Go to ground.news/universe to stay fully informed. Subscribe through my link to get 40% off unlimited access this month only. If you're wondering which bits are the "real" me and what's AI, the video is all an AI avatar of me, through the site www.heygen.com. HeyGen actually uses GANs not diffusion, but never mind! The audio is me speaking. However, that audio is then processed using audioenhancer.ai/ which makes a big difference, as you can tell at the end of the video.
Wouldn't have guessed, but knowing and looking for clues, the ai model squints and blinks quite regularly. At one pont there's even a dust particle flying around. I guess only the very last part was real, then? Pretty convincing. And frightening.
I hadn't noticed at all. Perhaps some of that was down to watching at 1.75x, perhaps some of it is being generally unperceptive about things like that, or perhaps the models are already that good!
That was so weird! It looked amazing but it didn't match your voice, except that bit at the very end. Your voice is unique in that it's very audible when you smile and the video didn't quite match. But for any given stretch of a few seconds it's uncanny.
It's very convincing. If I wasn't looking for it, I might not have noticed. Watching carefully, though, when your avatar blinks it sort of looks like your eyelids have glue on them and "stick" a little too long. Also, you're more animated in your body language.
It's amazing how much 'communication' happens non-verbally. If I had to guess, the last clip, where the sound quality changed, is the real LGU, and would be shocked if it was the other way round.
At the end one could understand why her often amazing face expressions were limited in the amount and repeating, maybe likely due to the limitation in the sample for the avatar generator. Glad to see she hasn't changed, it was "just" the avatar which looked incredible perfect
It did not take to long to spot the deep fake, at least on a TV. Blinking, lip movement and some "jerks" gave it away. But damn it's really good already. A bit spooky tbh
Mithuna, please do a follow-up vid on how you did this. The platform choices, the coding, the training data etc. The overall effort and cost. Plus, extrapolating to suggest how good it may become
I remember watching your videos as a high school student who thought I'd study quantum mechanics...here I am again after graduating college with an interest in AI. Full circle moment.
I wasn't watching very closely, but I noticed it was an AI version of you at 8:40 ish because at the end of the word "algorithm" your mouth didn't match at all and I rewatched it because I was confused
Hilarious!! There were points in this vid where I was concerned about your health, lol. At times it looked like you were on the verge of falling asleep. But honestly telling the difference was very much an 'empathetic' or intuitive exercise and not a clear, logical one. In addition to being clever, the clip was immensely informative for someone like me who's not a computer scientist.
Damn that ending, you definitely made your point well. Admttedly I wasn't paying super close attention to the visual element of the video but I didn't notice anything that looked or sounded wrong. It's extremnely concerning to consider how this tech will be misused in the very near future.
Blindingly obvious where the fakery was, even without paying too much attention. The voice and facial animation just didn’t look right, like the lips were glitching. If you can’t see that, that is concerning.
@@X22GJP You know some people don't have good vision right? They even said their attention wavered during the video, as it almost always does when any person watches a video. We are surrounded with distractions all the time. Unless people are prompted, it's completely unfair to say that missing these subtle cues is "concerning".
I've read some explainers about stable diffusion, but this is the first time I understood why it's called "stable diffusion!" I'm sure the cute dogs and robots with accessories helped 🌸
Interestng video, although I think you overstated how easy it is to train AI's - in reality many images have to be labeled manually (often by low paid workers) and also AI does not (in my opinion) 'learn' but instead alters the matrix multiplications (which is effectivly the same thing so I do understand your use of the term). I do hope you continue with the physics stuff, as I (and probably a lot of other subscribers to your channel) find it more interesting, rather than just following the hype of AI.
10:45 --> 10:59 "Not a deep-fake". The rest: look at the eyebrows and lip-syncing. It could be AI, or it could be a jittery camera. Even the "ground news" piece looks slightly off. Good ep, Mithuna!
Very informative video! As for the deep fake, I did notice that there were some instances where you looked like you were speaking but there was no voice. Definitely surprised me hah
love it apart from the fact you refer to de-noising as "fixing" and "recovering" an image in the beginning (0:38 etc.). i'd argue that this is not the case and it is a hallucination only a bit simpler but similar in that humans are not always able to recognize it.
I forgot the thumbnail so I noticed the that something is off half way through, real creepy! It kinda looks like the video is slowed down then sped up. Really clear explanation of the critical concepts. It's interesting how the basic ideas of modern AI are kinda simple yet produce such powerful results. It's becoming a big part of our world so I think it will be important even for people who didn't really want to care to know at least the basics.
Love the video! Just spurred on an hour conversation with our 15yo son. He is hearing towards art, game design, and music. Your video helped explain some nuances. Thank you! 😊
These videos are hard to watch for me as a former arts director. I was already disheartened to see so many ready made templates and stock images beinng used in the industry. Artist have always struggled to find fundings and the capitalist machine found a way to stop paying them for adverts, illustrations, marketing and films. However, I can also see the appeal as a programmer, so I'm really torn. It's helpful to see you talk about it, as I am used to you, having warched your videos for a long time. The familiarity and connexion with curiosity and generosity helps.
That is so cool and really well explained, perfect for the type of brain I have 😊. Kinda mind-blown here just at the idea of using a denoiser on an image of pure noise tbh. Reminded me of how we see images in clouds etc., with the prompts being like our brain's prior expectations of what it is likely to see, including being influenced by what's on our mind at the time, like in Rorschach tests and similar. Intrigued to know if it's possible to "help out" the iteration procedure by using an initial image that isn't pure noise, say an image of a golden blob plus noise to help it "see" the golden retreiver quicker? Sounds like that could help in the creative process, maybe even giving more precise control of where you want certain objects in the image? Don't know lol but consider me inspired 👍
I think this exact idea works! Have you seen those AI models where you do a sketch and it makes an image based on that? Very cool that you were able to guess that something like that would work
@@LookingGlassUniverse Ah, that is cool. I'll have to check them out, sounds fun 🙂. And thanks, good to know my own neural nerwork still works sometimes. I've a background in theoretical physics, bit rusty though. Loved the coding / algorithm design aspects the most and the theory behind them. I've always thought I could get totally lost in the machine learning / AI field, if life would allow lol. It is so fascinating. Time to chase down some of your sources and recommendations. Ta for making your cool vids btw 😉
I did notice that the apparent framerate dropped and the freckles on your nose suddenly popped, so guessing these are the AI generated frames. Great video!
That "clip" algorithm is still wild. I get that it makes image processing more efficient by converting written prompt to multidimensional vector. But how it associates "medium blonde hair dog" to "golden retriever" with no prior knowledge is beyond me, it seems more like "reasoning" than "chance", almost like it looks it up in a dictionary.
It's not "no prior knowledge", it's another neural network component that has to be trained. Once trained, it embeds the implication from the training data that prompts like that go with similar images even though they're not similar as raw strings.
I noticed that the video seemed a bit weird. It wasn't quite consistent enough from from frame to frame, almost like artefacts from a sophisticated compression algorithm. Pretty impressive, though.
AI denoisers are demonstrating remarkable progress in their ability to remove unwanted noise from various types of data, such as images, audio, and video files. This technology is becoming increasingly sophisticated, offering clearer and more precise outputs by filtering out background disturbances that can degrade quality. As these systems are further developed, they are finding widespread applications across diverse fields, ranging from medical imaging, where they help enhance the clarity of diagnostic scans, to entertainment, where they improve the audiovisual quality of movies and games. The potential of AI denoisers continues to expand, promising to revolutionize how we process and analyze data by making it cleaner and more accurate, thus enhancing the overall effectiveness of many technological applications.
Deepfake was only the visual part of you or the audio as well? If both, then the AI did a great job. Then again, I was not paying that much attention and whatever I have seen I attributed to editing (like cutting out short pauses to get a more uniform flow of words - which make some visual "glitches" to which I am used to see even in "real" clips). Maybe almost unnatural? If the AI is trained on such modified data, then it behaves like it. This remains a speculation, as I have no knowledge about how you do/did your videos and what the AI was fed in its training data.
I knew it, there was a repeating paterns of expression that seemed to have been amplified during the AI bit. It curiously pushed me to look at your face in more details. Probably a defensive system already in place in our brain. Althought the job was really well done and pleasant to look at it was emotionaly exacerbated like a loop of carcteristique not temporaly well adjusted as it was more random. I have to recognise though, I love your content and pay attention. If I was not paying that much attention I could have been fooled. It is really dangerously efficient.
Yeah... it was actually slightly creepy knowing I was watching an AI deepfake 'person', but not having been informed that the character was an AI rendition... which felt weirdly disingenuous.
Really helpful, easy (kinda) to understand video. I did wonder at a few points if you were real! I took another look and at about 3:44 the blink looks a little odd. ..reading the description it was all fake... wow...
That's a quite difficult challenge. 🤔 Something like this could become a new party game 😉 I think 4:15 and 5:00 are real footage, and 3:35 and 8:25 are fake, right? It might be much easier for someone who knows you better ... By the way: I heard of an AI that is trained to spot AI-generated content. Such an AI-spotting AI should become an essential browser plugin.
I realise this deep fake technology would be great for animating transitions between frames when you don't want to make a harsh jump cut. Nobody's job is destroyed, but you get better editing quality and better quality of life for the person shooting and for the person presenting.
As a still it's fully convincing IMO. Moving though it's slightly, I dunno, twitchy ? And it reminds me of a 'Star Trek' episode where Data deduces someone's actually an[other] android by noticing that the cadence of their blinks is driven by a Fourier series - there's just something _off_ about the eyes. Even then though, I could've easily assumed e.g. video glitches. Amazing _and_ a bit scary :).
We make art out of the lowest state of entropy which is a blank canvas and AI is making art out of the highest state of entropy which is pure noise. We function fundamentally different.
To be honest, I haven't noticed that you used deepfake at all... well, I had a weird feeling like your movements were a tiny bit unnatural, but brushed it off easily because I was sure it's all real. Dammit
Good video on diffusion, a little distracting with the weird face motions and the sound not fitting the lips but I can live with it for the fun of it. 🙂
Althought I was prompted by the something before the video that said you might be AI, it was really easy to spot. The way the AI blinked, it was almost like it was in reverse. Plus the AI doesnt do the normal facial gestures you or anyone does that well. Its trying but it doesnt look as normal as normal. That being said, if this was a video game Id be saying its the best graphics ever.
i noticed the start to look a bit odd and already thought that it probably was a deepfake but honestly i forgot about it watching midway until you reminded me of it again. crazy technology!
YESSSSSS I called it! Right at the beginning, I said this is odd, would be hilarious if this a deep fake made by Dr. Yoganathan to emphasize capabilities. Terminator ain't gonna get me boys!
If you hadn't said anything in the end, I wouldn't have noticed any AI generation at all. That is so creepy. Even after looking for it, I'm unsure, if I identified it correctly.
That is interesting🤖wow I had the thought why is she apparently using green screen for mundane room background!? Your surprise ending is why!!?🎬🎬🎬🎬🎬🤩♾🌈🤯🥰😘😍🚀🌌🗽💯💯💯🤖🤖🤖…
I was thinking that you were not quite hitting every beat and were coming across a little wooden and unsure of yourself than usual. When you said that most of it was a deepfake that all made sense.