A comprehensive examination of generative artificial intelligence reveals several significant concerns, including the unauthorized use of creative work, inherent biases within algorithms, and the substantial energy and water resources required for training these systems. Despite these challenges, generative AI exhibits impressive potential for developing prototype tools that may be valuable in various applications.
This potential was observed at Sundai Club, a generative AI hackathon occurring monthly near the MIT campus. A few months ago, the club, supported by the Cambridge nonprofit Æthos, which advocates for socially responsible AI usage, focused on tools beneficial to journalists. Participants of the club include students from MIT and Harvard, professional developers, product managers, and even a military personnel member.
The event initiates with a brainstorming session to generate project ideas, which are subsequently narrowed down to a final project to pursue. During the journalism-focused hackathon, notable suggestions included leveraging multimodal language models to monitor political content on TikTok, automatically generating freedom of information requests and appeals, and summarizing video clips of local court hearings to support local news coverage.
Ultimately, the participants elected to create a tool aiding reporters covering AI to identify notable papers on Arxiv, a widely used server for research paper preprints. This decision may have been influenced by a participant who expressed that finding intriguing research on Arxiv was a priority.
The team then developed a word embedding of Arxiv AI papers using the OpenAI API. This mathematical representation enabled analysis of the data to find relevant papers and to explore connections between different research areas. Additionally, by utilizing a word embedding of Reddit threads and conducting a Google News search, the team created a visualization illustrating research papers alongside Reddit discussions and pertinent news reports.
The prototype, named AI News Hound, is still in its early stages but demonstrates how large language models can facilitate information mining in innovative ways. A screenshot of the tool shows a search for “AI agents,” with green squares near news articles and Reddit clusters indicating research papers that might be pertinent to discussions on developing AI agents. This project exemplifies the creative output of the Sundai Club.
