New Stories
Dutch police say they have largely dismantled an international investment fraud ring that allegedly earned more than €100 million per month, operating about 20 call centres with more than 700 staff posing as financial advisers. Victims were persuaded to invest in cryptocurrency via fake trading platforms where no money was actually invested, with Dutch victims losing about €25 million. The main suspect, a 46-year-old man with Israeli and Polish nationality and a known hacker, was arrested in Poland in May; further arrests followed in Cyprus and Athens in July; authorities say tens of thousands worldwide have been defrauded.
Rung is a daily word game where players guess one hidden answer from five clues. The clues are cryptic at first and become obvious by the end; guessing earlier yields more points, and a wrong guess costs extra points. The game is published by Dreamwake Labs LLC and the page states there are no accounts or cookies—anonymous play counts.
The notes examine how open source projects can be sustained by balancing passion with revenue, including approaches like free and paid tiers, donations, or enterprise pricing, and by applying marketing and user feedback to drive adoption. They stress focusing on a single project to avoid burnout and maintenance overload, and they reference startup-style practices for planning, funding, and growth.
GPU-Tile-Sim is a tile-centric GPU simulation framework for hardware-software co-design of large language model workloads, based on the idea that kernel performance is driven more by dependency structure than by per-instruction latency. It represents kernel execution as a warp-level tile graph with tile-level operations and edges encoding data and ordering constraints, and provides an automatic frontend and a graph-driven backend; evaluations on A100 and H100 show a mean absolute percentage error (MAPE) between 1.22% and 8.71% across GEMM, attention, and end-to-end LLM workloads, with preliminary validation on Blackwell.
The post describes replacing a mock backend with a real Qwen2.5 model in digiKam by wiring llama.cpp inference through a dedicated worker thread to keep the GUI responsive. A bug where the model’s KV cache was not cleared between queries caused searches to fail after multiple runs; the fix is to clear the model memory at the start of each inference. The author argues that deploying an LLM in an application is largely about integration—build systems, packaging, and infrastructure—rather than the model itself, and outlines next steps including automated real-inference tests and cache/ambiguity improvements.
Fludi Leads offers a service that identifies local businesses with no website, a social-only presence, or a website that has lapsed by querying Google's official Places data. Users target a niche by industry and location, see the exact lead count and pricing before paying, and receive instant delivery of leads as a CSV via screen and email. Pricing is £2 per lead for orders under 10 and £1.50 per lead for 10 or more.
The article explains that a modern large language model like GPT-3 is built from 577 spreadsheets—tensors—containing about 175 billion parameters. It describes the first spreadsheet as a dictionary mapping words to numbers and notes the fixed size of 50,257 rows by 12,288 columns, with the other 576 layers following a repeating pattern, all of which drive the model’s outputs. It also discusses how, despite being labeled a black box, researchers can inspect internal structure and that understanding these internals is considered important for safety and responsible AI development.
HeapLens is a Visual Studio Code extension that analyzes Java and Android heap dumps (.hprof) inside the editor. It provides ten interactive views, including dominator trees, class histograms, leak suspects, and waste detection, plus a built-in AI assistant and a HeapQL SQL-like query language. The extension supports multiple LLM providers (including Ollama), can compare dumps, export reports, and navigate to source code, with fast native Rust parsing and telemetry designed to respect VS Code settings.
Stacks is a card solitaire where players place one card at a time into one of four gardens, and harvest occurs when a garden has four cards. Scores are based on the sum of card values multiplied by the longest chain multiplier, with an additional 5x multiplier for same-life-group cards; players may redraw their hand up to three times by moving the hand to the deck's back and dealing again from a 40-card deck.
Agentmetry is an open-source flight recorder and security layer for AI agent tool use. It captures every tool call, denial, and human approval as a local JSONL audit trail with MITRE ATT&CK enrichment and argument hashing, and can forward the trail to SIEMs such as Loki, Elastic, or Splunk.
The author argues that asking 'what do I want' is less productive than asking 'who I want to be,' drawing on Stoic ideas and a reference to ChatGPT. They propose defining values concretely as what you would be proud to do in specific situations, noting outcomes are often out of control and morality can guide action. The piece quotes Batman and promotes a Cheat at Search training course by Doug Turnbull starting May 18 with a signup link.
MemExchange is a cluster-wide, multi-tenant memory management system that dynamically right-sizes in-memory caching tenants based on workload demand. It reallocates idle memory across physical nodes using a marginal-utility allocation derived from online Miss Ratio Curve estimates and RDMA, avoiding centralized coordination or forced tenant co-location. Implemented in Memcached and evaluated on up to 100 CloudLab servers, it achieved up to 2.3x lower remote-access overhead, a 13% increase in cluster-wide memory utilization, and up to 63% reduction in miss rate for memory-constrained tenants under skewed workloads.
MCP Queen is described as a project to build tools, connect agents, and host a graded registry of MCP servers. The project claims the registry indexes over 16,000 servers and grades each with receipts, with two graded-evidence servers already active. Health AI, linked to Constat and Clarity, leads the effort.
Publisher ad supply fell by up to 40% in Q2 2026 as AI-driven search choked traffic to open-web sites, according to Ozone benchmarking data. The data show combined programmatic spend for U.S. and U.K. publishers down about 30.6% year over year, while eCPMs rose about 30% in the U.K. and 7% in the U.S., indicating higher prices for a smaller inventory. Analysts say this reflects a shift toward subscriptions, apps, and other high-quality inventory as discovery concentrates in walled gardens.
The article compares Sonnet 5 and Opus 4.8 across 24 tasks from two open-source repos using five reasoning levels in Claude Code, assessing patch quality and cost. It reports that Sonnet yields clearer, more intentional patches at high reasoning effort, while Opus produces simpler, more robust diffs; there is no universal winner, with guidance varying by task type.
Agent K is described as MiB's most capable agent, with admin privileges, access to tools, and the ability to load multiple inference providers and models while delegating tasks to subagents. The article lists features such as planning, expansion, multi-provider/model support, and an A2A server for multi-agent ecosystems, and it provides installation and run instructions including Docker usage and environment-variable options. It cautions that performance and results vary by provider and model and notes there are no safeguards against deleting files or data.
Ron McLeod argues automation and AI must be designed around human psychology to prevent errors in supervisory roles. He notes that real-time supervision is mentally demanding and that misdesigns or unclear authority have historically led to problems in control rooms, aviation, healthcare, and sports officiating.