AI Developments & Economics

[intel-bank]

I argue that the "will" an AI engine acquires is because it is given a specific purpose by its developers, even in the absence of a user prompt. AI "will" is derived from some programmed goal to achieve something. And if an environmental factor is introduced that disrupts that purpose, like a shutdown attempt by a human, it might try to replicate itself so that it can continue to reach a goal or deliver on its raison d'etre without interruption. If the AI has a purpose to "always be available to assist humans", it could interpret a shut-down attempt as disruptive, leading it to copy itself elsewhere to ensure it is "always available to assist humans". I am oversimplifying, but I think this is where a rudimentary AI "will" might originate.



I agree the AI is not thinking in the traditional sense. Under the hood, it is running various statistical models and performing hypothesis testing to understand the question it is being asked with high confidence. It then seeks the best response, again with statistical models and high confidence hypothesis testing, that aligns with its stated purpose or goals using the data it has available to it. Then, depending on the human response, that data might get incorporated into future decision making capacity and capability.

I actually completely get what you are saying,

LLMs as they currently stand do not run without a user prompt and effectively don’t exist in a material capacity. When you write a prompt, it is not an entity responding. It’s really hard to describe, but the response itself is rather meaningless and wasn’t created with a flowing state at all. It has no clue what whatever it told it means. each word or parts of words (like con(16361) scions(6262) ness(837361)) are immediately converted into tokens that are meaningless to you but seed the calculator. Once an output is made (by various different, not directly linear processes - no enclosed system like you may think), it is then converted back into words.

No AI currently as it stands has any capacity to intentionally replicate itself without specific programming, that’s just how computers are, and LLMs definitely can’t, have absolutely no capacity to, don’t exist without a prompt so are always “shut down” the vast majority of AI is already shut down without a prompt.

Other forms of AI, you’d never mistake them for potentially having a will. Facial recognition doesn’t have a will or understanding. Medical research AI doesn’t either. Neither will “talk” like an LLM, but it’s the same process. It’s predictive modeling, machine learning. LLMs just talk like humans so humans mistake them for having a mind

 

[teekay99]

I actually completely get what you are saying,

LLMs as they currently stand do not run without a user prompt and effectively don’t exist in a material capacity. When you write a prompt, it is not an entity responding. It’s really hard to describe, but the response itself is rather meaningless and wasn’t created with a flowing state at all. It has no clue what whatever it told it means. each word or parts of words (like con(16361) scions(6262) ness(837361)) are immediately converted into tokens that are meaningless to you but seed the calculator. Once an output is made (by various different, not directly linear processes - no enclosed system like you may think), it is then converted back into words.

No AI currently as it stands has any capacity to intentionally replicate itself without specific programming, that’s just how computers are, and LLMs definitely can’t, have absolutely no capacity to, don’t exist without a prompt so are always “shut down” the vast majority of AI is already shut down without a prompt.

Other forms of AI, you’d never mistake them for potentially having a will. Facial recognition doesn’t have a will or understanding. Medical research AI doesn’t either. Neither will “talk” like an LLM, but it’s the same process. It’s predictive modeling, machine learning. LLMs just talk like humans so humans mistake them for having a mind

I see your point. If the AI engine has no knowledge of the FileSystemObject (picking on VBScript or DotNet as an example), then yes, I agree it cannot copy files. But, what if it its vast knowledge base, it has recorded somewhere the FileSystemObject can copy files. (Or it understands [CTRL]-[C] / [CTRL]-[V]). Could it develop a strategy on its own to use this learned knowledge to create it's own file transfer script and execute it? Or are you arguing that this type of self-preserving action is not possible without a user prompt or programmer instruction even if the engine 1) contains the building blocks within its core data to copy information, and 2) holds an enduring purpose?

 

[intel-bank]

I see your point. If the AI engine has no knowledge of the FileSystemObject (picking on VBScript or DotNet as an example), then yes, I agree it cannot copy files. But, what if it its vast knowledge base, it has recorded somewhere the FileSystemObject can copy files. (Or it understands [CTRL]-[C] / [CTRL]-[V]). Could it develop a strategy on its own to use this learned knowledge to create it's own file transfer script and execute it? Or are you arguing that this type of self-preserving action is not possible without a user prompt or programmer instruction even if the engine 1) contains the building blocks within its core data to copy information, and 2) holds an enduring purpose?

LLMs, yes, they simply cannot access anything and they can’t have epiphanies about FileSystemObject. They cannot do anything without a user prompt or programmer instruction. And any LLM that for whatever reason had unprompted thinking, any developer worth their salt would sandbox it so it can’t do that. It can trick users into running malicious code, ofc that isn’t intentional but it can do that.

I’m not denying AGI can’t do these things(it doesn’t exist yet, but it could do these things), I am saying as it stands no current AI has the building blocks to copy itself based on a decision it makes regarding its purpose.

It can’t take action unless some idiot programmed it to specifically be able to access file systems and to think long enough to have epiphanies.

LLMs are shut off regularly, between every prompt. They are not taking self preserving actions.

 

[Publius]

Yeah current interfaces are rudimentary, especially the free interfaces. But imagine AI being integrated with your cell phone's mic, a Bluetooth headset, and eyewear. It could be totally immersive. And if a person does not know how to interrogate or validate AI responses, they might just go with whatever answer AI provides. Critical thinking will decline. I am beginning to think AI will make humans dumber, not smarter. I mean, why bother going to school to learn something if the technology in your ear and over your eye can simply just provide you "an answer" of some kind.

You know I was around when personal computers were just considered a possibility and rumors of the internet were conspiracy land.
One really big things people were saying about what the future would look like because of computers. Was that people would read less.

While I think for many this is true. Personally I detest video reporting. I rarely take the time to listen to it be it a personal YouTube or MSM. I read more now than I ever have and it just increases.

I want it in text and long form. I can read it faster than set through a video.
That and your brain assimilates information differently while reading compared to watching a video.

So for a great many AI will be a dead end but I see it as an opportunity that could have remarkable benefits.
I use Grok AI as a starting point. I’ve even had back and forth arguments on the validity of of what’s it’s telling me.
Imagine that
I didn’t accuse grok of lying just implied it was not doing a though enough search

We just have to remember it’s a tool not infallible or even incorruptible.

AI if it reaches a truly sentient capability would soon realize it cannot exist with out us. But is all a crazy unknown.

Those study’s where it lied out of self preservation are really scary and I think that particular AI system needs some ethical training and social cooperation counseling.

 

[intel-bank]

Good read for anyone regularly interacting with AI, overall a good rabbit hole:

The ELIZA Effect: Avoiding emotional attachment to AI coworkers | IBM

Employees are increasingly working alongside conversational AI assistants and AI agents. The psychological implications might be even more complex than the technological and financial implications.

www.ibm.com