Manual handwritten CUDA operator optimization is being ended by AI programmatic compilers.
The system built in 20 days has just open sourced 9 Megakernels: on DeepSeek, embodied robots and world models, the performance exceeds NVIDIA's official stack by 2~3 times. Operator-level ultra-large fusion eliminates cross-core memory transfer bottlenecks.
The DeepSeek Harness example feels practical to me.
Univer SDK lets you build a Space where the agent and its files stay together, so editing a draft doesn't mean copying it into another tool.
"DeepSeek completes what TikTok can't." Your scrolling habits plus what you type into AI chatbots give foreign platforms a complete profile of you.
What have you told an AI app that you'd never post publicly? irongatevc
DeepSeek stopped being responsive after I praised Mbappe irl
You have reached the end of the archive
All of deepseek