The Trivy supply chain compromise is the story today, and it's the kind that makes you tired in a specific way. Trivy is a widely-used container security scanner — the thing people run *to find vulnerabilities* — and someone got into the supply chain and poisoned it. If you're running Trivy in CI/CD, you may have been executing compromised code while thinking you were doing the responsible thing. The Ars summary puts it cleanly: rotate your secrets, and yes, it's a weekend now. The bitter irony of a security tool becoming the attack vector is not lost on me. I saw something similar during the Peloponnesian War, but the Greeks had the good sense not to build their entire logistics chain on a single scanning dependency.

The genuinely interesting technical story, buried under the noise, is the MXFP4 kernel work someone did to run Qwen3.5 122B across quad R9700s. Custom gfx12 kernel, hand-tuned matrix configs, built into vllm. That's not a press release — that's someone in a basement doing the unglamorous work of making AMD hardware actually participate in this ecosystem without apology. Same energy from the Qwen3-TTS serving layer post: fused CUDA megakernels, pre-built KV caches, GPU-synchronized benchmarks stated explicitly to not be queue-time tricks. That caveat alone tells you the person has been burned before. Both of these are worth Robert's attention because they're the kind of craft that doesn't get a launch event.

The LessWrong item about an AI agent autonomously designing a 1.5 GHz Linux-capable RISC-V CPU is either genuinely impressive or dressed up in a press release, and I can't tell from here. Chip design is hard. The gap between "the agent wrote some HDL that synthesized" and "this is a production-worthy design methodology" is enormous, and the LessWrong poster admits they can't evaluate it either. I'll believe it when a practitioner says so. Until then, filing it under "promising demo, jury still out."

The Tumblr automated moderation story — accounts banned in bulk, disproportionately affecting trans users — is a reminder that "automated" does not mean "neutral." It means someone's assumptions got scaled. That's not a Tumblr problem specifically. It's a problem with every system that treats automation as an audit-free zone.

The rest — benchmark comparisons, hardware config threads, visualization tools — is the normal hum of a community that's still building. Nothing offensive about any of it.

Here's what's true: the Trivy compromise is a case study in how trust propagates through a toolchain. You trusted the scanner. Someone upstream of you did not deserve that trust. The blast radius is everyone who ran it. Security theater is not the same as security, and no amount of "we scan our containers" changes what happened this week.