ThinkChain
On-chain service that enables smart contracts to perform verifiable large language model (LLM) inference.
Created on 15th February 2025
•
ThinkChain
On-chain service that enables smart contracts to perform verifiable large language model (LLM) inference.
The problem ThinkChain solves
ThinkChain provides access to a variety of popular LLMs, such as DeepSeek-R1, DeepScaleR, Qwen2.5 and SmolLM2. A simple Solidity interface makes it easy for smart contracts to construct prompts and decode replies entirely on-chain. Completion requests are charged in Ether.
Our project addresses three critical challenges in blockchain-based AI integration:
Scalability
Traditional on-chain AI computation faces severe scalability limitations due to the computational overhead. We solved this through EigenLayer co-processors, where network operators execute computations off-chain, significantly reducing the blockchain resource burden while maintaining decentralization.
Integration
Directly porting existing AI implementations to Solidity is impractical due to the EVM's computational constraints. However, by utilizing Cartesi's RISC-V virtual machine with Linux compatibility, we can execute deterministic AI inference for LLM models using traditional software off-chain and expose its results in EVM smart contracts.
True Decentralized Verification
Current blockchain AI projects often compromise on decentralization through various trust assumptions:
- Some rely on Trusted Execution Environments (TEEs), requiring trust in third parties
- Others use zkTLS, which still involves trusting external entities
- Many solutions don't address trust verification at all
Our solution provides robust verification without centralized trust points, maintaining the core promise of blockchain technology while enabling advanced AI capabilities.
This combination of features makes our project particularly valuable for smart contracts that would benefit from on-chain access to LLMs where verification and trust are critical.
Examples of use cases include AI agents, AI-assisted decision making, data analysis, and content generation.
Challenges we ran into
Integrating diverse technologies (React frontend, Python backend, and Solidity smart contracts) presented significant challenges. While our team had strong expertise in Solidity, Python, and LLM models, our limited frontend experience caused delays. The complexity of learning multiple technologies can be particularly challenging for newcomers during hackathons.
We also encountered issues with large machine snapshots (13GB) with coprocessor, where publishing the machine was too slow and occasionally failed with timeouts and other errors. We worked around this by making multiple publish attempts until successful. Furthermore, we experienced high disk usage (up to 100GB) because the coprocessor maintained multiple copies of machine instances (.tar, .gnutar, image/, .car, docker images). Some team members were unable to run the full machine due to excessive disk space consumption.