Cline is a coding agent that lets you approve every action inside VS Code and roll back at any time. It also detaches its core from the host via a protobuf/gRPC boundary (ProtoBus + HostBridge), so the same core runs in the editor and in a terminal CLI. We analyze Plan/Act approval, checkpoints, and the host abstraction, contrasting them with OpenCode's headless engine.
OpenCode externalizes model metadata to models.dev and even hand-rolls its own LLM protocol layer, so any provider attaches with a single line of data. A single Effect-based HTTP engine is shared by the TUI, web, desktop, Slack, and editors (ACP), while two generations — legacy and V2 — coexist. We analyze the structure against Qwen Code's single-vendor platform.
Paper notes on InstructGPT covering its core method — the three-step RLHF recipe (SFT → reward model → PPO) — the alignment result where a 1.3B model beats 175B GPT-3, gains in truthfulness and toxicity, the alignment tax and PPO-ptx, and the limitations.
Paper notes on GPT-3 covering its core ideas: 175B scaling, in-context learning (zero/one/few-shot), weighted-sampling training data, headline benchmark numbers, and the data-contamination and bias limitations.
A retrospective on building a macOS desktop app (Tauri + Rust + React) with an AI agent. The decision to make markdown files — not the database — the source of truth, the research I did before writing any code, the features I chose not to build, and the macOS permission bug that took the longest to crack. A story about judgment more than code.