MEGATRON hosts the ULTRATHINK research system -- a large-scale collection of spiking neural network (SNN) and efficient attention research papers, each with its own codebase, experiments, and dedicated Claude Code agent.
Metric Value
Total Projects 129+ Waves 8 Storage 234 GB (~/Papers/) Projects with CLAUDE.md 112 Projects with Code 16 Compute 2x RTX 3090 (48 GB VRAM total) Agent System Claude Code with per-project specialist agents Experiment Tracking MLflow Notebooks JupyterHub
Wave Papers Directory Range Theme
Core 6 paper/, paper3/, paper4/, paper6/, paper7/, paper8/ Original SNN research Extended (paper10-25) 16 paper10-25 ROSA extensions and variants Wave 2 10 paper26-35 Extended SNN architectures Wave 3 10 paper36-45 Advanced SNN applications Wave 4 15 paper46-60 Real-time audio/signal processing Wave 5 20 paper61-80 Thorpe TB trillion-neuron architecture Wave 6 15 paper81-95 Differentiable tree machines + SNN Wave 7 15 paper96-110 DeepSeek R2/RNN + SNN integration Wave 8 20 paper111-130 Domain-specific applications
Project Directory Purpose Status
ROSA ViT Papers/paper/ Rank-Order Spiking Attention for vision transformers Reviewed Grandmother Papers/paper3/ Invertible Rank-Order Autoencoder Reviewed, has code SpikeGPT Papers/paper4/ O(N) Attention via Spike Timing for language Reviewed Memory Engine Papers/paper6/ Hebbian Memory for Continual Learning Reviewed Imagination Papers/paper7/ SNN-based Video Generation Reviewed Audio Engine Papers/paper8/ SNN Audio Processing Reviewed
Project Directory Purpose Has Code
SpikeMoE Papers/paper26/ Mixture of Experts with Spike-Gated Routing Yes NeuroMamba Papers/paper27/ State Space Models with Spiking Dynamics Yes TempoBERT Papers/paper28/ Temporal Coding for Language Understanding No SpikeFlow Papers/paper29/ Normalizing Flows with Neuromorphic Primitives No ROSA-3D Papers/paper30/ Rank-Order Attention for Point Clouds No SpikeDreamer Papers/paper31/ World Models with Spiking Inference No NeuroRetrieval Papers/paper32/ Spike-Based Dense Retrieval No EventNeRF Papers/paper33/ Neural Radiance Fields from Event Cameras Yes SpikeCodec Papers/paper34/ Neuromorphic Video Compression No NeuroNav Papers/paper35/ Spiking Networks for Robot Navigation Yes
Project Directory Purpose Has Code
SpikeSSL Papers/paper36/ Self-Supervised Learning via Spike Timing No NeuroGAN Papers/paper37/ Spiking Generative Adversarial Networks Yes SpikeDiff Papers/paper38/ Diffusion Models with Neuromorphic Dynamics No TempoGraph Papers/paper39/ Temporal Graph Networks with Spike Communication Yes SpikeRAG Papers/paper40/ Retrieval-Augmented Generation with Spike Memory Yes NeuroSAT Papers/paper41/ Boolean Satisfiability via Spiking Networks Yes SpikeCLIP Papers/paper42/ Contrastive Vision-Language with Spike Alignment Yes TempoRL Papers/paper43/ Reinforcement Learning with Spike-Time Rewards No NeuroCast Papers/paper44/ Time Series Forecasting with Spiking Reservoirs No SpikeKAN Papers/paper45/ Kolmogorov-Arnold Networks with Spike Functions Yes
Wave Theme Count Key Projects
Wave 4 Real-time Audio/Signal 15 SpikeAudio-RT, NeuroVoice, SpikeDSP, SpikeSpeech Wave 5 Thorpe TB Architecture 20 ThorpeBrain-Core, SparseSpike-Trillion, NofM-Encoding Wave 6 Differentiable Tree Machines 15 SpikeTree-Core, TreeSpike-Parse, NeuroTree-Logic Wave 7 DeepSeek/RNN Integration 15 DeepSpike-R2, SpikeRNN-Hybrid, NeuroSeek-Reason Wave 8 Domain Applications 20 SpikeFluid-3D, NeuroCyber-Detect, SpikeWeather
See individual wave pages for full project listings:
Each project has a dedicated Claude Code agent (Sonnet-class) coordinated by a master orchestrator (Opus-class):
orchestrator-agent (Opus)
|
+-- rosa-agent (paper/)
+-- grandmother-agent (paper3/)
+-- spikegpt-agent (paper4/)
+-- memory-agent (paper6/)
+-- imagination-agent (paper7/)
+-- audio-agent (paper8/)
|
+-- spikemoe-agent (paper26/)
+-- neuromamba-agent (paper27/)
+-- ... (one agent per paper26-45)
|
+-- audio-snn-agent (Wave 4 coordinator, paper46-60)
+-- thorpe-tb-agent (Wave 5 coordinator, paper61-80)
+-- tree-snn-agent (Wave 6 coordinator, paper81-95)
+-- deepseek-snn-agent (Wave 7 coordinator, paper96-110)
+-- domain-snn-agent (Wave 8 coordinator, paper111-130)
See Agent Architecture for details.
Machine Role GPU RAM
MEGATRON Full training, distributed experiments 2x RTX 3090 (48 GB) 251 GiB Lambda Laptop Ablations, single-GPU experiments 1x RTX GPU 64 GB Mac Studio Inference, visualization, large models M2 Ultra (MPS) 512 GB MacBook M4 Development, quick tests M4 Max (MPS) 64 GB
Each project undergoes systematic review with scoring across four dimensions:
Dimension Weight Description
Concept & Design 25% Research hypothesis, novelty, related work Implementation 30% Code quality, tests, documentation Data & Experiments 25% Dataset setup, experiment configs Results & Analysis 20% Metrics, visualizations, paper readiness
Review artifacts per project: SUMMARY.md, CLAUDE.md, GAPS.md, PROPOSALS.md
# Check all project status
for dir in paper paper3 paper4 paper6 paper7 paper8 paper{26..45}; do
echo "=== $dir ===" && head -5 /home/lucas/Papers/$dir/REVIEW.md 2>/dev/null
done
# Run orchestrator
python /home/lucas/ultrathink_orchestrator.py status
python /home/lucas/ultrathink_orchestrator.py launch
python /home/lucas/ultrathink_orchestrator.py collect
# Run reviews
python /home/lucas/Papers/ultrathink_review_orchestrator.py status
python /home/lucas/Papers/ultrathink_review_orchestrator.py launch
# GPU status for training
nvidia-smi --query-gpu=index,memory.used,utilization.gpu --format=csv