Infrastructure
Every Infrastructure story we have explained, newest first — 13 stories so far.
Photo: Original: Utkarshraj Atmaram Vector: Pduive23 / Public domain · source Cloudflare Open Sources a Platform That Tracks What Agents Read
Cloudflare has open sourced Cloudflare OS, which tracks what data agents read and checks that anyone viewing their output is allowed to see it.
Via WIRED · source Military GPS Jamming Preceded a Fatal Medevac Crash in New Mexico
A US military GPS jamming exercise blinded civilian aircraft across New Mexico. What that had to do with a fatal medevac crash near Ruidoso.
Via Earendil · source A Coding Agent With Four Tools Outperformed Its Bigger Rivals
Databricks tested AI coding harnesses on its own codebase. The most minimal one, Pi, had the highest pass rate at a fraction of the cost of rivals.
Via GitHub · source A 304-Billion-Parameter Model Now Runs on One AMD GPU
A 304-billion-parameter DeepSeek model now runs in production on a single AMD MI300X GPU, unquantized, after hand-written fixes to the serving stack.
Photo: Computer Entertainer (Marylou Badeaux and Celeste Dolan) / CC BY 4.0 · source An Xbox Outage Blocked Games People Owned on Disc
An Xbox outage locked players out of disc games they owned. Here is why the plastic in the tray no longer guarantees you can play.
Three Million Amazonians Lived on Three Percent of the Forest
Archaeologists estimate three million people lived in the pre-Columbian Amazon, concentrated in about 3% of the forest. Why the clustering matters more.
Photo: Unknown authorUnknown author / Public domain · source Record Low Danube Forces Hungary to Shut Its Only Nuclear Plant
Hungary's Paks nuclear plant is shutting down entirely because the Danube has fallen below its cooling intakes — during a heatwave that spikes demand.
Lower Precision Lets Cloudflare Double AI Requests Per GPU
How Cloudflare stores AI models and their KV caches in fewer bits to double concurrent requests per GPU, and what it does to guard the shared memory.
Photo: Stanford Institute for Human-Centered Artificial Intelligence (permission obtained by email from the AI index research manager) / CC BY-SA 4.0 · source AirLLM runs 70-billion-parameter models on a single 4GB card
AirLLM runs 70B and even 671B-parameter AI models on consumer GPUs by streaming one layer at a time from disk. Here's why that works.
Photo: Original: Apple Inc. Vectorization: Totie / Public domain · source An 80-Billion-Parameter Model Now Runs on a Mac in 4.3 GB
Swiftlet runs an 80B-parameter Qwen model on a Mac using 4.3 GB of RAM by streaming expert weights from SSD. How the trick works, and what it costs.
Photo: Utkarshraj Atmaram / Public domain · source Tailscale Did Not Stop the Hugging Face Intrusion
Tailscale's post-mortem on the Hugging Face AI agent intrusion: no bug was exploited, but 181 rogue machines joined the network on one stolen key.
Google Dominated The Tools Behind RSS, Then Abandoned Them
Google ran the most-used RSS reader and the biggest feed service, then closed or froze both. A look at what that did to open web subscriptions.
Photo: U.S. Navy photo by Photographer’s Mate 2nd Class Jason Jacobowitz / Public domain · source AI firms are hiring electricians and carpenters by the thousands
AI firms are recruiting and training electricians, carpenters and pipefitters to build data centres. Why the boom is real — and why it may not last.