#llm

Grade an LLM against the classic linguistic layers — syntax, semantics, pragmatics, social — and the failure lands in one precise place: not meaning, but truth. Hallucination is semantically well-formed falsehood, and the layers that work are what makes the one that doesn't hard to spot.

Karpathy coined "vibe coding" in February 2025 as a fun way to throw away weekend projects. The 2026 audits show what happens when the workflow leaks into production — the debt is real, it is comprehension debt, and it was deferred rather than erased.

Prompt engineering was the 2023 job that mostly meant typing nicely at a chatbot. In 2025 Karpathy and Shopify's CEO renamed the real version of it — context engineering — and the rename was a promotion.

The real engineering discipline behind shipping reliable LLM features is not prompting — it is evals. They are to LLM apps what unit tests are to deterministic code, except you measure pass-rates and distributions instead of exact equality.

Three protocols showed up to standardise agentic AI — MCP for agent-to-tool, A2A for agent-to-agent, ACP for lightweight agent messaging. One of them is already dead.

'Agentic' was the word of the year, and Gartner reckons only about 130 of the thousands of vendors selling agents are shipping one. Here is the precise definition — a loop with tools, planning, state, and retry — the line that separates a real agent from a single LLM call wearing a cost...

The Model Context Protocol went from an Anthropic side-project in November 2024 to 97 million monthly SDK downloads by March 2026. Here is what it actually is — JSON-RPC, three primitives, a client-server split — why it won by turning M×N integrations into M+N, and why just a protocol...

A 2026 study found humans are quietly being gentler on AI-authored PRs than on human-authored ones. Single-reviewer review already broke at 400 lines per diff. The fix is two old ideas — distribute the review across the team, and pair every diff with a deterministic map of the code flow.

Brooks said adding people to a late project makes it later. Adding LLMs follows a similar pattern, except the cost lands on a different axis — output rises while shared understanding of the system erodes, and the resulting cognitive debt eventually has to be paid.

Apple's "do not hallucinate" prompt is funny. The same logic is why AI-written + AI-reviewed pipelines silently approve bugs, why long contexts compound errors, and why the real guardrails are deterministic.

Cursor, Kiro, Windsurf, Antigravity, Trae. All forks of VS Code, all wrapping one of three model APIs, all selling a long system prompt that does not move the model's ceiling. The vibe-coding tech-debt numbers were always pointing somewhere else.