The zkvm @ brevis_zk, which has been reposted and commented on by @ VitalikButerin in recent days, is very popular, especially since they recently launched real-time proofs with an average proof speed of 6.9 seconds. I have also shared this project before. Today, the system will share the importance of real-time proof for ETH and the overall research sharing of @ brevis_zk. Why is real-time verification important? What is the value of real-time proof? What is the relationship between ETH L1 expansion? 1) The calculation and capacity of ETH L1 are still limited, so the expansion of ETH L1 planned by the V God and Foundation team is on the verge of completion. Although there are different modules, ZKVM plays a more central role here. In the future, ETH L1 will be based on ZK as the core, and the consensus layer will avoid repeated calculations for verification. Instead, it will only be calculated once to generate proof, and other nodes will verify the proof and reach consensus through a new mechanism. And in order to avoid single node and implementation circuit vulnerabilities, ZKVM clients must support multiple implementation versions, so ZKVM clients like Brevis have a chance to get a share. 2) The mechanism of zkVM in L1 scaling: Traditional Ethereum L1 validators need to: -Re execute transaction (EVM interpreter) -Check if the state transition is legal This part has high computational costs and hinders expansion. 3) ZkVM allows "proof instead of recalculation": -ZkVM fully executes transaction logic off chain (EVM bytecode or higher-level VM instructions) -Generate a zk proof (to prove that the state transition is correct) -Verify the proof using constant time for the contract on L1 (Validator) -Even if EVM logic is complex, verification is still O (1) 4) Through the zkvm client, not all transactions need to be executed, only proofs need to be generated after execution, and other nodes can verify the proofs, avoiding the low efficiency of repeated execution by all nodes. This will greatly improve the efficiency of L1, which mainly requires the support of real-time proofs. How much real-time proof is called real-time proof? Let's take a brief look at some mechanisms of ETH's consensus layer: The slot time of the Ethereum consensus layer (Beacon Chain) is determined by slot time=12 seconds. Ethereum generates a new block (i.e. a time slot) every 12 seconds. During this time slot, the consensus client needs to complete: -Receive new blocks; -Execute transactions within the block; -Verify whether the execution result is correct; -Spread signatures, votes, and summaries online. Therefore, if the ZK proof system wants to replace the verification logic of the Execution Client, it must complete the proof generation within a 12 second window, otherwise the consensus layer cannot confirm the new block in real time. 5) The latest Pico Prism uses 64 RTX 5090 GPUs, with 99.6% of blocks proven within 12 seconds, averaging 6.9 seconds Below is a detailed study http://Brevis.network ： AI/Web3 verifiable computing layer supporting real-time proof http://Brevis.network Positioning: http://Brevis.network It is the verifiable computing layer of AI and Web3 http://Brevis.network There are three main products: 1) Pico RISC-V ZKVM (real-time proof supported): Pico is a RISC-V ZKVM, and in order to cope with the rapid development of proof systems and the constant changes in business logic, Pico is designed based on modular ZKVM, supporting pluggable backend proof systems and business co processors. Brevis' technology foundation comes from Pico, and the latest has been supported for real-time proof. In addition, compared to other ZKMVs, Brevis has two products that have been implemented in dozens of DEFI customers in actual business scenarios, namely Brevis data Coprocessor and Incentra. 2）Brevis data Coprocessor： Brevis data Coprocessor is an efficient zero knowledge proof coprocessor (Brevis Chain) that enables smart contracts to read the complete historical on chain data of supported blockchains and run customizable calculations in a completely trustless manner. Traditional methods either use centralized solutions or require expensive gas costs for on chain computing (which is basically unaffordable), while Brevis data Coprocessor supports data-driven DeFi, user retention and participation functions, trustless active liquidity management, identity based on full chain activities, and more with trustless historical data. The current cooperative clients include PancakeSwap ，Usual Money ，MetaMask，Linea ，Uniswap Foundation，OpenEden ，Trusta AI，0G Wait 3) Incentra's verifiable DeFi protocol incentive distribution platform: Through the verifiable incentive platform supported by zkCoprocessor, traditional airdrop calculations are black box and are privately calculated by the project party to distribute rewards. Incentra supports rewards to be calculated directly in the zero knowledge proof circuit and verified directly on the chain, ensuring trust free, transparent, and verifiable reward calculation. Users can independently verify the reward amount without relying on any centralized organization. The following mainly introduces the architecture design of the core infrastructure Pico RISC-V zkvm: Pico is an open-source zero knowledge virtual machine (zkVM), which is a modular design RISC-V ZKVM. Its core concept is a "glue and co processor"+modular architecture - that is, the VM itself acts as the "glue", responsible for general logic, control flow, and state transitions; High frequency or complex operations are processed by specialized coprocessors to improve performance, such as AI inference coprocessors and on chain data coprocessors. In addition, in recent years, it has been proven that the development of certification systems has been very rapid. Once the code is bound to a certification system, it may slow down the follow-up of the latest technological progress of the certification system. Therefore, Pico design is a pluggable modular architecture that supports pluggable certification systems and business circuit logic. Glue layer/VM core: Manage execution processes, state transitions, function calls, exceptions, memory&device access General, application agnostic logic is handled here, the core of VM Precompiled/Built in Functions: Provide efficient implementation for common operations such as hashing, encryption, signing, and arithmetic extensions. Expanded the RISC-V instruction set to accelerate underlying operations such as hash computation and signature verification. Pico supports developers to customize pre compiled modules according to their needs, and also provides out of the box optimization solutions to avoid these operations being deployed as general circuits, thereby reducing costs Coprocessor: Provide specialized circuit acceleration for specific scenarios such as on chain data access, transaction logic, complex algebraic operations, etc Can be customized or replaced as needed. Backend/Proof Pipeline (pluggable proof system): 1) Support multiple proof systems/prime fields/optimization strategies such as STARK (KoalaBear, BabyBear), CircleSTARK (Mersenne 31), Pico supports optional modular backends, which means developers can choose different proof systems, prime fields, or adjust the proof pipeline to optimize performance/proof size/validation time and other trade-offs 2) For example, the Poseidon2 hash function is mainly used for zkVM recursive proofs. Under the STARK system, KoalaBear's proof efficiency far exceeds BabyBear's, and performance can be significantly improved by simply replacing the proof backend without modifying the computational logic Pico also extends this architecture to verifiable AI inference (AI inference co processor): This is another potential direction for Pico. In many AI API businesses, a router service provider coordinates multiple models or automatically routes models based on cost. This is a good service in itself, but service providers often pass off inferior products as good based on profit and other factors. For example, if a user has called GPT5 10 times and the router service provider wants to increase profits, they will give you GPT5 6 times and another low-cost model 4 times. In this case, verifiable AI inference services are needed. Pico Prism's distributed/multi GPU architecture: Brevis has launched Pico Prism, a significant expansion of the original Pico architecture, featuring a distributed, multi GPU design to support larger scale, more real-time Ethereum block proofs. How Brevis participates: 1) Brevis Spark Points Activity: https://proving-grounds.brevis.network/qypDX6 2) Discor obtains roles Real time verification of research data: https://blog.ethereum.org/2025/07/10/realtime-proving