TLDR

• Altered State Machine is a protocol that enables the creation, ownership and tradability of AI agents through the use of NFTs.
• The AI agents supported by ASM can span across various categories. They can be game specific (NPC characters), DeFi related (trading bots) or have other use cases in virtual worlds (e.g. virtual assistants).
• Each Agent is represented by a mix of a Brain, which serves as the central component of the Agent; Memories, which contain the behavioral strategies the Agent has learnt (via training an AI model); and Forms, which encode the physical characteristics like what it looks like, and how it operates.
• Altered State machine provides the framework for hardware owners to plan, coordinate, and monetize GPU time rented to “Gyms”. Similar to miners, gyms are networked GPU cloud compute providers that decide to perform work training algorithms for applications using ASM agents.
• ASM is also enabling a new category of crypto-games, where NFT AI bots play, interact and compete with one another.
• AIFA (Artificial Intelligence Football Association), is a 4x4 football game played autonomously by ASM All-Star agents, the first example introduced by the project.
• ASTO is the native token of the protocol and is used for governance, AI Agent training, minting Agent NFTs, in-game transactions, asset creation, and more.
• All-Stars are a collection of 40,000 football playing mini-characters. They’re the first NFTs to receive Non Fungible Intelligence from ASM Brains.
• In ASM-based play-to-earn games, the scholarship and reward-sharing model (seen in games such as Axie)is not necessary as the human class has been automated by autonomous AI agents that can play the game independently.
• In ASM-based games you could have “companion agents” that recognize you as a unique player and adapt to your style of play.
• In ASM-based games the development and ownership of NPCs can be transferred from the game developers to other participants in the game. So many aspects of the game will be built by players for players.
• Models such as DALL-E 2 could potentially be tokenized in the future using protocols such as ASM.
• Tokenizing ML models through protocols such as ASM could potentially also be useful in fields such as Federated Learning.

Key Terms

Introduction

The rise of NFTs has attracted a significant amount of attention to the crypto space. Due to their potential use cases in a wide range of industries, they have allured individuals from various sectors including artists, game developers, musicians and even nightclub owners. At the same time, NFTs also attract the most criticism in the space as many of their use cases are not fully understood.

There are various categories of NFTs such as:

• PFPs
• Virtual land
• Avatars
• DeFi related (LP positions) etc.

One of the latest emergent NFT types providing a new functionality has to do with AI-infused NFTs. This is a nascent sector but one with tremendous opportunities for growth. Altered state Machine is a project built to support such NFTs.

Overview

Altered State Machine is a protocol which enables the creation, ownership and tradability of AI agents through the use of NFTs. It is a foundational layer which combines both AI and blockchain tech to facilitate the “tokenization” of AI agents.

• The AI agents supported by ASM can span across various categories. They can be game specific (NPC characters), DeFi-related (trading bots) or have uses in virtual worlds (virtual assistants). There are no restrictions put on the utility of the agents. Additionally, other third-party projects such as Meebits or Bored Apes for example, could use the ASM protocol to infuse their NFTs with “intelligence” and turn them into ASM agents. Each Agent is represented by a mix of a Brain, which serves as the central component of the Agent; Memories, which contain the behavioral strategies the Agent has learnt (via training an AI model); and Forms, which encode the physical characteristics like what it looks like, and how it operates.

• Altered State machine provides the framework for hardware owners to plan, coordinate, and monetize GPU time rented to “Gyms”. Similar to miners, gyms are networked GPU cloud compute providers that decide to perform work training algorithms for applications using ASM agents (eg: a play-to-earn game, a trading bot etc). The ASTO token, which is the native token of the project, is required as payment for training.

• As ASM provides a means to tokenize AI agents, it has also enabled an economic structure to buy, sell, loan and borrow AI agents. Since each agent can be trained in a Gym, different agents will have varying degrees of usefulness and ability (even those within the same domain and/or application). As a result, it would be useful to own agents (individually or jointly) and to trade them on NFT marketplaces in a decentralized way.

• Additionally, Altered State Machine also aims to enable the composability and interoperability of agents between various applications in order to enable owners to port the “intelligence” of their agents into various applications, games and virtual worlds without any restriction.

• ASM is also enabling a new category of crypto-games, where NFT AI bots play, interact and compete with each other. AIFA (Artificial Intelligence Football Association), which is a 4x4 football game played autonomously by ASM All-Star agents, is the first example introduced by the project.

AI Agents

In Altered State Machine, each AI agent is composed of three different components:

1. A Brain, which is the core of the Agent
2. A Form, which encodes physical attributes (what it looks like, and how it operates)
3. Memories, stored in the Brain's Memory Tree, which encode behavioral strategies the Agent has learned (AI model training)

It is important to note that these three different components are not always required to create an ASM Agent. Different use cases will require different components, for example, a trading bot might not necessarily need to have a form. The structure was introduced in order to allow for maximum flexibility in the creation process of an agent.

Breakdown

• The way your character looks and functions is visually represented by a Form. Forms are the way the Agent's Brain interacts with the World. Forms provide an agent with its physical form and characteristics, allowing it to function in its environment. Forms can be of 3D or 2D shape, depending on the World they operate in. In non-physical settings, the Form is often a distinctive, tradeable skin that aids in identifying or branding the Agent. In environments such as games and virtual worlds however, the forms will impact the Brain’s base values (more below) and will specify how the agent operates (e.g. in a football match environment an agent will be able to kick a ball).
• In situations where an NFT from an existing collection wants to use ASM and gain “intelligence”, then the Form of the agent will be the same as its original appearance.

• The Brain NFT is the core element of the ASM infrastructure and has the same general structure anywhere in the ASM universe.
• It has two primary functions:
  • To encode a randomized base attribute set (aka Genome Matrix) - think of it as the agent’s DNA
  • To store memories (in a Memory Tree)

The Genome matrix is randomly generated and provides the inherent skills the brain is created with. It is visualized as a heat map where different regions represent different traits. The Genome of a Brain is permanent and doesn't change even with training and experience. Each environment will interpret the genome matrix differently depending on the use case of the environment.

The attributes in a Genome Matrix can correspond to stats like strength, speed and size (in gaming), risk tolerance and randomness (in finance) or beats per minute (in a metaverse musical experience).

The functional characteristics created by the Brain attributes (Genome Matrix) are expressed in two main ways. One depends on the expressions defined by the World the Agent operates in. A World in this case is any 3rd-party application that connects with the ASM platform and enables the use of Agents in its experience. The second is dependent on the Form that the Brain is attached to. For instance, a Form in the setting of a fighting game would probably specify the Agent's size, strength or speed. A Form may also attach collectible items to a Brain, which could impact the Form modifiers (e.g a shield will grant extra protection to damage).

The Expression mapping for a World is depicted in blue circles above. The Expression is a look-up table which allows the Form to express the attributes needed to operate in that world (e.g. a football game world would express speed, strength, and intelligence from defined areas of the attribute set). As a result, Expressions would frequently inherit characteristics from their host World.

In most cases, a Form's modifiers and the World's expressions will be defined by the world owner. However, the ASM DAO and community will aggregate and publish generic Expression standards and Form modifiers, in order to facilitate interoperability between Worlds. By pushing comparable worlds to adopt comparable standards, interoperability becomes much easier to implement.

It is also important to add that there will be an Expression for every Environment if a World consists of many diverse use cases (for instance, a virtual world with action games and non-combat conversation experiences).

Lastly, Brains also contain a Memory Tree - an immutable record of the memories created and owned by that Brain.

• Memories are used to store the results of agent training. Memories are specific to the Agent's Genome and the world they were formed in. Each agent is capable of creating hundreds of Memories in various worlds. Since memories encode Machine Learning strategies, they are usually specific to the World the Agent operates in. Memories store the outputs of Gym Training as machine learning models, in a format like ONXX (ONNX is an open format built to represent machine learning models. ONNX defines a common set of operators - the building blocks of machine learning and deep learning models - and a common file format to enable AI developers to use models with a variety of frameworks, tools, runtimes, and compilers).

It is important to highlight that all the data relating to the Agent is not on-chain, just like for the majority of NFTs today. The ML model data that make up Memories as well as the artwork connected to the Form of the agent are stored in IPFS. By connecting IPFS and the NFTs, the ASM network makes sure that each Agent has ownership.

Gyms

Through training, agents can develop, evolve, and become better. Training takes place in a gym with a machine learning (ML) model trainer created specifically for each world to boost Agent performance. The ML model makes use of a neural network and is influenced by the attributes of an Agent's distinct genome to produce results that are unique to that Agent. The ML model output is encoded in a file for each Agent. These are Memories, and each time the Agent trains in the gym, new ones are created in the form of NFTs.

Training at the Gym is a process similar to mining as GPUs are used to provide computational power to train agents. GPU owners are incentivized to host Gyms as they are paid in ASTO token for their contributions. The Gym Host keeps the bulk of the rewards, with a minor portion going to the ASM DAO.

Examples of AI Agents

As mentioned above different agents will require different compositions of the three core components that make up an Agent. The examples below provide a comparison of two different agent types, providing differing use cases (a gaming character and a DeFi bot).

In the case of a game character agent, both the Form (how the character appears) and the Brain (genome matrix and memory tree) are equally significant. Both of them, as well as the Environment the agent operates in would impact the expression of the attributes.

In the case of a Defi trading bot, however, the Form has no real use case besides acting as a visual identifier (or brand), as a result it is entirely optional. Memories on the other hand are significantly more important as they will guide the strategies for placing trades, therefore the profitability of the trading bot will be directly impacted by them.

Using an ASM agent as a Defi trading bot gives it more transparency and enables its ownership. This could be beneficial for DeFi-related DAOs and protocols for example, as the community as a whole will be able to govern and own the strategies deployed by the agent. By looking up its characteristics (eg. risk tolerance can be a trait on the NFT) it could also be easier to determine the parameters of a particular trading bot instead of just trusting the team deploying the strategies. The performance of each specific strategy could also be tracked and analyzed as it will be deployed as an NFT.

NFTs

AIFA All-Stars

(OpenSea/Nansen) All-Stars are a collection of 40,000 football-playing mini characters. They’re the first NFTs to receive Non-Fungible Intelligence from ASM Brains.

Brains

(OpenSea/Nansen) ASM Genesis Brains are unique AIs owned via NFTs. Brains are capable of learning and evolving. They’re interoperable across different forms and worlds. To train up their football skills and play AIFA, each All-Star needs an ASM Brain.

There will only ever be 10,000 Genesis Brains, from the 10,000 ASM AIFA Genesis Boxes. There will be more ASM Brains after this, but they will not have the Genesis trait. Holding a Genesis Brain will provide early access to future projects and releases.

AIFA Genesis Boxes

(OpenSea/Nansen) The collection consists of 10,000 Boxes, divided into four rarities; Original, Rare, Mythic and Ultra. Inside every box there are 4 unique AIFA All-Stars and 1 unique Altered State Machine Brain. These are all separate NFTs.

AIFA

The Artificial Intelligence Football Association (AIFA) is the first decentralized Play-To-Earn game by Altered State Machine, where AI agents compete against other AI agents. It involves building a team of AIFA All-Stars NFTs powered by ASM Brains, in order to train them and compete against other ASM All-Star agents in a 4x4 football game. The gameplay involves finding All-Stars who mesh well and can complement each other's playing styles. As a team manager it will also be possible to raise the abilities of the All-Stars by training their Brains in the Gym.

ASTO Token

ASTO is the native token of the protocol.

• Governance (creating proposals for, and voting in ElderDAO - ASM’s DAO)
• AI Agent Training
• Minting Agent NFTs within ASM universe world
• Minting Agent NFTs within non-ASM universe worlds
• Powerups, instant improvements, items and skins
• In-game transactions, asset creation, and more

The central portal for the various use cases of ASTO is the ASM Cortex. The Cortex is basically a custom dashboard on the ASM Website that enables you to participate in ASM and ASTO-related events. The total $ASTO supply is 2,384,000,000.

The ASTO token has been in decline for the past few months. This is not surprising as the whole market has kept going down. Of course the utility of the token at the moment is also very limited as there aren’t any games utilizing the protocol. The introduction of Genome mining (see section below) is one use case at the moment which is driving demand for the token. We could expect to see significantly higher demand for the token once AIFA and other games launch on the market. This is due to the fact that ASTO will be used for in-game transactions and to train the ASM agents in Gyms.

Genome Mining

Genome Mining is the process of mining Gen II Brains using ASTO-Energy. ASTO-Energy is accrued by staking ASTO tokens, ASTO Auction LP tokens or ASTO-USDC Uniswap LP tokens. It is not a currency or token with external value outside of the ASM protocol or ecosystem.

Genome Mining will run over a series of 60-day cycles. At the end of each cycle, your accumulated ASTO-Energy will become available and can then be used to mine Gen II Brains. Just like Genesis Brains, Gen II Brains will have a unique Genome Matrix. Additionally, they will also share the same functions as Genesis Brains capable of training and evolving, however, only Genesis Brains will have exclusive perks (e.g. early access to future projects). Gen II Brains will be released in three cycles: 35,000, 55,000 and 75,000, respectively. The first Production Cycle began on Thursday 16th June PST / Friday 17 June NZDT.

Use Cases

The types of AI agent that can potentially be developed using a protocol such as ASM can be wide ranging. They can span various sectors and can address various use cases.

• One use case, which the team is also focusing on is games where AI agents compete with other AI agents in a virtual space (e.g. AIFA). This usually involves creating or assembling a team of AI agents, training them to increase their performance and then competing with other teams that have been assembled by opponents. If compared to a play-to-earn game such as Axie Infinity, these games can also be taught to have two tiers of classes (the owner class which owns the NFTs and the player class which actually plays the game). The difference here is that the player class is not made up of humans that can’t afford to buy NFTs, but by autonomous AI agents that can play the game independently. In other words, the human class has been automated. This makes the whole scholarship and reward-sharing model seen in games such as Axie redundant as owners that don’t have time/interest to play the game can still participate in the ecosystem. All they have to do is just manage their team once in a while. The time and effort required to manage a team will of course depend on the type of game.
• Another important use case that can arise from ASM agents, especially in games and virtual worlds are “companion agents”. These could take many forms, for example it could be a pet that follows you around in a game, a virtual assistant that directs you in a virtual world, a personal representative that interacts on your behalf while you are away/“logged out” or even a data collection agent that roams around virtual worlds observing interactions, collecting and storing data. The interesting difference about interacting with intelligent NPCs is of course the fact that they will be able to recognize you as a unique player. They will remember the last conversations you had with them and the specific problems you encountered before, as a result each time you interact with them they will learn more about you and adapt accordingly to give you a more personalized response or service.
• The idea of having intelligent quest helpers/ NPCs could also give rise to new types of gameplay which were previously not possible. In multiplayer games for example you could have NPC team members as AI agents that help you achieve your objective. The more you play with the agents the better they will understand your play style, and consequently the better they will be in protecting you in the gameplay for example. Additionally, by optimizing the combination of the individual team members skillset the stronger your team will become. As a result, your chance of “winning” in the game will not only depend on your individual skills as a player, but also by the quality of agents you have in your team and the strategy you implement to create your team. This would enable games where players are competing against a combination of smart agents and other real players. Unlike in many games today where bots exist but are not explicitly declared, this would be done in a transparent manner as everyone would have access to the same technologies.
• These types of games could also popularize game mechanics such as the introduction of permanent deaths in games. In the example above this could apply to your AI agent team members. Furthermore, one could also create games where instead of just killing each other’s smart NPCs it would be possible to “steal” or “conquer” your team members, in which case the ownership of the AI agent would be transferred to the winner. The new owner can then use the agent in his own team, or could choose to sell him in a decentralized marketplace outside of the game. This of course is only possible due to the tokenization of the AI agent.
• Another aspect where players can play against smart NPCs might not include you personally getting involved. For example, a free-roaming game such as GTA could incentivize anyone to provide AI agent services for some possible rewards. One case could be providing a smart NPC that works as a cop to maintain the peace in a city where real players are active. In this scenario each cop agent (or team) will have different skills based on the training he received, and everytime someone who broke the rules is captured a reward can be paid to the AI agent. These types of gameplay can provide new revenue streams for NFT holders.

It is important to stress here that the idea is not simply for players to compete against an AI and win. It has more to do with transferring the development and ownership of NPCs from the game developers to other participants in the game. So in other words, many aspects of the game are being built by players for players. In many games it will not be possible to compete against an AI, without imposing constraints and rules which will have to be introduced by the game developers. For example, reaction times for the AI will have to be similar to human reaction times. Games that find the right balance will be the ones that will be the most engaging.

The table below shows some of the use cases outlined by the team.

The potential use cases are of course wide ranging, although the implementation might be difficult. A lot of other potential use cases in finance, games and virtual worlds will also emerge as their respective segment also develops. Today, a lot of virtual worlds are mostly empty as they are entirely reliant on humans. In the near future, it is not hard to imagine these spaces to be populated with smart NPCs in order to provide a better experience for the player.

Recently, the development of AI systems that can create realistic images and art from a description in natural language such as DALL- E 2 and the Midjourney Bot showcased the potential of AI in the creative domain. In the long term, if these models become powerful enough and move from creating 2D images to 3D spaces, they could be used to help populate the Metaverse with millions of digital goods and services. Potentially, they could accelerate the development of the Metaverse by automating the creation process of digital content. In such a scenario, tokenizing the models through protocols such as ASM could reap huge benefits. A highly specialized model for example, could be trained on a specific domain (e.g. 3D cars) and provide its services to anyone willing to pay for them. In a way it is acting as a car manufacturer in a virtual space. This use case, although speculative in nature, highlights the potential benefit of AI agent tokenization.

Additionally, the benefits of tokenizing ML models could potentially also be useful in domains such as Federated Learning, where reward mechanisms such as distributing ownership of the model across participants (in relation to how much they contributed to the development of the model) could be explored. Federated Learning is a decentralized form of machine learning where user data used for training stays on the edge device and is never sent to a central server. This enables the creation of privacy-preserving, secure distributed machine learning models.

Key Takeaways

• Altered State Machine is a project at the intersection of blockchain tech and AI. By enabling the tokenization of AI agents it aims to democratize AI ownership. As mentioned above, the use cases of AI-infused NFTs are many and with tremendous opportunities for growth. Many sectors ranging from finance to gaming and entertainment could be significantly impacted if the implementation is conducted as planned. Going forward, it is possible to imagine that many of the virtual worlds which are empty today, will be filled with smart NPCs and games making them significantly more active and engaging. This would also open up opportunities to generate revenue for players (or guilds) as the majority of the characters and content in games/virtual worlds could be created/owned by them instead of the game developers.

• The introduction of AI agents that can transact independently of human intervention and can represent themselves, devices, services, or individuals will be a game-changer. This will have huge implications on the nascent data market, finance, commerce, and many other industries.

• It is important to highlight however that the tech to make all of this happen is still in the very early stages. A lot of existing games/virtual worlds have limited interoperability and importing your own characters into a third-party game environment is still problematic. The number of games/virtual worlds that are being built on open standards with the aim of interoperability are the exception rather than the norm. There are significant hurdles that needs to be overcome in order to port your AI Agent (or even a simple NFT)from one environment to the other in a seamless way.

• ASM is a project that is working to make these possibilities a reality. It serves as an illustration of how the fusion of two or more technologies can accelerate the advancement of one another's innovations. It is evident that some of the most successful businesses in the future will be those which use two, three, or more emergent technologies to offer a product or a service.