AI
Decision Models, Jev and Kev
Jev returns probabilities instead of text. Its confidence is arithmetic, not learned, and calibration belongs to the distribution, not the model.
LLM Calibration and Overconfidence
Calibration is knowing how often you'll be right. How ECE measures the gap, GPT-4's ECE going 0.007 to 0.074 after post-training, and what's known of why.
Harness (Systems) Engineering
After prompts gave way to context engineering, what comes next? Anthropic engineering points to harness design, evals, and containment.
How to Save AI Tokens
Measure AI coding token costs with a React POC and cut them using prompt caching, subagents, lean MCP, context engineering, and Cursor Composer.
How Tokens Work
How BPE creates tokens, embeddings feed LLMs, prefill and decode shape pricing, and KV cache and prompt caching reduce repeated-input costs.
Context Files
How agents load CLAUDE.md, AGENTS.md, SKILL.md, and Cursor rules, why instructions get lost, and what an ETH Zurich study says to put in context files.
Four Tiers of Code Intelligence
Four tiers of tools that cut an AI agent's code search cost: context packing like Repomix, tree-sitter repo maps, CodeGraph's graph, and LSP-based Serena.
MCP and Function Calling
How MCP differs from function calling: six primitives, stdio and Streamable HTTP, the tools/list to tool_use loop flow, and risks like Tool Poisoning.
The AI Frontend Engineer
How frontend engineers grow when AI writes the code: new skills in verification, specification, and judgment, drawn from Karpathy, Vercel v0, and METR.