Tech News

Amazon replaced Claude Sonnet 4.6 with its own decision model in its ticket system and cut the cost per ticket by 98%. This article explains what a decision model is, how it differs from a general LLM, and when it makes sense to make the same switch.

A broken object level authorization flaw let attackers control connected cars with nothing more than a license plate. The pattern is called BOLA, it tops OWASP’s API risk ranking, and it shows up in any IoT platform that forgets to check who actually owns a resource.

BenchUp is a free, open-source application for modeling furniture in 3D before cutting the wood. It runs in the browser or as a native app on Windows, macOS, and Linux, with no account or payment required.

A 284.33-billion-parameter MoE model on CPU runs in 64 GB of RAM without using a single GPU, thanks to aggressive quantization, selective expert offloading, and context caching. The llama-modchip project serves as a practical case study for understanding how these three techniques work together.

The OpenStick, a $10 ARM gadget with 512 MB of RAM, serves as an edge case for thinking through low-RAM self-hosting. With Debian, systemd, and Podman, several useful services run without maxing out memory, as long as you choose carefully what to install.

Discord showed off a new client that uses 70% less RAM than its Electron app. We explain what the Electron framework is, why every app bundles its own Chromium, and what alternatives like Tauri and WebView2 exist to avoid that.

Static recompilation translates a video game’s binary into C code without emulating the original console. NFSU2Recomp, N64: Recompiled, and XenonRecomp show how this technique preserves Xbox, Xbox 360, and Nintendo 64 games, even bringing them to WebAssembly.