Re:CriticGPT: Finding GPT-4's mistakes with GPT-4
by sva_ on Thursday June 27, 2024 @09:08PM
(#40816863)
Well the definition of deception, according to Google and how I understand it, is:
> deliberately cause (someone) to believe something that is not true, especially for personal gain.
Emphasis on the personal gain part. It seems like you have a different definition.
There's no point in arguing about definitions, but I'm a big believer in that if you can identify a difference in the definitions people use early into a conversation, you can settle the argument at that.
Re:CriticGPT: Finding GPT-4's mistakes with GPT-4
by advael on Thursday June 27, 2024 @10:52PM
(#40817286)
I both agree that it's pointless to argue about definitions and think you've presented a definition that fails to capture a lot of common usage of the word. I don't think it matters what the dictionary says when we are talking about how a word is used. Like we use "deceptive" to describe inanimate objects pretty frequently. I responded to someone who thought describing the outputs of a machine learning model as deceiving people implied it had agency, which is nonsense
2. While I agree that it's a stretch to call ChatGPT agentic, it's nonetheless "motivated" in the sense that it's learned based on an objective function, which we can model as a causal factor behind its behavior, which might improve our understanding of that behavior. I think it's relatively intuitive and not deeply incorrect to say that that a learned objective of generating plausible prose can be a causal factor which has led to a tendency to generate prose which often deceives people, and I see little value in getting nitpicky about agentic assumptions in colloquial language when a vast swath of the lexicon and grammar of human languages writ large does so essentially by default. "The rain got me wet!" doesn't assume that the rain has agency