"Web3 Carnival" short essay collection: What are people talking about? Which track is more popular?

2024 Hong Kong Web3 Blockchain Week was held at the Hong Kong Convention and Exhibition Center from April 6 to 9. The event covered 4 venues and discussed the core issues of each track of Web3. In the past few days, the world's smartest and most innovative minds have gathered in Hong Kong to share and discuss the latest Web3 technical solutions. (For the specific event process, please read: "Event Express丨 2024 Hong Kong Web3 Carnival will open on April 6, and the surrounding activities will be summarized" .) At the same time, many well-known investment institutions and crypto KOLs also participated. The event in Hong Kong is coming to an end. BlockBeats editor has compiled several "small essays" of the Hong Kong Web3 Carnival for everyone to see their latest discoveries and insights, so that you will not miss the cutting-edge hot spots even if you don't go to the scene.

Fred Dacong (crypto investor):

After coming to Hong Kong recently, I talked with some investors and found that everyone is paying attention to the AI and DePIN tracks. Share your recent thoughts on AI and DePIN, and discuss with everyone.

Four core questions:

1. Why do most decentralized computing power projects choose to do AI reasoning instead of AI training?

2. What is NVIDIA's strength? What is the reason for the difficulty of decentralized computing power training?

3. What will be the end of decentralized computing power (Render, Akash, etc.)?

4. What will be the end of decentralized algorithms (Bittensor)?

Next, let's unravel the mystery layer by layer:

1) Looking at this track, except for Gensyn, most decentralized computing power projects choose to do AI reasoning rather than training. The core lies in the different requirements for computing power and bandwidth.

To help everyone understand it simply, let's compare AI to a student:

- AI training: If we compare artificial intelligence to a student, then training is similar to providing artificial intelligence with a lot of knowledge, and examples can also be understood as what we often call data. Artificial intelligence learns from these knowledge examples. Since the nature of learning requires understanding and memorizing a lot of information, this process requires a lot of computing power and time.

- AI reasoning: So what is reasoning? It can be understood as using the knowledge learned to solve problems or take exams. In the reasoning stage, artificial intelligence uses the knowledge learned to answer, rather than activating new knowledge, so the amount of calculation required in the reasoning process is relatively small.

It is easy to find that the difference in difficulty between the two is essentially that large-model AI training requires a huge amount of data, and the bandwidth required for high-speed data communication is extremely high, so it is currently very difficult to implement decentralized computing power for training. While reasoning requires much less data and bandwidth, it is more likely to be implemented.

2) So where are the data and bandwidth bottlenecks? Why is decentralized training difficult to achieve?

This involves two key elements of large-model training: single-card computing power and multi-card parallel connection.

- Single-card computing power: At present, all centers that need to train large models are called supercomputing centers. To make it easier for everyone to understand, we can use the human body as an analogy. The supercomputing center is the tissue of the human body, and the underlying unit GPU is the cell. If the computing power of a single cell (GPU) is very strong, then the overall computing power (single cell × number) may also be very strong.

- Multi-card parallel connection: The training of a large model is often hundreds of billions of GB. For a supercomputing center that trains large models, at least 10,000 A100s are required. Therefore, it is necessary to mobilize these tens of thousands of cards for training. However, the training of large models is not a simple series connection. It is not training on the first A100 card and then on the second card. Instead, different parts of the model are trained on different graphics cards. When training A, the results of B may be needed, so multi-card parallel connection is involved.

Why is NVIDIA so powerful, and its market value has soared, while AMD and domestic Huawei and Horizon are currently difficult to catch up. The core is not the computing power of a single card itself, but lies in two aspects: CUDA software environment and NVLink multi-card communication.

- On the one hand, it is very important to have a software ecosystem that can adapt to the hardware, such as NVIDIA's CUDA system, and it is difficult to build a new system, just like building a new language, the replacement cost is very high.

- On the other hand, it is multi-card communication. In essence, the transmission between multiple cards is the input and output of information. How to connect in parallel and how to transmit. Because of the existence of NVLink, there is no way to connect NVIDIA and AMD cards; in addition, NVLink will limit the physical distance between graphics cards, requiring the graphics cards to be in the same supercomputing center, which makes it difficult to achieve decentralized computing power if it is distributed around the world.

The first point explains why it is difficult for AMD and domestic Huawei and Horizon to catch up; the second point explains why decentralized training is difficult to achieve.

3) What will be the end of decentralized computing power?

- Decentralized computing power is currently difficult to train large models. The core is that large model training is most concerned with stability. If the training is interrupted, it needs to be retrained, and the sunk cost is very high. It has very high requirements for multi-card parallel connection, and the bandwidth is limited by physical distance. NVIDIA uses NVLink to achieve multi-card communication. However, in a supercomputing center, NVLink will limit the physical distance between graphics cards, so the dispersed computing power cannot form a computing power cluster to train large models.

- But on the other hand, the demand for relatively low computing power requirements can be realized, such as AI reasoning, or some specific scenarios of vertical small and medium-sized model training is possible. When there are some relatively large node service providers in the decentralized computing power network, there is potential to serve these relatively large computing power requirements. And edge computing scenarios such as rendering are also relatively easy to implement.

4) What will be the end of the decentralized algorithm model?

The end of the decentralized algorithm model depends on the end of the future AI. I think the future AI war may have 1-2 closed source model giants (such as ChatGPT), plus a hundred flowers blooming models. In this context, application layer products do not need to be bound to a large model, but cooperate with multiple large models. In this context, Bittensor's model has great potential.

Original link

Dov Wo (crypto investor):

VC perspective on the Hong Kong conference essay(full of bias):

1. Bull market products are not important, narratives and emotions are more important. Meme has repeatedly proved this. At least half of the financing money is used to pull the market.

2. The most core abilities in this round of bull market: trading ability and shouting ability. If you don’t have trading ability, please practice shouting ability. I am also helping some project parties find KOLs for promotion. Welcome to take the business orders I give. There are many projects that have been invested by binance and OKX but cannot find suitable domestic/overseas promotion channels. False prosperity/buying is just a rigid demand.

3. VC’s status is getting worse and worse, with slow unlocking and high valuations, and retail investors don’t recognize it. If you can’t bring additional resources to the project party (such as exchange relations, resources and communities in a specific region, economic model design, promotion capabilities, etc.), most pure financial investment VCs can only be a big complaint.

Take a project that recently came to me as an example. The valuation of the KOL round is lower than that of VC, and the unlocking is better than VC. (This project is led by a top VC in the West)

KOL > VC is not groundless, it really exists. So I am also running a small boutique KOL agency to sell shovels.

4. The most popular activities are Berachain, Solana, and BTC; the ETH ecosystem is relatively inactive, and a big meme is Layer 69 (Solana's own funny ETH spoof video)

5. There are a lot of people at the Side Event, and there may be hundreds of Side Events at this event. On the contrary, there are not many people at the main venue; the project parties may need to reflect on whether the expensive main venue booth is still worth it; for example, Berachain and Solana do not have a main venue booth, but hold their own events. OKX has all booths and activities, and is in the top status, while Binance continues to have no events.

6. Ordinary people must become KOLs and find their own positioning. For example, if you want to shout "Tugou", just shout "Tugou". If you want to do good analysis, just do good analysis. You can get good business orders. I asked KOLs who shouted "Tugou" to shout orders, and the price of each order ranged from 200U-2000U. VCs who can't shout orders can't survive, and KOLs who can't shout orders can't make money. If you have traffic, a large number of project parties will actively look for you to cooperate, and you will have better information and more resources. Please read the second point again. Question the teacher who leads orders, understand the teacher who leads orders, and become a teacher who leads orders.

7. This year, HK talked less about compliance and licenses, and talked more about trading and accumulating resources (it may also be because I have reached a higher stage and can understand and participate). The largest booths at the main venue are OKX and DWF. Everyone knows who has resources and who has money.

ABCDE Lianchuang Du Jun commented on the tweet, saying that "it will be difficult for institutions without core investment and research capabilities."

Original link

Ethan yu (AC Capital partner):

Drank for 2 days and listened to some awesome talk:

1. Chinese institution Fomo Bitcoin ecology and Restaking, European and American/overseas institutions Fomo Solana DePin, a small number of institutions Fomo Cosmos ecology-related Bearchain, some institutions were trapped in the last round GameFi, so now we can only continue to push it hard, but all of them will end up with a big cut in the end, and the valuation is simply outrageous.

2. Almost all of the consensus is that there are too many gold dogs on Solana, and the get-rich-quick effect has been continuing. In the past few days, the local dog turnover that went to Base has returned to Solana.

3. It’s not that KOLs play a greater role than VCs, but because KOLs have more flexible funds and can shout orders and delete pushes at will, and the unlocking method is also more friendly. What do you do if the official push of the institution shouts orders? On the other hand, the institution is not stupid and rich. You start with a valuation of 100 million US dollars, and it will take 3 years and 36 months to fully exit. You don’t go to Binance or OKX. Let me invest 500,000 US dollars now. Are you stupid or the institution stupid?

4. There are indeed many newcomers and people who want to enter the industry. Young people are proactive, enthusiastic, and eager to learn. They have no industry baggage. They are optimistic about the all-in and just do it. However, there are not enough newcomers, so the bull market is still in its early stages.

Original link

If you also attended this blockchain week and have some insights, please submit your article to: [email protected] .

欢迎加入律动 BlockBeats 官方社群：

Telegram 订阅群： https://t.me/theblockbeats

Telegram 交流群： https://t.me/BlockBeats_App

Twitter 官方账号： https://twitter.com/BlockBeatsAsia