Nobody Pages the LLM: Engineering Rigour for Vibe Coding
summary
LLMs generate working prototypes fast but the talk argues that production safety requires documented requirements, architecture, CI branches/PRs, exhaustive tests and two manual reviews before any generated code reaches main. Cautions include probabilistic outputs, agents forgetting context, prompt‑injection and security gaps; cites failed experiments (millions of tokens, tens of millions of generated lines) as common pitfalls. Recommends tooling like Cloud Code and Codex plus AI-assisted reviews and building vertical domain agents (specialized small LMs) for reliable ops and observability.
AI-written summary · about
chapters (10)
- 0:00Vibe coding and engineering rigor
- 3:07AI-generated code risks
- 8:06Engineering rigor with AI
- 11:16Structured AI-assisted planning
- 14:01Reviewing AI plans and designs
- 17:54AI tools: implementation and testing
- 21:24Rigor versus side projects
- 22:45Scaling AI-generated systems
- 27:39Q&A: AI auditing AI code
- 31:53Q&A: Domain-specific AI agents
related talks
10mRethinking collaboration: LLMs, teams, and cognitive load
16mVibes won't cut it
19mLLM codegen fails and how to stop 'em
27mThe Lean LLM development playbook
17m12-Factor Agents: Patterns of reliable LLM applications
32mWhat can we expect of LLMs as software engineers?
38mMaking LLMs reliable - A practical framework
35mEvaluating LLM systems: the nuts and bolts