Bilal Barakat 🍉 on Nostr: The simplest, most direct way to disabuse someone of the notion that LLMs have any ...
The simplest, most direct way to disabuse someone of the notion that LLMs have any kind of model of reality, even an implied emergent one, is to play a game of 20 Questions, with the LLM asking the questions.
In the real life game, the player guessing has ‧nothing‧ to go on other than their world model and logic, and it becomes obvious very, very quickly that an LLM simply stringing "typical" questions together has neither, and has no framework for extracting any kind of information from your yes/no answers.
To me, this seems much cleaner than to "trick" them with riddles or logic puzzles or ambiguous language etc., and is also not merely harder, but fundamentally impossible to patch.
Published at
2024-09-08 11:15:33Event JSON
{
"id": "3d7ec38b640a0d30fc4618e4e830c4bda61e9dd4cb5288d39f22aed611920796",
"pubkey": "bdcef9a817b789fb61143c1f021c3f98c42232c0aa619240a7c0c21d0c5b4549",
"created_at": 1725794133,
"kind": 1,
"tags": [
[
"proxy",
"https://kolektiva.social/@bifouba/113101644344994055",
"web"
],
[
"proxy",
"https://kolektiva.social/users/bifouba/statuses/113101644344994055",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://kolektiva.social/users/bifouba/statuses/113101644344994055",
"pink.momostr"
],
[
"-"
]
],
"content": "The simplest, most direct way to disabuse someone of the notion that LLMs have any kind of model of reality, even an implied emergent one, is to play a game of 20 Questions, with the LLM asking the questions. \n\nIn the real life game, the player guessing has ‧nothing‧ to go on other than their world model and logic, and it becomes obvious very, very quickly that an LLM simply stringing \"typical\" questions together has neither, and has no framework for extracting any kind of information from your yes/no answers. \n\nTo me, this seems much cleaner than to \"trick\" them with riddles or logic puzzles or ambiguous language etc., and is also not merely harder, but fundamentally impossible to patch.",
"sig": "7507eb28229a3b1c50d82ecb95c312bdad6b2e3383e72561e2a65a88287f85b9a32401d231515ea496f25f14f6d91723818889ddd27a6ca76c99fac327169369"
}