Nvidia open-sources PAIR to pool home machines into one AI cluster
Free router fans agent jobs across every box on your LAN, halving a demo run. Plus: gh's Linux key expires tomorrow; Google's TimesFM-3 forecaster lands.
Copy markdown
One virtual router, every GPU in the house
PAIR sits in front of Ollama or LM Studio and forwards each request to whichever networked machine is free, then stitches the results back — no changes to your app or agent. Nvidia's demo ran a five-subagent task in ~9 minutes on a three-device cluster versus ~18 on a single laptop.
What you need, and what you could build this weekend
Works across GeForce RTX 20-series and newer, RTX Pro, DGX Spark, and Apple silicon M4+ on Windows, macOS, and Linux. It's a free beta and fully open source, so you can chain your gaming PC and MacBook to run parallel agents — or a model too big for any single box.
It stays on your own network
PAIR auto-detects compatible devices and encrypts all machine-to-machine traffic with mTLS, so nothing leaves your LAN and there's no cloud bill. The bottleneck it targets is the one builders hit now: fan-out agent workloads that stall on a single GPU.
Elsewhere: `gh` on Linux breaks after tomorrow unless you rotate a key
GitHub CLI's Debian/RPM signing key expires Sept 5; after that, apt and dnf throw EXPKEYSIG/NO_PUBKEY and won't update gh. Fix in one line: re-fetch githubcli-archive-keyring.gpg (APT) or re-add the gh-cli repo (DNF). Anyone who installed after Apr 8 already has both keys.
Elsewhere: TimesFM-3 forecasts many signals at once — but not commercially
Google's 330M zero-shot foundation model now does multivariate forecasting (blend sales, weather and other covariates) and tops Gift-Eval, FEV-Bench and TIME. The catch: the weights ship under a non-commercial license, with BigQuery the only sanctioned production path.