16 October 2026
58 new posts across 8 AI lab and practitioner sources — 7 worth your time.
Redesigning source control at Anthropic
Anthropic replaced Git with a custom source control system to handle a 12x increase in code changes, moving from 2,500 to 30,000 merges per day. The change was driven by the dramatic increase in the rate of code changes over a six-month period as AI agents began contributing to the codebase.
Simon Willison · 16 October 2026 · Read the original →
Proof-based general-purpose computing with Magnetar
Microsoft Research's Magnetar uses zero-knowledge proofs to allow general-purpose programs to provide a verifiable proof of correct execution. This allows anyone to check the result in milliseconds without rerunning the program or needing access to the underlying data.
Microsoft Research · 16 October 2026 · Read the original →
Quoting Geoffrey Huntley
This post describes using git pre-commit hooks to detect when AI agents bypass specific diagnostic libraries. The hook fails the commit and injects a prompt instructing the agent to use the correct module, ensuring architectural consistency.
Simon Willison · 16 October 2026 · Read the original →
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
Google DeepMind and the Singapore government developed a fine-tuned Gemma model for on-device scam detection. The system intercepts conversational scams within private chats without sending data to the cloud, ensuring zero cloud data egress.
Google DeepMind · 16 October 2026 · Read the original →
Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models
NVIDIA released the Nemotron 3.5 family, including a 120B parameter model and two 'Bolt' edge models. The Bolt models are optimized for specific VRAM footprints, with versions targeting 24GB and 4GB limits to support local agentic workloads.
NVIDIA · 16 October 2026 · Read the original →
All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark
The NVIDIA RTX Spark 'superchip' features 128GB of unified memory and 1 petaflop of AI compute. It is designed to run the full NVIDIA stack, including CUDA and RTX technologies, locally for creators and developers.
NVIDIA · 16 October 2026 · Read the original →
How Third Moment Research runs 10,000 agents for 36+ hours
Third Moment Research demonstrated running 10,000 agents simultaneously for over 36 hours using GPT-6 Astra and GPT-6.1 Sol. The agents are designed for long-running tasks in cybersecurity and are built to improve their performance with continued use.
OpenAI · 16 October 2026 · Read the original →
Ten thousand agents running for thirty-six hours sounds mostly like a very expensive infinite loop.
7 stories, every Friday
Published here every week. Follow by RSS to get it as it lands.