DeepMind, in collaboration with MIT and Harvard, has unveiled a groundbreaking advancement in scientific research: an AI system capable of autonomously generating expert-level empirical software for scientific problems. The team published a 71-page paper detailing how their system combines large language models (LLMs) with tree search algorithms to automate the creation and refinement of code that solves complex scientific tasks—tasks that traditionally take human researchers months or even years to develop. The AI system works by first using an LLM to generate candidate Python code based on a problem description, evaluation metrics, and relevant data. Each generated program is then executed in a sandbox environment, where it receives a quality score. A tree search algorithm—specifically, a novel variant called PUCT (Predictor + Upper Confidence bound applied to Trees), inspired by AlphaZero’s UCB method—evaluates these scores and guides the search by balancing exploration of new ideas with exploitation of promising paths. This iterative process enables the AI to evolve increasingly sophisticated solutions over time. A key innovation lies in the system’s ability to retrieve and integrate external scientific knowledge. It draws from research papers, textbooks, and AI-powered tools such as Gemini Deep Research and AI co-scientists, incorporating these insights directly into its prompts to guide code generation. This allows the AI to synthesize ideas across domains and discover novel strategies that human researchers might overlook. The system was rigorously tested on 16 Kaggle competitions from 2023, where it was compared against top human participants. Results showed that the tree search (TS) approach significantly outperformed both single LLM calls and even the best of 1,000 LLM-generated attempts. In several cases, the AI achieved “jump-like” improvements in performance, with cumulative gains leading to top-tier solutions. Adding expert advice into the prompt further boosted results, demonstrating the value of integrating domain knowledge into the AI’s reasoning. In scientific benchmarking, the system delivered superior performance across six diverse fields: bioinformatics, epidemiology, geospatial analysis, neuroscience, time series forecasting, and numerical analysis. In genomics, it excelled at removing batch effects in single-cell RNA sequencing (scRNA-seq), a critical challenge in high-dimensional, sparse data. Using the OpenProblems benchmark, the AI proposed 87 novel data analysis methods, 40 of which surpassed the best human-developed models. One standout method, BBKNN (TS), recombined elements from existing techniques like ComBat and BBKNN, achieving a 14% improvement over the previous state-of-the-art. In another test, the system tackled the ZAPBench challenge—predicting full-brain neural activity in zebrafish from over 70,000 neurons. It outperformed all baseline models, including the leading Unet-based video prediction models. The AI developed architectures incorporating time convolution, learned global brain states, neuron-specific embeddings, and even integrated a biophysical neural simulator (Jaxley) in a proof-of-concept experiment. This integration not only improved predictive accuracy but also enhanced model interpretability, suggesting a path toward AI systems that can embed scientific principles into their designs. The research demonstrates that AI can go beyond pattern matching—it can actively explore, combine, and innovate on scientific ideas. By automating the creation of high-quality empirical software, this system offers a scalable, general-purpose framework to accelerate scientific discovery across disciplines. The work marks a pivotal step toward AI as a true collaborator in science, capable of generating not just code, but new scientific methods.