Field notes on artificial intelligence
Models / tools / researchBuilding a local AI coding agent in Zsh: native JSON parsing, HTTP over TCP, incremental Ollama streaming, a responsive zdraw interface, and the engineering choices that made it practical.
A measurement-driven account of tuning Ollama and a small parade of local language models on a Ryzen laptop with an RTX 3070 8 GB—from unified-memory disasters and dense-model failures to MoE, MTP, context cliffs, and the models that finally worked.
Read article : Desperately Seeking LLMs: What Actually Works on an 8 GB Laptop GPUA hands-on breakdown of four multi-agent AI experiments focused on autonomous constitution-building, including failures, philosophical pivots, and the technical lessons learned.
Read article : Forging AI Autonomy: Inside Four Experiments in Multi-Agent Constitutional ConsensusA detailed developer retrospective on designing, iterating, and debugging a Python web server that generates real-time, single-file websites using both remote and local large language models.
Read article : Building Aether Architect: Lessons from Creating a Live AI Website Generator with Local LLMsA technical review of recent research revealing how LLM providers might overcharge users through tokenization, why transparency alone isn't enough, and how per-character billing could close the loophole.
Read article : LLM Billing Exposed: How Tokenization Obscures True Costs and What to Do About ItThis article presents a rigorous benchmark of 18 leading LLMs tasked with generating a constraint-heavy children’s story, analyzing prompt engineering protocols, compliance metrics, editing effort, and the evolution of system prompts for maximal rule adherence and editorial efficiency.
Read article : System Prompts Versus User Prompts: Empirical Lessons from an 18-Model LLM Benchmark on Hard Constraints
