I build experimental software around AI, music technology, local intelligent systems, and unusual ideas worth testing.
I like taking ideas that sound slightly ridiculous, building them, measuring what actually happens, and keeping the failures when they teach us something useful.
TinyBrain is my open research project exploring a non-Transformer recurrent AI architecture with explicit associative memory and continual-learning experiments.
Current research areas include:
- recurrent / GRU-based learning
- explicit delta associative memory
- continual learning and catastrophic forgetting
- selective memory replacement
- fast-weight memory
- compositional generalisation
- representation learning
The repository preserves successful, negative, and inconclusive experiments rather than presenting only the wins.
TinyBrain has now reached an open research problem: it can store and selectively update structured facts, but it does not yet reliably generalise to completely unseen entity/relation/value combinations.
👉 Read the research
👉 Help investigate the current problem
Cadenza is a free, open-source software arranger workstation for Windows, built with C++ and JUCE.
Play chords with your left hand and Cadenza generates a full backing arrangement in real time — drums, bass, accompaniment and arranger sections — while you play melody on top.
It includes work around:
- live auto-accompaniment
- Intro / Main / Fill / Ending arranger sections
- MIDI and style playback
- Yamaha style support
- native style editing
- MIDI-to-style conversion
- SoundFont / FluidSynth playback
- optional AI-assisted style generation
- Artificial intelligence beyond standard LLM architectures
- Continual learning and memory
- Recurrent neural systems
- Local AI and AI agents
- Experimental machine learning
- Music software and digital instruments
- MIDI, synthesis and arranger systems
- Building useful things and testing weird ideas properly
Python · PyTorch · C++ · JUCE · CMake · Git · Windows · AI coding agents
I'm especially interested in people who want to contribute to:
- TinyBrain — continual memory, recurrent learning, associative memory, compositional generalisation
- Cadenza — arranger algorithms, MIDI, music theory, style playback and music software
Criticism and failed replications are welcome too. If an experiment is wrong, I'd rather know.
Can a small recurrent system learn concepts and relationships instead of simply memorising familiar combinations?
That's the problem TinyBrain is currently leaving open for anyone interested enough to attack it. 🧠