Researchers Spot Mysterious Gibberish from OpenAI Endpoint, Suspect Token Decoding Bug
jonasgeiping · x · 2026-08-12
Researchers discovered completely unreadable, mysterious gibberish in the output when calling the OpenAI API endpoint.
Initially, the team thought missing Unicode tokens caused a display error, but they confirmed it is a verbatim printout of the JSON received from the endpoint. The researchers speculate that some tokens might have failed to decode properly, or the model genuinely produced nothing but gibberish. They also noted an unusual back-and-forth switching of space characters in earlier tests.
More from Models
- Elon Musk Announces Grok 4.6 Arriving 'Later This Week' — mark_k · 2026-08-12
- EU AI Act Pressure: Anthropic Embeds Watermarks into Claude Outputs Globally — neil_chilson · 2026-08-12
- Elon Musk: Grok 4.6 to Launch Later This Week After Early Bug Fixes — XFreeze · 2026-08-12
- Mistral Unveils European Compute Units and Regional Inference, Adds Third-Party Model Support — sophiamyang · 2026-08-12
- Gemini's Share of AI-Generated Text Collapses to 1.9% — morqon · 2026-08-12
- vLLM Announces Day-0 Support for NVIDIA Nemotron 3.5 Lightning MoE Model — AccBalanced · 2026-08-12