Several researchers have perceived scientific progress in the last 50 years and there has been the same disturbing conclusion: scientific performance decreases. It takes more time, more funds and larger bands to make discovering which were once faster and cheaper. Although various explanations have been offered for slowdown, one is that as research becomes more complex and specialized, scientists must spend more time browsing publications, designing sophisticated experiments and analyzing data.
Now, philanthropically financed from the Research FutureHouse laboratory, he is trying to speed up the scientific research with the AI ​​platform designed to automate many critical steps on the way to scientific progress. The platform consists of a series of AI agents specializing in tasks, including information search, information synthesis, chemical synthesis project and data analysis.
The founders of FutureHouse, Rodriques, doctorate '19 and Andrew White, believe that by giving each scientists access to their AI agents, they can break the greatest bottlenecks in science and help solve some of the most smoking problems of humanity.
“Natural language is a real language of learning,” says Rodriques. “Other people build models of biology foundations in which machine learning models speak in DNA or proteins, and this is powerful. But discoveries are not represented in DNA or proteins. The only way we know how to represent discoveries, hypothesis and reason is natural language.”
Finding big problems
For his doctoral studies in Mit Rodriques, he tried to understand the internal functioning of the brain in the laboratory of Professor Ed Boyden.
“The whole idea for Futurehouse was inspired by the impression I got during the doctorate in myth that even if we had all the information that we had to know about how the brain works, we do not know about it because no one has time to read all the literature,” explains Rodriques. “Even if they could read all this, they would not be able to put it in a comprehensive theory. It was a fundamental piece of the future puzzle.”
Rodriques wrote about the need for new types of large research as the last chapter of his doctoral dissertation in 2019 and although he spent some time running a laboratory at the Francis Crick Institute in London after graduation, after graduating, he turned to broad learning problems, which no single laboratory can take.
“I was interested in how to automate or scale science and what new organizational structures or technologies would unlock higher scientific performance,” says Rodriques.
When Chat-Gpt 3.5 was published in November 2022, Rodriques saw the path to stronger models that could generate scientific observations on their own. Around this time, he also met Andrew White, a computing chemist at the University of Rochester, who received early access to chat-gpt 4. White built the first large language agent for science, and scientists joined forces to set up Futurehouse.
The founders began to want to create separate AI tools for tasks such as literature search, data analysis and generation of hypotheses. They started by collecting data, ultimately releasing Paperq in September 2024, which Rodriques calls the best AI agent in the world to recover and summarize information in scientific literature. Around the same time, they released everyone, a tool that allows scientists to determine whether someone conducted specific experiments or examined certain hypotheses.
“We just sat with the question:” What are the questions that we as scientists ask all the time? ” – recalls Rodriques.
When FutureHouse officially launched his platform on May 1 this year, he changed the brand of some of his tools. Paper Qa is now a crow and now that someone is called Owl. Falcon is an agent capable of compilation and reviewing more sources than CROW. Another new agent, Phoenix, can use specialized tools to help researchers plan chemistry experiments. And Finch is an agent designed for automation discoveries based on data in biology.
On May 20, the company demonstrated a multi -level flow of scientific work in order to automate the key stages of the scientific process and identify the new therapeutic candidate for age -related macular degeneration (DAMD), the main cause of irreversible blindness around the world. In June, FutureHouse released Ether0, a 24b open reasoning model for chemistry.
“You really have to think about these agents as a larger system,” says Rodriques. “Soon literature search agents will be integrated with a data analysis agent, hypothesis generation agent, experiment planning agent and everyone will be designed for cooperation without any problems.”
Agents for everyone
Today, everyone can access FutureHouse agents on the platform.futurehouse.org. The launch of the company's platform caused emotions in the industry, and stories began to appear about scientists using agents to accelerate research.
One of the FutureHouse scientists used the means to identify the gene, which can be associated with polycystic ovary syndrome and develop a new hypothesis of the treatment of the disease. Another researcher from Lawrence Berkeley National Laboratory has used CROW to create an AI assistant who can search the PUBMED research database to obtain information related to Alzheimer's disease.
Scientists from another research institution used agents to conduct systematic gene reviews relevant to Parkinson's disease, stating that FutureHouse agents are working better than general agents.
Rodriques says that scientists who think about agents less like Google Scholar, and more intelligent scientist's assistant to fully use the platform.
“People who are looking for speculation have a greater course of deep tap tapt o3, while people who are looking for really loyal reviews of literature usually come out of our agents,” says Rodriques.
Rodriques also believes that FutureHouse will soon reach the point where its agents can use raw data from research to test the playback of their results and verify the conclusions.
In a longer run, to maintain the scientific progress of marching, Rodriques claims that FutureHouse is working on embedding his agents with silent knowledge to be able to conduct more sophisticated analyzes, while giving agents the opportunity to use computational tools to discover hypotheses.
“There are so many progress in scientific foundations models and around language models for proteins and DNA that we must now provide our agents with access to these models and all other tools that people commonly use for learning,” says Rodriques. “Building infrastructure to enable agents to use more specialized learning tools will be crucial.”
