[ExI] LLM's cannot be concious

[Stuart LaForge]

Hi, Giovanni. Thanks for noticing. :) I tried to make the problem easy  
to solve, but hard to understand, because it was a test for  
understanding. I don't know what proportion of humans could solve this  
problem but I would wager that most could provided that they took the  
time read it all the way through.

Stuart LaForge

Quoting Giovanni Santostasi <gsantostasi at gmail.com>:

> Your test was pure genius.
> I wonder how many humans can actually solve this simultaneous simple and
> confusing problem (at least for some given the chain of actions involved).
> Very well done. Yes, understanding of several concepts is necessary to
> solve this problem.
> Giovanni
>
> On Sat, Mar 18, 2023 at 5:22 PM Stuart LaForge via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>>
>> 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
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> _______________________________________________
>> extropy-chat mailing list
>> extropy-chat at lists.extropy.org
>> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>>