
头雁|Sep 01, 2025 11:05
0G @ 0G_labs TGE is probably coming soon. I have written 0G several times before, mainly focusing on the decentralized computing power module and some basic components. I did not have a systematic and complete look at its core features. Today, I will write in detail, which is a bit long, and I have researched it manually.
Positioning: 0G is a modular decentralized AI public chain
Main research points:
My favorite thing about research projects is to see what their positioning is and the team situation. Today, I won't write about the team, but focus on the characteristics and design after positioning, with a particular emphasis on researching:
-What are the value points of decentralized AI
-0G Multi Consensus Protocol and Verifiable Computing Layer
-Innovation in Computing Power (Low Communication Computing Power Layer)
Current entry perspective for the decentralized AI market:
At present, most of the AI infrastructure in the mainstream market is focused on Agent frameworks, model and community alignment, and decentralized computing power alone. There is no complete design from the perspective of deep integration of AI and blockchain. 0G is a relatively complete decentralized AI public chain that combines the scalability of blockchain and deep integration with AI.
Why do we need decentralized AI?
One is the monopoly of computing power:
Nowadays, the cost of training an AI model for a small team can range from several million to tens of millions of dollars, which is difficult for small companies and teams to afford. This has led to large models becoming a monopolistic industry, which can result in some companies controlling the dangerous AGI process. If decentralized computing power can lower the threshold for model development by gathering dispersed and idle low-cost computing power, it will indirectly promote industry development.
One is hindering innovation:
Another point you can see is how popular Deepseek was before and after this year's Spring Festival. The core reason is that it has open-source researched many details and models, as well as open-source training related code.
And the implementation of the inference part of OPENAI has been open sourced (this part has always been covered by OpenAI, and the open-source community has done a lot of analysis and possibilities, which has been leading the open-source community for six months to a year), which has triggered a large number of subsequent iterations and innovations in the open-source AI community, leading to the rapid development of RL training. It can be seen that open source and openness can better drive industry development, while CLOSEAI is holding back, resulting in RL training technology being delayed for 6-12 months to gain more expansion.
One is the lack of transparency:
When companies like OPENAI shift from non-profit organizations to for-profit entities, they do not completely prioritize industry or group interests in some core areas, and do not share their core goals. Here, you can see the value conflicts within OPENAI in the past year (such as Ilya Sutskever's desire to launch AGI at a safer pace, which led to information asymmetry between SAM and the board of directors to accelerate the launch of AGI products with commercial products as the core), as well as conflicts of interest. Some core members resign due to unfair distribution of computing power for a certain group. This indicates that when a company that controls the lifeblood of industry development cannot do better, it is necessary for them to... Transparent governance of these relationships will bring huge risks to the development of the industry, especially AGI This revolutionary product.
So blockchain can better coordinate and govern these relationships, making governance more transparent and centered around the interests of community groups. Of course, this also requires the establishment of a good community, because tools are always just tools, and it still depends on the community that uses them.
Solution and technical architecture design for 0G, architecture diagram:
0G adopts a complete modular design concept, divided into several parts:
-0G Chain: 0G Chain adopts advanced modular design, which clearly separates consensus and execution. It is an L1 blockchain designed specifically for AI.
-0G computing: providing decentralized computing power for management, the biggest problem with decentralized AI is still early performance issues.
-0G storage: 0G storage is a distributed data storage system designed on chain to incentivize storage nodes to store data on behalf of users. Anyone can run storage nodes and receive rewards for maintaining them
-0G DA: The availability of 0G DA data is mainly used to prove that the data is easy to access, verifiable, and retrievable. For example, in the computation verification of decentralized AI, it is necessary to retrieve DA data to verify whether the relevant calculations are truly remembered.
-0G DEPIN: A hardware layer that distributes computing resources equally in terms of computing power and nodes.
In the following in-depth research, there are two interesting parts of 0G design: the consensus part of 0G chain and the computation part of 0G:
0G Multi Consensus Part:
In terms of consensus design, it is difficult to have a universal consensus protocol. It can be seen that every new blockchain will constantly innovate on consensus protocols, but it is still necessary to find that the requirements for consensus protocols are constantly evolving and changing. 0G adopts multiple consensus protocols, and the underlying layer shares security guarantees and validator sets. Because in modular AI blockchain, there are storage networks, 0G data availability networks, and inference networks, all of which will establish separate consensus protocols based on these network characteristics.
Verification mechanism for classification:
-Reasoning proof (0G computing network):
0G supports the computation of decentralized GPU networks, which can be verified and prevent these decentralized computers from doing evil.
--Support opML (Optimistic Proof Validation for Machine Learning)
--ZkML (Zero Knowledge Proof Validation for Machine Learning)
--TEEML (Machine Learning Based on TEE Hardware for Proof Validation)
Here, 0G is different from most others that support OPML as the core. 0G supports multiple verification mechanisms, which can fully consider the trade-off between technology maturity and cost. Especially when zkml development is not particularly mature (some recent progress has been made at https://(GitHub)/ICME Lab/bolt atlas, and this team has optimized a lot of performance), it is a good design solution.
-Proof of Storage:
PoRA requires storage nodes to quickly access randomly stored data segments to prevent them from falsely reporting the actual stored data. The system does not punish wrongdoers, but rewards honest storage providers through competitive mining processes. This mechanism includes multiple safeguard measures to ensure that only genuine storage providers can effectively participate.
0G calculation part:
-Verifiable computing layer:
0G computing is a decentralized network organization of GPU machines, so there are various machine providers, excluding whether they can perform computing tasks according to the needs of users/developers. Therefore, the core of 0G computing should be able to verify that the computing is as required and return the results that are not malicious to the demand side. Architecture diagram:
The entire architecture is divided into two parts: on chain and off chain:
-Smart contracts on the chain include service registration, user accounts, and settlement logic.
Smart contracts are mainly responsible for verification (signature, nonce, balance) and fund settlement (transfer)
The overall process of Smart Contract is divided into 5 steps:
Service registration: Service providers register "verifiable services" in on chain smart contracts, including service types URL、 Price.
User pre deposit: Users deposit a portion of the fee into the smart contract for subsequent service calls.
User request: The user sends a signed request (including metadata such as address, nonce, service type, etc.) to the service provider.
Service provider response: The service provider verifies the user's balance and the validity of the request, and returns the result. After receiving the result, the user verifies it themselves, and if it fails, they can stop calling again.
Settlement: The service provider will submit a zero knowledge proof containing "request trajectory+user signature" to the on chain smart contract.
Smart contract verification proof → Check signature nonce、 Balance → Transfer funds to the service provider.
-Off chain (Provider&User/Developer)
User/Developer: Initiate a request that includes a signature and metadata (address, nonce, service type, etc.).
Provider: Receive requests, store critical data, generate verifiable proofs (zero knowledge proofs/TEEs), and submit them to the chain for settlement.
Core logic sorting:
On chain: Ensure the safety of funds and the reliability of settlement.
Off chain: Responsible for actual service execution and generating verifiable proofs.
Key mechanism:
Pre deposit: Users recharge first to avoid default.
Signature&nonce: Ensure that the request source is genuine and cannot be replayed.
Calculation verification mechanism: ZKML/TEEML/OPML: enables on chain confirmation of request validity without executing service logic.
This is a trusted service invocation architecture where users pre deposit fees, service providers execute requests off chain, and submit the request results to on chain settlement using zero knowledge proof/TEE proof/OP proof.
Low communication computing power layer (key innovation of 0G):
Due to the large amount of communication required between AI training nodes to exchange data, the traditional decentralized training track is mostly story and narrative, making it difficult to truly implement. And the earliest ones were several papers in 2023 https://arxiv.org/abs/2311.08105 , https://arxiv.org/abs/2502.12996 They are mainly training models for heterogeneous low-speed Internet GPUs. At first, the models were relatively small.
But it should have inspired many people, and the company that brought even greater breakthroughs was Prime Intellect, which released Intellect-1, a model that basically verified the feasibility of decentralized training.
And 0G released a model that uses decentralized training to train an order of magnitude larger. Based on the DiLoCoX framework implemented in the paper, the framework has successfully trained a basic model containing 107B parameters in a bandwidth limited decentralized network environment.
The main innovations include:
-Pipeline parallelism divides the model across nodes and overlaps computation and communication to maximize utilization.
-The dual optimization strategy allows nodes to perform local updates while still maintaining consistency with the global model objectives.
-Single step delay overlap allows training to continue without waiting for all nodes to synchronize, thereby reducing idle time.
-Adaptive gradient compression reduces the size of transmission updates while maintaining accuracy.
These technologies make it possible to train models with over 100 billion parameters in accessible and bandwidth limited environments with minimal convergence degradation. There are many innovative points in this part of the paper, but I don't see much open source coming out. I hope that in the future, 0G can have more open source code, open source training data, etc., so as to give the community greater transparency. Otherwise, it will still be similar to closed source models such as OpenAI.
Research materials:
DiLoCoX paper: https://arxiv.org/abs/2506.21263
Research paper on deepmind with low communication training:
https://arxiv.org/abs/2311.08105
https://arxiv.org/abs/2502.12996
Latest developments in zkml:
https://((github.com))/ICME-Lab/jolt-atlas
https://blog. (icme.io)/sumcheck-good-lookups-good-jolt-good-particularly-for-zero-knowledge-machine-learning/
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink