Bilingual

The gauge broke: devs felt 20% faster with AI, measured 19% slower
20 days ago
- A controlled trial found experienced developers felt 20% faster but were actually 19% slower when using AI tools in familiar codebases.
- AI speeds up typing, which isn't the bottleneck for experts, and adds overhead like prompting, waiting, and reviewing often subtly wrong outputs.
- Team-level data shows increased pull request volume and size but no net delivery gain, with more code churn and less refactoring, indicating verification has become the new bottleneck.
- Tool builders are shifting focus to agent-first IDEs, emphasizing verification over generation, aligning with the identified bottleneck in development workflows.
- The feeling of speed is misleading; successful teams must measure actual production outcomes and restaff review stages, rather than relying on perceived velocity.
Software Has Changed
8 days ago
- Software development has changed drastically and quickly, leaving experienced engineers shell-shocked.
- Development is now faster but not necessarily better, focusing more on quantity over quality.
- Flow states have been replaced by 'thinking...' states, and code is often discarded without thought due to cheap experiments.
- Good developers can now do more, using saved time to review, polish, refactor, and add extra features.
- Technical debt can be addressed before it reaches the main codebase, improving overall code quality.
- The field has changed overnight, requiring adaptation to build faster and think at a higher level to avoid tunnel vision.
- A key question arises: if the end result is solid, thoroughly reviewed, and trusted code, does the creation process truly matter?
The Human-in-the-Loop Is Tired
5 days ago
- LLMs make programming genuinely useful but also destabilizing, leading to burnout if ignored.
- Development with LLMs introduces 'fatigue of supervision': reviewing AI-generated code lacks the satisfaction of traditional programming and feels lonely.
- AI increases work intensity and parallel task initiation but doesn't improve thoughtful completion, creating a 'human reward function problem'.
- LLM-assisted programming shifts skills rather than eliminates them, emphasizing taste, nuance, and architectural expertise as key differentiators.
- New skills are emerging, like pre-mortem analysis and encoding expertise into LLM instructions, showing evolution rather than death of craft.
IDEs are done, and they are not coming back
7 hours ago
- The author created a list of 30 developer experience issues in JavaScript tooling upon joining VoidZero in January.
- A friend's question led the author to realize they no longer needed both Sublime Text and their IDE, motivating them to remove Sublime Text.
- The author stopped writing code by hand months later, making the previously identified paper cuts irrelevant and forgotten.
- The author has now deleted their todo list and plans to delete VS Code, declaring that IDEs are obsolete and will not return.

Hasty Briefsbeta

#developer experience

The gauge broke: devs felt 20% faster with AI, measured 19% slower

Software Has Changed

The Human-in-the-Loop Is Tired

IDEs are done, and they are not coming back