The world is drowning in scientific research. Over 2 million scientific papers are published each year. That’s 5,500 per day! If you worked 24 hours a day with no breaks, you’d have to read 1 paper every 16 seconds just to keep up. That’s a new theory on solar geoengineering, a new AI technique, a new way of engineering proteins, a new approach to treating cancer, every 16 seconds.
This unprecedented rate of discovery is an unqualified good. Every modern marvel, from the phone in your hand to the internet delivering you this post to the medicine you take, started as a science experiment. Fundamental research is the fuel of human progress and the discoveries are accelerating. Today more PhD’s are toiling and more money is being spent on research than at any point in human history.
The problem, however, is that the explosion in research has not led to a proportional surge in breakthrough products or inventions — most notably in pharmaceuticals. While there have been incredible advances in fields like autonomous vehicles, quantum computing, and longevity, in many other areas the gains are less profound, require greater resources, take longer, and cost more to produce than they used to.
This dynamic has many facets, but it starts with the “too much research problem.” The first step in any experiment is the literature review. This is a tedious, painstaking process by which a researcher sets out to first understand a problem, what work has been done before, the questions that have been answered, the ones that haven’t, and to begin to think about how one might structure an experiment in virgin territory to fill in any gaps.
This winnowing process, typically, relies on tools such as Google Scholar, which procure lists of PDFs of previous papers — some of them relevant, many of them not — based on basic keyword searches. These must then be read assiduously, and will often lead to new strands as a researcher follows a citation to a new paper, which in turn leads to another, and another. In a world where new papers pop up every 16 seconds, this winnowing process is hugely time-consuming and deeply inefficient. The problem of information overload is a giant speed bump on the road to progress. A survey back in 2016, when academic output was 40% lower than today, found that researchers spent at least 7 hours per week simply looking for information — ~20% of their work week!
What if there was a way to use AI to automate not just that first, painful step, but much of the rest of the process of scientific discovery as well? What if we could hand every bioengineer and economist and physicist a superpower that unshackles them from the drudgery that slows down their work and delays human progress?
Enter Elicit, which has developed an AI assistant that leverages large language models to automate research. The spinout from Ought, a nonprofit research lab, launched the first version of its product two years ago and has by word-of-mouth alone built up a monthly user base of 200,000 researchers, ranging from MIT and Stanford scientists to researchers at industry leaders like Novartis, Genentech, and OpenAI.
Elicit has built one-click literature review. And it’s working. Elicit’s language model responds to natural language prompts — "What are all of the effects of creatine?" or "What are all of the datasets that have been used to study solar geoengineering?” — to produce and summarize relevant papers. The results are based not on keywords but on concepts, and often link research from other disciplines that traditional searches would miss. In a pilot project, researchers using Elicit halved the time and cost of data extraction and analysis that today is still done by hand.
This isn’t easy (just ask Meta). Hallucinations plague LLMs. If you’re creating an image or writing marketing copy, a hallucination every now and then is annoying but tolerable. But if you’re going to invest tens of millions of dollars to develop a new cancer therapy based on LLM output, hallucinations are a gamestopper. To overcome this, the Elicit team pioneered iterated task decomposition, a new approach to language model deployment. This technique breaks complex tasks into more easily understood pieces, which then allow for statistical guarantees for each piece. By doing so, Elicit gets more understanding of how often different models hallucinate and in what context, which allows better checking. By forging new ground on Process Supervision, Elicit is able to uniquely synthesize vast quantities of scientific data without falling prey to one of the more hilarious and annoying failure modes of LLMs — confidently claiming something as fact that’s absolutely not the case.
Literature review at warp speed is just the beginning. Once a literature review is done, researchers must often then carry out a systemic review process, also known as meta-analysis. This is a deeper-dive that yields a rigorous empirical examination of the chosen domain of inquiry. The process, which leads to a written manuscript, typically takes five researchers 41 weeks to complete. Elicit will soon automate this as well.
Elicit is building the power loom of scientific inquiry — a system that turns inefficient, hours-long processes into tasks that can be handled in an instant. In this dawning age of artificial intelligence, Elicit is pioneering a new paradigm in which thinking and reasoning will, increasingly, move from the human mind to machines. If today Elicit can automate the brute force task of plowing through the world’s compendium of scientific knowledge, it’s not hard to imagine a future where it becomes a true co-pilot — a thought partner to the world’s greatest minds that helps devise experiments to test a given hypothesis.
Imagine one wanted to know if metformin could treat Alzheimer's. With a click, Elicit can find and summarize all the relevant research. It will soon be able to create a meta-analysis that synthesizes all the studies into a full picture of what’s known to date. If a researcher decided things looked promising, today they’d assemble a team to design experiments and clinical trials to prove or disprove their hypothesis. This is extremely time consuming and expensive. In the future, one could imagine Elicit handles that as well. From a curiosity to a literature review to a meta-analysis to fully designed experiments — something that would today take a team of researchers many months could be done by a single person in hours. That would radically accelerate scientific discovery.
This is the future that Elicit co-founder & CEO Andreas Stuhlmüller has been thinking about since he was a teenager building chatbots with QBasic. And he’s well prepared for the challenge — he holds a PhD in Cognitive Science from MIT, was a Computational Cognitive Science postdoc researcher at Stanford, and co-created WebPPL, a programming language for probabilistic machine learning. In 2017 he started Ought, an AI research nonprofit, to address the concern that this powerful technology would primarily be used for unimportant things like meme generation or stock trading. He wanted to use AI to advance human knowledge, propel science, and influence policy for the better.
To start Elicit, Andreas teamed up with co-founder Jungwon Byun, who holds an economics degree from Yale and has long been focused on maximizing positive impact. This focus led her first to microfinance in Africa and then to Upstart, where she was head of growth in the years leading up to its IPO. Lured by a shared passion to use artificial intelligence for societal good, she joined Andreas at Ought, which then spun out Elicit as a public benefit corporation.
Together, they have built the tools to supercharge the scientific endeavor. The vision, however, goes much further. As Andreas says, "Over the next 3 to 30 years, AI systems will have the capacity to do human-like reasoning at superhuman quality and quantity, leading to a radical shift in our society, with most thinking happening in machines instead of human brains." Elicit wants to massively scale up good reasoning.
At Fifty Years, our sweet spot is supporting founders at the earliest stages of building deep tech companies that can generate huge financial outcomes and create massive positive impact.
Deep Tech: Preventing hallucinations is hard. The Elicit team has pioneered new approaches in AI like iterated task decomposition to do this. By building the critical infrastructure of the AI-enabled research age, Elicit will accelerate the entire scientific endeavor.
$1B yearly revenue potential: Big Pharma alone spends hundreds of billions of dollars on research and development and is desperate for technologies to make that money go further. This is in addition to the countless university researchers and industry leaders in semiconductors, energy, materials science, and beyond who are all bedeviled by the same problem: the slow, expensive, by-hand processes that are the foundation of all research.
Massive positive societal impact: Elicit is making research systematic, transparent, and unbounded:
Systematic: They’re running batch processing over millions of papers.
Transparent: They’re taking the best practices in people’s heads and transforming them into software that can be shared and improved.
Unbounded: They’re turning static PDFs into living models that anyone can extend and customize.
Human progress is a story of technological and scientific advancements going back to harnessing fire and crafting stone tools. Fundamental research is the driver of those advances today and Elicit is bestowing practitioners of this vital work with superpowers that will unleash a new age of breakthroughs.
Inspired by Elicit’s vision and execution, Fifty Years was proud to lead their seed round along with Fifty Years LPs Arash Ferdowsi, co-founder of Dropbox, Tom Preston-Werner, co-founder of Github, Ilkka Paananen, co-founder of Supercell, and Jeff Dean, Chief Scientist at Google, as well as Basis Set Ventures and angels like Thomas Ebeling, the former CEO of Novartis. We look forward to working with Andreas and Jungwon as they build the warp drive for research.
I love using Elicit!
Do you guys have funds? Asking on behalf of a fund of funds in Spain