Daywire Radio.

Hello everyone, and welcome to Daywire. Let's make sense of the tech news. It's October twenty six, two thousand twenty six. We begin with our lead story. Cerebras says its new WSE‑3T wafer can hold a single 16 GB dense or linear‑attention layer, and it strings several wafers together to cover trillion‑parameter models. By using linear‑attention, which skips the key‑value cache, two 16 GB layers fit on each wafer, and a compiler places the layers while an RDMA‑enabled runtime shuttles token activations between wafers in about two microseconds per hop. The result is low‑latency inference that hits 88.3 percent on Terminal‑Bench two point one and 91.2 percent on BrowseComp. Engineers can try it now via Cerebras’ free cloud API, but they have to respect the 16 GB‑per‑layer limit and the cache‑free trade‑off when scaling across wafers.

Next on the bulletin. POSIX wordexp, the little‑known C library function for expanding strings, actually hands the input to a system shell unless you explicitly set WRDE_NOCMD. That means a payload like "$(whoami)/foo" gets executed, not just expanded. Both musl and macOS’s libSystem still call out to external helpers – musl to the shell, macOS to a wordexp‑helper – and those helpers can miss patterns, so even the WRDE_NOCMD flag isn’t a guarantee. Here’s why that matters: if you ever feed user‑supplied data into wordexp, you’re opening a command‑injection door. The safe move is to skip wordexp entirely and use a parser that never invokes a shell.

Turning to our next story. A hobbyist got the inexpensive ESP32‑S3 to run the FAST corner detector’s pre‑test using its hidden SIMD unit. By hand‑crafting assembly and a tiny register allocator called basm, they squeezed signed‑8‑bit compares—the only vector compare the chip supports—into the algorithm. The trick relies on forcing the central pixel to be aligned and reinterpreting the image bytes as signed data. That means you can prototype lightweight computer‑vision pipelines on sub‑$5 hardware, provided you respect the alignment and signed‑only comparison quirks.

Next on the bulletin. Goose, a new open‑source systems language, says it can run code up to three point three times faster than C++ while using less memory. It compiles to C, frees memory automatically at scope exit, and avoids a garbage collector or heap allocation by using bump‑pointer allocation and fixed‑size array types. The developers measured 1.2 to 1.9 times lower memory use in their own benchmark suite. Here's why that matters: if you need deterministic memory management without lifetimes, Goose might be worth a test, but you should verify the numbers on your own workloads before committing.

And finally, our last story. Vim Classic 8.3.0 hit the streets on June two, twenty twenty‑six, marking the first stable fork that locks in Vim eight point x without any generative‑AI bells and whistles. The community runs the code on SourceHut, keeps the docs online, and fields questions on its mailing lists, so you get the same Vim feel with a predictable maintenance schedule. Here's why that matters: teams that need a steady editor lifecycle can adopt Vim Classic and dodge future AI‑driven changes while keeping full Vim functionality. And that's your Daywire briefing. You'll find the full stories and sources in today's newspaper. Thanks for listening. Take care, everyone.