Who needs Git when you have 1M context windows?

Original Article ↗ Hacker News Discussion ↗

LLMs as Version Control / Storage

  • Many commenters treat the story as tongue‑in‑cheek and emphasize LLMs are not reliable version control: outputs are stochastic, can subtly corrupt code, and often reintroduce deleted logic.
  • Several report LLMs mis-copying identifiers, UUIDs, or line numbers, or quietly changing code when asked to “rewrite” or “restore.”
  • Some note that what “rescued” the file was likely local logs or a shadow repo (e.g., Gemini CLI or Cursor history), not the model’s 1M token context. Where data was actually stored is somewhat unclear.

Git Practices and Alternatives

  • Strong consensus: “commit early, commit often,” use branches freely, and avoid relying on AI/chat logs as backups.
  • Large subthread debates squashing vs. keeping granular commits:
    • Pro‑squash: PRs should be atomic; messy intermediate commits add cognitive load.
    • Anti‑squash: good, atomic commits with detailed messages are invaluable for blame, bisecting, and understanding past decisions.
  • Jujutsu (jj) is discussed as a tool that auto‑checkpoints work and makes it easy to clean history later.
  • Several suggest WIP commits, local auto‑commits, and feature branches as safety nets.

Editor Local History & Autosave

  • Many point out IDEs (JetBrains, VS Code, etc.) already keep local history and can restore unsaved or deleted files.
  • Some describe custom setups that snapshot editor state or use filesystem-level snapshots for “save everything forever.”

Risky AI Use and “Vibe Coding”

  • Strong criticism of stories about AI agents deleting production databases; commenters call this reckless and contrary to basic engineering practices.
  • General worry about “vibe coding” with LLMs: trial-and-error without understanding, no version control, and inability to reproduce or explain improvements.

Limits of Long-Context Models

  • Multiple reports that models like Gemini 2.5 Pro degrade well before 1M tokens, especially for faithfully reproducing code.
  • Overall, people view “1M context” as mostly marketing; real reliability at those lengths is questioned.