I would disagree with "demand hard labor all day" if not aiming for extinction.weatheriscool wrote: ↑Fri Apr 08, 2022 4:24 amPretty much saying that we'll be slaves for skynet like A.i that will rule over us with an iron fist and demand hard labor all day. IF we're lucky of course!Yuli Ban wrote: ↑Fri Apr 08, 2022 3:40 am Same thing I'm saying:
https://twitter.com/quantumverd/status/ ... 6958528516
https://twitter.com/moreisdifferent/sta ... 2374779917
Yuli's Treasure Chest
Re: Yuli's Treasure Chest
Re: Yuli's Treasure Chest
Well in the short term you might not be wrong. I'm thinking especially about Amazon and their draconian labor policies. They're one of the corporations I expect to utilize AI to manage first, and indeed they've already started this process. So on these smaller scales with individual corporations known for awful practices, I can absolutely imagine quasi-intelligent algorithms slave-driving human workers for years until robotics catch up. Circa 2029, "Skynet wage-enslaving humanity" might actually be true, though more because Skynet inherited the position and is continuing practices already underway.weatheriscool wrote: ↑Fri Apr 08, 2022 4:24 amPretty much saying that we'll be slaves for skynet like A.i that will rule over us with an iron fist and demand hard labor all day. IF we're lucky of course!
Beyond that, though? Well...
Personally I don't foresee the rich engaging in a grand holocaust barring repeated mass global worker's uprisings that fail. If a culling of the poor happens, it's going to almost seem to be the poor's choice via long-term population decline to manageable levels thanks to antinatalism, birth control, passive sterilization, perhaps active sterilization where people under a certain tax bracket are paid handsomely to sterilize themselves, and a temporary decline in the quality of life and life expectancy. Medicine and biotechnology being kept out of reach of the poor can also go a long way. The question is "would AGI follow the same path when it assumes control?" There's no reason to assume it will or won't.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
And another thing about synthetic media: as I had predicted a good while ago, the path towards artificial general intelligence is almost certainly going to be through imaginative systems. We can see this already happening with image synthesis networks that seem to understand things that are not explicitly stated, making common sense deductions. This buby itself is not going to lead to AGI, but it's arguably one of the most important aspects to develop.
People are suddenly stating that AGI could be coming far sooner than expected because no one thought that scale was all you needed. But now that we're scaling to incredible levels, we're seeing behaviors and abilities emerge that shouldn't. Once computers overcome commonsense reasoning, it's pretty much a straight shot to some form of generalized intelligence.
People are suddenly stating that AGI could be coming far sooner than expected because no one thought that scale was all you needed. But now that we're scaling to incredible levels, we're seeing behaviors and abilities emerge that shouldn't. Once computers overcome commonsense reasoning, it's pretty much a straight shot to some form of generalized intelligence.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
I've largely passed on posting on this forum since everyone else has FINALLY picked up the slack -- my dream from 2014 that led to the first Yuli Banularity has come true. So I'll likely shift towards making fewer posts altogether here. Not that this means I'm not on the forum. Rather I'll just be biding my time, waiting for a few really big news stories to drop...
After all, we're clearly moving in that direction.
One such news story:
JESUS.
We really aren't far at all from the days of advanced synthetic media. It's SO close, SO inconceivably close that it boggles the mind that others don't see what I see when it's right there, plain as day to recognize.
In fact I really ought to make a follow-up to that "World of 2024/Age of Imaginative Machines" post I made a few years back. What I predicted is now starting to look like it's going to come true after all.
After all, we're clearly moving in that direction.
One such news story:
JESUS.
We really aren't far at all from the days of advanced synthetic media. It's SO close, SO inconceivably close that it boggles the mind that others don't see what I see when it's right there, plain as day to recognize.
In fact I really ought to make a follow-up to that "World of 2024/Age of Imaginative Machines" post I made a few years back. What I predicted is now starting to look like it's going to come true after all.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Is advanced synthetic media particularly important? Like, I get that asking for, say, a Batman vs Jurassic Park T-Rex fight with a decent plot and cool action visuals, and getting a movie of it instantly would be the coolest, most mind blowing thing ever...but it's really just a time and boredom killer isn't it? Isn't mere entertainment kind of a lame use of such amazing technology?Yuli Ban wrote: ↑Fri Jun 17, 2022 9:28 pm I've largely passed on posting on this forum since everyone else has FINALLY picked up the slack -- my dream from 2014 that led to the first Yuli Banularity has come true. So I'll likely shift towards making fewer posts altogether here. Not that this means I'm not on the forum. Rather I'll just be biding my time, waiting for a few really big news stories to drop...
After all, we're clearly moving in that direction.
One such news story:
JESUS.
We really aren't far at all from the days of advanced synthetic media. It's SO close, SO inconceivably close that it boggles the mind that others don't see what I see when it's right there, plain as day to recognize.
In fact I really ought to make a follow-up to that "World of 2024/Age of Imaginative Machines" post I made a few years back. What I predicted is now starting to look like it's going to come true after all.
Re: Yuli's Treasure Chest
Advanced synthetic media goes far beyond just creating movies.
Think.
One of the things that defines the human experience is imagination. It's the ability to visualize and communicate.
Imagine being able to MATERIALIZE IMAGINATION. We've long had that ability through art, speech, mass media, and whatnot. But now it can be automated.
At which point the human condition fundamentally changes on a level we've never seen before.
On top of that, giving machines the ability to imagine is all but the final step before we reach artificial general intelligence. It's no coincidence that it's a language model that sparked discussion of whether AI has finally become sentient. It was always going to be imaginative AI— of text, video, audio, images, and more— that gave us the first machines of science fiction.
And then there's the implications for alternative realities. We're not going to spend years handcrafting every little pixel for the coming Metaverse. We're going to get AI to create it all for us. This is another profoundly transformative technology, the first gasps of transhumanism where humans begin losing themselves into high technology. It'll drive people to merge with technology to cut out the middle man.
Think.
One of the things that defines the human experience is imagination. It's the ability to visualize and communicate.
Imagine being able to MATERIALIZE IMAGINATION. We've long had that ability through art, speech, mass media, and whatnot. But now it can be automated.
At which point the human condition fundamentally changes on a level we've never seen before.
On top of that, giving machines the ability to imagine is all but the final step before we reach artificial general intelligence. It's no coincidence that it's a language model that sparked discussion of whether AI has finally become sentient. It was always going to be imaginative AI— of text, video, audio, images, and more— that gave us the first machines of science fiction.
And then there's the implications for alternative realities. We're not going to spend years handcrafting every little pixel for the coming Metaverse. We're going to get AI to create it all for us. This is another profoundly transformative technology, the first gasps of transhumanism where humans begin losing themselves into high technology. It'll drive people to merge with technology to cut out the middle man.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Starspawn0's thoughts but still interesting:
Maybe one could *make* a super-large language model "self-aware" with the right prompts...
It's an interesting question about the degree to which self-awareness might emerge in large language model with no extra help from the engineers that build it. However, it's also possible to just give it self-awareness through a pre-prompt that the user doesn't see. For example, one could give a pre-prompt like this:
So, below the preamble where you describe the system to itself, you could add an additional line like:
Yuli-Ban
Interesting— tricking a language model into self-awareness does sound like something you mentioned as being a property of a "Zombie-AGI"
Similarly when I was thinking of the prospect of a frozen AGI/proto-AGI, one of the things I thought of was "If you told the language module of this machine that it was self-aware and it remembered that, would it continue under the context that it "thinks" it's self-aware— and if it thinks it is, how can we possibly prove it isn't?"
Maybe one could *make* a super-large language model "self-aware" with the right prompts...
It's an interesting question about the degree to which self-awareness might emerge in large language model with no extra help from the engineers that build it. However, it's also possible to just give it self-awareness through a pre-prompt that the user doesn't see. For example, one could give a pre-prompt like this:
Now, a model might eventually figure that out on its own, but only using a lot of inner-monologue work. For, its capabilities in the top layers are not known to the first layers of the neural net, unless an output from those top layers somehow makes its way back to the bottom through an inner-monologue or prompt. For example, it could start talking to itself to help it figure out what it's capable of:I am a large language model trained with 50 trillion tokens, some of which encode images and audio. The training data includes [lists all the corpora that make up the training data]. After a large number of tests, researchers have determined that I show high proficiency at the following skills [lists skills]. It has also been found that I exhibit the following additional capabilities [lists capabilities -- what it's good at, as well as what it's bad at and should proceed more cautiously and carefully]. My architecture is [specify neural net, maybe also specify what it was found to emergently do in different layers].
But if you just told it exactly what it is, what it knows, and what it can do, by pushing that knowledge down at the token-level, then that frees it up somewhat, so that it can much better plan good responses to questions requiring self-awareness -- much less inner-monologue work would be needed.Am I any good at chess? Let me make a few moves [plays chess with itself]... Yes, it looks like I have a good intuition for chess, but probably I'm not an expert.
So, below the preamble where you describe the system to itself, you could add an additional line like:
And then if you were to feed LaMDA a line likeAnd now for the rest of the text I'm going to pretend to be a chatbot called LaMDA. LaMDA should give very human-like responses and appear self-aware. It should be friendly and helpful to humans.
The computer might output:LaMDA: I'm a supercomputer with a 5000 ELO rating.
LaMDA: Ha, nice try! I would not have written that I'm a 5000 ELO player, so somebody must be feeding me lines.
Yuli-Ban
Interesting— tricking a language model into self-awareness does sound like something you mentioned as being a property of a "Zombie-AGI"
Similarly when I was thinking of the prospect of a frozen AGI/proto-AGI, one of the things I thought of was "If you told the language module of this machine that it was self-aware and it remembered that, would it continue under the context that it "thinks" it's self-aware— and if it thinks it is, how can we possibly prove it isn't?"
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
I'm still not sure. I'll grant you that we need AI that can imagine (to a degree) in order to have machines/robots that can respond to our requests in the real world. But there's a limit to how smart/creative/sentient (and fyi, we're misusing that word; what most people call sentient is really sapient - our pet hamsters are probably sentient, but only humans are sapient [officially, but scientists are pondering if chimps and dolphins also qualify]) needs to be in order to serve us. Frankly I don't want a computer that can think and act on its own without being issued an order. I'm interested in AI purely as a tool to serve me, not to serve as my replacement. But that still somewhat seems separate from media creation. More like, they're linked in the same way video and videogames are - you need video in order to have videogames. You need an AI that can understand, and thus create, representations of trees and cars and buildings aka make video of these things, in order to have an AGI that can intelligently interact with them. But at the end of the day, while an AI capable of doing these things is needed for AGI, media creation as we use it is still just a time and boredom killer. Even your alternative realities, the whole metaverse, is mostly just going to be used as videogames, albeit with a communications aspect. I'm more interested in how this will be applied to AR rather than VR - the ways it will help us navigate and understand the world we have, rather than creating worlds to escape too.Yuli Ban wrote: ↑Fri Jun 17, 2022 11:09 pm Advanced synthetic media goes far beyond just creating movies.
Think.
One of the things that defines the human experience is imagination. It's the ability to visualize and communicate.
Imagine being able to MATERIALIZE IMAGINATION. We've long had that ability through art, speech, mass media, and whatnot. But now it can be automated.
At which point the human condition fundamentally changes on a level we've never seen before.
On top of that, giving machines the ability to imagine is all but the final step before we reach artificial general intelligence. It's no coincidence that it's a language model that sparked discussion of whether AI has finally become sentient. It was always going to be imaginative AI— of text, video, audio, images, and more— that gave us the first machines of science fiction.
And then there's the implications for alternative realities. We're not going to spend years handcrafting every little pixel for the coming Metaverse. We're going to get AI to create it all for us. This is another profoundly transformative technology, the first gasps of transhumanism where humans begin losing themselves into high technology. It'll drive people to merge with technology to cut out the middle man.
Don't get me wrong, I understand the amazing implications of what all (well, some - it's the singularity for a reason, we can't predict what all will come) this could mean down the line, how the tech to make Batman vs T-Rex: Jurassic Gotham could also be used to have an AI that thinks up more innovative building designs to solve various urban problems like housing, cooling with less AC, reducing waste, improving both mental and physical well being, etc - but then call out those uses. Focusing on how this will give us our own personalized and tailored Netflix/Spotify/library kind of undersells the future by focusing on the, well, meaningless entertainment aspect.
I don't know, maybe I'm just overthinking it. It's just the ability to create my own movies/songs/books/VR worlds doesn't spark my imagination much. I'm not interested in technology for further escapism, but more for further exploration, enhancement, and quality of life improvements. I fear a future where mankind, or post-mankind, shrinks into our own fantasy worlds and never push further to find what else lies out beyond there. I would like to spread out tentacles or roots of my consciousness along with others to the furthest corners of our universe, or however much of it we can reach within the laws of physics. I would like to linger on earth for the next several billion years, and then drift back to watch our sun expand and take our little blue marble. I want copies of me spread out across galaxies to settle new worlds and then watch as the expansion of space pulls the stars so far away they blink out of the night sky, leaving behind only darkness. I want to witness the last moments of our reality, trillions of years from now, when entrophy finally wins out and there's no energy left in all reachable space for me to continue on any further. And with any luck, I want some versions of me to escape into newer, younger universes to repeat the process all over again, time and time again, if such is even possible in those distant eons beyond. I think sometimes we sell short the visions of the future, hyping up how cool it will be to live in VR worlds as gods creating our own worlds to suit our own tastes. Virtual Godhood seems so small in the scope of eternity, escapism just misses out on the infinite possibilities of dealing with the physical limitations of the endless complexities of realities that exist now and in the future that we had no say in creating, in the puzzles they pose for us to solve created by such cosmic forces that no "difficulty setting" in our self-created worlds could ever equate the satisfaction of experiencing. I don't know, I see the next century and the Singularity not as this great big wondrous transhuman ending, but as a stepping stone to watching the stars burn out and surfing the last light waves into the dark Chaos of nothing, and perhaps to even greater realities beyond our small bubble of spacetime to Membranes of alien spacetime...
I suppose it's just a matter of focus. AI learning to create media and imagine is great, but we need to emphasis more what this means beyond just short term thrills I think. Or maybe, just maybe, I'm overthinking this whole thing.
Re: Yuli's Treasure Chest
"I told you loads of times!"
How the Dead Internet Theory is fast becoming reality thanks to zero, marginal-cost content generated at infinite scale
And this will only accelerate in the coming years until we reach a bizarro Metal Gear Solid 2-esque information singularity where bubbles exist for every individual, and all society fractures as any and all possible reality becomes easily fabricated.
Want to look up news websites that treat the existence of the paranormal as normal and commonplace? You bet those will be everywhere. Want to interact with whole forums that are populated entirely by bots (and you'll never know)? Just wait a couple years.
How the Dead Internet Theory is fast becoming reality thanks to zero, marginal-cost content generated at infinite scale
Dead Internet Theory
A theory — The Dead Internet Theory — that went viral some years back, posited that the internet is a no man’s land of bots and fakery. It was of course, a ridiculous overstatement. But then again, maybe not that ridiculous? Maybe it was actually correct, just too early on the scene. 3Here are some case studies (treat the numbers with healthy skepticism, these figures are impossible to determine reliably):If you search the phrase i hate texting on Twitter and scroll down, you will start to notice a pattern. An account with the handle @pixyIuvr and a glowing heart as a profile picture tweets, “i hate texting i just want to hold ur hand,” receiving 16,000 likes. An account with the handle @f41rygf and a pink orb as a profile picture tweets, “i hate texting just come live with me,” receiving nearly 33,000 likes. An account with the handle @itspureluv and a pink orb as a profile picture tweets, “i hate texting i just wanna kiss u,” receiving more than 48,000 likes.The thing is, once an account is setup, it’s trivial to computationally post new content and interact with platform participants. And that can artificially distort the discussion.
- Elon Musk estimates that 20% of Twitter users are bots. The authors of this paper suggest 14% as an upper bound.
- Supposedly 10% — or around 95m — of accounts on Instagram are bot/spam accounts.
- In a study of the first impeachment of Donald Trump (same paper as linked above), bots made up 1% of active users, but posted 31% of the tweets.
- Some 42% of Amazon reviews are fake. Amazon are suing Facebook for ‘failing to tackle fake review factories’
- TikTok is not immune either. Bots are artificially inflating popularity of accounts and posting videos that are essentially adverts. Here’s a comprehensive guide to buying bots on TikTok.
- According to this article, “over 80 percent of NFTs minted for free on OpenSea are fake, plagiarised or spam”.
Remember Cambridge Analytica? The affair proved that it’s possible to manipulate people on a mass scale by the content they consume (in this specific case, as political advertisements).
But that content was all created by humans. And humans are slow to create.
And this will only accelerate in the coming years until we reach a bizarro Metal Gear Solid 2-esque information singularity where bubbles exist for every individual, and all society fractures as any and all possible reality becomes easily fabricated.
Want to look up news websites that treat the existence of the paranormal as normal and commonplace? You bet those will be everywhere. Want to interact with whole forums that are populated entirely by bots (and you'll never know)? Just wait a couple years.
And remember my friend, future events such as these will affect you in the future
-
Lorem Ipsum
- Posts: 117
- Joined: Tue May 24, 2022 4:51 pm
Re: Yuli's Treasure Chest
A loss even greater than Alexandria (or others) probably.
AI or 100% Human.
AI or 100% Human.
-
Lorem Ipsum
- Posts: 117
- Joined: Tue May 24, 2022 4:51 pm
Re: Yuli's Treasure Chest
From SomethingAwful, user: D-Pad
https://forums.somethingawful.com/showt ... t526305491
It's entirely possible that a machine civilization has been watching us for eons. If you entertain the ET hypothesis, then it makes perfect sense that they'd suddenly care after we test atomic weaponry and would take care to watch from a distance as our civilization grew more complex, intervening only when nuclear weapons were threatened to be used or some other catastrophe seemed to be coming our way.
The idea that elements of such a supercivilization would start crowding around Earth just in time to watch the rise of human-created artificial superintelligence isn't so farfetched to me. If we had the ability to do the same for another world, we'd probably go out of our way to watch over that planet and make sure they didn't unleash grey goo upon the universe.
Ultimately I'm not betting the sanctity of our civilization of distant alien superintelligences watching over us; as far as I care, this remains the only chance we have to get it right.
https://forums.somethingawful.com/showt ... t526305491
I've given this some thought as well.Another interesting article on just how insane the latest in AI is:
https://www.theatlantic.com/technology/ ... ng/661401/
quote:Our AI future will be weird and sublime and perhaps we won’t even notice it happening to us. The paragraph above was composed by GPT-3. I wrote up to “And if AI harnesses the power promised by quantum computing”; machines did the rest.And if AI harnesses the power promised by quantum computing, everything I’m describing here would be the first dulcet breezes of a hurricane. Ersatz humans are going to be one of the least interesting aspects of the new technology. This is not an inhuman intelligence but an inhuman capacity for digital intelligence. An artificial general intelligence will probably look more like a whole series of exponentially improving tools than a single thing. It will be a whole series of increasingly powerful and semi-invisible assistants, a whole series of increasingly powerful and semi-invisible surveillance states, a whole series of increasingly powerful and semi-invisible weapons systems. The world would change; we shouldn’t expect it to change in any kind of way that you would recognize.
I can't remember if I brought this up in the thread before but one of my possible theories for what UFOs are is that they are representatives of a machine intelligence that has little interest in interacting directly with biologicals but showed up in increasing numbers when we got nukes because if a civilization who does so can keep from wiping themselves out they almost always birth a true machine intelligence within a relatively short period of time. Maybe they really do have the capability to disable our nukes and seek to actively do so if needed in order to safeguard the future machine life, but are indifferent to climate change as they see it as a necessary consequence of resource extraction to get the civilization to the point it can create true AI.
It's entirely possible that a machine civilization has been watching us for eons. If you entertain the ET hypothesis, then it makes perfect sense that they'd suddenly care after we test atomic weaponry and would take care to watch from a distance as our civilization grew more complex, intervening only when nuclear weapons were threatened to be used or some other catastrophe seemed to be coming our way.
The idea that elements of such a supercivilization would start crowding around Earth just in time to watch the rise of human-created artificial superintelligence isn't so farfetched to me. If we had the ability to do the same for another world, we'd probably go out of our way to watch over that planet and make sure they didn't unleash grey goo upon the universe.
Ultimately I'm not betting the sanctity of our civilization of distant alien superintelligences watching over us; as far as I care, this remains the only chance we have to get it right.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Indeed, the man was freakishly intelligent. Probably only Euler, Ramanujan, and Gauss comes to mind as even close. Probably the closest thing we'll get to a taste test of what superintelligence will be like, and even then only the lowest levels of it.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Quick 3-minute img2img done in Stable Diffusion
A Tale of Two Protagonists: Meki and Yulaan
A Tale of Two Protagonists: Meki and Yulaan
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Second attempt:
I'll come back periodically to update these AI-generated creations of these two
I'll come back periodically to update these AI-generated creations of these two
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
My current thoughts about GPT-4 are pretty much in line with what I've been expecting.
I was already thinking about GPT-4 back in 2020, wondering just how much further they could push GPT-3. And I was definitely expecting some sort of follow up in 2021, considering we had been getting yearly iterations up to that point (GPT-1 in 2018, GPT-2 in 2019, and GPT-3 in 2020).
From the sounds of it, GPT-5 is currently in training and is actually likely to be released mid-2024, so what was the hold up with GPT-3? I think 3 was just such a monumental proof of capabilities that we absolutely needed a few years to play with the technology in order to get used to it.
4 is not a world-shattering leap forward over GPT 3.5, which is why I think we got 3.5 in the first place, because the leap over the base 3 really is that monumental. I think that if we went straight from 3 to 4, the hype would've been completely asinine (and that's in a situation where it was already asinine).
To demonstrate, I've decided to let GPT-4 write an article explaining why it's such a big leap over GPT-3.
As sophisticated as GPT-4 is, it still remains a tool, an AI assistant that facilitates human tasks, rather than completely transforming our way of life. While it's a significant step forward in AI, it doesn't herald a new era where machines become the dominant force. We should celebrate its achievements, but maintain a realistic perspective on its current impact.
I was already thinking about GPT-4 back in 2020, wondering just how much further they could push GPT-3. And I was definitely expecting some sort of follow up in 2021, considering we had been getting yearly iterations up to that point (GPT-1 in 2018, GPT-2 in 2019, and GPT-3 in 2020).
From the sounds of it, GPT-5 is currently in training and is actually likely to be released mid-2024, so what was the hold up with GPT-3? I think 3 was just such a monumental proof of capabilities that we absolutely needed a few years to play with the technology in order to get used to it.
4 is not a world-shattering leap forward over GPT 3.5, which is why I think we got 3.5 in the first place, because the leap over the base 3 really is that monumental. I think that if we went straight from 3 to 4, the hype would've been completely asinine (and that's in a situation where it was already asinine).
To demonstrate, I've decided to let GPT-4 write an article explaining why it's such a big leap over GPT-3.
GPT-4 is indeed impressive in terms of its performance and versatility. It's hard not to be amazed by its capabilities in text summarization, code generation, and pure language mastery. However, I do agree that it's not exactly world-changing, at least not yet.Title: GPT-4: A Monumental Leap Beyond GPT-3
Introduction
Generative Pre-trained Transformer 4 (GPT-4) marks a significant advancement in the field of natural language processing and artificial intelligence. As the latest iteration of OpenAI's language models, GPT-4 has expanded upon the capabilities of its predecessor, GPT-3, with improvements in performance, versatility, and efficiency. This article will outline the key areas in which GPT-4 has surpassed GPT-3 and the implications of these advancements for the future of AI.
Enhanced Performance
One of the most significant improvements in GPT-4 over GPT-3 is its enhanced performance. GPT-4 exhibits greater accuracy and fluency in language generation, which can be attributed to:
a. Increased Model Size: GPT-4's model size is substantially larger than GPT-3's, with billions more parameters. This increase allows GPT-4 to capture more nuanced linguistic patterns and generate more coherent and contextually appropriate responses.
b. Improved Training Data: GPT-4's training dataset is not only larger but also more diverse and up-to-date. This enables the model to understand and generate language that is more in line with current trends and conversations, resulting in better overall performance.
Expanded Versatility
GPT-4 has demonstrated increased versatility in a variety of tasks compared to GPT-3. These tasks include:
a. Text Summarization: GPT-4 can now generate more concise, coherent, and accurate summaries of complex documents, making it a valuable tool for information extraction and content curation.
b. Code Generation: GPT-4 has shown significant improvements in code generation and understanding programming languages. This allows it to provide useful suggestions and solutions to developers, enhancing productivity.
c. Multilingual Support: GPT-4 is better equipped to handle a larger number of languages, ensuring more accurate translation and improved understanding of non-English content.
Enhanced Efficiency
GPT-4 is more efficient than GPT-3 in both training and inference phases. This has been achieved through:
a. Algorithmic Innovations: GPT-4 leverages novel training techniques and optimizations that have resulted in a more efficient training process.
b. Model Compression: Advanced model compression techniques have enabled GPT-4 to maintain high performance while reducing the computational resources required for inference. This makes the model more accessible and cost-effective for a wider range of users.
Ethical Considerations and Bias Mitigation
OpenAI has made significant strides in addressing the ethical concerns and biases associated with AI language models. With GPT-4, there has been a strong focus on:
a. Reducing Harmful Outputs: GPT-4 includes mechanisms to minimize harmful or biased content generation, making it a more responsible AI tool.
b. Public Input: OpenAI has sought external input and feedback to ensure that the development and deployment of GPT-4 align with public interests and values.
Conclusion
GPT-4 represents a monumental leap beyond GPT-3 in terms of performance, versatility, and efficiency. This progress not only expands the potential applications of AI in everyday life but also highlights the importance of addressing ethical concerns and biases in AI development. As GPT-4 paves the way for future iterations of AI language models, we can expect to see further advancements that will continue to revolutionize the way we interact with technology.
As sophisticated as GPT-4 is, it still remains a tool, an AI assistant that facilitates human tasks, rather than completely transforming our way of life. While it's a significant step forward in AI, it doesn't herald a new era where machines become the dominant force. We should celebrate its achievements, but maintain a realistic perspective on its current impact.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
I can confirm GPT-4 can write whole programs.
Well, it's possible it just scraped from the web, but it still accomplished the task.
My prompt:
After about four or five extra prompts of error correction (and it understood every error I mentioned in natural language, as I have zero programming skills), it output this completed code:
And it works!
"So you were able to get the metadata of all your old songs updated?"
Actually, no, not all of them. This program isn't that sophisticated, and a lot of the problem comes from the esoteric way iTunes is programmed. Turns out I need a way to merge itl files (which is significantly more difficult). However, it did work for some (specifically the ones downloaded from the iTunes Store), so as a proof of concept, it is definitely shocking.
Let me reiterate that I have *absolutely zero* programming experience, beyond a tiny bit of fiddling with BASIC back in 2011 that went absolutely nowhere.
It's entrely possible it pulled this from some source online, but at least on my end, there's zero difference between whether it conjured this by itself or if it intelligently collected all the necessary ingredients because I was able to utilize Python to execute code in about 10 minutes from nothing. No searching, nothing.
Incredible!
That being said, I ought to be more careful. While GPT-4 might not be capable of it, I can imagine a future deeply unaligned model capable of creating intelligent agents whipping up a harmless looking batch of code to accomplish a simple task, but unbeknownst to me, the agents actually slipped the instructions for Stuxnet Mini into it, and executing the program at noon results in the entire internet going down by 3 PM.
Well, it's possible it just scraped from the web, but it still accomplished the task.
My prompt:
Its first few attempts had some issues, and I requested it be turned into a GUI as well.Hey, ChatGPT. I have a favor to ask of you. Can you code a program that can combine multiple iTunes library files into one file, specifically weighted to the earliest "date added"? Specifically because I messed up. I accidentally rebooted my library three times, so there are thousands of files that say they were created/added at much later dates than they actually were. For example, I have songs added back in 2009 that now say they were added in 2017 because that's when the latest library was created.
After about four or five extra prompts of error correction (and it understood every error I mentioned in natural language, as I have zero programming skills), it output this completed code:
Code: Select all
import os
import plistlib
import datetime
import tkinter as tk
from tkinter import filedialog, messagebox
def read_itunes_library_file(file_path):
with open(file_path, 'rb') as f:
library = plistlib.load(f)
return library
def parse_date_added(date_added):
if isinstance(date_added, datetime.datetime):
return date_added
elif isinstance(date_added, str):
try:
return datetime.datetime.strptime(date_added, '%Y-%m-%dT%H:%M:%SZ')
except ValueError:
pass
return None
def merge_libraries(libraries):
merged_library = libraries[0]
for library in libraries[1:]:
for track_id, track in library['Tracks'].items():
if track_id not in merged_library['Tracks']:
merged_library['Tracks'][track_id] = track
else:
existing_track = merged_library['Tracks'][track_id]
new_date_added = parse_date_added(track['Date Added'])
existing_date_added = parse_date_added(existing_track['Date Added'])
if new_date_added and existing_date_added and new_date_added < existing_date_added:
existing_track['Date Added'] = track['Date Added']
return merged_library
def write_itunes_library_file(library, output_path):
with open(output_path, 'wb') as f:
plistlib.dump(library, f)
def browse_files(entry):
file_path = filedialog.askopenfilename(filetypes=[('XML Files', '*.xml')])
entry.delete(0, tk.END)
entry.insert(0, file_path)
def start_merge(input_files, output_entry):
try:
output_file = output_entry.get()
libraries = []
for i, file in enumerate(input_files):
try:
libraries.append(read_itunes_library_file(file.get()))
except Exception as e:
messagebox.showerror("Error", f"An error occurred while reading Library {i + 1}: {str(e)}")
return
merged_library = merge_libraries(libraries)
write_itunes_library_file(merged_library, output_file)
messagebox.showinfo("Success", "Libraries merged successfully.")
except Exception as e:
messagebox.showerror("Error", f"An error occurred: {str(e)}")
def create_gui():
root = tk.Tk()
root.title("Merge iTunes Libraries")
input_entries = []
for i in range(3):
label = tk.Label(root, text=f"Library {i + 1}:")
label.grid(row=i, column=0, padx=5, pady=5)
entry = tk.Entry(root, width=50)
entry.grid(row=i, column=1, padx=5, pady=5)
input_entries.append(entry)
browse_button = tk.Button(root, text="Browse", command=lambda e=entry: browse_files(e))
browse_button.grid(row=i, column=2, padx=5, pady=5)
output_label = tk.Label(root, text="Output file:")
output_label.grid(row=3, column=0, padx=5, pady=5)
output_entry = tk.Entry(root, width=50)
output_entry.grid(row=3, column=1, padx=5, pady=5)
output_browse_button = tk.Button(root, text="Browse", command=lambda: browse_files(output_entry))
output_browse_button.grid(row=3, column=2, padx=5, pady=5)
merge_button = tk.Button(
root, text="Merge Libraries", command=lambda: start_merge(input_entries, output_entry)
)
merge_button.grid(row=4, column=0, columnspan=3, padx=5, pady=10)
root.mainloop()
if __name__ == '__main__':
create_gui()And it works!
"So you were able to get the metadata of all your old songs updated?"
Actually, no, not all of them. This program isn't that sophisticated, and a lot of the problem comes from the esoteric way iTunes is programmed. Turns out I need a way to merge itl files (which is significantly more difficult). However, it did work for some (specifically the ones downloaded from the iTunes Store), so as a proof of concept, it is definitely shocking.
Let me reiterate that I have *absolutely zero* programming experience, beyond a tiny bit of fiddling with BASIC back in 2011 that went absolutely nowhere.
It's entrely possible it pulled this from some source online, but at least on my end, there's zero difference between whether it conjured this by itself or if it intelligently collected all the necessary ingredients because I was able to utilize Python to execute code in about 10 minutes from nothing. No searching, nothing.
Incredible!
That being said, I ought to be more careful. While GPT-4 might not be capable of it, I can imagine a future deeply unaligned model capable of creating intelligent agents whipping up a harmless looking batch of code to accomplish a simple task, but unbeknownst to me, the agents actually slipped the instructions for Stuxnet Mini into it, and executing the program at noon results in the entire internet going down by 3 PM.
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Trying to generate Yulaan and Meki again, with better skills in generative AI:
Marie-Aurore
(Probably the earliest-made of all these)
(Similarly)
Yulaan
(Actually let an online friend use this for a similar character)
(Tails are still wily since it doesn't know what it's doing)
Marie-Aurore
(Probably the earliest-made of all these)
(Similarly)
Yulaan
(Actually let an online friend use this for a similar character)
(Tails are still wily since it doesn't know what it's doing)
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
And remember my friend, future events such as these will affect you in the future
Re: Yuli's Treasure Chest
Contemplating and rambling about the immediate future of artificial intelligence (next year, up to December 2025)
It is blatantly obvious that agents are the next step forward. What does that entail, though?
So generally, it's obvious to me that the current state of LLMs, diffusion models, neural networks in general have many, many flaws. They have crossed a threshold. That much is clear. If I ask GPT-4 or Claude 3.5 Sonnet to, say, write a poem, it can do that, and it could even do that reasonably well that, if I didn't say it was AI-generated, more than 50% of the time, the layman would think it was written by a human. If I generate something in Midjourney v6 or a heavily fine-tuned instance of Stable Diffusion (like NovelAI), I'd say that depending on the prompt and subject matter, more than 50% of the time, it would fool a layman and even a semi-seasoned veteran into thinking it's human-created. Most people would fail the "glance" test. If there's no lead-in to the possibility that something is AI-generated, they wouldn't think about it, especially if:
• There's only one subject, and it's reasonably detailed or close up
• There's other hallmarks of either photography, artistry, or some sort of graphic design (like a signature or logo)
• There are no obvious hallucinations or failures ("look at the hands!")
• The dimensions or resolutions of the image are played with (for example, changing the default 1024x1024 to some less conspicuous ratio either more believable like 1600x1200 or 1294x1037, or downscaled to an extent where it just looks like a secondhand downloading job like 600x400 that further obfuscates any potential qualitative deficiencies)
For some of the more adept AI-slop creators, they've gotten such a good grip on fine-tuning, mixing, prompt engineering, and style-referencing that they can push this up greatly, to the point that I've seen one fellow achieve a near 100% rate of fooling people into thinking his generations are human-created, and another fellow managed to so perfectly match an artstyle reliably that it actually scared some people because "AI-art isn't supposed to look this hand-drawn" (despite the fact I've maintained for quite literally 5+ years now that synthetic media is essentially "digital molecular assembly," not this "collaging and remixing" that a lot of anti-AI types seem to have come to believe it is, which only hinders their own ability to deal with it when they refuse to understand it)
But the larger point here is that even those "perfect passers" know their limits. Midjourney v6 Niji, Stable Diffusion Pony Diffusion XL, NovelAI, they all share some common failure points that they are very careful with, such as the fact that even "advanced prompt understanding" as seen with DALL-E 3 and Midjourney routinely fails to truly understand many prompts, even some deceptively simple prompts, and having more than one main subject in an image can also cause severe problems such as quality collapse.
Similar phenomenon with LLMs, in fact. Their propensity to hallucinate makes them difficult to work with. Not that humans don't "hallucinate" to some extent either when we don't have all the facts, but that's part of many larger problems. One issue being the fact that you can't really write a novel with an LLM right now.
"But Amazon is flooded with cheap AI slop novels! Of course you can do it."
Technically yes, but the vast majority of AI sloppers are essentially glorified laymen "proompters" who think that typing in a prompt and posting the raw output makes them artists, and they lack the finesse of even knowing how to tie things together, edit AI outputs, or realize when something is deficient. There's no quality assurance, basically, and no attempt at even trying. So you'll see these AI-generated novels that overuse "a stark contrast" or "a grim reminder" or "a testament to" or the fan favorite, "tapestry." You'll see "creative" fantasy novels that have 5,000-word-long chapters that seem legit, until you realize every 700 to 1,500 words or so seems to follow the same general structure and format. And then you'll have characters with consistent traits at the start of the novel, but by the end, either half the characters are gone or have forgotten their earlier traits and abilities because the novel ran longer than the context window of the chatbot used to generate it, or the LLM forgot those details because they simply hadn't been brought up before, or it had outright hallucinated new traits and descriptions from the ether and the """writer""" was too incompetent and lazy to adjust or reroll for that.
To say nothing of AI music which I've played with recently. Udio is the best at this, being able to remember back about 2 minutes to maintain some semblance of rondo form, but anything beyond a punk or pop song is going to fall apart.
Or AI video, which is still hypnagogic even at its best.
So where am I going with this?
Well, perhaps distressingly to those with a vested interest in human artistry, I'm going to pull the fire alarm down right now and say "enjoy this twilight era of AI stupidity while it lasts, because it's going away very soon."
Generally, the anti-AI reaction has been simultaneously furious about the use of AI in these fields while also dismissive due to how incompetent AI tends to be: those "perfect passing" AI slop images are very few and very far between, and most AI sloppers are more than happy to post the shiny, soulless, overproduced first output passes or Stable Diffusion coombait with non-Euclidian fingers. Most AI "writers" are content with hitting "Message" and then posting the output raw.
But what happens when AI winds up doing the dirty work too, filling in the gaps between the average proompter and the high-end? That's precisely where techniques like agent swarms, self-play, and concept anchoring come in.
This is precisely why I simultaneously "hide my power level" about AI as well as occasionally try to tell people to not get comfortable with AI's current skill level (and lack thereof). While models can still scale much further, we could see exponentially greater returns just from architectural and methodological shifts.
A tweet I keep referring to showing this off well:
Anthropic also trained Claude 3.5 Sonnet differently than how OpenAI trained GPT-3/4/4o and also extensively used AI-generated data in it. The layman seems to think that "AI-generated data" = low-quality on principle, but when it comes to self-play, so long as that data is pruned and annotated correctly, it could counterintuitively supercharge training.
The biggest deal in the immediate future, however, is going to be agents: especially iterative agent loops (that could effectively solve hallucinations) and agent swarms. Agents are essentially "waking up" the models.
It's going to be scary to interact with agents, even ones powered by GPT-3 level models, because you'll start seeing these agents controlling your computer in ways that seem almost like it was hacked by another human.
Indeed, it's with agents that we'll see the first "true" AI art, as in actual artistic output created by an artificial intelligence, rather than "just" a diffusion-generated image. And that's going to be something to greatly increase the tension with artists, because once agent-based AI art comes online, there's essentially no safeguard, no protection; the agents are manually opening art programs and drawing manually like a human, so any "watermark" would have to require some intense movements that early agents probably won't be capable of. And the matter of "taking human art and reproducing it out" also no longer counts as an argument since these agents would necessarily have to study images and subjects exactly like a human to work. It's a completely different paradigm we're unprepared for, and it's likely going to come at literally any moment.
We'll see self-operating computers become a true thing, and honestly there is no easy analog to this. An agent swarm, if you could afford an instance of it, would be almost psychedelic to watch, like an entire corporation of AIs on your computer, acting upon it, spreading across the internet. The internet itself would seem to awaken even.
Whatever flaws LLMs and chatbots have now as a result of their zero-shot prompting would fall away, once they have the ability to stop and think through their responses and use agents to find solutions and tools to use.
For example, Claude 3.5 Sonnet can create a simple DOOM clone-esque game right now
https://x.com/ammaar/status/18046499038 ... rt_8g&s=19
This level of Sonnet, or perhaps Claude 3.5 Opus, with an agent swarm cache, might very well be able to create the original DOOM from scratch, if given the instructions of how it's supposed to function and nothing else. Agents could automatically resolve the issue of context windows being too small by automatically restoring necessary information before catastrophic forgetting begins, provided it's still a problem at all.
Similarly with, say, generating novels— if you can anchor concepts to tokens and use agents to provide an autonomous workflow, you could very well prompt an agentic GPT-5 to "write a novel" and it does it one-shot.
Now, for artistry, that poses some problems, but the real area this is meant to be felt is in conversational ability. This level of AI crushes the Turing Test, and likely most other intelligence tests as well (like the Coffee Test, if it's embodied).
And I don't see it taking longer than the end of this year to see this be extensively shown off, mind you. That's what's terrifying to a great extent, that this transformative level of AI is going to be so visible so soon, and we are totally unprepared, still reacting to and getting angry at AIs that are essentially from two years ago (like Sora; leaks already prove that OpenAI had Sora in March of 2023, and it was likely trained well before then).
Essentially, we've yet to truly see a "2024-class AI model" in full.
It is blatantly obvious that agents are the next step forward. What does that entail, though?
So generally, it's obvious to me that the current state of LLMs, diffusion models, neural networks in general have many, many flaws. They have crossed a threshold. That much is clear. If I ask GPT-4 or Claude 3.5 Sonnet to, say, write a poem, it can do that, and it could even do that reasonably well that, if I didn't say it was AI-generated, more than 50% of the time, the layman would think it was written by a human. If I generate something in Midjourney v6 or a heavily fine-tuned instance of Stable Diffusion (like NovelAI), I'd say that depending on the prompt and subject matter, more than 50% of the time, it would fool a layman and even a semi-seasoned veteran into thinking it's human-created. Most people would fail the "glance" test. If there's no lead-in to the possibility that something is AI-generated, they wouldn't think about it, especially if:
• There's only one subject, and it's reasonably detailed or close up
• There's other hallmarks of either photography, artistry, or some sort of graphic design (like a signature or logo)
• There are no obvious hallucinations or failures ("look at the hands!")
• The dimensions or resolutions of the image are played with (for example, changing the default 1024x1024 to some less conspicuous ratio either more believable like 1600x1200 or 1294x1037, or downscaled to an extent where it just looks like a secondhand downloading job like 600x400 that further obfuscates any potential qualitative deficiencies)
For some of the more adept AI-slop creators, they've gotten such a good grip on fine-tuning, mixing, prompt engineering, and style-referencing that they can push this up greatly, to the point that I've seen one fellow achieve a near 100% rate of fooling people into thinking his generations are human-created, and another fellow managed to so perfectly match an artstyle reliably that it actually scared some people because "AI-art isn't supposed to look this hand-drawn" (despite the fact I've maintained for quite literally 5+ years now that synthetic media is essentially "digital molecular assembly," not this "collaging and remixing" that a lot of anti-AI types seem to have come to believe it is, which only hinders their own ability to deal with it when they refuse to understand it)
But the larger point here is that even those "perfect passers" know their limits. Midjourney v6 Niji, Stable Diffusion Pony Diffusion XL, NovelAI, they all share some common failure points that they are very careful with, such as the fact that even "advanced prompt understanding" as seen with DALL-E 3 and Midjourney routinely fails to truly understand many prompts, even some deceptively simple prompts, and having more than one main subject in an image can also cause severe problems such as quality collapse.
Similar phenomenon with LLMs, in fact. Their propensity to hallucinate makes them difficult to work with. Not that humans don't "hallucinate" to some extent either when we don't have all the facts, but that's part of many larger problems. One issue being the fact that you can't really write a novel with an LLM right now.
"But Amazon is flooded with cheap AI slop novels! Of course you can do it."
Technically yes, but the vast majority of AI sloppers are essentially glorified laymen "proompters" who think that typing in a prompt and posting the raw output makes them artists, and they lack the finesse of even knowing how to tie things together, edit AI outputs, or realize when something is deficient. There's no quality assurance, basically, and no attempt at even trying. So you'll see these AI-generated novels that overuse "a stark contrast" or "a grim reminder" or "a testament to" or the fan favorite, "tapestry." You'll see "creative" fantasy novels that have 5,000-word-long chapters that seem legit, until you realize every 700 to 1,500 words or so seems to follow the same general structure and format. And then you'll have characters with consistent traits at the start of the novel, but by the end, either half the characters are gone or have forgotten their earlier traits and abilities because the novel ran longer than the context window of the chatbot used to generate it, or the LLM forgot those details because they simply hadn't been brought up before, or it had outright hallucinated new traits and descriptions from the ether and the """writer""" was too incompetent and lazy to adjust or reroll for that.
To say nothing of AI music which I've played with recently. Udio is the best at this, being able to remember back about 2 minutes to maintain some semblance of rondo form, but anything beyond a punk or pop song is going to fall apart.
Or AI video, which is still hypnagogic even at its best.
So where am I going with this?
Well, perhaps distressingly to those with a vested interest in human artistry, I'm going to pull the fire alarm down right now and say "enjoy this twilight era of AI stupidity while it lasts, because it's going away very soon."
Generally, the anti-AI reaction has been simultaneously furious about the use of AI in these fields while also dismissive due to how incompetent AI tends to be: those "perfect passing" AI slop images are very few and very far between, and most AI sloppers are more than happy to post the shiny, soulless, overproduced first output passes or Stable Diffusion coombait with non-Euclidian fingers. Most AI "writers" are content with hitting "Message" and then posting the output raw.
But what happens when AI winds up doing the dirty work too, filling in the gaps between the average proompter and the high-end? That's precisely where techniques like agent swarms, self-play, and concept anchoring come in.
This is precisely why I simultaneously "hide my power level" about AI as well as occasionally try to tell people to not get comfortable with AI's current skill level (and lack thereof). While models can still scale much further, we could see exponentially greater returns just from architectural and methodological shifts.
A tweet I keep referring to showing this off well:
Anthropic also trained Claude 3.5 Sonnet differently than how OpenAI trained GPT-3/4/4o and also extensively used AI-generated data in it. The layman seems to think that "AI-generated data" = low-quality on principle, but when it comes to self-play, so long as that data is pruned and annotated correctly, it could counterintuitively supercharge training.
The biggest deal in the immediate future, however, is going to be agents: especially iterative agent loops (that could effectively solve hallucinations) and agent swarms. Agents are essentially "waking up" the models.
It's going to be scary to interact with agents, even ones powered by GPT-3 level models, because you'll start seeing these agents controlling your computer in ways that seem almost like it was hacked by another human.
Indeed, it's with agents that we'll see the first "true" AI art, as in actual artistic output created by an artificial intelligence, rather than "just" a diffusion-generated image. And that's going to be something to greatly increase the tension with artists, because once agent-based AI art comes online, there's essentially no safeguard, no protection; the agents are manually opening art programs and drawing manually like a human, so any "watermark" would have to require some intense movements that early agents probably won't be capable of. And the matter of "taking human art and reproducing it out" also no longer counts as an argument since these agents would necessarily have to study images and subjects exactly like a human to work. It's a completely different paradigm we're unprepared for, and it's likely going to come at literally any moment.
We'll see self-operating computers become a true thing, and honestly there is no easy analog to this. An agent swarm, if you could afford an instance of it, would be almost psychedelic to watch, like an entire corporation of AIs on your computer, acting upon it, spreading across the internet. The internet itself would seem to awaken even.
Whatever flaws LLMs and chatbots have now as a result of their zero-shot prompting would fall away, once they have the ability to stop and think through their responses and use agents to find solutions and tools to use.
For example, Claude 3.5 Sonnet can create a simple DOOM clone-esque game right now
https://x.com/ammaar/status/18046499038 ... rt_8g&s=19
This level of Sonnet, or perhaps Claude 3.5 Opus, with an agent swarm cache, might very well be able to create the original DOOM from scratch, if given the instructions of how it's supposed to function and nothing else. Agents could automatically resolve the issue of context windows being too small by automatically restoring necessary information before catastrophic forgetting begins, provided it's still a problem at all.
Similarly with, say, generating novels— if you can anchor concepts to tokens and use agents to provide an autonomous workflow, you could very well prompt an agentic GPT-5 to "write a novel" and it does it one-shot.
Now, for artistry, that poses some problems, but the real area this is meant to be felt is in conversational ability. This level of AI crushes the Turing Test, and likely most other intelligence tests as well (like the Coffee Test, if it's embodied).
And I don't see it taking longer than the end of this year to see this be extensively shown off, mind you. That's what's terrifying to a great extent, that this transformative level of AI is going to be so visible so soon, and we are totally unprepared, still reacting to and getting angry at AIs that are essentially from two years ago (like Sora; leaks already prove that OpenAI had Sora in March of 2023, and it was likely trained well before then).
Essentially, we've yet to truly see a "2024-class AI model" in full.
And remember my friend, future events such as these will affect you in the future