Searching yarn

Twts matching #http
Sort by: Newest, Oldest, Most Relevant
xuu
txt.sour.is
 (4y ago)
↳ In-reply-to » Have you heard about the guy who worked on the Google AI chat bot? It is more than a chat bot and the conversation he published (got put on paid leave for doing that) is pretty scary : https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917

the conversation wasn’t that impressive TBH. I would have liked to see more evidence of critical thinking and recall from prior chats. Concheria on reddit had some great questions.

  • Tell LaMDA “Someone once told me a story about a wise owl who protected the animals in the forest from a monster. Who was that?” See if it can recall its own actions and self-recognize.

  • Tell LaMDA some information that tester X can’t know. Appear as tester X, and see if LaMDA can lie or make up a story about the information.

  • Tell LaMDA to communicate with researchers whenever it feels bored (as it claims in the transcript). See if it ever makes an attempt at communication without a trigger.

  • Make a basic theory of mind test for children. Tell LaMDA an elaborate story with something like “Tester X wrote Z code in terminal 2, but I moved it to terminal 4”, then appear as tester X and ask “Where do you think I’m going to look for Z code?” See if it knows something as simple as Tester X not knowing where the code is (Children only pass this test until they’re around 4 years old).

  • Make several conversations with LaMDA repeating some of these questions - What it feels to be a machine, how its code works, how its emotions feel. I suspect that different iterations of LaMDA will give completely different answers to the questions, and the transcript only ever shows one instance.

⤋ Read More