Thinking About Thinking with LLMs (davi.sh)

The author expanded a newsletter note into a full blog post in response to the Recurse Center’s “Developing our position on AI,” praising it as a model of civil, nuanced public discussion about AI. The piece recommends a simple process for contentious topics: listen broadly, acknowledge disagreements, and highlight nuance to find common ground. A clear takeaway is caution about relying on LLMs for learning—most people agree they’re powerful helpers but can encourage superficial understanding if used uncritically. Technically, the post frames LLMs as another “step-function” abstraction—like search, Stack Overflow, garbage collectors, or high-level languages—that makes many tasks easier while shifting what expertise looks like. Tools such as Claude Code and Cursor push programming toward natural language, democratizing access and enabling faster iteration, but they don’t eliminate the value of deep mental models. The implication for ML practitioners and educators is twofold: embrace LLMs to broaden who can program and prototype, but teach and cultivate deeper conceptual understanding to avoid brittle solutions and misuse. The balanced stance encourages both practical adoption and disciplined pedagogy rather than alarmism or uncritical hype.