Articles
Product
Career
『巨大なのに、動かすのは安い』——DeepSeekやKimiに代表されるMoE(Mixture of Experts)モデルは、数千億〜数兆の総パラメータを持ちながら、各トークンはごく一部のエキスパートだけを通るため、実際に動くパラメータは総容量のほんの一部に抑えられます。本スライドでは、この仕組みの基本から、ルーターの負荷分散が抱える課題、DeepSeekが導入した補助損失なしのロードバランシングとその発展形までを、図解で解説します。
Leído
Siguiente lectura ↓Le enviaremos un correo electrónico cuando se publique un nuevo artículo.
Masahiro Taima
Founder | CEO | Ph.D.
Doctorado en Ingeniería Urbana de la Escuela de Graduados en Ingeniería, Universidad de Tokio. Posteriormente, profesor asistente especial en la Universidad de Keio, profesor asistente especial en la Universidad de Kyoto, e investigador visitante en la Universidad de Londres y la Universidad de Pekín. Se especializa en planificación y análisis de arquitectura e ingeniería civil.
Harsh Chandra
AI Research Scientist | Ph.D.
After graduating from the Indian Institute of Technology, completed a Ph.D. in Mechanical Engineering at the Graduate School of the University of Tokyo. Following research at Caltech and a research internship at NEC, worked on AI/ML development including edge AI, multilingual VLM/OCR, and foundational vision models at Fast Accounting and AIRUCA.
Sozoro
Building Review Practice Lead
Engaged in building permit reviews at a designated confirmation inspection agency with experience reviewing over 5,000 cases. Has approximately 40,000 social media followers. Operates the building code specialty media "Building Standards Law Tora no Maki." Has authored books including "Building Code for Residential Design," edited building code-related legal collections, and supervised building code educational materials. At mign, leverages specialized building code knowledge and review experience in AI agent development.
Jikai Li
AI Research Scientist | Ph.D.
Completed Ph.D. in Electrical and Information Engineering at Niigata University Graduate School. After a research internship at NTT Research Laboratories, worked as a data scientist at Toggle Holdings. Specializes in deep learning, image processing, and GIS.