Recently, with the listing of AI-based tokens such as io.net and Aethir on top exchanges such as Binance and OKX, more and more projects combining with AI have emerged in Web3. Can Web3 use its own technological advantages to promote the AI revolution and subvert the industry? What are the application scenarios that can be implemented in the process of combining Web3 with AI? The author will share with you the AI applications in Web3 through a series of articles.
1. Competition in AI Big Models
The competition for large AI models mainly focuses on three aspects: computing power, algorithms, and data.
If traditional technology giants want to enter the AI big model field, they must first consider the high training and debugging costs of the big models in the early stage. Tian Qi, chief scientist of Huawei Cloud's artificial intelligence field, mentioned in his speech at the AI Big Model Technology Summit that the single cost of big model development and training is as high as 12 million US dollars. OpenAI CEO Sam Altman also mentioned that the training cost of GPT-4 exceeds 100 million US dollars, of which GPU computing power costs account for the majority of the entire training cost.
Not long ago, Musk's AI-focused startup xAI raised $6 billion in its latest round of financing, and claimed that in the future xAI will mainly use the funds raised to expand GPU clusters. Meta CEO Zuckerberg also announced that he would invest up to $30 billion to purchase 350,000 NVIDIA H100 GPUs to develop its artificial general intelligence (AGI). The competition among technology giants on large AI models has become an arms race for GPUs. Whoever has enough funds to purchase more computing power will have a first-mover advantage.
At the core algorithm level, the industry's technological openness is relatively high, and many domestic companies basically build algorithms based on the GPT2 open source code and GPT3/3.5 paper architecture.
However, at the data level, due to the differences in their own business scenarios, the accumulated training data of each company is significantly different, so data will become the core competitive factor of AI products. For technology giants with large amounts of multi-dimensional business data, they have obvious advantages in model training and optimization. After the early products are launched, they rely on the feedback from their own user groups and product interactions to tune the models, thereby generating a data flywheel effect and continuously optimizing and iterating, which will become the moat of the product in the future.
2. Web3 Sora-Livepeer
1. Introduction to Livepeer
Video is the most frequently used form of media display on the Internet, and it consumes 80% of the Internet bandwidth. Videos need to be transcoded before they are distributed on the Internet. Transcoding is the process of reformatting the original video file. After transcoding, the video can be adapted to end users with different bandwidths and devices.
But video streaming is extremely expensive, especially for real-time video streaming. If a video startup uses cloud services such as Amazon to transcode each video, it costs $3, which means that for a media server, the cost of video transcoding per month is as high as $4,500.
To solve the high cost of video streaming and transcoding raw video files, Livepeer has developed a more scalable and cost-effective video streaming solution. Founded in 2017, Livepeer is a protocol built for decentralized video streaming, and the project was originally a P2P infrastructure running on Ethereum. Essentially, the Livepeer network transcodes video files into the corresponding format in a decentralized manner and reduces costs by up to 50 times.
2. Business architecture
Livepeer's ecological participants mainly include developers, users, and video media.
Developers can use Livepeer to enhance video capabilities when building real-time or on-demand video applications; users can use Livepeer-based applications to stream videos, games, live broadcasts, and other content in real time; and some video media can use Livepeer to reduce the cost of live video broadcasts.
During the operation of the Livepeer network, there are two roles: Orchestrator and Delegators to ensure the quality of video tasks.
Anyone can become a coordinator by installing the livepeer node software. The coordinator uses the GPU, CPU, bandwidth and other resources of its own device to transcode and distribute videos. Users of the Livepeer network will pay a certain fee, part of which will be paid to the coordinator in the form of ETH or LPT.
Livepeer network token holders can stake LPT tokens to their recognized coordinators to become delegators, thereby obtaining benefits from the protocol through staking.
Livepeer protocol system workflow
3. Team and financing
core team:
Doug Petkanics, founder and CEO of Livepeer, has a computer science degree from the University of Pennsylvania and was previously co-founder and VP of engineering at startup Wildcard and co-founder of startup Hyperpublic, which was acquired by Groupon in 2012.
Eric Tang, co-founder and core developer of Livepeer, previously founded Wildcard and is also the chairman of the Carolyn Faye Kramer Fan Club.
Financing History:
Livepeer's investors include well-known institutions such as Tiger Global, DCG, Coinbase Ventures, and angel investor Alan Howard. Alan Howard is the co-founder of Brevan Howard, a famous European macro hedge fund. He was keenly aware of the risks of subprime loans in 2008 and still made a profit of more than 20% during the 2008 financial crisis. In 2020, he again made an accurate bet on the macro of the epidemic and made a huge profit of 100%. He is known as the "God of Macro".
Data from cryptorank.io
4. Token Economics
The Livepeer protocol token LPT was issued on Ethereum on May 1, 2018, and subsequently migrated to ARB. The total issuance was 24,918,514, of which about 12.35% were held by the founding team and unlocked after three years. About 19% of the tokens were sold in the pre-sale, while 63.437% were allocated to public trading. 0.213% were issued to several early advisors and contributors who helped Livepeer get started, and 5% were used to maintain and develop the Livepeer platform. Currently, all the tokens issued in the early stage have been circulated, with a total circulation of 33,005,680.
Data from cryptorank.io
LPT adopts an inflationary economic model, and its inflation rate is not fixed, but depends on the pledge rate of the Livepeer network. The pledge rate here refers to the proportion of the number of tokens pledged by the delegators to the coordinators to the circulation.
The Livepeer protocol assumes that the staking rate of 50% is the balance point between network security and token liquidity.
When the pledge rate is lower than 50%, the LPT inflation rate increases by 0.00005%.
Conversely, when the pledge rate exceeds 50%, the LPT inflation rate will drop by 0.00005%.
The issuance of inflation tokens is based on the ETH block cycle. At the beginning, each round of LPT token issuance was 5760 Ethereum blocks. With the merger and upgrade of ETH, each round became 6377 Ethereum blocks. The average block time of ETH is 12 seconds, so currently a round of LPT issuance is carried out in about 21 hours, and the number of LPT issuances in each round is 16090.
Summary: The lower the pledge rate, the higher the inflation rate of LPT. Delegators will pledge more and more LPT, which is beneficial to the price of LPT tokens. The project party and the community Live Pioneers are also encouraging ecological participants to pledge tokens to reduce the inflation rate.
3. Livepeer AI New Narrative
With the launch of Open AI's video generation model Sora, the threshold for video creation has been lowered to just entering text prompts. The top open source AI video model Stable Diffusion has grown to more than 10 million users in two months. There will be more and more AI video computing needs in the future, and there will also be problems with expensive and scarce computing resources such as GPUs.
On May 21, 2024, Livepeer launched the AI subnet, the world's first decentralized video processing network with AI computing capabilities. The Livepeer AI subnet will leverage Livepeer's open network of thousands of GPUs to provide low-cost, high-performance processing of AI video computing tasks.
Livepeer.ai
Livepeer's AI network mainly consists of AI Orchestrator Nodes and AI Gateway Nodes.
AI Gateway Nodes mainly manage task flows and direct tasks to appropriate Orchestrator Nodes based on capabilities and loads, ensuring efficient task allocation and system scalability. AI Orchestrator Nodes are responsible for executing AI tasks.
Livepeer AI subnet task processing flow chart
Compared with centralized servers, the LivepeerAI subnet assigns tasks to decentralized, distributed GPU networks. According to analysis and research, NVIDIA provides 92% of the GPU computing resources used by AI data centers, while Microsoft and OpenAI account for 69% of the basic model and platform market. Decentralized networks can avoid single point failures caused by centralized services due to company closures or regulatory reasons. The access rights of the Livepeer AI subnet are open and permissionless. Entrepreneurs from any country can use it on demand to avoid being affected by regulatory policies.
Comparison of the AI video computing workflow between the centralized system and the Livepeer decentralized system
In addition, Livepeer has become the first decentralized AI infrastructure project to join C2PA, an open technical standard that provides publishers, creators, and consumers with the ability to track the sources of different types of media. The C2PA specification can help users understand the identity of the video creator, creation time, creation tools and other information.
C2PA members already include TikTok, Adobe, Google, Sony, Intel, BBC, Microsoft, and OpenAI. The upcoming Livepeer AI mainnet will solve the problem of content authenticity through native cryptographic signatures that show the source of the video, so that Livepeer will introduce global standards into the Web3 world while maintaining decentralization and openness.
Currently, the Livepeer AI subnet has been launched and is being continuously optimized. More than 20 high-performance AI Orchestrators are already active. It is expected that Livepeer will launch the mainnet in the third quarter of 2024.
Livepeer AI Video Roadmap
IV. Future and Risks
After the Livepeer AI subnet was launched, the community also approved two financial proposals, namely investments in the Aquareum and Realtime Interactive Video projects.
Aquareum aims to build a video application that can run on IOS, Android, and the web based on the Livepeer protocol, similar to the video live streaming platforms Twitch or YouTube. After Aquareum goes online, in addition to helping the growth of Livepeer network users, it will also use C2PA signatures for transcoding, so that the video clips transcoded by Livepeer can be traced back to the creator information, creation time, and other information of the video. It truly realizes the implementation of the C2PA technical standard in the decentralized video protocol application and solves the problem of the authenticity and reliability of decentralized video processing content.
Project founders Eli Mallon and Adam Soffer are both long-term contributors to the Livepeer protocol. Eli is also the main architect of Livepeer Studio. Adam played an important role in building Livepeer Subgraph, Livepeer Explorer, and Livepeer Studio player and front-end. Both worked at Livepeer Inc before developing the Aquareum project.
Realtime Interactive Video has significantly increased demand for the Livepeer network by attracting new users with the launch of a real-time interactive video product and growing protocol revenue.
Realtime Interactive Video's potential revenue expectations for the agreement
The project founder Evgeny previously worked at market maker Citadel Securities. He is also a loyal member of the Livepeer community and holds about 32,000 LPT tokens. In the previous pre-proposal, the project mentioned that the project token TEASY will be launched in August this year, which will be issued in the form of IDO, and most of the TEASY will be airdropped to the Livepeer community to continue to empower the Livepeer protocol.
In addition, Livepeer has also established a community organization Live Pioneers, which was inspired by Graphtronauts, the largest unofficial community of GRT holders. The main work is to build new Livepeer telegram groups and promote Livepeer in media channels such as youtube, X, and Discord.
Currently, the focus of Live Pioneers’ work is to guide and educate LPT holders to delegate staking. This also shows that the project is actively working to increase the token delegation staking rate, reduce the token inflation rate, and empower the token economy.
In the process of continuous construction of the Livepeer ecosystem, there are also some regulatory pressures that decentralized projects may encounter, as well as the pressure on the token economy caused by the long-term low entrusted pledge rate. After the Livepeer AI mainnet is launched, with the official release of Sora at the end of the year, more and more AI video teams from different countries and regions will join the ecosystem, and there will be regulatory risks such as lack of review of video content, resulting in intellectual property infringement, content violations and illegalities.
According to the current token inflation rate, nearly 6 million LPTs will be issued each year. If the promotion and construction of subsequent projects are not as expected and the pledge rate cannot be improved, the excessive issuance of tokens will undermine the confidence of investors and users of the project.
