SPQA: The AI-Based Architecture That'll Replace Most Existing Software (danielmiessler.com)

A new AI application architecture called SPQA (STATE, POLICY, QUESTIONS, ACTION) is being proposed as the successor to traditional “circuit-based” software: instead of hardwired inputs/outputs, systems will be built around large GPT-style models that “understand” context (STATE), organizational goals and guardrails (POLICY), the prompts you ask (QUESTIONS), and the operations they perform (ACTION). Practically this looks like stacking a general base LLM (e.g., a future GPT-6) with a custom STATE model trained on company telemetry and a separate POLICY model that encodes mission, constraints and anti-goals; QUESTIONS drive ACTION to produce artifacts, configurations, analyses and code. The author argues this will let tasks that today take months and dozens of specialists be done in minutes, while updates become trivial iterations of POLICY and QUESTIONS. For AI/ML practitioners this frames a clear technical path and product implications: multimodel stacks, real‑time connectors (Splunk, Slack, GSuite, Salesforce) to stream STATE, modular sub‑models for hot vs cold data, and separate POLICY models for frequent policy changes. Key caveats include hallucinations (human oversight required), training cost and latency for hundreds of terabytes, and the need to design moats around unique POLICY and data. The shift from rule-based to understanding-based systems makes prompt/question design and policy engineering the strategic differentiators — not just model scale — and forces product teams to rethink integration, APIs and competitive defensibility.