If you ask ChatGPT how to run an instance of itself on the blockchain, until recently, the answer would be along the lines of "Um, you know smart contracts don't work like that, right?"
Everything changes with Cartesi rollups. Powerful, distributed non-web3 applications are now possible - and ChainGPT is here to test the boundaries.
ChainGPT is an open, verifiable and decentralised implementation of a chatbot AI agent similar to ChatGPT, using Stanford's Alpaca Large Language Model (LLM).
Why do we need cryptographic truth when using LLMs? Because LLMs always give an answer, and don’t actually know what’s true. Examples:
ChainGPT is a 🤯 C++ application 🤯 that is guaranteed to give the trusted, expected result - providing economic security for businesses that leverage AI chatbots. The security of ChainGPT's output comes from it being run on multiple "validator nodes" (which are just regular old computers in anyone's house). The fact that ChainGPT results are reproducible means that together these validators ensure the correctness of each response. These results can be referenced to settle claims against the behaviour of chat bots.
Users can interact with ChainGPT knowing it's backed by the security of the blockchain. This is because ANY discrepancy between what two or more validators say the result should be, are mediated on the blockchain. In the absence of any discrepancies, the blockchain is still being used to compare results, but it's not expensive. Combining this with the long-running audit chain of previous ChainGPT interactions is what allows it to provide auditing and accountability for AI chatbots in all fields.
ChainGPT is the fusion of two main technologies: an existing pure C++ application to run Stanford's Alpaca large-language AI model (call alpaca.cpp), and the Cartesi Operating System, which allows for nearly any app that can be run in a RISC-V linux architecture Docker container to be decentralised on the blockchain.
Alpaca.cpp, however, does not currently support RISC-V architecture (ask me how I know). Porting it so that the fusion of these technologies can happen is the technical novelty in this project. Indeed, this technical difficulty was the motivation for the project in the first place: while ChainGPT is a tool that can be USED by people of all ages, all professions, living in any country, to develop it requires a specific skillset at the intersection of web3 and machine learning.
The community contributions that have come from the development of ChainGPT are:
"Blockchain-backed ChatGPT, is that even possible?!" ChainGPT proves that with Cartesi: yes, yes it is. 🔥