Foresight Ventures: Best Attempt at a Decentralized AI Marketplace

By: Ian@Foresight Ventures

TL;DR

• A successful decentralized AI marketplace needs to closely combine the advantages of AI and Web3, take advantage of the added value of distributed, asset rights confirmation, income distribution and decentralized computing power, lower the threshold for AI applications, and encourage developers to upload and Share models while protecting users' data privacy rights, and build a developer-friendly AI resource trading and sharing platform that meets user needs.
• Data-based AI marketplace has greater potential. The marketplace of the dead model needs the support of a large number of high-quality models, but the lack of user base and high-quality resources on the early platform makes it difficult for excellent model providers to attract high-quality models; while the data-based marketplace is decentralized and distributed Collection, incentive layer design and data ownership guarantee can accumulate a large amount of valuable data and resources, especially private domain data. However, data marketplaces also need to address the challenge of data privacy protection, and solutions include designing more flexible policies that allow users to customize the settings of privacy levels.
• The success of the decentralized AI marketplace relies on the accumulation of user resources and strong network effects. Users and developers can get more value from the market than they can get outside the market. In the early days of the market, the focus is on accumulating high-quality models to attract and retain users, and then after establishing a high-quality model library and data barriers, it turns to attract and retain more end users. Moreover, an excellent AI marketplace needs to find a balance between the interests of all parties, and properly handle factors such as data ownership, model quality, user privacy, computing power, and incentive algorithms.

1. AI Marketplace of Web3

1.1 Review of the AI track in the web3 field

First, review the two general directions of the combination of AI and crypto that I mentioned earlier, ZKML and decentralized computing power network👇

ZKML

ZKML makes the AI model transparent + verifiable, which means to ensure that the three factors of model architecture, model parameters and weights, and model input can be verified on the entire network. The significance of ZKML is to create the next stage of value for the web3 world without sacrificing decentralization and trustless, and to provide the ability to undertake wider applications and create greater possibilities.

Foresight Ventures: AI + Web3 = ？

Computing Power Network

Computing resources will be a major battlefield in the next decade, and future investment in high-performance computing infrastructure will increase exponentially. The application scenarios of decentralized computing power are divided into two directions: model reasoning and model training. The demand for AI large model training is the greatest, but it also faces the greatest challenges and technical bottlenecks. Including the need for complex data synchronization and network optimization issues. There are more opportunities to implement model reasoning, and the future incremental space that can be predicted is also large enough.

**1.2 What is AI Marketplace? **

AI marketplaces are not a very new concept, and Hugging Face is arguably the most successful AI marketplace (except for the absence of transactions and pricing mechanisms). In the field of NLP, Hugging Face provides an extremely important and active community platform where developers and users can share and use various pre-trained models.

From the success of Hugging Face, it can be seen that an AI marketplace needs to have:

a. Model resources

Hugging Face provides a large number of pre-trained models that cover a variety of NLP tasks. This richness of resources has attracted a large number of users. Therefore, it is the basis for forming an active community and accumulating users.

b. Open source spirit + spread and share

Hugging Face encourages developers to upload and share their models. This spirit of openness and sharing enhances the vitality of the community and enables the latest research results to be quickly utilized by a large number of users. This is based on the accumulation of excellent developers and models, accelerating the efficiency of research results being verified and promoted.

c. Developer friendly + easy to use

Hugging Face provides easy-to-use API and documentation, enabling developers to quickly understand and use the models it provides. This lowers the threshold for use, improves user experience, and attracts more developers.

Although Hugging Face does not have a transaction mechanism, it still provides an important platform for the sharing and use of AI models. Therefore, it can also be seen that the AI marketplace has the opportunity to become a valuable resource for the entire industry.

Decentralized AI marketplace in short:

Based on the above elements, the decentralized AI marketplace is based on blockchain technology, allowing users to have ownership of their own data and model assets. The value brought by Web3 is also reflected in the incentive and transaction mechanism. Users can freely choose or match the appropriate model through the system, and at the same time, they can also put their own trained models on the shelves to obtain benefits.

Users have ownership of their own AI assets, and the AI marketplace itself does not have control over data and models. Instead, the development of the market is dependent on the user base and the consequent accumulation of models and data. This accumulation is a long-term process, but it is also a process of gradually establishing product barriers. The market development is supported by the number of users and the quantity/quality of models and data uploaded by users.

**1.3 Why pay attention to the AI Marketplace of Web3? **

1.3.1 Compatible with the general direction of computing power application

Due to communication pressure and other reasons, it may be difficult to implement decentralized computing power on the training base model, but the pressure on finetune will be much smaller, so it has the opportunity to become one of the best scenarios for the implementation of centralized computing power network.

A little background knowledge: why the fine-tuning stage is easier to land

Foresight Ventures: A Rational View on Decentralized Computing Power Network

AI model training is divided into pretraining and fine-tuning. Pre-training involves a large amount of data and a large number of calculations. For details, please refer to the analysis in my above article. Fine-tuning is based on the base model, using task-specific data to adjust model parameters so that the model has better performance for specific tasks. The computing resources required in the model fine-tuning stage are much smaller than those in the pre-training stage. There are two main reasons:

1. Data volume: In the pre-training phase, the model needs to be trained on a large-scale data set to learn a general language representation. For example, the pre-training of the BERT model is carried out on Wikipedia and BookCorpus, which contain billions of words. In the fine-tuning stage, the model usually only needs to be trained on a small-scale dataset for a specific task. For example, a fine-tuned dataset for a sentiment analysis task may only have a few thousand to tens of thousands of reviews.
2. Number of training steps: The pre-training stage usually requires millions or even billions of training steps, while the fine-tuning stage usually only needs thousands to tens of thousands of steps. This is because the pre-training stage needs to learn the basic structure and semantics of the language, while the fine-tuning stage only needs to adjust a part of the parameters of the model to adapt to a specific task.

For example, taking GPT3 as an example, the pre-training phase uses 45TB of text data for training, while the fine-tuning phase only needs ~5GB of data. The training time for the pre-training phase takes weeks to months, while the fine-tuning phase only takes hours to days.

1.3.2 The starting point of the intersection of AI and crypto

To judge whether a web3 project is reasonable, one of the most important points is whether it is crypto for crypto, whether the project maximizes the value brought by web3, and whether the addition of web3 brings differentiation. Obviously, the added value that web3 brings to this AI marketplace cannot replace rights confirmation, income distribution and computing power

I think an excellent Web3 AI marketplace can closely integrate AI and crypto. The most perfect combination is not what applications or infra the AI market can bring to web3, but what web3 can provide to the AI market. Obviously, for example, each user can have the ownership of their own AI model and data (such as encapsulating the AI model and data as NFT), and they can also trade them as commodities, which makes good use of web3 can play value. It not only motivates AI developers and data providers, but also makes the application of AI more extensive. If a model is good enough, the owner has a stronger incentive to upload it to others for sharing.

At the same time, the decentralized AI marketplace may introduce some new business models, such as models, data sales and leasing, task crowdsourcing, etc.

1.3.3 Lower the threshold of AI application

Everyone should and will be able to train their own artificial intelligence models, which requires a platform with a sufficiently low threshold to provide resource support, including base models, tools, data, computing power, etc.

1.3.4 Demand and Supply

Although large models have powerful reasoning capabilities, they are not omnipotent. Often, fine-tuning for specific tasks and scenarios will achieve better results and have stronger practicability. Therefore, from the demand side, users need an AI model market to obtain useful models in different scenarios; for developers, they need a platform that can provide great resource convenience to develop models, and gain benefits through their own professional knowledge .

Second, model-based vs data-based

2.1 Model market

model

With tooling as the selling point, as the first link of the link, the project needs to attract enough model developers in the early stage to deploy high-quality models, so as to establish supply for the market.

In this mode, the main points that attract developers are the convenient and easy-to-use infra and tooling. The data depends on the developer's own ability and why some experienced people in a certain field can create value. The data in this field needs Developers collect and fine-tune models with better performance.

think

Recently, I have seen a lot of projects about the combination of AI marketplace and web3, but what I think is: Is creating a decentralized AI model market a false proposition?

First of all, we need to think about a question, what value can web3 provide?

It is far from enough if it is only the incentive of the token, or the narrative of the ownership of the model. From a practical point of view, high-quality models on the platform are the core of the entire product, and excellent models usually mean extremely high economic value. From the perspective of model providers, they need enough motivation to deploy their high-quality models to the AI marketplace, but can the incentives brought by token and ownership meet their expectations for the value of the model? For a platform that lacked a user base in the early days, it is obviously far from reaching. Without an extremely good model, the entire business model will not be established. So the question becomes how to make enough revenue for model providers in the absence of end users in the early stage.

2.2 Data Marketplace

model

Based on decentralized data collection, through the design of the incentive layer and the narrative of data ownership onboard more data providers, as well as users who label data. With the blessing of crypto, the platform has the opportunity to accumulate a large amount of valuable data within a certain period of time, especially the private domain data that is currently lacking.

What excites me the most is that this bottom-up development model is more like a crowdfunding game. No matter how experienced people are, it is impossible to have complete data in a field, and one of the values that web3 can provide is permissionless and decentralized data collection. This model can not only concentrate expertise and data in various fields, but also provide AI services to a larger user group. Compared with a single user's own data, these crowdfunding data are collected from a large number of real users' actual scenarios, so they can better reflect the complexity and diversity of the real world than data collected from a single source, which can greatly enhance The generalization ability and robustness of the model enable the AI model to perform well in many different environments.

For example, a person may have a lot of experience in nutrition and has accumulated a lot of data, but personal data alone is far from enough to train an excellent model. While users share data, they can also access and utilize valuable data contributed by other users in the same field and across the network on the platform, so as to achieve better fine-tuning effects.

think

From this perspective, it may also be a good attempt to build a decentralized data market. As a "commodity" with lower thresholds, shorter production links, and wider provider density, data can make better use of the value that web3 can provide. Incentive algorithm and data confirmation mechanism can provide motivation for users to upload data. Under the current model, data is more like a one-time commodity, meaning that it has little value after being used once. In the decentralized AI model market, user data can be used repeatedly and benefited, and the value of data will be realized in a longer period of time.

It seems to be a good choice to use data as an entry point to accumulate users. One of the core and barriers of the big model is high-quality and multi-dimensional data. After a large number of data providers are onboard, these people have the opportunity to further transform into end users or model provider. The AI marketplace based on this can indeed provide the underlying value for excellent models, and give algorithm engineers the motivation to contribute models on the platform from the perspective of training models.

This dynamic is a change from 0 to 1. Now that large companies have massive amounts of data, they can train more accurate models, which makes it difficult for small companies and individual developers to compete. Even if a user has very valuable data in a certain field, it is difficult for this small part of data to be of value without the cooperation of data on a larger set. However, in a decentralized market, everyone has the opportunity to obtain and use data, and these experts join the platform with valuable incremental data. Therefore, the data quality and quantity of the platform have been further improved, which is It makes it possible for everyone to train excellent models and even promote AI innovation.

Data itself is indeed well-suited to be a barrier to competition in this kind of AI marketplace. First of all, an excellent incentive layer and secure privacy guarantees allow more retail investors to participate in the entire protocol and contribute data. And, as the number of users increases, so does the quality and quantity of data. This will generate community and network effects, making the market able to provide greater value and wider dimensions, so it will be more attractive to new users. This is the process of establishing barriers for the market.

So fundamentally, to make a data-driven AI marketplace, the most important thing is the following 4 points:

1. Incentive layer: Design an algorithm that can effectively motivate users to provide high-quality data, and it is necessary to balance the strength of incentives and the sustainability of the market.
2. Privacy: Protect data privacy and ensure efficient use of data.
3. Users: Accumulate users quickly and collect more valuable data in the early stage.
4. Data quality: Data come from various sources, and effective quality control mechanisms need to be designed.

**Why is the provider of the model not listed as a key factor by me in this scenario? **

The main reason is based on the above four points, and it is logical to have an excellent model provider to join.

2.3 The value and challenges of the data market

Private Domain Data

The value of private domain data lies in its unique and difficult-to-obtain information in a specific domain, which is especially important for fine-tuning AI models. Using private domain data can create more accurate and personalized models that outperform models trained on public datasets in specific scenarios.

Now the construction process of the basic model can obtain a large amount of public data. Therefore, the focus of the web3 data market is not on these data. How to obtain and add private domain data during training is currently a bottleneck. By combining private domain data with public data sets, the adaptability of the model to diverse problems and user needs and the accuracy of the model can be increased.

For example, taking medical and health scenarios as an example, AI models using private domain data can usually increase the prediction accuracy by 10% to 30%. Referring to Stanford's research, the deep learning model using private domain medical data is 15% more accurate in predicting lung cancer than the model using public data

Data Privacy

Will privacy become the bottleneck restricting AI + Web3? Judging from the current development, the landing direction of AI in web3 has gradually become clear, but it seems that every application cannot avoid the topic of privacy. Decentralized computing power needs to ensure the integrity of data and models in both model training and model reasoning. Privacy; a condition for zkml to be established is to ensure that the model will not be abused by malicious nodes.

The AI marketplace is built on the basis of ensuring that users control their own data. Therefore, although user data is collected in a decentralized and distributed manner, all nodes should not directly collect, process, store, use, etc. access raw data. The current encryption methods are facing bottlenecks in the use, taking fully homomorphic encryption as an example:

1. Computational complexity: FHE is more complex than traditional encryption methods, which greatly increases the computational overhead of AI model training under fully homomorphic encryption, making model training extremely inefficient, even Not feasible. Therefore, fully homomorphic encryption is not ideal for tasks that require a lot of computing resources, such as deep learning model training.
2. Calculation error: During the calculation process of FHE, errors will gradually accumulate as the calculation proceeds, which will eventually affect the calculation results and affect the performance of the AI model.

Privacy is also divided into levels, no need to worry too much

Different types of data have different levels of privacy requirements. Only, for example, medical records, financial information, sensitive personal information, etc. require a high level of privacy protection.

Therefore, the diversity of data needs to be considered in the discussion of decentralized AI marketplace, the most important thing is balance. In order to maximize user participation and resource richness of the platform, it is necessary to design a more flexible strategy that allows users to customize the privacy level settings, not all data requires the highest level of privacy.

3. Reflections on the decentralized AI Marketplace

**3.1 Users have the right to control assets, will the withdrawal of users lead to the collapse of the platform? **

The advantage of the decentralized AI marketplace lies in the user's ownership of resources. Users can indeed withdraw their resources at any time, but once users and resources (models, data) accumulate to a certain extent, I think the platform will not be affected** . Of course, this also means that a lot of money will be spent on stabilizing users and resources in the initial stage of the project, which will be very difficult for a start-up team.

Community Consensus

Once the decentralized AI marketplace forms a strong network effect, more users and developers will become sticky. And because an increase in the number of users leads to an increase in the quality and quantity of data and models, making the market more mature. Users driven by different interests gain more value from the market. Although a small number of users may choose to leave, the growth rate of new users in this case will not slow down theoretically, and the market can continue to develop and provide greater value.

Incentives

If the incentive layer is properly designed, as the number of participants increases and various resources accumulate, the benefits obtained by all parties will increase accordingly. A decentralized AI marketplace not only provides a platform for users to trade data and models, but may also provide a mechanism for users to profit from their own data and models. For example, users get paid by selling their own data or by letting others use their own models.

For model developers: Deploying on other platforms may not have enough data to support finetune a model with better performance;

For data providers: Another platform may not have such a complete data foundation, and a small piece of data for users alone cannot exert value and obtain sufficient usage and benefits;

summary

Although in the decentralized AI marketplace, the project party only plays the role of matching and providing a platform, the real barrier lies in the accumulation of data and models brought by the accumulation of the number of users***. Users do have the freedom to withdraw from the market, but a mature AI Marketplace will often make them get more value from the market than they can get outside the market, so users have no incentive to withdraw from the market.

However, if most users or some high-quality model/data providers choose to withdraw, the market may be affected. This is also consistent with the existence of dynamic changes and adjustments in the entry and exit of users in various economic systems.

3.2 Which came first, the chicken or the egg

Judging from the above two paths, it is difficult to say which one will eventually come out, but it is clear that the data-based AI marketplace is more make sense, and the ceiling is much higher than the first one. The biggest difference is that the data-based market is constantly enriching barriers, and the process of accumulating users is also the process of accumulating data. In the end, the value endowed by web3 is to enrich a huge decentralized database. This is a positive cycle. . At the same time, in essence, this kind of platform does not need to retain data, but provides a lighter market for contributing data. In the end, this is a large data mart, and this kind of barrier is difficult to replace.

From the perspective of supply and demand, an AI marketplace needs to have two points at the same time:

1. Lots of excellent models
2. END USER

From a certain point of view, these two conditions seem to be interdependent. On the one hand, the platform needs to have enough users to provide motivation for the providers of models and data. Only when there are enough users accumulated can the incentive layer play a role. For the greatest value, the flywheel of data can also be turned, so that more model providers can deploy models. On the other hand, enough end users must come for a useful model, and the user's choice of the platform is largely a choice of the quality and capabilities of the platform model. Therefore, without accumulating a certain number of excellent models, this demand does not exist. No matter how advanced the routing algorithm is, routing without good models is empty talk. This is like the premise of the apple store is that apple

Therefore, a better development idea is:

Initial Strategy

• **Accumulate high-quality models, **The most noteworthy thing in the initial stage is to build a high-quality model library. The reason is that no matter how many end users there are, without high-quality models for them to choose and use, the platform will not be attractive, and users will not have stickiness and retention. By focusing on building a high-quality model library, the platform can ensure that early users can find the models they need, thereby building brand reputation and user trust, and gradually building community and network effects.

Expansion Policy

• Attract end users, After building a high-quality model library, turn to attract and retain more end users. A large number of users will provide sufficient motivation and interest for model developers to continue to provide and improve the model. In addition, a large number of users will also generate a large amount of data, which will further improve the training and optimization of the model.

summary

What is the best attempt at an AI marketplace? *** In a word, the platform can provide enough high-quality models, and can efficiently match users with suitable models to solve problems ***. This sentence solves two contradictions. First, the platform can provide enough value for developers (including model developers and users), so that there are enough high-quality models on the platform; second, these "commodities" can provide users with efficient local solutions, thereby accumulating more users and providing protection for the interests of all parties.

Decentralized AI Marketplace is an easy direction for AI + web3 to land, but a project must figure out what the real value this platform can provide and how to onboard a large number of users in the early stage. Among them, the key is to find a balance point of the interests of all parties, and at the same time handle multiple elements such as data ownership, model quality, user privacy, computing power, and incentive algorithms, and finally become a sharing and trading platform for data, models, and computing power.