Robotic AI Scientists - Tetsuwan Scientific

A Serendipitous Encounter Sparks Innovation

Cristian Ponce, dressed as Indiana Jones, first connected with his future co-founder, Théo Schäfer, at a Halloween event in 2023. This party was hosted by Entrepreneur First, a program designed to facilitate connections between potential founders prior to idea inception.

Ponce recalls a quick rapport forming with Schäfer, who possessed a strong academic background. Schäfer held a master’s degree in underwater autonomous robotics from MIT and had previously contributed to research at NASA’s Jet Propulsion Laboratory, focusing on the search for life on Jupiter’s moons. “Truly remarkable work,” Ponce remarked with a smile.

The Frustrations of Manual Lab Work

Their conversation centered around the challenges inherent in laboratory work. Ponce, having studied bioengineering at Cal Tech and worked with E. coli, particularly voiced his frustrations with the repetitive manual tasks involved in genetic engineering.

A significant portion of a lab technician’s time can be consumed by precisely transferring liquids between tubes using a scientific syringe, often referred to as “pipetting.”

Automation Challenges and Existing Limitations

Despite efforts to automate this process, widespread adoption has been hindered. Existing robotic solutions are often highly specialized, prohibitively expensive, and demand substantial programming expertise.

Whenever experimental parameters require adjustment – a frequent occurrence – scientists often face delays while waiting for a programmer to modify the robot’s code and resolve any resulting issues. In many instances, manual operation proves more efficient, cost-effective, and accurate.

Tetsuwan Scientific: Bridging the Gap

Driven by these challenges, Ponce and Schäfer established Tetsuwan Scientific, with the initial goal of adapting readily available, lower-cost lab robots to address the automation gap.

However, a pivotal moment arrived in May 2024 during OpenAI’s unveiling of its multi-modal product line, an event notable for the controversy surrounding a voice resembling Scarlett Johansson.

The demonstration of conversational interaction with the OpenAI model proved to be the crucial element Tetsuwan Scientific had been seeking.

The Power of Large Language Models

“We were witnessing the rapid advancement of large language models and their burgeoning scientific reasoning abilities,” Ponce explained.

Ponce tested GPT-4 by presenting it with an image of a DNA gel. The model not only correctly identified the image but also detected an unintended DNA fragment – a “primer dimer” – and offered a detailed, scientifically sound suggestion for preventing its formation.

This realization – a “light bulb moment,” as Ponce described it – highlighted the potential of LLMs to diagnose scientific outcomes, yet they lacked the “physical agency” to implement their recommendations.

Developing Robotic AI Scientists

The concept of robotic AI scientists isn’t entirely new, with roots tracing back to Ross King’s “Adam” and “Eve” robots in 1999. However, recent years, particularly since 2023, have seen a surge in academic research in this field.

Tetsuwan’s research revealed a critical missing component: software capable of translating scientific objectives into robotic actions. The robots require an understanding of the physical properties of the liquids they handle.

“The robot needs contextual awareness. It must recognize whether a liquid is viscous or prone to crystallization. We must provide this information,” Ponce stated. He further noted that audio LLMs, enhanced with RAG to mitigate hallucinations, can effectively manage “hard-to-code” scenarios.

Current Progress and Future Vision

Tetsuwan Scientific’s robots are not designed to be humanoid; instead, they feature a square glass structure. Their development focuses on enabling them to independently evaluate results and adjust parameters, mirroring the actions of a human scientist.

This involves creating software and integrating sensors to facilitate understanding of calibration, liquid class characterization, and other relevant properties.

Currently, Tetsuwan Scientific is collaborating with La Jolla Labs, a biotech company specializing in RNA therapeutic drugs, as an alpha customer. The robots are assisting in measuring and assessing dosage effectiveness.

The company recently secured $2.7 million in an oversubscribed pre-seed funding round, led by 2048 Ventures, with participation from Carbon Silicon Ventures, Everywhere Ventures, and several prominent biotech angel investors.

Ponce envisions a future where autonomous AI scientists automate the entire scientific method, from hypothesis generation to the production of reproducible results. “This is the most impactful work we could undertake,” he asserted. “Any technology that automates scientific inquiry will be a catalyst for exponential growth.”

Other organizations pursuing similar goals include the non-profit FutureHouse and Seattle-based Potato AI.