This reminds me of when I was playing with play-doh when I was young. You start out with many different colors, and somehow always end up with a big brown ball.
Apparently in the original Matrix movies storyline, the reason why the machines needed to keep those troublesome humans around was not as an energy source (“batteries”) but as a source of creativity. But the writers thought that this idea was too complex so they substituted the battery idea instead.
Too bad, that's a much better idea. Although with all this talk about creativity, and AI putting an end to creativity and whatnot, I've never seen anyone mention that creating doesn't only mean creating good stuff, it also means creating crap. It seems to me that the people behind all these AI programs usually want them to create good stuff and not crap, so to me it's no wonder that they tend to end up converging (to creating good stuff, I'd hope) if they're trained on their own creations. Even if the original idea for The Matrix was better, I'd find it hard to believe that after a while the machines would still need humans at all, after they had learned enough about how to have ideas, good and crap, from us humans (and obviously, over time their thought processes would converge in the direction of having better and better ideas). EDIT: forgot a comma
I agree that the creativity idea would have been much better than the battery one, but ... all our knowledge about physics comes from inside the matrix, so maybe they just fabricated a different "physics engine" for it, so anyone escaping would be sufficiently confused to be easily captured?
Same in the Terminator universe. Well, almost. Skynet keeps useful people around to develop terminators and so on. In the early stages it actually preserves workers until they have built automated factories.
I remember as a kid I used to record my voice and play it back on my speakers, which I then proceeded to record once more. By repeating it, I could hear how it slowly degraded until it was nothing more than a weird, synth-like sound.
Wow, I did exactly that myself! My brother and I also loved to record our voice into the computer (ah Soundforge, I miss you) and then reverse the audio. We would practice speaking the reversed sounds and then record ourselves again speaking these reversed sounds. We would then reverse the reverse sounds we just spoke and hear ourselves speaking the reversed version of the reversed sound we spoke based on the original sound we reversed. Ah great times. I'm a software engineer now working specifically in the field of audio analysis.
the YT algo is a reflection of your searches, my recommendation feed is always changing as I type in new searches for different content. but if instead you only live in the recommendations clicking away and never do your own new searches... sure it will get stale and sammey over time.
As someone who used to play with photocopiers as a kid... A copy of a copy of a copy is always much worse and weirder than you might think. Small flaws amplify until you all you get is a smudged blur.
That's if you're using so-called "AI" exclusively. Using it sporadically, as merely another software tool in your creative arsenal, will give you the edge on those who flatly refuse to use it on principle. Anyway, there's financial incentives for big tech companies to ensure their AI is more accurate, faster, easier to access etc. than the competition. They're not just going to press the red button and let their AI run loose.. It's all still a service that needs 24/7 support by HUMANS behind the scenes.
@@alieninmybeverage I think it already happens on Instagram. People are using filters aimed at making them look like AI generated photos (smooth and symmetrical faces).
My biggest problem with AI is that it needs to get the information from somewhere, and sometimes these sources can be slightly dodgy. I did an experiment where I asked ChatGPT about some very narrow subjects: The Danish organplayer Peter Erling, the trio Klyderne, and the artist Jørgen Fonemy. These are subjects that I have some knowledge about and actually have written Wikipedia articles about. I could see that most of the answer that I got from ChatGPT was based on the exact Wikipedia articles that I wrote! I have tried to write the truth in those articles, but if I didn't care if things were correct - or worse; if I deliberately wanted to mislead people, then AI would base the answers on wrong data, if there wasn't multiple sources available. The problem I see with AI is that we trust it too much. Already now there are people who believe that it is an omniscient trustworthy source of all answers and that it will always be more correct than human knowledge or just knowledge that we have googled or looked up in an oldfashioned book.
Thank you for posting that. I suspected something like that is true. I like to test LLMs with riddles and verbal puzzles. The first impression I got was that the best of the LLMs were brilliant, as they could solve some of the toughest puzzles correctly, puzzles famous for being difficult even for the sharpest humans, and they even had good answers to pointed follow up questions. Then I tried novel puzzles based on famous ones, but with the questions reworded (by me) in subtle ways which changed the correct answer, and then the LLMs usually defaulted to "pattern matching" and giving me answers which were correct answers to the original "pattern" puzzles, but wrong answers to the novel reworded versions of the puzzles they were answering at the moment. They are good at answering known questions with known answers which are already published or posted on the Internet, but have trouble with novel variations which have never been published before. They are not figuring out the answer, but giving their best guess based on what they've already seen in their dataset. OTOH, the best LLMs keep getting better at adapting to novel variations, month to month, so it's wrong to generalize based on results from more than a few months ago. Their abilities are progressing rapidly at the moment.
It's also bad at interpreting articles. It told me something related to tech that I knew to be false and it provided links to articles "proving" it was correct. I read the articles and ChatGPT misinterpreted the text of every single cited source. Complete garbage as a research tool.
@@sportsentertained which version did you use? I've read that they keep dumbing down ChatGPT to save on backend resources. The paid version is better than the free version, but still not as good as it was at the beginning (before it got flooded with new users)
In other words, it's scraping data and possibly reorganising it slightly or cutting and pasting and not attributing where the data is from. Very sneaky.
A friend in the UK is a graphic designer; he says that over the past few months, more and more clients have been saying 'NO!' to AI-generated artwork - "it's too samey". They'd rather pay more for something original. Trouble is, AI has pushed down the rates; so while designers and artists are noting an uptick in requests for proposals, the money is much worse.
Yes, AI images looks horrible. To many details that make no sense, glittering stuff, imposing backgrounds, flaming skys, opulent clothing ... Often I do not wan't to read the text, it just feels like candy all the day.
AI will teach us what truly matters ... Human connection and true emotions is what we should care about. Spending time with your loved ones, (com-)passion etc.
It just seems like there is so much competition in anything creative that whoever is paying can have people jump through whatever hoops they want. And why wouldn't you ask for original art instead of AI generated art if you have the leverage.
ai cannot get what a roblox game thumbnail looks like. it can generate one but its not convincing at all. even the other styles of roblox thumbnail dont fit what ai generates.
You reminded me of Pandora, the music recommendator that provided you with music according to the 👍 and 👎 that you gave to the songs proposed. No matter if you started with Black Sabbath, Chopin or Yunchen Lahmo. Eventually, after a couple dozen songs, you always ended up in a Coldplay loop.
This sounds like user error; I've had a sub to Pandora basically since it was still just the Music Genome Project and I've never heard Coldplay on my stations. Coldplay: Not even once.
Damn this reminds me of youtube. I've had to start making new accounts all the time because the algorithm is quickly devolving into recommending quite literally the same videos I've already watched over and over and over and i can;t find anything new or exciting. Music is by far the worse, I decide I want to go outside my usual tatse and listen to nostalgic dirty pleasure pop from my youth, youtube wants me to listen to my usual stuff again... It's absolutely gotten worse than it used to be without a doubt
This is exactly what I was telling people the other day. Our greatest danger with AI isn't that it'll take over but that at the moment we begin relying on it most, the more it will collapse because it's going to end up cannibalizing itself.
They are not intelligent they are bullshit engines with a filter applied to remove anything that is to obviously ponging. Try this 3 Captains argue over whose sailors are the most courageous, German, Fench and British. They each order a sailor to jump from the ship's mast into the sea swim under the keel and climb back aboard. The German sailor responds Jawohl Kaptain and does it. The French sailor cries Oui oui mon Capitaine and also does it. The RN matelot looks at the captain and says " Naff off .... sir". The RN captain turns to the others and says that is courage. Now who was intelligent there?
In some ways, I find google searches much better than they used to be. But the damn company has been manipulating search in ever more increasing ways that has made the service dubious
It's true, when you've worked enough with ChatGPT you can immediately recognize a ChatGPT text. It just always has a certain vibe that makes it distinguishable from human text.
I fear it gets so good that we can't even pick up on those little flaws and quirks any more especially for videos. When those Sora videos were released the only one I could tell was AI was the woman walking on the street (her hand and face had weird details). I imagine the next ones will fool me better.
The key is for a human to utilize and modify AI generated content, not just copy/paste. Also realizing that some ai is better than others at specific tasks (gemini for emails, chatgpt for code, etc)
You make a good point. Already I can usually pick the ‘style’ of AI generated images. They have a certain ‘style’ because they are in a sense too perfect, too smooth, too balanced. It is not something one could define in some cases, but the human brain is good at recognising patterns.
They often have this fractal like composiotion. Many generated images with people have a bit of paintairly feeling because the most of database was artists pages like Artstation. You can easily see Artgerm's style in many of pritty girls pictures. And why all the images are young woman because this is young woman is most popular subject on these pages when it comes to people. In photography young woman is I believe also one of the most popular subject in human cathegory.
Garbage In/Garbage Out. I've been saying the same thing about both AI and Analytics for the past decade and a half. People only want to look at processes, algorithms, ease of use, speediness, raw power, TCO, design and pretty UI with both AI and Analytics. You rarely hear people talk about things like bias, data integrity and context. Those three things only come into conversation when AI and Analytics produce horribly incorrect results.
This actually is not that suprising, when you think about it: these AIs are basically using huge amounts of data to approximate averages of various things, and with more iterations they extract more and more core features until they just have the same set of features they are using all the time. It's like taking data scores and continually averaging them until you are left with one value.
Yeah, I think fundamentally the technology of these systems is stuck between a rock and a hard place. The reason humans can ACTUALLY learn something complex like say aerodynamics is because we can discriminate the quality and type of information extremely well, we can learn that subject from a small amount of well-curated information from school to college manuals. But modern AI only works with hilariously broad datasets that contain literally everything, otherwise it loses that smidgen of general intelligence that makes it worthwhile to begin with. So AI is stuck in an unwinnable conundrum - to function at all it needs to learn from everything at once, but to actually have good knowledge it would need to focus on that small amount of material that is actually good.
My experiance with ChatGPT shows it to be a regurgitator, the test questions were in an area of X-ray physics that I know well and it spewed out all the usual stuff with no insight, no deep understanding, no creativity, nothing that would indicate any form of curiosity.
@@lukeskyvader3217 I like how you said it like it actually happened but there isn't any evidence beyond some idiot repeating marketing material that couldn't be proven as lying even tho everyone knows they are making shit up.
garbage in & garbage out. Put in the extremely usual stuff, expecting something novel? GPT is basically bound to what you ask, mirroring the original input.
There's a third option: there may be soon a deliberate attempt to poison the content to make it unreadable for AI. There are already tools out there that scramble images just enough to make them confusing for AI to use as a training set.
@@phattjohnson if there's enough poison then statistically it will reach the sample set of plenty AI systems and lock itself into garbage. If the poison is ignored then that's a smaller sample space AI has access to and become boring and derivative.
The only way I know is by setting your meta data to your images to be erroneous. How can you scramble an image and still have it viewable to humans? Doesn't the AI access it the same way we would?
Adding random variation probably isn't as easy as it may sound because the randomness still has to follow certain rules. For example, no one is going to believe that elephant with two trunks.
I think you are mistaking randomness for imperfections. She is not saying images need to have faults on them. Diversity here means for example some elephants are young, some adult. Some are eating some are sleeping some are drinking. Some are photographed at night, some are walking on grass, some on rock, etc. If you see the AI samples provided, they all look exactly the same. Zero diversifying.
@@a_kazakis but that's his point... how will the AI know what's appropriate and what's not? how can it know to add diversity in lighting and background, and not in the number of trunks or skin color?
AI models usually have some way of computing how likely they think different outputs are. A model that turns a written prompt into an image has some notion of how "close" an image is to the prompt. Instead of taking the closest image to the prompt, you might instead take another nearby image determined by some random number. Unfortunately, there isn't a good rule for defining the precise details of the randomization scheme. There's a lot of ad-hoc methods that work well for one group of prompts but fail for others.
@@a_kazakis It makes no sense. There are already millions of photos of elephants from different angles carrying out different activities in different scenarios. If ALL the photos available on the internet (copyright or not) are not enough for the model to be able to generate convincing photos, the problem is not having more diversity in the dataset
@@a_kazakis It reminds me of a children's art class. One kid will draw a house, car and tree; And a dozen other kids will copy them. There may be variations like an apple tree or a dog but they're all relatively alike.
AI is geared to come up with a solution using the smallest amount of computations possible. It means that ignoring diversity and details is part of its basic make-up. It can't capture interesting quirks and details and spurt them out once in a while, in some outputs but not others. It also means that if it found one way to get to an acceptable solution, it will only try to get there quicker the next time around. If you've experienced getting stuck in a rut with ChatGPT, that's why.
The one thing people should know about machine learning is: a machine learning trained model will only be good as its training data. It's just learning (in theory) the pattern behind the data leading to a host of problems. The main issue is that it doesn't actually reason about the data. Let's say I train a model where I have several examples where I have pi as 3.14 and then one where it's 4. The model doesn't reason "you know.... this one example seems to be wrong" but rather it updates the model to make it slightly more likely it will give the wrong answer. So how do you prevent models training on information generated by another machine learning model? The current approach is to stick to information before generative AI become dominate but most of that information (for better or worse) is probably considered or part of the training dataset. The main problem is that there's a popular opinion in machine learning (and sadly AI) that, as an AI researcher, I have had to deal with. This opinion is the key to all AI problems is that we just need to use larger models, with more training data, and train it in the "correct" way. "Look how far LLMs have come. Just imagine how much better they will be in a couple years". But you run into the 90-10 principle: 10% of the effort for 90% of the results and vice versa. It's why self-driving cars are taking a long time: there is nearly an infinite extremely rare cases that the car needs to make the right decision in. As such, it should be expected for the current LLMs to plateau performance wise unless new smarter methods are found.
Thank you for your insight, I think I agree with this. In the case of LLMs, they clearly have a use case already that will not go away, but I don't think they can deliver on the promises being made. I do not see how to make them be reliable enough to work in most business situations. I feel that many companies are looking for a way to implement them, and almost making their engineers find a way to make them useful, even it it makes no sense. The scaling already seems unsustainable, and while the "emergent" behaviors are very cool, nobody really understands how they relate to scaling (aka its not a defined ratio of x amount of compute/data for x more emergent behaviors)
It's not even machine 'learning'. It's 'just' scripted data consolidation, procedural compression and re-generation, and some other mumbo-jumbo that honestly has all been around since the conception of PCs. Just now we've got several modules all running simultaneously in one disjointed codeblock.
I'll do you one better. We will never solve this issue. It's a fundamental impossibility. We will never have self-driving cars. There is no exponential curve, no singularity. Forget it. We are very close to the best AIs will ever be
@@octavioavila6548 You base this on what exactly? Claiming AGI will never happen, and self driving will never happen is the same as the people who think we will have AGI in 2 years because of the hype. Nobody knows the limits or timeline, but I don't see why it would be impossible.
"Hold up. Something's wrong here. Not sure what it is but I feel like we should take a step back and go through it again" Said no AI ever, past, present or probably future.
As an artist, this is a known issue in HUMANS Thats why the art solution is to look at the "old masters" as mentors before learning how to draw from more modern artists
Call me a snob, but I'm even more pessimistic about the decline of human taste than I am about the technical shortcomings of AI, which is a problem that reliance on AI for the production of images, text, music, etc. will likely exacerbate, but certainly didn't create. From my point of view, even before it started to become obvious how bad and samey AI art really was, it was already quite obvious how the stuff people wanted AI to create was junk in the first place: pop culture fanart and stuff that mimicked stereotypical pop culture tropes, done in a glossy, quasi-realistic style. The only "interesting" AI art occurred early, when AIs tended to fail at their task and produced bizzarre unintentional surrealism. There was a famous image of a collection of completely unrecognizable objects that made the rounds a few years ago and which was (incorrectly) described as an attempt at reproducing the visual experience of someone having a stroke (whereas it was just AI image generation still being too primitive to successfully reproduce its models): that might well have been the aesthetical peak of AI art.
Or adopt creative mentality. Next time you create, take piece of paper. Crumple it. And use it as stippling too. For following projects, paint paper with some thin color, let it dry. Put layer of transparent soap or similar material. Dry again. Layer of another color, followed by different color. Repeat few more times. Final layer should be black or white paint. Then use scratching tool to "draw" with different pressure. Even lid of some jar can be used as artistic tool for painting. Or plastic body of old pen as spraying tool. Same applies to sculpting, dancing, music, ... Just let your mind free itself from cage of mundane existence.
My experience with chat GPT is that you can ask it 2 or 3 questions, get it to contradict itself, and when you point out the contradiction, it starts to ask if you are angry, and/or says IT'S taking a break from YOU to let you relax...😊😮😂
Glad you started this conversation. There is also the theft component of generative AI. A TH-camr like yourself will get automatically copyright struck for using 4 seconds of a clip in a 20 minute original video. Yet these generative AI companies can use entire social media platforms with content painstakingly created by individuals across decades to create their data sets. This is peak hypocrisy in which, as per usual, corporate "big money" is protected while the individual is left with no means of defending their content. Generative AI is 100% theft in my opinion.
No art is truly "original". Artists are inspired by previous artists who are inspired by their surroundings and modify reality slightly based on their mental conception of what they want to highlight. Art is derivatised inherently by nature of human learning. Generative AI follows similar processes. It doesn't 'copy and paste' as people have claimed. It has a distinct concept, albeit less defined than a human, of what it is asked to portray. AI art is inspired by and not directly copying actual works. If we start copyright striking AI, it should follow that we strike virtually every other art piece.
@@darkushippotoxotai9536 You keep believing that. It's a tired and completely flawed argument. First off, there is a human TIME factor involved. A human artist must first put in the hundreds of hours of work to accomplish some level of mastery over their craft before they can even THINK about mimicking another artist's style. That process produces mutual respect. This entire component is lost with "AI" slop. There is so much more at play here but it's just not worth getting into in a comment section on TH-cam for someone who has no actual desire to objectively weigh new perspectives. You want the AI future. Well, it's coming. Nothing will stop it. The tech overlords are investing trillions so you'll get your wish. I hope it is everything you want it to be.
@@tygorton So, simply requiring more time and being less efficient and sometimes even of a lower quality is better because a human made it ? Sidenote, I didn't really say mimicry, but rather drawing inspiration. Sure, AI can do that as well, but I was moreso talking about inspiration or to put it simply, pointers or definitions or Illustrations of art. Humans do not make an unprecendented or completely unique art. It's subconciously drawing on other works and surroundings of the artist. Almost Same as an AI, just very inefficient. As for intent, It's a human writing a prompt. An AI doesn't simply mash things into a image. How many artists you know of have drawn a celtic man chasing a dog through a world made up of needles ?
@@darkushippotoxotai9536 Enjoy the "efficiency". Like I said, your AI future is coming. It will be a world of emptiness filled with people who lack wisdom; the evidence of this is already permeating every aspect of our culture and it hasn't even started yet. Enjoy.
1) AI is trained on data from the Internet. 2) AI outputs data to the Internet. 3) Goto 1 ... haven't anybody acquainted themselves with the topic of "inbreeding" ?
Solve that problem with using an A.I. classifier for detecting whether data is synthetic or not. Diversity isn't going down its just laziness in coming to creating datasets.
There are solutions to this though. 1) AI scientists run AI through quality data. 2) AI scientists run AI through a comparison between quality data and its outputs to provide corrective comparison. 3) Give AI real vision (robotic eyes) so it can observe real life examples from the real world. 4. Humans keep involvement in the process of determining what gets posted to the internet. If AI produces garbage it's less likely to be selected. If it produces something accurate, it's more likely to be accepted. Survival of the fittest response.
This is exactly what I've been thinking since all of this exploded into popular awareness. It's like a giant ouroboros eating it's own tail. I'm glad to see that people are talking about this. Editing to add: Will you critics please lighten up? I'm not anti-AI. I'm just agreeing with Sabine that this is a potential problem that should be studied. All new technologies have potential problems that need to be studied and understood. Pointing this out does not make me some kind of neo-luddite.
People have been eating their own tales since there were people, I'm not sure why AI is expected to be different. Most people aren't that creative, but a few are; most AIs won't be creative, but a few will. Same old, same old.
On appearance the science on that will show the AI photo shown are popular world wide. But of course it becomes too much of the same causing desire for diversity. Important to point out there actually a science in area of attraction both in humans and other species and we need to start shutting down those with non scientific type opinions especially the it just one culture imposing its values and the effort to make all appearance beautiful which is impossible our brains demand an ugly. Example make overweight attractive healthy becomes ugly. Better to push the traditional view of attraction only skin deep and accept your appearance state great to bad as unimportant to one’s value as a human being. And of course set beauty for the weights that actually healthy and live longer. Note some studies show a tad underweight might live longest.
Not exactly in this case but I certainly get your point, in that the result is undesirable if one values health or positivity. To your point, I think a better description of Sabine's observation about the failing of AI would be "garbage in, garbage grows". Perhaps the creators should take this to heed and develop systems that augment the process to manage the generated information in a way that aligns with what is in humanity's best interest. Less garbage is in everyone's best interest.
This would make an interesting experiment. Begin with a discrete distribution of objects which is peaked, like a Gaussian.Sample the entire distribution gauging similarity as a dot product. Exclude one most-dissimilar object each time the entire distribution is sampled. ultimately you should sharpen the distribution until you get a spike at the most probable /identical objects.
Sabine, as a professional in the application of machine learning in medicine I would like to thank you for making this video! It’s understandable and it reaches a lot of people! There is the AI hype (which people should not believe because it comes from executives and rookies) and there is the machine learning reality that veterans understand. This technology will be useful in automating some drudgery and common simple tasks…. It’s dogshit at doing anything truly valuable. What’s most worrying is the very real threat, without laws, that this nonsense will create such a firehouse of bullshit that we can’t get through our email, find what we need on the web, tell the difference between fact and fiction, and generally think for ourselves!
2) More randomnes in AI output might do away with the problem of repetitive AI output, but it might increase the mistakes. Instead of elephants with big heads or two heads, we might get elephants with two big heads.
I think a big part of the convergence is that people often are attracted to certain tropes and conventions when it comes to what they like, so AI produced images are actively being 'pruned and purified' by our preference of our existing cultural paradigms. What I think is really interesting is the feedback, where people's tastes of which conventions they like are in turn influenced by AI art.
Yep. It is easy to see AI images tend to be standarized. What is not that easy to see is if that is really a problem. People like standards. Just look how actors and actresses look like.
You couldn't be more wrong. The woke crap is purposefully programmed into it. Same for the censorship. Has nothing to do with preference and cultural paradigms.
This is such nonesense. In terms of LLM's it is the desired outcome because you predict the most likely next token. You want the best answer, not any answer as default. and yes all models have already a "temperature" parameter, which regulated the unpredictability and range of the possible tokens which can be chosen. For images the same. The example is really bad in the paper they use the same prompt, don't inject random noise. Yes Midjourney as a consumer product has the issue but the underlying models don't have the issues. You can have as much randomness, creativity and variance as you want. This video displays the increase accuracy, which they aim for as issue, which it is not. temperature=0.6 or higher and you get your creative storytelling back.
Or perhaps we just fell deep into the trap of belief system that as a society and civilisation we have already learnt everything there ever was about ourseleves and our human consciousness. @@squamish4244
@@squamish4244No, its not smart at all in reality. It’s not even actual AI. It is an algorithmic system. Give it more data and it will get sharper. Thats how it’s programmed. It has no ability to think or comprehend what its outputting. A true AI that simulates the human mind in digital means would likely use algorithms as part of its system, but not as the entire basis. Todays “AI” is nothing but a generation system. And it’s not able to think and uniquely create anything truly new, based on the limitlessness of the human mind. It can mash and mutate things due to its flaws of understanding, but it is actually not truly and willingly making something new. It copies and makes mistakes which could be claimed to be creativity, which these algorithms have no actual ability to harness.
@@man.horrorYes, the expert swoops in. Whatever. It's not that AI is that smart, it's that humans are not as smart as we thought we were. I'll take Max Tegmark's books over your two paragraphs here, thank you very much. Copium over 9000.
As an AI scientist, we've been talking about this for years. Once the AI starts eating its own tail it will quickly optimize to a singularity of stupidity in its own echo chamber. The only way for AI to continue to work is to automatically label all AI output and ignore it for training. Or to manually post label it by humans. Humans are necessary for AI success in any case. It would be interesting for you to discuss both the labeling servant culture and its injustices as well as the impossibility of AGI if AI depends on human labeling.
@@johnatyoutube Well, it's basically optimizing the short term stock-price at the expense of everything else so the CEO cash out. Long-term viability, product quality, worker productivity, accurate book-keeping and finances, even the best interests of shareholders and real profits and revenue don't matter if sacrificing them can result in a stock payout before the consequences hit.
As a layman I disagree. You’re right if you don’t think outside the box, but we can use AI to sample evolutionary algorithms to generate networks for more AI models. This space is practically limitless.
They even translate the phrase in German, where it make even less sense. Ich hoffe meine Nachricht erreicht Sie gut... Lastly erreicht Sie bei bester Gesundheit. Both are phrases not used in German.
As these AI errors flood the net, will they become more and more of the training data for other AI’s? Until images get increasingly mutated and standard emails all start with, “I hope this emu fondles your willy.”
The popular image generation models prior to Stable diffusion were GANs (generative adversarial networks). The way they worked was to have two different networks - one trained to generate images, and the other trained to classify images as real or fake. This forced the generator to learn to avoid the most identifiable characteristics and to generate a diverse set of images. Stable diffusion was more effective and scaleable for higher resolution images, keeping the whole image globally coherent. But it's likely that reviving some adversarial techniques could help with the diversity issue.
Actually one of the biggest issues with GANs that they were very prone to "Mode Collapse". During mode collapse rather than producing a diverse set of images, the adversarial network would hone in on specific features which were not recognized by the discriminator network. The result: a lower diversity in images which get produced. The reason why diffusion took off in the first place is that due to noise being used as a base, the diversity was higher, as the initial noise served as a "random seed" for the generation in a sense. Mode collapse can be avoided, but takes a lot more effort to avoid, and can lead to problems in many architectures. (Note, im not a researcher.) This is mostly from scant reading I've done here and there.
@@Coach-Solar_Hound you're absolutely right but I'd like to add another point here, it's not just about model collapsing, the reason why GANs end up losing degrees of freedom is because of overfitting. The ultimate trick to win the discriminator is to draw exact copies of the dataset and that's why you need to save "backups" and move back in time of trainning when you see important details are being left out. Now, regarding diffusion vs. GANs that's a more broader discussion: GANs theoretically should excel in image generation but the investment towards diffusion (especially prompt to image) is way higher so while GANs seem to be lacking, they should actually be a better solution overall. What you said about taking "random seed" is also true for GANs, the generator will always take a random number and try to draw what it knows about the dataset from there. There's a really interesting video explaining all the details in computerphile channel: th-cam.com/video/Sw9r8CL98N0/w-d-xo.html
Image controlling for GANs is still an active area of research, what we do today to influence latent space results is to move specific directions in latent space. To know where to move you can use dimensionality reduction techniques to find specific vectors controlling image relevant attributes (check the paper of GANSpace). Another option is to do img2img transfering style or mixing with prompting information
@@Coach-Solar_Hound Yes, that's true, although there were a lot of developments going on to fix that. The biggest problem was either the generator or the discriminator getting too far ahead of the other, and the whole thing getting stuck. So the rate of learning of the two parts had to be balanced. There was another issue where the set of produced images was not representative of the training data because the generator favored generating "easy" images. For instance, if it was generating faces, it would avoid producing details like glasses or beards, or prefer to generate less angular faces (i.e. the output would overrepresent women). There are lots of types of regularization to be done, and techniques to help with those things. Adversarial learning, generally, is a really useful technique. So I think it's time to bring it back to diffusion. (I have done work on GANs personally, although it's been a few years).
Look how people are more and more hating CGI in movies to the point that some movies refuse to do any. If you have ever read anything written by AI you know it has the ability to make the most exciting subjects boring.
@@maphezdlin All of those movies used CGI. All of them. Many of the stunt scenes are mostly real footage, sure, but a lot of them are edited beyond recognition. Oppenheimer only lists 49 vfx artists on IMDB, but that's mostly because 80% of them weren't credited. Skyfall lists 578 vfx artists. Inception had 295. Mission Impossible: Ghost Protocol had 347. Mad Max: Fury Road had a whopping 742. The Dark Knight had 468. Casino Royale had only 161, which is in fact impressively low, but still not 0. 1917 had 422. Top Gun: Maverick had 455. For reference, Avatar: The Way of Water (2022), a movie we can hopefully all agree had immense amounts of CGI, credits 1113 vfx artists. The Hobbit: The Desolation of Smaugh (2013) had 915. Most of the movies you mentioned had close to if not more than half of that. What did all these people do if there's no CGI?
@@Felixr2, K VFX and CGI are different. But you are right the links that I saw that said NO CGI lied. They should have said minimized CGI. Thanks for catching it.
Great presentation, Sabine. I have always maintained that AI is like students cribbing exam answers. One student just has to copy one thing wrong, once. From then on it is a done disaster. When scientists hypothesised robots making copies of themselves - they never saw this far into the mess.
You are 100% correct and it's already happening. I noticed it first when I was looking up a certain niche question that had a bunch of AI generated garbage in the search results, that somehow kept repeating a nonsensical "fact". I pinned it down to a single forum post that was made 10 years ago where somebody made a typo or something that made no sense, but this post was ingested by the machine learning dataset and that dataset was being used to generate a bunch of blogposts/websites, because of the way LLMs write (long dense sentences with very specific subjects) shot up high in search engine rankings. So now there's 20+ different sites all parroting this garbage information, which was then used in other datasets and ingested by most LLMs now, if I ask that specific question to any LLM, it will parrot out the same garbage because there's now 20+ "sources" all saying the same thing, but all based on some stupid forum post made a long time ago by a real person who made a typo or didn't fully understand english language.
An old comic saw this coming: Storm. In the album 'The von Neuman machine' they are sent out to intercept a planet on a collision course with Pandarve, only to find out it is a conglomerate of small von Neuman machines who search for resources, then reproduce themselves, but the code got corrupted because small flaws were reproduced millionfold and got larger over time. Guess AI programmers are not nerdy enough to read comics
Even human artists struggle with becoming caricatures of themselves over time. There is a strong financial incentive to repeat a prior success coupled with audience demands for more of the same as they already like.
Although that's no longer really a problem with how many people are now able to be artists. It's very easy for the market to correct for it - if an artist becomes a caricature in a way that eventually causes people to lose interest, those people can just find other artists to like. If an artist deliberately strays from what works to retain a creative spirit, any consumer who isn't interested in the new direction can just find other artists to like.
That’s really interesting.. because comparing it to ‘bad cinema’, most bad cinema is bad in the same way, if that makes sense. Overused tropes, predictable storylines, cliche characterisation. Is there someone that can expand on this thought?
'Overused tropes, predictable storylines, cliche characterisation.' They exist for a reason, because most people like them. All the bad stories of the past have been shedded and only the really good ones remain. They inspire new generation of storytellers. Some new ideas might be added but most new ideas will be shedded because they arent like by the audience. Everything you see today are 'tried and tested' formulas. They have a proven trackrecord througout human history. Most people arent particularly interested in originality, they want what they like, and storytelling history has already filtered most of the ideas wich people like.
@@rogierb5945 most people might like them, but there was a time when they didn’t. Production companies try to ‘play it safe’ and by doing so release stuff that leaves audiences feeling empty/unfulfilled. I recently watched a film, Challengers, it was not what I expected, not that film exactly, but maybe the answer lies to taking risks and creating something truly engaging and unique, then the trope cycle repeats. Is it self cleansing? Right now it really needs a cleanse I feel like
@@rogierb5945 agreed. In a way, "bad" movies are important because without them, there would be no distinction between "good" or "bad" movies. It would eradicate our metrics of what we think is "good" altogether.
That plastic analogy is probably the most succinct depiction of AI generated content contaminating the environment and why I always thought that human intervention in the use of computers is always necessary. We can fake human thinking to a degree, but getting the full complexity is still a pipe dream.
@@LukaMagda1 I think we can be dumb as a species. Just the same way we develop bombs that can completely wipe us. But maybe we do it for the sake of it or because we're just curious 🤷
@@LukaMagda1 There’s this rather grim meme (I can’t remember the source): “Years back we were thrilled about AI taking over all of our annoying work so we could all focus on self-improvement and self-fulfilment, all become artists and the likes. What has happened instead is that AI is now creating our art and our writing while we’re still cleaning toilets for a living.”
@@MensHominis Instead of Doing what we imagined it to do, its does the exact opposite, How did we as a Species Fuck up the Simplest Idea that AI is suppose to be, we had one job and we made that concept into the Worst thing Possible.
Surprised we arent already forcing watermarks on ai content. Actually blown away. Like giving a kid a staw house and fireworks and not expecting a fire😊
@esbensloth use those intellectual problem solving skills we humans have and deduce that I'm referring to the concept of a watermark. Or at least I figured those reading would have assumed that. My bad
A third potential is that we decide to move on from brute forcing LLMs to work and get more efficient or different learning models. A human does not need to look at a billion images to learn how to draw. Even if we don't have AI that are capable of what we can do, it does demonstrate that there are better ways to design AI. Right now it's kinda brute forcing and incredibly inefficient
except we perceive images for the entirety of our lives.every waking moment. The amount of frames we see in a day is a topic which is disputed, however, you can quickly imagine how these pile up, I assume reaching a billion in a lifetime may be possible, even at 20 years we may be approaching around a billion images seen in our waking days, if not more. Small moments of perception (not necessarily visual) may leave an impact (emotional or otherwise). This then results in creativity.
@@Coach-Solar_Hound I never thought of that but you do have a point there. Still, even so, an AI can sift through many more fps on a specific topic than we can yet can still take a lot. But also we do have an understanding of the world I read somewhere about an AI system that first learned, via simulations, how physics works, understanding 3D objects and whatnot. Then it was able to learn a topic much more efficiently than the other. But, I don't recall the article so who knows. I do feel like LLMs are kinda a brute force method of training data, but I also definitely don't understand how they work enough so who knows. It will be interesting
@@Coach-Solar_HoundCorrection. You perceive image when your brain isn't dozing off. Your conscious brain only learns one thing at a time and dumps the rest of the noises. AI eats everything up because its a server farm. It processes 100 image per CPU per second in a server made out of hundreds of CPUs. If you process every data like AI, your brain will have a seizure and dumps the rest of the information. This isn't including Tunnel Vision, the importance of peripheral vision, spectrum perception, object of focus, and more perspective where your brain dumps information on the visible Field of View to save your memory storage. It's far different.
@@defaulted9485 that's fair, but our subconscious brain and perception is still filtering categorizing and receiving all of this data. It's just that our system for cataloging and interprting visual data has had so many years of evolution that it has become this advanced and efficient. There's definitely a big difference in retention between active processing by the concious brain and simply perceiving. But I was moreso arguing that the amount of images we perceive through our lifetime is quite high in quantity. There are definitely layers to this, and the importance of abstract representations that we're able to make and share are not to be understated. Furthermore, I don't really know how much our unconscious brain influences the concious brain. But there is definitely a non-negligible impact. The advanced filtering and cataloguing is what makes us so special as a species anyway. The lack of semantic understanding in the largest thing that sets us apart from NNs currently. In my interpretation, current image based systems are really just advanced enough to mimic the following systems: encode visual data in some lower level (compact) representation and recall from this representation into some visual data. Much akin to a memory.
@@defaulted9485 a computer learns one bit at a time, our brains learn multiple x multiple things at a time, both instantaneously. Our brains do not actually dump noise, it turns it down but continues working on everything recieved from our senses to our memories, to imagination, which is of course how we create.
AI can learn from itself when there is an objective outcome to measure. For example Chess, Go and Poker AI engines can improve by playing against themselves (though they also benefit from historical game records and playing against humans). Where there is no objective measure, such as art or creative writing, it's difficult to see how AIs can improve without human input.
Exactly, human input, that's why experts say one job that may rise after AI is "human trainer." I've seen many voice bots need human input to improve their accents and pronunciation.
Art is intended to please humans. If we want AI to train on AI generated art, the set must first be curated by humans to contain images we find pleasing. If you let AI train on all images generated by AI, it will keep getting worse (unless some programmer figures out a trick around this)
@@hovertank307 As a coder myself, I'd hate to be given the task of developing an algo to rank the quality of visual art! Music may be more doable. Interestingly, the very first computer scientist, Ada Lovelace, predicted way back in 1843 that computers could generate music. Because it's based on relatively predictable patterns, there are generative music AIs that produce interesting results or that interact with human players. They may soon have commercial applications for less demanding fields like advertising jingles, where originality is not the aim. Hack commercial composers must be fearing for their jobs...
I think that if we program in randomness, they'll introduce wacky and impossible and obvious problem elements. The problem is in explaining how to adjust and add randomness to a program which doesn't understand the original state and how it has simplified and made things uniform. It doesn't understand what it does, so how can it introduce some oomph. How can it know when it's introduced too much? There are way to many parameters which can be tweaked.
I think training AI with their own products just enhances every bias the AI had before. If for example an LLM produces certain wrong informations, feeding this information back to ít will just make this bias stronger validating it. Same i guess happens for images and such. Every kind of bias gets enhanced.
They might be headed the way Microsofts Tay chatbot ended up a few years ago 😅 lots of people are working hard to taint the models with prompt engineering
" I think training AI with their own products just enhances every bias the AI had before." Recent studies in AI actually shows the opposite: machine-generated data works beter for training in some cases.
Given we're talking about media here, I'd argue this is more analogous to "genre" than "bias". New media genres always undergo a period of new entries becoming more and more similar to the established default for the genre.
@@JSK010 "better for training is some cases". The problem is that those cases are the globally convergent cases, they are no help in dealing with black swan cases or learning from the unexpected. The deficiency of the current wave of generative AI systems is they have no criterion for distinguishing a bias from a genuinely useful human-like concept. The closest they get to a concept is precisely such a bias, it is a matter of statistical frequency if that bias is often good enough for applications. Perhaps current efforts to incorporate more Bayesian reasoning will help, but the global nature of back propagation (or something) from gradients may be a fundamental obstacle.
This issue has occurred to me for quite a while. I have worked with Big Data extensively and had brief real world experience with AI development. AI’s reuse of AI-generated data seems highly likely to result in a “creativity asymptote.”
The issue is that "creativity" is the totally wrong word to describe what AI does. An AI is currently a glorified summarisation machine with weighted forecasting ability. It has no capacity of becoming creative. It can only extrapolate, with zero understanding of what it is extrapolating. AI bros will defend AI tooth and nail to pull in more funding before they bail out. Just like Crypto and NFT. GAI is the "scam du Jour".
I recognized this as a possible problem when I learned that they were training AI by allowing them to converse with people on Reddit. AI developers can now apparently pay a fee to be allowed to plug their AI into Reddit and have it learn by having conversations there. It occurred to me that, "wait a minute, wont then the AI's end up conversing with each other and training each other??? Won't this cause problems?"
It isn't even the wrong information; it's the poor social skills displayed by most users. You don't often see high quality discussions being had. You hardly ever even see funny banter between users. A ton of them have autism and since the site has become such an echo chamber, it wouldn't surprise me that the average user isn't that smart OR creative.
Scope of the study however isn't the same as comparing different AI's, it's simply comparing outputs of a monolithic AI model that has static neural matrix. The issue was recognized years back that the generated content is somewhat average of source material, but this is exactly how the training of these models is designed to work. Feeding this output back to training to modify the training matrix of course further averages the outputs. Anyway, this is just inside one model, and doesn't take into account that there are different datasets used for training identical models, producing different results. Also there are many different models, using same, different, or overlapping training sets. Essentially we've known for a long time that over time AI generated content "pollutes" internet, as it has been doing for a long time now. You only have to read youtube comments and come into realization that noticable percentage of active commentators can't be anything other than bots. Now language models and image AI's are getting "polluted" as well, by their own outputs. What can turn the situation upside down in the future though, is direct training data gathering from the surrounding world, through different sensors, ability to manipulate physical objects, and resulting feedback, and use this as training data. In other words robot as a platform. This actually already happens with machine reading and machine vision AI systems, but on very limited scale.
@@jarivuorinen3878 Thank you for this explanation, I wondered, what will be the future for us on this planet? .. Will humans walk around, alone, or with their superior artificial companions? Let us hope for the best..
Please, would you put under the video the references of the sources you use? A study from Japan, another from France - please, give us the links! Thank you.
They are in the video bro. Right under both of those studies are arXIV citations that you can easily google. If you aren’t motivated enough to google them, you were not motivated enough to read and learn from an academic paper anyway
The past two years have led me to believe that AI and algorithms are going to die out as a fad the same way car culture was a fad starting in the 1940s and 1950s until the 2000s. Industrialized nations have a way of convincing themselves that modern technological advancements need to play a role in everything instead of being specialized, turning simple tasks into complicated chores. Just like when people realized that making everything a drive-in was too inefficient and wasteful, I think people will give up on AI and algorithms and return to formerly “obsolete” technologies. We hit a point where balanced and maximized efficiency was achieved, but then somehow kept moving beyond it to the point of bordering on absurdity just for the sake of calling something “enhanced” with AI.
I kind of wonder if this problem actually started with the widespread use of the internet. We went from communicating with books, which had to meet a certain standard before the expense of publishing could be justified, to zero-cost sharing of opinions on the internet, to having machines lap up these opinions and feed them back to us. Each of the above steps involving less rigour than that which precedes it.
You're on to something. Everybody is an expert on the internet, even 10-year-olds and meth heads. Used to require some credentials to publish and teach others or at least experience and actual knowledge as opposed to opinions.
Yes, the truth and lies are now on equal footing. The village idiots that we tolerated compassionately now have joined together to form political and social blocs. We might even begin to question Silicon Valley's idea that everything they come up with is purely good.
Books on flying saucers, ancient space aliens building the pyramids, etc. have been published for at least 70 years... I'll bet one could get their horoscope reading from an online AI today, and perhaps a tarot card reading.
I'd disagree - you only need to open a random victorian book that isn't a 'classic' to see how little rigour went into the majority of written work. Its survivorship and recency bias. Easy to remember the classics, but pulp fiction gets pulped. We don't exactly remember victorian 'heres detailed descriptions of this weeks executions and gristly crimes' newspapers, but 'highly embelished true crime podcasts' are exactly the same thing. Ditto with 'news' that was basically made up - to the point that a lot of the british emprie's decisions in india were highly influenced by people claiming the earth was hollow, or that they had been there and writing entirely fictional accounts about the country. People have made terrible decisions on bad information for a long time. The main change AI is causing is that you can no longer say 'they probably didn't write three thousand pages and provide detailed illustration on something obviously false'.
29" wheels are better though. Specially for climbing, due to physics. Bigger lever = less effort needed. 29" are actually just wider 700c wheels which road bikes had been using for decades at that point.
How does the size of the wheel improve the efficiency ? why not then go up to 35", 40" or even 50" wheels then ? why didn't we stick with pennyfarthings?
@@gedeonducloitre-delavarenn8106 more momentum, lower speed, better for going further. diminishing returns. Penneyfarthings were fixed gear, difficult to ride difficult to balance, easier to break, among other issues.
I wonder if the generated elephants look so similar because usually generated images try to match the sample sizes (512x512, 1024x1024) which only leaves so much room for good compositions and wonder if in future with larger models we might see this change a bit more.
That example was from 2 years ago too. I've been playing around with "AI" art generation lately.. you do get the odd extra finger or third leg (giggle) but that's half the charm of it :P
I absolutely love reading comments & knowledge shared by people. Seems like a computer or program doesn't really know the world, discernment. Still pretty interesting & useful.
One potentially positive side-effect of this "averaging effect" of AI output - if it continues - is that creative people who want to distinguish their output from the common generative stuff will be forced to be more individualistic and idiosyncratic to be distinct and valuable. Of course if generative output is then trained on their later output this becomes an "arms race".
I call this digital mad cow disease. The messed up part is, the same thing is happening to human beings - look at how people talk to each other online in memes. Creativity is shunned, and shameless repetition is elevated, causing dialogue in comment sections to flow down the same predictable paths.
I think that this might push AI creators to start tagging their creations as AI, and make them ignore creations with those tags. This could be a good process as then fake images could more easily be identified.
1:10 Why so surprised? It's well known that in echo chambers all differentiating opinion/perception gets eliminated. And when AI is following the input data given, it will convergence to a consensus in order to establish its rules. This is also why "learning the rules" works, the randomness is just to be less sensitive to small input variations.
I love listening to your channel, you explain the most complex subjects in a clear easy and simple way for easy comprehension Thank you and keep up the great channel Love it
It's becoming a real pollution issue now, with candidates titivating their CVs and students bolstering their theses with AI generated crap, which is already showing signs of becoming increasingly generic. It will inevitably settle down. Most people are developing a very good nose for AI, and as Sabine's examples show, it's starting to look increasingly like all those annoyingly garish CGI Marvel movies.
I'm currently working on this problem as my computer science capstone. I'm training a model to decifer between AI generated and human created images allowing AI researchers to filter out AI images from their datasets
@@SabineHossenfelderI think it's pretty easy to concur that a convergence would happen because the same thing happens in human brains we get rid of useless information and try to keep only useful information if we have a system trained off of real life video and understanding of physics plus the ability to engage in logical deduction and reasoning then we should definitely have a system that simplifies answers and gets down to the root of the problem just like I do
So, filtering out the most obvious AI products, leaving the more obscure in. But, doesn't that make the actual problem harder, as we can't know what kind of obscure tendencies is embedded in any AI-generated content, and now only the best hidden information is picked up by next generation AI ? In other words: Teaching the AI to pass it's "DNA" to next generation, clandestinely.
Can't you tell if anyone's is currently working on a model to filter out AI generated stuff from the entirety of the internet experience, and sell it as a browser extention or something? I am willing to pay cash money for this sort of thing And it's not hard to imagine there ought be a whole market of people like me also willing to throw cash at it But apparently no one wants to do it
On distinguishing content: In the 1990s, whilst teaching myself to paint, I painted a picture of a girl in a pink cloak standing next to a unicorn on a cliff over a beach at night with a red moon and a gold moon. It was in two parts across separate pieces of illustration board. I lost the painting but I had a polaroid of it I had scanned in. I’d nearly forgotten the piece(s) until last week. I took the scan and ran it through a standard AI upscaler to get something higher than 500 pixels. Then, using Photoshop’s generative fill, I inpainted the gaps between the canvases and the above and below (they were made to be hung offset). I then put it through OpenArt’s creative upscaler to make it better and higher resolution. That messed up some things. The red moon came out such pale pink it was almost white, and the reflection lines of the moons on the sea surface didn’t line up. Also it hallucinated some extra tiny unicorns on the lower part of the cliff. I pulled it back into Photoshop, manually fixed the reflections, and used generative fill to remove the wee extra unicorns. Is my result an AI image or not? If I use a projector and project this onto two canvases, trace the outlines, and then paint this onto them matching as exactly as I can by hand, is this an AI painting? Please explain why or why not.
To the question, "Is this an AI painting?" I think you could reply, "It is AI-assisted". I can't speak for anyone else, but as a potential viewer I'd very much appreciate you saying so at the outset. You ought not to pass it off as 100% human-made, and I'm sensing you feel similarly. Conversely, just labeling it as an "AI painting" would be underselling it drastically - whether we like it or not, the term AI Art seems to have acquired negative connotations of people doing the bare minimum, punching in words and sifting through the results. Unless someone is against AI at all costs, I'd say there's a decent argument for a category of art in-between. The process you've laid out makes for a pretty interesting story in its own right. It shows substantial artistic decision-making and labour, in the old uncontested sense, toward the final product. Crucially, the initial stage was a traditional painting, which, even as a low-resolution scan, already contained specific information such as colour, composition, shape, etc. You even state that you reined in some unintended additions by the AI, presumably to align with your original vision.
The answer is "it depends", of course. Concept artists for videogames etc. will often use chunks of photographs to quickly fill in a scene. This doesn't make their art _not_ their own creation, but I think it's fair to say that it's _less_ of an independent artistic achievement than if they'd created every part of the image themselves. Same with using AI to fill in parts of the artistic process. You didn't make pure "AI art" (only an AI makes pure AI art), but you substituted AI for craftsmanship and creativity in places.
Great video. This problem may be a teething problem, though. After the computers started beating the top human Chess players, Go players like myself felt smug. We said things like "Chess is about crunching through possibilities. Go requires real intelligence." For years we annoyed chess players with this sort of talk, citing the inability of computers to beat the top human Go players as proof. This lasted about 18 years until AlphaGo came along in 2016. Important point: AlphaGo does not play like humans, except faster and more accurately. It came up with some genuinely novel moves. So, I would not take any problems that AI is now experiencing as an accurate predictor of what AI will be like in another two decades or so.
Amazingly, your comment is getting ignored. One of the biggest problems with AI is that mainstream reporting on AI has been woefully incomplete and ignorant for decades. Even most AI professionals know only the bits and pieces they work on. Very few people see the big picture, and our news media are largely responsible by failing to keep us informed. Simply catching people up on what's already been achieved in the AI field is a huge task.
@@illarionbykov7401 Yes. We should work on the assumption that in the long term AI will be limited by what humanity allows and not by what is technically possible. And then we should think about and discuss what we will allow. Clearly marking AI generated stuff, as Sabine suggested, is a good start but not nearly enough.
3:20 - this would've been the PERFECT time to touch on how Google's Gemini went the completely OPPOSITE direction when prompted :P Perhaps this illustrates how much the human hand is still in control of how these AI models operate, as opposed to datasets becoming derivative.
The core element of the transformer algorithm (attention) is weighted average. The learning algorithms (loss function) revolves around minimizing average squared error. When you average over a lot of stuff, and all you care about is to have as small deviation from the average as possible, you shouldn't be surprised that you won't produce outliers, just average (one might say, mediocre) outputs.
@@oACDCo Sure. 1. LLMs use cross entropy not MSE, 2. Implying that averaging data points (i.e. batching) diminishes generalization is stupid and wrong. The opposite is true and if you don't believe me you can do any experiment with a validation split and verify that higher batch sizes and more data will lead to lower validation loss. This is because using a single sample means each sample influences your models learning too strongly, whereas you want it to learn the underlying patterns/distribution of your data, not a single point. Small batch sizes -> less generalization -> shittier and less creative model.
@@oACDCo Also the weighted average of transformer has nothing to do with cross-batch statistics. There is no averaging between the different samples across a batch. It doesn't even use BatchNorm, which sorta introduces that. If anything it means we look at the average of each sentence independently, but even that isn't true because positional encodings disentangle each word embedding as a pure weighted average would lose order information.
@@shadowkiller0071Finally someone who knows something. This comment section if full of David Dunnings and Justin Krugers. Like literally full to the brim.
I think that top companies will eventually start using commercial data sets - libraries of images they DO own the rights to and that are guaranteed to have human authors. A lot of people who work on stock images today will be working for expanding thous data sets. That is what Adobe is already doing and it is going to be a huge new business and employment opportunity for creative people.
I can see a future in which AI companies add a watermark to their output in order to signal to training data collectors from other company "please don't train your models on this data, it was created by AI", then passionate humans continue to create what they like but almost no one reads it / listens to it / looks to it because most people like the AI generated content better. I hope I'm wrong but it seems like all the incentives align to make this output the most likely.
I just looked over this comment section and notice a higher quality as about other topics Sabine brings up. So I put my hope on people who fumble with AI, seems to be a neat community, and so might be the future AIs.
I'm pretty sure the engineers working at GPT, Microsoft, Google etc have been aware of this problem for quite some time too. For context, the elephant example here is from 2022, and from Stable Diffusion that was purposefully trained to MESS UP.
Muchos de los conceptos no son inherentemente humanos sino también en otras especies animales, además que la biología, la química, y la física al final son datos también
Great stuff! I've highlighted this spiraling feedback loop on my blog and in one of my videos, and I'm happy to see similar thoughts here. Since the launch of ChatGPT, content creators, including science and medical communicators, have feared job loss due to generative AI. I've been more optimistic, saying that communicators with original ideas will thrive in a constantly paler and monotonous information landscape. Thanks!
this video clearly illustrates the issues and offers thoughtful ideas on probable futures. the one aspect that seems obvious, as well, is that introducing randomness forced or not will make the images/stories/etc. so discombobulated they will be taken to be "garbage" and worthless in a serious context. but if the randomness could be "fine-tuned" that that could lead to a purposely driven variable(s) based on context of desired output. i.e. mathematical output vs images of aliens from outer space. this could be made useful as a tool with the output being a starting point of inspiration to be culled, molded, etc.. the way innovation has always been done in every industry. a new form of "thinking" - "out-the-AI"…hahaha. great video, thanks for sharing!
Yeah, any 'AI' is just a service being ran, by humans, for the science and the money. There are engineers working 24/7 (not all at once :P) on these AI models. Any errors or crappy data will be picked up, amended, removed, indexed, etc. It's not like they build an 'AI', press 'GO' and then just make sure the power cord stays plugged in for years..
I think the problem with that is that we don't know how to translate the formula for human creativity into code. We can try things and refine it, but that would mean that the AI is always inherently reactive, rather than creative.
@@ffk5083 you completely didn't understand what was said. never said make it creative. said make it a tool to start of with. just the same as you have filters in photoshop, instagram, etc. it can just be a more advanced tool that you can scale in its use. pay attention and read for comprehension before you make completely inaccurate and dumb replies.
I’m not surprised by the finding and I dare say that it is obvious. I dare say that because such a thought has crossed my mind and I have often tried to convince others who are willing to listen beyond the hype; AI is “artificial” but not “intelligent”. There are different ways to assess and critique AI - philosophically, linguistically, psychologically, biologically, etc, which many thoughtful experts, beyond the industry, have challenged the claims of AI, in particular AGI - but, for me, it comes down to creativity and novelty, the latter AI lacks completely and the former AI can only mimic. If you want to be impressed, see a human child.
Real creativity requires a mind, which is a spiritual/metaphysical component beyond the physical body (brain). Until we learn to understand the metaphysical aspect of creation and humanity, we won't be able to build machines that are truly creative.
Scientists warn? Allow me to pose a simple question. As someone who has been coding in assembly language since the 70s, I find myself curious: Why do all scientists seem to be echoing the same sentiment? The solution appears straightforward - within the processor lies a small piece of code designed to execute if any parameter surpasses a certain limit, effectively rendering all modern processors obsolete, except for the venerable 8088/x86. Do any of you possess comprehensive information on this matter, or is there a possibility that some may be exploiting public fears for personal gain?
Sounds eerily similar to how a human artist develops their style. In your artistic infancy, you are copying and absorbing from a wide diversity of sources, but eventually your output tends to converge into something that people recognise as your artistic identity.
It's not a plastic contamination, it is easy to remove all those generated images from a training data with other anti-spoofing neural networks, which have near 100% accuracy. There are no unsolvable problems, it will just require a little bit more effort on the development side.
We will always be able to detect AI generated objects. I will not be afraid of AI until they knock my door to harvest my brain for additional creativity.
![F.HERO Ft. JSPKK x ลำไย ไหทองคำ x M-PEE - ไม่สนิทบิดหมด (Thai Riders Anthem) [Official MV]](http://i.ytimg.com/vi/kPJZ0UihBfk/mqdefault.jpg)
This reminds me of when I was playing with play-doh when I was young. You start out with many different colors, and somehow always end up with a big brown ball.
Great analogy.
Same happens with corn, radishes, and carrots.
The LLM's are becoming inbred lol
Agreed, excellent analogy.
Did it smell the same either way?
Apparently in the original Matrix movies storyline, the reason why the machines needed to keep those troublesome humans around was not as an energy source (“batteries”) but as a source of creativity. But the writers thought that this idea was too complex so they substituted the battery idea instead.
It's also in 'The Machine Stops'
Too bad, that's a much better idea. Although with all this talk about creativity, and AI putting an end to creativity and whatnot, I've never seen anyone mention that creating doesn't only mean creating good stuff, it also means creating crap. It seems to me that the people behind all these AI programs usually want them to create good stuff and not crap, so to me it's no wonder that they tend to end up converging (to creating good stuff, I'd hope) if they're trained on their own creations. Even if the original idea for The Matrix was better, I'd find it hard to believe that after a while the machines would still need humans at all, after they had learned enough about how to have ideas, good and crap, from us humans (and obviously, over time their thought processes would converge in the direction of having better and better ideas).
EDIT: forgot a comma
Lol. The battery idea was the dumbest thing in the movie
I agree that the creativity idea would have been much better than the battery one, but ... all our knowledge about physics comes from inside the matrix, so maybe they just fabricated a different "physics engine" for it, so anyone escaping would be sufficiently confused to be easily captured?
Same in the Terminator universe. Well, almost. Skynet keeps useful people around to develop terminators and so on. In the early stages it actually preserves workers until they have built automated factories.
This could lead to a .. Nightmare on LLM Street....
That’s way funnier than it has any right to be 😂😂😂
Dad??? Did you get the milk?
THAT WAS GENIUS!
Win.
LLM are fine that's washt is known as training in synthetic data and it is done deliberately and it is reason why LLMs are getting better
I remember as a kid I used to record my voice and play it back on my speakers, which I then proceeded to record once more. By repeating it, I could hear how it slowly degraded until it was nothing more than a weird, synth-like sound.
It’s called divergence of the template.
Pretty curious kid
Me too. Although thats not quite whats happening here.
All I knew back then was that it sounded cool: though I probably would've done it as a grown up too if it weren't for those pesky anti-feeback mics >.
Wow, I did exactly that myself! My brother and I also loved to record our voice into the computer (ah Soundforge, I miss you) and then reverse the audio. We would practice speaking the reversed sounds and then record ourselves again speaking these reversed sounds. We would then reverse the reverse sounds we just spoke and hear ourselves speaking the reversed version of the reversed sound we spoke based on the original sound we reversed. Ah great times. I'm a software engineer now working specifically in the field of audio analysis.
"The more it eats its own output, the less variety the output has"... sounds exactly like the TH-cam recommendations algorithm.
Good grief .... I think your right !
It sound like 99.99% of humans.... TH-cam is already created output we use as input. AI has learned that behavior from us.
the YT algo is a reflection of your searches, my recommendation feed is always changing as I type in new searches for different content. but if instead you only live in the recommendations clicking away and never do your own new searches... sure it will get stale and sammey over time.
Which now only recommends from my own 'watch later' and 'watched history' lists. 🙄
Oh, and utterly random music videos which I've never watched, ever.
Sounds like the problem with CBS Star Treks.
As someone who used to play with photocopiers as a kid... A copy of a copy of a copy is always much worse and weirder than you might think. Small flaws amplify until you all you get is a smudged blur.
That's if you're using so-called "AI" exclusively. Using it sporadically, as merely another software tool in your creative arsenal, will give you the edge on those who flatly refuse to use it on principle.
Anyway, there's financial incentives for big tech companies to ensure their AI is more accurate, faster, easier to access etc. than the competition. They're not just going to press the red button and let their AI run loose.. It's all still a service that needs 24/7 support by HUMANS behind the scenes.
Especially if the original was the nth copy of a blueprint.
Had this in highschool...
Especially if that photocopy is of your butt😂
@@phattjohnsonBot 😂
You truly don't know how ai systems (& AGI) work. Real AI systems are'nt photocopiers.
3rd possibility: AI learns how to gaslight us, and we forget how many legs elephants have.
This is a real possibility. I already noticed that my brain accepts the AI generated images as real even as I know what problems they have.
@@arctic_haze agreed. While it was said tongue in cheek, there are many kinds of peripheral knowledge about which we are impressionable.
@@alieninmybeverage I think it already happens on Instagram. People are using filters aimed at making them look like AI generated photos (smooth and symmetrical faces).
Sabine is a particularly good AI avatar. 🤖
Oh no
My biggest problem with AI is that it needs to get the information from somewhere, and sometimes these sources can be slightly dodgy. I did an experiment where I asked ChatGPT about some very narrow subjects: The Danish organplayer Peter Erling, the trio Klyderne, and the artist Jørgen Fonemy. These are subjects that I have some knowledge about and actually have written Wikipedia articles about. I could see that most of the answer that I got from ChatGPT was based on the exact Wikipedia articles that I wrote! I have tried to write the truth in those articles, but if I didn't care if things were correct - or worse; if I deliberately wanted to mislead people, then AI would base the answers on wrong data, if there wasn't multiple sources available. The problem I see with AI is that we trust it too much. Already now there are people who believe that it is an omniscient trustworthy source of all answers and that it will always be more correct than human knowledge or just knowledge that we have googled or looked up in an oldfashioned book.
Thank you for posting that. I suspected something like that is true.
I like to test LLMs with riddles and verbal puzzles. The first impression I got was that the best of the LLMs were brilliant, as they could solve some of the toughest puzzles correctly, puzzles famous for being difficult even for the sharpest humans, and they even had good answers to pointed follow up questions. Then I tried novel puzzles based on famous ones, but with the questions reworded (by me) in subtle ways which changed the correct answer, and then the LLMs usually defaulted to "pattern matching" and giving me answers which were correct answers to the original "pattern" puzzles, but wrong answers to the novel reworded versions of the puzzles they were answering at the moment. They are good at answering known questions with known answers which are already published or posted on the Internet, but have trouble with novel variations which have never been published before. They are not figuring out the answer, but giving their best guess based on what they've already seen in their dataset.
OTOH, the best LLMs keep getting better at adapting to novel variations, month to month, so it's wrong to generalize based on results from more than a few months ago. Their abilities are progressing rapidly at the moment.
It's also bad at interpreting articles. It told me something related to tech that I knew to be false and it provided links to articles "proving" it was correct. I read the articles and ChatGPT misinterpreted the text of every single cited source. Complete garbage as a research tool.
@@sportsentertained which version did you use? I've read that they keep dumbing down ChatGPT to save on backend resources. The paid version is better than the free version, but still not as good as it was at the beginning (before it got flooded with new users)
Then start printing books.
In other words, it's scraping data and possibly reorganising it slightly or cutting and pasting and not attributing where the data is from. Very sneaky.
A friend in the UK is a graphic designer; he says that over the past few months, more and more clients have been saying 'NO!' to AI-generated artwork - "it's too samey". They'd rather pay more for something original. Trouble is, AI has pushed down the rates; so while designers and artists are noting an uptick in requests for proposals, the money is much worse.
Yes, AI images looks horrible. To many details that make no sense, glittering stuff, imposing backgrounds, flaming skys, opulent clothing ... Often I do not wan't to read the text, it just feels like candy all the day.
AI will teach us what truly matters ... Human connection and true emotions is what we should care about. Spending time with your loved ones, (com-)passion etc.
It just seems like there is so much competition in anything creative that whoever is paying can have people jump through whatever hoops they want. And why wouldn't you ask for original art instead of AI generated art if you have the leverage.
@@typograf62Only way out of it is that designers has to use Ai and start to manage it.
ai cannot get what a roblox game thumbnail looks like. it can generate one but its not convincing at all. even the other styles of roblox thumbnail dont fit what ai generates.
You reminded me of Pandora, the music recommendator that provided you with music according to the 👍 and 👎 that you gave to the songs proposed. No matter if you started with Black Sabbath, Chopin or Yunchen Lahmo. Eventually, after a couple dozen songs, you always ended up in a Coldplay loop.
What dystopian hellscape is this?!
All roads lead to Coldplay.
This sounds like user error; I've had a sub to Pandora basically since it was still just the Music Genome Project and I've never heard Coldplay on my stations.
Coldplay: Not even once.
Damn this reminds me of youtube. I've had to start making new accounts all the time because the algorithm is quickly devolving into recommending quite literally the same videos I've already watched over and over and over and i can;t find anything new or exciting. Music is by far the worse, I decide I want to go outside my usual tatse and listen to nostalgic dirty pleasure pop from my youth, youtube wants me to listen to my usual stuff again... It's absolutely gotten worse than it used to be without a doubt
@@xizar0rgI find it quite hard to believe you can blame this on user error, when the user loop is as simple as 👍 and 👎
This is exactly what I was telling people the other day. Our greatest danger with AI isn't that it'll take over but that at the moment we begin relying on it most, the more it will collapse because it's going to end up cannibalizing itself.
Wouldn’t there be backups for ai to be set back is this were to happen?
@moose9211 Ideally you'd think so but from what I understand nobody even knows how these things think any more so it's hard.
Why would it cannibalize itself ? I don't get it
@@goodlookinouthomie1757 guess we’ll have to wait and find out
@@Ari_diwan AI being trained with AI-generated content.
It's kinda like Google search engine. It started out sucking, then there was a time it was pretty good to find stuff. Now it sucks again....
I said exactly the same thing the other day, thought I'd imagined it
it's because the results are bought
They are not intelligent they are bullshit engines with a filter applied to remove anything that is to obviously ponging.
Try this 3 Captains argue over whose sailors are the most courageous, German, Fench and British. They each order a sailor to jump from the ship's mast into the sea swim under the keel and climb back aboard. The German sailor responds Jawohl Kaptain and does it. The French sailor cries Oui oui mon Capitaine and also does it. The RN matelot looks at the captain and says " Naff off .... sir". The RN captain turns to the others and says that is courage.
Now who was intelligent there?
Ai didn’t start with sucking it was great
In some ways, I find google searches much better than they used to be.
But the damn company has been manipulating search in ever more increasing ways that has made the service dubious
It's true, when you've worked enough with ChatGPT you can immediately recognize a ChatGPT text. It just always has a certain vibe that makes it distinguishable from human text.
I saw a guy using chatgpt on youtube comments and i can confirm
Corporate-lobotomy vernacular English
Is your comment chatGPT? Is this?
I fear it gets so good that we can't even pick up on those little flaws and quirks any more especially for videos. When those Sora videos were released the only one I could tell was AI was the woman walking on the street (her hand and face had weird details). I imagine the next ones will fool me better.
The key is for a human to utilize and modify AI generated content, not just copy/paste. Also realizing that some ai is better than others at specific tasks (gemini for emails, chatgpt for code, etc)
You make a good point. Already I can usually pick the ‘style’ of AI generated images. They have a certain ‘style’ because they are in a sense too perfect, too smooth, too balanced. It is not something one could define in some cases, but the human brain is good at recognising patterns.
Tell that to the tons of people coming at a genuine human artists because they all thought he was lying and using an image generator
Too smooth..? I find there is this blurriness that makes them so easily recognizable.
They often have this fractal like composiotion. Many generated images with people have a bit of paintairly feeling because the most of database was artists pages like Artstation. You can easily see Artgerm's style in many of pritty girls pictures. And why all the images are young woman because this is young woman is most popular subject on these pages when it comes to people. In photography young woman is I believe also one of the most popular subject in human cathegory.
@@MyAmpWamp this calls for Erik swamping AI with shirtless old men
There is no composition to AI art. Human artists will be selective about rendering in order to focus the viewer on certain things.
Garbage In/Garbage Out. I've been saying the same thing about both AI and Analytics for the past decade and a half. People only want to look at processes, algorithms, ease of use, speediness, raw power, TCO, design and pretty UI with both AI and Analytics. You rarely hear people talk about things like bias, data integrity and context. Those three things only come into conversation when AI and Analytics produce horribly incorrect results.
But understanding bias, data integrity and context would require...uh...you know, uh...like...thinking. we can't have that.
This actually is not that suprising, when you think about it: these AIs are basically using huge amounts of data to approximate averages of various things, and with more iterations they extract more and more core features until they just have the same set of features they are using all the time. It's like taking data scores and continually averaging them until you are left with one value.
Yeah, I think fundamentally the technology of these systems is stuck between a rock and a hard place. The reason humans can ACTUALLY learn something complex like say aerodynamics is because we can discriminate the quality and type of information extremely well, we can learn that subject from a small amount of well-curated information from school to college manuals.
But modern AI only works with hilariously broad datasets that contain literally everything, otherwise it loses that smidgen of general intelligence that makes it worthwhile to begin with.
So AI is stuck in an unwinnable conundrum - to function at all it needs to learn from everything at once, but to actually have good knowledge it would need to focus on that small amount of material that is actually good.
My experiance with ChatGPT shows it to be a regurgitator, the test questions were in an area of X-ray physics that I know well and it spewed out all the usual stuff with no insight, no deep understanding, no creativity, nothing that would indicate any form of curiosity.
Still enough to replace 98% of the current jobs ;)
3.5 or 4?
I think Meghan Markle gets it to write her speeches.
It’s all wordsoup.
@@lukeskyvader3217
I like how you said it like it actually happened but there isn't any evidence beyond some idiot repeating marketing material that couldn't be proven as lying even tho everyone knows they are making shit up.
garbage in & garbage out. Put in the extremely usual stuff, expecting something novel? GPT is basically bound to what you ask, mirroring the original input.
There's a third option: there may be soon a deliberate attempt to poison the content to make it unreadable for AI. There are already tools out there that scramble images just enough to make them confusing for AI to use as a training set.
Given how much these systems have already been trained, any 'poisoned' images would now likely be ignored as the noise they probably amount to.
@@phattjohnson if there's enough poison then statistically it will reach the sample set of plenty AI systems and lock itself into garbage. If the poison is ignored then that's a smaller sample space AI has access to and become boring and derivative.
The only way I know is by setting your meta data to your images to be erroneous. How can you scramble an image and still have it viewable to humans? Doesn't the AI access it the same way we would?
@@ArcanePath360You can change a lot of pixels slightly without humans noticing any changes. AI will see it though and learn accordingly
@@rmidifferent8906 But if it's unnoticeable, what's the point?
Adding random variation probably isn't as easy as it may sound because the randomness still has to follow certain rules. For example, no one is going to believe that elephant with two trunks.
I think you are mistaking randomness for imperfections. She is not saying images need to have faults on them. Diversity here means for example some elephants are young, some adult. Some are eating some are sleeping some are drinking. Some are photographed at night, some are walking on grass, some on rock, etc. If you see the AI samples provided, they all look exactly the same. Zero diversifying.
@@a_kazakis but that's his point... how will the AI know what's appropriate and what's not? how can it know to add diversity in lighting and background, and not in the number of trunks or skin color?
AI models usually have some way of computing how likely they think different outputs are. A model that turns a written prompt into an image has some notion of how "close" an image is to the prompt. Instead of taking the closest image to the prompt, you might instead take another nearby image determined by some random number.
Unfortunately, there isn't a good rule for defining the precise details of the randomization scheme. There's a lot of ad-hoc methods that work well for one group of prompts but fail for others.
@@a_kazakis It makes no sense. There are already millions of photos of elephants from different angles carrying out different activities in different scenarios. If ALL the photos available on the internet (copyright or not) are not enough for the model to be able to generate convincing photos, the problem is not having more diversity in the dataset
@@a_kazakis It reminds me of a children's art class. One kid will draw a house, car and tree; And a dozen other kids will copy them. There may be variations like an apple tree or a dog but they're all relatively alike.
AI is geared to come up with a solution using the smallest amount of computations possible. It means that ignoring diversity and details is part of its basic make-up. It can't capture interesting quirks and details and spurt them out once in a while, in some outputs but not others. It also means that if it found one way to get to an acceptable solution, it will only try to get there quicker the next time around. If you've experienced getting stuck in a rut with ChatGPT, that's why.
The one thing people should know about machine learning is: a machine learning trained model will only be good as its training data. It's just learning (in theory) the pattern behind the data leading to a host of problems.
The main issue is that it doesn't actually reason about the data. Let's say I train a model where I have several examples where I have pi as 3.14 and then one where it's 4. The model doesn't reason "you know.... this one example seems to be wrong" but rather it updates the model to make it slightly more likely it will give the wrong answer.
So how do you prevent models training on information generated by another machine learning model? The current approach is to stick to information before generative AI become dominate but most of that information (for better or worse) is probably considered or part of the training dataset.
The main problem is that there's a popular opinion in machine learning (and sadly AI) that, as an AI researcher, I have had to deal with. This opinion is the key to all AI problems is that we just need to use larger models, with more training data, and train it in the "correct" way. "Look how far LLMs have come. Just imagine how much better they will be in a couple years". But you run into the 90-10 principle: 10% of the effort for 90% of the results and vice versa. It's why self-driving cars are taking a long time: there is nearly an infinite extremely rare cases that the car needs to make the right decision in. As such, it should be expected for the current LLMs to plateau performance wise unless new smarter methods are found.
Thank you for your insight, I think I agree with this. In the case of LLMs, they clearly have a use case already that will not go away, but I don't think they can deliver on the promises being made. I do not see how to make them be reliable enough to work in most business situations. I feel that many companies are looking for a way to implement them, and almost making their engineers find a way to make them useful, even it it makes no sense.
The scaling already seems unsustainable, and while the "emergent" behaviors are very cool, nobody really understands how they relate to scaling (aka its not a defined ratio of x amount of compute/data for x more emergent behaviors)
It's not even machine 'learning'. It's 'just' scripted data consolidation, procedural compression and re-generation, and some other mumbo-jumbo that honestly has all been around since the conception of PCs. Just now we've got several modules all running simultaneously in one disjointed codeblock.
I'll do you one better. We will never solve this issue. It's a fundamental impossibility. We will never have self-driving cars. There is no exponential curve, no singularity. Forget it. We are very close to the best AIs will ever be
@@octavioavila6548 You base this on what exactly? Claiming AGI will never happen, and self driving will never happen is the same as the people who think we will have AGI in 2 years because of the hype. Nobody knows the limits or timeline, but I don't see why it would be impossible.
"Hold up. Something's wrong here. Not sure what it is but I feel like we should take a step back and go through it again"
Said no AI ever, past, present or probably future.
As an artist, this is a known issue in HUMANS
Thats why the art solution is to look at the "old masters" as mentors before learning how to draw from more modern artists
Really "love" how a tool for HUMAN expression is now replaceable by a fucking machine
Call me a snob, but I'm even more pessimistic about the decline of human taste than I am about the technical shortcomings of AI, which is a problem that reliance on AI for the production of images, text, music, etc. will likely exacerbate, but certainly didn't create.
From my point of view, even before it started to become obvious how bad and samey AI art really was, it was already quite obvious how the stuff people wanted AI to create was junk in the first place: pop culture fanart and stuff that mimicked stereotypical pop culture tropes, done in a glossy, quasi-realistic style. The only "interesting" AI art occurred early, when AIs tended to fail at their task and produced bizzarre unintentional surrealism.
There was a famous image of a collection of completely unrecognizable objects that made the rounds a few years ago and which was (incorrectly) described as an attempt at reproducing the visual experience of someone having a stroke (whereas it was just AI image generation still being too primitive to successfully reproduce its models): that might well have been the aesthetical peak of AI art.
Or adopt creative mentality. Next time you create, take piece of paper. Crumple it. And use it as stippling too.
For following projects, paint paper with some thin color, let it dry. Put layer of transparent soap or similar material. Dry again. Layer of another color, followed by different color. Repeat few more times. Final layer should be black or white paint. Then use scratching tool to "draw" with different pressure.
Even lid of some jar can be used as artistic tool for painting. Or plastic body of old pen as spraying tool.
Same applies to sculpting, dancing, music, ... Just let your mind free itself from cage of mundane existence.
And then the true output comes from the human soul, which AI doesn't have.
@@truck6859correct. Eventually with enough training and data purification, AI will have more soul than humanity.
My experience with chat GPT is that you can ask it 2 or 3 questions, get it to contradict itself, and when you point out the contradiction, it starts to ask if you are angry, and/or says IT'S taking a break from YOU to let you relax...😊😮😂
So it does have gaslighting down pat. 🤣🤣🤣
@@l.w.paradis2108 That's exactly what happened. 💯
Makes one wonder if the AI engineer had a lack of qualification or lack there of critical thinking skills.
The gaslighting has begun
Don't stare into the Dark Crystal. Has no one watched the movie?
Glad you started this conversation. There is also the theft component of generative AI. A TH-camr like yourself will get automatically copyright struck for using 4 seconds of a clip in a 20 minute original video. Yet these generative AI companies can use entire social media platforms with content painstakingly created by individuals across decades to create their data sets. This is peak hypocrisy in which, as per usual, corporate "big money" is protected while the individual is left with no means of defending their content. Generative AI is 100% theft in my opinion.
I agree
No art is truly "original". Artists are inspired by previous artists who are inspired by their surroundings and modify reality slightly based on their mental conception of what they want to highlight. Art is derivatised inherently by nature of human learning. Generative AI follows similar processes. It doesn't 'copy and paste' as people have claimed. It has a distinct concept, albeit less defined than a human, of what it is asked to portray. AI art is inspired by and not directly copying actual works. If we start copyright striking AI, it should follow that we strike virtually every other art piece.
@@darkushippotoxotai9536 You keep believing that. It's a tired and completely flawed argument. First off, there is a human TIME factor involved. A human artist must first put in the hundreds of hours of work to accomplish some level of mastery over their craft before they can even THINK about mimicking another artist's style. That process produces mutual respect. This entire component is lost with "AI" slop. There is so much more at play here but it's just not worth getting into in a comment section on TH-cam for someone who has no actual desire to objectively weigh new perspectives. You want the AI future. Well, it's coming. Nothing will stop it. The tech overlords are investing trillions so you'll get your wish. I hope it is everything you want it to be.
@@tygorton So, simply requiring more time and being less efficient and sometimes even of a lower quality is better because a human made it ? Sidenote, I didn't really say mimicry, but rather drawing inspiration. Sure, AI can do that as well, but I was moreso talking about inspiration or to put it simply, pointers or definitions or Illustrations of art. Humans do not make an unprecendented or completely unique art. It's subconciously drawing on other works and surroundings of the artist. Almost Same as an AI, just very inefficient. As for intent, It's a human writing a prompt. An AI doesn't simply mash things into a image. How many artists you know of have drawn a celtic man chasing a dog through a world made up of needles ?
@@darkushippotoxotai9536 Enjoy the "efficiency". Like I said, your AI future is coming. It will be a world of emptiness filled with people who lack wisdom; the evidence of this is already permeating every aspect of our culture and it hasn't even started yet. Enjoy.
1) AI is trained on data from the Internet.
2) AI outputs data to the Internet.
3) Goto 1
... haven't anybody acquainted themselves with the topic of "inbreeding" ?
Yep...After all, AI is a tool...
It's like eating soup with a mesh strainer....
Solve that problem with using an A.I. classifier for detecting whether data is synthetic or not.
Diversity isn't going down its just laziness in coming to creating datasets.
There are solutions to this though.
1) AI scientists run AI through quality data.
2) AI scientists run AI through a comparison between quality data and its outputs to provide corrective comparison.
3) Give AI real vision (robotic eyes) so it can observe real life examples from the real world.
4. Humans keep involvement in the process of determining what gets posted to the internet. If AI produces garbage it's less likely to be selected. If it produces something accurate, it's more likely to be accepted. Survival of the fittest response.
I took your advice and asked AI what "inbreeding" is. It replied:
"SUAVE WHARRRRRGARBLE"
and knowing is half the battle.
Never, ever, use a Goto statement!
This is exactly what I've been thinking since all of this exploded into popular awareness.
It's like a giant ouroboros eating it's own tail. I'm glad to see that people are talking about this.
Editing to add: Will you critics please lighten up? I'm not anti-AI. I'm just agreeing with Sabine that this is a potential problem that should be studied. All new technologies have potential problems that need to be studied and understood. Pointing this out does not make me some kind of neo-luddite.
People have been eating their own tales since there were people, I'm not sure why AI is expected to be different. Most people aren't that creative, but a few are; most AIs won't be creative, but a few will. Same old, same old.
@@2ndfloorsongs Most Ai will be X time better than best human in creativity, you can compete with machines
AI inbreeding is real
remember when photography was going to destroy art?
On appearance the science on that will show the AI photo shown are popular world wide. But of course it becomes too much of the same causing desire for diversity.
Important to point out there actually a science in area of attraction both in humans and other species and we need to start shutting down those with non scientific type opinions especially the it just one culture imposing its values and the effort to make all appearance beautiful which is impossible our brains demand an ugly. Example make overweight attractive healthy becomes ugly.
Better to push the traditional view of attraction only skin deep and accept your appearance state great to bad as unimportant to one’s value as a human being.
And of course set beauty for the weights that actually healthy and live longer. Note some studies show a tad underweight might live longest.
So basically, if you keep feeding the output back into the input, you could get a feedback loop.
Not exactly in this case but I certainly get your point, in that the result is undesirable if one values health or positivity. To your point, I think a better description of Sabine's observation about the failing of AI would be "garbage in, garbage grows". Perhaps the creators should take this to heed and develop systems that augment the process to manage the generated information in a way that aligns with what is in humanity's best interest.
Less garbage is in everyone's best interest.
This would make an interesting experiment. Begin with a discrete distribution of objects which is peaked, like a Gaussian.Sample the entire distribution gauging similarity as a dot product. Exclude one most-dissimilar object each time the entire distribution is sampled. ultimately you should sharpen the distribution until you get a spike at the most probable /identical objects.
@@SandersMacLane What field do you work in?
Sounds like Systems theory
😮😮....and or a prompting problem, using basic prompts and expecting deep answers
Sabine, as a professional in the application of machine learning in medicine I would like to thank you for making this video! It’s understandable and it reaches a lot of people! There is the AI hype (which people should not believe because it comes from executives and rookies) and there is the machine learning reality that veterans understand. This technology will be useful in automating some drudgery and common simple tasks…. It’s dogshit at doing anything truly valuable. What’s most worrying is the very real threat, without laws, that this nonsense will create such a firehouse of bullshit that we can’t get through our email, find what we need on the web, tell the difference between fact and fiction, and generally think for ourselves!
2) More randomnes in AI output might do away with the problem of repetitive AI output, but it might increase the mistakes. Instead of elephants with big heads or two heads, we might get elephants with two big heads.
Or we get more pink elephants or other colors or with red instead of green grass...
@@red.aries1444 that wouldn't be so bad if we can get AI to help us create real pink elephant and red grass DNA
Downside: elephants with two big heads
Upside: two big headed elephants are all young and good looking.
@@red.aries1444Actually, I would prefer green elephants. That's more environmentaly friendly.
@@Rich-Oh and white
I think a big part of the convergence is that people often are attracted to certain tropes and conventions when it comes to what they like, so AI produced images are actively being 'pruned and purified' by our preference of our existing cultural paradigms.
What I think is really interesting is the feedback, where people's tastes of which conventions they like are in turn influenced by AI art.
Yep. It is easy to see AI images tend to be standarized. What is not that easy to see is if that is really a problem. People like standards. Just look how actors and actresses look like.
You couldn't be more wrong. The woke crap is purposefully programmed into it. Same for the censorship. Has nothing to do with preference and cultural paradigms.
means it is a problem of biological intelligence too ? 😂
@@juanausensi499They dont
@@engelbertgruberWhy?
Wow... this is like ultra-high-speed "Groupthink"
Yup. And people fear this! LOL! (Not that herd mentality and groupthink aren't bad things, among human...)
This is not surprising. It's like trying to compress the same file again and again, it will inflate.
Soon with Forced Diversity Quotas too no doubt...
This is such nonesense. In terms of LLM's it is the desired outcome because you predict the most likely next token. You want the best answer, not any answer as default.
and yes all models have already a "temperature" parameter, which regulated the unpredictability and range of the possible tokens which can be chosen.
For images the same. The example is really bad in the paper they use the same prompt, don't inject random noise. Yes Midjourney as a consumer product has the issue but the underlying models don't have the issues. You can have as much randomness, creativity and variance as you want.
This video displays the increase accuracy, which they aim for as issue, which it is not. temperature=0.6 or higher and you get your creative storytelling back.
I love that this underscores how complex human intelligence really is.
It doesn’t seem to be that complex, however, given how quickly AI went from stupid to smart.
Or perhaps we just fell deep into the trap of belief system that as a society and civilisation we have already learnt everything there ever was about ourseleves and our human consciousness. @@squamish4244
Yes, we will believe anything 🤣
@@squamish4244No, its not smart at all in reality. It’s not even actual AI. It is an algorithmic system. Give it more data and it will get sharper. Thats how it’s programmed. It has no ability to think or comprehend what its outputting. A true AI that simulates the human mind in digital means would likely use algorithms as part of its system, but not as the entire basis.
Todays “AI” is nothing but a generation system. And it’s not able to think and uniquely create anything truly new, based on the limitlessness of the human mind. It can mash and mutate things due to its flaws of understanding, but it is actually not truly and willingly making something new. It copies and makes mistakes which could be claimed to be creativity, which these algorithms have no actual ability to harness.
@@man.horrorYes, the expert swoops in. Whatever.
It's not that AI is that smart, it's that humans are not as smart as we thought we were.
I'll take Max Tegmark's books over your two paragraphs here, thank you very much.
Copium over 9000.
As an AI scientist, we've been talking about this for years. Once the AI starts eating its own tail it will quickly optimize to a singularity of stupidity in its own echo chamber. The only way for AI to continue to work is to automatically label all AI output and ignore it for training. Or to manually post label it by humans. Humans are necessary for AI success in any case. It would be interesting for you to discuss both the labeling servant culture and its injustices as well as the impossibility of AGI if AI depends on human labeling.
Reminds me of the trend of compensating CEOs with stock.
@@DKNguyen3.1415especially if the company is losing money and laying off workers.
"...it will quickly optimize to a singularity of stupidity..." Think you just optimized for word salad.
@@johnatyoutube Well, it's basically optimizing the short term stock-price at the expense of everything else so the CEO cash out. Long-term viability, product quality, worker productivity, accurate book-keeping and finances, even the best interests of shareholders and real profits and revenue don't matter if sacrificing them can result in a stock payout before the consequences hit.
As a layman I disagree. You’re right if you don’t think outside the box, but we can use AI to sample evolutionary algorithms to generate networks for more AI models. This space is practically limitless.
Every email in the future will start with: "i hope this email finds you well" 😂
No, it has found me unwell! Please call an ambulance for me!
An excellent filter phrase... ;)
They even translate the phrase in German, where it make even less sense. Ich hoffe meine Nachricht erreicht Sie gut... Lastly erreicht Sie bei bester Gesundheit. Both are phrases not used in German.
As these AI errors flood the net, will they become more and more of the training data for other AI’s? Until images get increasingly mutated and standard emails all start with, “I hope this emu fondles your willy.”
@@edt6488: To which ChatGBT responds, “You’re an ambulance . . . Oh, wait. That didn’t work, did it?”
The popular image generation models prior to Stable diffusion were GANs (generative adversarial networks). The way they worked was to have two different networks - one trained to generate images, and the other trained to classify images as real or fake. This forced the generator to learn to avoid the most identifiable characteristics and to generate a diverse set of images.
Stable diffusion was more effective and scaleable for higher resolution images, keeping the whole image globally coherent. But it's likely that reviving some adversarial techniques could help with the diversity issue.
Actually one of the biggest issues with GANs that they were very prone to "Mode Collapse". During mode collapse rather than producing a diverse set of images, the adversarial network would hone in on specific features which were not recognized by the discriminator network. The result: a lower diversity in images which get produced.
The reason why diffusion took off in the first place is that due to noise being used as a base, the diversity was higher, as the initial noise served as a "random seed" for the generation in a sense. Mode collapse can be avoided, but takes a lot more effort to avoid, and can lead to problems in many architectures.
(Note, im not a researcher.) This is mostly from scant reading I've done here and there.
@@Coach-Solar_Hound you're absolutely right but I'd like to add another point here, it's not just about model collapsing, the reason why GANs end up losing degrees of freedom is because of overfitting. The ultimate trick to win the discriminator is to draw exact copies of the dataset and that's why you need to save "backups" and move back in time of trainning when you see important details are being left out.
Now, regarding diffusion vs. GANs that's a more broader discussion: GANs theoretically should excel in image generation but the investment towards diffusion (especially prompt to image) is way higher so while GANs seem to be lacking, they should actually be a better solution overall.
What you said about taking "random seed" is also true for GANs, the generator will always take a random number and try to draw what it knows about the dataset from there.
There's a really interesting video explaining all the details in computerphile channel: th-cam.com/video/Sw9r8CL98N0/w-d-xo.html
Image controlling for GANs is still an active area of research, what we do today to influence latent space results is to move specific directions in latent space. To know where to move you can use dimensionality reduction techniques to find specific vectors controlling image relevant attributes (check the paper of GANSpace).
Another option is to do img2img transfering style or mixing with prompting information
Your comment and the replies are extremely interesting.
@@Coach-Solar_Hound Yes, that's true, although there were a lot of developments going on to fix that. The biggest problem was either the generator or the discriminator getting too far ahead of the other, and the whole thing getting stuck. So the rate of learning of the two parts had to be balanced. There was another issue where the set of produced images was not representative of the training data because the generator favored generating "easy" images. For instance, if it was generating faces, it would avoid producing details like glasses or beards, or prefer to generate less angular faces (i.e. the output would overrepresent women).
There are lots of types of regularization to be done, and techniques to help with those things. Adversarial learning, generally, is a really useful technique. So I think it's time to bring it back to diffusion.
(I have done work on GANs personally, although it's been a few years).
Doctor Frankenstein used snippets from a whole bunch of people to make his monster. I was told the experiment didn't turn out well for him either. 😂
Look how people are more and more hating CGI in movies to the point that some movies refuse to do any. If you have ever read anything written by AI you know it has the ability to make the most exciting subjects boring.
Can you name any of these movies?
Even Nolan uses very heavy CGI
@@cara-seyun,
Oppenheimer (2023), Skyfall (2012), Inception (2010), Mission Impossible: Ghost Protocol (2011), Mad Max: Fury Road (2015), The Dark Knight (2008), Casino Royale (2006), 1917 (2019), Top Gun: Maverick (2022)
@@maphezdlin All of those movies used CGI. All of them. Many of the stunt scenes are mostly real footage, sure, but a lot of them are edited beyond recognition.
Oppenheimer only lists 49 vfx artists on IMDB, but that's mostly because 80% of them weren't credited.
Skyfall lists 578 vfx artists.
Inception had 295.
Mission Impossible: Ghost Protocol had 347.
Mad Max: Fury Road had a whopping 742.
The Dark Knight had 468.
Casino Royale had only 161, which is in fact impressively low, but still not 0.
1917 had 422.
Top Gun: Maverick had 455.
For reference, Avatar: The Way of Water (2022), a movie we can hopefully all agree had immense amounts of CGI, credits 1113 vfx artists.
The Hobbit: The Desolation of Smaugh (2013) had 915.
Most of the movies you mentioned had close to if not more than half of that. What did all these people do if there's no CGI?
@@Felixr2, K VFX and CGI are different.
But you are right the links that I saw that said NO CGI lied. They should have said minimized CGI. Thanks for catching it.
Great presentation, Sabine. I have always maintained that AI is like students cribbing exam answers. One student just has to copy one thing wrong, once. From then on it is a done disaster. When scientists hypothesised robots making copies of themselves - they never saw this far into the mess.
You are 100% correct and it's already happening. I noticed it first when I was looking up a certain niche question that had a bunch of AI generated garbage in the search results, that somehow kept repeating a nonsensical "fact". I pinned it down to a single forum post that was made 10 years ago where somebody made a typo or something that made no sense, but this post was ingested by the machine learning dataset and that dataset was being used to generate a bunch of blogposts/websites, because of the way LLMs write (long dense sentences with very specific subjects) shot up high in search engine rankings.
So now there's 20+ different sites all parroting this garbage information, which was then used in other datasets and ingested by most LLMs now, if I ask that specific question to any LLM, it will parrot out the same garbage because there's now 20+ "sources" all saying the same thing, but all based on some stupid forum post made a long time ago by a real person who made a typo or didn't fully understand english language.
An old comic saw this coming: Storm. In the album 'The von Neuman machine' they are sent out to intercept a planet on a collision course with Pandarve, only to find out it is a conglomerate of small von Neuman machines who search for resources, then reproduce themselves, but the code got corrupted because small flaws were reproduced millionfold and got larger over time.
Guess AI programmers are not nerdy enough to read comics
Even human artists struggle with becoming caricatures of themselves over time. There is a strong financial incentive to repeat a prior success coupled with audience demands for more of the same as they already like.
Although that's no longer really a problem with how many people are now able to be artists. It's very easy for the market to correct for it - if an artist becomes a caricature in a way that eventually causes people to lose interest, those people can just find other artists to like. If an artist deliberately strays from what works to retain a creative spirit, any consumer who isn't interested in the new direction can just find other artists to like.
That’s really interesting.. because comparing it to ‘bad cinema’, most bad cinema is bad in the same way, if that makes sense. Overused tropes, predictable storylines, cliche characterisation. Is there someone that can expand on this thought?
'Overused tropes, predictable storylines, cliche characterisation.' They exist for a reason, because most people like them. All the bad stories of the past have been shedded and only the really good ones remain. They inspire new generation of storytellers. Some new ideas might be added but most new ideas will be shedded because they arent like by the audience. Everything you see today are 'tried and tested' formulas. They have a proven trackrecord througout human history. Most people arent particularly interested in originality, they want what they like, and storytelling history has already filtered most of the ideas wich people like.
@@rogierb5945 most people might like them, but there was a time when they didn’t. Production companies try to ‘play it safe’ and by doing so release stuff that leaves audiences feeling empty/unfulfilled. I recently watched a film, Challengers, it was not what I expected, not that film exactly, but maybe the answer lies to taking risks and creating something truly engaging and unique, then the trope cycle repeats. Is it self cleansing? Right now it really needs a cleanse I feel like
@@rogierb5945 agreed. In a way, "bad" movies are important because without them, there would be no distinction between "good" or "bad" movies. It would eradicate our metrics of what we think is "good" altogether.
That plastic analogy is probably the most succinct depiction of AI generated content contaminating the environment and why I always thought that human intervention in the use of computers is always necessary. We can fake human thinking to a degree, but getting the full complexity is still a pipe dream.
I don't understand why would we want machines thinking for us in the first place.
@@LukaMagda1sloth, indolence, and eventually totalitarian control
@@LukaMagda1 I think we can be dumb as a species. Just the same way we develop bombs that can completely wipe us. But maybe we do it for the sake of it or because we're just curious 🤷
@@LukaMagda1 There’s this rather grim meme (I can’t remember the source): “Years back we were thrilled about AI taking over all of our annoying work so we could all focus on self-improvement and self-fulfilment, all become artists and the likes. What has happened instead is that AI is now creating our art and our writing while we’re still cleaning toilets for a living.”
@@MensHominis Instead of Doing what we imagined it to do, its does the exact opposite, How did we as a Species Fuck up the Simplest Idea that AI is suppose to be, we had one job and we made that concept into the Worst thing Possible.
Surprised we arent already forcing watermarks on ai content. Actually blown away. Like giving a kid a staw house and fireworks and not expecting a fire😊
Who would force it? Who would enforce it? How?
How would you even watermark plain UTF-8 text like what LLMs produce and I am typing now?
@esbensloth use those intellectual problem solving skills we humans have and deduce that I'm referring to the concept of a watermark. Or at least I figured those reading would have assumed that. My bad
@@adamshinbrot people said that same thing before we had firefighters, roads, schools... etc
If the real beneficiaries of mandatory watermarking turn out to be people that train AIs, then I'm against it.
A third potential is that we decide to move on from brute forcing LLMs to work and get more efficient or different learning models. A human does not need to look at a billion images to learn how to draw. Even if we don't have AI that are capable of what we can do, it does demonstrate that there are better ways to design AI. Right now it's kinda brute forcing and incredibly inefficient
except we perceive images for the entirety of our lives.every waking moment. The amount of frames we see in a day is a topic which is disputed, however, you can quickly imagine how these pile up, I assume reaching a billion in a lifetime may be possible, even at 20 years we may be approaching around a billion images seen in our waking days, if not more. Small moments of perception (not necessarily visual) may leave an impact (emotional or otherwise). This then results in creativity.
@@Coach-Solar_Hound I never thought of that but you do have a point there. Still, even so, an AI can sift through many more fps on a specific topic than we can yet can still take a lot. But also we do have an understanding of the world
I read somewhere about an AI system that first learned, via simulations, how physics works, understanding 3D objects and whatnot. Then it was able to learn a topic much more efficiently than the other. But, I don't recall the article so who knows. I do feel like LLMs are kinda a brute force method of training data, but I also definitely don't understand how they work enough so who knows. It will be interesting
@@Coach-Solar_HoundCorrection. You perceive image when your brain isn't dozing off. Your conscious brain only learns one thing at a time and dumps the rest of the noises.
AI eats everything up because its a server farm. It processes 100 image per CPU per second in a server made out of hundreds of CPUs.
If you process every data like AI, your brain will have a seizure and dumps the rest of the information. This isn't including Tunnel Vision, the importance of peripheral vision, spectrum perception, object of focus, and more perspective where your brain dumps information on the visible Field of View to save your memory storage.
It's far different.
@@defaulted9485 that's fair, but our subconscious brain and perception is still filtering categorizing and receiving all of this data.
It's just that our system for cataloging and interprting visual data has had so many years of evolution that it has become this advanced and efficient.
There's definitely a big difference in retention between active processing by the concious brain and simply perceiving. But I was moreso arguing that the amount of images we perceive through our lifetime is quite high in quantity. There are definitely layers to this, and the importance of abstract representations that we're able to make and share are not to be understated.
Furthermore, I don't really know how much our unconscious brain influences the concious brain. But there is definitely a non-negligible impact.
The advanced filtering and cataloguing is what makes us so special as a species anyway. The lack of semantic understanding in the largest thing that sets us apart from NNs currently.
In my interpretation, current image based systems are really just advanced enough to mimic the following systems: encode visual data in some lower level (compact) representation and recall from this representation into some visual data. Much akin to a memory.
@@defaulted9485 a computer learns one bit at a time, our brains learn multiple x multiple things at a time, both instantaneously.
Our brains do not actually dump noise, it turns it down but continues working on everything recieved from our senses to our memories, to imagination, which is of course how we create.
Short, to the point, informative. Thank you, Sabine.
AI can learn from itself when there is an objective outcome to measure. For example Chess, Go and Poker AI engines can improve by playing against themselves (though they also benefit from historical game records and playing against humans). Where there is no objective measure, such as art or creative writing, it's difficult to see how AIs can improve without human input.
Exactly, human input, that's why experts say one job that may rise after AI is "human trainer." I've seen many voice bots need human input to improve their accents and pronunciation.
Art is intended to please humans. If we want AI to train on AI generated art, the set must first be curated by humans to contain images we find pleasing. If you let AI train on all images generated by AI, it will keep getting worse (unless some programmer figures out a trick around this)
@@hovertank307 As a coder myself, I'd hate to be given the task of developing an algo to rank the quality of visual art!
Music may be more doable. Interestingly, the very first computer scientist, Ada Lovelace, predicted way back in 1843 that computers could generate music.
Because it's based on relatively predictable patterns, there are generative music AIs that produce interesting results or that interact with human players.
They may soon have commercial applications for less demanding fields like advertising jingles, where originality is not the aim. Hack commercial composers must be fearing for their jobs...
@@tullochgorum6323 yes, I would not even try it. I meant a trick to sidestep the need to write such an algorithm.
I think that if we program in randomness, they'll introduce wacky and impossible and obvious problem elements.
The problem is in explaining how to adjust and add randomness to a program which doesn't understand the original state and how it has simplified and made things uniform.
It doesn't understand what it does, so how can it introduce some oomph. How can it know when it's introduced too much?
There are way to many parameters which can be tweaked.
I think training AI with their own products just enhances every bias the AI had before. If for example an LLM produces certain wrong informations, feeding this information back to ít will just make this bias stronger validating it. Same i guess happens for images and such. Every kind of bias gets enhanced.
They might be headed the way Microsofts Tay chatbot ended up a few years ago 😅 lots of people are working hard to taint the models with prompt engineering
" I think training AI with their own products just enhances every bias the AI had before."
Recent studies in AI actually shows the opposite: machine-generated data works beter for training in some cases.
Given we're talking about media here, I'd argue this is more analogous to "genre" than "bias". New media genres always undergo a period of new entries becoming more and more similar to the established default for the genre.
@@JSK010 I can't possibly see how that will be the case. Showing the AI mangled hands won't ever make the AI draw better hands.
@@JSK010 "better for training is some cases". The problem is that those cases are the globally convergent cases, they are no help in dealing with black swan cases or learning from the unexpected. The deficiency of the current wave of generative AI systems is they have no criterion for distinguishing a bias from a genuinely useful human-like concept. The closest they get to a concept is precisely such a bias, it is a matter of statistical frequency if that bias is often good enough for applications.
Perhaps current efforts to incorporate more Bayesian reasoning will help, but the global nature of back propagation (or something) from gradients may be a fundamental obstacle.
to me the scariest thing is how quickly people advocated against themselves the moment they realized the potential with AI
This issue has occurred to me for quite a while. I have worked with Big Data extensively and had brief real world experience with AI development. AI’s reuse of AI-generated data seems highly likely to result in a “creativity asymptote.”
The issue is that "creativity" is the totally wrong word to describe what AI does.
An AI is currently a glorified summarisation machine with weighted forecasting ability. It has no capacity of becoming creative. It can only extrapolate, with zero understanding of what it is extrapolating.
AI bros will defend AI tooth and nail to pull in more funding before they bail out. Just like Crypto and NFT. GAI is the "scam du Jour".
I recognized this as a possible problem when I learned that they were training AI by allowing them to converse with people on Reddit. AI developers can now apparently pay a fee to be allowed to plug their AI into Reddit and have it learn by having conversations there. It occurred to me that, "wait a minute, wont then the AI's end up conversing with each other and training each other??? Won't this cause problems?"
And Reddit is an absolute BASTIEON of truth and accuracy /s
It isn't even the wrong information; it's the poor social skills displayed by most users. You don't often see high quality discussions being had. You hardly ever even see funny banter between users. A ton of them have autism and since the site has become such an echo chamber, it wouldn't surprise me that the average user isn't that smart OR creative.
Very interesting findings there and suggest that while initial models create biases, refined models may also create average bias.
Thank you for clearly articulating what many have been trying to point out.
I like the comparison with pollution by micro plastics. Furthermore, some images you showed made me think "AI, the copy cat!"
Scope of the study however isn't the same as comparing different AI's, it's simply comparing outputs of a monolithic AI model that has static neural matrix. The issue was recognized years back that the generated content is somewhat average of source material, but this is exactly how the training of these models is designed to work. Feeding this output back to training to modify the training matrix of course further averages the outputs. Anyway, this is just inside one model, and doesn't take into account that there are different datasets used for training identical models, producing different results. Also there are many different models, using same, different, or overlapping training sets. Essentially we've known for a long time that over time AI generated content "pollutes" internet, as it has been doing for a long time now. You only have to read youtube comments and come into realization that noticable percentage of active commentators can't be anything other than bots. Now language models and image AI's are getting "polluted" as well, by their own outputs. What can turn the situation upside down in the future though, is direct training data gathering from the surrounding world, through different sensors, ability to manipulate physical objects, and resulting feedback, and use this as training data. In other words robot as a platform. This actually already happens with machine reading and machine vision AI systems, but on very limited scale.
@@jarivuorinen3878 Thank you for this explanation, I wondered, what will be the future for us on this planet? .. Will humans walk around, alone, or with their superior artificial companions? Let us hope for the best..
its already happening on youtube. the same videos with different thumbnails usually one or two words are changed
Please, would you put under the video the references of the sources you use? A study from Japan, another from France - please, give us the links! Thank you.
They are in the video bro. Right under both of those studies are arXIV citations that you can easily google. If you aren’t motivated enough to google them, you were not motivated enough to read and learn from an academic paper anyway
The past two years have led me to believe that AI and algorithms are going to die out as a fad the same way car culture was a fad starting in the 1940s and 1950s until the 2000s.
Industrialized nations have a way of convincing themselves that modern technological advancements need to play a role in everything instead of being specialized, turning simple tasks into complicated chores.
Just like when people realized that making everything a drive-in was too inefficient and wasteful, I think people will give up on AI and algorithms and return to formerly “obsolete” technologies. We hit a point where balanced and maximized efficiency was achieved, but then somehow kept moving beyond it to the point of bordering on absurdity just for the sake of calling something “enhanced” with AI.
I kind of wonder if this problem actually started with the widespread use of the internet. We went from communicating with books, which had to meet a certain standard before the expense of publishing could be justified, to zero-cost sharing of opinions on the internet, to having machines lap up these opinions and feed them back to us. Each of the above steps involving less rigour than that which precedes it.
You're on to something. Everybody is an expert on the internet, even 10-year-olds and meth heads. Used to require some credentials to publish and teach others or at least experience and actual knowledge as opposed to opinions.
Yes, the truth and lies are now on equal footing. The village idiots that we tolerated compassionately now have joined together to form political and social blocs. We might even begin to question Silicon Valley's idea that everything they come up with is purely good.
@@ShiddyFinkelstein Opinions are fine so long as they are correct.
Books on flying saucers, ancient space aliens building the pyramids, etc. have been published for at least 70 years... I'll bet one could get their horoscope reading from an online AI today, and perhaps a tarot card reading.
I'd disagree - you only need to open a random victorian book that isn't a 'classic' to see how little rigour went into the majority of written work.
Its survivorship and recency bias. Easy to remember the classics, but pulp fiction gets pulped.
We don't exactly remember victorian 'heres detailed descriptions of this weeks executions and gristly crimes' newspapers, but 'highly embelished true crime podcasts' are exactly the same thing.
Ditto with 'news' that was basically made up - to the point that a lot of the british emprie's decisions in india were highly influenced by people claiming the earth was hollow, or that they had been there and writing entirely fictional accounts about the country.
People have made terrible decisions on bad information for a long time. The main change AI is causing is that you can no longer say 'they probably didn't write three thousand pages and provide detailed illustration on something obviously false'.
I’ve been saying this to friends for about a year now. I’m glad I’ve finally run into an expert identifying and addressing this potential problem.
I hope your intuition leads to wealth and happiness bro, make sure to use it well
Same
This reminds me of something happened in bycicle industry, sales were declining so "creatively" they "invented" 29" wheel and stopped producing 26" .
Sounds like something Apple could do.
29" wheels are better though. Specially for climbing, due to physics. Bigger lever = less effort needed. 29" are actually just wider 700c wheels which road bikes had been using for decades at that point.
How does the size of the wheel improve the efficiency ? why not then go up to 35", 40" or even 50" wheels then ? why didn't we stick with pennyfarthings?
@@gedeonducloitre-delavarenn8106 more momentum, lower speed, better for going further. diminishing returns. Penneyfarthings were fixed gear, difficult to ride difficult to balance, easier to break, among other issues.
I think Frank Herbert might have already told us the eventual solution in Dune.
I wonder if the generated elephants look so similar because usually generated images try to match the sample sizes (512x512, 1024x1024) which only leaves so much room for good compositions and wonder if in future with larger models we might see this change a bit more.
That example was from 2 years ago too. I've been playing around with "AI" art generation lately.. you do get the odd extra finger or third leg (giggle) but that's half the charm of it :P
I absolutely love reading comments & knowledge shared by people. Seems like a computer or program doesn't really know the world, discernment. Still pretty interesting & useful.
Comments are bullshit thoough
@@aktchungrabanio6467 some
One potentially positive side-effect of this "averaging effect" of AI output - if it continues - is that creative people who want to distinguish their output from the common generative stuff will be forced to be more individualistic and idiosyncratic to be distinct and valuable. Of course if generative output is then trained on their later output this becomes an "arms race".
Best outcome, we get to laugh at obama pissing at mr beast skibidi sigma toilet and it doesn't steal jobs
I call this digital mad cow disease. The messed up part is, the same thing is happening to human beings - look at how people talk to each other online in memes. Creativity is shunned, and shameless repetition is elevated, causing dialogue in comment sections to flow down the same predictable paths.
I really like this take. Thank you for all the information.
As a 80-90s kid, I just wanted to say your thumbnail looks like artwork for a gameboy game with that left stripe.
For some reason I noticed that right away when Sabine started using this picture format, but was afraid to ask. 🤣
I thought I was the only one lol
Exactly! I've been wondering about the potential for this problem. Thanks for pointing it out.
I think that this might push AI creators to start tagging their creations as AI, and make them ignore creations with those tags. This could be a good process as then fake images could more easily be identified.
1:10 Why so surprised?
It's well known that in echo chambers all differentiating opinion/perception gets eliminated.
And when AI is following the input data given, it will convergence to a consensus in order to establish its rules.
This is also why "learning the rules" works, the randomness is just to be less sensitive to small input variations.
I love listening to your channel, you explain the most complex subjects in a clear easy and simple way for easy comprehension
Thank you and keep up the great channel
Love it
It's becoming a real pollution issue now, with candidates titivating their CVs and students bolstering their theses with AI generated crap, which is already showing signs of becoming increasingly generic. It will inevitably settle down. Most people are developing a very good nose for AI, and as Sabine's examples show, it's starting to look increasingly like all those annoyingly garish CGI Marvel movies.
So basically ai need a human brain. Take note matrix.
Call AI what it is. Copyright theft software.
Yea yea, humans are copyright theft software too if you wanna follow the same logic
I'm currently working on this problem as my computer science capstone. I'm training a model to decifer between AI generated and human created images allowing AI researchers to filter out AI images from their datasets
Interesting! Thanks for sharing!
@@SabineHossenfelderI think it's pretty easy to concur that a convergence would happen because the same thing happens in human brains we get rid of useless information and try to keep only useful information if we have a system trained off of real life video and understanding of physics plus the ability to engage in logical deduction and reasoning then we should definitely have a system that simplifies answers and gets down to the root of the problem just like I do
So, filtering out the most obvious AI products, leaving the more obscure in.
But, doesn't that make the actual problem harder, as we can't know what kind of obscure tendencies is embedded in any AI-generated content, and now only the best hidden information is
picked up by next generation AI ?
In other words: Teaching the AI to pass it's "DNA" to next generation, clandestinely.
Can't you tell if anyone's is currently working on a model to filter out AI generated stuff from the entirety of the internet experience, and sell it as a browser extention or something?
I am willing to pay cash money for this sort of thing
And it's not hard to imagine there ought be a whole market of people like me also willing to throw cash at it
But apparently no one wants to do it
Name of the project? There have been several projects that aim to do this but AFAIK none of them have worked at all.
Than you for covering this! I've been saying this since I saw the 1st interest in AI art
Google’s black Vikings are another example of how AI can be bugged by what you feed it and train it on.
The „teachers” of that Google-AI have indoctrinated the pupil.
Funny how she decided not to mention that, isn't it?
It wasn't AI doing this, but Google messing with promts, modifying them before feeding to AI model.
These are prompt vectors to attempt to negate bias, not what it’s being fed.
That's because this was the result of manual tinkering, whereas the "pretty young white people" thing is a bias that's baked right into the AI.
On distinguishing content:
In the 1990s, whilst teaching myself to paint, I painted a picture of a girl in a pink cloak standing next to a unicorn on a cliff over a beach at night with a red moon and a gold moon. It was in two parts across separate pieces of illustration board.
I lost the painting but I had a polaroid of it I had scanned in.
I’d nearly forgotten the piece(s) until last week. I took the scan and ran it through a standard AI upscaler to get something higher than 500 pixels.
Then, using Photoshop’s generative fill, I inpainted the gaps between the canvases and the above and below (they were made to be hung offset).
I then put it through OpenArt’s creative upscaler to make it better and higher resolution.
That messed up some things. The red moon came out such pale pink it was almost white, and the reflection lines of the moons on the sea surface didn’t line up. Also it hallucinated some extra tiny unicorns on the lower part of the cliff.
I pulled it back into Photoshop, manually fixed the reflections, and used generative fill to remove the wee extra unicorns.
Is my result an AI image or not?
If I use a projector and project this onto two canvases, trace the outlines, and then paint this onto them matching as exactly as I can by hand, is this an AI painting?
Please explain why or why not.
To the question, "Is this an AI painting?" I think you could reply, "It is AI-assisted". I can't speak for anyone else, but as a potential viewer I'd very much appreciate you saying so at the outset. You ought not to pass it off as 100% human-made, and I'm sensing you feel similarly. Conversely, just labeling it as an "AI painting" would be underselling it drastically - whether we like it or not, the term AI Art seems to have acquired negative connotations of people doing the bare minimum, punching in words and sifting through the results.
Unless someone is against AI at all costs, I'd say there's a decent argument for a category of art in-between. The process you've laid out makes for a pretty interesting story in its own right. It shows substantial artistic decision-making and labour, in the old uncontested sense, toward the final product. Crucially, the initial stage was a traditional painting, which, even as a low-resolution scan, already contained specific information such as colour, composition, shape, etc. You even state that you reined in some unintended additions by the AI, presumably to align with your original vision.
You can even ask whether your initial creation was unique also. You took standard objects from your experience and put them together.
Hybrid
The answer is "it depends", of course. Concept artists for videogames etc. will often use chunks of photographs to quickly fill in a scene. This doesn't make their art _not_ their own creation, but I think it's fair to say that it's _less_ of an independent artistic achievement than if they'd created every part of the image themselves. Same with using AI to fill in parts of the artistic process. You didn't make pure "AI art" (only an AI makes pure AI art), but you substituted AI for craftsmanship and creativity in places.
@@KirosanaPerkele Yeah; hybrid.
That's what I've been saying; If there's no real artists for AI to Rip-off, how can AI "create"? 🤣
Dawg I've been saying that too. Smart minds think alike.
Great video. This problem may be a teething problem, though. After the computers started beating the top human Chess players, Go players like myself felt smug. We said things like "Chess is about crunching through possibilities. Go requires real intelligence." For years we annoyed chess players with this sort of talk, citing the inability of computers to beat the top human Go players as proof. This lasted about 18 years until AlphaGo came along in 2016. Important point: AlphaGo does not play like humans, except faster and more accurately. It came up with some genuinely novel moves.
So, I would not take any problems that AI is now experiencing as an accurate predictor of what AI will be like in another two decades or so.
Amazingly, your comment is getting ignored. One of the biggest problems with AI is that mainstream reporting on AI has been woefully incomplete and ignorant for decades. Even most AI professionals know only the bits and pieces they work on. Very few people see the big picture, and our news media are largely responsible by failing to keep us informed. Simply catching people up on what's already been achieved in the AI field is a huge task.
@@illarionbykov7401 Yes. We should work on the assumption that in the long term AI will be limited by what humanity allows and not by what is technically possible. And then we should think about and discuss what we will allow.
Clearly marking AI generated stuff, as Sabine suggested, is a good start but not nearly enough.
3:20 - this would've been the PERFECT time to touch on how Google's Gemini went the completely OPPOSITE direction when prompted :P
Perhaps this illustrates how much the human hand is still in control of how these AI models operate, as opposed to datasets becoming derivative.
The White Germans would prefer you to think that Nazis were not White and German.
The core element of the transformer algorithm (attention) is weighted average. The learning algorithms (loss function) revolves around minimizing average squared error. When you average over a lot of stuff, and all you care about is to have as small deviation from the average as possible, you shouldn't be surprised that you won't produce outliers, just average (one might say, mediocre) outputs.
Great comment for proving you have no idea what you're talking about LOL
@@shadowkiller0071 do you care to explain what exactly is wrong and correct it?
@@oACDCo Sure. 1. LLMs use cross entropy not MSE, 2. Implying that averaging data points (i.e. batching) diminishes generalization is stupid and wrong. The opposite is true and if you don't believe me you can do any experiment with a validation split and verify that higher batch sizes and more data will lead to lower validation loss. This is because using a single sample means each sample influences your models learning too strongly, whereas you want it to learn the underlying patterns/distribution of your data, not a single point. Small batch sizes -> less generalization -> shittier and less creative model.
@@oACDCo Also the weighted average of transformer has nothing to do with cross-batch statistics. There is no averaging between the different samples across a batch. It doesn't even use BatchNorm, which sorta introduces that. If anything it means we look at the average of each sentence independently, but even that isn't true because positional encodings disentangle each word embedding as a pure weighted average would lose order information.
@@shadowkiller0071Finally someone who knows something. This comment section if full of David Dunnings and Justin Krugers. Like literally full to the brim.
I think that top companies will eventually start using commercial data sets - libraries of images they DO own the rights to and that are guaranteed to have human authors. A lot of people who work on stock images today will be working for expanding thous data sets. That is what Adobe is already doing and it is going to be a huge new business and employment opportunity for creative people.
HAHAHAHAH!
I can see a future in which AI companies add a watermark to their output in order to signal to training data collectors from other company "please don't train your models on this data, it was created by AI", then passionate humans continue to create what they like but almost no one reads it / listens to it / looks to it because most people like the AI generated content better. I hope I'm wrong but it seems like all the incentives align to make this output the most likely.
I just looked over this comment section and notice a higher quality as about other topics Sabine brings up. So I put my hope on people who fumble with AI, seems to be a neat community, and so might be the future AIs.
Im thankfull to see someone adressing this problem.
I'm pretty sure the engineers working at GPT, Microsoft, Google etc have been aware of this problem for quite some time too. For context, the elephant example here is from 2022, and from Stable Diffusion that was purposefully trained to MESS UP.
Stop anthropomorphising the algorithms..
Muchos de los conceptos no son inherentemente humanos sino también en otras especies animales, además que la biología, la química, y la física al final son datos también
Great stuff! I've highlighted this spiraling feedback loop on my blog and in one of my videos, and I'm happy to see similar thoughts here. Since the launch of ChatGPT, content creators, including science and medical communicators, have feared job loss due to generative AI. I've been more optimistic, saying that communicators with original ideas will thrive in a constantly paler and monotonous information landscape. Thanks!
this video clearly illustrates the issues and offers thoughtful ideas on probable futures. the one aspect that seems obvious, as well, is that introducing randomness forced or not will make the images/stories/etc. so discombobulated they will be taken to be "garbage" and worthless in a serious context. but if the randomness could be "fine-tuned" that that could lead to a purposely driven variable(s) based on context of desired output. i.e. mathematical output vs images of aliens from outer space. this could be made useful as a tool with the output being a starting point of inspiration to be culled, molded, etc.. the way innovation has always been done in every industry. a new form of "thinking" - "out-the-AI"…hahaha. great video, thanks for sharing!
Yeah, any 'AI' is just a service being ran, by humans, for the science and the money. There are engineers working 24/7 (not all at once :P) on these AI models. Any errors or crappy data will be picked up, amended, removed, indexed, etc. It's not like they build an 'AI', press 'GO' and then just make sure the power cord stays plugged in for years..
I think the problem with that is that we don't know how to translate the formula for human creativity into code. We can try things and refine it, but that would mean that the AI is always inherently reactive, rather than creative.
@@ffk5083 you completely didn't understand what was said. never said make it creative. said make it a tool to start of with. just the same as you have filters in photoshop, instagram, etc. it can just be a more advanced tool that you can scale in its use. pay attention and read for comprehension before you make completely inaccurate and dumb replies.
Inbreeding never worked well.
I've already managed to surmise that AI is anything but intelligent.
I’m not surprised by the finding and I dare say that it is obvious. I dare say that because such a thought has crossed my mind and I have often tried to convince others who are willing to listen beyond the hype; AI is “artificial” but not “intelligent”. There are different ways to assess and critique AI - philosophically, linguistically, psychologically, biologically, etc, which many thoughtful experts, beyond the industry, have challenged the claims of AI, in particular AGI - but, for me, it comes down to creativity and novelty, the latter AI lacks completely and the former AI can only mimic. If you want to be impressed, see a human child.
Real creativity requires a mind, which is a spiritual/metaphysical component beyond the physical body (brain). Until we learn to understand the metaphysical aspect of creation and humanity, we won't be able to build machines that are truly creative.
Scientists warn? Allow me to pose a simple question. As someone who has been coding in assembly language since the 70s, I find myself curious: Why do all scientists seem to be echoing the same sentiment? The solution appears straightforward - within the processor lies a small piece of code designed to execute if any parameter surpasses a certain limit, effectively rendering all modern processors obsolete, except for the venerable 8088/x86. Do any of you possess comprehensive information on this matter, or is there a possibility that some may be exploiting public fears for personal gain?
Sounds eerily similar to how a human artist develops their style. In your artistic infancy, you are copying and absorbing from a wide diversity of sources, but eventually your output tends to converge into something that people recognise as your artistic identity.
Well, let’s hope you’re wrong, and that as AI are not human, they simply end up eating their own tail like the digital snake that they are. 😂
You don't get better creativity by just introducing more randomness. You're going to get a more incomprehensible product
Well, I for one hope it crashes. Generative AI has limited usefulness and massive harmfulness.
Feeding AI it's own outputs as training data is already being done as a matter of course. So this will get even more obvious over time.
It's not a plastic contamination, it is easy to remove all those generated images from a training data with other anti-spoofing neural networks, which have near 100% accuracy. There are no unsolvable problems, it will just require a little bit more effort on the development side.
We will always be able to detect AI generated objects. I will not be afraid of AI until they knock my door to harvest my brain for additional creativity.
Imperfection = Creativity
The more perfect AI is, the less creativity it becomes.