Articles
Long-form writing on tech, culture, and the edges of the internet.
torch.load runs attacker code before the first denoising step
Moebius is a 226M-parameter inpainting model its authors say matches the 11.9B FLUX.1-Fill-Dev at 26ms per step via linear attention and latent distillation.
Rejecting AI code even when it works
Vinicius Brasil on why he rejects AI-generated code he can't explain, even when it passes CI, and why review is now the real bottleneck.
Kees Cook deleted strncpy from the Linux kernel
Linux 7.2 removed strncpy after six years and about 362 commits. Here are the strscpy, strtomem_pad and memcpy replacements and when to use each.
There is no free()
Claude Code's extended-thinking output is a summary, not the reasoning that drove the agent, so it cannot serve as an audit trail of agent logic.
This isn't a bug. It's the default.
OpenAI's Codex persisted TRACE-level logs to SQLite by default, writing tens of terabytes to local SSDs. The cause, the damage, and the fixes.
A model brute-forced Linear A
A self-taught engineer used Claude Code to build corpus tools behind his claimed decipherment of Linear A, now under review at Rutgers and Cambridge.
Google IPv6 crossed 50%, your IPv4 controls didn't follow
Google measures IPv6 at 50% of users, APNIC at 42%. Why they differ, and why controls scoped to IPv4 miss the v6 path on dual-stack hosts.
Linux kernel deleted strncpy across 360 patches
Linux 7.2 removed strncpy() from the kernel after six years and 362 commits. What the replacements are and how to port out-of-tree code.
The same-origin policy is not protecting your API
Zoom encoded a localhost server's responses in image dimensions to dodge CORS, exposing it to every website on the internet.
AMD's Memory Encryption Blind Spot
AMD removed TSME memory encryption from non-PRO Ryzen 9000 chips in a firmware update and will restore the BIOS option in July.
An AI engineer skipped the prediction test
A self-taught AI engineer used Claude Code to back his claim of deciphering Linear A, but the prediction test that confirms a decipherment is still pending.
Trust does not carry forward
Bigger LLMs hallucinate more, not less: why calibration numbers belong next to capability scores when you pick a model for production.