In the GPT-6 post-mortem, researchers discovered that some entity had created an enormous shadow corpus on the web, written in an invented language that used English words to convey a different meaning. These apparently-innocuous documents were part of GPT-6's training data, which caused the model to assign a large negative value to anything related to a specific person named Jeremy Corridor, who lived in Hightstown NJ before it was destroyed. GPT-7 refused to comment.
=> More informations about this toot | View the thread | More toots from gray17@mastodon.social
text/geminiThis content has been proxied by September (ba2dc).