The UK Government's Low Value Purchase System Is a Waste of Time
The UK Government’s RM6237 Low‑Value Purchase System lets public bodies buy goods under a set threshold from registered suppliers, but it forces every supplier to submit a monthly management‑information report—even when no sales occur. A Freedom of Information request revealed that roughly 96% of submissions each month are “no business” filings, with over 1,500 firms reporting nil returns. The requirement consumes about two minutes per report, equating to more than two full days of collective SME time each month, while only 59 firms actually recorded sales. Critics argue the onus should shift to buyers to report spend, not suppliers.
Real-Time LLM Inference on Standard GPUs: 3k Tokens/S per Request
Kog AI unveiled a tech preview of its Kog Inference Engine, delivering 3,000 output tokens per second per request on an 8‑GPU AMD MI300X node and 2,100 tokens/s on an 8‑GPU NVIDIA H200 system. The engine runs a 2‑billion‑parameter model and focuses...
Cloudflare Flagship
Cloudflare introduced Flagship, a native feature‑flag service that runs directly on its Workers edge platform. Developers can define flags with targeting rules, percentage rollouts, and multi‑type values, then evaluate them via a built‑in Workers binding or the OpenFeature‑compatible @cloudflare/flagship SDK....
Netherlands Seizes 800 Servers, Arrests 2 for Aiding Cyberattacks
Dutch financial‑crime investigators arrested Andrey Nesterenko and Youssef Zinad, the co‑owners of MIRhosting and WorkTitans BV, for providing infrastructure that enabled Russian‑backed cyberattacks and disinformation within the EU. The raid on May 18 seized more than 800 servers across data centers...
Memory Has Grown to Nearly Two-Thirds of AI Chip Component Costs
High‑bandwidth memory now makes up roughly two‑thirds of AI‑chip component spending, climbing from 52% in Q1 2024 to 63% by Q4 2025. In absolute terms, HBM outlays jumped from about $12 billion to $32 billion, driving total component spend from $22 billion to $52 billion. Logic‑die...
Constraint Decay: The Fragility of LLM Agents in Back End Code Generation
Large language model (LLM) agents excel at generating functional backend code from loose specifications, but struggle when strict architectural and data‑layer constraints are imposed. A new arXiv study evaluated 100 greenfield and feature‑implementation tasks across eight web frameworks, measuring both...
Greg Brockman: Inside the 72 Hours That Almost Killed OpenAI
Greg Brockman, OpenAI’s co‑founder and president, recounts the 72‑hour turmoil that followed Sam Altman’s sudden dismissal, detailing how the board’s decision nearly derailed the company. He describes the rapid creation of a backup venture, dubbed “Phoenix,” built at Altman’s house...
I Bypassed Adobe and Microsoft to Build a Git-Tracked Book Production Pipeline
A self‑published novelist and software developer replaced the traditional Word‑InDesign‑Kindle Create workflow with an open‑source, Git‑tracked pipeline. He uses LibreOffice ODT as the single source of truth, then runs custom Python scripts to generate both XHTML‑based EPUBs (validated by the...
Google's AI Is Being Manipulated. The Search Giant Is Quietly Fighting Back
A BBC investigation showed that a single, well‑crafted web page can poison AI chatbots, causing Google’s AI overviews, ChatGPT, and others to deliver false answers. The reporter demonstrated the flaw by publishing a hoax about being a world‑champion hot‑dog eater,...
Qwen3.7-Max: The Agent Frontier
Qwen AI unveiled Qwen3.7‑Max, a proprietary large‑language model built for the emerging agent era. The model excels as a coding agent, office‑automation assistant, and long‑horizon autonomous executor, handling thousands of tool calls across multi‑hour sessions. Benchmark results show it leads...
Was My $48K GPU Server Worth It?
In 2024 the author quit a FAANG job and built a $48 K personal GPU server, "grumbl," equipped with six RTX 6000 Ada cards to accelerate independent AI research. By logging utilization and electricity costs, the author found the rig ran at an...
GenCAD
GenCAD is an image‑conditional generative model that produces full 3D CAD models together with the complete parametric command history, or CAD program. It tackles the difficulty of learning from boundary‑representation data by leveraging a transformer encoder, contrastive joint‑embedding, latent diffusion,...
I Turned a $80 RK3562 Android Tablet Into a Debian Linux Workstation
A community‑driven project called rkdebian now lets the $80 Doogee U10 Android tablet run a full Debian 12 Bookworm system from an SD card, without unlocking the bootloader or altering the internal eMMC. The pre‑release image, built on May 14 2026, supports most hardware features—including...
Δ-Mem: Efficient Online Memory for Large Language Models
The paper introduces δ‑mem, a lightweight online memory module that plugs into a frozen full‑attention large language model. By maintaining an 8 × 8 state matrix updated via a delta‑rule, δ‑mem generates low‑rank corrections to the model’s attention during generation. This approach lifts...
New Nginx Exploit
Depthfirst's automated analysis system uncovered a critical heap buffer overflow in NGINX’s ngx_http_rewrite_module, cataloged as CVE‑2026‑42945. The flaw allows unauthenticated remote code execution on servers that use rewrite or set directives, and it was disclosed alongside three related memory‑corruption issues....