Andrej Karpathy is right: the biggest AI capability gap is not between people who write better prompts. It is between workflows with durable context and workflows that still force models to start from zero.
Two blind copywriting tests for a project called CodexWork showed that AI slop is rarely about bad English. It is a failure to read where the visitor is standing: which page they are on, what they already know, and what the next click should feel like.
Generic AI reviewers fail where enterprise context begins. The next valuable AI code reviewer will behave like an internal digital staff engineer: platform-aware, company-aware, human-looped, and accountable for merge trust.
The AI industry's pursuit of a single omnipotent God Model is a dead end. Due to the mathematical paradoxes of RLHF and alignment, models are inevitably bifurcating into specialized hemispheres—and the future belongs to Agentic Frontal Lobe orchestration.
In early 2025 I started repo-graph-rag to build token-free, traversable repository intelligence. The recent code-graph wave validates the problem, but also explains why I moved upstream anyway.
After building DocMason with AI agents, I stopped treating vibe coding as prompt craft. The durable layer is repository governance: Dao, Fa, Qi, Shu, AGENTS.md, skills, planning, verification, and closeout.
I grew an AI code review GitHub App to 500+ installs and 4,000+ repos. Here is why the rise of vibe coding forced me to abandon the paid tier and pivot to building local context infrastructure.
Most ChatGPT Plus users never touch Codex. DocMason turns that unused capacity into a local, evidence-first knowledge base for decks, spreadsheets, PDFs, and emails on Mac.
Most AI code review tools generate 10-20 comments per PR. The problem? 80% are noise. Here's a framework for measuring signal-to-noise ratio in code reviews - and why it matters more than you think.
How LlamaPReview's repository-wide context analysis caught a production-breaking bug that human reviewers missed—and why the future of code understanding lies in graph-based RAG.