DispatchByteDance CUDA Agent: AI Writing GPU Kernels at 98.8% Pass RateImport AI·high signalXBlueskyLinkedInCopy linkFine-tuned Seed 1.6 230B model for GPU kernel generation. 98.8% pass rate, 2.11x speedup vs torch.compile. AI-for-AI-development loop closing.SourceSource pageImport AI↳ Follow the threadPolicy dependency / Stack layerA controlled ablation on a production science agent finds the model choice dominates topology and prompting, and a PPO policy nearly matches it for freearXivPolicy dependency / Stack layerLemonade Now Fronts 15 Inference Engines Behind One Base URL With Semantic and Policy Routingr/LocalLLaMA (Lemonade maintainers)Policy dependency / Stack layerCherry Studio v2.0.9 unifies tool approval into one declarative policy and lets the provider catalog hot-update without an app releaseGitHubStack layer / ContrastParitok-4B compresses coding-agent context to 25.7% of size with a 264 MB adapter, and shows gpt-5 as a compressor is net-negativearXivStack layer / ContrastCyberFactory Turns CVEs From the Wild Into Executable Training Tasks; Aegis Hits 52.4% Pass@1 on CyberGym, +22.8 Over Its BasearXiv 2608.23181Policy dependency / Stack layerr/LocalLLaMA Is Revolting Against Its Own Megathread Policy, Arguing It Killed Release-Day Discussionr/LocalLLaMAStack layer / ContrastClaude Code ships /claude-api cost-optimize, a skill that profiles an existing project's API spend one measured change at a timeClaude Code changelogStack layer / Threat patternZ.ai Held GLM-5.3's Weights Back Two Weeks Over Its Own Cyber Scores, With a Cursor Bug as the Headlinepaddo.dev