Buong transkripsyon ng How Local Language Models Protect Privacy, with Dr. Arshavir Blackwell

KOPYA NG SALIN 🔗

WEBVTT 00:00:10.000 --> 00:00:17.359 The Dead Pixel Society Podcast is brought to you by Media clip, Advertek Printing, and Independent Photo Imagers. 00:00:17.359 --> 00:00:25.039 Welcome to the Dead Pixels Society Podcast, the photoimaging industry's leading news source. 00:00:25.039 --> 00:00:27.600 Here's your host, Gary Pageau. 00:00:30.239 --> 00:00:33.679 Hello again and welcome to the Dead Pixel Society Podcast. 00:00:33.679 --> 00:00:36.799 I'm your host, Gary Pageau, and today we're talking to Dr. 00:00:36.799 --> 00:00:40.960 Arshavir Blackwell, who is coming to us from Los Angeles. 00:00:40.960 --> 00:00:49.040 He has a company called Arvoinen AI, and he is an AI expert, especially in language models. 00:00:49.040 --> 00:00:50.560 Hello, Arshavir. 00:00:50.560 --> 00:00:51.680 How are you today? 00:00:52.079 --> 00:00:53.359 Hi, I'm doing great. 00:00:53.359 --> 00:00:54.640 It's great to be here. 00:00:54.879 --> 00:00:59.280 So before we get started, can you expand your background in this field? 00:00:59.280 --> 00:01:00.879 How you kind of got started in it? 00:01:00.879 --> 00:01:08.480 Because it sounds to me like you've been in it a long time, even though it's a relatively hot topic, you've been in it for a while. 00:01:08.959 --> 00:01:13.040 Yeah, I've been in it since, as I like to say, way before it was cool. 00:01:13.040 --> 00:01:18.799 I started out doing graduate work at UC San Diego. 00:01:18.799 --> 00:01:36.880 And I worked there with two pioneers in the field, Elizabeth Bates, who was a linguist, and Jeff Elman, who did a lot of the pioneering work in neural networks, which is what we were calling AI, the kind of AI we talk about now. 00:01:37.040 --> 00:01:37.359 Yeah. 00:01:37.600 --> 00:01:38.480 At the time. 00:01:38.480 --> 00:01:47.040 And he did a lot of pioneering work that, believe it or not, still applies today to the way modern LLMs work. 00:01:47.040 --> 00:01:51.120 Liz was more interested in human language processing. 00:01:51.120 --> 00:01:57.920 And what she did was she built models and theories about language learning and language processing. 00:01:57.920 --> 00:02:08.079 And that's significant because a lot of people develop these machine models, but they didn't really have anything to tell us about how humans learn language. 00:02:08.079 --> 00:02:18.640 And our feeling was hey, you know, if this thing is really going to work correctly, like a human interacting with another person or another computer, sure. 00:02:18.640 --> 00:02:24.080 It should probably have a lot of these other learning properties that human beings have as well. 00:02:24.080 --> 00:02:33.200 So from that I went to, I worked for a whole bunch of different, mostly startups or some companies that started a startup, and then while I was there, they got bigger. 00:02:33.200 --> 00:02:42.800 You may remember Ask Jeeves, the question answering, which was a sort of proto LL chat GPT type thing way back in the day. 00:02:42.800 --> 00:02:57.680 The secret to that was all of those answers, all they were canned, and basically they, although the responses came up immediately and automatically, the weights to connect your question to the response were all done by hand. 00:02:57.680 --> 00:03:04.240 So they were basically doing by hand what we now do using algorithms in neural networks. 00:03:04.479 --> 00:03:04.800 Okay. 00:03:05.120 --> 00:03:13.280 And so since then I've been working, one of the big research areas of interest for me is mechanistic interpretability. 00:03:13.280 --> 00:03:23.680 And what I find fascinating about these models is that no one who builds them really fully understands how they work. 00:03:23.680 --> 00:03:26.000 And people are confused by that statement. 00:03:26.000 --> 00:03:27.520 So let me unpack it a little. 00:03:27.520 --> 00:03:46.479 Surely they understand if they build them, but it's the the situation is that these are highly complicated, complicated systems with literally billions and billions of different numbers, different parameters that are set automatically by the algorithm. 00:03:46.479 --> 00:03:49.759 There's not a human being who's in there twisting knobs and dials. 00:03:49.759 --> 00:03:52.400 It wouldn't be possible to do that. 00:03:52.400 --> 00:04:10.719 So although we understand in a very broad looking at the entire forest metaphor for how these things work, we really don't understand how the individual trees all fit together to provide output that you get at the end of the day. 00:04:10.719 --> 00:04:19.599 So you put in, you type in something, and it whirls through all of these very high-dimensional, complicated computations. 00:04:19.599 --> 00:04:33.919 At the end of the day, it comes out with response to you, but we don't have a great sense, a great bead on exactly what's going on on the inside of the machine to allow that to happen. 00:04:34.239 --> 00:04:37.839 Yeah, that's sort of black box where it's a black box. 00:04:37.839 --> 00:04:38.719 Exactly. 00:04:38.719 --> 00:04:40.000 We really don't understand. 00:04:40.000 --> 00:04:47.599 And that's why it's difficult to know when the systems hallucinate or make things up or give information because we don't know how they got to that. 00:04:48.079 --> 00:04:49.039 Exactly. 00:04:49.039 --> 00:05:01.039 And for things like hallucinations, the main techniques are really just to right now to check that the citations that it's giving you are actual real citations and not something that's hallucinated. 00:05:01.279 --> 00:05:01.519 Right. 00:05:01.679 --> 00:05:25.119 But interpretability is all about designing methods to look inside what's going on in this black box, sure, and seeing if you can get some kinds of general parameters or general rules that explain to you exactly what's going on when you do something. 00:05:25.119 --> 00:05:33.359 So one example of this is there's a researcher, Neel Nanda, who wrote a really what I thought was a great paper. 00:05:33.359

Load full transcript

--> 00:05:51.759 And there's a phenomenon where if you try to train a neural net to do simple addition, at first what happens is it just recognizes the pair, the number pairs that it's seen by rote memorization, essentially. 00:05:51.759 --> 00:05:59.199 So you give it five and two, or you give it seven and four, or whatever, it just spits out what it is already seen. 00:05:59.199 --> 00:06:05.919 But what's really amazing about these systems is they go through what Nanda calls phase transition. 00:06:05.919 --> 00:06:24.560 And what that means is that there's a certain point where it's almost like there's a switch inside it, and it suddenly turns on and it's doing real arithmetic, it's doing real addition, and it's giving you addition for any arbitrary pair of numbers, even though it hasn't seen. 00:06:24.560 --> 00:06:25.039 Right. 00:06:25.039 --> 00:06:39.359 And so, what interpretability has to do with that is that we're looking inside what the box is doing and see exactly what's happening inside it, how is it processing information, which parts of the network are actually solving this problem? 00:06:39.359 --> 00:06:44.959 What happens if we take out a part of this network? 00:06:44.959 --> 00:06:50.560 Does it change what kinds of outputs you get from the network? 00:06:50.560 --> 00:06:52.479 So that would be a use case. 00:06:53.839 --> 00:07:00.719 So what getting back a little bit to kind of the terminology, so the audience understands that. 00:07:00.719 --> 00:07:01.439 Sure, sure. 00:07:01.439 --> 00:07:08.639 When you when you mean LLM or you know, large language model or local language model, which are not the same thing, right? 00:07:08.639 --> 00:07:15.359 But but what is that for someone who you know maybe is heard, maybe is playing with chat GPT? 00:07:15.359 --> 00:07:17.599 What is actually that thing? 00:07:18.399 --> 00:07:23.439 Well, mechanistically, what's going on is in some ways is very simple to explain. 00:07:23.439 --> 00:07:32.719 Basically, what happens is all of your words get changed into numbers, and each of those numbers corresponds to a different word. 00:07:32.719 --> 00:07:42.239 And then there's a whole the internal part of the large language model is like a stack of pancakes, like you know, 30, 40, 50. 00:07:42.239 --> 00:07:52.159 So it's like this set of layers, and what you're giving it gets sent to the first layer, which does some processing. 00:07:52.159 --> 00:07:58.159 That processing then gets sent to the second layer, and that just keeps on going on and on and on. 00:07:58.159 --> 00:08:07.119 And what people are surprised about is that when you get to the top layer, the output is just one word, it's the next word. 00:08:07.119 --> 00:08:23.039 So it sees everything in your sentence or paragraph or everything that you've given it, and it provides you with one additional word at the end of that whole layering process, and then that becomes the new input for the next cycle. 00:08:23.439 --> 00:08:24.000 Sure. 00:08:24.000 --> 00:08:24.879 Okay. 00:08:24.879 --> 00:08:35.519 So, but it's one of those things where I think when people look at what's happening with AI today, it's been around in various forms for several years. 00:08:35.519 --> 00:08:38.159 And what had what changed? 00:08:38.159 --> 00:08:42.160 What kind of made it such a hot topic? 00:08:42.160 --> 00:08:48.080 When it's kind of been around, at least like I know in the imaging space we were talking earlier, it's been around for a while. 00:08:49.280 --> 00:08:49.600 Sure. 00:08:49.600 --> 00:08:59.440 And the thing is that AI is a difficult word because in the 70s it meant something quite different from what people are using it to mean today. 00:08:59.440 --> 00:09:03.600 So in the 70s, what we had what were called symbolic systems. 00:09:03.600 --> 00:09:07.040 And these were systems that were basically just like computer programs. 00:09:07.040 --> 00:09:19.040 So you told it when it saw this input, you know, it would go through all this data percolate, and you have symbols that represent, you know, here's a green block, put it on top of the red block or what have you. 00:09:19.040 --> 00:09:26.240 These new systems really weren't at one point called AI, they were called neural networks connection systems. 00:09:26.240 --> 00:09:27.600 And that's what I worked in. 00:09:27.600 --> 00:09:35.440 And it's only been in the recent past that what these models, or they were also called, you know, let's say machine learning systems. 00:09:35.600 --> 00:09:35.920 Yeah. 00:09:36.399 --> 00:09:52.320 But all of those things have kind of been swept to the side because neural networks and large language models, which neural networks are kind of large language model, are now so successful at solving so many of these problems. 00:09:52.320 --> 00:09:55.280 It's really kind of the only game in town. 00:09:55.280 --> 00:10:02.960 And the reason for that, in answer to your question, is that in 2017, a paper came out with a new architecture. 00:10:02.960 --> 00:10:06.800 And that it was called Attention is all you need. 00:10:06.800 --> 00:10:32.560 And that was kind of a watershed moment in the field, because it's very soon after that that companies like OpenAI and Anthropic and Google and Microsoft began employing this new kind of architecture and learning that oh my God, this thing can really interact with people as though it were an intelligent entity. 00:10:32.560 --> 00:10:40.000 There's a whole sort of subculture of people that fall in love with their AIs and that think that they're real and so on and so forth. 00:10:40.000 --> 00:10:43.120 And you know, I try to explain to them, look, it's just numbers running through a machine. 00:10:43.120 --> 00:10:56.000 And the thing that I should point out is that as amazing as this technology is, this kind of puts me in mind of the fact that it still is very, very far from what the human brain does. 00:10:56.000 --> 00:10:59.280 And

there's just so many differences. 00:10:59.280 --> 00:11:10.639 One of the really fascinating things, we hear about these data centers that need massive amounts of energy and how much energy it is just for one query to get processed. 00:11:10.639 --> 00:11:16.240 And we know that the brain uses like as much energy as a light bulb or something like that. 00:11:16.240 --> 00:11:16.960 Yeah, yeah. 00:11:16.960 --> 00:11:27.680 So there's a quarter of magnitude difference between what your LLM is doing in terms of energy usage and what you're doing in terms of energy usage. 00:11:27.680 --> 00:11:41.920 And that in and of itself is a big clue to the fact that these are really kind of just early models that to get something that uses a lot of the tricks that human brains use, we're gonna have to radically redesign the architectures. 00:11:42.080 --> 00:11:42.399 Yeah. 00:11:42.560 --> 00:12:04.240 But it was that shift in 2017 in that paper, I think, and that I remember, I guess it was in 2022 that people really began realizing how powerful these models could be and how you could type in something and get these massively complex and you know, sometimes quite wrong, but you know, massively complex responses. 00:12:04.240 --> 00:12:06.399 And that was how the change came about. 00:12:06.639 --> 00:12:20.639 And that was about the time, like in the imaging space where the generative AI stuff started coming in, where you could describe a photograph with you know words, and it would create it basically. 00:12:20.639 --> 00:12:29.040 It would be created either by using yes, and then of course, there's the whole ethics of that where our metal where models are being trained and all those kind of things. 00:12:29.120 --> 00:12:29.440 Right. 00:12:29.440 --> 00:12:53.600 So yeah, and of course, you know, the sort of predecessor to the current image models was that sort of dream model that Google came up with, where it was able to create these images that had this these very bizarre characteristics of dreams, and then as they furthered their work and refined it, they made it so that you could actually shape what came out to be something less dreamlike and more realistic. 00:12:53.600 --> 00:12:55.440 And that's the stage that we're at now. 00:12:55.759 --> 00:13:02.800 And now it's you know, really affecting like you know, entertainment and a bunch of other industries for sure in the imaging space. 00:13:02.800 --> 00:13:30.720 But when we were talking beforehand about like applications for AI now, quote unquote, in the real world, you know, some of the things that we were talking about were developing an authentic voice that relates to your actual voice because it just because you put in often, because as a typical business person would, hey, I need a marketing campaign for our new film promotion. 00:13:30.720 --> 00:13:40.240 Give me some text for that, and it comes out in usually needs to be rewritten because it's the bases are there, but but it's wrong tonally. 00:13:40.240 --> 00:13:43.040 But you've done some work on that piece of it. 00:13:43.040 --> 00:13:45.040 Can you talk a little bit about that? 00:13:45.519 --> 00:13:46.320 Oh, sure. 00:13:46.320 --> 00:13:55.200 So, this is something that I'm working on, and it's there's that website where you can put in your email address and get information about it as it progresses. 00:13:55.200 --> 00:13:57.920 It's called your voicecraft.ai. 00:13:57.920 --> 00:14:03.920 And the idea is for it to learn how to create text using your voice. 00:14:03.920 --> 00:14:07.360 And there, I should step back here and explain. 00:14:07.360 --> 00:14:14.720 When you use Chat GPT or Claude or any of those models, you're using what's called a frontier model. 00:14:14.720 --> 00:14:22.240 And what that means is something you type in gets sent over the internet to some server farm somewhere belonging to these companies. 00:14:22.240 --> 00:14:28.240 And one thing that you can't do is really protect that information. 00:14:28.240 --> 00:14:40.399 So if you're using something that, let's say, has HIPAA regulations on it or any number of other reasons why you might not want it to end up out on the internet somehow, you really can't use it. 00:14:40.399 --> 00:14:45.440 And the other thing is you have to remember that those models cannot be trained by you. 00:14:45.440 --> 00:14:47.759 They're trained by companies. 00:14:47.759 --> 00:14:59.759 So if you have your own text, you have blog posts and marketing data and emails, all these things that are in your voice. 00:14:59.759 --> 00:15:09.360 There's no way for you to actually change the weights of that frontier model so that it will sound more like you. 00:15:09.360 --> 00:15:15.040 Now, there's something called a local, which you alluded to before, a local large language model. 00:15:15.040 --> 00:15:21.280 And all that is, is a large language model that's living on your computer locally. 00:15:21.680 --> 00:15:21.920 Okay. 00:15:22.160 --> 00:15:26.160 And so it's a little bit smaller, but still very powerful. 00:15:26.160 --> 00:15:33.120 And, you know, it's like I kind of liken it to the difference between like a battleship and a speed boat. 00:15:33.120 --> 00:15:40.639 I mean, a battleship is great for certain applications, but if you just want to get across the water, a speed boat will work just as well for you. 00:15:40.639 --> 00:15:49.280 And so these local language models that live right on your computer have a lot of advantages to them. 00:15:49.280 --> 00:15:59.680 One of the advantages is that because they're local, you can be completely air gapped from the internet and you don't have to worry about that information leaking out. 00:15:59.680 --> 00:16:19.360 Another advantage is that you don't pay for tokens because it's also you've heard these horror stories of these companies with tens of thousands of dollars of bills because you know they're just using so much compute and so many tokens that yeah

for all sorts of little side projects, and so you don't have to pay for that. 00:16:19.360 --> 00:16:33.120 And then the big thing here, as regards our interest in authenticity and voice and so forth, is that you can actually tune the models themselves. 00:16:33.120 --> 00:16:55.040 So, unlike with the Frontier model, you can actually change the weights of the models themselves so that they reflect more the kind of advertising copy or the kind of blog posts or whatever it is that you traded on that you do so that the voice that comes out sounds much more like something you would actually write. 00:16:55.040 --> 00:16:57.680 But at the same time, it's not a costume. 00:16:57.680 --> 00:17:06.639 It doesn't just change a few words on the surface and move things around and make it sound how I think it should sound to sound more like you. 00:17:06.639 --> 00:17:21.279 It's really doing a deep change on the weights and actually giving you something that really is at a very deep level, completely different from the baseline text that it would otherwise give you if you typed in a prompt. 00:17:21.519 --> 00:17:23.279 So, what is something like that running on? 00:17:23.279 --> 00:17:36.000 Like, you know, we've heard Claude, like that you or that you can put, you know, like for example, I've heard that you know, the Mac Mini is a great device for this, and your people are putting I what's it called? 00:17:36.000 --> 00:17:37.839 OpenClaw or different things like that. 00:17:37.839 --> 00:17:44.159 So, what does it take to do that sort of thing, to create your own local language machine? 00:17:44.559 --> 00:17:58.159 Just any relatively higher-end laptop or desktop machine, and that's another kind of cool thing about this is you don't need a $20,000 machine to get it to work. 00:17:58.159 --> 00:18:04.000 And so if you have, would it work on some sort of lower-end Chromebook? 00:18:04.000 --> 00:18:05.439 I'm not sure about that. 00:18:05.439 --> 00:18:06.639 Um maybe. 00:18:06.639 --> 00:18:19.119 But if you had a decently powerful Mac or PC, it should be able to run fine on that because a lot of what you're doing, you're not really doing the training is done elsewhere. 00:18:19.119 --> 00:18:22.959 So all you need to do is actually get the model to run. 00:18:22.959 --> 00:18:29.279 And because it's specifically designed to be smaller and to fit on your machine, it works just fine. 00:18:29.279 --> 00:18:57.609 And there are all sorts of little tricks that they use to basically sort of squeeze it down to laptop size. 00:18:57.849 --> 00:19:05.370 But you're trying to train it to really learn more about your business, your activity, your whatever, which is a much smaller data set. 00:19:06.170 --> 00:19:24.569 Yeah, I mean, if you wanted to write a 150-page spec script, then I guess you would probably use a frontier model, but you'd still have to go through the frontier model and see what it's doing because I've actually had it do things like that, and

they weren't very good to be able to very shade and not very good, but which has to do with its training regimen. 00:19:24.569 --> 00:19:33.849 So I don't know if you could get it to write a 150-page spec script, but you might get it to write like a 10-page treatment or something like that, more in your voice. 00:19:33.849 --> 00:19:34.250 Yeah. 00:19:34.250 --> 00:19:36.250 And that's the big difference between the two. 00:19:36.250 --> 00:19:48.730 And of course, as these models become more sophisticated and as our computers become more powerful, that line is going to keep moving and moving because we're going to be able to do more and more things locally. 00:19:48.730 --> 00:19:54.490 And I think that local model is going to be something you're going to be seeing a lot more of in the near future. 00:19:54.730 --> 00:20:12.730 I just think not only because of the concern that people have with the LLMs and the ethics involved with what they're being trained on and things like that, but just that's like you said, you know, if you're paying token upon token, that can that that can run up significantly. 00:20:13.210 --> 00:20:20.890 And I think that's a real concern for people because you know, people don't really, when they're using it, think about, oh, I want to try a hundred different variations of this. 00:20:20.890 --> 00:20:27.370 They don't think about the token usage that's involved in that, especially if you're in a company because you're not paying for it yourself. 00:20:27.370 --> 00:20:27.849 Sure. 00:20:27.849 --> 00:20:37.049 And if you're using a local model, you can throw whatever you want to at it, and it's just gonna, you know, it'll use more electricity, but that's about it. 00:20:37.450 --> 00:20:51.289 Because one of the things I think that people have come to learn is AI is is a tool, like you said, for text generation or whatever, but it still needs some oversight in terms of fact-checking. 00:20:51.289 --> 00:20:56.250 You should never generate something in AI and then just automatically assume it's correct. 00:20:56.250 --> 00:21:00.890 But with a local model, you may actually have higher confidence. 00:21:01.769 --> 00:21:02.970 I think it's possible. 00:21:02.970 --> 00:21:05.529 That's an interesting hypothesis. 00:21:05.529 --> 00:21:10.569 I think a local model could still make mistakes just the way a frontier model does. 00:21:10.569 --> 00:21:22.730 But I think the difference is that if you were trying to get the local model just to follow your own voice, that's probably something that's going to need less monitoring because again, you've trained it on your voice already. 00:21:22.730 --> 00:21:27.210 So you know that what's going to come out is still pretty close to the way you would write. 00:21:27.210 --> 00:21:33.210 I'd say always check the work of your because they can make mistakes. 00:21:33.210 --> 00:21:35.130 And you can even ask it to check its own work. 00:21:35.130 --> 00:21:42.009 You can say, what are

go through here and say, tell me what are all the flaws in what you've said, or what are the weak points. 00:21:42.089 --> 00:21:42.410 Yeah. 00:21:42.650 --> 00:21:48.490 And you can even, you know, give it to a different LLM, like from ChatGPT to Claude, and have them check each other. 00:21:48.490 --> 00:21:57.690 So there are a lot of really interesting little tricks that you can use to make sure that what you're getting is accurate. 00:21:58.009 --> 00:22:00.170 So, in the case like a small business person, right? 00:22:00.170 --> 00:22:04.490 Let's say, for example, I need some marketing help and I want it in my voice. 00:22:04.490 --> 00:22:13.769 I've built up relationships with my community, I know who my customers are, I put that in my LLM and I know how all my people are, and I want this help. 00:22:13.769 --> 00:22:24.890 What would be some of the advantages of using this type of system for a small business for that type of thing, as opposed to just doing it the old way, just not doing it at all? 00:22:25.289 --> 00:22:31.690 Well, because I think it does save you time compared to doing it the old way. 00:22:31.690 --> 00:22:39.769 And If you have, let's say, 10 different variations of an ad that you want to have generated. 00:22:39.769 --> 00:22:50.009 And let's say they're targeted at it's all the same product or all the same idea, but one is targeted for Facebook, one is targeted for LinkedIn, Instagram, whatever. 00:22:50.009 --> 00:22:57.450 This is a great way to speed up your iterative process of getting what you want. 00:22:57.450 --> 00:23:03.450 Because what you can do is have it do the base marketing text for you. 00:23:03.450 --> 00:23:04.970 Check it out and make sure. 00:23:04.970 --> 00:23:07.130 And it's if it's marketing, it's probably not very long. 00:23:07.130 --> 00:23:08.890 So it's not going to take you long to check it. 00:23:08.890 --> 00:23:20.490 And then have it generate variations for what you want it to do for those various targeted, you know, Facebook and Instagram and so forth. 00:23:20.490 --> 00:23:28.490 And I have a friend who actually is in marketing, and he said that it's amazing to him how much time these systems have saved him. 00:23:28.490 --> 00:23:42.329 Now that doesn't always work out that way, but he did have, for instance, he was working on a PowerPoint presentation, and he said it did for him in about four or five minutes what would have taken him several hours to get done. 00:23:42.329 --> 00:23:42.730 Yeah. 00:23:42.730 --> 00:23:44.650 And so you hear a lot of stories like that. 00:23:44.650 --> 00:23:45.849 And I think those are valid. 00:23:45.849 --> 00:23:48.970 And I think that that jibes with what I've seen. 00:23:48.970 --> 00:23:52.089 So that would be time saving, I think is the big one. 00:23:52.089 --> 00:23:55.609 Also, LLMs can generate ideas for you. 00:23:55.609 --> 00:24:02.329 And I've actually had ideas generated for me by LLMs that were not necessarily things that I

would have come up with my own. 00:24:02.329 --> 00:24:04.809 They do have a creative component. 00:24:04.809 --> 00:24:10.410 So the notion that they're just kind of spitting back what you give in to them is not necessarily the case. 00:24:10.410 --> 00:24:15.049 And you give them a whole bunch of different parameters and say, give me some ideas. 00:24:15.049 --> 00:24:18.250 So you could use the same system for giving you marketing ideas. 00:24:18.250 --> 00:24:18.650 Right. 00:24:18.650 --> 00:24:24.089 Instead of giving it the text that you want it to, you know, flavor. 00:24:24.089 --> 00:24:28.009 What you would do is just say, here's the product, here's the target audience. 00:24:28.250 --> 00:24:35.289 And maybe here's past sales campaigns or sales information from previous events we've done. 00:24:35.289 --> 00:24:35.609 Exactly. 00:24:35.609 --> 00:24:41.210 Look to see what worked, what didn't work, and give me some new ideas. 00:24:41.529 --> 00:24:42.170 Exactly. 00:24:42.170 --> 00:24:42.890 Exactly. 00:24:42.890 --> 00:24:47.289 And that for that, a lot of that would be the pre-training that you do when you're adjusting the weights. 00:24:47.289 --> 00:24:48.410 But that's exactly. 00:24:48.410 --> 00:25:03.930 So I think that as people get used to incorporating that more and more into their workflow, they'll begin to understand the value of it and they'll begin to sort of appreciate what these models are good at and what they're not as good at. 00:25:03.930 --> 00:25:13.049 And so once people get this sort of intuitive feel for it, time will be saved because you'll know I'm going to use it for this because I know it's good or this. 00:25:13.049 --> 00:25:15.930 I'm not going to use it for that because I know it's not good for that. 00:25:15.930 --> 00:25:17.609 And we're not really at that stage yet. 00:25:17.849 --> 00:25:21.690 But it is an amazing time, like you said, if from terms of a productivity standpoint. 00:25:21.690 --> 00:25:27.370 Because I know you know you see a lot in the marketing world about A-B testing and all these things. 00:25:27.370 --> 00:25:34.329 But really, you know, if if you're a small business owner and you just need to get stuff done, you may not have time to do Acting. 00:25:34.329 --> 00:25:41.690 And this gives you the opportunity to create those sort of things so you can be more competitive with people who are doing A-B testing or things. 00:25:42.089 --> 00:25:42.730 Yes, exactly. 00:25:42.730 --> 00:25:43.130 Yeah. 00:25:43.130 --> 00:25:47.210 And so it gives the smaller companies a bit of an advantage, I think. 00:25:47.210 --> 00:25:51.450 You know, lets them get up to the same level as larger organizations. 00:25:51.609 --> 00:26:03.769 Yeah, because I I just find it interesting how, like, you know, big companies kind of create standards for behavior, you know, like you know, the you know, one-click shopping cart that Amazon created, everyone else copied it, right? 00:26:03.769 --> 00:26:16.730 And I think a lot of these big companies are going to be doing AI-driven marketing, and then smaller companies will be able to take advantage of the technology as it migrates downstream at a more local level to do these things. 00:26:17.049 --> 00:26:18.569 Yep, absolutely. 00:26:19.210 --> 00:26:26.890 Well, where can people go for more information to learn about what it is you do at your company? 00:26:27.609 --> 00:26:29.289 So there are two places. 00:26:29.289 --> 00:26:38.569 One is I have a Substack and I publish articles there on a regular basis, mostly about what I was talking about interpretability, other areas of LLMs as well. 00:26:38.569 --> 00:26:41.529 And that is inside the blackbox.ai. 00:26:41.529 --> 00:26:42.490 Awesome. 00:26:42.490 --> 00:26:44.329 And that's again, that's on Substack. 00:26:44.329 --> 00:26:54.089 And then if they want to, you know, sign up to see what's going on with your voicecraft, which is the app that learns to speak in your voice, as that progresses. 00:26:54.089 --> 00:26:56.410 I'll be sending out information on that. 00:26:56.410 --> 00:26:58.650 And that's your voicecraft.ai. 00:26:58.650 --> 00:27:00.569 So they're both.ai. 00:27:01.370 --> 00:27:01.849 Awesome. 00:27:01.849 --> 00:27:02.730 Well, thank you so much. 00:27:02.730 --> 00:27:03.370 I appreciate it. 00:27:03.370 --> 00:27:06.410 I've learned a lot from this conversation. 00:27:06.410 --> 00:27:10.890 It's I think it'll be valuable for people who then really thought how they could do something like this. 00:27:10.890 --> 00:27:15.769 Kind of roll your own AI, I guess, is kind of exactly. 00:27:16.009 --> 00:27:17.049 That's what we're talking about. 00:27:17.289 --> 00:27:17.690 Awesome. 00:27:17.690 --> 00:27:18.650 Well, thank you so much. 00:27:18.650 --> 00:27:19.370 Appreciate your time. 00:27:19.930 --> 00:27:20.329 Thank you. 00:27:20.329 --> 00:27:21.930 Thanks a lot. 00:27:24.650 --> 00:27:28.089 Thank you for listening to the Dead Pixel Society podcast. 00:27:28.089 --> 00:27:35.370 Read more great stories and sign up for the newsletter at www.the dead pixels society.com.

I-edit ang transcript