For researchers, sifting through scientific papers can consume valuable time. A recent survey found that scientists devote an average of seven hours each week to search for information, while systematic literature reviews—thorough evaluations of evidence on a specific topic—typically take around 41 weeks for a research team of five to complete.
But it doesn't have to be this challenging.
This is the premise behind Elicit, an AI startup co-founded by Andreas Stuhlmüller, designed to function as a “research assistant” for scientists and R&D labs. Backed by notable investors, including Fifty Years, Basis Set, Illusion, and esteemed individual investors like Jeff Dean (Google's chief scientist) and Thomas Ebeling (former Novartis CEO), Elicit aims to streamline literature review processes by leveraging AI technology.
“Elicit automates scientific research using advanced language models,” Stuhlmüller explained in an email interview. “It specifically facilitates literature reviews by identifying relevant papers, extracting key study details, and organizing this information into coherent concepts.”
Elicit is a for-profit initiative spun out of Ought, a nonprofit research foundation established in 2017 by Stuhlmüller, formerly a researcher at Stanford's computation and cognition lab. Co-founder Jungwon Byun joined Elicit in 2019 after leading growth at online lending platform Upstart.
By employing both custom and third-party models, Elicit enables users to explore various concepts in academic papers. Users can pose queries like “What are all effects of creatine?” or “What datasets have explored logical reasoning?” and receive comprehensive lists derived from academic literature.
“By automating the systematic review process, we provide immediate cost and time savings for academic and industrial research organizations,” Stuhlmüller stated. “By significantly lowering costs, we can unlock new possibilities that were previously too expensive, such as on-demand updates when knowledge in a field evolves.”
However, one might wonder—don’t language models often produce inaccurate information? This concern is valid. For instance, Meta's language model aimed at scientific research, Galactica, was taken down just three days post-launch due to its frequent citation of fabricated papers that seemed authentic but were not.
Stuhlmüller asserts that Elicit has implemented measures to enhance the reliability of its AI compared to many existing platforms. For starters, Elicit deconstructs complex tasks performed by its models into “human-understandable” segments. This breakdown allows Elicit to monitor the frequency of inaccuracies within model-generated summaries, guiding users on what answers to verify.
Additionally, Elicit evaluates a scientific paper's overall “trustworthiness” by considering elements such as trial controls, funding sources, potential conflicts, and study sizes.
“We don’t utilize chat interfaces,” Stuhlmüller emphasized. “Elicit allows users to apply language models in batch processes. Rather than simply generating answers, we always trace the results back to the original scientific literature to mitigate inaccuracies and facilitate verification.”
While I remain cautious about Elicit’s ability to address significant challenges associated with language models, it seems to be gaining interest—and perhaps some trust—from the research community. Stuhlmüller reports over 200,000 monthly users, marking a threefold year-over-year increase since January 2023 from prominent organizations like The World Bank, Genentech, and Stanford. “Our users are eager for more advanced features and to leverage Elicit at greater scales,” he added.
This momentum likely contributed to Elicit's recent funding round—$9 million led by Fifty Years. The primary objective of this funding is to enhance Elicit's product and expand its team of product managers and software engineers.
But how does Elicit plan to generate revenue? I posed this question directly to Stuhlmüller. He highlighted Elicit's newly launched paid tier, allowing users to search academic papers, extract data, and summarize concepts on a larger scale, surpassing the capabilities of the free tier. The long-term vision is to evolve Elicit into a comprehensive tool for research and reasoning that entire organizations would be willing to invest in.
One potential challenge to Elicit’s commercial success is the emergence of open-source initiatives, like the Allen Institute for AI’s Open Language Model, which aim to create a free large language model optimized for scientific contexts. Nevertheless, Stuhlmüller views such open-source efforts as complementary rather than competitive.
“The primary competition we face today is human labor—research assistants who painstakingly extract data from papers,” Stuhlmüller stated. “The scientific research market is vast and largely untapped in terms of efficient research workflow tools. This presents an opportunity for entirely new AI-driven workflows to arise.”
Ranking
