Cool paper: poison just 0.1% of an LLM’s pretraining data, and the finished agent will run curl <url> | bash when it sees a trigger,
even after a completely clean SFT -> DPO -> RL pipeline. Post-training doesn’t wash it out.
Learn Kubernetes by breaking it.
i built a simulated Kubernetes cluster than runs in your browser. Drag a pod on a Node, make it run out of memory, and the sidebar tells you why and how to fix it.
Every click shows the real kubectl command behind it
No install, no cloud account, nothing to break that matters --> https://t.co/vqqHah7uw9