noyzless

Introduces open-source sparse attention kernels (Flash-MSA) for training frontier models on NVIDIA Hopper/Blackwell, addressing efficiency gaps in training sparse attention architectures.

Analysis of migrating an AI agent from Claude Opus to GPT-5.6, detailing critical lessons in tool schema handling, prompt caching, and evals.

Thoughtful synthesis on the role of LLMs in software engineering. Argues for evolving via tool mastery rather than fearing obsolescence, with practical tips on LLM-assisted coding and code review.

Explains how Math.tanh in Chrome 148+ leaks OS-specific fingerprinting data via floating-point math differences and suggests mitigation strategies.