AI (actually intelligent) smart contracts: the Aizel thesis
the significance of LLM-powered smart contracts and why they will run on teeML (instead of zkML or opML)
(M31 Capital led Aizel’s recent pre-seed funding round)
Aizel Network is building a novel platform for on-chain smart contracts to leverage off-chain GPU, TEE/MPC verification & DAG scalability to bring trustless LLM functionality on-chain much more practically than zkML or opML.
Smart contracts aren’t actually very smart.
Blockchains are designed to be decentralized ownership ledgers; since all nodes must execute and download each transaction, on-chain smart contract functionality is limited to fairly simple logic to manage compute and storage requirements (optimizing network participation, i.e., decentralization). Therefore, more complex computation, such as LLMs, cannot be performed on-chain (by design).
However, smart contracts can leverage off-chain LLM compute – but only if the model’s output is verifiable, ensuring the blockchain’s state remains trustless. No one has been able to solve this problem in a timely and economically viable manner to-date, which has greatly limited the potential of smart contract functionality.
DeFi Summer Case Study
DeFi offers an interesting case study in unlocking the potential of smart contract utility. During Ethereum’s first few years of existence, on-chain functionality was limited, which proved to be a huge bottleneck for user adoption. In mid-2019 Chainlink (the decentralized oracle network) launched its mainnet, which finally enabled developers to connect on-chain smart contracts with off-chain price data in a trust-minimized manner. This development ignited an explosion in DeFi innovation and adoption and was the driving force behind the “DeFi Summer” of 2020 (arguably the single greatest breakthrough in Web3 application utility).
We believe bridging off-chain LLMs with on-chain smart contracts in a trust-minimized manner will similarly have a significant impact on Web3 application functionality and adoption.
LLM Verification Challenges
The most hyped and popular approach to LLM verification to-date has been zkML, attracting the majority of the market’s mindshare and capital. However, ZKPs are highly complex calculations, and verifying LLMs today costs up to 1000x the price of the original inference computation and can take weeks to complete. Even the loudest proponents of the technology agree it won’t be practical for at least another 3-5 years (at which point the tech will still likely be limited to simple LLMs). Despite the hype, we don’t believe zkML will bring meaningful AI functionality to smart contracts.
Enter Aizel Network
Aizel Network offers permissionless off-chain GPU compute for on-chain smart contracts to run complex LLMs, then returns the output back to the smart contract in a trust-minimized manner. It does so by verifying the integrity of the LLM computational process, then submitting this verification on-chain for anyone to audit. But as opposed to zkML, Aizel’s teeML design leverages a unique combination of TPM (trusted platform module) and MPC (multi-party computation) technologies. The architecture enables quick LLM verification at a fraction of the original inference cost.
Aizel Investment Thesis
zkML is overhyped and not feasible for mass adoption in the next 3-5 years.
Aziel offers a unique full-stack solution (GPU processing and inference verification) for on-chain LLM inference with technology that works today.
The platform’s modular design can also integrate with ZKPs in the event the technology becomes meaningfully more efficient and accessible in the future.
The network is highly scalable through its parallelized DAG execution architecture.
Trustless LLM infrastructure will become extremely valuable, materially opening up the design space for on-chain smart contracts & Web3 applications.
Aizel Technical Design
Parallelized Merkle DAG
General-purpose blockchains group transactions into blocks then add them to the chain in an ordered sequence so the state can be identical across all nodes. This design is optimized for single-threaded virtual machines, but unfeasible for compute-heavy ML applications. Instead, Aizel uses a Merkle DAG (Directed Acyclic Graph) architecture, enabling transaction parallelization. Since most ML inferences are context-free, multiple requests can be executed in parallel by different nodes and eventually verified and added to the Merkle DAG. Aizel will therefore be able to process a large amount of inference requests at the same time.
Node Architecture
Data Nodes: Off-chain nodes that contribute storage and data persistence for pre-trained LLMs, inference output, and all other transaction history information.
Inference Nodes: Off-chain nodes that receive inference requests from on-chain smart contracts, process the input through the LLM program stored on Data Nodes, verify (process detailed below) & encrypt the output, then send the output and verification data to the relay nodes.
Relay Nodes: Bridge data from off-chain inference nodes back to on-chain smart contracts in a trust-minimized manner by broadcasting verified transactions and corresponding inference output onto other blockchain networks.
Prover (future development if ZK tech is adopted): For certain scenarios where the integrity of the inference process is critical, Aizel is also designed to support ZKPs, allowing users to choose the verification method that best suits their needs and costs.
Inference Verification Technology
Trusted Platform Module (TPM): A physical, secure cryptographic computer chip that is used to attest the integrity of a system by a remote party. Even if the physical chip ends up in the hands of adversaries, it’s nearly impossible to successfully remove the private keys to hack the system. Although the risk is low, Aizel also leverages decentralized MPC technology to further secure the network.
Multi-Party Computation (MPC): The decentralized MPC ensures the attestation process is achieved in a highly secure manner, collectively done by a group of nodes chosen at random. This additional layer ensures the trustlessness of the remote attestation is nearly as high as the zkML approach.
Potential Use Cases
The CTO of Space and Time, the leading coprocessor for trustless off-chain data query & indexing, recently tweeted about what he believes are the most exciting potential use-cases for LLM-powered smart contracts:
As with any new technology, the killer use-cases built on top will likely be novel and hard to predict today. However, some additional ideas include:
On-Chain Biometric KYC: RWAs are one of the hottest narratives in Web3 today, however, one major bottleneck to adoption is the KYC process needed to ensure the legitimacy of corresponding wallets. ML models running on Aizel Network can identify bio features such as fingerprints or faces, and directly verify wallets/DIDs on-chain, enabling RWAs to finally cross the chasm to mainstream adoption.
On-Chain Security Computation: Web2 commonly leverages AI to proactively monitor various data networks, however the cost of computation has always limited the potential to do the same in Web3. Using Aizel Network, security-focused AI logic can directly interact with smart contracts, breaking the current limitation of on-chain security computation.
Human-Blockchain Interaction Engine: Web3 is infamous for its complex and difficult to use front-end interactions. For Web3 to gain mass adoption, UX/UI needs to improve by at least 10x. With Aizel, blockchain users can directly interact with natural language processing (NLPs) models, like ChatGPT, without needing to deal with the technical & clunky blockchain interfaces of today.
On-chain AI Generated Content (AIGC): AIGC has exploded in popularity over the past year, driven by the overwhelming success of OpenAI’s ChatGPT and DALL-E products. Applications leveraging the Aizel Network will finally be able to confirm and verify AI generated content's provenance and uniqueness on-chain.
Autonomous Oracles: Oracles can leverage AI models to provide off-chain data in a verifiable way, which could disrupt the current oracle model. Chainlink, the industry’s leading oracle network, understands this dynamic and is currently building out its ZK capabilities (DECO), validating the potential Aizel use case.
Private LLM Computation for Web2: Aizel isn’t limited to blockchain-based applications - it can also be leveraged for Web2 applications that require sensitive information to interact with LLMs in a private and secure manner, such as financial and healthcare data.
Competition
Although zkML has received the most market hype, its high costs and long processing times aren’t practical for the majority of use-cases. Interestingly, the Aizel team actually pursued zkML initially, but ultimately concluded it wasn’t economically or functionally practical. (Given Aizel’s modular design, if zkML does sufficiently improve in the medium-term, the network will be able to seamlessly integrate ZKPs into their tech stack anyway)
opML also offers an interesting alternative to zkML, however, costs are still multiples of the original inference computation, and the system lacks data privacy.
We therefore believe Aizel’s teeML architecture provides the most practical and efficient solution to power trustless on-chain LLM functionality.
We believe protocols like Aizel will unlock the next step-function in Web3 application utility. If you are building or researching the intersection of on-chain smart contracts and off-chain LLM compute, please reach out for partnership and/or collaboration opportunities!