Hacker News Reader

It's amazing to me that people actually thought GPT2 produced "human-passing" text, while I'm still tripping over obvious LLMisms in the output of recent models on a daily basis.

(It's also amazing to me that it took mere minutes for this observation, deep in a sub-thread, to get downvoted without any reply, with no obvious reason for it.)

View on HN

People may perceive you to be cheaply mischaracterizing the argument.

Nobody believed or suggested that GPT2 could do longform or produce novel text that stood up against careful scrutiny as insightful or well informed. But because the capabilities were novel, people would have no strong alternative than to believe some person wrote it.

You current tripping over LLMisms is irrelevant. You have years of antibodies, both personal and herd-immunity (eg, the many, many articles and comments that describe LLMisms).

LLM-isms are much less prevalent in base models, which is what GPT-2 was. It had significant problems with maintaining coherence, but GPT-2 generated text did not have the obvious tells of today's LLMs.

Reply on news.ycombinator.com