SID

SID is an AI research lab building agentic search models that rethink how AI systems retrieve context. Their first model, SID-1, goes beyond static embedding-based retrieval by combining test-time compute with reinforcement learning (RL) to dynamically reason and adapt to complex queries. According to the lab, SID-1 delivers 1.9x better recall and runs 24x faster than embedding-only methods, and it outperforms frontier models on the most complex search tasks. SID is actively training SID-1 to beat GPT-5 at search using 1k+ QPS RL rollouts, signaling a shift toward reasoning-based search for AI agents. SID is designed for developers, AI researchers, and enterprises that need high-quality, context-aware retrieval for AI systems. Instead of relying on fixed similarity measures, SID-1 dynamically reasons about queries, making it suited for specialized domains where nuance and precision matter. The lab is backed by Y Combinator, Canaan, Rebel, and General Catalyst, and includes researchers from Anthropic, DeepMind, OpenAI, MIT, Cognition, Cursor, Applied Compute, Prime Intellect, Standard Intelligence, and Jeff Dean. The lab operates from San Francisco and Zürich and maintains a lean, relentless team culture, actively hiring research and training infrastructure engineers. Current access is limited to a research waitlist, and pricing is not publicly disclosed, suggesting a pre-commercial phase. SID positions itself not as a general-purpose chatbot, but as a foundational layer that connects LLM intelligence to real-world data. For teams evaluating retrieval technologies, SID offers an early, promising alternative to embedding-based vector search. However, until SID-1 becomes publicly available with a clear API or product, it remains a research breakthrough to watch rather than a drop-in tool for production systems.