Actually human children acquire knowledge faster than AI because they learn context and understanding. Comprehensive. Which AI isn’t capable of doing.
And even compared to llms. A toddler child can learn verbal language faster.
They hear a word being used one time and can apply that word appropriately a short time later in a new situation. Because they learn context and understanding of the word properties.
Llms need literally millions of stolen literature to form sentences. And it’s just copy paste based on probabilities.
That’s not how humans work. We need way less exposure to have deep learning and comprehension. Sometimes only requiring a single experience of a word or response to know how to apply it to new scenarios.
No large training sets? The baby is updating their training data set every second with new sounds, images and other sensory inputs (smell, touch). Current data centers would be overwhelmed with the amount of data that one child processes every day.
No AI has ever attempted to eat its own foot, or to even have a foot. I’m not sure if that’s a sign of intelligence or not.
Difference is that a child/baby can recognize an item simply from ONE “video” (vision) of it. It can also recognize drawn versions and similar versions of the same item. For instance an elephant can be viewed once and for ever be recognized in multiple forms
Also the use of tools. Once a human that’s never seen a glass or a mug before, child or not, sees someone drinking out of a glass, it knows EXACTLY how glasses work and how to use them. It’s not learning through repetition the way an LLM does, it’s learning through understanding.
Yes, but only after learning about types of objects and types of visualisations on thousands of examples. For example they see people (parents, siblings and others), they see drawings of people in books, they see people as toy figurines, that way they learn concepts like drawing and sculpting, they can recognise that concept and copy it on different ideas (for example they can imagine how crocodile looks like even though they only see it in a book). That actually is similar to AI image generation, where you can add images of a person (only photos) and examples of an art style. And AI will be able to generate images of that person using that art style (with various success).
Humans can turn information into knowledge. AI can only extract predigested knowledge. That’s why they need such large data sets.
This is also why AGI is still a pipe dream. Most of the subsections already exist, but without a knowledge engine at its core, it’s like a person with a massive concussion. It sounds reasonable to simple conversation, but goes off on weird tangents.
I always find it fascinating to watch children learn and grow. The mistakes they make are often completely logical, they just have holes in their knowledge, and lack critical information. It leads to wonderful conclusions, which are also completely wrong.
Oh, so that’s why the training part is so ridiculously inefficient. No amount of data is enough, and that sounds like a massive design flaw to me.
It is. Unfortunately, no-one has come up with a viable self teaching system. They are trying to brute force it, but have run into a wall. Without a knowledge engine, you can’t close the loop. Without closing the loop, self reflection is impossible. Without self reflection, hallucination filtering is almost impossible. Such a system can’t understand the difference between a discussion down the pub, and a lawyer’s legal brief. It can see the difference, and mimic it’s structure, but not understand the why, and account for it.
Who is upvoting this crap??
Genuine question, why is this crap?
Hey chat, the world is going to shit and I’m going to make a new person to experience it. Go me right?
It’s just going to be so interesting to see how the person I’m bringing into this shithole copes with all this shit.
There really is so much shit here. I hope my child ignores all the shit. Yes, my child will be special and ignore all this terrible shit everywhere.
So chat, what do we think? I’m going to be such a good parent, right?
I am not asking for praise. What made you believe I did?
When did I insinuate “being a good parent”?
You might disagree with my choice with being a parent, but that is a separate matter. It is still interesting to see a human develop and experience the world, and how humans differ from machines
Eh, depends on the baby. For all we know it could grow up to be a flat-earther or something.
AI models don’t really “know” anything
My baby learn without large training sets.
What does this mean?
A baby can see an elephant and without seeing thousands/millions of them, learn to recognize it every time. Also drawn, different types etc.
Ahh diggit!
Unless you home school then how will you know they won’t be taught with AI?
She can fold clothes, which AI can’t do.
Baby can be flexible, adaptable, feeling, loving, hoping, dreaming. Enjoy her ComradePenguin, she is a treasure. “AI” may be many things one day. Please don’t assume it can’t do something (intellectual). One of the big negatives is that a human intelligence runs on a few hundred watts, whole body. An “AI” uses probably several orders of magnitude more, just for the “I” part which is not good for the humans (pollution, etc). If only they could try to improve “AI” efficiency instead of brute-forcing. USA logic though. Brute force.
Judging by its parents grasp of the English language I don’t think it has much of a fighting chance
Judging by its parents’ grasp of the English language, I don’t think it has much of a fighting chance.
Having babies now is moronic. They’re doomed.
You must be a real hit at social gatherings.
But why can’t you make an AI that constantly adjusts the weights as new data becomes available? You wouldn’t have to dump 20 MW of power on it during the training period every year. Instead, just keep running it at a lower power and have it generate and learn simultaneously.
Adjusting the weights is how training works.
If you are thinking about adjusting the number of weights, that’s also a technique. The issue is that (afaik) it can only be trained through genetic algorithms, which take longer than backpropagation.
(This information might be incomplete/incorrect.)