Self-Routing: Parameter-Free Expert Routing from Hidden States

Self-Routing is a research-stage algorithmic contribution (5 days old, no adoption signals) that proposes eliminating learned router parameters in MoE layers by directly using hidden state subspaces as expert logits. This is a clever incremental innovation over standard MoE routing (e.g., from Switch Transformers, Base Layers, or GShard), combining existing MoE concepts with a parameter-reduction insight. DEFENSIBILITY: Score of 3 reflects zero traction (0 stars/forks), prototype maturity, and algorithmic nature. No moat exists—the core idea is easily implementable as a 20-line modification to any MoE layer. No community, no data gravity, no network effects. The contribution is intellectually sound but commoditizable. FRONTIER RISK: HIGH. Frontier labs (OpenAI, Anthropic, Google DeepMind) are actively shipping MoE-based models (GPT-4 experts, Claude's architecture patterns, PaLM variants). Parameter-free routing directly addresses model efficiency—a core optimization vector for both training and inference. Frontier labs would either (a) integrate this as a feature into their MoE implementations, (b) publish a similar idea simultaneously, or (c) discover it during routine ablations. This is not a specialized niche—it's a direct infrastructure optimization competing with platform-level improvements. The paper is likely weeks away from a similar ArXiv submission from a major lab. NOVELTY: Novel combination. The subspace-as-logits insight is new, but it's an incremental refinement of learned routing, not a breakthrough. Combines standard MoE design with a parameter-removal observation. COMPOSABILITY: Pure algorithm. Implementation is a reference prototype; integration would be 1-2 lines in existing MoE codebases (swap router layer for subspace projection). RISK FACTORS: (1) Algorithm is platform-agnostic and easily stolen; (2) Frontier labs have existing MoE infrastructure and can absorb this in days; (3) No moat—success depends entirely on adoption/citation, not defensibility; (4) Research-only (no product/service layer); (5) Zero current traction means no lock-in or community. The arXiv paper is the only asset, and ideas have no IP protection. A polished paper + open-source reference could drive citations, but will be superseded by frontier lab improvements within 6-12 months.