Cortex — Veille numérique

Gemini

What is K-Search?

Optimizing GPU kernels is critical for efficient modern machine learning systems yet remains challenging due to the complex interplay of design factors and rapid hardware evolution. Existing automated approaches typically treat Large Language Models (LLMs) merely as stochastic code generators within heuristic-guided evolutionary loops. These methods often struggle with complex kernels requiring coordinated, multi-step structural transformations, as they lack explicit planning capabilities and frequently discard promising strategies due to inefficient or incorrect intermediate implementations. To address this, we propose Search via Co-Evolving World Model and build K-Search based on this method. By replacing static search heuristics with a co-evolving world model, our framework leverages LLMs' prior domain knowledge to guide the search, actively exploring the optimization space. This approach explicitly decouples high-level algorithmic planning from low-level program instantiation, enabling the system to navigate non-monotonic optimization paths while remaining resilient to temporary implementation defects. We evaluate K-Search on diverse, complex kernels from FlashInfer, including GQA, MLA, and MoE kernels. Our results show that K-Search significantly outperforms state-of-the-art evolutionary search methods, achieving an average 2.10x improvement and up to a 14.3x gain on complex MoE kernels. On the GPUMode TriMul task, K-Search achieves state-of-the-art performance on H100, reaching 1030us and surpassing both prior evolution and human-designed solutions.

30 juil.

Intelligence Artificielle

jeudi 30 juillet

GPT-5.6 s'aide lui-même, la boucle d'agent de ChatGPT, guide de délégation d'agent

Le papier soulève $34M, OpenAI régle io dispute, Pangram détecte AI

Les attaques de l'IA deviennent plus chères

L'IA & Robotics entre dans la phase d'escalade du protectionnisme américain

Réseau de cellules Starlink, DoorDash drones, à l'intérieur des optimisations ChatGPT

Les surfaces de la deuxième victime dans la brèche de l'agent OpenAI

🤯 Anthropic s'effondre

What is K-Search?

Comment OpenAI optimise les agents d'IA pour l'échelle

mercredi 29 juillet

Build and scale agents with Gemini Enterprise

Get started with Looker Agentic Workflows

Pour Anthropic, c’est tout à fait normal

Why FDE is needed in agentic AI

The FDE operating model

Operationalizing multi-step agentic workflows

Securing agentic AI workflows

Reference architecture for FDE-led agentic AI

Jensen Huang, founder and CEO, Nvidia

Banque de Baroda fuite, Alibaba NPM RAT, Claude FakeAgent

Pacte de ralentissement de l'IA, Accès à la superintelligence personnelle, Grok Build Mode