Tetsuwan Scientific is developing robotic AI scientists that can conduct experiments on their own

Cristian Ponce wore an Indiana Jones costume while meeting co-founder Theo Schaefer. It was launched in 2023 by Entrepreneur First, a startup program that introduces founders to each other before launching an idea at a Halloween party.

The two hit it off, Ponce recalls. Schäfer trained at MIT to master autonomous underwater robots and worked at NASA’s Jet Propulsion Laboratory, probing Jupiter’s moons for aliens. “Crazy stuff,” Ponce smiles. “I was coming from Cal Tech, doing bioengineering,” where he worked on E. coli.

The two bonded over stories about the hardships of being a lab technician. Ponce (above left) especially complained about all the manual labor involved in genetic engineering. A low-tech lab can spend hours with a scientific syringe “pipette” manually moving liquids from tube to tube.

Attempts to automate the process have not been successful, as robots capable of doing so are specialized, expensive, and require special programming skills. Every time scientists need to change the parameters of an experiment—which happens all the time—they ask a programmer to program the bot, debug it, and so on. they will have to wait. In most cases, it is easier, cheaper and more accurate to use a human.

The company they founded, Tetsuwan Scientific, set out to solve this problem by replacing cheaper white-label lab robots.

But in May 2024, the co-founders were watching OpenAI’s multi-model product launch. He tagged Scarlett Johansson same voice). OpenAI showed people talking to the model.

This was the missing link that Tetsuwan Scientific needed. “We’re seeing before our eyes the maddening progress of big language models, their ability to be scientifically justified,” Ponce said.

After the demo, Ponce fired up GPT 4 and showed him an image of a DNA gel. Not only did the model successfully interpret what the image was, it actually identified a problem—an unintended DNA fragment known as a primer dimer. He then presented a very detailed scientific proposal on what causes it and how to change the conditions to prevent it.

Ponce explained that the LLM models were already capable of diagnosing scientific results, but “had no physical agency to carry out the propositions they were making.”

The co-founders were not alone in exploring the use of artificial intelligence in scientific discoveries. Robot AI scientists can be traced back 1999 with Ross King’s robot Adam and Evehowever really hit off with the series of academic papers Starting in 2023.

But the problem, Tetsuwa’s research showed, was that there was no software that “translated” scientific intent—what the experiment was looking for—into robotic execution. For example, the robot has no way of understanding the physical properties of the liquids it pipettes.

“This robot has no context to know. Perhaps it is a viscous liquid. Maybe… it will crystallize. That’s why we have to say this,” he said. Audio LLMs with hallucinations suppressed by RAG can work with things that are “hard to code”.

Tetsuwan Scientific’s robots are not humanoid. As shown in the photo, they are a square glass structure. But they are built to evaluate results and make changes on their own, just as a person does. This involves building software and sensors so robots can perform calibration, fluid class characterization and understand other properties.

Tetsuwan Scientific currently has an alpha customer, La Jolla Laboratories, a biotech working on RNA therapeutics. Robots help measure and determine dose effectiveness. It also raised $2.7 million in an oversubscribed round led by 2048 Ventures with participation from Carbon Silicon, Everywhere Ventures and some prestigious biotech angel investors.

Ponce’s eyes light up when he talks about the ultimate goal of this work: independent AI scientists that can be used to automate the entire scientific method, from hypothesis to reproducible results.

“This is the craziest thing we can work on. “Any technology that automates the scientific method is a catalyst for hyperbolic growth.”

He is not the only one who thinks this way. Others working on AI scientists include a commercial organization FutureHouse and is located in Seattle Potato.