- Using AI like everyone else can make our work look more similar, even when the individual results improve. One experiment found that AI helped people write better-rated stories, but those stories became more alike; another found that developers learning a new library with AI scored 50% on an immediate unaided quiz, compared with 67% without it. Getting something working and learning how it works can come apart. For skills we want to develop we should attempt the problem ourselves first, asking AI for hints and questions, and checking whether we can explain the result without its help x.com/MilesCranmer/status/2111105654211608847 github.com/brandonlovejoy/cranmer-ai-brain-rot-sources
- Getting better starts with knowing one small area deeply, then expanding from there. Becoming the person who understands a particular system gives you opportunities to handle incidents, make decisions, and take ownership before you know how everything else works. AI makes it easier to ship without developing that understanding, so it can be worth slowing down early to learn how the system works, why it was built that way, and where it breaks vikramsingh.me/increase-your-slope
- Writing is thinking, and choosing the words is part of figuring out what we mean. Letting an agent do all of it can mean giving up that practice, along with the particular voice and imperfections that make something ours. Being able to express an idea clearly also helps us direct agents: knowing the vocabulary of a discipline gives us more precise ways to ask for what we want x.com/poteto/status/2109885001584292091
- Code review is also how more than one person knows what changed and why. As agents produce more code, repeated feedback should become lint rules and build failures, with human attention going to consequential changes, especially anything that involves money or is difficult to roll back. Agree on what done means before review starts, and make sure someone can explain the change when it breaks at 2am. x.com/AjeyGore/status/2109896105093922984
- With agents, much of the important thinking happens in the session: what to try, which approach to reject, and what to leave out. A reviewer looking only at the final diff misses those decisions, and both people lose some of the learning that used to happen through reviewing each other’s code. Planning together, sharing relevant parts of sessions, and explaining deviations from the plan can make that thinking visible again without asking reviewers to read hours of conversation just-be.dev/blog/code-review-sucks-now
- Even work we enjoy for its own sake depends on other people. A shared body of work gives us something to build on, while someone reading, using, or responding to what we make helps us keep going when motivation fades. If AI replaces those exchanges, we may lose part of what makes us want to learn and contribute in the first place. just-be.dev/blog/code-review-sucks-now
What I’ve been reading (W44, 2026)
Learning with AI, making engineering decisions visible and why we need other people