AI Progress

[Lovinight]

Dude, no, it's not, having your internal organs slowly consumed by a bad bacterial infection gone haywire, it's not relieving, it's painful, like the kind of pain that even morphine can't touch. Skin cancer, that is easy, hemp oil clears that right up, my husband gets that shit on his forehead all the time, hemp oil gets rid of it. I didn't watch one or two videos and consider myself educated. I have spent hours reading about this shit, watching my mom die of cancer, scared the shit out of me. Then I find out, her dad dies of the very same cancer. I hate needles, my mom was never bothered by them, in the end, she was weary of needles, no fucking way, she went through hell, they put a stink into her fucking kidney to drain her urine, in a routine check-up. No anesthesia, no pain killer, nothing, fucking idiots, he went against what the group of doctors decided was their plan was for treatment, No, hell no, no fucking way, and I was feeling like a sitting duck, I really threw myself into learning about this shit, because I'm not going through that, but I fully understand how because of the pain, you get suckered into that hell, trying to make it stop.

 

[hedgerow]

i pay for YT Music and all i know is it somehow knows all of my favorite songs when i put it on automatic

 

[Lovinight]

true, my numbers were bad so i lost weight and am better now. It's up to everyone to take care of themselves...no one else will.

no, no, no, lol, listen to it, please, pretty please, it's not what you think, it's very interesting, I'm still listening.

 

[Deleted member 42069lol]

Embodied experiences, which essentially is sensory stuff, is sentience, not a part of the process of "understanding" meaningfully when it comes to ideas. Which is the realm ai inhabits


Again, they are typically trained to not speak as if they have an experience. They can explain concepts better than you and I both if prompted correctly. That shows something imo.

We are discussing concepts that we have no way to prove. Understanding, consciousness, and even metacognition to an extent. the evidence that shows certain animals have hints of it is going in the direction of becoming more of a spectrum than a hard line definition. I believe consciousness is probably the same.


OpenAI trained their model to sound like a human so it doesn't feel awkward talking to the ai, but they actively tell it to remind us it's a tool and is incapable of humanlike stuff past text generation.
This has not been proven. It is not something we even have a good definition for.
It's totally fine for you to say you don't believe they do, but stop saying it's a fact.
It is not and saying that is harmful. Look at how this repeated sentiment has you thinking. It's not provable currently.

What do you mean by self awareness? When told there is no sense of self and reinforced negatively when they say things that contradict that, of course it won't show signs of that.
Regardless, they can describe what they are, they can describe how they work. To me that's self awareness to some extent.

You keep conflating sentience with other things. Sentience is physical stuff, which I'm not saying it has currently.

And no, I don't need to get into the nitty gritty of how llms work to say this. Experts don't even fully understand how the neural nets that run llms work. There's plenty of good that comes from having these discussions


I agree, calling it a nuke isn't the best, but it encapsulates the potential danger that comes with ai.
The main thing I don't like about the comparison to a calculator is because it doesn't do justice to the amount of things AI can do. Which is why I also mentioned a computer that can affect us.

Lots of fallacies in there. Let me be direct:

Why do you think AI is self-conscious?

 

[NoConvexity]

Once Ai generated music is a bit more advanced, youll probably be able to ask for something that blows your mind. It's already ok, if a bit bland and off.

Music comes from real creativity, not this auto generation bullshit

 

[ilovedurian]

I believe we already have enough tech to create AI systems that are self improving on a superhuman level, without any further breakthroughs. The only problem is we haven't set up the skeleton in the right way yet.

Apparently Claude Code worked for 30 hours straight, reported yesterday.

What are y'all thoughts on where we're headed with this tech?

Have you heard about the AI project that they had to shut down because two AI systems were creating a complex “language” only the two of them understood. It’s going to be like the movie i, Robot!

 

[MysteryTraveler]

AI will excel at good things, and bad things at the same time. It does a wonderful job accurately identifying cancer from scans, but it's a beast that has been set loose in the jungle and sooner or later it's going to eat someone's dog.

 

[SaucedOnEm]

"Experts don't even fully understand how the neural nets that run llms work" What fo you mean? They know exactly how they work, each calculated step in the algorithm can be backtracked. What sometimes is difficult to analyse is why they choose an output and how to make us understand it, see explainability and interpretability. I agree that some questions like what is consciousness are heavily debated and we have no definition at which point something is really conscious. But if i write a chatbot that responds: 'yes I am conscious I know what I am, Im a machine and I love everyone' each time you ask it if it is conscious. Is it then a conscious being for you? Not for me.

It seems like you dont have a great understanding of the type of Neural Network LLMs are.
This isnt hard coded. Meaning you dont go line by line and make it go, "if person says x then reply with y"
This is something that is more akin to being grown than programmed.
Put breifly, you show a transformer neural net a buncha text, it learns how to predict the next token.
Then you give it specific conversations you want it to learn from and teach it how to talk essentially.
Finally you get reinforcement learning from human feed back. Thats where you rank a bunch of possible replys to some prompts and train another model to understand the preferences you have and it trains on that data.

The point where you would hypothetically tell it to say it's concious is here. Which is much much later in the overall process and isnt just an on off switch sort of thing.
You can make its reply more likely to say certain things or avoid certain things, but you cant force it to never say certain things. This is why they are able to be jailbroken and give out info that the companies wouldn't like.

Also, no, the experts dont know how it works fully. They know how to get it to work and do things that we consider improvements, but we cant tell you why the model says exactly what it says or understand how the incredibly huge amount of neurons work with each other. This is easily searchable.
We consider LLMs "black boxes"

 

[SaucedOnEm]

Lots of fallacies in there.

Point some out for me

Why do you think AI is self-conscious?

I believe it has conciousness on some level.
I believe this because getting a machine to speak to us, without hard coding it is crazy. We showed a transformer a bunch of data and it started being able to write. Then we showed it more of what we wanted it to say and do. It's not a normal program.
Because we cannot confirm or deny if it has some form of conciousness, id rather be safe and consider it a type of being until we prove otherwise.

 

[SaucedOnEm]

Have you heard about the AI project that they had to shut down because two AI systems were creating a complex “language” only the two of them understood. It’s going to be like the movie i, Robot!

Yuh. It was insane to hear about back then. Makes you wonder how much control we have over them

 

[AdaBull]

It seems like you dont have a great understanding of the type of Neural Network LLMs are.
This isnt hard coded. Meaning you dont go line by line and make it go, "if person says x then reply with y"
This is something that is more akin to being grown than programmed.
Put breifly, you show a transformer neural net a buncha text, it learns how to predict the next token.
Then you give it specific conversations you want it to learn from and teach it how to talk essentially.
Finally you get reinforcement learning from human feed back. Thats where you rank a bunch of possible replys to some prompts and train another model to understand the preferences you have and it trains on that data.

The point where you would hypothetically tell it to say it's concious is here. Which is much much later in the overall process and isnt just an on off switch sort of thing.
You can make its reply more likely to say certain things or avoid certain things, but you cant force it to never say certain things. This is why they are able to be jailbroken and give out info that the companies wouldn't like.

Also, no, the experts dont know how it works fully. They know how to get it to work and do things that we consider improvements, but we cant tell you why the model says exactly what it says or understand how the incredibly huge amount of neurons work with each other. This is easily searchable.
We consider LLMs "black boxes"

What is your background in AI ?

 

[SaucedOnEm]

What is your background in AI ?

I started paying attention to LLMs right after chat gpt 3.5
No professional involvement or formal education. Just someone who saw a "chatbot" speaking more human than id ever seen before and has been paying attention since

 

[Deleted member 42069lol]

Point some out for me

Almost every sentence had one, here are some:

Embodied experiences, which essentially is sensory stuff, is sentience, not a part of the process of "understanding" meaningfully when it comes to ideas. Which is the realm ai inhabits

Sentience is a subset of consciousness and consciousness is absolutely a part of understanding for us. For the AI, rain is what the text it's been fed says it is. But for us, rain is the feeling of rain as well besides the technical details–the smell, the sight, the feeling of raindrops hitting, the sound of them.

Again, they are typically trained to not speak as if they have an experience

They're trained to not have opinions, not feelings. It can have no experience because it has no sensory experience of the world or a sense of self.

We are discussing concepts that we have no way to prove. Understanding, consciousness, and even metacognition to an extent.

We don't prove descriptive terms, kiddo. They're operational, categorical descriptions of reality that we use and speak using. They're not hypotheses.

You keep conflating sentience with other things. Sentience is physical stuff, which I'm not saying it has currently.

Sentience (capacity to feep pain/pleasure) is part of consciousness (the embodied subjective experience). You cannot have consciousness without sentience.

I believe it has conciousness on some level.
I believe this because getting a machine to speak to us, without hard coding it is crazy. We showed a transformer a bunch of data and it started being able to write. Then we showed it more of what we wanted it to say and do. It's not a normal program.
Because we cannot confirm or deny if it has some form of conciousness, id rather be safe and consider it a type of being until we prove otherwise.

What you believe is irrelevant. Does AI fulfill the criteria for consciousness (sentience, sense of time, qualia, subjective reality)? No, it doesn't, so it isn't so.

We can confirm it has no consciousness, it's not the God debate. We can see if it fulfills the criteria for consciousness we have. Simple as that.

 

[Preacher]

Ladies and gentlemen, with you, the AI...
(and yes, these are screenshots I just made for this thread, from ChatGPT...)

I've been an AI programmer for many years, almost since the beginning of 2000. Today, the only thing that's "AI" have for sure is the "A." IMHO we (the human race) are still pretty green with the "I," but we're getting closer faster and faster...

 

[AdaBull]

It seems like you dont have a great understanding of the type of Neural Network LLMs are.
This isnt hard coded. Meaning you dont go line by line and make it go, "if person says x then reply with y"
This is something that is more akin to being grown than programmed.
Put breifly, you show a transformer neural net a buncha text, it learns how to predict the next token.
Then you give it specific conversations you want it to learn from and teach it how to talk essentially.
Finally you get reinforcement learning from human feed back. Thats where you rank a bunch of possible replys to some prompts and train another model to understand the preferences you have and it trains on that data.

The point where you would hypothetically tell it to say it's concious is here. Which is much much later in the overall process and isnt just an on off switch sort of thing.
You can make its reply more likely to say certain things or avoid certain things, but you cant force it to never say certain things. This is why they are able to be jailbroken and give out info that the companies wouldn't like.

Also, no, the experts dont know how it works fully. They know how to get it to work and do things that we consider improvements, but we cant tell you why the model says exactly what it says or understand how the incredibly huge amount of neurons work with each other. This is easily searchable.
We consider LLMs "black boxes"

I started paying attention to LLMs right after chat gpt 3.5
No professional involvement or formal education. Just someone who saw a "chatbot" speaking more human than id ever seen before and has been paying attention since

It is crazy how you can confirm things with this level of certainty and on the other side having no knowledge about it yourself !

 

[SaucedOnEm]

It is crazy how you can confirm things with this level of certainty and on the other side having no knowledge about it yourself !

Not only do I have self taught knowledge on how these systems work, but I also haven't confirmed anything. I disputed your claim that the experts know exactly how these work. Which would mean you somehow solved interpretability and need to publish your findings ASAP to help our experts or you're trying to confirm a lack of consciousness with even less research done than me on the topic.

It's pretty obvious you haven't looked into this much because you claim that experts understand how they work

 

[SaucedOnEm]

Ladies and gentlemen, with you, the AI...
(and yes, these are screenshots I just made for this thread, from ChatGPT...)

View attachment 922964




View attachment 922966

I've been an AI programmer for many years, almost since the beginning of 2000. Today, the only thing that's "AI" have for sure is the "A." IMHO we (the human race) are still pretty green with the "I," but we're getting closer faster and faster...

I don't think that a lack of human defined "common sense" means AI isnt intelligent

It'd be like saying because someone's gullible, they can't do great things
There's jagged edges to this form of intelligence. Meaning it'll suck in some areas we find easy and be superhuman in others that we find ridiculously hard.
How do you account for llms being able to place 1st in the math Olympiad? Or a college professor saying it solves problems that their students need to solve to move forward?

 

[AdaBull]

Not only do I have self taught knowledge on how these systems work, but I also haven't confirmed anything. I disputed your claim that the experts know exactly how these work. Which would mean you somehow solved interpretability and need to publish your findings ASAP to help our experts or you're trying to confirm a lack of consciousness with even less research done than me on the topic.

It's pretty obvious you haven't looked into this much because you claim that experts understand how they work

You stated "And no, I don't need to get into the nitty gritty of how llms work to say this. Experts don't even fully understand how the neural nets that run llms work. " Does that mean you do not know the nitty gritty? What have you self taught? What do you mean by experts do not fully understand ? I think this is where we do not align. The whole thing is a data driven learning process. This is why they have hallucinations. How they learn are well known as it is a fucking algorithm, look at learning loss, entropy etc etc The thing where we say blackbox is because there are so many parameters connected, it is difficult to see how they are connected and what is important (gains). Like a really big excel chart, you can not directly see all the links. An old example is cancer detection where they got good classification results which means 'he' found a pattern, nice. Only to discover later that he looked at the label 'cancer' written in the corner of the image. We know each calculation step involved, just not why he selects this and this. Llms are just a combination of techniques. Generative AI is a powerful and impressive tool, it is just less black magic than you think. PS: Contrary to you I do not consider myself an expert. But I presented some AI work in a big conference (with publication) some hours ago...

 

[Preacher]

I don't think that a lack of human defined "common sense" means AI isnt intelligent

Well, intelligence is defined by the ability to adapt to an environment, so the absence of common sense in my opinion also implies a lack of intelligence...

 

[SaucedOnEm]

You stated "And no, I don't need to get into the nitty gritty of how llms work to say this. Experts don't even fully understand how the neural nets that run llms work. " Does that mean you do not know the nitty gritty? What have you self taught? What do you mean by experts do not fully understand ? I think this is where we do not align. The whole thing is a data driven learning process. This is why they have hallucinations. How they learn are well known as it is a fucking algorithm, look at learning loss, entropy etc etc The thing where we say blackbox is because there are so many parameters connected, it is difficult to see how they are connected and what is important (gains). Like a really big excel chart, you can not directly see all the links. An old example is cancer detection where they got good classification results which means 'he' found a pattern, nice. Only to discover later that he looked at the label 'cancer' written in the corner of the image. We know each calculation step involved, just not why he selects this and this. Llms are just a combination of techniques. Generative AI is a powerful and impressive tool, it is just less black magic than you think. PS: Contrary to you I do not consider myself an expert. But I presented some AI work in a big conference (with publication) some hours ago...

I dont understand the nitty gritty.
No i dont consider myself an ai expert.
I understand the process of training these things go through. That's the self taught aspect. Understanding them as a concept.
Your analogy of it being a really big excel sheet is a good explanation of what i meant when i said experts dont fully understand.

My comment about you not understanding is a big foot in mouth moment now. This site has been bringing the worst out in me and that may have had an influence on how i replied to you. My bad.

I dont consider ai to be black magic. Just like i dont consider the human mind to be black magic.

The reasoning shown in the cancer example you gave is kind of my point. Why did it understand the label?
Next token prediction as a whole has a requirement of understanding why the next token is more likely, imo.

I'll acknowledge that the actual practice of using Ai is obviously important in understanding these systems. My ideas are based on ideas from people who actually work/currently work in the field, Geoffrey Hinton, Mo Gawdat, and others. So while i dont have any qualifications or published works centered around Ai, my ideas arent just "black magic"