People who think AI is "like a child growing up", what was Eliza?

[PinealisGlandia]

Please, Log in or Register to view URLs content!

I've heard/read a lot of people compare LLMs to human children, justifying their errors as "like a child learning how to behave". I'd like to hear from these people. What is special about LLMs that make them "childlike consciousnesses" compared to the generations of AI that came before them, which were all recognized as code emulating human behavior. I admit, it's been a bit of a peeve of mine the last few years. For a solid 70 years, every advancement in AI was met with "it's not true AI", until LLMs, and now suddenly it counts as AI? I don't get it.

In all honesty, my take is if you see an LLM as a child, it's because you don't understand the tech very well. But that's why I started this thread, I'm trying to cultivate empathy by understanding someone else's point of view. So please, if you hold this point of view, add your explanation of how an LLM works to your post, that I might change my assumptions. Thank you for your time.

 

[Robert Ramsay]

I do not hold this point of view and the reason is simple:

"I heard some professor put googly eyes on a pencil and waved it at his class saying "HI! I'm Tim the pencil! I love helping children with their homework but my favorite is drawing pictures!" Then, without warning, he snapped the pencil in half. When half his college students gasped, he said "THAT'S where all this AI hype comes from. We're not good at programming consciousness. But we're GREAT at imagining non-concious things are people."

So it's actually our empathy that's allowing AI to fool us.

 

[Morell]

Thanks for that dose of cold observation!

I definitely agree. In fact it's common for us humans to personalize objects. Probably every ship on sea has a name and pronounces she. A lot of locomotives have names too. Some cars even. Naming blades is ancient too. And we occultists even spiritualize objects, connect them with spirits with names, etc. Religious statues we in fact create to be bodies through which we connect with forces, sometimes conscious.

I don't think it is exactly empathy. Empathy is about understanding. This is sort of human-natural familiarizing.

 

[Robert Ramsay]

Thanks for that dose of cold observation!

I definitely agree. In fact it's common for us humans to personalize objects. Probably every ship on sea has a name and pronounces she. A lot of locomotives have names too. Some cars even. Naming blades is ancient too. And we occultists even spiritualize objects, connect them with spirits with names, etc. Religious statues we in fact create to be bodies through which we connect with forces, sometimes conscious.

I don't think it is exactly empathy. Empathy is about understanding. This is sort of human-natural familiarizing.

Yes, technically it's anthropomorphising, but empathy is about the ability to put yourself in another's shoes, and it happens even when the 'other' doesn't have any shoes

 

[Nagaram]

The problem with viewing current gen LLMs like this is that they don't work like this at all.

It's not a learning system. It's a tool that has to be trained on correct answers and it will probably give correct answers. It's quite interesting from a Computer Science perspective, but it's not going to be any more revolutionary than autocorrect. Unless they start giving it the ability to shoot missiles or something. That would be crazy! ...anyways

If this worked like Chappy at all then I think we'd have a more interesting conversation on the nature of consciousness than we currently do.

 

[Swampdweller900]

LLMs get called "AI" because it's shorthand that sells. It's not true AGI, and even the term AGI moves the goalposts on what sounds like the less fun and sexy Machine Learning.

Children learn and develop in iterative steps, which includes (should, anyway) introspection. LLMs get new model upgrades which are an external process imposed upon them. The Andon Labs experiment where they let the Big 4 models run a radio station shows that changes in models drastically and quickly change the personality, vocabulary, and logic.

I've trained customized LLMs as part of some freelancing work. You have to build up examples to build up patterns to train the models. No different than training a dog or a rat to run a maze.

 

[KjEno186]

A child's first experience with apples may happen when its mother feeds it some applesauce from a jar. The sweet taste and smooth texture make an impression upon the child even before it knows how to speak. After the child has enough teeth, its parent may give it slices of apples and whole apples. The feeling of the fruit in its hands, the smell of the fruit in its nose, the textures of the fruit on its lips and tongue, those bits of apple skin that get caught in the teeth, and of course the taste of the apple, all inform the child of the physical reality of the fruit we call an apple. The child could eventually decide to become a connoisseur of all things related to apples, like ciders and pastries.

An LLM can be prompted to provide a wealth of information on all aspects of apples, but it will never have seen, felt, or tasted one. I've heard of current tech leaders who claim they want to "upload" their consciousness to an AI. Even if they were somehow agentic, I think the AGI minds of said tech leaders could be tortured to madness (if AGI could "go mad") by knowing that they'd never taste another apple again.