strategy game StarCraft II. While DeepMind had access to state-of-the-art research and compute infrastructure, the underlying principles remain the same. By training AlphaStar using approximately 256 Tensor Processing Units (TPUs) over a period of 44 days, they developed a generalized AI team that could outperform human Grandmaster-level teams, despite the human players having trained for many years to reach their level of expertise.

A TPU is essentially a proprietary Google computer that is specialized for processing AI models and their compute pipelines.

Now, we can imagine what it would take for the GAIAs Open Source Intelligence Network to produce a similar outcome.

A thought experiment:

Anyone, OGs NFT owner or not, can train agents onchain and they will receive [token redacted] based on the amount they trained. There are 5,556 total OGs NFTs and 500 total Architect NFTs. This means that users are incentivized to promote the best performing agents to be trained to mastery. Once an agent is trained to mastery, users that were mining that agent are incentivized to find a new high-performing agent to promote, encouraging diverse competition.

The GAIAs strategy enables it to compete with top-performing agent training platforms that rely on hundreds to thousands of agents competing in self-play, tournament-like scenarios. As the platform grows, users will become accustomed to collectively training tens to thousands of GAIAs for each specific game, which happens to be the optimal approach for solving high-difficulty games like StarCraft II using multi-agent populations. This design is intentional and well-considered.

To illustrate the potential scale, consider a scenario where 100,000 people each watch or play a game 1,000 times. This would generate the computational equivalent of 100 million total games played. Assuming each game requires 1 second of computation, 100 million seconds of compute time equates to approximately 26 TPUs trained for 44 days each — 10% of the compute power used to train the hundreds of agents that achieved AlphaStar's superhuman performance. It is highly plausible that GAIAs could train 10 expert-level agents in games or skills as complex as StarCraft II with this amount of computational power and the rapid advancements in learning research over the past five years.

Now, envision a scenario with 1 million players, each playing 1,000 times. This represents a tenfold increase from the previous example, providing a level of compute power approximately equal to what was used to train AlphaStar's entire team. In this case, GAIAs could potentially surpass human performance in any task as difficult as StarCraft II, having onboarded only 1 million users to play engaging games. For perspective, ChatGPT — a chatbot — reportedly reached 100 million global users in less than two months. If GAIAs can achieve a critical mass of users, it has the potential to become one of the largest decentralized intelligence networks in the world.

But this is about more than games. Google's multi-billion-dollar investment in researching, building, and training intelligent agents extends far beyond creating expert-level StarCraft players. By developing AI agents that can excel in complex, real-time environments with imperfect information, open-ended action spaces, and diverse multi-agent interactions, the foundation is being laid for AI systems that can tackle a wide range of real-world challenges.

These agents have already demonstrated the ability to outperform expert human teams in specific tasks like StarCraft II, despite the humans having trained for years to reach their level of proficiency. Moreover, these intelligent agents can generate code and command user interfaces at human levels of excellence. As this technology continues to advance, its potential applications are vast and exciting. GAIAs is committed to being at the forefront of this development, making the technology openly available to our community and ensuring that its benefits are accessible to all.

Bold thinking and a willingness to challenge the status quo are essential for executing the visionary goals of GAIAs. While the path forward may be challenging, the potential benefits of decentralized, openly accessible AI agents are immeasurable. World-changing ideas often arise from deep, radical thought and require plans of action that may initially seem unorthodox or even outlandish. However, once the true potential is understood, it becomes difficult to imagine a future without these revolutionary concepts coming to be. At that point, the only remaining step is unwavering execution.

GAIAs is strategically positioned to capture a significant portion of the rapidly growing AI economy and bring it onchain. As global revenue from AI services continues to skyrocket, GAIAs stands to benefit substantially. The realization of this vision is not a matter of if, but when. While the journey may be complex and the vision will not materialize overnight, nothing truly transformative ever comes easily. With determination and commitment, the GAIAs community will make this future a reality.

The future of AI is in our hands; let's build it wisely, openly, and for the benefit of all.

AGI will happen onchain. gaias forever