1.5x faster MoE training with custom MXFP8 kernels
Achieving a 3.5x MoE layer speedup with a complete rebuild for Blackwell GPUs.
research ·
Iterating with shadow workspaces
Hidden windows and kernel-level folder proxies to let AIs iterate on code without affecting the user.
research ·
More problems
Several exciting problem areas for the next phase of AI-programming.
research ·
Editing Files at 1000 Tokens per Second
A new model and inference method for high-accuracy full-file edits at 1000 tokens/s.
research ·
Our problems
A list of problems we are excited to solve for Cursor.
research ·
Inference characteristics of Llama
A primer on inference math and an examination of the surprising costs of Llama.
research ·
Prompt design
Prompting is like web design. Let’s call it prompt design, and build better tools for it.
research ·