Patent attorney: physical AI race will be won in the patent office

Offrun runs coding-agent CLIs locally on Apple Silicon Macs, giving each agent its own git worktree so parallel agents never touch the same files. It uses your existing agent logins, moving a chat to another account when one hits its limit.

A developer spent the past week building a strategy combat game described as "fantasy airship FTL," with crew dungeon runs, cargo quests, and mage crew members. It is playable for free.

A llama.cpp expert-streaming setup on Windows hits ~11.5 tok/s with the UD-IQ3_XXS quant of Qwen3.8-Flash-Next 177B, up from roughly 7 tok/s on the inherited setup. Hardware is an RTX 5070 with 12GB VRAM plus 32GB DDR4 RAM.

A LocalLLaMA user built a dedicated RAG/search node for Qwen3.8 27B using a cheap AliExpress mini PC that arrived with fraudulent hardware instead of the promised Intel N150 with DDR4/DDR5 support.

The list covers Caddy, which turns scattered phone info into calendar events and reminders and has been in public beta since April 2026, plus Fambot, an SMS-based family "chief of staff" for school comms, sports, meals and calendars. Instinct is cited as the buzziest of the space after a $1B round at a $10B valuation.