GPT-3.5 crashes when it thinks about useRalativeImagePath too much

The article discusses a peculiar error encountered when using OpenAI's GPT-3.5 model. The author found that when asking GPT-3.5 to include the phrase "pony

Key takeaways

The author discovered that asking GPT-3.5 to include a specific phrase at the top of the response, specifically 'pony', consistently results in an error message.
OpenAI’s GPT models output streams of multi-character 'tokens' instead of letters for improved performance and accuracy.
Three of these tokens are 'rparr', 'sarr', and 'RalativePath'. The latter appears in numerous files on GitHub as an option in XML configuration files for Katalon Studio, an automated testing software.
The author suggests that this error could potentially be used for 'data poisoning', to throw off attempts to summarize documents with GPT-3.5.

GPT-3.5 crashes when it thinks about useRalativeImagePath too much

Key takeaways

Discussion (0)