Conversation

Notices

1. Embed this notice
   Marcus Hutchins :verified: (malwaretech@infosec.exchange)'s status on Monday, 20-Mar-2023 22:08:26 JST Marcus Hutchins :verified:

   Apparently I accidentally reinvented the Chinese Room thought experiment whilst trying to explain to someone why ChatGPT doesn't actually "understand" language.

   My scenario was: say you downloaded every question from the Chinese version of Quora. For each thread, you create a folder. The folder's label is the thread title (the question). Each page of the folder is a user's reply to the thread (the answer). You ask native Chinese speakers to write letters to you containing their questions. You respond to their letters by comparing the symbols in the letter to the folder labels, then send back a random page from the corresponding folder. The Chinese speakers are convinced you are fluent in Chinese and all-knowing. Are you?

   • Embed this notice
     Marcus Hutchins :verified: (malwaretech@infosec.exchange)'s status on Monday, 20-Mar-2023 22:08:25 JST Marcus Hutchins :verified:

     in reply to

     Born too late to be a philosopher, too early to experience AGI, but just in time to shitpost on social media.

     clacke repeated this.
   • Embed this notice
     📄 Mehdi.doc (mehdi_benadel@mastodon.balamb.fr)'s status on Monday, 20-Mar-2023 22:08:41 JST 📄 Mehdi.doc

     in reply to

     @malwaretech Aren't philosopher's just ancient shitposters?

   • Embed this notice
     Marcus Hutchins :verified: (malwaretech@infosec.exchange)'s status on Monday, 20-Mar-2023 22:08:43 JST Marcus Hutchins :verified:

     in reply to

     • Giovanni Campagna

     @gcampax Evidence where?

   • Embed this notice
     Giovanni Campagna (gcampax@mastodon.social)'s status on Monday, 20-Mar-2023 22:08:44 JST Giovanni Campagna

     in reply to

     @malwaretech yeah careful with the analogies, there's evidence that LLMs do not just retrieve from memory. At least, there is evidence of reasoning and compositionality. Besides, you could implement literally the retrieval model you describe and QA performance would be far worse than modern LLMs.

   • Embed this notice
     Liminimal (liminimal@aus.social)'s status on Monday, 20-Mar-2023 22:18:00 JST Liminimal

     in reply to

     @malwaretech That always seemed trivially easy to refute by referring to earlier parts of the conversation in a subsequent statement. Without understanding there would be no way to respond correctly. A conversation is not a solvable problem like a game of chess