Research
Research
We pursue architectures grounded in physical realism, biological fidelity, and resource efficiency.
Organismic AI (OAI) & BIB 2
Can artificial intelligence possess agency, continuous sensorimotor adaptation, and allostatic survival without backpropagation or cloud GPU clusters?
Organismic Artificial Intelligence (OAI) moves beyond discrete token prediction toward continuous-time dynamical systems. Our flagship substrate, BIB 2 (Biologically Inspired Brain 2), is a complete 1:1 macro-anatomical replica of the human central, peripheral, autonomic, and enteric nervous systems. Running at >540 Hz on standard CPU, it combines 14 biological subsystems, local 3-factor Hebbian corticostriatal plasticity, and Slow-Wave Sleep (SWS) memory consolidation.
In continuous Squash Pong, BIB 2 learns from scratch via retinotopic optic nerve signals (CN_II) and spinal proprioception (C5), achieving a 96.2% net interception rate (-83.3% error reduction) across 5 multi-seed trials and 100% on small-paddle stress tests.
ATOMIC
How do we solve the quadratic memory wall of Transformer attention on consumer laptops?
ATOMIC introduces Gated Linear Recurrent Attention (GLRA) and Zero-GEMM Token Shifting. It delivers an invariant 192KB recurrent state (256× compression over KV caches) and up to 2.62× CPU inference speedup at 4,096 tokens, proving that multi-step Chain-of-Thought reasoning can be trained from scratch on a consumer CPU.
BIM Lineage
The ancestral substrate exploring continuous biological learning from a blank slate. BIM 0 through BIM 4 pioneered Sparse Distributed Representations (16,384 columns, 0.39% sparsity), surprise-scaled Jaccard plasticity, and hippocampal pattern completion without backpropagation, laying the algorithmic groundwork for BIB 2 and Project Cambrian.
MAUL
Specialist routing beats monolithic generalization. MAUL routes incoming requests across dedicated specialists—symbolic math, sandboxed code execution, language reasoning, and external retrieval—achieving >95.9% accuracy on GSM8K.