With Sinclair Davidson and Jason Potts. Available at SSRN
Abstract: Generative AI is a very powerful new computing technology, but the problem of how to make it economically useful (Alice: “hello LLM, please send an email to Bob”) is limited by its inherent unpredictability. It might send the email, but it might do something else too. As a consequence, the large language models that underpin generative AI are not safe to use for most economically useful and valuable interactions with the world. This is the ‘economic alignment’ problem between the AI as an ‘agent’ and the human ‘principal’ who wants the LLM to interact in the world on their behalf. The answer we propose is smart contracts that can take LLM outputs and filter them as deterministic constraints. With smart contracts, LLMs can interact safely in the real world, and can unlock the vast economic opportunity of economically aligned and artificially intelligent agents.