CourseLarge Language Models: The Complete Course · Module 3: Inference and Decoding · part 26 of 103
Part 26 · Module 3: Inference and Decoding

Incident: A sampling bug turns replies into word salad

7 min read·13 Oct 2026
Module 3 · Incident file · About 30 minutes

What happened

OPENAI STATUS PAGE, UTC 20 Feb 23:40 investigating 23:47 identified 21 Feb 00:59 monitoring 07:14 normal 22 Feb 01:02 write-up about 7.5 hours of unexpected responses
From the first status post (red) to 'operating normally' (green) took about seven and a half hours; the write-up (blue) came a day later.

At 23:40 UTC on 20 February 2024 OpenAI's status page posted: "We are investigating reports of unexpected responses from ChatGPT." Seven minutes later: "The issue has been identified and is being remediated now." It moved to monitoring at 00:59 on 21 February and was marked "ChatGPT is operating normally" at 07:14 UTC, about seven and a half hours after the first post (OpenAI status page, incident "Unexpected responses from ChatGPT").

WHAT ONE USER SAW (THE REGISTER, 21 FEB 2024) prompt: "What is a computer?" "It does this as the good work of a web of art for the country, a mouse of science..."
Every word is English and the reply means nothing (red): the failure users called word salad.

What users saw was reported by The Register on 21 February 2024 (Richard Speed). Asked "What is a computer?", ChatGPT gave one Reddit user a reply that began "It does this as the good work of a web of art for the country, a mouse of science, an easy draw of a sad few..." Others reported answers in a different language or pure gibberish; some called it "word salad". OpenAI's write-up gives no count of affected users or conversations, so this file does not guess one.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside