Christian Salem, a backup quarterback at Northwestern University outside Chicago from 2012 to 2016, graduated with a degree in economics. “We’re technical professionals who secretly want to be scientists…but we don’t have the time or focus to read scientific papers.”
Originally from Sudbury, co-founder Eric Olson played light tackle at Northwestern University and earned a master’s degree in predictive analytics. Olson then worked as a data scientist at Boston-based sports betting company DraftKings, while Salem became his product manager for the National Football League.
Olson and Salem are not the first to experiment with scientific search engines. Since 2015, the Allen Institute for Artificial Intelligence in Seattle has operated Semantic Scholar, a website that summarizes the content of academic research papers.
And in mid-November, tech giant Meta, Facebook’s parent company, released a public beta of a tool called Galactica, trained to summarize 48 million academic papers. However, Galactica was shut down after just two days because the results were often gibberish.
“At the time of release, we outlined the limitations associated with large language models such as Galactica, including the possibility of producing inaccurate and unreliable output,” Meta said in a statement. I’m here. “Given that large language models such as Galactica tend to produce text that may look authentic but is inaccurate . . . we have decided to remove the demo from publication.”
Karl Bergstrom, a biology professor at the University of Washington who has tried it, said Galactica’s AI often gives completely wrong answers even when they look believable. Bergstrom said, “When you give something, it doesn’t know it.”
Instead of trying to summarize academic papers, consensus simply identifies and highlights key findings. Bergstrom says the results are much more useful. “I really, really, really hate Galactica,” he said, “and I like this one.”
Olson and Salem have been thinking about a scientific search engine for years. But when they learned that traditional search engines like Google could easily spread false information, they got serious. The problem, Olson said, is that Google’s algorithm tends to favor the most popular internet sources instead of the most trusted ones.
“Exploring the thoughts of the experts…it’s really, really, really hard,” Olson said. “Google was not designed to do this for us.”
Olson and Salem, on the other hand, have found that AI systems have become much more powerful in recent years, able to understand words and phrases in a wider context. This convinced them that they could train the AI to break down the document into important sections and show only the most important parts.
In 2021, the two men hired three software engineers and raised $1.3 million. The largest chunk comes from Winklevoss Capital, founded by Tyler Winklevoss and Cameron Winklevoss, best known for its bitter legal battle with Meta CEO Mark Zuckerberg over Facebook’s founding.
Olson and Salem started with an existing AI model already trained on scientific documents written by humans. We then hired scientists to read and annotate about 100,000 academic papers in various fields. We then used these papers and scientists’ notes to train the AI to recognize specific features found in all academic papers, especially the part that summarizes the authors’ conclusions.
The Consensus search engine links to a database of 200 million public domain scholarly articles. When a user asks a question, Consensus responds in about five seconds with a long list of excerpts from scientific journal articles. The system does not aim to provide simple answers, but to provide users with various academic studies on the topic. Asks the consensus, “Is nuclear power safe?” And offers citations from multiple papers, some in favor of nuclear power, others more skeptical. But all citations are from reputable academic publications, not dogmatic amateurs.
For now, Consensus does not have access to the millions of academic papers locked behind paywalls by subscription-only publishers. Salem said Consensus plans to work with publishers to negotiate a solution. It is said that Additionally, the Biden administration announced in August that he must make all federally funded research papers available to the public starting in 2025.
About 15,000 people have logged on since Consensus opened to the public in September. Salem said the service has received many inquiries from health and fitness enthusiasts, parents seeking parenting advice, and students seeking homework help. Salem and Olson plan to offer a premium his version that provides a more detailed overview of each paper, including information about the organizations that funded the research. But for now, Olson said, “We’re completely free, and we want to keep some of our products free forever.”
Hiawatha Bray can be reached at [email protected]. follow him on twitter @GlobeTechLab.
