Research
Research
Three active threads. BIM is the main one. MAUL is running. WebAI is an early experiment.
BIM
BIM asks whether a system can learn language from a blank slate. No pre-training, no backprop, no GPU. Only mechanisms found in biological brains.
BIM 0 proved Hebbian learning and interaction loops work. BIM 1 added Sparse Distributed Representations: 16,384 columns, 64 active per token (0.39% sparsity), 131,072 cells total, 1-shot memorization at 100% recall on novel 20-token sequences, ~160 TPS on CPU. BIM 2 merged both, adding real-time weight updates from interaction, synaptic homeostasis, winner cell disambiguation, and a concept graph that crystallizes named patterns from repeated co-activations.
BIM 2 is complete. We ran a small internal benchmark: 100 multiple-choice questions about 2026 events. BIM 2 scored 87. Frozen LLMs on the same test start at around 25%. They have no 2026 training data. BIM 2 had learned the answers.
Next: inference from chained facts. Facts BIM was never told directly, but can derive from what it learned. That experiment hasn't been published for a CPU-native system.
MAUL
One idea: a model that only does math, one for code, one that searches. If a router coordinates them, do they outperform one model doing everything? MAUL is how we test that.
MAUL-2 has five specialists. Code runs in a sandbox. Math goes to SymPy. The other three handle reasoning, language and tone, and conversation/search. Before picking a specialist, it parses what you actually asked.
MAUL-1 with Hydrogen-1: 95.68% on GSM8K. MAUL-2 with Hydrogen-2: 95.90% on GSM8K. The long plan: route between actual trained specialist models, not specialist prompts.
WebAI
Most AI tooling needs a server. WebAI asks what happens when you take it out of the equation.
WebLM-1 is 30M parameters: a decoder-only transformer (10 layers, 8 heads, 512 dimensions) that trains and runs in a browser tab using TensorFlow.js and WebGPU. BPE tokenizer, streaming generation, model export and import, training dashboard with live loss curves. No installation. No API key.
The core loop works. It's an experiment. We're figuring out what to build on top of it.