WEBVTT 00:00:00.000 --> 00:00:03.000 Hello, I am here with Emily Young. 00:00:03.000 --> 00:00:09.400 And what do you do for the company? I am a writer. I used to be a benchmarker, but that's lab's job now. 00:00:09.400 --> 00:00:14.440 Excellent, excellent. So I'm just asking everyone in the company here what their thoughts on AI is, like the state of it. 00:00:14.440 --> 00:00:18.360 So what are your thoughts? We don't have it. 00:00:18.360 --> 00:00:23.040 No, it's not. What we refer to as AI right now, 00:00:23.040 --> 00:00:27.240 I think it would be better to describe it as like, 00:00:27.240 --> 00:00:31.280 have you ever played Mass Effect? 00:00:31.280 --> 00:00:35.120 Two, but a long time ago. OK, so in the lore of Mass Effect, 00:00:35.120 --> 00:00:39.640 AIs are like, you don't do that. 00:00:39.640 --> 00:00:43.960 So instead, what they have is what's called a VI, a virtual intelligence. 00:00:43.960 --> 00:00:49.680 And that is a lot like what we consider to be like chatGPT and all that kind of stuff right now. 00:00:49.680 --> 00:00:52.880 So it's like, it's able to answer questions. 00:00:52.880 --> 00:00:55.880 It's able to carry a conversation. 00:00:55.920 --> 00:00:59.560 It's able to do a lot of things, but it's not able to think. 00:00:59.560 --> 00:01:07.560 It doesn't have actual thoughts or feelings. So what it's doing is, in terms of like chatGPT or whatever, 00:01:07.560 --> 00:01:11.080 it's basically, it's learning patterns. 00:01:11.080 --> 00:01:16.880 So it's just learning patterns and spinning out what it thinks is the next thing in the pattern, more or less. 00:01:16.880 --> 00:01:22.920 At least that's my overly simplistic explanation of it. 00:01:25.920 --> 00:01:29.720 So I don't think we have AI in the traditional sense right now. 00:01:29.720 --> 00:01:35.800 At least that's, I'm talking about generative AI that people think is what AI is. 00:01:35.800 --> 00:01:41.400 But it can be pretty convincing. So then it sounds like you're kind of not liking AI. 00:01:41.400 --> 00:01:45.360 Is that right to assume or? 00:01:45.360 --> 00:01:49.480 AI is difficult. There are things that AI is really good for. 00:01:49.480 --> 00:01:53.200 I mean, it's basically like, say for example, AI image 00:01:53.200 --> 00:01:57.200 scaling or whatever. You train it to do whatever it needs to do, and it does it. 00:01:57.200 --> 00:02:00.920 Basically by failing as many times as is necessary 00:02:00.920 --> 00:02:04.440 to fail the correct way. 00:02:04.440 --> 00:02:09.560 And that's actually really cool, because it does things in a way that is very difficult to do 00:02:09.560 --> 00:02:13.080 in traditional software. 00:02:13.080 --> 00:02:17.160 When you get into marketing AI as, well, I mean, 00:02:17.160 --> 00:02:22.000 I keep saying AI. It's more like machine learning. So when you get into marketing it as like AI, 00:02:22.040 --> 00:02:26.080 and you say, oh, well, chatGPT is going to take over the world, well, that's overselling it. 00:02:26.080 --> 00:02:30.280 And I think it's by design that they're doing that, 00:02:30.280 --> 00:02:34.200 because they want people to invest. They want people to think that we're close enough, 00:02:34.200 --> 00:02:37.720 and that these technology companies that are investing in it 00:02:37.720 --> 00:02:40.440 are more powerful than they really are. 00:02:43.320 --> 00:02:49.000 But that being said, AI is a tool like any other tool. It can be used for good or evil. 00:02:49.000 --> 00:02:53.720 There are ways to generate images, for example, that are not copyright infringing. 00:02:53.720 --> 00:02:57.920 I think Adobe's is potentially non-infringing, 00:02:57.920 --> 00:03:03.120 but I know that stable diffusion and what have you are, 00:03:03.120 --> 00:03:06.280 as far as copyright goes. 00:03:06.280 --> 00:03:11.280 Not tested in courts, obviously, but if you take somebody's art and you put it into the generator 00:03:11.280 --> 00:03:16.040 and it just blends it up with everything else, then what is that? 00:03:16.040 --> 00:03:21.760 That's the question that we have to answer. And so far, we don't have one. 00:03:21.760 --> 00:03:26.960 So yeah, it's not a good or bad thing. It's just that the way we use it is really 00:03:26.960 --> 00:03:30.320 what determines whether it's good or bad. 00:03:30.320 --> 00:03:34.480 OK, OK. So then as I creative, you do write. 00:03:34.480 --> 00:03:39.560 And I think one of the things people talk about when AI, what topic of AI is, is that, oh, 00:03:39.560 --> 00:03:44.560 it's going to take over writer's jobs. Do you believe that? Do you think that's a thing in the distant future? 00:03:44.560 --> 00:03:48.840 Or do you think it's inevitable in the next five years like that, or? 00:03:48.840 --> 00:03:55.040 The problem with that is that if no new writing is made, 00:03:55.040 --> 00:03:58.160 like if, say, for example, writing is completely no longer 00:03:58.160 --> 00:04:03.320 a thing because AI is taken over, what you then end up with 00:04:03.320 --> 00:04:07.960 is, what do they call it? Like noise amplification or generational noise. 00:04:07.960 --> 00:04:10.960 That basically, as you iterate on things that are already 00:04:10.960 --> 00:04:15.480 iterated on, you kind of lose bits and pieces of it. 00:04:15.520 --> 00:04:21.480 So it'll be difficult to continue indefinitely. 00:04:21.480 --> 00:04:26.000 You might get away with it a little bit for a little while. And people are already getting away 00:04:26.000 --> 00:04:29.560 with writing paragraphs or whatever with chat 00:04:29.560 --> 00:04:36.560 GPT or what have you. But it still requires an editing pass. 00:04:36.560 --> 00:04:42.240 It still requires people who know what they're doing because I don't know if you or people who are watching know. 00:04:42.280 --> 00:04:47.200 But AI, especially like chat GPT, has a very short memory. 00:04:47.200 --> 00:04:52.520 So if you're trying to tell a story, like in a movie or something, keeping that thing on the rails 00:04:52.520 --> 00:04:59.040 is going to be a really tough operation. 00:04:59.040 --> 00:05:04.560 So you basically have to keep feeding it like the background information because, what do they call it? 00:05:04.560 --> 00:05:10.920 I forget what they call it. But basically, you only have a certain number of basically 00:05:10.920 --> 00:05:15.240 things that it can remember indefinitely. 00:05:15.240 --> 00:05:19.480 And once you run out, it starts dropping them 00:05:19.480 --> 00:05:23.000 and adds more. And eventually, you have like, oh, well, 00:05:23.000 --> 00:05:27.520 you started out in New Zealand. And now you're in, I don't know, Turkey 00:05:27.520 --> 00:05:35.280 because it just chose another location because it forgot what location it was, for example. 00:05:35.280 --> 00:05:38.600 I've tried using chat GPT to write code. 00:05:38.600 --> 00:05:41.680 And it's really, really handy for small bits of code. 00:05:41.680 --> 00:05:45.320 But if you try to get it to do something substantial 00:05:45.320 --> 00:05:48.680 with that code, no. 00:05:48.680 --> 00:05:51.840 Yeah, that's kind of an assumption. Like I kind of give you a lot of people prior. 00:05:51.840 --> 00:05:55.400 And it's a good starting point, I guess. 00:05:55.400 --> 00:05:59.600 But then you need to have an actual person to look over it 00:05:59.600 --> 00:06:02.600 and expand on it further after that. Yeah. 00:06:02.600 --> 00:06:07.040 And obviously, it's going to get better. And it's going to be able to handle more and more things. 00:06:07.040 --> 00:06:10.720 And I mean, eventually, it might get to the point where it could write a screenplay or an operating 00:06:10.720 --> 00:06:17.080 system or something. The problem then becomes, well, how 00:06:17.080 --> 00:06:24.160 do we determine the copyright of that? How do we determine whether it's all actually good? 00:06:24.160 --> 00:06:36.680 Sometimes the code that it spits out isn't super conventional. 00:06:36.720 --> 00:06:43.280 So if it spits out something that is completely and totally unintelligible, and it turns out that there's a bug later on, 00:06:43.280 --> 00:06:47.920 like fixing that bug becomes a little difficult. 00:06:47.920 --> 00:06:54.640 You could use AI to fix that bug then. But then you have compounding factors 00:06:54.640 --> 00:06:58.160 that you're not quite sure how the system even works. 00:06:58.160 --> 00:07:02.640 So I think that for the foreseeable future, 00:07:02.640 --> 00:07:05.640 there's at least going to need to be some kind of review 00:07:05.680 --> 00:07:10.120 process. So you might build a function in AI. 00:07:10.120 --> 00:07:13.280 And then you can put that in. Then you build another function. You put that in. 00:07:13.280 --> 00:07:18.440 And basically, you stitch it together yourself, right? That's probably the best way to do it. 00:07:18.440 --> 00:07:21.440 Yeah. Yeah. 00:07:21.440 --> 00:07:26.880 Sorry. OK, so then as we are a tech company and a creative company, 00:07:26.880 --> 00:07:31.440 do you think that we should be using AI at any level? Or do you think we should be like, 00:07:31.440 --> 00:07:36.680 we don't want to touch that at all? Depends on the AI. And depends on what we're using it for. 00:07:36.680 --> 00:07:43.920 If we're talking chat GPT, I don't know that it's not copyright infringing. 00:07:43.920 --> 00:07:49.200 So that's one thing. The fact that we don't know what the actual data sets 00:07:49.200 --> 00:07:55.520 that they train them on are is a bit of a big, giant question 00:07:55.520 --> 00:07:58.680 mark in terms of the ethics and the, I guess, 00:07:58.680 --> 00:08:06.120 legality, ultimately, of using AI for generative purposes. 00:08:06.120 --> 00:08:13.240 On the other hand, say, for example, I'm like 95% of the way through a script. 00:08:13.240 --> 00:08:16.440 And you're like, you don't know how to close it out. 00:08:16.440 --> 00:08:20.200 You could sit there and you could scratch your head for an hour or two. 00:08:20.200 --> 00:08:25.880 Or you could ping some concepts off of chat GPT or something 00:08:25.880 --> 00:08:31.480 and see what it looks like. And I think that is the best way to go for it. 00:08:31.480 --> 00:08:36.040 So say if you were a visual artist, for example, 00:08:36.040 --> 00:08:41.800 what I would say is a relatively benign use of stable diffusion 00:08:41.800 --> 00:08:47.040 or what have you would be is laying out a scene 00:08:47.040 --> 00:08:50.320 or something like that. So yeah, I want to have this here and this here and this here 00:08:50.320 --> 00:08:56.040 and this is what the perspective should look like. And then when you have that, then you can paint over it. 00:08:56.040 --> 00:09:02.040 So I feel like that's the kind of thing that ultimately we should be using it for. 00:09:02.040 --> 00:09:05.960 I know that many people will be using it for and already are 00:09:05.960 --> 00:09:10.640 using it for complete works, which is a little less 00:09:10.640 --> 00:09:15.880 defensible. Yeah, I think that Disney's sequence invasion 00:09:15.880 --> 00:09:20.400 was a big example of like, because from what I am aware, 00:09:20.400 --> 00:09:25.720 they just put it in the prompt and that's it, right? They didn't change anything, right? 00:09:26.160 --> 00:09:30.600 I'm not sure. Yeah, OK. Yeah, I haven't seen that. Yeah, I just heard about it and I was like, oh. 00:09:30.600 --> 00:09:34.440 But then I wasn't like that invested during that time, but I was like the last two weeks. 00:09:34.440 --> 00:09:39.880 I'm like, I'm getting more into it, but. Yeah, I mean, the few things that I've seen, like, oh, well, 00:09:39.880 --> 00:09:44.440 AI is going to completely just take over the movie industry. Every frame was generated of this thing. 00:09:44.440 --> 00:09:49.280 I don't remember what it was, but it's like, it looks not good yet. 00:09:49.280 --> 00:09:55.640 And I guess you can see that it's there, sort of. But it doesn't really have, I guess, 00:09:55.640 --> 00:10:00.080 the memory to create a plot, and it still 00:10:00.080 --> 00:10:03.320 can't do fingers quite perfectly. 00:10:03.320 --> 00:10:12.120 So you've got like after images of stuff, you've got like six fingers or 12, stuff like that, 00:10:12.120 --> 00:10:16.720 or like teeth on teeth kind of thing, where it thinks stuff should go because that's 00:10:16.720 --> 00:10:22.120 where the pattern is, right, the pattern that it was fed. So it's like, OK, so teeth go here. 00:10:22.120 --> 00:10:29.200 And oh, yeah, that looks like teeth might go there, too. And right, so that's kind of how that works. 00:10:29.200 --> 00:10:33.040 So we're just going to wrap it up here. Is there anything that we didn't talk about 00:10:33.040 --> 00:10:36.640 that you wanted to bring it up, or just want to say anything to the float planers? 00:10:36.640 --> 00:10:40.880 I'm trying to coin that term now. People who subscribe, float planers, yeah. 00:10:40.880 --> 00:10:44.240 You think it's a good name? That's what we've been calling them internally for a while. Really? 00:10:44.240 --> 00:10:47.640 Yeah, float planers. All right, float planers, you heard it first. 00:10:48.080 --> 00:10:52.680 Is there anything you want to say about the topic, or anything you want to mention in general, or? 00:10:52.680 --> 00:10:56.880 Just reiterate that everything, all tools, 00:10:56.880 --> 00:10:59.880 are inherently not good or evil, right? 00:10:59.880 --> 00:11:03.200 It is what you use it for that is the most important. 00:11:03.200 --> 00:11:06.880 So AI image upscaling is amazing, for example. 00:11:06.880 --> 00:11:10.640 It's just one of those things that's just so universally 00:11:10.640 --> 00:11:14.800 useful that as it gets better, it 00:11:14.800 --> 00:11:20.040 becomes less and less likely that we'll ever not use it. 00:11:20.040 --> 00:11:23.960 Whereas those generative techniques, I don't know. 00:11:23.960 --> 00:11:28.520 It's one of those things where we might actually end up saying, OK, so we need to take a step back 00:11:28.520 --> 00:11:33.400 and ask ourselves some tough questions about how we want to do this. 00:11:33.400 --> 00:11:38.480 As far as AI, the concept of actual artificial intelligence 00:11:38.480 --> 00:11:44.800 goes, where are ways out from that, from what I understand? 00:11:44.800 --> 00:11:51.080 But I mean, we can simulate it pretty well with what we have now. And soon, it might not even matter. 00:11:51.080 --> 00:11:55.200 If you simulate it close enough, then what's the difference, right? 00:11:55.200 --> 00:12:00.160 So I guess the ultimate takeaway is be nice to your computer. 00:12:00.160 --> 00:12:04.880 Shut it down properly. You heard it first, don't unplug it randomly. 00:12:04.880 --> 00:12:08.120 All right, thank you very much, Emily, for your time. Thank you.