(TNS) — Artificial intelligence can now run science experiments with a few simple prompts from researchers, saving them time and speeding the pace of discovery.
The new system, Coscientist, was developed by Gabe Gomes, a chemistry professor at Carnegie Mellon University, and his two doctoral students, Daniil Boiko and Robert MacKnight, whose research was funded by the National Science Foundation. They published initial findings last week in the journal Nature.
Boiko and Gomes are currently building a startup to commercialize the technology.
The autonomous system can essentially do its own research after receiving a written prompt from a chemist or other physical scientist. In a demonstration prior to the Nature release, Gomes showed a 5-second video of a robot mixing chemicals by shaking a bunch of test tubes. It didn’t look like much, he acknowledged. But the yield — 55% — was “very impressive.”
“We’re able to go from an English prompt to fully realize complex scientific experiments,” he said.
While the tool has the potential to turbocharge the scientific method, scientists might not fully understand how they’re arriving at new conclusions. Coscientist is powered by a set of proprietary large language models (LLMs), including OpenAI’s GPT-4, whose operations aren’t particularly transparent.
But scientists don’t need to understand exactly what decisions it’s making if the output is successful, Gomes said.
“If you can’t tell the difference, does it matter?” he asked, quoting the dystopian show Westworld. “I don’t think it does.”
His team tested several AI systems but chose GPT-4 because it was the most capable. Gomes acknowledged the technology could be misused but said its potential for breakthroughs makes the risks worthwhile. He also said the system can correct itself when it makes mistakes.
Kate Kleinbaum asked during the Tuesday presentation if her lab could use the new technology. She is chief of staff at the Austin-based Emerald Cloud Lab, which is controlled remotely by software.
Yes, Gomes assured her. But the barrier to entry is even lower.
“You don’t have to have robots or automation systems,” he said. “You can use this for chemistry, biology, material sciences, you name it. There’s nothing preventing a human scientist in the lab from being informed by Coscientist about what they should do next.”
The tool has the potential to “ democratize” access to science, he said.
DJ Kleinbaum, the CMU alum who built the Emerald Cloud Lab, commended the Coscientist team.
“The work that Gabe and his team are doing here is some of if not the most cutting edge work,” he said. “You’re actually able to build a closed loop between a problem statement and a scientific result that can then iterate on itself.”
At his cloud lab, Kleinbaum said software performs thousands of checks before running an experiment, to make sure it’s not building a bomb or making a solvent that would melt the plastic tubing it’s housed in. Gomes said his team built similar guardrails.
Boiko, his graduate student and now business partner, said he was looking forward to adding more value. MacKnight said he was looking forward to seeing how other AI researchers responded to the Nature article.
