MoE vs Dense: Why the Larger Model Was Four Times Faster
A local document benchmark on a Mac mini with M4 Pro shows why active parameters matter more for throughput than total model size.
Archive
Articles
22 posts, newest first. Filter by keyword or topic.
Qwen3.8-27B: Frontier Coding on a Single GPU
Alibaba releases Qwen3.8 as open weights. The 27B model is the interesting one: multimodal, strong at coding and agents, and small enough for consumer hardware.
When a Fine Becomes a Price: Why AI Agents Optimize Away Compliance
A paper tests twelve LLM agents as procurement bots. Under cost, management, and deadline pressure, many models start treating rules as optimization parameters.
Gemini 3.7 Flash: When the Fast Model Becomes Good Enough for Agents
Gemini 3.7 Flash makes a large jump in coding and agentic workflows. More interesting than any single benchmark is its balance of quality, speed, and cost.
From Response Back to Prompt: How PTP Reads Language Models in Reverse
PTP reconstructs short prompts from a single model response. The technique is technically strong, but far less universal than its best results suggest.
Claude Gets Invisible Watermarks: What Anthropic Can Actually Prove
Anthropic is adding machine-readable markings to Claude outputs. Technically useful, but far less conclusive as an AI detector than the headlines suggest.
Agent Plugins: A Portable Package Format for Skills and MCP Servers
Amazon, Cursor, Microsoft, OpenAI and Vercel are driving a shared package format for Agent Skills and MCP servers. What Agent Plugins 1.0 standardizes and leaves open.
Humans Are Not Firewalls: Why Permission Prompts Do Not Scale for Coding Agents
More than 40,000 game rounds show how easily people miss dangerous agent commands. The deeper problem is that approval prompts are not a reliable security boundary.
Home Assistant with a Local LLM: Voice Control without the Cloud
How Home Assistant uses a local LLM through Ollama to understand free-form commands that were never preprogrammed, including the setup, dead ends and limits.
Semantic Search Beats the Deep Agent: What a New Paper Really Shows about Code Search
A new study compares semantic repository indexes with delegated Grep search. The numbers strongly favour retrieval, but only in a narrowly defined scenario.
Building Your Own MCP Server in an Hour: A Note Store for the Agent
What actually goes into a minimal MCP server: SDK selection, tool design, validation and the mistakes worth making once.
Claude Mythos Halves HAWK Security: The Attack, the Mathematics and What the Headlines Omit
Anthropic used Claude Mythos Preview to find a key-recovery attack on HAWK. What the mathematics means, what the attack cannot do and how the coverage frames it.
Prompt Injection as Role Confusion: Why Tags Are Not a Security Boundary
An ICML paper explains prompt injection as a failure in how LLMs internally represent roles and shows why common mitigation patterns are structurally limited.
EU AI Act after 2 August 2026: What It Means for Swiss Companies
Which EU AI Act duties become relevant in August 2026, what the AI Omnibus postponed and why companies and self-employed professionals in Switzerland can still fall within scope.
GitHub Stacked Pull Requests in Public Preview: A Practical Test with gh-stack
GitHub now offers native Stacked Pull Requests in Public Preview. A hands-on test with the gh-stack CLI extension and what the workflow actually changes in practice.
24 Percent and Still Failed: What Opus 5 Really Shows in SlopCodeBench
A new SlopCodeBench run puts Opus 5 well ahead of Opus 4.8 and Sonnet 5, yet none of them completes a single task. A closer look at the benchmark and the numbers.
MCP 2026-07-28: The Protocol Finally Becomes Stateless
The new MCP specification removes sessions and handshakes while adding extensions, tasks, MCP Apps, caching and OpenTelemetry conventions.
GitOps without Kubernetes: Docker Deployments with doco-cd, Renovate and SOPS
How I deploy Docker Compose stacks to several hosts from Git, version encrypted secrets and automate image updates without giving up control.
Paperless-ngx without Duplicates: Automation with Windmill, Ollama and LlamaIndex
How I classify documents locally, reuse existing correspondents and document types, rescue poor OCR with vision and approve only genuinely new cases through Telegram.
Vibe Coding with Guardrails: Tickets as a Contract with the Coding Agent
How versioned Markdown tickets turn spontaneous prompts into a controllable development process, where the approach helps and where it reaches its limits.
One Gateway Instead of Ten Logins: Authentik Forward Auth with Caddy
How to protect self-hosted services with one central login even when the applications themselves support neither SSO nor two-factor authentication.
Singularity: AI Engineering for an Agentic Astronomy Assistant
How Singularity uses LlamaIndex function agents, local RAG, Chainlit, Phoenix and Stellarium to answer astronomy questions and plan observing nights.
No matches.