Leap Labs

We gave Claude a frontier science task: analyse experimental catalyst data (Meta OCx24) and give us three insights about what makes a good CO₂ reduction catalyst. On the surface, Claude looked pretty smart – as requested, we got three insights about catalyst composition, complete with plots to back them up and relevant references. But there was one big problem: none of the ‘insights’ were actually supported by the data. This isn’t Claude’s fault! All LLMs do this – and humans do it too. All too often we look for the patterns that we expect to see, overgeneralise, or fixate on outliers – and miss real insight. So we tried again – same dataset, same prompt – but this time gave Claude access to Leap's Discovery Engine. Our tool for systematically finding novel patterns in data. The result was quite different. We found many non-obvious relationships (e.g. negative synergies between certain metals), and Claude did a spectacular job of synthesising the results from Discovery Engine and providing useful guidelines for catalyst composition. This is a toy example, with one model, single-shot prompted, on a single dataset. But I think it points to an important lesson: if we want LLMs to help with science, we need to give them the right scaffolding. Language models need data-driven discovery just as much as humans do.

We’re announcing three new scientific papers today – in plant biology, immunology, and meteorology – all co-authored with domain experts, all powered by our Discovery Engine. It feels like a milestone for us at Leap, and I wanted to share why: Most science still relies on a person manually analysing data, testing one hypothesis at a time. It’s really slow, and it misses things. We’ve been building something different: a system that makes discovery systematic, 100x faster, and reproducible. Scientific data in, empirical insight out – automatically. And it works! These new papers are based on real discoveries that came out of Discovery Engine – results that scientists were excited to publish. Many of them would have taken months to uncover using standard methods. Some might never have been found at all. All of these came from the same system. Discovery Engine is domain-agnostic, automated, and doesn’t rely on LLMs – just data, deep learning, and interpretability. This is very exciting for the Leap team, but I think the implications are huge for scientific inquiry in general. We can make more science happen! We can make R&D investment a much safer bet. Gathering data, through experiment or observation, can be a huge investment. And right now it’s a risky one – return on it is uncertain and slow to realise. But how much more science would get done if we could squeeze every drop of value out of that investment? How much more could we invest in science if returns were less uncertain? How much more productive would the average researcher be, with Discovery Engine at their fingertips? Leap is now open to industry pilots – so if you want to see what’s hiding in your data (novel insights, major breakthroughs, or even just confirm that there’s nothing new there, so you can move on) get in touch. DMs are open, or book a call: https://lnkd.in/epSAKyMC Read more here: https://lnkd.in/e9jJchWT

Hack the Sciences - a hackathon for scientists who code ⚙️ 🛠️ June 7th-8th, London 🚀 48 hours to build tools that accelerate the natural sciences and promote interdisciplinary exploration. 👩💻👨💻 Come join us in London if you are a researcher, entrepreneur, or simply curious about building at the intersection of science, data and compute. Super excited to be hosting this alongside: Advanced Research + Invention Agency (ARIA) x CompMotifs x Leap Labs Bring your ideas, your drive and get ready to meet some exceptional people 🔥 📌 London, UK 🗓️ 7th-8th June '25 ➡️ Apply here: https://lnkd.in/eb3WFGRq Ines Ullmo Isabel Zhang Kieran Didi Wojtek Treyde Danyal Akarca Rory Byrne Kristina Kordova Austin Mroz Puria Radmard Eva Sevenster Jakub Lála Martin Buttenschoen Filip Szczypiński Aleksy Kwiatkowski

📢  We are very excited to announce that Leap Labs is joining our AI for Longevity Hackathon! They will be running a workshop on: “𝐀𝐈 𝐓𝐨𝐨𝐥𝐬 𝐟𝐨𝐫 𝐒𝐜𝐢𝐞𝐧𝐜𝐞” How AI is changing science is the new playbook for scientific discovery in the age of artificial intelligence: from LLMs that mine millions of papers in minutes to autonomous robots running experiments 24/7. Join for a fast-paced tour of the state-of-the-art in scientific AI, including tools that you can use today – and a look at the future of discovery, where machines not only automate science but invent entirely new ways to ask and answer its biggest questions. 👨💻 👩💻 PLUS: Leap Labs is launching their Open Data Challenge Got a scientific dataset (longevity or anything else)? Submit it for a chance to: 🏆 Win $1K, $2K and $10K 📜 Chance to co-publish 📅 Deadline: June 1 🔗 Submit your data: https://lnkd.in/euXjvWSa 𝐁𝐫𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐜𝐮𝐫𝐢𝐨𝐬𝐢𝐭𝐲. 𝐁𝐫𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐝𝐚𝐭𝐚. 𝐉𝐨𝐢𝐧 𝐭𝐡𝐞 𝐡𝐚𝐜𝐤𝐚𝐭𝐡𝐨𝐧. 𝐒𝐢𝐠𝐧 𝐮𝐩 𝐡𝐞𝐫𝐞: https://lnkd.in/e6Nttg55 Brought to you by GetSeen Ventures and London Longevity. Jessica Rumbelow Jugal Patel #LongHack #LondonLongevity #GetSeenVentures

My team mates have been grinding for months to build our Discovery Engine. And we have made our first scientific breakthroughs! Case studies linked. The Engine is incredibly data efficient. Our first novel discovery is in plant biology from a dataset of less than 1000 rows and 7 columns. We identified a novel genotype/nutrient combination that contributes to a desired root architecture. Our second novel discovery is in meteorological science. It is being validated by domain experts and could update every meteorological model, potentially improving forecasting by a few percentage points. You can imagine the downstream effects of this. Publications upcoming. This is just the beginning! We are working with computational chemists on reaction yield, materials scientists to develop a new catalyst for hydrogen production, a bioinformatician to discover which antigens attach to specific T-cell receptors. We are accelerating science by enabling researchers to make discoveries 100x faster. Scientists spend weeks to months analyzing their data, often missing complex, nonlinear, combinatorial patterns. Disco works in hours and surfaces them with ease. This is where scientific enquiry begins. If you are working with any observational or experimental data and want to increase the efficiency and output of your R&D team give me a shout.

"If an AI can outperform tools designed by humans in a given scientific field, it suggests that the AI system is representing something about the world currently unknown to us." - Open Problems in Mechanistic Interpretability (led by Apollo Research) Deep neural networks can encode complex, non-linear relationships and extract meaningful features without human intervention. Interpretability enables us to gain insight into these predictors and utilize them for knowledge discovery. Leap Labs co-founder Jessica Rumbelow expands on this idea in her section on Microscopic AI. It is a core enabler of our mission to accelerate scientific discovery. #ML4Science #Interpretability #DataDrivenDiscovery #KnowledgeDiscovery https://lnkd.in/eHbzQ5k6