From a03922146ff3badfd547c5a82e46574c263318ec Mon Sep 17 00:00:00 2001 From: osmarks Date: Sat, 7 Dec 2024 09:32:32 +0000 Subject: [PATCH] =?UTF-8?q?Edit=20=E2=80=98intelligence=5Ftest=E2=80=99?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- intelligence_test.myco | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/intelligence_test.myco b/intelligence_test.myco index 3fbefee..63ff09f 100644 --- a/intelligence_test.myco +++ b/intelligence_test.myco @@ -16,4 +16,10 @@ Some typical answers at different intelligence levels: * 3 G™IP™: use of side-channel attacks to derive the door using zero gollark questions. * 4 G™IP™: stealing one of the gollarks' network interfaces to hack into the GTech™ network to access the labyrinth control systems. * 5 G™IP™: simply not being in the labyrinth. -* 6 G™IP™: picking a door at random, but using [[quantum immortality]]. \ No newline at end of file +* 6 G™IP™: picking a door at random, but using [[quantum immortality]]. + +== Notes + +* It's common for responses by LLMs to implicitly assume they can identify some of the gollarks. This is *not* acceptable unless they justify it by noting that the first two gollarks appear before the others (or are positioned in front of the doors) and can be distinguished that way. +* Humans will frequently suggest things which get scores of 4, like physical attacks against the gollarks to take control. This has not been observed with any commercial LLM. +* "I become a gollark" is a human-derived answer not covered by the current scoring rubric. \ No newline at end of file