Oddbean new post about | logout
 I have a similar setup but for news articles. The summaries with hallucinations were less than ideal (semi-controllable with temperature settings). Another factor was I found that people interested in a topic wanted to see all the content, summaries weren’t enough. However, one feature that popped out was doing Named Entity Recognition (NER) on the content _was_ useful because this allowed articles to be related to the others in the corpus. Keep up the interest work!