WEBVTT 00:00:03.120 --> 00:00:09.120 Bit Defender sponsored this video so I can show you that I can juggle computers 00:00:07.040 --> 00:00:14.160 and tell you about our new amazing product, Linus Strength Pills. Watch as 00:00:11.920 --> 00:00:20.360 I make David watch all of 2019's [music] cats. Wait, not like this. 00:00:18.340 --> 00:00:20.360 [screaming] 00:00:23.580 --> 00:00:25.600 [laughter] 00:00:34.719 --> 00:00:41.600 Thanks to Linus Pills, I'm unstoppable. Start your subscription today by 00:00:39.200 --> 00:00:45.840 emailing me at Linus.Linus Linus.com. I don't remember recording 00:00:44.000 --> 00:00:49.280 any of this. Bit Defender [music] did sponsor a video about protecting 00:00:47.200 --> 00:00:52.800 yourself from deepface, though, because now these have gotten so good that you 00:00:51.360 --> 00:00:59.399 didn't even realize that I'm not the real Lionist either. 00:00:55.680 --> 00:00:59.399 What are you guys doing? 00:01:07.439 --> 00:01:14.799 Fully AI generated videos, at least for now, still have some easy to spot tells, 00:01:13.200 --> 00:01:19.119 especially if they're long or if they involve a lot of movement. But a simple 00:01:16.960 --> 00:01:22.560 head replacement deep fake or a stationary figure at a desk telling you 00:01:20.799 --> 00:01:27.439 all about their latest get-rich quick scheme. Those have gotten shockingly 00:01:24.960 --> 00:01:32.640 convincing over the last 5 years. Let's talk about both how we deep faked and 00:01:30.000 --> 00:01:37.280 how we fully generated me for this video using just commodity hardware. But 00:01:34.799 --> 00:01:41.360 before we do that, an important message. If you've got a loved one who's going to 00:01:38.880 --> 00:01:44.560 need some help recognizing what's AI and what's not, send them to this timestamp. 00:01:43.491 --> 00:01:50.320 [music] We've got you. Starting with our deep fake. The process 00:01:48.240 --> 00:01:55.680 is still pretty similar to how we did it last time, just 5 years refined. First, 00:01:53.920 --> 00:01:58.960 find an actor with a similar body shape to your target. You can see why this 00:01:57.600 --> 00:02:03.840 matters in this test that we did with Plof, who owns a display. When you blow 00:02:01.439 --> 00:02:07.920 it up on a TV, it's pretty easy to spot anomalies like his hat or his beard 00:02:05.920 --> 00:02:12.239 leaving a soft outline. But on your phone, especially with poor eyesight, 00:02:10.319 --> 00:02:15.760 mostly just looks like a guy with a slim face on a chunky body. Oh my god, am I 00:02:14.239 --> 00:02:22.239 allowed to say that? It's okay. I wrote the script. >> Okay. So, we parsed the subreddit then 00:02:19.040 --> 00:02:24.959 for the latest is this Lionus meme and 00:02:22.239 --> 00:02:29.760 chose Chase for our real attempt. We used Deep Face Lab to train a model on 00:02:27.280 --> 00:02:37.120 about 7,000 recent images of my face. And boom, much more convincing. And this 00:02:33.840 --> 00:02:39.440 is both super cool and super scary. It 00:02:37.120 --> 00:02:43.040 was at least 100 times easier than when we did it [music] last time. Syncing up 00:02:41.440 --> 00:02:47.680 the lips with the audio is still pretty tough. But if you keep your clips short 00:02:45.519 --> 00:02:51.840 and you edit them together in a punchy manner with alternate angles and 00:02:49.200 --> 00:02:55.760 close-ups like you typically would, you can splice [music] together something 00:02:53.120 --> 00:03:00.720 pretty convincing and pretty long, one 5 to 7second shot at a time. With that 00:02:58.560 --> 00:03:05.040 said, I'm sure that a lot of you guys could still tell cuz realistically you 00:03:03.360 --> 00:03:09.599 dwell the same corners of the internet that I do and you've been watching with 00:03:06.720 --> 00:03:14.640 a sense of awe and with dread as deep fakes and video generation have gone 00:03:11.200 --> 00:03:16.319 from [snorts] obviously fake garbage to 00:03:14.640 --> 00:03:24.000 oh this requires a little bit of scrutiny. But you savvy viewer are not 00:03:20.000 --> 00:03:26.640 my concern. Many, maybe even most people 00:03:24.000 --> 00:03:31.519 can't tell anymore. and online scams and fraud are increasing [music] every year 00:03:28.400 --> 00:03:34.560 to the point where in 2024 losses are 00:03:31.519 --> 00:03:35.680 estimated at over $1 trillion. 00:03:34.560 --> 00:03:41.840 That number came from Bit [music] Defender, who is in the business of knowing these things. And the scariest 00:03:39.519 --> 00:03:47.519 part is that [music] a huge portion of the scam industry is still using old 00:03:44.159 --> 00:03:49.599 school text to voice phone calls. I 00:03:47.519 --> 00:03:54.640 mean, imagine if you could hit victims with something more like [music] this. 00:03:51.519 --> 00:03:56.799 Hey, I'm in a jam. Can I borrow $5,000? 00:03:54.640 --> 00:04:02.159 And I know some of you guys are probably thinking, "Tough break, boomers. You had 00:03:59.519 --> 00:04:06.159 enough of the wealth anyway." But that's easy to say until it's your family 00:04:04.640 --> 00:04:10.640 member. [music] And we have multiple people in this office who have had 00:04:07.920 --> 00:04:16.000 family members impacted. And again, the scariest part is how easy it was to 00:04:13.200 --> 00:04:20.479 generate that clip. All we needed for it and for the ones at the start of this 00:04:17.440 --> 00:04:22.880 video were start and end key frames. 00:04:20.479 --> 00:04:26.560 Once those were captured, or in many cases just scraped off of an existing 00:04:24.960 --> 00:04:31.360 video or Facebook photo. >> My name is Nicholas Plove and I hate 00:04:30.080 --> 00:04:38.080 displays. >> We got an advanced subscription to openart.ai and it was off to the races. 00:04:36.479 --> 00:04:42.960 Now, to be clear, we're not actually recommending OpenArt.ai. They've gotten 00:04:40.320 --> 00:04:47.120 a lot of deserved hate. It's just that we used it for this project because Sora 00:04:44.880 --> 00:04:51.360 2, which was our first choice, won't let us generate anything with me in it. and 00:04:49.280 --> 00:04:55.680 Open Art allows us to quickly experiment with many different models. Choosing a 00:04:53.759 --> 00:05:00.240 model, by the way, is not as simple as just grabbing the latest one. While 00:04:57.840 --> 00:05:04.000 newer models like Google's V3 might be generally more convincing, they also 00:05:02.240 --> 00:05:08.240 tend to have stricter content guidelines. So, for certain prompts, an 00:05:06.720 --> 00:05:13.600 older model might give a more satisfactory result. Once we settled on 00:05:10.800 --> 00:05:19.120 mostly V3 with a sprinkling of WAN 2.5 and Cling 2.1, we ran into our second 00:05:16.000 --> 00:05:21.120 hurdle. A lot of our prompts, completely 00:05:19.120 --> 00:05:26.800 by accident, I assure you, got flagged as not safe for work. Okay, listen. I 00:05:24.800 --> 00:05:31.759 think it's pretty obvious that I wanted this. But whether it's from past 00:05:29.360 --> 00:05:36.000 experience or from the training data, our video generator thought that I 00:05:33.600 --> 00:05:41.520 wanted this. Anywh who, we got around it by using 00:05:38.400 --> 00:05:43.919 cloud AI to help us create more AI 00:05:41.520 --> 00:05:48.080 friendly prompts to help bypass these guardrails. From there, the biggest 00:05:46.240 --> 00:05:52.240 constraint was just how much money we wanted to burn on tokens. We ended up 00:05:50.160 --> 00:05:55.680 throwing away about five video clips for every [music] one that we were able to 00:05:53.520 --> 00:05:59.680 use. Now, I'm sure a few of you are wondering, why not just create this on 00:05:57.840 --> 00:06:03.919 your own hardware? And that's a totally valid question. With Comfy UI and some 00:06:02.080 --> 00:06:10.080 of the open source models out there, you can create videos, but the DIY ones are 00:06:08.080 --> 00:06:14.479 not as convincing yet, and the performance is pretty rough on consumer 00:06:12.240 --> 00:06:17.440 hardware. With that said, things are moving. so fast that by the time you 00:06:15.840 --> 00:06:23.360 watch this, it'll probably have improved. That leads us to hurdle number 00:06:19.840 --> 00:06:28.160 three. See, a generated actor, unlike a 00:06:23.360 --> 00:06:29.759 deep fake actor, doesn't actually speak. 00:06:28.160 --> 00:06:34.880 So, you've got to line up your clips with separately generated audio. Huh, 00:06:33.039 --> 00:06:39.360 that is enough of a challenge when the subject is stationary. You throw in some 00:06:36.880 --> 00:06:44.160 walk and talk or walk and carry your colleague and things get pretty rough. 00:06:42.319 --> 00:06:48.160 Unlike this WAN deskpad from ltstore.com, we use it all the time and 00:06:46.479 --> 00:06:53.759 it's still nice and soft to the touch. But back on subject, look at this clip. 00:06:50.400 --> 00:06:57.199 Not too bad. Add audio so I can show you 00:06:53.759 --> 00:07:00.400 that I can juggle computers. [music] 00:06:57.199 --> 00:07:02.960 That's a yikes. Now, we tried lip 00:07:00.400 --> 00:07:07.280 syncing services, but those completely fell apart when it wasn't a simple 00:07:04.560 --> 00:07:11.360 talking head shot. So, our editing supervisor extraordinaire Emily used 00:07:09.440 --> 00:07:16.479 Fish Audio to generate a bunch of different audio versions and then we 00:07:12.880 --> 00:07:18.800 picked the closest matches. Perfect. No, 00:07:16.479 --> 00:07:22.639 but this was a just for fun video to see what we could do with limited time and 00:07:20.479 --> 00:07:26.715 with the tools at hand. Scammers, on the other hand, they can spend a lot more 00:07:24.560 --> 00:07:29.440 time and money on these clips if they [music] think they can make money on 00:07:27.919 --> 00:07:33.280 them. That's where the sponsor of today's video comes in. That is one 00:07:31.520 --> 00:07:37.280 thing fake liners was right. It's Bit Defender and their message today is 00:07:35.199 --> 00:07:42.639 pretty simple but very important [music] and it's one that I agree with. Common 00:07:39.759 --> 00:07:46.240 sense is not always enough anymore. What we showed you guys today is just the 00:07:44.479 --> 00:07:50.880 beginning for bad actors out there. Scammers are constantly innovating their 00:07:48.319 --> 00:07:54.400 tactics and rapidly adopting AI, making it more difficult than ever to 00:07:52.560 --> 00:07:59.440 distinguish between what's a real emergency and what's a manufactured call 00:07:56.960 --> 00:08:02.400 that might fool a scared loved one. If you're interested in protecting yourself 00:08:00.800 --> 00:08:06.160 and your loved ones from all sorts of scams, you can get 90 days of Bit 00:08:04.479 --> 00:08:10.639 Defender's premium security product absolutely free at the link below, which 00:08:08.240 --> 00:08:14.479 includes scam protection, your AI powered defense against online fraud. 00:08:12.639 --> 00:08:18.720 Now, let's talk about how to spot a deep fake or a generated scam video. 00:08:16.879 --> 00:08:22.479 Professor Hannie Fared does this for a living and gave a pretty cool TED talk 00:08:20.639 --> 00:08:26.975 earlier this year here in Vancouver highlighting cuttingedge strategies for 00:08:24.240 --> 00:08:30.639 detecting AI imagery. Some of them [music] like analyzing image noise 00:08:29.360 --> 00:08:35.360 probably not going to help the average person that much. But what we can do is 00:08:33.760 --> 00:08:41.680 look at things like shadows and vanishing points. See, we live in a 3D 00:08:38.880 --> 00:08:47.680 space. Whoa. But AI is creating 2D images in an 00:08:45.120 --> 00:08:51.680 attempt to simulate a 3D space. [music] And it's doing that without really a 00:08:49.920 --> 00:08:56.560 proper understanding of the laws of physics that govern light. Let's say you 00:08:54.560 --> 00:09:02.320 stick a light in a room. You know intuitively [music] that shadows will be 00:08:58.399 --> 00:09:04.320 cast away from it. AI kind of gets this 00:09:02.320 --> 00:09:08.160 and tries to get it right. But if you try to align the shadows with the light 00:09:06.080 --> 00:09:12.720 source, odds are that they're not going to converge properly. And the same goes 00:09:10.880 --> 00:09:17.680 for something that human artists have understood for centuries, perspective 00:09:15.040 --> 00:09:21.839 drawing and vanishing points. See, in a real picture, every object eventually 00:09:20.080 --> 00:09:29.279 converges towards what's called the vanishing point on the horizon like this 00:09:24.880 --> 00:09:30.959 and this and this. But AI fails to do 00:09:29.279 --> 00:09:37.440 that and is going to draw lines that don't converge [music] properly. Is this 00:09:33.680 --> 00:09:39.360 always easy or convenient to check? No, 00:09:37.440 --> 00:09:42.399 because unfortunately we are far beyond the days of Will Smith's face melting as 00:09:41.200 --> 00:09:46.000 he eats his [music] non-convergent spaghetti. And depending on the model 00:09:44.560 --> 00:09:50.640 and how much effort someone would put in, sure, you might spot an extra limb 00:09:48.160 --> 00:09:54.720 or lack thereof and instantly know the image is fake. But it's gotten much 00:09:52.959 --> 00:09:59.920 harder to tell right away [music] and is advancing at a rapid pace. So quickly, 00:09:57.200 --> 00:10:04.160 in fact, that these physics quirks today could probably be ironed out in the 00:10:01.440 --> 00:10:09.519 coming months rather than years. So what can you count on? Well, for now, your 00:10:06.880 --> 00:10:14.480 best bet is to be more skeptical about any content that you see, especially on 00:10:11.680 --> 00:10:19.279 social media. Or at least don't rely on it for trustworthy news. And if you see 00:10:16.560 --> 00:10:23.600 anything suspicious, do some digging. Find out if it's real or just more AI 00:10:21.200 --> 00:10:27.839 propaganda. Use other people to help you. Have a look at the discussion. Even 00:10:26.079 --> 00:10:31.040 just looking a little closer to see if there's kind of a weird shimmer around 00:10:29.360 --> 00:10:35.120 an object or if some of the lighting doesn't look quite right. That can help 00:10:32.720 --> 00:10:39.440 you spot the slop. And whatever you do, please don't answer any urgent requests 00:10:37.440 --> 00:10:44.000 for your information or for your money or click on any sketchy links. If 00:10:41.839 --> 00:10:47.360 someone reaches out to you, the safest thing to do is say, "Hey, I'm going to 00:10:45.920 --> 00:10:50.959 call you back at the number that I already have for you." And confirm if 00:10:49.519 --> 00:10:55.760 this is you. Thanks again to Bit Defender for sponsoring this video. They're a global leader in cyber 00:10:53.760 --> 00:10:59.920 security. They've got over 17 years of AI innovation under their belts starting 00:10:57.519 --> 00:11:04.800 in 2008 when they introduced AI and machine learning based threat detection. 00:11:02.160 --> 00:11:08.560 While AI's ability to reproduce and scale tactics is a threat, its 00:11:06.720 --> 00:11:11.760 replication patterns can also present opportunities for detection. We're going 00:11:10.320 --> 00:11:16.079 to have a link for their services in the video description. And I sincerely hope 00:11:13.519 --> 00:11:20.000 this video helps you and your loved ones avoid getting fooled by [music] AI. 00:11:18.640 --> 00:11:26.560 Thanks for watching, guys. If you like this video, maybe check out the last time we tried to deep fake [music] me 00:11:23.040 --> 00:11:28.800 over 5 years ago. It really has come a 00:11:26.560 --> 00:11:33.200 long way. Not even necessarily in terms of the convincingness, but in terms 00:11:30.377 --> 00:11:33.200 [music] of the ease.