Skip to content
ENTITY PROFILE / KNOWLEDGE GRAPH

K-Search

RSS

K-Search has a backend for Apple Silicon’s MLX framework K-Search uses Gemini 3.5 Pro Preview to reason about GPU kernel optimizations and generate candidate kernels Shiyi Cao is the developer of K-Search K-Search was evaluated on CUDA kernels from FlashInfer K-Search was compared against OpenEvolve K-Search was compared against ShinkaEvolve K-Search achieves a 0.97× speedup over the native MLX Attention kernel on Apple Silicon K-Search achieves a 20× prefill speedup over the community mlx‑lm implementation on the Mamba SSM kernel K-Search optimizes CUDA kernels by using existing CUDA kernel knowledge as a starting point Dr. Marily Nika is affiliated with AI Product Academy OpenAI produced a proof of Stein's restriction conjecture· Entity summary generated by Graphiti

RELATIONS
11Relations
CURRENT
11Current
EVENTS
2Events
REPORTS
3Source reports
KNOWLEDGE ATLAS0 entities · 0 relations
No public neighbors yet
Explore all published links of a node; previously discovered relations remain visible.

Relation timeline

BY VALID-FROM DATE
  1. OpenAIPRODUCED PROOF FORK-Search

    OpenAI produced a proof of Stein's restriction conjecture

    Current1 sourceOpenAI
  2. Microsoft ThinkingBoxAFFILIATED WITHK-Search

    Dr. Marily Nika is affiliated with AI Product Academy

    Current1 sourceMicrosoft ThinkingBox
  3. K-SearchEVALUATED ONFlashInfer

    K-Search was evaluated on CUDA kernels from FlashInfer

    Current1 sourceFlashInfer
  4. K-SearchCOMPARED WITHShinkaEvolve

    K-Search was compared against ShinkaEvolve

    Current1 sourceShinkaEvolve
  5. K-SearchOPTIMIZESCUDA

    K-Search optimizes CUDA kernels by using existing CUDA kernel knowledge as a starting point

    Current1 sourceCUDA
  6. K-SearchACHIEVES SPEEDUP OVERMLX-lm implementation

    K-Search achieves a 20× prefill speedup over the community mlx‑lm implementation on the Mamba SSM kernel

    Current1 sourceMLX-lm implementation
  7. K-SearchHAS BACKEND FORMLX

    K-Search has a backend for Apple Silicon’s MLX framework

    Current1 sourceMLX
  8. K-SearchCOMPARED WITHOpenEvolve

    K-Search was compared against OpenEvolve

    Current1 sourceOpenEvolve
  9. K-SearchACHIEVES SPEEDUP OVERMLX Attention kernel

    K-Search achieves a 0.97× speedup over the native MLX Attention kernel on Apple Silicon

    Current1 sourceMLX Attention kernel
  10. Shiyi CaoDEVELOPER OFK-Search

    Shiyi Cao is the developer of K-Search

    Current1 sourceShiyi Cao
  11. K-SearchUSESNVIDIA Omniverse libraries

    K-Search uses Gemini 3.5 Pro Preview to reason about GPU kernel optimizations and generate candidate kernels

    Current1 sourceNVIDIA Omniverse libraries

Relations are facts Graphiti extracted from reports, not verbatim quotes. An end date means later reporting showed the relation no longer holds.

Events 2

Source reports 3