[00:00.000 --> 00:03.420]  All right, so this is Stable Diffusion. This is an AI?
[00:03.420 --> 00:05.040]  Yeah, this is Stable Diffusion.
[00:05.040 --> 00:06.380]  This is the AI?
[00:06.380 --> 00:08.080]  This is one of the AIs.
[00:08.100 --> 00:09.280]  And what's this one do?
[00:09.280 --> 00:12.480]  This is actually ControlNet bolted on top of Stable Diffusion.
[00:12.480 --> 00:16.320]  So you have, in this case, the Nike Swish.
[00:16.480 --> 00:21.920]  It's going to control the Stable Diffusion to generate a edge map right here.
[00:21.920 --> 00:27.380]  So you can get... this is how you get the images with certain patterns built into them.
[00:27.380 --> 00:29.640]  So have you seen the TikTok filters with people like...
[00:29.640 --> 00:32.300]  Dancing like a... they look like a...
[00:32.300 --> 00:34.040]  I totally do those videos.
[00:34.040 --> 00:36.460]  Yeah, this is how it's built.
[00:36.460 --> 00:40.120]  Okay, so it does the edging around it and then it can do things around it.
[00:40.480 --> 00:43.460]  So this one does edging, but then you can do it with a normal map,
[00:43.460 --> 00:47.520]  you can do it with a pose map. There's all sorts of different ways of doing it.
[00:47.520 --> 00:50.800]  Do you always have to feed an image in or can you also do text in?
[00:50.800 --> 00:54.540]  This is a text image. So this is a human baby lobster,
[00:55.120 --> 00:59.400]  but we can do it with this image or this... various other images.
[00:59.640 --> 01:04.040]  Okay. If you don't feed it an image, are you able to just generate something?
[01:04.040 --> 01:08.540]  It's going to see what happens with human baby lobster without an image fed into it.
[01:08.540 --> 01:12.380]  This one's still got the Nike Switch. I haven't swapped it over.
[01:14.500 --> 01:18.880]  That is remarkably creepy for some of these.
[01:18.900 --> 01:22.840]  Yeah, this one has improved drastically since we did this last year.
[01:22.840 --> 01:27.340]  We had Latent Diffusion and if you asked it to generate an image like this,
[01:27.340 --> 01:32.600]  it would generate like outer nightmare fuel. I think I have some nightmare fuel on this computer,
[01:32.600 --> 01:36.420]  but it was like the most disgusting tentacle monster we've ever seen.
[01:36.420 --> 01:40.680]  Oh my gosh. Some of these are still like, they're really good.
[01:40.680 --> 01:42.080]  I like the ribs.
[01:42.080 --> 01:44.860]  Kind of nuts at the same time.
[01:45.220 --> 01:51.350]  So as a curiosity, the other item on the scav hunt list is crashing AI.
[01:51.840 --> 01:52.660]  Yeah, we can...
[01:52.660 --> 01:54.400]  How do you crash an AI?
[01:54.400 --> 01:58.420]  This particular one I crash usually by going too big.
[01:59.300 --> 02:05.160]  So this model was trained to do 512x512.
[02:05.280 --> 02:09.400]  So you're making the image way too big for it to process and it will crash.
[02:09.400 --> 02:14.040]  Oh, well this won't crash, but I want to turn off the control net stuff.
[02:14.100 --> 02:15.680]  And you're expecting this one to crash?
[02:15.680 --> 02:19.840]  No, this won't crash, but the model is going to start generating some really weird stuff.
[02:20.340 --> 02:22.900]  But what happened is I pushed it up.
[02:22.920 --> 02:28.580]  Now it's consuming nearly 36 gigs, 36 gigabytes of VRAM.
[02:28.760 --> 02:32.840]  I have an A6000, so I have nearly 50 gigs of VRAM.
[02:33.640 --> 02:36.520]  But normally this would crash just out of memory.
[02:36.560 --> 02:37.140]  Okay.
[02:37.140 --> 02:39.560]  And if I pushed these numbers up a little bit higher,
[02:39.560 --> 02:41.160]  I would have crashed it without a memory.
[02:41.160 --> 02:44.100]  But what happens now is it's going to take 10 minutes to generate.
[02:44.100 --> 02:48.060]  But usually when you push it this high, you get a result.
[02:48.060 --> 02:53.500]  So it doesn't crash crash, but because I've pushed the image range so high,
[02:53.500 --> 02:56.280]  it doesn't generate a sensible image anymore.
[02:56.280 --> 02:58.240]  So if you want a horse with eight legs,
[02:58.240 --> 03:00.420]  and you want to guarantee a horse with eight legs,
[03:00.420 --> 03:04.080]  you push the image resolution too high.
[03:04.260 --> 03:08.440]  And then it will generate horses with eight legs and all sorts of fun things
[03:08.440 --> 03:15.340]  that you don't get when you generate normal images on Hugging Face or Dali 2.
[03:15.340 --> 03:18.700]  Because they want you to get good images that are usable,
[03:18.700 --> 03:20.720]  and I'm trying to currently push an image that's
[03:20.720 --> 03:23.340]  through this thing that is just going to break the system.
[03:23.340 --> 03:27.540]  And what causes it? Okay, maybe it's not a crash, but what causes it to go goofy?
[03:27.540 --> 03:30.880]  It's that it runs out of memory, so in an attempt to save memory,
[03:30.880 --> 03:32.920]  it's like offloading, and what's happening?
[03:32.920 --> 03:35.240]  The model was trained to generate 512x512.
[03:35.240 --> 03:39.360]  So in this case, the model was trained to generate 256x256 images,
[03:39.360 --> 03:41.740]  and then upscale that to 512x512.
[03:41.760 --> 03:44.280]  And the upscaling is literally like CSI Enhanced.
[03:44.280 --> 03:46.600]  That's what it is, just CSI Enhanced.
[03:46.600 --> 03:51.560]  I feel like every time I use CSI Enhanced, I do have to, in a deep voice, go, Enhanced.
[03:51.560 --> 03:57.060]  Yeah, this parameter right here is saying, oh, we're going to do Enhanced 30 times.
[03:58.280 --> 03:59.900]  But it does the upscaling.
[04:00.540 --> 04:03.860]  But what it was trained to do is have that field of view,
[04:03.860 --> 04:06.260]  and so suddenly you have given it this field of view.
[04:06.520 --> 04:12.140]  And so it's going to make the same, it just tiles that process over and over again.
[04:12.140 --> 04:14.080]  So it's going to try to generate a horse here,
[04:14.080 --> 04:16.380]  and then it's going to go over here and try to generate a horse.
[04:16.380 --> 04:20.840]  And they overlap, so you get a horse, so the horses meld together.
[04:20.840 --> 04:24.540]  Is this one of the reasons why some of the early things with stable diffusion
[04:24.540 --> 04:27.660]  made people look so weird, or is that a separate problem?
[04:27.660 --> 04:34.100]  I don't actually know myself, but one of the creators is around,
[04:34.100 --> 04:39.580]  I don't think she's here right now, but like that particular thing, I think was a,
[04:39.580 --> 04:42.660]  basically that story of like, I'm going to generate one set of finger,
[04:42.660 --> 04:45.240]  it knows how to generate a finger, and then generates another finger,
[04:45.240 --> 04:49.200]  and then it just keeps going. And it just accidentally makes six fingers,
[04:49.200 --> 04:53.660]  because it knows how to generate small patches. But it's got the same thing.
[04:53.840 --> 05:00.940]  I think you're right, but I'm not sure. Because I do ML security with malware models,
[05:00.940 --> 05:03.500]  and spam, and all sorts of other weird things.
[05:03.500 --> 05:09.360]  So let's answer the question for Scav Hunt, which is, how do you crash AI?
[05:09.960 --> 05:15.620]  Well, in this case, if you want to crash stable diffusion, you can allocate way too much RAM.
[05:15.740 --> 05:19.760]  Most of the time I crash AI, it's out of memory.
[05:21.300 --> 05:27.920]  With a large language model, the number of tokens, the number of tokens,
[05:27.920 --> 05:33.560]  the amount of memory you need scales with the square of the number of tokens, and that is bad.
[05:33.560 --> 05:37.520]  So is that what happens when you get too many tokens, it loses the original con?
[05:37.520 --> 05:41.320]  It does one of two things. Either A, it loses the original context,
[05:41.320 --> 05:43.920]  and this is one of the ways a lot of jailbreaks work, right?
[05:43.920 --> 05:48.620]  Where it fills up the context, it forgets about its original security parameters,
[05:48.620 --> 05:51.340]  and now it's doing things it's really not supposed to do.
[05:51.360 --> 05:57.240]  Like the invention of a lot of the jailbreaks, where you can now, all of a sudden, recipes
[05:57.240 --> 06:01.180]  for drinking bleach and things like that, where it's saying things you're definitely
[06:01.180 --> 06:06.360]  not supposed to do as recipes. You either have the option of losing context, or...
[06:06.360 --> 06:07.760]  You start running out of memory.
[06:07.760 --> 06:10.420]  You run out of memory, and it officially crashes the process.
[06:10.420 --> 06:14.340]  Yeah, but they control the number of tokens. That's why
[06:15.180 --> 06:18.020]  Anthropic was so excited about beginning to 10,000 tokens.
[06:18.440 --> 06:23.700]  They came up with a new way of doing it to drastically expand the amount of tokens
[06:23.700 --> 06:25.540]  without messing with the memory too much.
[06:25.540 --> 06:31.020]  Okay. So with this then, first, thank you for teaching me how to crash AI.
[06:31.020 --> 06:35.000]  I have that now, that feather in my cap. Thank you so much for your time.
[06:35.000 --> 06:40.060]  I really, really appreciate it. Thank you for watching. And as always, hack on.
[06:40.280 --> 06:46.680]  That is super hot right now. And if it doesn't crash, it might catch fire.