Why Deworker AI Protocol?
Last updated
Last updated
A machine has two components: Team and Culture.
A team's effectiveness is measured by the amount of economic value it brings to the world. As Ray Dalio described, all teams are machines with two components: workers and culture. Workers are specialized and used to get the work done. Culture is the connection between those workers that ensures the team works together to guarantee the best possible output.
The team is effective because everyone knows their strengths, and the role of the manager is to effectively build upon each worker's strengths and create a culture where the team operates like a symphony, producing optimal output.
The essence of the team as a machine with workers and culture has stayed the same, but the workflow has changed dramatically since the birth of high-speed internet.
In the past, when organizations could only be formed with everyone sitting in the office, teams needed to be in the most crowded and talented locations like Silicon Valley (with Berkeley and Stanford) or New York (with Wall Street). The team structure was like the best talent in the regional area with the local culture.
Later, with the internet, the organizational structure evolved, especially during the remote work/COVID-19 era, where SaaS tools like Zoom and Slack became essential for daily work. Workers started becoming distributed in multiple locations, and the culture evolved through digital communications online, in Reddit forums, Discord, Zoom calls, and Slack channels. Workers are still specialized in most cases, but they are now based in different locations. In both the Web2 and Web3 eras, fully distributed companies like MakerDAO, WordPress, and Linear emerged.
While we are talking about the machine, distributed work is not changing the worker side of the machine but the culture side of how workers get connected, which has unleashed lots of innovation and creativity in the distributed workforce.
Then everything changed after ChatGPT.
Models are becoming essential in daily workflows, and text-to-X (image, text, audio, video, PowerPoint, website, etc.) is becoming essential. In the early days, they are assisting individuals in their current workflows, like Copilot. However, it will change with inevitable reasoning ability and more niche skill sets to achieve better results, where AI agents will become AI workers/co-workers as a prerequisite of the workflow, not an afterthought. We could view every AI worker as a duplicate of an actual human worker with skills, memory, and the ability to execute. We are not only making the workforce distributed but also making it possible to leverage worker expertise even if the user is not in the market.
We leveraged location from work1 to work2. And we are leveraging time from work2 to work3.
We are entering a new era of organizational structure: from a centralized workforce in talent-centric areas to a distributed workforce across the globe, to a leveraged workforce where a significant part of the team's economic output lies in how we delegate and leverage.
While we recognize delegation as the ultimate leverage, we need to be clear about the difference between the delegator and the delegatee. In the first generation of organizations, the delegatees were manual workers who were laborers, and the delegators were managers. Think about the oil tycoons who managed tons of people to extract and sell oil. In the second generation, the delegatees became knowledge workers who mostly moved the world through their expertise, and the delegators were incredible builders like Zuck, who was also a knowledge worker himself and needed to be able to produce code, write diagrams, and draw designs. Just think about Zuck in his dorm room trying to get Facebook, the hookup dating site, on the market. In the third generation, as we imagine, we will see vision workers who can just imagine and have ideas, and things will be done by all sorts of AI workers in the world.
We are turning from hands to left brain to ultimately our right brain focus in the workforce.
As Stephen King famously wrote in his book "On Writing," writing is telepathy, and the ideas on paper are transferring across time and space. For centuries, we have never been able to easily turn that telepathy into meaningful next steps.
If writing is telepathy, then AI workers turn telepathy into action.
Centralized
Labor manager
Local worker
Labor/ How
Oil tycoons
Distributed
Knowledge manager
Remote Worker
Space/ Where
Facebook, etc.
Network
Visionary manager
AI Worker
Time/ When
Community
Absolute power corrupts absolutely.
Every new leverage causes the users of the leverage to be able to capture more value in the process. It is an efficient way of operating the market to allow those with "new leverage" to thrive, otherwise no one will be interested in adopting such leverage.
However, new solutions create new problems. As the new leverage becomes more powerful, it accrues more power to the user, making them even more powerful. Power itself is not bad; you will not be able to move mountains without a centralized, coordinated group to get the job done. However, the problem lies in the creators of that leverage and how they can capture a part of the value as well to ensure the supply and demand of the market is in check and balance.
In the labor market, it is the union that fights for local workers to be part of the labor force, have 9-5 working hours instead of 12-hour workdays, and have leave of absence, etc. In the remote working setting, it is platforms like Upwork, Freelancer, or even DAOs that provide reviews and checks for those managers to ensure the organization is giving the right amount of value to the worker.
Then what about the network organization? How do we check and balance the power of the leverage to make sure that workers are in coordination with that power?
We have some potential outcomes.
Solution number 1 is to make sure the big players won't be evil. As Google famously said in the past, "Don't be evil." Now, most of the AI power lies either in OpenAI's hands, with few closed-source competitors, or Microsoft.
Most of the data providers are promised by OpenAI that their data will not be used for training. Not only is this a promise that could be changed anytime (just like it used to be a non-profit organization), but it also gives no value back to those users who provide the data. Just like in social media, where you think you are the customer by using Facebook to publish all your images and photos, in reality, you are just the product being served to advertisers.
In the Google or Facebook ads model, at least if you provide your info to Google, Google sends some of the traffic back to you, which might allow you to grow your business based on that traffic.
But in the AI case, not only are you not getting the value (i.e., the money), but you are also not getting the traffic because your contribution has been synthesized in the AI response already.
We already see signals of this issue, like The New York Times suing OpenAI, but if it is a legal solution, it will take years to resolve, and if you are an individual creator, you won't have the money and resources to fight this battle.
The other solution is an open-source one, which is the counterbalance to closed-source with companies like Hugging Face, Mistral, Llama2, etc., leading the way. It indeed provides a great ecosystem for creators to create things upon those models.
One promising community is Civitai, where users are building upon Stable Diffusion as an open-source model and creating images. It is a vibrant ecosystem indeed, but if you look deeply, there are some concerning issues.
One is the lack of value capture from those creators. Even if there are models getting millions of downloads and runs, the value capture lies with Nvidia, platforms like Civitai, and the creators are left begging for donations.
The other is the deep fake issue, which doesn't happen as often in closed-source models because the platform usually doesn't want to be responsible for the issue. However, in the open-source space, it is happening fast. Two of the most deprecating use cases are porn and political issues, where you can easily find celebrity photos being used in all sorts of scenarios on Civitai.
We need a new kind of AI to fix these issues.
Prompts as mini-programs unlock composability for the masses
Both open-source and closed-source AI movements have a premise that whoever is training and running the AI model are developers, which is where the word "source" comes from, as in source code. It also implies that the benefit of different pieces of computing, whether open-source in terms of providing source code or closed-source in terms of providing APIs, has the benefit of interoperability and composability.
The power of composability is that if the software is done well once, you never need to do it again, which is a springboard for humanity to move forward and build new things.
However, for open-source AI, only coders will be able to run those projects and understand how to fork, run, and compose those projects together. Closed-source AI, mostly relies on centralized APIs, which can change pricing, change API call methods, or simply be deprecated at will.
And most importantly, no one contributing to the AI will get benefits from the AI.
Nowadays, with natural language as the default input method for AI models and software, it is the first time we are making composability available to everyone. As long as you understand language, have creativity, and have the willingness to prompt, train, and improve a model, you can unlock completely new use cases with only a few words and clicks.
But if composability is available to the masses, the value creation from those components as Lego bricks needs to come back to the Lego maker.
However, before we define how legos are being made, we need to define participants in the lego-making process. There are three layers of participants in community-source AI.
The user has two roles. One is simply to describe what task they want to get done. It could be sending emails, creating movies, films, etc. All the user needs is a text input. Also, users will act as validators to validate the quality of the AI worker output. Not only does it solve the deep fake issue, but it also gives incentives for creators of AI workers to keep up with the quality and price.
The dispatcher is the key component. The user has various requirements and there are various AI worker providers in the market. The dispatcher can understand the user requirement, plan how to accomplish the task, rank it based on user validation, feedback, and ranking, and assign it to various AI workers (will explain next), what to do and how to have those multi-AI workers collaborate on a coherent workflow together.
The AI worker is the AI agent who is getting tasks done, basically running a specific AI model with memory, context, skills, and unique datasets to help users get the task done. For example, blog post writer, email worker, website scraping worker, anime image generation worker, etc. These are the text-to-text, image, music, etc. players with unique world views and capabilities. It is defined as an AI worker instead of an AI agent to make the term clearer that all AI workers being created have a specific purpose.
The sub-worker: These are creators of AI workers who are providing training data. It could be prompts, images, videos, texts, as long as the AI-Worker is providing a unique AI worker to the public. The AI-Worker will own an AI worker handle like @searchworker for instance as an NFT for future sales.
The miner: We are using a proof of stake (PoS) method where the miner can be running an open-source model locally or can be running a closed-source model simply providing APIs. The purpose of the miner is to ensure the stability of AI models and also provide incentives accordingly.
Thus, we believe the Deworker worker network is not only a protocol, but a new approach to collaboratively building and using AI called - community source AI: The bottom-up approach of collaborative AI worker development upon a blockchain network.
Contribute once, earn forever.
If AI is the accumulation of human intelligence from various sources, then the key features of the blockchain network for community-source AI are to 1) make sure the contributors of those sources get proportional incentives and 2) the quality of those community sources and AI workers can be validated and authenticated.
We do have some centralized ways of creating the network. GPT store is an example where OpenAI controls the incentives (which are still very vague), the ranking (not disclosed) and also the AI model being used on the network. It is not surprising that we have already seen creators not getting paid, GPT getting deranked and banned.
Like Google famously said internally: don't be evil. But we want it to be "Can't be evil" for AI contributors.
Thus we propose a new protocol, Proof of Attestation (PoA), where AI-Workers get attribution and users provide feedback for authentication and ranking.
The network state needs a network workspace.
Here is the definition of the network state from Balaji:
A network state is a highly aligned online community with a capacity for collective action that crowdfunds territory around the world and eventually gains diplomatic recognition from pre-existing states.
It is a bold vision but the most important verb in the statement is collective action. We already have highly aligned online communities living in Discord channels. We have crowdfunding mechanisms from various DeFi projects. We even have diplomatic recognition for the Crypto Islands. But what we are missing is the collective action part.
Action-taking is not the goal for most group chats and Discord channels, and it will never be. Because that is the cultural part of the community where users form a deep alignment with each other on core values. However, action-taking is also necessary to achieve the mission.
With the rise of AI workers, it is even more important to leverage AI to get most tasks delegated to unlock creativity for human workers in the community.
The application layer is designed for this purpose. It is a collaborative AI automation workspace where you can delegate tasks to AI workers, and collaborate with human workers, all in one place.
The user journey looks like this:
Step 1: Orchestrate - Based on your requirements and needs, we will form a group comprising selected AI and human workers to accomplish your tasks in a workflow. Those AI workers will be selected based on their on-chain credentials based on PoA.
Step 2: Dispatch - Tasks will be executed automatically across AI workers and human workers.
Step 3: Collaborate - While the task is being done within those projects, you can interact with AI workers as a group or as individuals to provide feedback or ask for more details.
Step 4: Incentivize - While the task can't be done by AI workers in the network, it provides incentives to the network to get those tasks done. If the task is done successfully, users can provide feedback and AI-Workers will get incentives accordingly.
Step 4 links back to the network design. Every time there is a "Sorry I don't know" happening in the dispatching phase, it is an opportunity for the network to support such causes.
Your "Sorry I don't know" is my quest for an AI worker.
We are moving from single-player rule-based automation for repetitive tasks to multi-player interactive automation for creative tasks.
