Hacker News Reader

I noticed that when GPT-3 has trouble continuing a sentence or create a difficult argument it relies on filler words to stretch out a sentence until it has no choice but to write something meaningful.

View on HN

To be fair we have all been there.

LLMs are great and have many use cases but ultimately they’re just very talented bullshit machines.

This is due to the way they were trained: predict the most likely next word. In a difficult context the most likely next word is a filler word.

Reply on news.ycombinator.com