Ryan Hibbs

Self-taught AI researcher in British Columbia, working independently through MINA Labs Research Inc. I built a complete training stack for state space language models from scratch in C++ and CUDA — tokenizer, corpus pipeline, data-parallel trainer, CUDA inference backend, evaluation harness — and used it to train a 139.5M-parameter Mamba to convergence on consumer GPUs. That work surfaced previously undocumented two-phase representational structure in Mamba layers, found via CKA analysis; a paper is in progress. My current research is cooperative dual-architecture systems: two language models sharing continuously-running randomized reservoirs, where every write to shared state carries its provenance — human input, internal inference, recalled memory — routed through distinct projections so origin is dynamically distinguishable rather than metadata. The goal is an architecture whose reasoning is auditable by construction. I work at the systems level deliberately. Building the trainer myself is what made the CKA finding visible: I added per-group gradient logging because I couldn't afford to waste a run on consumer hardware, and that instrumentation later caught a gradient-clipping confound in a replication study that a black-box workflow would have missed. Interests: state space models, reservoir computing, dendritic computation, interpretability, cooperative multi-agent architectures.

LinkedIn

LinkedIn

Twitter / X

Twitter

Projects

No grants recorded.