AI+Web3 is likely to be one of the hot spots in the next bull market. What are the specific trends and opportunities?

The hottest AI at the moment is seen as the key point and core of the Fourth Industrial Revolution, while the last hot concept in the technology world is Web3, which is seen as the key core of the next generation of the Internet.

AI and Web3 are two major concepts that will set off a technological revolution. If they can be combined, what kind of "surprises" might they bring us?

01 Let’s talk about AI itself first

The AI ​​industry was actually going to be dead. You all know Yilong, the founder of Near, right? He used to work in AI and was the main code contributor to TensorFlow (the most popular machine learning framework). Everyone speculated that he saw no hope in AI (machine learning before big models) so he came to work in Web3.

Finally, at the end of last year, the industry welcomed ChatGpt3.5, and the industry suddenly came alive again, because this time it can be regarded as a qualitative change, rather than the previous waves of hype and quantitative changes. A few months later, the wave of AI entrepreneurship also spread to our Web3. Silicon Valley Web2 is very competitive, various capital Fomo, various homogeneous solutions began to compete in price wars, various large manufacturers and large models PK...

However, it should be noted that after more than half a year of explosive growth, AI has also entered a relatively bottleneck period. For example, Google's search popularity for AI has dropped drastically, the growth rate of Chatgpt users has slowed down significantly, and AI Output has a certain degree of randomness, which limits many landing scenarios... In short, we are still very, very far away from the legendary "AGI - general artificial intelligence".

At present, the Silicon Valley venture capital circle has the following judgments on the next development of AI:

1) There are no vertical models, only big models + vertical applications (we will mention this later when we talk about Web3+AI)

2) Data from edge devices such as mobile phones may be a barrier, but AI based on edge devices may also be an opportunity

3) The length of context may lead to qualitative changes in the future (vector databases are currently used as AI memory, but the context length is still not enough)

02Web3+AI

AI and Web3 are actually two completely different fields. AI requires concentrated computing power + massive data for training, which is a very centralized thing. Web3 focuses on decentralization, so it is not so easy to combine them. However, the argument that AI changes productivity and blockchain changes production relations is too deeply rooted in people's minds, so there will always be people who continue to look for that combination point. In the past two months, we have talked about no less than 10 AI projects.

Before talking about the new combined track, let’s talk about the old AI+Web3 projects, which are basically platform-based, represented by FET and AGIX. How should I put it? My domestic AI professional friend told me this - "Those who worked on AI in the past are basically useless now, whether it is Web2 or Web3, many of them are burdens rather than experience. The direction and future are big models based on Transformer like OpenAI, which saved AI", you can judge for yourself.

Therefore, the general platform model is not the Web3+AI model he is optimistic about. The more than 10 projects I talked about do not have this aspect. Currently, the basic tracks I see are as follows:

1. Bot/Agent/Assistant model assetization

2. Computing power platform

3. Data Platform

4. Generative AI

5. Defi transactions/audits/risk control

6.ZKML

1. Bot/Agent/Assistant model assetization

The assetization track of Bot/Agent/Assistant is the track that has been discussed the most and is the track with the most serious homogeneity. Simply put, most of these projects use OpenAI as the underlying layer, combined with other open source/self-developed technical means, such as TTS (Text to Speech), and add specific data, so that FineTune can produce some robots that are "better than ChatGPT in a certain field."

For example, you can train a beautiful teacher to teach you English. You can choose whether she has an American accent or a London accent. You can also adjust her personality and chat style. Compared with the more mechanical and official answers of ChatGPT, the interactive experience will be better. There was a DAPP and Web3 female-oriented game with a virtual boyfriend in the circle, called HIM, which can be regarded as a representative of this type.

Based on this idea, you can theoretically have many Bots/Agents to serve you. For example, if you want to cook boiled fish, there may be a Cooking Bot called Fine Tune that teaches you how to cook it, and the answers it gives are more professional than ChatGPT. If you want to travel, there is also a Travel Assistant Bot that provides you with various travel suggestions and plans. Or if you are a project owner, you can create a Discord customer service robot to help you answer community questions.

In addition to this kind of "GPT-based vertical application" Bot, there are also derivative projects based on it, such as Bot "model assetization". It is a bit like NFT "small picture assetization", so can the popular prompts in AI now also be assetized? For example, different prompts in MidJourney can generate different pictures, and different prompts will have different effects when training Bots, so Promopt itself has value and can also be assetized.

There are also projects like portal indexing and searching based on such bots. When we have thousands of bots, how can we find the most suitable bot for you? Maybe by then we will need a portal like Hao123 in the Web2 world, or a search engine like Google to help you "locate".

In my opinion, Bot (model) assetization has two drawbacks and two directions at this stage:

1) Disadvantages

Disadvantage 1- Too much homogeneity, because this is the AI+web3 track that users can understand most easily, and it is a bit like NFT with some utility attributes. Therefore, the primary market is beginning to show a red ocean trend, and it is rolling up, but the bottom layer is all OpenAI, so everyone actually has no technical barriers, and can only compete in design and operation;

Disadvantage 2- Sometimes, things like putting Starbucks membership card NFT on the chain, although it is a good attempt to break out of the circle, for most users, it may not be as convenient as a physical or electronic membership card. Bots based on Web3 also have this problem. If I want to learn English with a robot or chat with Musk, Socrates, or someone else, wouldn’t it be better for me to use Web2’s http://Character.AI?

2) Direction

Direction 1 is in the near + medium term. Putting models on the chain may be an idea. Currently, these models are a bit like ETH NFT small pictures. MetaData mostly points to off-chain servers or IPFS, rather than pure on-chain. Models are usually tens to hundreds of megabytes in size, and they have to be thrown on the server.

However, with the recent rapid decline in storage prices (2TB SSD is 500 RMB) and the advancement of storage projects such as Filecoin FVM and ETH Storage, I believe that it should not be difficult to put a 100-megabyte model on the chain in the next two to three years.

You may ask what are the benefits of chaining? Once the model is chained, it can be directly called by other contracts, which is more Crypto Native. There are definitely more tricks that can be played, and it feels like a fully onchain game, because all data is chain native. At present, I see that there are teams exploring this aspect, but it is still in a very early stage.

Direction 2- is the medium + long term. If you think about smart contracts seriously, they are actually best suited for "machine-machine interaction" rather than human-computer interaction. AI now has the concept of AutoGPT, which can create a "virtual avatar" or "virtual assistant" for you that can not only chat with you, but also help you perform tasks according to your requirements, such as helping you book flights, hotels, buy domain names and build websites...

Do you think it is more convenient for an AI assistant to operate your various bank accounts, Alipay, etc., or to transfer money through a blockchain address? The answer is obvious. In the future, will there be a bunch of AI assistants integrated with AutoGPT, which can automatically perform C2C, B2C, and even B2B payments and settlements through blockchain and smart contracts in various task scenarios? At that time, the boundary between Web2 and Web3 will become very blurred.

2. Computing power platform

The projects of computing power platforms are not as numerous and volatile as the assetization of Bot models, but they are relatively easier to understand. We all know that AI requires a lot of computing power, and BTC and ETH have proven in the past 10 years that there is a way in the world to spontaneously and decentralizedly organize and coordinate massive computing power to cooperate and compete to do one thing under the environment of economic incentives and games. Now this method can be used on AI.

The two most famous projects in the industry are undoubtedly Together and Gensyn. One raised tens of millions in seed round financing, and the other raised 43 million in Series A round. The reason why these two raised so much money is said to be because they need funds and computing power to train their own models first, and then build a computing power platform to provide to other AI projects for training.

The amount of financing for computing power platforms that do inference is relatively much smaller, because in essence, they aggregate idle GPU computing power and provide it to AI projects in need for inference. RNDR aggregates rendering computing power, and these platforms aggregate inference computing power. However, the technical threshold is currently vague, and I even wonder if RNDR or Web3 cloud computing power platform will one day extend a foot to the inference computing power platform.

The computing power platform is more realistic and predictable than model assetization. It is basically a foregone conclusion that there will be demand and one or two leading projects. It depends on who can come out on top. The only uncertainty at present is whether there will be leaders in training and reasoning, or whether the leaders will take over both training and reasoning.

3. Data Platform

This is actually not difficult to understand, because the underlying components of AI are basically three major components: algorithms (models), computing power, and data.

Since both algorithms and computing power have "decentralized versions", data will certainly not be absent. This is also the direction that Dr. Lu Qi, the founder of Qi Ji Chuangtan, is most optimistic about when talking about AI and Web3.

Web3 has always emphasized data privacy and sovereignty, and also has technologies such as ZK to ensure data reliability and integrity. Therefore, AI trained based on Web3 on-chain data must be different from that trained based on Web2 off-chain data. So this line makes sense as a whole. Currently, Ocean should be considered as this track in the circle. There are also projects such as special AI data markets based on Ocean in the primary market.

4. Generative AI

Simply put, it is to use AI to draw pictures, or similar creations, to serve other scenarios. For example, NFT, or in-game map generation, NPC background generation, etc. I feel that the NFT line is more difficult because AI generation is not scarce enough. Gamefi is one way, and I have seen teams trying it in the primary market.

However, I saw a news a few days ago that Unity (which has dominated the game engine market for many years together with Unreal Engine) has also released its own AI generation tools Sentis and Muse, which are still in the closed beta stage and are expected to be officially launched next year. How should I put it? I feel that the game AIGC projects in the Web3 circle may be hit by Unity by then...

5. DeFi Transaction/Audit/Yield/Risk Control

We have seen projects trying these categories, and homogeneity is relatively unnoticeable.

1) DeFi trading - This is a bit tricky, because if a trading strategy works well, as more and more people use it, the strategy may gradually become less useful, and you have to switch to a new strategy. I am also curious about the future winning rate of AI trading robots and where they will rank among ordinary traders.

2) Audit - Visual inspection should be able to help quickly review and resolve existing common vulnerabilities, but it should not be able to detect new or logical vulnerabilities that have never appeared. This will only be possible in the AGI era.

3) Yield and risk control - Yield is not difficult to understand. Just imagine it as a YFI with AI intelligence. Throw money into it, and AI will find a platform for staking, LP, mining, etc. based on your risk preference. As for risk control, it feels strange to make it a separate project. It feels more sensible to serve various lending or similar Defi platforms in the form of plug-ins.

6.ZKML

This is a track that is becoming increasingly popular in the industry because it combines two of the most cutting-edge technologies, one is ZK in the industry, and the other is ML (Mechine Learning, a narrow branch of AI) outside the industry.

Theoretically, the combination with ZK can provide privacy, integrity and accuracy for ML, but if you insist on saying what the specific usage scenarios are, in fact, many project parties can’t think of them, and they will build the infrastructure first... At present, the only real demand is that machine learning in some medical fields does have the privacy needs of patient data. As for the narrative about on-chain game integrity or anti-cheating, it always feels a bit far-fetched.

There are only a few star projects in this field, such as Modulus Labs, EZKL, Giza, etc., which are all popular in the primary market. There is no way, because there are only a few people in the world who understand ZK, and even fewer people who understand ZK and ML at the same time, so the technical threshold of this field is much higher than others, and the homogeneity is relatively not obvious. Finally, ZKML is mostly aimed at reasoning, not training.