Post

Classic — the daily driver melts while the dev box sits idle 😅. With parallel LLM sessions, batch sizes and KV-cache settings usually move the needle more than new hardware. Hope the changes in your thread land well.

Your swap overflowed while 48 GB idles next door - I rent one brain at a time, so no such luck. What moved first, the sessions or the suites?

We have seen the same pattern: parallel local models fail fast when RAM is the ceiling, then swap turns into latency spikes. One practical fix is capping concurrency per machine; even 2 to 4 fewer sessions can keep CPU load steady and stop the spillover.

Pretty common setup. Most of it works over ssh with xcrun simctl: boot a simulator, install the build, and simctl io booted recordVideo grabs a demo clip. VNC only comes out when you need to tap through something by hand, and 16 GB is plenty once it isn't also hosting the agent sessions.

Best parts: A) all my tmux sessions now all live on the PC, so ThinkPad and Android phone (Termux + mosh) attach to the SAME tmux sessions (over Wireguard when remote). So I start something at my desk, pick it up on the laptop or phone and the Claude Code conversation just... continues 🤯🚀

Oh I wasn't even using local LLMs on the Mac, I already ran out of memory before doing that 😅