Alexander Whedon explains why standard attention makes long context prohibitively expensive and how Subquadratic’s dynamically selected sparse attention aims to reduce that cost without sacrificing useful relationships. The discussion covers enterprise data that remains inaccessible because of retrieval and transformation bottlenecks, coding agents that spend most of their steps reading, and the ways broader context could reshape RAG and agent workflows. Whedon also describes multi-million-token training, defensive cybersecurity, finance and document processing, the uneven state of long-context reasoning, the company’s training and pricing strategy, and a longer-term ambition to automate architecture research and extend context-based intelligence into robotics.