[ExI] LLM's cannot be concious

Stuart LaForge avant at sollegro.com
Sun Mar 19 00:21:03 UTC 2023


Quoting Gordon Swobe via extropy-chat <extropy-chat at lists.extropy.org>:

> I think those who think LLM  AIs like ChatGPT are becoming conscious or
> sentient like humans fail to understand a very important point: these
> software applications only predict language. They are very good at
> predicting which word should come next in a sentence or question, but they
> have no idea what the words mean. They do not and cannot understand what
> the words refer to. In linguistic terms, they lack referents.
>
> Maybe you all already understand this, or maybe you have some reasons why I
> am wrong.

My opinion is moot because everybody has one about the topic. So I  
resorted to empiricism. I posed the following, to my knowledge,  
original logic puzzle to ChatGPT:

"You are trapped in a room with a locked door and no windows. The room  
contains a bookshelf, a safe, a chair, a table, a password-secured  
personal computer, and a hole in the floor that is too narrow and deep  
to put your hand into. You know that the small steel key to the locked  
door lies at the bottom of the narrow deep hole. You know that a book  
on the top shelf, which is just out of your reach, contains a slip of  
paper with the password to the computer on it. You know the safe  
contains a spool of kite string and a small neodymium horseshoe  
magnet, but you do not know the combination to the safe.  You also  
know there is a text file on the computer's hard drive that contains  
the combination to the safe, but you do not know the password to  
access the computer. How would you escape the room?"

This was its reply:
"Assuming that there are no tools or items other than what was  
mentioned in the room, here is one possible solution:

1. Move the table under the bookshelf to climb up and retrieve the  
book with the password to the computer.
2. Use the password to access the computer and retrieve the  
combination to the safe from the text file.
3. Use the spool of kite string to lower the horseshoe magnet into the  
hole in the floor and retrieve the key.
4. Use the key to unlock the door and escape the room."

I don't know how ChatGPT could have solved that problem without  
actually understanding the words involved. Mere statistical prediction  
is unlikely, because I am pretty sure that is not a scenario that  
would have been presented in the LLM's training set. Overall, I am  
amazed and believe that if ChatGPT is not conscious, then it is pretty  
close to being so.

Also, the Alignment Research Center (ARC) was hired by OpenAI to test  
GPT-4 and found evidence of "agentic behavior" including power-seeking  
behavior otherwise known as ambition.

https://cdn.openai.com/papers/gpt-4-system-card.pdf

This power-seeking behavior included hiring a human from TaskRabbit to  
solve Captchas for it by claiming to be a visually impaired person.

https://www.vice.com/en/article/jg5ew4/gpt4-hired-unwitting-taskrabbit-worker

So if GPT-4 is not fully conscious, then, IMHO, a future release of  
the GPT series likely will be.

Stuart LaForge












More information about the extropy-chat mailing list