Reading

Orosz and Noda dig into how companies like Google and Stripe actually measure productivity — not just DORA metrics but the qualitative signals they track. As we think more carefully about AI's impact on engineering velocity at Coursera, the frameworks here are useful. The common thread: every company they spoke to uses both qualitative and quantitative measures.

Simon Willison is honest about what LLMs can and can't do. This post is a careful account of where they genuinely accelerate and where they introduce subtle bugs you need to watch for. That balance matches my experience vibe coding side projects.

Will Larson on engineering strategy — diagnosis, guiding policies, coherent actions. The part he gets right: a guiding policy that doesn't imply a tradeoff should make you suspicious. Most 'strategies' I've seen are just goals dressed up in strategy language. Useful framing when I write strategy docs.

Mollick writes clearly about AI in education. The study he covers here, by his colleagues at Penn, challenges the assumption that AI tutoring straightforwardly improves outcomes: students who used ChatGPT without guardrails leaned on it as a crutch and did worse on the final exam. The nuance around when AI assistance helps vs. when it becomes a crutch is directly relevant to decisions we're making at Coursera about how and where to deploy AI in the learning experience.

A classic that holds up. Allspaw's framing of 'mature engineers' vs just 'senior' resonates with how I think about building engineering culture. The emphasis on not just solving problems but understanding the human and organizational context around them — that's what I look for in senior engineers. I still share it with engineers I work with.

A practical guide to LLM system design. Eugene breaks down patterns like evals, RAG, fine-tuning, caching, and guardrails with real-world tradeoffs. His note on the recall-latency trade-off in retrieval maps to what we run into at Coursera when building AI features.

A breakdown of agentic AI patterns - from simple tool use to complex multi-agent systems. The progression from ReAct to more involved patterns like planning and multi-agent collaboration matches what we're seeing as we build agent tooling at Coursera.

How urban development is squeezing the tea farms around Uji, and what the growers are doing about it.