Misunderstood and Misaligned (bozhao.substack.com)

The piece argues that the core problem for large language models isn’t bigger networks but “knowledge alignment”: creating the equivalent of philosophical common knowledge between humans and machines. Using David Lewis’s notion of common knowledge, the author shows how even obvious facts can fail to become mutually assured through chains of doubt, and explains that the same fragility crops up when an LLM must infer task-relevant context from imperfect prompts. Prompt and context engineering (task descriptions, few‑shot examples, RAG, multimodal data, tools, state/history and compacting) are difficult, costly, and brittle—easy to prototype with “vibe coding” but hard to scale into maintainable, trustworthy systems. Isoform’s announced approach reframes the problem as conversational alignment rather than model improvement: instead of making coding agents produce code first, they prioritize building a platform that lets humans “show” intent through sustained, enjoyable conversation so the machine accumulates shared context. The implication for the AI/ML community is a shift from maximizing model capacity to engineering persistent, interactive context pipelines and UX that create durable intent understanding—reducing guesswork, lowering long‑term costs, and improving trust and maintainability in production LLM applications.