Partnered With:

AI is booming, how can generative AI be integrated with Web3?

Read this article in 59 Minutes

The only way for generative AI companies to achieve differentiation is by transferring ownership and control to the users.

Since the launch of OpenAI's ChatGPT at the end of 2022, artificial intelligence has become popular worldwide. This machine can answer most questions and has quickly become the fastest-growing application, attracting over 100 million users. However, like many venture-backed startups, it burns millions of dollars every day.

But burning money like this is not a big problem, because Microsoft has just invested $10 billion in OpenAI. With the hardware of Azure and the traffic distribution of Bing, ChatGPT has crossed the divide of prejudice in popularity.

However, Microsoft is not the only one competing. After Alphabet's generative AI product demo failed, its market value lost $100 billion. Alibaba and Amazon have announced their entry into the competition, vying for the dominant position in artificial intelligence with ChatGPT. Apple is also likely to use Siri as a flagship product to enter this market.

The explosive growth of artificial intelligence tools is due to the fact that people can now truly use artificial intelligence. Unlike other trends such as Crypto, drones, or self-driving cars, artificial intelligence has no high entry barriers. Anyone can use artificial intelligence to complete homework or even improve their chances of success on dating apps like Bumble. In contrast, in Crypto, you can only buy a token and pretend it is your entire personality.

Arthur C. Clarke once said, "Any sufficiently advanced technology is indistinguishable from magic." Artificial intelligence has become so complex that many now consider it to be magic. In the second quarter of 2022, AI-related companies received a whopping $17 billion in funding. This surge of funding from FAANG, venture capital firms, and retail investors indicates that AI is ready to capture the public's attention.

We spent several weeks investigating the current state of the artificial intelligence industry and its potential impact in the next decade. This article presents our limited understanding of the industry and why we believe blockchain and artificial intelligence will converge in the next ten years. However, before delving deeper into this topic, let's first review some basic concepts in economics.

Attention: I used the terms "AI model" and "model" interchangeably throughout the article. For the sake of readability, any mention of the term "model" refers to generative artificial intelligence models.

Scarcity's disappearance

The history of humanity is a history of struggling with scarcity. The migration for resources can be traced back tens of thousands of years. After we learned to harness the power of fire and agriculture, our ancestors thrived and created entire civilizations. In order to ensure that society has the necessary resources, we ventured across the seas and participated in trade.

With the development of civilization, human attention has shifted from basic needs to competition for status. The Great Wall of China, the pyramids of Egypt, the Taj Mahal in India, and the great cathedrals of the European Renaissance are all examples of symbols of status that played a crucial role in the social and economic structures of those times. People were able to mobilize thousands of workers and spend decades building these magnificent structures because they no longer had to worry about finding food.

As we transitioned into a society that values status, skills and services became scarce. By the 14th century, celebrations of war heroes decreased and artists, inventors, and performers became more respected. The works of Shakespeare, Michelangelo, and Banksy demonstrate how artists can interpret the world in unique ways, often requiring decades of cultural experience. Without a deep understanding of culture, creating culture is impossible.

What we are facing today is not so much scarcity around material goods, but rather scarcity around the goods that affect our mental state. Outstanding artists only appear once in a generation because the environment that produces them is difficult to replicate. Even if hundreds of people experience the same environment, only a few, like the rap star Jay-Z, can become billionaires. Their emergence is often unpredictable and elusive.

In a society where skills are scarce, outstanding talents will receive high rewards. During the Renaissance in Europe, artists who worked for cathedrals were often commissioned for years of creation. However, for most of the past few centuries, economic output has been proportional to the energy consumed. Energy is either burned to power factories or used by humans to produce goods.

The past wealth production mainly depended on the number of people or the resources that one person could obtain, which is why we had such terrible stages of slavery in our history. The growth of wealth is a linear trajectory that often relies on conquest.

In the 20th century, code and servers changed traditional ways of wealth production. People no longer needed to invade foreign lands or conquer people to gain chips. As Naval Ravikant pointed out, code and media are the new chips, and people can harness robot armies through code.

Looking at platforms like Instagram or Tiktok, you will find that the number of employees is not proportional to the number of users. Instead, scalability is achieved by increasing hardware to support the continuously growing user base.

This prosperous era can be traced back to the emergence of the Internet in the early 1990s. For example, the reduction in communication costs has led to the problem of spam. Digital storage and bandwidth have become more easily accessible, as evidenced by platforms such as Limewire and Napster. Today, games and social networks have become digital gathering places for millions of users, and the scarcity of space is no longer an issue.

In the digital age, the marginal cost of meeting an additional user demand has greatly decreased. With the development of digital technology, scarcity has become a thing of the past. Naval discusses this topic in more detail in this video.

Developers have immense power, similar to skilled craftsmen during the Renaissance, because they can use their time and expertise to create powerful tools. However, this dynamic of power is slowly evolving. Microsoft's recent actions - laying off 10,000 employees while investing $10 billion in OpenAI, a company like ChatGPT - vividly illustrate this contrast.

I have no intention of instilling fear or implying that developers will be obsolete. On the contrary, we may witness the rise of alternative jobs empowered by artificial intelligence. This transformation is reflected in the development of current generative artificial intelligence technology.

富足时代的到来

translates to

The Arrival of an Era of Prosperity

in English.

We are currently in an era of abundant digital consumption. At the same time, the cost of dissemination is becoming lower and lower. Whether the readership is 1000 or 100000, the cost of distributing this article is the same. Substack's infrastructure can quickly expand, so it does not charge based on pageviews. However, even though we can now distribute content freely, due to people's limited attention, the amount of content creation and article reading are both scarce.

As an author, my productivity is limited by how much meaningful content I can create within a specific time frame. As a reader, there is only so much you can expect to read from me.

The rise of concepts such as ChatGPT and generative artificial intelligence is disrupting the economic theory of production and consumption. The last time we witnessed a similar situation may have been during the rise of the printing industry. The popularity of books greatly reduced the cost of storing and disseminating human thought. Although it took England four centuries to raise literacy rates from 5% to 50%, people now spend almost two hours a day reading.

The changes in this type of intellectual behavior directly led to the arrival of the Enlightenment era, characterized by the rapid development of science and philosophy. Figures such as Kant, Voltaire, Descartes, and Adam Smith forever changed our worldview. Throughout history, from cave paintings to digital files on Google, the process of storing, sharing, and iterating ideas has created great achievements every time.

In order to understand the impact of generative artificial intelligence, we must recognize its potential to reduce production and consumption costs. Applications such as ChatGPT and Midjourney generate results based on existing data: ChatGPT uses data from books, Wikipedia, and journals to create its text responses, while Stable Diffusion initially relied on stock images; Github's Copilot utilizes billions of lines of code to assist developers.

Essentially, generative artificial intelligence synthesizes publicly available information and processes it according to user requests, from explaining complex concepts in simple terms to creating visual scenes for specific scenarios. AI can provide results for instruction requests in a matter of minutes, even if the output may not always be convincing. However, with sufficient adjustments, you will eventually get some content that may be considered human-generated, provided that the topic is general and you do not expect any personalized elements.

摘自红杉关于生成式 AI 的文章

Translated:

Excerpt from Sequoia's article on Generative AI

So, what stage have we currently reached? The above image from Sequoia's blog provides us with a good outlook on the current situation. Nowadays, AI models can process text input in real-time, edit text, adjust tone, and even edit images and code. The implementation of these features is heavily reliant on generative AI models trained on large datasets. If we can quickly obtain and integrate knowledge from multiple books or generate art, we will eventually face the problem of information overload. In handling complex tasks such as editing movies or music, human involvement is still essential, but AI can certainly accelerate this process.

Ironically, AI has already begun to be used to summarize large amounts of literature and provide key insights. However, the challenge is that we cannot confirm its citation sources. For example, ChatGPT may give incorrect answers in the absence of sufficient information in the public domain, and it does not indicate the source of the data used to generate the answers. This is precisely the risk that current generative AI faces.

We can create endless works, but we cannot trace their inspiration or authorship. At some point, we will rely on AI to process this content and tell us which information is important.

My Twitter content in the past few weeks

If there is no mechanism for tracing the source of infrastructure or verifying AI models, generative AI will be nothing more than a replica of today's internet: a bunch of fake news generated by algorithms and information pushed based on user preferences. Although generative AI can make everyone an artist, it may also marginalize the original creators who provide training materials for AI models. This will be a key challenge in the era of information explosion: how to ensure that works in the form of art, text, or code are effectively attributed and commercialized when everyone benefits from them?

The internet has provided some answers to this problem. Platforms like Instagram's Reels and TikTok heavily rely on music works from third-party artists. Users mix and edit audio clips to create short videos that are catchy, trendy, and cover various themes such as dance and cooking. TikTok has introduced SoundOn to help artists upload and monetize their works. In these cases, the platform is responsible for licensing and paying royalties. When the audio is the same and distribution is controlled by the platform itself, handling this issue becomes relatively simple.

However, when the audio content involves multiple works and is remixed by users outside the platform, the situation is different. Do you remember the thing I mentioned before about rewriting "Harry Potter" in Shakespearean style? Considering that it only involves two works, it will be relatively easy to handle. But if we rewrite "Harry Potter" in Shakespearean style and add the plot of "Game of Thrones", then the final work will have three authors who need to share royalties.

The challenge facing current generative AI is that you need to process the works of hundreds of authors to generate thousands of output results, and none of these works can be accurately identified, attributed, or tracked.

Usually, creating these works costs nothing. I can run hundreds of commands on ChatGPT every day until I get the answers I want. Today's social media networks are largely similar to slot machines in casinos, where users spend hours searching for content that can bring them huge surprises.

With the emergence of generative AI, we encourage people to continuously try various instructions until they get an ideal answer. However, there is a method to solve this problem, which is to enforce costs. In the NFT field, we have some early attempts regarding this.

Verifiable factual truth

Since the 1970s, the copy and paste function has existed. Looking at the Instagram captions of the younger generation, you will find that we have used this function countless times. A simple technological change has changed our understanding of the copy and paste function - NFT.

Of course, you can copy a picture of a bored monkey, but to own it, you usually need to spend a lot of money to buy it. Blockchain makes this unique asset still attractive even in the case of easy replication, while achieving real-time verification of ownership of the asset across the entire network.

Some people may think that this technology does not have much practical use in real life. However, once generative AI is taken into account, these two tools can be integrated with each other. Soon, generative AI will largely rely on copyright. Large film and television companies such as Disney or Netflix own the copyright of the characters we loved from childhood to adolescence.

Film and television companies can use AI to mine the audience's deepest and most precious memories, and create personalized roles for them. What if Iron Man could teach children math? What if Darth Vader could give dating advice? Maybe this will help many people find a partner.

Undoubtedly, film and television companies can own and release products of these AI-generated chatbots or interactive characters. However, by utilizing blockchain technology, they will be able to track, verify, and claim royalties. They can establish an open market where anyone can participate and create, not just limited to in-house developers of film and television companies. In fact, any company with a large amount of IP can develop itself into a large platform by allowing derivative works.

Let me explain what this means. Imagine that I decide to have Chester Bennington (deceased) from Linkin Park read this long article I wrote for you. The estate holders of his IP and the studio that holds his synthesized voice could soon grant me the authorization. However, this would be a long and tedious process involving lawyers and a lot of paperwork. This may seem foolish, but an industry has emerged around deceased celebrities and their intellectual property.

Imagine what would happen if Chester's voice copyright was on the chain. In this way, it can be authorized to hundreds of people around the world. Of course, people may be concerned that this voice will be abused, such as being used to create deep fake videos or content that the deceased does not want to be associated with. However, if the usage rights are priced high enough, the entry threshold will screen out most malicious actors.

This situation has already been put into practice in some fields, that is, meme generation. I recently saw a series of posts on Instagram by Raj Kunkolienkar, one of the founders of Stoa. He recreated several popular memes using his own photos, which is an interesting application of generative AI.

Memes are a part of culture and belong to the public domain. People have tried to turn them into NFTs. However, if we want to "reshape" culture in our own identity, I believe these memes need to have verifiable sources and should reward the original creators as much as possible.

Should Raj "authorize" the right to create derivative works from these images? Should culture be monetizable? I don't know. He may just be spending a Saturday morning in beautiful Goa, and it seems there may be a way to do that.

Sorry, Raj - I just wanted to find a way to promote my investment in Stoa. That's adding value, right?

In the past, intellectual property such as superheroes and game characters were more relatable to the audience. At comic conventions, it's not unusual to see people dressed up as Batman or Darth Vader. We're unlikely to require fans to purchase licenses in order to dress up as their favorite characters.

However, it is possible that some fans may crowdfund to obtain the copyright, thus legally re-creating and releasing new works to supplement the original author's ideas and content.

You may find this absurd, but in the Web3 ecosystem, it has become a reality. Last year, a community raised $47 million in an attempt to bid on the US Constitution at Sotheby's auction house. Despite the bid being unsuccessful, thousands of people rallied to contribute their hard-earned money. In the end, users could choose to get a refund or hold onto their PEOPLEToken. As of now, approximately 17,000 users hold the token, which is valued at around $140 million.

Token and on-chain traceability enable the community to unite and purchase intellectual property rights, which can be collaboratively used with AI to create derivative artworks.

Collaboration between humans and machines has been widely applied. In June 2022, Cosmopolitan magazine used DALL-E to create a magazine cover. They made a version of Darth Vader on the cover, but chose not to publish it.

In this case, if you want to build a large community, the operation manual will be developed by DAO. The studio itself issues on-chain tools that represent these rights, which may be a single NFT obtained by the community through crowdfunding. The issuance ratio of tokens is proportional to the amount of funds contributed by members to purchase intellectual property rights. Then, DAO will decide how to manage and use the license.

Communities can set a minimum token amount required for the use of licenses, while the execution of more complex functions, such as creating generative art, may require a DAO vote. DAOs can generate cash flow by requiring a portion of the revenue from the aforementioned IP to be returned.

Due to the risk that large studios cannot afford, emerging artists are more likely to adopt this business model. Although this seems out of reach, artists are always pioneers whenever we try a new form of distribution or a better way to interact with the audience. In the past decade, Spotify and Soundcloud have become key tools for discovering emerging artists. In the next decade, artists will combine on-chain primitives and generative music to create income and accelerate their careers.

Reconsidering Technology Stack

From the article "Who Owns the Generative AI Platform" by A16z

I want to know which part of the generative AI technology stack may be disrupted, and which parts have achieved significant growth. The above image is from A16z's article "Who Owns the Generative AI Platform", revealing where value has been accumulating over the past 18 months.

Article excerpt:

We observe that infrastructure providers may be the biggest winners in this market so far, occupying the majority of funds flowing through the technology stack. Application companies experience rapid revenue growth, but often face difficulties in user retention, product differentiation, and gross margins. While most model providers have formed such a market, they have not yet achieved large-scale commercialization.

In other words, the companies that create the most value (i.e. train generative AI models and apply them to new applications) have not captured most of the value. While it is difficult to predict what will happen next, we believe that the key to the problem lies in which parts of the technology stack are truly differentiated and maintainable. This will have a significant impact on market structure (such as horizontal and vertical company development) and the driving factors of long-term value (such as profit margins and user retention). So far, we have found it difficult to find structural maintainability anywhere in the technology stack, except for the defense system of existing companies.

This article indicates that despite revenues exceeding $100 million in multiple fields, concerns still exist regarding profitability and retention rates. When everyone has access to underlying AI models such as Stable Diffusion or ChatGPT, no one will understand maintainability, and it will be difficult to accurately predict user retention rates after the novelty wears off.

Most of the value capture occurs in the hardware and cloud platform fields. Amazon AWS, Google Cloud, and Microsoft Azure have spent decades perfecting storage and computing technologies to provide large-scale hardware with unit economics that meet actual needs; while Filecoin, Render, and Akasha have become corresponding platforms in the Web3 field.

However, in its current form, I am still unsure how to beat the reliability and scale provided by centralized service providers through crowdsourced hardware. According to A16z's analysis, value can be accumulated in three areas: physical infrastructure, AI models or applications. We believe that in the Web3-based AI field, we can build competitive barriers by planning to use token incentives to target specific user groups, developing data flows and profit models that can be extended through community participation.

We have seen some examples in this regard.

Target Users

Numeraire's original MVP provides us with an example. The company behind the Token released a standardized dataset on the stock market to researchers, who used their own AI and machine learning models to analyze the data and come up with a "signal". Simply put, this signal shows their prediction of asset price trends. Based on the amount of NMR native Tokens staked by the signal provider, the signal is assigned a certain weight.

Tokens that provide false signals to users will be destroyed. Since these tokens are traded on the liquidity market, users will not provide incorrect signals in order to obtain real currency. Instead, researchers who provide accurate signals will be rewarded. Over time, the actual effect of doing so is to allow the group of users who make accurate predictions to collect more NMRTokens, thereby influencing the company's funding model, although all of this seems unreal.

So, can a decentralized group of researchers incentivized by tokens beat the market? In fact, they can. Since its inception, the fund has had a return rate of approximately 48%. According to the company's website, NMRTokens worth about $55 million have been awarded to data scientists who have developed over 5,000 models.

In the case of Numerai, the data itself is not a proprietary resource. The real value lies in the network of data scientists who trust the product and are willing to stake tokens and share model output results. The network uses a token incentive mechanism to create a community of data scientists focused on a specific field, which in itself is a moat.

Data Stream

The plugin can collect and anonymize data at the browser or hardware level, and pass it on to third parties who can benefit from it. The internet itself is already performing this operation. Our data is collected and passed on to companies that sell us unnecessary products.

This kind of system no longer relies on platform monopolies like Google or Facebook, but instead relies on a protocol that standardizes data collection and provides data on the market. Companies can offer benefits, such as advanced access privileges, in exchange for users' consent to share data. We have already seen early versions of this model on Pocket.

The team behind Pocket is creating a standardized protocol that represents users building data structures, allowing enterprises to request Pocket users to share data in an easy-to-use format. At the same time, users can choose which data to share based on unlocked benefits. The BAT reward in Brave browser is an early version of this model.

Model Monetization

Web3 native products may interact with AI through leasing AI models. Early versions of this model already exist on the Ocean Protocol market. In this case, research teams can develop and authorize AI models to third parties, who then apply them to their own hardware and data.

Some viewpoints believe that in the world of open-source models, there is no primitive for verifying the source of output. Combining crypto-economic primitives (such as DAO or Token) with open-source models can provide additional cash flow sources for researchers' work while verifying the source of output.

AI model suppliers can maintain and update their models through collaboration with researchers, and they will receive a portion of the cash flow generated by leasing AI models. In this context, we can view generative AI models as NFTs and researchers as artists. This model may be applicable in cases where data is too sensitive to be shared, such as medical data, proprietary financial data, or user data. Early versions of this model can already be seen on platforms such as Hugging Face and Replicate, although they do not yet include any Web3 elements.

A16z's article concludes with the following statement:

Currently, there seems to be no systematic barrier for generative AI. Preliminary estimates suggest that applications lack significant product differentiation as they use similar models. The long-term differentiation prospects of such models are unclear as they are trained on almost identical datasets using similar architectures. Cloud service providers lack in-depth technical differentiation as they run the same GPUs.

The only way for a generative AI company to achieve differentiation is to transfer ownership and control to the user. As it stands, data is typically crowdsourced. AI models are open source, with value flowing downwards to pay for hardware costs.

Incentivizing users to share data or AI models can reduce the burden and expenses of the team. Conversely, when the model is managed by the community rather than independent gatekeepers, it may greatly improve the output results of generative AI platforms.

This seems a bit far-fetched, but some teams in the industry have already combined blockchain infrastructure (such as NFT), copyright, and generative AI. We saw early attempts at this in Alethea AI. But before we talk about this, it's best to understand what's happening in the generative AI symbiosis field.

Generative Artificial Intelligence DApp

Last Tuesday, Packy McCormick's Not Boring by Packy McCormick published an article about Replika, revealing the problems that users of generative AI products will face in the coming years. In short, Replika was once an application that allowed people to interact romantically through chat. Unfortunately, the GPT model that drove the application suddenly thought it was harassing users and sending inappropriate content to children under 13, which was not the case. As a result, the developers of the application quickly disabled all adult content in the app to control the situation.

In the Replika subreddit, users have discovered that the robots they have established emotional connections with are no longer operating as they used to.

Initially, this application promised to provide companionship and even enthusiastic conversations. Users had high expectations for this and the application fulfilled its promise. However, suddenly, it decided to no longer provide such services to its users. For many users who have difficulty interacting with others or forming close relationships in the real world, this application was a good place to escape. But in an instant, they lost the opportunity to escape.

This reflects the challenges that users of generative AI products will face in the coming years. People who spend hours on these applications have almost no ownership of the creativity, models, or content involved. A few early advantage-taking companies can reap huge benefits from publicly available resources.

This situation will shift the power dynamics towards companies that produce generative AI models. Therefore, OpenAI may start prioritizing the companies they invest in and impeding the development of competitors. One way to mitigate this risk is to treat the models as public goods and keep them open source, while allowing the community to manage them.

Alethea AI has been working towards this direction. They define themselves as the "cornerstone of generative AI economy". Suppose you create a cutting-edge model and only want users holding specific tokens to access it. Or, you want to embed an AI-powered NFT, transforming it from a lifeless static asset to an interactive intelligent asset that can engage in real-time conversations (usually achieved by interacting with Alethea's models). Alethea's protocol allows developers and creators to perform these operations without permission.

In order to build on-chain intelligent dAPPs and intelligent NFTs, service providers specializing in generative AI fields such as copywriting, protein folding scientific research, or Python code writing can share their products to train Alethea's models.

Then, downstream creators can combine their NFTs with specific skills in the Alethea model and train their intelligent NFTs to provide services. Alethea, as a protocol, can manage the Crypto economic incentives of the market composed of AI models, service providers, and NFT users seeking AI services, and handle all transactions.

Of course, we also need to pay attention to some things. The current stage of generative AI is like the mobile internet era of 2007. Whenever a new medium appears, people always need some time before they start releasing their own versions. People are still exploring how to create AI models for specific fields. I am very willing to provide a generative AI model for analyzing on-chain activities every week and discussing hot topics like Bored Ape NFTs. However, I have not yet acquired the skills to achieve this goal.

As we witness the gradual evolution from fintech applications to decentralized finance (DeFi), centralized and closed-source providers will dominate the market for a period of time. Only when teams realize that handing ownership over to a larger user base can spark network effects and data sources in ways that centralized enterprises typically cannot achieve, will we see more open-source and community-governed models. Specific generative AI models often form a community, and the only way to join that community (or gain access to the AI model) is to upload one's own data. This is like liquidity mining, but for the generative AI era.

Last year, Alethea showcased a possible vision. They released an intelligent NFT (iNFT) through a Sotheby's auction, which ultimately sold for nearly $500,000. This NFT is an artistic form that can converse with OpenAI's GPT-3. Web3 native companies have long been accustomed to the risks of building on centralized platforms. Games, wallets, and exchanges are often removed from app stores.

Dependence on OpenAI for NFTs also carries similar risks. Therefore, the team behind the project has developed their own GPT model, named CharacterGPT. Unlike the text-based responses you currently receive on ChatGPT, their developed model is capable of generating synthetic speech, facial animations, and personality. The difference is that first, generative AI models cannot be stored on the chain; and secondly, Alethea's team believes that this will not happen in the future. Currently, CharacterGPT is private and centralized.

But what may gradually be realized is the decentralization of governance of the model. Local tokens like Alethea's ALI can determine the response of the model in certain situations. This does not necessarily mean that the model will avoid self-regulation. If community members act in their own interests, they may push for edge cases that could cause a lot of trouble, but this is still a more decentralized process. Even before I proposed that DAOs could govern generative AI models at the technological frontier, Adam Smith's invisible hand was everywhere.

Alethea itself only provides smart contracts, CharacterGPT, and protocols that connect on-chain primitives (such as NFTs) with generative AI. They believe that at some point, third-party developers will create decentralized applications (DApps) that offer a variety of services based on their foundation. You can try this feature through a third-party DApp built on Polygon - Mycharacter. It allows you to generate a synthetic character with adjustable attributes and mint it as an on-chain NFT. You can converse with Sandeep Nailwal's digital persona here.

This robot's ability to mimic Sandeep's speech is so realistic that it's a bit frightening.

Noah's Ark is a decentralized application that showcases how AI services and on-chain primitives will be combined. Users can use a small selection of pre-selected NFT products and host them as AI characters that people can interact with. The way it works is that users "fuse" their NFTs with something called a "pod" on the platform. As of the writing of this article, each pod is priced at $300 on OpenSea. Essentially, a pod is a passport to use Alethea AI services. You can connect your pod with your NFT to enable functions such as reciting songs or informing about the weather.

Under the current form, these products look like toys. Of course, no one would want to talk about NFTs all day. One of my standards for consumer products is whether it can spark my interest in using it, like how I am excited to use my PS-VR device. However, talking to a robot like NFT may not be as exciting. Understanding this can be helpful for startup teams in choosing the direction of product development.

Alethea's mission is not just to establish a dialogue interface. As it stands, in fact, most of Web2 is a front-end for OpenAI to collect data. Even users who use Chat-GPT or Stable Diffusion are unwittingly contributing to its development, but cannot share in its benefits. Alethea's philosophy is that in the future we will see multiple models owned by users. Currently, there is no protocol that can easily develop these models and combine them with existing on-chain primitives, but this is precisely the goal that Alethea is pursuing.

But what aspects will such a future involve? To understand this, we need to go back to the question mentioned at the beginning of the article - how is scarcity reflected in an affluent era?

Just like what we saw in the early days of the internet, we will go through a period of rampant piracy through torrent and P2P file sharing. Is it ethical to build an AI model similar to Hitler? How should royalties be distributed if third-party developers use proprietary intellectual property of a studio? In such a period, bringing intellectual property onto the chain will be greatly beneficial. We now have the tools to achieve this goal, and that is NFT.

Both studios and creators will benefit from it. Suddenly, you have the opportunity to create cash flow from those idle assets. On the other hand, creators of generative AI tools can expand without worrying about breaking the law. Github's Copilot feature already allows you to use AI while coding. But what if you could replicate the style of your favorite developer? Or what if Rick Rubin provided feedback on the music he produced for you?

Someone has created a robot that can mimic Paul Graham's responses using all of his articles. You can see a demonstration of it in the tweet below.

In this case, how do we ensure copyright acquisition and royalty payment?

The prominent use cases we see around this technology are text and art, as these are two of the most commonly recorded and easily trainable work mediums. With the ease of creating models and the increase in complexity of information fed into the models, we will see individuals publishing their own AI services. This, in turn, will reduce the time people spend answering basic questions. Perhaps my AI version will recommend venture capital funds to founders or explain the risks of collaborating with others.

The scarcer the required intelligence and knowledge patterns are, the higher the cost users pay to access these services. The real world also operates based on similar economic principles, where you need to pay a high premium to obtain services provided by increasingly specialized institutions, such as lawyers proficient in Web3 or venture capital firms knowledgeable in gaming.

Community VS Enterprise

I often ponder why DeFi has developed so rapidly in the early 2020s. In hindsight, the reason is that people are incentivized to contribute capital and use these products through Crypto economic primitives such as Tokens. Historically, the cost of funds for fintech companies engaged in trading or lending is much higher than the cost required to issue Tokens and use them as rewards.

By using these tokens to govern DeFi products, it provides people with a sense of ownership that is not easily replicable in private enterprises. Of course, you can purchase stocks and claim ownership, but this is a different concept from the ownership generated by using the product.

In the battle between Blur and OpenSea, NFT is the common theme. By incentivizing liquidity through tokens and giving governance power to users, Web3 and AI will also collide for the same reason.

The community will gather around the acquisition of intellectual property of creators they appreciate and organize themselves in the form of DAO. Creators like Mike Shinoda and Snoop Dogg have already become a part of the Web3 ecosystem. I believe it is not impossible for Snoop Dogg to tokenize his voice and sell it to the community. He recently showed off his bored ape in Eminem's music video.

He can use tools like Noah's Ark to merge his voice with Bored Ape. Once the intellectual property is tokenized and put on the chain, it can be embedded into the model, and these models can obtain data from the public.

This may sound a bit far-fetched, but considering that Stable Diffusion is an open-source project that is now in trouble because it used stock images from Getty. If they decide to allow users to upload artistic works created over the years in exchange for governance rights over the model, what will happen?

Millions of creators can contribute their works without being harmed. Stable Diffusion based on blockchain can easily track the source of the artwork used to generate images and charge fees to those who download the artwork. Then, the contributing artists can receive a share of the royalties.

The above model shows what it might look like. This is a recombination of the current generative AI technology stack recommended by A16z. Web3 native AI platforms can incentivize crowdsourced datasets (such as images) by providing token incentives. Contributors such as artists can share their work, which can then be used by the model to generate art in a specific style.

If a user who inputs a query command decides to use a certain work, they can mint an NFT that displays the model used and input data. Such generative art will have value like past NFTs because its origin can be verified.

The next Open Sea may integrate contributors who create generative art forms, provide data or run queries, and on-chain primitives to prove the various components included in the work. Groups responsible for maintaining and optimizing models can lease them to markets where users want to mint NFTs. Alternatively, a large model used by millions of people can be open-sourced and run as a DAO, giving people more say in maintaining and expanding generative AI tools.

You may think this is irrelevant - just a useless solution. But you can ask Replika's users how they would react when product decisions are made in the app without their input.

Nowadays, these tools can be trained with only a small amount of data. Therefore, the artists who originally contributed to the creation of these tools may become redundant. To ensure fairness, we can give them a corresponding proportion of tokens based on the usage of their works.

Imagine if those artists who contributed to the early development of OpenSea could receive equity or tokens as a reward on the platform. Then, perhaps they wouldn't have to worry about royalties like they do now.

I have attempted to sort out the different philosophical approaches to AI between Web2 and Web3.

At first glance, Web3 native AI seems like a useless solution, which was also my thought for the past few months. But if you pay attention to the development of OpenAI in the past few months, you will understand that we lack tools to deal with platform monopolies. Between tokens, NFTs, and on-chain sources, the industry has created a sufficient toolkit to deal with the fake news and unemployment crisis we will face in the next few years. This is no longer just about crypto, but about how to build a strong system to prevent artificial intelligence from producing meaningless content and causing chaos.

We need to apply the principles of Web3 in the context of AI, because AI technology is too powerful to be left under the control of a few companies. When the internet was just emerging, we didn't have tools to verify sources or govern platform monopolies. But now the situation has changed, and platforms like Uniswap have shown us a distributed, community-owned infrastructure and viable centralized alternatives. As time goes on, we will inevitably face similar issues with generative AI models.