Show HN: Pulse-Field – O(N) AI Architecture (12x faster than Transformers) (github.com)

Pulse-Field is a radically different AI architecture announced as an event-driven, field-based alternative to Transformers that claims linear O(N) scalability and effectively “infinite” context windows. Benchmarks versus a GPT-2–style Transformer on equivalent hardware report dramatic wins: a purported perplexity-like “Energy Defect” of 1.55, accuracy 95.0% vs 67.5%, latency 5 ms vs 60 ms (≈12× faster), and a 20 MB model footprint versus 300 MB. The team (self-described as an “orchestra of AIs” coordinated by a human) also reports stable performance up to 100k tokens with only small latency and memory increases, plus a full test suite (60/60 pytest passing) and stress/coverage checks. Technically, Pulse-Field abandons dense matrix layers for sparse, directed routing of dynamic data packets called Impulses across a graph of specialized nodes (“Crystals”). Impulses carry a semantic payload and an energy budget; routing decisions cost energy and coherent reasoning preserves energy while incoherence increases an “Energy Defect,” which the system uses to suppress hallucinations. Hybrid Crystals perform fuzzy neural reasoning, deterministic symbolic ops, and memory retrieval (HNSW archive) enabling RAG-like “forever” context and end-to-end interpretability via tractable traces. If validated independently, the approach could shift cost/latency trade-offs, enable true edge LLM reasoning on CPUs, and reframe model design toward sparse, physics-inspired computation.