Where Nature
Inspires
Intelligence
NatureCast is an open research initiative studying how biological systems — from neurons to ecosystems — can transform the design of artificial intelligence. We publish code, models, benchmarks, and ideas.
Bridging biology and machine learning
Billions of years of evolution have produced extraordinarily efficient information-processing systems. We study these systems — and translate their principles into better AI.
Neuroscience-Inspired AI
Drawing on decades of neuroscience research to design neural architectures that mirror biological learning — attention, memory consolidation, and predictive coding.
Nature-Inspired Algorithms
Evolutionary computation, swarm intelligence, and biological optimization strategies applied to the hardest open problems in machine learning.
State-of-the-Art LLMs
Rigorous benchmarking, interpretability studies, and efficiency research on frontier large language models and multimodal systems.
Figure 1. NatureCast research pipeline — from biology to AI implementation.
A systematic pipeline from biology to code
We study biological intelligence at multiple scales — from single neurons to whole ecosystems — and systematically extract computational principles that can improve artificial systems.
Every project ships working code, reproducible benchmarks, and clear documentation so others can build on our work.
Literature review and experimental analysis of biological mechanisms.
Distill computational principles that generalise beyond biology.
Build, benchmark, and openly publish models and code.
Active Projects
NeuroSynth-SNN
Biologically-inspired spiking neural network framework implementing leaky integrate-and-fire neurons with STDP learning rules. Achieves competitive accuracy at 10× lower energy cost on temporal tasks.
EvoSearch-NAS
Evolutionary neural architecture search using CMA-ES and genetic programming. Discovers efficient transformer variants inspired by natural selection and co-evolution dynamics.
OmniEval-LLM
Comprehensive evaluation framework for large language models across reasoning, coding, science, and instruction-following benchmarks. Reproducible, hardware-aware comparisons.
From the Blog
The LLM benchmark landscape is a mess — inconsistent prompting, unreported hardware, cherry-picked tasks. We present OmniEval, a reproducible evaluation...
Spiking neural networks (SNNs) represent information the same way biological neurons do — in discrete spikes through time. After years in the shadow of ...
The transformer's attention mechanism was revolutionary — but it bears only a surface resemblance to biological attention. We explore the neuroscience o...