Today, Future House is announcing WikiCrow, our first automated system for synthesizing scientific knowledge. As a demo, we are releasing a cited technical summary for every gene in the human genome that previously lacked a Wikipedia article: 15,616 in all.
As scientists, we stand on the shoulders of giants; but most scientific knowledge today is still locked up in impenetrable technical reports. One of the ways AI will help to accelerate science is by enabling us to rapidly synthesize existing scientific knowledge. WikiCrow generates Wikipedia style articles on scientific topics by summarizing full-text scientific articles that we can access through our academic affiliations. In the past few days, we used WikiCrow to generate high quality, cited summaries for all 15,616 protein-coding genes that previously lacked high-quality Wikipedia articles, with information sourced from over 14 million pages of more than 850,000 relevant papers. WikiCrow is much more reliable than human authors at providing citations for information, and makes inaccurate statements about 9% of the time. We don’t usually observe hallucinations, but the model sometimes confuses gene names and details like the N and C terminus. We expect these errors to become rarer as we mature our systems. As they do, resources such as these will make it easier for researchers to understand and contextualize the results of their experiments, and will be a key resource for the AI Scientists we build in the future.
The articles are hosted on our website. Read our blog post here: www.futurehouse.org/wikicrow
and read our paper on PaperQA, the underlying agent, here: storage.googleapis.com/fh-public/paperqa/PaperQA__Retriev…