FutureHouse

Just published at Asimov Press: An inside look at FutureHouse, a nonprofit in San Francisco that is working to build an AI Scientist. Bryan Duoto sat down with both Samuel G. Rodriques and Andrew White to learn: 1. Which problems were surprisingly simple to solve, and which turned out to be deceptively hard? 2. What kinds of data are still needed to build an AI scientist? 3. How are their wet lab automation efforts going? What are the bottlenecks. Check it out (and subscribe!): https://lnkd.in/eZwRiNbn

One half of an AI scientist is rejecting or accepting hypotheses. ScienceMachine and FutureHouse just put out ~300 novel hypotheses from ~50 published papers along with ground-truth data. Humans take 4.2 hours to solve these and frontier models get 10-20% correct. This is like SWE-bench for comp bio - so if you get a good score, you make new discoveries instead of closing issues in Django's github. Here's an example hypotheses (+ a dataset) Truncating ASXL1 mutations will lead to specific gene expression changes in blood that reflect alterations in hematological processes, such as T cell and neutrophil activation. And the open answer ground-truth: Gene ontology (GO) analysis of differentially expressed genes (DEGs) in Bohring-Opitz syndrome blood samples revealed significant enrichment for hematological processes, including T cell activation (p-adj = 3.23E-8) and neutrophil activation (p-adj = 1.90E-5). This suggests that ASXL1 mutations notably impact immune-related pathways in blood samples. It also comes in T/F and MCQ variants, if you like that kind of eval. See the benchmark here: https://lnkd.in/gKhrjRBE The arxiv paper: https://lnkd.in/gwBWGrgs And an overview: https://lnkd.in/grSzZHRv

?? Unlocking New Frontiers: 2nd LLM Hackathon for Applications in Materials and Chemistry ?? Join us on May 8-9th for the 2nd Large Language Model Hackathon for Applications in Materials and Chemistry! Last year's hackathon showcased the promise of LLMs, with participants showcasing new projects in structured information extraction, property prediction, novel software interfaces, education, and more. This year, with even more powerful models at our disposal, we invite you to push the boundaries of what's possible. ?? Paper from last year's event: https://lnkd.in/gxayv2j7 ?? Registration: https://lnkd.in/gs7VjCtH Whether you join us in person or virtually, you'll have the opportunity to network, learn, and contribute to cutting-edge research that could transform the future of materials and chemistry. Don't miss this chance to be part of something extraordinary. Registration is open now. Stay tuned for more details on our featured speakers, judges, and in-person locations! Save the date: May 8-9th, 2024. Together, let's unlock new frontiers in materials and chemistry research! Stay tuned for more details on our featured speakers, judges, and in-person locations! Save the date: May 8-9th, 2024. Together, let's unlock new frontiers in materials and chemistry research! Registration: https://lnkd.in/gs7VjCtH