How does Discovery Engine differ from traditional scientific analysis?

Traditional methods are often hypothesis-driven and path-dependent, limiting scope. Discovery Engine is relatively hypothesis-free, systematically and unbiasedly analyzing data to uncover patterns that might not have been initially hypothesized.

Can you give an example of a scientific discovery made with Leap Labs' tools?

Yes, they identified novel markers associated with T-cell receptors, found a synergistic effect between manganese and plant genotype impacting root architecture, and challenged a foundational assumption in meteorological modeling.

How does Discovery Engine compare to language models like Claude for data analysis?

Language models alone struggle with large numerical datasets. When given access to Discovery Engine, Claude's ability to synthesize information for scientific discovery was significantly enhanced, highlighting the need for specialized tools.

What are the potential risks of current AI models in scientific discovery?

There's a concern that AI models can generate highly plausible but incorrect scientific papers. This poses a challenge to the scientific literature's reliability, especially in frontier research where ground truth is hard to verify.

What kind of data can Discovery Engine analyze?

The tool is domain-agnostic and can analyze any data. Leap Labs is expanding to multimodal datasets (vision and tabular) and offers a free dashboard service for academics and open-source data.

How much faster is Discovery Engine compared to manual analysis?

Leap Labs claims their system is about 100 times faster, reducing the process from months of iterative work to a few hours of systematic pattern extraction.

Where can I learn more about Leap Labs and Discovery Engine?

You can find more information on their website, leaplabs.com. They are also launching a free dashboard service for academics.

Key Moments

⚡️Automating Scientific Discovery - Jessica Rumbelow, Leap Labs

Latent Space Podcast

Science & Technology3 min read28 min video

Nov 2, 2025|1,881 views|57|6

Save to Pod

Key Moments

TL;DR

Leap Labs' Discovery Engine automates scientific discovery by finding hidden patterns in data, accelerating research.

Key Insights

Traditional scientific discovery often involves tedious manual analysis and is limited by human intuition and existing literature.

Leap Labs' Discovery Engine uses neural networks to identify novel patterns in data that humans might miss.

The engine automates the process of training models, extracting patterns, and contextualizing them with existing research.

It has successfully identified novel markers in immunology, synergistic effects in plant biology, and challenged foundational assumptions in meteorology.

Language models alone struggle with raw scientific data analysis; integrating them with tools like the Discovery Engine enhances their capability.

Leap Labs offers a free web app for academics and researchers to utilize the Discovery Engine for open-source data analysis.

THE LIMITATIONS OF TRADITIONAL SCIENTIFIC RESEARCH

Current scientific research often relies on manual data exploration, guided by intuition and existing literature, which can be prone to biases and replication issues. Researchers spend significant time on data wrangling and analysis, accessing only a fraction of potential discoveries within their datasets. This manual, hypothesis-driven approach, while historically fundamental, is inefficient and limits the scope of exploration. The process can take months, exploring only what is suspected to exist, rather than uncovering entirely new phenomena.

THE ORIGINS OF LEAP LABS AND THE QUEST FOR INTERPRETABILITY

Jessica Rumbelow's journey to Leap Labs began with an interest in interpretability, stemming from her academic work in histopathology. She built neural networks to diagnose cancers and identify immune cells, but a key insight emerged when a model could identify immune cells in a way human pathologists couldn't understand or replicate. This sparked the question: what signals or patterns has the model learned that are hidden from human observation? This led to the idea of using neural networks not just for automation, but as a tool for discovery and a lens to reveal unseen patterns in data.

INTRODUCING THE DISCOVERY ENGINE

Leap Labs' Discovery Engine is an end-to-end system designed to revolutionize scientific discovery. It ingests arbitrary scientific datasets, automatically trains numerous neural networks, and systematically extracts learned patterns using proprietary interpretability methods. These patterns are then contextualized with existing literature, ranked by novelty and prevalence, and presented to human researchers in an understandable format, enabling the identification of genuinely new scientific insights at unprecedented speed and scale.

SUCCESSFUL APPLICATIONS AND REAL-WORLD DISCOVERIES

The Discovery Engine has yielded significant scientific advancements. In immunology, it identified novel markers predictive of T-cell receptor reactivity to tumors. In collaboration with a plant biologist, it uncovered a synergistic effect between manganese content and crop genotype on root architecture, crucial for drought or flood resilience. Furthermore, it challenged a foundational assumption in meteorological modeling (surface layer theory) by finding it invalidated in a significant percentage of coastal locations, a discovery with immense implications for weather prediction and infrastructure planning.

LANGUAGE MODELS AND THE NECESSITY OF SPECIALIZED TOOLS

While powerful, large language models (LLMs) alone are not inherently suited for analyzing complex, arbitrary numerical datasets. When tasked with scientific data analysis without specialized tools, LLMs tend to hallucinate, overgeneralize, and struggle with the nuanced, data-driven requirements of frontier science. However, when LLMs, like Claude, are given access to tools like the Discovery Engine, their capabilities are dramatically enhanced. They excel at synthesis and understanding human priorities, making them powerful partners for data-driven discovery when augmented with robust analytical engines.

THE FUTURE OF AUTOMATED DISCOVERY

Leap Labs is actively expanding its platform to include support for multimodal datasets (e.g., vision and tabular data) and is launching industry pilots for R&D acceleration. They aim to make the Discovery Engine freely accessible via a web app for academics publishing open-source data. The core value proposition remains accelerating discovery by processing data orders of magnitude faster than manual methods, reducing months of iterative analysis to a few hours. This advancement promises to democratize advanced data analysis and hasten scientific progress across various domains.

Mentioned in This Episode

●Supplements

●Software & Apps

●Companies

●Organizations

●Concepts

Common Questions

Leap Labs has developed the Discovery Engine, an AI system that trains neural networks on scientific data to find novel patterns that humans might miss. It uses interpretability methods to extract and contextualize these patterns.

Topics

Meteorology Plant Biology

Mentioned in this video

Concepts

Aggregate cement ratio

A factor in the concrete compressive strength dataset that, when combined with other conditions, has a synergistic effect on strength.

Concrete Compressive Strength dataset

An open-source dataset used to demonstrate the Discovery Engine dashboard, which has been analyzed many times previously.

Surface layer theory

A foundational assumption in meteorological modeling that Leap Labs' Discovery Engine found to be untrue in 20% of coastal samples.

Ash to cement ratio

A factor in the concrete compressive strength dataset that, when combined with other conditions, has a synergistic effect on strength.

T-cell receptors

Identified as important for understanding why certain T-cells are reactive to tumors; Leap Labs found novel markers associated with them.

Meteorological modeling

The field that underpins decisions on offshore wind farms and hurricane path prediction; Leap Labs' engine found issues with a foundational assumption.

Software & Apps

Discovery Engine

Leap Labs' end-to-end system that takes scientific data, trains neural networks, and extracts novel patterns using interpretability methods.

leaplabs.com

The website for Leap Labs, where users can find more information and potentially access the Discovery Engine.

arXiv

An online repository where Jessica Rumbelow is concerned about the increasing number of plausible but potentially untrue scientific papers being submitted.

Lane Space blog

Where Ben Horwitz's concept of 'thinking with tools' was discussed.

Media

Solid Gold Magikarp

A project Jessica Rumbelow is famously known for, which brought attention to tokenization and Reddit.

Companies

Leap Labs

The company founded by Jessica Rumbelow focused on automating scientific discovery through interpretability tools.

Organizations

National Center for Atmospheric Research

The institution where collaborators Patrick and Casey work; they collaborated with Leap Labs on the meteorological study.