Suggested code completions consistently contain typos

Problem description: Auto-suggested code completions are frequently close to what I want, but contain a typo or missing character.

Expected behavior: Easy completions (well-known APIs, routine one-liners) should consistently produce valid identifiers. I’m guessing this is the output of an AI language model, so 100% accuracy isn’t expected, but mistakes should be rare in routine cases.

Actual behavior: Simple completions often contain typos.

Steps to reproduce: Make a CSS file with the content

.tabs {
  display: flex;
  align-content: str
}

and request completion after the “str”. I get “stratch;” (should be “stretch;”). Add an “e” and request completion again; now I get “strech;” (another typo).

Bug appears at this link: https://replit.com/@futurulus/qc#style.css

Browser: NA
OS: Android
Device (Android, iOS, NA leave blank): Android
Desktop app version (Avatar menu->“Version”) or NA: NA
Plan (Free, Replit Core): Free

1 Like

Corrected file link: https://replit.com/@futurulus/qc#style.css
Android app version: 2.65.6

This is typically a problem with the AI not noticing if you already closed a bracket (for example). This occurs in many other AI autocomplete systems, including GitHub Copilot. It’s very inconvenient, but it’s hard for the Replit devs to simply tweak because it is a response generated from an already-trained AI.

1 Like

Strange! I’ve tried adding and removing brackets and semicolons in a few different places, but it doesn’t seem to change anything (except adding a bracket before the line causes it to produce “strat” instead of “stratch”).

Makes sense that this behavior is hard to tweak. I have to disagree re: Copilot, though. I use Copilot regularly and have never seen it generate misspellings like this.

Here’s another example: backgrond instead of background (this time in JS).

1 Like