It's striking how narrow the research scope becomes once you constrain yourself to vLLM/SGLang, especially if you try to move even epsilon outside the architectures those backends support. Stepping outside that at scale very quickly turns into serious technical debt. Saying this

James Alcorn Alex Shan I flipping love these guys.

Our goal at Reflection AI is to train the best open models in the world. People like Behrooz Behrooz Ghorbani help make that a reality. Having previously pioneered reasoning models at OpenAI, Behrooz will now lead science of scaling at Reflection. Come find us at Neurips.

Come meet some of the Reflection research team and our founders Ioannis Antonoglou and Misha Laskin at our NeurIPS booth.