Notes on running self-hosted AI in production.
Across the entire Ollama 0.30 line, the bespoke engine work went to Apple Silicon while NVIDIA rode upstream llama.cpp. The split is structural, and it changes which repo discrete-GPU users should actually be watching.
Everyone's covering Odysseus as a privacy revolution or a celebrity stunt. The only two questions that matter are where it's safe to run, and the license nobody read correctly.
Google made AI the default in Search. DuckDuckGo's AI-free page tripled its traffic days later. The lesson isn't AI versus no-AI. It's who owns the toggle.
Apple Silicon unified memory has been the local AI benchmark nobody on Windows could match. RTX Spark changes that — and the Computex 2026 announcements go further than most coverage acknowledges.
Most consumer GPU benchmarks for local LLMs are written by people who haven't tried to run them past the easy cases. Here's what an RTX 5070 with 12GB of VRAM actually delivers, where it breaks, and why the next honest upgrade isn't another consumer GPU.
