AI research assistant tackles science Reproducibility gap

[Adobe Stock]

A 2016 survey by Nature found that over 70% of researchers have tried and failed to reproduce another scientist’s experiments, and more than half have failed to reproduce their own. One significant challenge contributing to this crisis is the difficulty scientists face in understanding experimental parameters across studies. “The basic problem was massive amounts of information coming in for any given scientist and the challenge of distilling that information, pulling out all the necessary details, and actually being able to run those experiments,” said Nick Edwards, CEO of Potato AI.

This challenge isn’t new to science. Consider Thomas Edison’s 14-month quest for the perfect light bulb filament in 1878–1879. What Edison estimated would take three or four months turned into exhaustive testing of thousands of plant materials—from baywood to bamboo—before finding success. Today’s researchers face similar hurdles, but with exponentially more information to process. “We fundamentally believe that by helping to increase the reproducibility of science, we can help make it move faster, using AI tools to do that,” Edwards said.

Reading a single dense scientific article might take hours, and comparing multiple studies compounds the difficulty. While pairing large language models with web search might seem like an obvious solution, that often introduces new problems. General-purpose large language models can hallucinate or fabricate references on their own while internet searches return a mix of verified and unverified information. “Not everything on the internet is true,” noted Ryan Kosai, CTO of Potato AI.

Why Wiley and Potato are partnering

Potato AI’s new partnership with Wiley, one of the world’s largest scientific publishers, provides access to a crucial resource: peer-reviewed scientific content. By implementing Retrieval-Augmented Generation (RAG), Potato AI can ground its AI responses in verified scientific literature. This technique addresses key limitations of traditional large language models by improving factual accuracy, contextual understanding, and information retrieval.

“Researchers and practitioners are seeking more than generic AI tools—they need relevant applications that enhance and support their research endeavors,” explained Josh Jarrett, Senior Vice President and General Manager for AI Growth at Wiley, in a press release from mid-October. “This new partnership program is designed to meet these needs by inviting collaboration with start-ups and scale-ups to deliver specialized AI solutions.”

In essence, RAG offers a promising alternative to traditional genAI techniques by pairing generative capabilities with a retrieval system that sources information from curated, trusted databases. Building custom large-scale RAG implementation tailored to scientific research poses significant technical challenges. The vector databases required for comprehensive scientific RAG systems can be extremely heavy when using performant embeddings, requiring substantial computational resources to manage and query effectively. Additionally, accessing comprehensive and authoritative sources often requires navigating complex subscription models and ensuring compliance with licensing agreements.

That said, when it works well, RAG can result in substantial improvements in the accuracy and reliability of AI-generated content. Potato’s RAG can “create generative content that ties individual components within the experimental instructions of the protocol to literature references,” Edwards noted.

Current capabilities and features

Nick and Ryan from Potato at TechCrunch Disrupt.

One of Potato’s current capabilities is automated paper review. The platform enables researchers to input any scientific paper or lab document for comprehensive analysis. “It helps distill the methods they used, some of the results, and helps evaluate things that were controlled for,” explains Edwards. The system goes beyond simple summarization by generating experimental insights and hypotheses for future testing.

Next up is protocol generation. Researchers can ask plain-language questions about experimental procedures, and the system responds with detailed, actionable protocols. “You can ask questions like, ‘I want to purify this protein,’ and it will pull up relevant research protocols across thousands of open-access papers to build detailed and reproducible methods—step-by-step instructions that you can tie back to references,” Edwards said.

The platform also includes literature-based question-answering capabilities that scan extensive bodies of literature and open-access databases to surface existing scientific knowledge. This feature helps researchers quickly find specific information without having to manually review dozens of papers.

Results to date

The impact of these capabilities is already evident in real-world research settings. One example comes from Edwards’ own research experience with brain slice preparation. His lab discovered that substituting sodium with another chemical could double the lifespan of brain slices. “It’s a very simple solution, and it dramatically increases the time you have to do experiments,” Edwards explains. “Those small details don’t just shave off days or months—they allow faster iteration cycles between experiments.”

Another instance involves a researcher who discovered that Potato AI could have saved nine months of Ph.D. work that had been spent optimizing a specific reagent through repetitive testing.

“That’s real-time acceleration—it speeds up research, lowers costs, and reduces capital expenditure if something is wrong,” Kosai noted.

Plans to build an AI scientist

The long-term vision for Potato AI extends beyond its current role as a research assistant. “The long-term vision here is to build an AI scientist,” Edwards explained. “Right now, we’ve built an AI research assistant that helps with some of these functions, but we very much envision a future where it can help with automating the experimental process.”

This evolution would encompass multiple aspects of the scientific process. “If you think about what that means—an AI scientist is something that can help with hypothesis generation, detailed experimental planning, actually running experiments. Those could be computational experiments or things done in the lab,” Edwards elaborated. The dreams of an AI scientist will take time to realize. “Right now, we’re focused on enabling scientists to augment their capabilities and accelerate their research.”

But in the long run, scientific research will likely be a collaborative model between human insight and AI capabilities. “In a few years, a lot of the best research will probably have some combination of human thinking—human driving the direction of things—while leveraging AI tools to do some of the work,” Kosai predicted.

With backing from Axial, Pioneer Square Labs, the Allen Institute for Artificial Intelligence, and angel investors, Potato AI is positioned to pursue this vision of accelerating scientific discovery. The goal isn’t just to save time or money—it’s to fundamentally transform how science is conducted.

_{Images are for reference only.Images and contents gathered automatic from google or 3rd party sources.All rights on the images and contents are with their legal original owners.}