OpenAI Triton: writing GPU kernels in Python becomes practical

Writing fast NVIDIA GPU code is historically hard: you need to know CUDA, manage shared memory, sync, coalescing. A craft of its own.

OpenAI releases Triton, a language (and compiler) that looks like Python and lets you write custom GPU kernels without being an expert CUDA engineer, with performance close to hand-tuned code.

Meaning AI researchers and ML devs can optimize specific layers of their models (attention, normalization, custom losses) without depending on whoever maintains PyTorch. Triton becomes the foundation of FlashAttention a year later, and a stable part of the PyTorch stack.