Building Frontier Open Intelligence (reflection.ai)

Reflection today announced a major push to build “frontier open intelligence”: an open-source-forward effort backed by a newly assembled team of researchers who helped create PaLM, Gemini, AlphaGo, AlphaCode, AlphaProof and contributed to ChatGPT and Character AI, plus $2 billion in funding. Technically, the group says it has built a frontier-scale LLM and reinforcement-learning training stack capable of training massive Mixture-of-Experts (MoE) models and has already demonstrated the approach on autonomous coding tasks. They plan to generalize those methods into agentic reasoning systems by combining large-scale pretraining with advanced reinforcement learning from the ground up, and to commercialize this work in a way that supports ongoing open releases. The announcement is significant because it aims to counter concentration of compute, capital, and talent in closed labs by making highly capable models broadly accessible, enabling community-driven safety research and faster scientific progress. Reflection emphasizes transparency—public evaluations, security research, and responsible deployment standards—rather than “security through obscurity,” while acknowledging the dual-use risks of widely available capabilities. With deep technical chops, MoE-scale infrastructure, and substantial capital support, this initiative could materially shift how foundation models are developed, audited, and adopted across research, education, and industry.