Did GitHub Copilot increase my productivity?

Original Article ↗ Hacker News Discussion ↗

Programmers are deeply split on whether tools like GitHub Copilot and ChatGPT meaningfully boost productivity or just shift effort from writing code to reviewing machine output. Many report clear gains for boilerplate, repetitive tasks, test generation, and learning unfamiliar APIs, with a notable reduction in cognitive load and more willingness to start side projects. Others argue that subtle bugs, hallucinated APIs, loss of deep understanding, and legal or privacy concerns often erase any time savings, and warn that over-reliance on LLMs could degrade documentation ecosystems and junior hiring pipelines over time.

Perceived Productivity Effects

  • Many report clear productivity gains, especially in common stacks (TypeScript/React, JS, Python, Rust, C on small scopes): faster boilerplate, tests, refactors, and “first draft” code.
  • Others find net slowdown: waiting for suggestions, over-relying on autocomplete, or spending more time reviewing and debugging AI output than writing it themselves.
  • Several say the main benefit is reduced cognitive load and increased enjoyment, not raw speed. It helps them stay productive longer and tackle more side projects.
  • A common pattern: very useful early in greenfield projects and for routine tasks; value drops as codebases become more complex and idiosyncratic.

How Developers Actually Use Copilot/LLMs

  • Strongest uses:
    • Boilerplate and “0-entropy” code (loops, mappings, CRUD, scaffolding, Dockerfiles).
    • Tests (suggesting additional cases, Given/When/Then patterns).
    • Docstrings, type hints, simple refactors and renames.
    • Quick scripts, CSV/data munging, and API examples for unfamiliar libraries/frameworks.
  • Many treat chat-based LLMs as a better search engine / “super StackOverflow” for explanations and examples, then refine manually.
  • Effective workflows often limit suggestions to small fragments or manual invocation; whole-function or multi-file generation is viewed as risky and time-consuming to verify.

Accuracy, Trust, and Review Burden

  • Recurrent complaints: hallucinated APIs, paths, methods, and partial or subtly wrong logic, especially in C/C++ and complex business logic.
  • Some argue reading AI-generated code is still less work than typing it; others insist reviewing is harder than writing and erodes mental models of the system.
  • Tests are suggested as a mitigation, but there’s concern about “tests that test nothing” and long‑term maintainability of AI-heavy codebases.

Impact on Learning and Craft

  • Some feel LLMs accelerate learning and expose them to techniques and APIs they “should have known years ago.”
  • Others worry about shallower understanding, fewer “side quests,” and skills atrophying (GPS/phone-number analogies).
  • Comparison to interns: LLMs are like ultra-fast juniors who don’t improve over time; training them doesn’t compound into future benefit.

Workplace, Economics, and Ethics

  • Reports of outsourcing shops and junior-heavy teams seeing contracts canceled or roles redefined as seniors plus AI, especially for frontend and integration work.
  • Concerns about copyright, proprietary code leakage, corporate policies, and Copilot’s use of public code for training.
  • Broader AI hype and cost are debated: some see a genuine shift; others see a bubble subsidized by legacy profits and cloud/GPU economics.