I’m not sure why but it seems like very few of the ChatGPT threads I’ve read across on Twitter/HN showing “very accurate and intelligent” results never experiment with ablation which is very important in prompt engineering and when evaluating DL models.
I’m very excited about transformer based LLMs and use them extensively in my research and I think their generated content is good, but no where near being perfect or an imminent threat to society as people are worried about.
There are several fundamental limitations in their training that prevent this from happening and can’t be solved simply with “more training” as they still lack knowledge.
Recent work by researchers such as Dr. Percy Liang and Michiro Yasanuga (i.e. DRAGON, LinkBERT) has begun tackling this problem with KG injection (rather than supervised learning in ChatGPT) but we’re still a ways away.
https://openai.com/blog/chatgpt/ https://arxiv.org/abs/2210.09338