「Web3 Carnival」Short Essay Collection: What is everyone talking about? Which track is more popular?

The 2024 Hong Kong Web3 Blockchain Week will be held at the Hong Kong Convention and Exhibition Center from April 6th to 9th. This event covers 4 venues and focuses on the core topics of various Web3 tracks. Over the past few days, the world's smartest and most innovative minds have gathered in Hong Kong to share and discuss the latest technical solutions in Web3. (For specific event details, please read: "Event Express | 2024 Hong Kong Web3 Carnival will open on April 6th, summary of surrounding activities".) At the same time, many well-known investment institutions and crypto KOLs are also participating. As the event in Hong Kong is coming to a close, BlockBeats has compiled a few "short essays" on the Hong Kong Web3 Carnival for everyone to see their latest discoveries and insights, so you don't miss out on the cutting-edge hot topics.

Big Onion Fred (Crypto Investor):

After arriving in Hong Kong recently, I chatted with some fellow investors and found that there is quite a high level of interest in the AI and DePIN tracks. Let's share some recent thoughts on AI and DePIN and discuss them together.

Core discussion of four questions:

1. Why do most decentralized computing power projects choose to do AI inference rather than AI training?

2. What makes NVIDIA so outstanding? What are the challenges in decentralized computing power training?

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

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

Next, let's peel back the layers:

1) Looking at this track, apart from Gensyn, most decentralized computing power projects choose to do AI inference rather than training, mainly due to different requirements for computing power and bandwidth.

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

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

- AI Inference: So what is inference? It can be understood as using the acquired knowledge to solve problems or take exams. In the inference stage, artificial intelligence uses the learned knowledge to answer questions, rather than acquiring new knowledge. Therefore, the computational requirements during inference are much lower.

It is easy to see that the fundamental difference in difficulty between the two lies in the fact that training large model AI requires a huge amount of data and high bandwidth requirements for high-speed communication of data. Therefore, the current difficulty in using decentralized computing power for training is extremely high. In contrast, inference requires much less data and bandwidth, making it more achievable.

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

This involves two key elements of training large models: single-card computing power and multi-card parallelism.

- Single-card computing power: Currently, all supercomputing centers that require training large models, which we call supercomputing centers. For easy understanding, we can use the human body as an analogy, where the supercomputing center is like the body's organization, and the basic unit GPU is like a cell. If the computing power of a single cell (GPU) is strong, then the overall computing power (single cell × quantity) may also be strong.

- Multi-card parallelism: Training a large model often involves hundreds of billions of GB. For supercomputing centers that train large models, they need at least tens of thousands of A100s as a foundation. Therefore, it is necessary to mobilize these tens of thousands of cards for training. However, training a large model is not simply serial, it is not training on the first A100 card and then on the second card, but different parts of the model are trained on different graphics cards, where training A may require the result of B, thus involving multi-card parallelism.

Why is NVIDIA so powerful, with its market value soaring, while AMD and domestic companies like Huawei and Horizon are currently struggling to catch up? The core lies not in the single-card computing power itself, but in two aspects: the CUDA software environment and NVLink multi-card communication.

- On one hand, having a software ecosystem that can adapt to hardware is crucial, such as NVIDIA's CUDA system, and building a new system is difficult, akin to creating a new language, with very high replacement costs.

- On the other hand, it is about multi-card communication. Essentially, the transmission between multiple cards is the input and output of information, how to parallelize, how to transmit. Because of the existence of NVLink, it is impossible to connect NVIDIA and AMD cards; furthermore, NVLink will limit the physical distance between cards, requiring the cards to be in the same supercomputing center, making it difficult for decentralized computing power to be distributed globally.

The first point explains why AMD and domestic companies like Huawei and Horizon are currently struggling to catch up; the second point explains why decentralized training is difficult to achieve.

3) What will the endgame of decentralized computing power look like?

- Decentralized computing power is currently difficult to train large models, mainly due to the emphasis on stability in training large models. If training is interrupted, it needs to be retrained, incurring high sunk costs. Its requirements for multi-card parallelism are high, and bandwidth is limited by physical distance. NVIDIA uses NVLink to achieve multi-card communication, but within a supercomputing center, NVLink restricts the physical distance between cards, making it difficult for dispersed computing power to form a computing cluster for training large models.

- However, on the other hand, demands with relatively low computing power requirements are achievable, such as AI inference, or training of small to medium-sized models in specific scenarios. When there are relatively large node service providers in the decentralized computing power network, there is potential to serve these relatively large computing power needs. Additionally, scenarios like rendering in edge computing are relatively easy to achieve.

4) What will the endgame of decentralized algorithm models look like?

The endgame of decentralized algorithm models depends on the endgame of future AI. I believe the future AI battle may involve 1-2 closed-source model giants (such as ChatGPT), along with a variety of models. In this context, application-layer products do not need to be tied to a single large model, but can collaborate with multiple large models. In this context, the potential for models like Bittensor is still significant.

Original article link

Dov Wo (Crypto Investor):

VC Perspective on Hong Kong Conference (Full of Bias):

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

2. The most crucial abilities in this bull market: trading skills and signal-calling abilities. If you lack trading skills, please practice signal-calling abilities. I am also helping some projects find KOLs for promotion. Many projects invested by Binance, OKX, etc., cannot find suitable domestic/overseas promotion channels, false prosperity/buying pressure is a must.

3. The status of VCs is deteriorating, slow unlocking, high valuation, retail investors do not recognize. If VCs cannot bring additional resources to projects (such as exchange relationships, resources and communities in specific regions, economic model design, promotion capabilities, etc.), most pure financial VCs can only be big complainers.

Take a recent example of a project that approached me, the valuation of the round given to KOLs is lower than that of VCs, and the unlocking is better than VCs. (This project is led by a top VC in the West)

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

4. The hottest activities are Berachain, Solana, BTC; the ETH ecosystem is relatively less active, with a big meme being Layer 69 (Solana's quirky parody of ETH in a video).

5. There are a lot of people at the Side Event, with possibly hundreds of Side Events this time. On the contrary, there are not many people at the main venue; project parties may need to rethink whether the expensive main venue booth is still worth it; for example, Berachain and Solana do not have main venue booths, but organize their own events. OKX has all booth activities, in the top state, while Binance continues to have an event without one.

6. Ordinary people must become Key Opinion Leaders (KOLs), find their own positioning, for example, if you shout about Dogecoin, just focus on that, if you are good at analysis, then focus on that. You can all receive good business opportunities. I found KOLs who shout about Dogecoin to promote, with prices ranging from 200U to 2000U per post. VCs who don't promote won't survive, and KOLs who don't promote won't make money. If you have traffic, many project parties will actively seek to cooperate with you, providing you with better information and more resources. Please reread the second point. Question the promotion teacher, understand the promotion teacher, become the promotion teacher.

7. This year, there are fewer discussions about regulations and licenses in Hong Kong, and more discussions about trading and accumulating resources (possibly because I have reached a higher stage and can understand and participate). When you enter the main venue, the largest booths are from OKX and DWF, everyone knows who has resources and who has money.

ABCDE co-founder Du Jun commented on this post, stating that "institutions without core research capabilities will find it very difficult."

Original post link

Ethan Yu (AC Capital Partner):

After drinking for 2 days, I heard some amazing things:

1. Chinese institutions Fomo on the Bitcoin ecosystem and Restaking, while institutions with European/American genes Fomo on DePin on Solana, and some institutions are focusing on the Cosmos ecosystem-related Berachain. Some institutions are stuck in GameFi from the last round, so they can only continue to push hard, but in the end, they will all end up the same way and have to make significant cuts, the valuations are simply outrageous.

2. Almost everyone agrees that there are too many "golden dogs" on Solana, the effect of getting rich quick continues, and these days the flow of funds from Dogecoin on Base has returned to Solana.

3. The impact of KOLs is greater than VCs, not because KOLs have more flexible funds and can easily promote or delete posts, but also because their unlocking methods are more user-friendly, what will you do when institutional officials promote and you don't? On the other hand, institutions are not foolish with money, if you come in with a valuation of 100 million USD right from the start, and it takes a full 36 months to fully exit, and you're not on Binance or OKX, would I be foolish to invest 500,000 USD now, or is it the institution that's foolish?

4. There are indeed many newcomers and those who want to enter the industry, young people are enthusiastic and eager to learn, without the burden of the industry, they just go for it if they see potential, but there are not enough newcomers yet, so the bull market is still in its early stages.

Original post link

If you also attended this blockchain week and have insights, feel free to submit them to: [email protected].

Join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia