Forward Deployed Engineers, AWS Pain, Dead Framework Theory, and Vibe Coding Unit Tests

The Rise of Forward Deployed Engineers

• 🚀 A new AI-related tech role, Forward Deployed Engineer (FDE), has seen an 800% increase in job postings over the last 9 months.
• 💡 FDEs bridge the gap between general-purpose AI models and specific client needs, transforming them into scalable solutions.
• 🎯 Companies like Anthropic and OpenAI are actively recruiting FDEs to embed within customer and product teams, moving beyond traditional coding.

Generational Differences in Cloud Adoption

• ☁️ AWS is described as a painful experience for newer developers, contrasting with smoother platforms like Vercel.
• 🧠 The author suggests this pain point might be generational, with younger developers (Gen Z) preferring platforms not designed as tests of skill.
• 👴 However, the author also notes a preference among older developers for Heroku-style deployments over AWS, indicating it's not solely a generational issue.

Understanding LLM Agents

• 🤖 To truly understand LLM agents, the recommendation is to build one yourself, as it offers a surprising and insightful programming experience.
• 🧩 Building an agent provides clarity on what is likely the most important developer-facing technology of the decade.

The Persistence of Frameworks

• 🔄 The prediction that LLMs would abstract away framework choice has been revised; React is now seen as a platform rather than just another framework.
• 📈 New frameworks and libraries must now contend with a self-reinforcing feedback loop involving LLM training data, system prompts, and developer output, making displacement of React nearly impossible.
• 🛠️ Tools are even hard-coding React into system prompts because developers need maintainable code, and for most, that means React.

Caution on AI-Generated Unit Tests

• 🧪 While LLMs are often considered good for boilerplate and tests, they can produce unconstructive, noisy, brittle, and bad unit tests.
• ⚠️ The advice is to stop 'vibe coding' unit tests with AI; current methods for getting good tests from LLMs require generating them one at a time, which is time-consuming.