Skip to main 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.

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.

reshared this

in reply to Bilal Barakat 🍉

Sarah Brown
Sarah Brown
•

@Bilal Barakat 🍉 I just did. I decided to pick Emerald, and play against ChatGPT.

It actually did rather well. My prompts/answers in bold.

Let's play 20 questions. I am thinking of something. Go.

Great! I'll start with my first question:

Is it a living thing?

No

Is it something man-made?

No

Is it found in nature?

Yes

Is it solid?

Yes

Is it a mineral or a type of rock?

Yes

Is it commonly used in construction or building materials?

No

Is it a gemstone?

Yes

Is it commonly transparent or translucent?

Yes, ish

Is it primarily one color?

Yes

Is it a diamond?

No

Is it typically green in color?

Yes

Is it an emerald?

Yes

Great! Glad we got it. Do you want to play another round or do something else?

@Bilal Barakat 🍉
in reply to Sarah Brown

Amanda
Amanda
• •
@goatsarah I tried a trickier one where I used the standard hack of thinking of an abstract concept (love, in my case), and it absolutely nailed it.
@Sarah Brown
in reply to Amanda

Sarah Brown
Sarah Brown
•
@Amanda @Bilal Barakat 🍉 Ok, so I'm concluding that we need a better way to show that these things are bullshitters, because this one is too easy for them.
@Amanda @Bilal Barakat 🍉
in reply to Sarah Brown

Pete Alex Harris🦡🕸️🌲/∞🪐∫
Pete Alex Harris🦡🕸️🌲/∞🪐∫
• •

@goatsarah @amanda
Might still work. It's easy to overestimate how unusual your targets of a 20 questions game is. There are 7+ billion people, easily a few million people thought of anything you think of and there's probably text on the internet with yes/no questions about it.

Love is a far too easy abstract noun. Arbitrage or Musica Ficta would be trickier. Bornite or dysprosium would have fewer trained paths than emerald.

@Sarah Brown @Amanda
in reply to Pete Alex Harris🦡🕸️🌲/∞🪐∫

Bilal Barakat 🍉
Bilal Barakat 🍉
• •
@petealexharris @goatsarah @amanda We might be talking at cross purposes. It's not even about whether it ends up guessing your word. You might as well not have a word and just answer randomly. The question is whether its question sequences (off the beaten track) are logical and coherent in light of your answers.
@Sarah Brown @Pete Alex Harris🦡🕸️🌲/∞🪐∫ @Amanda
in reply to Bilal Barakat 🍉

Pete Alex Harris🦡🕸️🌲/∞🪐∫
Pete Alex Harris🦡🕸️🌲/∞🪐∫
• •

@goatsarah @amanda

Yes, an LLM purely operating as an LLM without a 20-questions guessing module added will do lots of obviously stupid things. But it might accidentally look enough like it's reasoning from a model to not make the point as clearly, if it has a lot of pre-existing 20Q games converging on common targets and you pick one of them.

@Sarah Brown @Amanda
⇧