WEBVTT 00:00:00.000 --> 00:00:07.040 dita fender sponsored this video so i can show you that i can juggle computers 00:00:07.040 --> 00:00:13.880 and tell you about our new amazing product Linus strength pills watch as i make david watch all of 2019 00:00:15.000 --> 00:00:17.000 not this 00:00:30.000 --> 00:00:36.160 Thanks to Linus Pills. 00:00:36.160 --> 00:00:42.960 I'm unstoppable. Start your subscription today by emailing me at Linus.Linus.Linus.com. 00:00:42.960 --> 00:00:49.080 I don't remember recording any of this. Bitdefender did sponsor a video about protecting yourself from deep phase though. 00:00:49.080 --> 00:00:52.600 Because now, these have gotten so good that you didn't even realize that I'm not the 00:00:52.600 --> 00:00:55.840 real Linus either. 00:00:55.840 --> 00:01:07.520 What are you guys doing? 00:01:07.520 --> 00:01:13.520 Fully AI generated videos, at least for now, still have some easy to spot tells, especially 00:01:13.520 --> 00:01:20.600 if they're long or if they involve a lot of movement. But a simple head replacement deepfake, or a stationary figure at a desk telling you 00:01:20.600 --> 00:01:23.600 all about their latest get rich quick scheme. 00:01:23.600 --> 00:01:27.280 Those have gotten shockingly convincing over the last five years. 00:01:27.280 --> 00:01:32.840 Let's talk about both how we deepfaked and how we fully generated me for this video using 00:01:32.840 --> 00:01:37.320 just commodity hardware. But before we do that, an important message. 00:01:37.320 --> 00:01:41.280 If you've got a loved one who's going to need some help recognizing what's AI and 00:01:41.280 --> 00:01:44.680 what's not, send them to this timestamp. 00:01:44.680 --> 00:01:50.400 We've got you. Starting with our deepfake, the process is still pretty similar to how we did it last 00:01:50.400 --> 00:01:53.600 time, just five years refined. 00:01:53.600 --> 00:01:56.840 First find an actor with a similar body shape to your target. 00:01:56.840 --> 00:02:01.100 You can see why this matters in this test that we did with Ploof, who owns the display. 00:02:01.100 --> 00:02:05.440 When you blow it up on a TV, it's pretty easy to spot anomalies like his hat or his 00:02:05.440 --> 00:02:12.520 beard leaving a soft outline. But on your phone, especially with poor eyesight, mostly just looks like a guy with a slim face 00:02:12.520 --> 00:02:16.840 on a chunky body. Oh my god, am I allowed to say that? It's okay, I wrote the script. 00:02:16.840 --> 00:02:23.400 So we parsed the subreddit then for the latest is this Linus meme and chose chase for our 00:02:23.400 --> 00:02:31.680 real attempt. We used deepface lab to train a model on about 7000 recent images on my face and boom, much 00:02:31.680 --> 00:02:36.880 more convincing. And this is both super cool and super scary. 00:02:36.880 --> 00:02:41.160 It was at least a hundred times easier than when we did it last time. 00:02:41.160 --> 00:02:45.600 Sinking up the lips with the audio is still pretty tough, but if you keep your clips short 00:02:45.720 --> 00:02:50.400 and you edit them together in a punchy manner with alternate angles and close ups, like 00:02:50.400 --> 00:02:55.040 you typically would, you can splice together something pretty convincing and pretty long, 00:02:55.040 --> 00:02:58.120 one 5-7 second shot at a time. 00:02:58.120 --> 00:03:03.040 With that said, I'm sure that a lot of you guys could still tell, because realistically 00:03:03.040 --> 00:03:06.680 you dwell the same corners of the internet that I do, and you've been watching with 00:03:06.680 --> 00:03:13.040 a sense of awe and with dread as deepfakes and video generation have gone from obviously 00:03:13.040 --> 00:03:17.520 fake garbage to oh, this requires a little bit of scrutiny. 00:03:17.520 --> 00:03:21.520 But you savvy viewer are not my concern. 00:03:21.520 --> 00:03:25.520 Many, maybe even most people, can't tell anymore. 00:03:25.520 --> 00:03:31.200 And online scams and fraud are increasing every year to the point where in 2024 losses 00:03:31.200 --> 00:03:34.760 are estimated at over $1 trillion. 00:03:34.760 --> 00:03:38.800 That number came from Bitdefender, who is in the business of knowing these things. 00:03:38.800 --> 00:03:44.640 And the scariest part is that a huge portion of the scam industry is still using old school 00:03:44.640 --> 00:03:51.200 text to voice phone calls. I mean, imagine if you could hit victims with something more like this. 00:03:51.200 --> 00:03:54.800 Hey, I'm in a jam. Can I borrow $5,000? 00:03:54.800 --> 00:03:59.360 And I know some of you guys are probably thinking, tough break boomers. 00:03:59.360 --> 00:04:05.280 You had enough for the wealth anyway. But that's easy to say until it's your family member. 00:04:05.280 --> 00:04:09.920 We have multiple people in this office who have had family members impacted. 00:04:09.920 --> 00:04:14.880 And again, the scariest part is how easy it was to generate that clip. 00:04:14.880 --> 00:04:20.680 All we needed for it and for the ones at the start of this video were start and end keyframes. 00:04:20.680 --> 00:04:26.040 Once those were captured, or in many cases just scraped off of an existing video or Facebook 00:04:26.040 --> 00:04:31.440 photo. My name is Nicholas Pluff and I hate displays. 00:04:31.440 --> 00:04:36.680 We got an advanced subscription to openart.ai and it was off to the races. 00:04:36.680 --> 00:04:39.840 Now to be clear, we're not actually recommending openart.ai. 00:04:39.840 --> 00:04:46.640 They've gotten a lot of deserved hate. It's just that we used it for this project because Sora 2, which was our first choice, 00:04:46.640 --> 00:04:51.320 won't let us generate anything with me in it and openart allows us to quickly experiment 00:04:51.320 --> 00:04:57.680 with many different models. Choosing a model, by the way, is not as simple as just grabbing the latest one. 00:04:57.680 --> 00:05:02.360 While newer models, like Google's VO3, might be generally more convincing, they also tend 00:05:02.360 --> 00:05:10.080 to have stricter content guidelines. So for certain prompts, an older model might give a more satisfactory result. 00:05:10.080 --> 00:05:15.480 Once we settled on mostly VO3, with a sprinkling of WAN 2.5 and Cling 2.1, we ran into our 00:05:15.480 --> 00:05:22.680 second hurdle. A lot of our prompts, completely by accident I assure you, got flagged as not safe for 00:05:22.680 --> 00:05:27.760 work. I think it's pretty obvious that I wanted this. 00:05:27.760 --> 00:05:32.720 But whether it's from past experience or from the training data, our video generator 00:05:32.720 --> 00:05:42.280 thought that I wanted this. Anywho, we got around it by using Cloud AI to help us create more AI-friendly prompts 00:05:42.280 --> 00:05:45.280 to help bypass these guardrails. 00:05:45.280 --> 00:05:49.660 From there, the biggest constraint was just how much money we wanted to burn on tokens. 00:05:49.660 --> 00:05:54.660 We ended up throwing away about five video clips for every one that we were able to use. 00:05:54.660 --> 00:05:58.940 Now I'm sure a few of you are wondering, why not just create this on your own hardware? 00:05:58.940 --> 00:06:05.500 And that's a totally valid question. With CompFee UI and some of the open source models out there, you can create videos. 00:06:05.500 --> 00:06:12.060 But the DIY ones are not as convincing yet, and the performance is pretty rough on consumer 00:06:12.060 --> 00:06:17.100 hardware. With that said, things are moving so fast that by the time you watch this, it'll probably 00:06:17.100 --> 00:06:20.900 have improved. That leads us to hurdle number three. 00:06:20.900 --> 00:06:28.300 See a generated actor, unlike a deepfake actor, doesn't actually speak. 00:06:28.300 --> 00:06:33.220 So you've got to line up your clips with separately generated audio. 00:06:33.220 --> 00:06:36.220 That is enough of a challenge when the subject is stationary. 00:06:36.220 --> 00:06:41.500 You throw in some walk and talk or walk and carry your colleague and things get pretty 00:06:41.500 --> 00:06:46.980 rough. Unlike this WAN desk pad from lttstore.com, we use it all the time and it's still nice 00:06:46.980 --> 00:06:50.460 and soft to the touch. But back on subject, look at this clip. 00:06:50.460 --> 00:06:55.500 Not too bad. Add audio so I can show you that I can juggle computers. 00:06:55.500 --> 00:06:59.020 Ooh, that's the yikes. 00:06:59.020 --> 00:07:04.340 Now we tried lip syncing services, but those completely fell apart when it wasn't a simple 00:07:04.340 --> 00:07:11.500 talking headshot. So our editing supervisor extraordinaire Emily used fish audio to generate a bunch of different 00:07:11.500 --> 00:07:14.620 audio versions, and then we picked the closest matches. 00:07:14.620 --> 00:07:20.620 Perfect. No. But this was a just for fun video to see what we could do with limited time and with the 00:07:20.620 --> 00:07:26.620 tools at hand. Scammers on the other hand, they can spend a lot more time and money on these clips if 00:07:26.620 --> 00:07:31.140 they think they can make money on them. That's where the sponsor of today's video comes in. 00:07:31.140 --> 00:07:37.460 That is one thing fake liners was right. It's bit defender. And their message today is pretty simple, but very important. 00:07:37.460 --> 00:07:42.580 And it's one that I agree with. Common sense is not always enough anymore. 00:07:42.580 --> 00:07:46.540 What we showed you guys today is just the beginning for bad actors out there. 00:07:46.540 --> 00:07:51.820 Scammers are constantly innovating their tactics and rapidly adopting AI, making it more difficult 00:07:51.820 --> 00:07:56.540 than ever to distinguish between what's a real emergency and what's a manufactured 00:07:56.540 --> 00:08:03.020 call that might fool a scared loved one. If you're interested in protecting yourself and your loved ones from all sorts of scams, 00:08:03.020 --> 00:08:07.300 you can get 90 days of bit defenders premium security product absolutely free at the link 00:08:07.300 --> 00:08:12.580 below, which includes scam protection, your AI powered defense against online fraud. 00:08:12.580 --> 00:08:16.900 Now let's talk about how to spot a deep fake or a generated scam video. 00:08:16.900 --> 00:08:21.020 Professor Hany Fareed does this for a living and gave a pretty cool TED talk earlier this 00:08:21.020 --> 00:08:26.180 year here in Vancouver, highlighting cutting edge strategies for detecting AI imagery. 00:08:26.180 --> 00:08:30.780 Some of them, like analyzing image noise, probably not going to help the average person 00:08:30.780 --> 00:08:36.740 that much. But what we can do is look at things like shadows and vanishing points. 00:08:36.740 --> 00:08:41.740 See we live in a 3D space, whoa. 00:08:41.740 --> 00:08:48.620 But AI is creating 2D images in an attempt to simulate a 3D space, and it's doing that 00:08:48.620 --> 00:08:54.180 without really a proper understanding of the laws of physics that govern light. 00:08:54.180 --> 00:09:00.020 Let's say you stick a light in a room. You know intuitively that shadows will be cast away from it. 00:09:00.020 --> 00:09:05.460 AI kind of gets this and tries to get it right, but if you try to align the shadows 00:09:05.460 --> 00:09:10.060 with the light source, odds are that they're not going to convert properly. 00:09:10.060 --> 00:09:14.940 And the same goes for something that human artists have understood for centuries, perspective 00:09:14.940 --> 00:09:22.220 drawing and vanishing points. See in a real picture, every object eventually converges towards what's called the vanishing 00:09:22.220 --> 00:09:27.660 point on the horizon, like this, and this, and this. 00:09:27.660 --> 00:09:33.180 But AI fails to do that and is going to draw lines that don't converge properly. 00:09:33.180 --> 00:09:36.540 Is this always easy or convenient to check? 00:09:36.620 --> 00:09:41.180 So, because unfortunately we are far beyond the days of Will Smith's face melting as 00:09:41.180 --> 00:09:45.460 he eats his non-convergent spaghetti, and depending on the model and how much effort 00:09:45.460 --> 00:09:50.260 some would put in, sure you might spot an extra limb or lack thereof and instantly 00:09:50.260 --> 00:09:55.060 know the image is fake, but it's gotten much harder to tell right away and is advancing 00:09:55.060 --> 00:10:01.260 at a rapid pace. So quickly in fact that these physics quirks today could probably be ironed out in the 00:10:01.260 --> 00:10:05.940 coming months rather than years. So what can you count on? 00:10:06.020 --> 00:10:11.380 For now, your best bet is to be more skeptical about any content that you see, especially 00:10:11.380 --> 00:10:16.060 on social media. Or at least, don't rely on it for trustworthy news. 00:10:16.060 --> 00:10:19.500 And if you see anything suspicious, do some digging. 00:10:19.500 --> 00:10:24.180 Find out if it's real or just more AI propaganda. Use other people to help you. 00:10:24.180 --> 00:10:29.340 Have a look at the discussion. Even just looking a little closer to see if there's kind of a weird shimmer around an 00:10:29.340 --> 00:10:33.860 object or if some of the lighting doesn't look quite right. That can help you spot the slop. 00:10:33.860 --> 00:10:38.620 And whatever you do, please don't answer any urgent requests for your information or 00:10:38.620 --> 00:10:41.740 for your money or click on any sketchy links. 00:10:41.740 --> 00:10:46.020 If someone reaches out to you, the safest thing to do is say, hey, I'm going to call 00:10:46.020 --> 00:10:50.380 you back at the number that I already have for you and confirm if this is you. 00:10:50.380 --> 00:10:54.180 Thanks again to Bitdefender for sponsoring this video. They're a global leader in cybersecurity. 00:10:54.180 --> 00:10:58.740 They've got over 17 years of AI innovation under their belts starting in 2008 when they 00:10:58.740 --> 00:11:02.340 introduced AI and machine learning based threat detection. 00:11:02.340 --> 00:11:07.540 While AI's ability to reproduce and scale tactics is a threat, its replication patterns 00:11:07.540 --> 00:11:13.420 can also present opportunities for detection. We're going to have a link for their services in the video description and I sincerely hope 00:11:13.420 --> 00:11:18.620 this video helps you and your loved ones avoid getting fooled by AI. 00:11:18.620 --> 00:11:23.580 Thanks for watching guys. If you liked this video, maybe check out the last time we tried to deep fake me over five 00:11:23.580 --> 00:11:27.620 years ago. It really has come a long way. 00:11:27.620 --> 00:11:31.620 Not even necessarily in terms of the convincingness, but in terms of the ease.