Retrieval Is Not State
Challenge the assumption that improvements in context size, retrieval quality, model scale, and memory systems will eventually produce something equivalent to long-term memory, understanding, and reliable reasoning. …
Blog
Evaluating Local Models as VSCode Agents
Evaluating local models on repository investigation and planning tasks and comparing Qwen3.6, Gemma4, GLM-4.7-Flash, and GPT-OSS-20B, across multiple dimensions of repository-agent quality. Our results show that …
Local Web Research with SearXNG and Crawl4AI
This article describes a local research stack that combines SearXNG and Crawl4AI to provide current web search and page retrieval for local language models. It explains how the components interact, how they are …
Local Vision-Language OCR Benchmark
We benchmark open-weight vision-language models for local document OCR and semantic page reconstruction. The goal is to identify a practical production configuration that balances extraction quality and operational …
Fast LLM Judging with No-Think Mode
We explore the use of “no-think” mode for LLM judges in a local evaluation harness. No-think mode suppresses the reasoning phase, which can improve latency and compliance for strict-output tasks. Our …
Re-evaluating Ollama's LLM Performance on Apple Silicon
Ollama’s local LLM performance on Apple Silicon reveals that only a subset of models delivers practical latency for short-response tasks. The results highlight execution characteristics over theoretical …
Was there an Actual Shift in AI Model Performance
In early 2026, developers across the open-source ecosystem observed a sudden improvement in the usefulness of AI-generated bug reports. This article explores whether this shift was due to a leap in model capability or …
Trendslop: Why LLMs Struggle with Strategy
A recent study reveals that large language models, when tasked with generating strategic advice, tend to produce generic, consensus-aligned recommendations rather than context-specific insights. This phenomenon, dubbed …
Embedding Models on Affordable Cloud VMs and Apple Silicon
A benchmark of CPU-only embedding inference from low cost DigitalOcean droplets to Apple Silicon, focusing on runtime overhead, CPU tier effects, and scaling limits.
Security Gating as a Control Problem
A practical approach to securing LLM agents by treating tool invocation as a control boundary. Why fail-closed gates, monotonic risk checks, and adversarial benchmarks are more robust than prompt-only defenses for …