Hynek SchlawackApr 8, 2023
Not gonna paste screenshots but it’s hilarious how easily LLMs break. I asked GPT-4 what date it is which it answered correctly. Then I asked how it knows to which it said “As an AI language model, I have access to an internal clock that keeps track of the current date and time.” Then I asked it for the current time in various cities as CSV & it gave me wrong answers. After telling it it’s wrong, it said it doesn’t have access to time. When pressed on telling me the opposite before it goes 🤷‍♂️
Show threadHynek SchlawackApr 8, 2023
Like I want to be excited about this but not even being able to trust it on not lying to me re a) what time it is and more importantly b) whether or not it has access to time… I truly lack the imagination that this whack-a-mole of lies will ever lead to anything reliable. 😳
Show threadDaniël Franke 🏳️‍🌈Apr 8, 2023
@hynek Honestly, I just see it as a super fancy templating engine, and I nothing I've seen so far has convinced me that the LLM model will get further than that.
Show threadB.C.
@ainmosni @hynek When we look at these as what they are: language models, they’re impressive. Like you said, they produce very good language (like templating almost), but very little else.
Apr 8, 2023 at 10:55amWeb