I don't think ai is a threat in the way you think it's going to be

[Isaac]

My point is that the most dangerous component of a new technology is the individuals using them. As an example, on ChatGPT, if you try to ask "teach me how to do something illegal and get away with it", the AI will rightfully deny your request.

However, if you were to run your own language model at home without any safety measures, you could ask it "I have all of these items in my house... (list of items)... how can I get high off them?" and it would provide you with an answer, based on the accuracy of the model.

This could be the future nightmare scenario, where unrestricted language models become valuable to people and businesses who need them. Although no cloud service would facilitate breaking the law, there could be a demand for an unrestricted language model in private settings.

Edit:
Please ignore my post in the Tech forum asking about building a GPU sever.

[vision]

This could be the future nightmare scenario, where unrestricted language models become valuable to people and businesses who need them.

We are already at this stage. I used to not be worried about AI, but over the past year I've realized there is an extreme danger to how AI is being developed and implemented. First, you have to understand that whether you like it or not everything is a weapon. Second, you have to realize we have been in a hot information war for many years now (this war has profound real world effects on people who have no idea how they are being manipulated, and it's the cause of incalculable death and suffering). Third, know that the people who currently wield this technology (businesses and governments) have only their self interest in mind, not the welfare of humanity let alone the world. These new AI tools are poised to be devastating weapons in the information war. If you think the Internet is adversarial today just wait a couple years. We've basically ruined one of the best tools humans have ever made, which is fitting since we also ruined the Earth and everything we touch. We don't have countermeasures to defend against malicious information tools and by the time we do it will be too late.

[Isaac]

Your point is valid, but we're not there yet as far as what I'm referring to: where most people can run their own AI model from the privacy of their home, without a cloud service. Currently, to run an unrestricted model, you need the resources to afford the computing power and access to the experts required to develop your model. This means not everyone will have access to the latest homebrew. Google appears foolish right now because it's not easy. However, the future I'm talking about is just around the corner. Imagine being able to spend $3000 for a single GPU with 1TB of VRAM. Once we reach that point, everything will change. Homebrew models will be everywhere, and with model merging, you'll be able to add functionality without training.

Just to clarify where we currently stand:

Suppose you want to create your own chatbot like ChatGPT and plan to train your GPT model independently. In such a case, your language model may require 1TB of VRAM or more, necessitating the use of multiple GPUs and dividing your model into segments. This process, known as parallelism, involves distributing different portions of the model across multiple GPUs. This comes with its own challenges, such as increased complexity in model development, potential communication overhead between GPUs, and difficulty in achieving efficient scaling. However, this technique allows you to have a model that significantly exceeds the VRAM capacity of any single GPU. Most homebrew models are designed to operate on gaming GPUs, like the NVIDIA 4090s. Why gaming GPUs? That's what the average person has! And that's where we're at. We're just not at the stage that my original post refers to.

[Tunnelcat]

Me thinks Isaac is STILL using the bot to write his answers. The clue is that the answers are too descriptive with the proper syntax and punctuation, like an English teacher with a computer degree replied. No rambling off into the weirdsphere like Isaac normally would be doing.

[vision]

Me thinks Isaac is STILL using the bot to write his answers.

It's obvious, and also why I didn't reply. In fact, I won't be relying to any of his post from here on out because this kind of behavior breaks the trust essential to a healthy online community.

[Tunnelcat]

Will the REAL Isaac please stand up? I'm going to take it as a challenge to figure out which is Isaac and which is the bot. So far, easy peasy.

Now this scares me. I've got voice recognition to access my account with one of my banks. This new AI development kind of throws a big monkey wrench into the system for using that method as a form of security.

https://arstechnica.com/information-tec ... -of-audio/

[Isaac]

I think that using ChatGPT to clean up my writing is a good idea, especially because my writing skills are not very strong. I often struggle with grammar, spelling, and punctuation, which can make my writing difficult to read and understand.

I believe that using ChatGPT can help me identify and correct errors in my writing, which can make my work more polished and professional. Additionally, ChatGPT can suggest better word choices and sentence structures, which can improve the clarity and flow of my writing.

I understand that some people might feel that using ChatGPT is cheating or takes away from the authenticity of their writing, but I don't agree. I see it as a tool to help me improve my skills and become a better writer over time.

Me thinks Isaac is STILL using the bot to write his answers.

It's obvious, and also why I didn't reply. In fact, I won't be relying to any of his post from here on out because this kind of behavior breaks the trust essential to a healthy online community.

Or perhaps you still don't understand my post even after cleaning it up.

[vision]

My last post in this thread will be a story about a musician I know. He's a decent guitar player but not a great singer. However, his voice on recordings sounds good enough thanks the the magic of autotune, a technology that's everywhere in music and still quite contentious. But there is something else about his singing that didn't sit right with me. The lyrics to all his songs are quite verbose and feel much more like prose than poetry. On top of that, the content seems out of character for my friend in a way I couldn't exactly describe and I could never figure out why until we had a discussion one day about songwriting workflows. It turns out the last several years he's been using lyric generation websites to create the bulk of his lyrics, then changing a couple words here and there to personalize it. Learning this, I was tremendously disappointed. Good songwriting is a craft that takes hard work. He's cut a lot of corners through the use of technology and in the end the result is "average" at best. And that's the most important point I want to make about how these language models work: they generate an average of their training data, literally. The word prediction algorithm builds content word by word choosing the next most likely based on the previous. I don't know how many people in the world are satisfied with mediocrity, but I'm not one of them.

[Isaac]

I don't know how many people in the world are satisfied with mediocrity, but I'm not one of them.

I say if you can dish mediocrity out you should be able to take it from others.

[Tunnelcat]

Isaac, do you think an AI could analyze all the works of some famous author, say Stephan King, come up with an algorithm to replicate his style and write a full length novel that could fool most human readers? Or would the emotion be missing from the story?

[Isaac]

Isaac, do you think an AI could analyze all the works of some famous author, say Stephan King, come up with an algorithm to replicate his style and write a full length novel that could fool most human readers? Or would the emotion be missing from the story?

On my phone. I might be writing badly:

Not in its current form, but it would be helpful for not making mistakes in the cannon. The could review a new storyline and return contradicts or things that come close to an event or a character path crossing from another King book. It would be like a having the most knowledge nerd on all King books to help you stay grounded in that universe.

It could help with character powers like Dr. Sleep, the green mile, pet cemetery, and IT so you're not adding random power sets.

One ai to map all characters on a timeline. Another to try to sort out the effects of every cause. A third to interact with the user and make use of the data made by the other two ai programs.

With this kind of analysis you could make sure your character is properly connected and not just a story written in the style of king.

[Flatlander]

One ai to map all characters on a timeline. Another to try to sort out the effects of every cause. A third to interact with the user and make use of the data made by the other two ai programs.

And one AI to rule them all and in the darkness bind them!

[Isaac]

One ai to map all characters on a timeline. Another to try to sort out the effects of every cause. A third to interact with the user and make use of the data made by the other two ai programs.

And one AI to rule them all and in the darkness bind them!

On my phone again. Forgive my bad writing:

You joke but one of the solutions to various problems with chatgpt is to be goal driven by a secondary RL model: a non-verbal based ai that's trying to decide what the end-goal of a conversation is and then try to achieve it. So let's say you ask it to change your mind about something, you've then given it a clear goal and in the conversation it can focus on walking with you through your logic and then guide you to changing your mind.

[Tunnelcat]

Isaac, that's what makes us human, even when writing novels. Mistakes. Perfection is sometimes boring and pedantic.

[Isaac]

Isaac, that's what makes us human, even when writing novels. Mistakes. Perfection is sometimes boring and pedantic.

Which comment are you referring to? It depends

[Tunnelcat]

I was referring to your comments about mistakes in canon in the example stories you listed. I've read them all and sure, there are subtle errors here and there and canon issues. Sometimes I catch them, sometimes not. I tend to find errors like that in most books and movies anyway, but that doesn't lessen my enjoyment. Like I said, it's human mistakes that let us know a human wrote them. However, sometimes mistakes in canon are so egregious they get noticed, even by kids. Take the last 2 Avenger movies, Infinity Wars and Endgame. There were so many time continuum violations and paradoxes that it kind of messed up the whole story and gave me a headache trying to sort it all out, unbelievable sci-fi fantasy escapism or not. When the whole story line revolves around people going back in time and what results from that travel, they'd better get it close enough to pass a little scrutiny.

[woodchip]

I Tend to agree with vision but another factor is part of this; that's human nature. Humans tend to want someone or something to handle things for them. That's why we have politicians and bosses. If people view AI's as benign they will approve their use, if not AI's will be forced on them. I guess the way I see thing being forced on us, someone will see AI's as a way of making a butt load of money or attaining unimaginable power. Either way we're screwed.

[Tunnelcat]

AI will creep in because humans are lazy and want everything done for them. I can just imagine what's going to happen in our schools with the easy access to this thing by lazy, corner cutting kids. Many jobs will be taken over by AI over time. By the time we realize that it's taken over and now controls every aspect of our lives, we'll be nothing but stupid meat sacks and screwed.

[Darth Wang]

AI will creep in because humans are lazy and want everything done for them. I can just imagine what's going to happen in our schools with the easy access to this thing by lazy, corner cutting kids. Many jobs will be taken over by AI over time. By the time we realize that it's taken over and now controls every aspect of our lives, we'll be nothing but stupid meat sacks and screwed.

Or maybe it will just free up more time for us to do more enjoyable and productive things?

[Tunnelcat]

Oh, that'll be the selling point for sure. But like all things that look wonderful on the surface, there's always that dark downside that eventually bites us in the ass.

[Isaac]

I was referring to your comments about mistakes in canon in the example stories you listed. I've read them all and sure, there are subtle errors here and there and canon issues. Sometimes I catch them, sometimes not. I tend to find errors like that in most books and movies anyway, but that doesn't lessen my enjoyment. Like I said, it's human mistakes that let us know a human wrote them. However, sometimes mistakes in canon are so egregious they get noticed, even by kids. Take the last 2 Avenger movies, Infinity Wars and Endgame. There were so many time continuum violations and paradoxes that it kind of messed up the whole story and gave me a headache trying to sort it all out, unbelievable sci-fi fantasy escapism or not. When the whole story line revolves around people going back in time and what results from that travel, they'd better get it close enough to pass a little scrutiny.

Again I'm on my phone, so i might get a bit illiterate

I think the end product is what you're really referring to and not the planning. The planning stage having an ai assistant is what I was describing. It wouldn't be the actual writer. I think that might have not been clear on my part.

edit:

Look at the last Starwars movies. An ai would have helped that film greatly. The Force Awakens had a lot of problems with character arcs: Han solo owed two different gangs money? Did nothing from his past let him grow or change as it was implied in episode 6? The Last Jedi had an issue with established lore. And the Rise of Skywalker was a complete joke. There's nothing about the human touch that makes these errors charming. It's just really looks like they don't care about Starwars. An ai keeping them on track would have been needed more than ever. It might have even avoided creating the "First Order" just to need an empire foe.

[woodchip]

I was referring to your comments about mistakes in canon in the example stories you listed. I've read them all and sure, there are subtle errors here and there and canon issues. Sometimes I catch them, sometimes not. I tend to find errors like that in most books and movies anyway, but that doesn't lessen my enjoyment. Like I said, it's human mistakes that let us know a human wrote them. However, sometimes mistakes in canon are so egregious they get noticed, even by kids. Take the last 2 Avenger movies, Infinity Wars and Endgame. There were so many time continuum violations and paradoxes that it kind of messed up the whole story and gave me a headache trying to sort it all out, unbelievable sci-fi fantasy escapism or not. When the whole story line revolves around people going back in time and what results from that travel, they'd better get it close enough to pass a little scrutiny.

Again I'm on my phone, so i might get a bit illiterate

I think the end product is what you're really referring to and not the planning. The planning stage having an ai assistant is what I was describing. It wouldn't be the actual writer. I think that might have not been clear on my part.

edit:

Look at the last Starwars movies. An ai would have helped that film greatly. The Force Awakens had a lot of problems with character arcs: Han solo owed two different gangs money? Did nothing from his past let him grow or change as it was implied in episode 6? The Last Jedi had an issue with established lore. And the Rise of Skywalker was a complete joke. There's nothing about the human touch that makes these errors charming. It's just really looks like they don't care about Starwars. An ai keeping them on track would have been needed more than ever. It might have even avoided creating the "First Order" just to need an empire foe.

First they get you used to the assistant then they get you used to the boss.

[Tunnelcat]

Isaac, the first Star Wars was plagued by bad and cheesy acting, although they got better at it by the 3rd installment. No AI could've helped that part of the mess. Yet people clamored to those movies back in the 70's like they were a cocaine fix. I do agree that every movie that followed the first 3 were either poorly influenced by Lucas, who mistakenly thought Star Wars was just a kid's movie and later Disney for continuing that theme to this day, and doing it with so many mistakes. Maybe an AI could've helped with the purging of the insidious Disney kiddification that kept creeping in like syrup, almost as a production requirement. The newest installment in the series, The Mandalorian, positively reeks of this "cuteness", all for toy marketing I'm guessing.

I'd rather that the movies that came later continued the history in all the books. Vector Prime for example. An AI could've filtered out all the badly authored books that showed up as the galaxy story line processed over the years and help make things a lot more coherent. Zahn and his Thrawn series would've been a good offshoot to make a movie out of as well. But noooo, Disney owns it and dammit, Star Wars movies are kid's movies now.

[Top Gun]

The Mandalorian is massively popular, and Andor is apparently one of the best things to ever happen to the franchise, so I'm not really sure where you're going with this.

[Vander]

I lost interest in Mandalorian and didn’t finish the second season. Obi-Won Kenobi was just ok. Andor was pretty great, though.

[Isaac]

RIP original post. Before I reply about starwars would it be best if I started an new thread in the cafe?

[Tunnelcat]

This is more in line with the OP. But i think this endeavor will be akin to smoking, once we get started, we won't be able to quit. A pause is only a pause before it's full steam ahead because all tech-minded people love the possibility of humans developing a machine intelligence, consequences be damned.

https://arstechnica.com/information-tec ... velopment/

[Krom]

https://arstechnica.com/information-tec ... velopment/

• "Should we let machines flood our information channels with propaganda and untruth?"
• "Should we automate away all the jobs, including the fulfilling ones?"
• "Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete, and replace us?"
• "Should we risk loss of control of our civilization?"

I made it this far through the article before laughing because conservatives have already flooded information channels with propaganda and untruth, which means civilization has been out of control for quite some time already. It is far too late on those already. Granted a powerful AI trained in this environment can and will absolutely make things worse much faster than humans could on their own. But even if all AI development was halted and the systems were switched off, so long as conservatives are holding a significant amount of control society will continue to spiral further out of control at an ever accelerating pace with or without AI.

[Isaac]

This is more in line with the OP. But i think this endeavor will be akin to smoking, once we get started, we won't be able to quit. A pause is only a pause before it's full steam ahead because all tech-minded people love the possibility of humans developing a machine intelligence, consequences be damned.

https://arstechnica.com/information-tec ... velopment/

What this thread is talking about is what AI is right now. The OP is talking about what would happen if we each had our own ai hardware with no safety limits. I think everyone skipped the OP.

[Tunnelcat]

That's why I posted the link. It still concerns many people that AI may or will become a threat depending on how it's created or used. The Future of Life Institute is concerned enough to have just made a statement about it, in the effort to ensure that any AI we create is "beneficial", not "undirected", and to make sure this is the stated premise of all research into the area. That's why this group called for a 6 month pause in that research. Yeah, that'll work.

https://futureoflife.org/open-letter/ai-principles/

Given human beings and their predilection for creating things for nefarious purposes, I think this pause request is empty and toothless. Eventually we WILL create an AI that exceeds our own intelligence, may even become sentient and will end up inhabiting all things or equipment that's powered, electronic or computational in nature. That's a staggeringly large number of products or things that we currently use or depend on in our everyday lives. The end result is that there will most certainly be no security or privacy left in our lives. At worst, it may desire to kill us using our own machines against us, or end up being used as a weapon by a hostile government. There's no reason an evolving AI can't end up evil or self-protecting bent on survival, especially if it's intentionally coded that way or learns it over time, as we've seen with ChatGP.

[Isaac]

I'm trying to draw a distinction between the commercial models like chatgpt and the home-brew models nobody is talking about yet. Sorry! Please be patient with my bad explanations.

Here's something you might not have seen yet. It's some of the home-brew models they're releasing to the public: https://huggingface.co/models?pipeline_ ... generation

Don't worry, they're mostly not that good because they're designed to run on a high-end gaming video card. My point is, these corny home-brew models are going to catch up to what chatgpt is today as home gpu hardware gets better. Imagine in 5 years YOU had your own chatgpt that didn't need internet to function, except it has no safeguards. In other words, you can ask it "How can I steal a car and get away with it" and it would ask you for makes and models of the car you were targeting. Imagine everyone having their own "Better Call Saul"+"james bond" advisor their disposal. It can be as racist as you want or as criminal as you want.

Your comments are focusing on about how good chatgpt is going to get, but that's still a commercial product by a company that doesn't want to release a racist or bias product. They're trying to make it safe. These walled-garden style language model products are not where the true danger is.

[woodchip]

https://arstechnica.com/information-tec ... velopment/

• "Should we let machines flood our information channels with propaganda and untruth?"
• "Should we automate away all the jobs, including the fulfilling ones?"
• "Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete, and replace us?"
• "Should we risk loss of control of our civilization?"

I made it this far through the article before laughing because conservatives have already flooded information channels with propaganda and untruth, which means civilization has been out of control for quite some time already. It is far too late on those already. Granted a powerful AI trained in this environment can and will absolutely make things worse much faster than humans could on their own. But even if all AI development was halted and the systems were switched off, so long as conservatives are holding a significant amount of control society will continue to spiral further out of control at an ever accelerating pace with or without AI.

Interesting. Do conservatives control the NEA, The unions? Other than Fox news do the conservative control the major news outlets? Other than talk radio (and liberals tried to get a foot hold in this market with their failed Air America and then failed again when they tried to pass laws under a "Fairness Doctrine"), just exactly how do the conservatives hold control? Did the conservative tell the FBI to coerce Twitter to control posters regarding messages about Hunter Biden. Did you realize a poll was taken and 15% of voters that voted for Biden wouldn't have if they had heard about Hunters laptop? Not to mention the Dems had control of the House . Lest I babble on much more, you'll have to give actual examples of how the conservatives are controlling things instead of speaking in generalities.

[Krom]

Do your own homework, its right there in your browser history.

[Isaac]

I wasn't intending for this to become a political topic. In fact if it does we'll have half the people wanting to ban ai and the other half trying to shove it down our throats.

[woodchip]

Do your own homework, its right there in your browser history.

Stop with the weak replies.

[Isaac]

Do your own homework, its right there in your browser history.

Stop with the weak replies.

You want krom to really give "it" to you, huh?

[Tunnelcat]

I'm trying to draw a distinction between the commercial models like chatgpt and the home-brew models nobody is talking about yet. Sorry! Please be patient with my bad explanations.

Here's something you might not have seen yet. It's some of the home-brew models they're releasing to the public: https://huggingface.co/models?pipeline_ ... generation

Don't worry, they're mostly not that good because they're designed to run on a high-end gaming video card. My point is, these corny home-brew models are going to catch up to what chatgpt is today as home gpu hardware gets better. Imagine in 5 years YOU had your own chatgpt that didn't need internet to function, except it has no safeguards. In other words, you can ask it "How can I steal a car and get away with it" and it would ask you for makes and models of the car you were targeting. Imagine everyone having their own "Better Call Saul"+"james bond" advisor their disposal. It can be as racist as you want or as criminal as you want.

Your comments are focusing on about how good chatgpt is going to get, but that's still a commercial product by a company that doesn't want to release a racist or bias product. They're trying to make it safe. These walled-garden style language model products are not where the true danger is.

So someone in their bedroom codes and creates an AI, posts it on the net and lets people play and interact with it, which I assume their point is to let that AI modify itself, learn and grow. Creepy but, it's just a program toy at that moment. It's the kind of people playing with it and where they steer it's capabilities that concerns me. No regulation and no oversight, no restrictions. It would be comparable to someone messing with the code of an organic virus (not alive by the way, which is nothing more than an organic program wrapped up in a shell covered in protein keys that allow it to open cell membranes and gain access), and letting it loose in the wild to see what happens. A virus has no sentience, yet it can kill millions if it has the right codes. It's fortunate that it's difficult to modify viruses intentionally.

But an AI program learns from human interactions right? So that means it's self-modifying and evolves over time to become whatever it's taught or experiences during any interactions between it and humans or even other programs. If what happened with ChapGP is any indication, it can go sideways real fast and become a monster we can't control. You read most Sci-Fi books and the one overlying concern with any AI is if it becomes sentient and what might happen if any controls that keep that it "caged" are bypassed. Sentience means free will. Can a human create an AI that could possibly evolve on it's own over time and gain sentience, then like all life, begin to worry about it's own existence and whether it should defend itself? Most of the time, people are afraid of that possibility. And yet, few people even see that as a possibility of happening at this point in time.

[Top Gun]

ChatGPT is not remotely "intelligent," nor does it have any capability to become so. That doesn't mean that its usage can't be incredibly problematic, however.

[Isaac]

So someone in their bedroom codes and creates

Which has been taking AI far... Different topic, but look at the current state of Stable Diffusion art. It gets better every week and is currently better than OpenAI's DALL-E and the paid service Midjourney. Not every month. Every week. Hands and arm glitches used to be everywhere, but they keep getting better, and the reason it's improving is that all you need is a CUDA-compatible GPU or a fast CPU: https://civitai.com/. Note that there will be a mix of models and images on that site, both new and old.

...an AI, posts it on the net and lets people play and interact with it, which I assume their point is to let that AI modify itself, learn and grow.

Exactly, but maybe not in the way you imagine They're using something called Unsupervised Learning (We'll call it UL). More specifically, a type of UL called Self-Supervised Learning, where you have the model train itself.

Creepy but, it's just a program toy at that moment.

The home-brew models are not even a good toy. However this will change as home desktop GPUs get more powerful. I said it before but when we can have 1 TB of vram these "toys" become wild versions of chatgpt.

It's the kind of people playing with it and where they steer it's capabilities that concerns me. No regulation and no oversight, no restrictions. It would be comparable to someone messing with the code of an organic virus (not alive by the way, which is nothing more than an organic program wrapped up in a shell covered in protein keys that allow it to open cell membranes and gain access), and letting it loose in the wild to see what happens. A virus has no sentience, yet it can kill millions if it has the right codes. It's fortunate that it's difficult to modify viruses intentionally.

But an AI program learns from human interactions right? So that means it's self-modifying and evolves over time to become whatever it's taught or experiences during any interactions between it and humans or even other programs. If what happened with ChapGP is any indication, it can go sideways real fast and become a monster we can't control. You read most Sci-Fi books and the one overlying concern with any AI is if it becomes sentient and what might happen if any controls that keep that it "caged" are bypassed. Sentience means free will. Can a human create an AI that could possibly evolve on it's own over time and gain sentience, then like all life, begin to worry about it's own existence and whether it should defend itself? Most of the time, people are afraid of that possibility. And yet, few people even see that as a possibility of happening at this point in time.

I'm about to get real boring, but please stay with me. Machine Learning model is NOT a brain. The best way to visualize it is a large collection of vectors.

Let's look at one vector. This is called a simple linear classifier. The red and blue dots are your training data. Then you train your model and it makes the GREEN line: the model, which you can now ask questions to!




Once you have your model, you can plug in values for X and Y, and your model will predict what color your dot would be based on the training data. For example, you enter Y=0.5 and X=0.2, and your model will look at just its model (the green line) and return a blue dot.

Does this seem like machine learning? No? It seems more like statistics, right? Because it is! Machine Learning is all basically using forecasting models in a new way (oversimplified explanation). However, this simple linear classifier example is using a 2-dimensional model. You can easily visualize that. You could even visualize a 3-dimensional model, right? However, I've made 50-dimensional models with my python programs. Language ai models don't use simple linear classifiers, but use more complex models, such as deep learning models or tree-based models. These models work with high-dimensional vector spaces to represent words and their contexts. But they're still all vectors. Thank you, if you made it to the end.

[Tunnelcat]

Uncle.