Independent AI Researcher & Complexity Scientist
I build agents that don't train. No neural networks, no gradient descent, no reward optimisation. Instead: thermodynamic perception, entropy gradients, and bio-inspired heuristics.
Zero-training entropy agent vs DeepMind's Melting Pot benchmark. 350 lines of NumPy beats trained RL (ACB, VMPO) on Clean Up without ever seeing a reward signal. DOI: 10.5281/zenodo.18717202
Emergent symbolic language through entropy-driven evolution. 40 Unicode glyphs, logistic map chaos. Key finding: selection destroys measurable grammar while creating real structure (compression ratio proves it).
Multi-agent ALife simulation — 13 genes, 12 emergent types, 42 scenarios. Shared language creates tribes through mutual recognition, not cooperation advantage.
Bio-inspired network simulation: 5 tree species communicating through mycelial networks. MycoNet daemon with Hebbian learning achieves 62% lower path cost than OSPF. 239 tests.
Entropy-driven artificial life — 12 behavioural types emerge from simple energy and spatial rules. Weather, gravity, physics, and biome variants.
Mapping emergent psychological disorders in LLMs to DSM-5 categories. 8 AI disorders, 10 manipulation types, 7 control structures.
Translating spiritual language into field-theoretic coherence: Kuramoto model, Shannon entropy, phase-error dynamics applied to religious metaphor.
- Thermodynamic approaches to intelligence
- Multi-agent cooperation without training
- Complexity science and emergent behaviour
- Bio-inspired computation (slime mould, jellyfish, mycorrhizal networks)
- Symbolic cognition and language emergence