Workshop Presentation Group 5: BCI(s) for AI Alignment @ WBE Workshop 2023

Participants interested in exploring the project (+potential role, e.g. driver, advisor, funder, etc)Sumner NJudd RAndy Tamuz HodDiana S – network for fundingWilliam ENiccolo Zanichelli Davidad – advisorJoanne Peng This talk summary discusses a workshop presentation focused on Brain-Computer Interfaces (BCI) for AI alignment and neurotech. The objective is to identify BCI approaches that can enhance the probability of successful AI alignment. The project acknowledges the limitations in this field such as limited access to human data, funding constraints, and a lack of testable hypotheses and experiments. To address these limitations, the project suggests exploring multiple BCI approaches simultaneously using the upper confidence bounds algorithm, as this subfield is still in its early stages. The estimated cost for testing BCI approaches is relatively low, and specific experiments are proposed to test hypotheses related to BCI for AI safety. The experiments include increasing human capacity using BCIs, using autoencoders as a choke point to prevent ASI from representing non-human brain states, and training a generative model based on human brain data as a proxy for human evaluation during AI alignment. The project emphasizes the need to conduct experiments to identify effective solutions, considering their cost-effectiveness and potential impact on AI alignment.