From foundational models to novel filesystems. We ask one question: how far can on-device intelligence go.
"The bottleneck is not model size. It is training objective. A model trained to orchestrate tools will outperform a larger model trained to recite. That is what our reinforcement learning work keeps confirming."
Anshul Kulkarni
Undergraduate Research Scientist at GPT-Research
Foundational modeling. Reinforcement learning for tool orchestration. Agentic Harness design. Novel filesystem architectures. The goal is frontier-class on-device intelligence.
Learn About Our Research
Every system GPT-Research releases traces to a specific experimental finding. The research determines direction. Releases are evidence.
Explore Our API
