1 00:00:04,280 --> 00:00:07,880 Speaker 1: Hello, Hello, Welcome to Smart Talks with IBM, a podcast 2 00:00:07,960 --> 00:00:13,560 Speaker 1: from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. This season, 3 00:00:13,600 --> 00:00:16,919 Speaker 1: we're diving back into the world of artificial intelligence, but 4 00:00:17,040 --> 00:00:22,919 Speaker 1: with a focus on the powerful concept of open its possibilities, implications, 5 00:00:23,079 --> 00:00:26,400 Speaker 1: and misconceptions. We'll look at openness from a variety of 6 00:00:26,440 --> 00:00:30,560 Speaker 1: angles and explore how the concept is already reshaping industries, 7 00:00:30,600 --> 00:00:34,199 Speaker 1: ways of doing business and a very notion of what's possible. 8 00:00:34,560 --> 00:00:38,040 Speaker 1: In today's episode, I sat down with mo Duffy, software 9 00:00:38,080 --> 00:00:41,840 Speaker 1: engineering manager at red Hat, who works on instruct Lab, 10 00:00:42,080 --> 00:00:46,040 Speaker 1: a project co developed by red Hat and IBM. Most 11 00:00:46,040 --> 00:00:50,800 Speaker 1: shared with me how this new initiative is revolutionizing AI training, 12 00:00:51,120 --> 00:00:54,400 Speaker 1: making it not only more accessible, but also more inclusive. 13 00:00:55,280 --> 00:00:59,280 Speaker 1: This project, unique in the industry, allows developers to submit 14 00:00:59,440 --> 00:01:04,800 Speaker 1: incremental contributions to one base AI model, creating a continuous 15 00:01:04,840 --> 00:01:09,200 Speaker 1: loop of development, much like normal open source software. By 16 00:01:09,319 --> 00:01:14,280 Speaker 1: leveraging community contributions and IBM's cutting edge granite models, Mo 17 00:01:14,520 --> 00:01:17,520 Speaker 1: in the team of ibmrs and red hatters are paving 18 00:01:17,560 --> 00:01:20,920 Speaker 1: the way for a future where AI development is a 19 00:01:20,959 --> 00:01:25,920 Speaker 1: communal endeavor. Our insights into open source software extend beyond 20 00:01:26,040 --> 00:01:31,520 Speaker 1: technical proficiency to the profound impact of collaborative effort. At 21 00:01:31,520 --> 00:01:35,440 Speaker 1: the heart of Moe's work is a belief in democratizing technology, 22 00:01:36,000 --> 00:01:40,440 Speaker 1: ensuring that AI becomes a tool accessible to all. So 23 00:01:40,560 --> 00:01:44,360 Speaker 1: let's explore how Mo, red Hat and IBM are empowering 24 00:01:44,360 --> 00:01:48,560 Speaker 1: individuals and businesses alike to reshape the future of technology 25 00:01:48,760 --> 00:01:57,640 Speaker 1: through collaboration and innovation. Mo, thank you for joining me 26 00:01:57,640 --> 00:02:01,000 Speaker 1: today so much, for you have just about the most 27 00:02:01,040 --> 00:02:06,760 Speaker 1: Irish name ever. I do very proud you weren't born 28 00:02:06,760 --> 00:02:07,360 Speaker 1: in Ireland. 29 00:02:07,560 --> 00:02:09,320 Speaker 2: No, my grandparents or your grandparents. 30 00:02:09,480 --> 00:02:10,560 Speaker 1: I see, where did you grow up? 31 00:02:10,960 --> 00:02:11,960 Speaker 2: New York Queens? 32 00:02:12,280 --> 00:02:15,520 Speaker 1: Oh you're a see. So tell me a little bit 33 00:02:15,560 --> 00:02:18,280 Speaker 1: about how how you got to red hat? What was 34 00:02:18,320 --> 00:02:18,800 Speaker 1: your path? 35 00:02:19,280 --> 00:02:21,800 Speaker 2: When I was in high school, it was a chatty girl, 36 00:02:22,160 --> 00:02:24,640 Speaker 2: teenage girl on the phone. We had one phone line. 37 00:02:24,960 --> 00:02:27,760 Speaker 2: My older brother was studying at the local state college 38 00:02:27,800 --> 00:02:30,360 Speaker 2: computer science, and he had to tell that end to 39 00:02:30,400 --> 00:02:33,240 Speaker 2: compile his homework. One phone line and I'm on it 40 00:02:33,280 --> 00:02:36,239 Speaker 2: all the time. He got very frustrated and he needed 41 00:02:36,240 --> 00:02:38,799 Speaker 2: a compiler to do his homework. So he bought red 42 00:02:38,840 --> 00:02:43,240 Speaker 2: hat Linux from a CompUSA, brought it home and that 43 00:02:43,360 --> 00:02:45,320 Speaker 2: was on the family computer. So I learned Linux and 44 00:02:45,320 --> 00:02:48,040 Speaker 2: I started playing around with it. I really liked it 45 00:02:48,120 --> 00:02:52,000 Speaker 2: because you could customize everything, like the entire user interface. 46 00:02:52,360 --> 00:02:54,560 Speaker 2: You could actually modify the code of the programs you 47 00:02:54,560 --> 00:02:56,840 Speaker 2: were using to do what you wanted. And for me, 48 00:02:56,880 --> 00:02:59,079 Speaker 2: it was really cool because, especially when you're a kid 49 00:02:59,120 --> 00:03:00,880 Speaker 2: and like people tell you this is the way things 50 00:03:00,880 --> 00:03:02,639 Speaker 2: are and you just have to deal with it, it's 51 00:03:02,720 --> 00:03:04,280 Speaker 2: nice to be like, I'm going to make things the 52 00:03:04,320 --> 00:03:07,600 Speaker 2: way I want, modify the code and playing. Yeah, it 53 00:03:07,680 --> 00:03:09,520 Speaker 2: was amazing and it was just such a time and 54 00:03:09,560 --> 00:03:12,600 Speaker 2: like before it was cool, I was doing it and 55 00:03:13,360 --> 00:03:15,320 Speaker 2: what I saw on that is sort of the potential 56 00:03:15,360 --> 00:03:18,400 Speaker 2: like number one of like a community of people working together. 57 00:03:18,440 --> 00:03:21,480 Speaker 2: And like the Internet existed, it was slow, it involved modems, 58 00:03:21,880 --> 00:03:23,840 Speaker 2: but there were people that you could talk to who 59 00:03:23,840 --> 00:03:27,600 Speaker 2: would give you tips and you'd share information, and this 60 00:03:27,720 --> 00:03:31,400 Speaker 2: collaborative building something together is really something special. Right. I 61 00:03:31,440 --> 00:03:35,240 Speaker 2: could file a complaint to whatever large software company made 62 00:03:35,240 --> 00:03:37,440 Speaker 2: whatever software I was into, or I could go to 63 00:03:37,480 --> 00:03:39,640 Speaker 2: an open source software community and be like, hey, guys, 64 00:03:39,680 --> 00:03:42,000 Speaker 2: I think we should do this. I'm like, yeah, okay, 65 00:03:42,080 --> 00:03:44,200 Speaker 2: I'll help. I'll pitch in so you don't feel powerless. 66 00:03:44,200 --> 00:03:45,600 Speaker 2: You feel like you can have an impact, and that 67 00:03:45,640 --> 00:03:49,560 Speaker 2: was really exciting to me. However, open source software has 68 00:03:49,680 --> 00:03:52,920 Speaker 2: a reputation for not having the best user interface, not 69 00:03:52,960 --> 00:03:57,480 Speaker 2: the best user experience. So I ended up studying computer 70 00:03:57,560 --> 00:04:01,760 Speaker 2: science and electronic media, and then I did human computeraction 71 00:04:01,920 --> 00:04:04,920 Speaker 2: as my master's And my thought was, wouldn't it be 72 00:04:05,000 --> 00:04:08,960 Speaker 2: nice if this free software accessible to anybody, if it 73 00:04:09,000 --> 00:04:11,120 Speaker 2: was easier to use, some more people could use it 74 00:04:11,160 --> 00:04:14,440 Speaker 2: and take advantage of it. And so, long story short, 75 00:04:14,600 --> 00:04:16,919 Speaker 2: I ended up going to Red Hat saying, Hey, I 76 00:04:16,920 --> 00:04:18,520 Speaker 2: want to learn how you guys work. Let me embed 77 00:04:18,560 --> 00:04:21,159 Speaker 2: in your team draft out of my graduate program, and 78 00:04:21,200 --> 00:04:22,880 Speaker 2: I'm like, I want to do this for a living. 79 00:04:23,000 --> 00:04:25,560 Speaker 2: This is cooler. So I thought this is the way 80 00:04:25,600 --> 00:04:27,240 Speaker 2: to go, and I've been there ever since. They haven't 81 00:04:27,240 --> 00:04:28,080 Speaker 2: been able to get rid of me. 82 00:04:30,160 --> 00:04:32,880 Speaker 1: To backtrack just a little bit, you were talking about 83 00:04:32,920 --> 00:04:36,640 Speaker 1: the sense of community that surrounds this way of thinking 84 00:04:36,680 --> 00:04:40,400 Speaker 1: about software. Talk a little bit more about what that 85 00:04:40,400 --> 00:04:43,240 Speaker 1: community is like, the benefits of that community, why it 86 00:04:43,240 --> 00:04:44,080 Speaker 1: appeals to you. 87 00:04:44,560 --> 00:04:46,880 Speaker 2: Sure, well, you know part of the reason I actually 88 00:04:46,960 --> 00:04:50,040 Speaker 2: ended up going to the graduate school track. Suddenly you're 89 00:04:50,080 --> 00:04:53,240 Speaker 2: a peer of your professors and you're working side by 90 00:04:53,320 --> 00:04:56,360 Speaker 2: side with them. At some point they retire and you're 91 00:04:56,400 --> 00:04:59,839 Speaker 2: in the next generation. So it's sharing information, building on 92 00:04:59,839 --> 00:05:02,640 Speaker 2: the work of others in sort of this cycle that 93 00:05:02,839 --> 00:05:07,240 Speaker 2: extends past human lifespan and in the same way, like 94 00:05:07,279 --> 00:05:10,719 Speaker 2: the open source model is very similar, but you're actually 95 00:05:10,760 --> 00:05:13,400 Speaker 2: you're building something, and it's something in me. I'm just 96 00:05:13,800 --> 00:05:16,200 Speaker 2: really attracted, Like I don't like talking about stuff. I 97 00:05:16,320 --> 00:05:20,280 Speaker 2: like doing stuff with open source software. The software doesn't 98 00:05:20,279 --> 00:05:23,919 Speaker 2: cost anything, the code is out there, generally uses open 99 00:05:23,960 --> 00:05:27,359 Speaker 2: standards for the file formats. I can open up files 100 00:05:27,400 --> 00:05:30,039 Speaker 2: that I created and open source tools as a high 101 00:05:30,040 --> 00:05:33,599 Speaker 2: school student today because they were using open formats and 102 00:05:33,640 --> 00:05:36,120 Speaker 2: that software still exists, I can still compile the code 103 00:05:36,240 --> 00:05:39,400 Speaker 2: and it's an active community project. Like these things can 104 00:05:39,440 --> 00:05:42,120 Speaker 2: outlast any single company in the same way that the 105 00:05:42,160 --> 00:05:44,760 Speaker 2: academic community has been going on for so many years, 106 00:05:44,760 --> 00:05:47,320 Speaker 2: and hopefully we'll continue moving on. So it's sort of 107 00:05:47,360 --> 00:05:50,240 Speaker 2: like not just the community around it, but just the 108 00:05:50,320 --> 00:05:53,440 Speaker 2: knowledge sharing and also bringing up the next generation as well. 109 00:05:53,480 --> 00:05:55,120 Speaker 2: Like all of that stuff really appealed to me. And 110 00:05:55,160 --> 00:05:58,279 Speaker 2: also at the center of it the fact that we 111 00:05:58,320 --> 00:06:01,680 Speaker 2: could democratize it by following this open source process and 112 00:06:01,800 --> 00:06:03,760 Speaker 2: feel like we have some control. We're not at the 113 00:06:03,760 --> 00:06:06,839 Speaker 2: mercy of some faceless corporation making changes and we have 114 00:06:06,920 --> 00:06:08,840 Speaker 2: no impact. Like that really appealed to me too. 115 00:06:09,640 --> 00:06:14,719 Speaker 1: For those of us who are not software phisionados, take 116 00:06:14,720 --> 00:06:19,120 Speaker 1: a step backwards and give me a kind of description 117 00:06:19,240 --> 00:06:22,120 Speaker 1: of terms. What's the opposite of open to proprietary? 118 00:06:22,200 --> 00:06:23,680 Speaker 2: Proprietary is what we say, So. 119 00:06:24,520 --> 00:06:28,320 Speaker 1: Specifically and practically, the difference would be what between something 120 00:06:28,320 --> 00:06:30,160 Speaker 1: that was opened us in something that was proprietary. 121 00:06:30,240 --> 00:06:32,839 Speaker 2: Sure, so there's a lot of difference. So with open 122 00:06:32,880 --> 00:06:36,839 Speaker 2: source software you get these rights. When you're given the software, 123 00:06:36,920 --> 00:06:39,320 Speaker 2: you get the right to be able to share it. 124 00:06:39,400 --> 00:06:41,839 Speaker 2: And depending on the lot, different licenses that are considered 125 00:06:41,880 --> 00:06:44,839 Speaker 2: open source have different little things that you have to 126 00:06:44,839 --> 00:06:50,640 Speaker 2: be aware of. With proprietary code, it's one copyright the company. 127 00:06:50,680 --> 00:06:53,040 Speaker 2: Even a lot of times, when you sign your employment 128 00:06:53,040 --> 00:06:55,360 Speaker 2: contract for a software company and you write code for them, 129 00:06:55,680 --> 00:06:57,320 Speaker 2: you don't own it. You sign over your rights to 130 00:06:57,360 --> 00:06:59,280 Speaker 2: the company, So if you leave the company, the code 131 00:06:59,279 --> 00:07:01,599 Speaker 2: doesn't go with you. It stays in the ownership of 132 00:07:01,600 --> 00:07:04,520 Speaker 2: that company. So then one one company buys out another 133 00:07:04,560 --> 00:07:06,440 Speaker 2: and kills a product that code's gone. 134 00:07:06,560 --> 00:07:10,480 Speaker 1: It's gone. For a business, Why would a business want 135 00:07:10,560 --> 00:07:14,360 Speaker 1: to be have open source code as opposed to proprietary. 136 00:07:13,760 --> 00:07:15,800 Speaker 2: Well, for the same reasons. Like say you're a business. 137 00:07:16,280 --> 00:07:20,160 Speaker 2: You've invested all this money into this software platform, right, 138 00:07:20,640 --> 00:07:23,800 Speaker 2: and you've upskilled your employees on it, and it's a 139 00:07:23,800 --> 00:07:26,640 Speaker 2: core part of your business, and then a few years 140 00:07:26,680 --> 00:07:30,000 Speaker 2: later that company goes out of business or something happens, 141 00:07:30,160 --> 00:07:33,520 Speaker 2: or even something less drastic. You really need this feuture. 142 00:07:33,840 --> 00:07:36,960 Speaker 2: But for the company that makes the software, it's not 143 00:07:37,000 --> 00:07:39,840 Speaker 2: in their best interests. It's not worth the investment. They're 144 00:07:39,840 --> 00:07:41,760 Speaker 2: not going to do it. How do you get that feature? 145 00:07:42,120 --> 00:07:44,680 Speaker 2: You either have to completely migrate to another solution, and 146 00:07:44,680 --> 00:07:46,600 Speaker 2: this is something that's core at your business, that's going 147 00:07:46,640 --> 00:07:48,760 Speaker 2: to be a big deal to migrate. But if it's 148 00:07:48,800 --> 00:07:52,560 Speaker 2: open source, you could either hire a team of experts. 149 00:07:52,760 --> 00:07:55,320 Speaker 2: You could hire software engineers who are able to go 150 00:07:55,400 --> 00:07:58,640 Speaker 2: do this for you. Go in the upstream software community, 151 00:07:59,120 --> 00:08:02,120 Speaker 2: implement the future that you want, and it'll be rolled 152 00:08:02,160 --> 00:08:04,720 Speaker 2: into the next version of that company software. So even 153 00:08:04,760 --> 00:08:07,920 Speaker 2: if that company didn't want to implement the feature, if 154 00:08:07,920 --> 00:08:10,960 Speaker 2: they did it open source, they would inherit that feature 155 00:08:11,000 --> 00:08:13,520 Speaker 2: from the upstream community is what we call it, so 156 00:08:13,560 --> 00:08:16,200 Speaker 2: you have some control over the situation. If it's open source, 157 00:08:16,280 --> 00:08:20,440 Speaker 2: you have an opportunity to actually affect change in the product, 158 00:08:20,720 --> 00:08:22,760 Speaker 2: and you could then pick it up or pay somebody 159 00:08:22,760 --> 00:08:24,840 Speaker 2: else to pick it up, or another company could form 160 00:08:24,880 --> 00:08:27,480 Speaker 2: and pick it up and keep it going. So there's 161 00:08:27,480 --> 00:08:29,960 Speaker 2: more possibilities. If it's open source, it's more like it's 162 00:08:30,000 --> 00:08:31,560 Speaker 2: like an insurance policy almost. 163 00:08:31,600 --> 00:08:35,160 Speaker 1: So innovation from the standpoint of the customer, innovation is 164 00:08:35,200 --> 00:08:38,200 Speaker 1: a lot easier when you're working in an open source environment. 165 00:08:38,360 --> 00:08:39,000 Speaker 2: Absolutely. 166 00:08:39,400 --> 00:08:43,160 Speaker 1: Yeah. So now at RedHat you're working with something called 167 00:08:43,160 --> 00:08:45,920 Speaker 1: instruct lab. Tell us a little bit about what that is. 168 00:08:46,360 --> 00:08:48,600 Speaker 2: So the thing that really excites me about getting to 169 00:08:48,640 --> 00:08:51,000 Speaker 2: work on this project is AI is sort of that 170 00:08:51,080 --> 00:08:54,200 Speaker 2: has been this scary thing for me because it's one 171 00:08:54,240 --> 00:08:57,360 Speaker 2: of those things like in order to be able to 172 00:08:57,440 --> 00:09:02,520 Speaker 2: pre train a model, you have to have unobtainium GPUs, 173 00:09:03,240 --> 00:09:07,560 Speaker 2: you have to have rich resources, It takes months, it 174 00:09:07,559 --> 00:09:12,520 Speaker 2: takes expertise. There's a small handful of companies that can 175 00:09:12,559 --> 00:09:16,559 Speaker 2: build a model from pre train to something usable, and 176 00:09:16,800 --> 00:09:19,160 Speaker 2: it kind of feels like those early days when I 177 00:09:19,240 --> 00:09:22,000 Speaker 2: was kind of delving in software in the same way. 178 00:09:22,000 --> 00:09:25,320 Speaker 2: I think if more people could contribute to AI models, 179 00:09:26,120 --> 00:09:29,720 Speaker 2: then it wouldn't be just influenced by whichever company had 180 00:09:29,760 --> 00:09:33,000 Speaker 2: the resources to build it. And there's been a lot 181 00:09:33,000 --> 00:09:36,400 Speaker 2: of emphasis on pre training models, so taking massive terabytes 182 00:09:36,480 --> 00:09:40,439 Speaker 2: data sets, throwing them through masses of GPUs over months 183 00:09:40,480 --> 00:09:43,760 Speaker 2: of time, spending hundreds of millions of dollars to build 184 00:09:43,800 --> 00:09:46,640 Speaker 2: a base model. But when instruct lab does is say okay, 185 00:09:46,880 --> 00:09:49,760 Speaker 2: you have a base model, we're going to fine tune in. 186 00:09:49,800 --> 00:09:53,240 Speaker 2: On the other end, it takes less compute resources. The 187 00:09:53,280 --> 00:09:55,559 Speaker 2: way we've built in struck lab, you can play around 188 00:09:55,559 --> 00:09:58,080 Speaker 2: with the technology and learn it on it off the 189 00:09:58,080 --> 00:10:01,520 Speaker 2: shelf laptop that you can actually buy. So in this 190 00:10:01,640 --> 00:10:05,400 Speaker 2: way we're enabling a much broader set of people to 191 00:10:05,480 --> 00:10:08,200 Speaker 2: play with AI, to contribute it, to modify it. And 192 00:10:08,200 --> 00:10:11,160 Speaker 2: I'll tell you one story from red Hat Succi, who 193 00:10:11,240 --> 00:10:15,680 Speaker 2: is our chief diversity officer, very interested in inclusive language 194 00:10:15,679 --> 00:10:18,600 Speaker 2: and open source software, doesn't have any experience with AI. 195 00:10:19,360 --> 00:10:21,400 Speaker 2: We have a community model that we have an upstream 196 00:10:21,440 --> 00:10:24,000 Speaker 2: project around for people to contribute knowledge and skills to 197 00:10:24,040 --> 00:10:26,160 Speaker 2: the model. She's like, I want to teach the model 198 00:10:26,480 --> 00:10:29,440 Speaker 2: how to use inclusive language like replace this word with 199 00:10:29,480 --> 00:10:31,280 Speaker 2: this word, or this word with this word. Oh my, oh, 200 00:10:31,320 --> 00:10:34,199 Speaker 2: that's so cool. So she paired up with Nicholas who 201 00:10:34,280 --> 00:10:37,080 Speaker 2: is a technical guy at red Hat, and they built 202 00:10:37,200 --> 00:10:40,240 Speaker 2: and submitted a skill to the model that you can 203 00:10:40,320 --> 00:10:42,480 Speaker 2: just tell the model, can you please take this document 204 00:10:42,559 --> 00:10:45,000 Speaker 2: and translate this language to more inclusive language and it 205 00:10:45,040 --> 00:10:47,439 Speaker 2: will do it. And they submitted it to the community. 206 00:10:47,520 --> 00:10:49,240 Speaker 2: They were so proud. It was like, that's the kind 207 00:10:49,280 --> 00:10:51,720 Speaker 2: of thing that like, you know, maybe a company would 208 00:10:51,720 --> 00:10:54,160 Speaker 2: be incentivized to do that, but if you have some 209 00:10:54,280 --> 00:10:57,880 Speaker 2: tooling that's open source and something that anybody could access, 210 00:10:58,040 --> 00:11:00,640 Speaker 2: then those communities could actually get together and build that 211 00:11:00,760 --> 00:11:01,960 Speaker 2: knowledge into AI models. 212 00:11:02,200 --> 00:11:06,720 Speaker 1: Just so understand, what you guys have is the structure 213 00:11:06,800 --> 00:11:11,080 Speaker 1: for an AI system, and in other cases, individual companies 214 00:11:11,240 --> 00:11:15,199 Speaker 1: own and train their own AI systems. It takes enormous 215 00:11:15,200 --> 00:11:18,240 Speaker 1: amount of resources. They hoover up all kinds of information, 216 00:11:18,800 --> 00:11:22,000 Speaker 1: train it according to their own hidden set of rules, 217 00:11:22,040 --> 00:11:26,480 Speaker 1: and then a customer might use that for some price. 218 00:11:26,840 --> 00:11:28,760 Speaker 1: What you're saying is, in the same way that we 219 00:11:28,840 --> 00:11:32,680 Speaker 1: democratize the writing of software before, let's democratize the training 220 00:11:32,720 --> 00:11:36,400 Speaker 1: of an AI system. So anyone can contribute here and 221 00:11:36,800 --> 00:11:40,480 Speaker 1: teach the model the things that they're interested in teaching 222 00:11:40,480 --> 00:11:43,400 Speaker 1: the model. I'm guessing correct me. On the one hand, 223 00:11:43,920 --> 00:11:46,000 Speaker 1: this model, at least in the beginning, is going to 224 00:11:46,040 --> 00:11:49,160 Speaker 1: have a lot fewer resources available to it. But on 225 00:11:49,200 --> 00:11:50,959 Speaker 1: the other hand, it's going to have a much more 226 00:11:51,320 --> 00:11:53,280 Speaker 1: diverse set of inputs. 227 00:11:53,760 --> 00:11:56,840 Speaker 2: That's right. And the other thing is that IBM, basically 228 00:11:56,840 --> 00:11:59,840 Speaker 2: as part of this project, has something called the granite model, 229 00:12:00,720 --> 00:12:03,040 Speaker 2: and they've donated some granite models. So these are the 230 00:12:03,040 --> 00:12:06,400 Speaker 2: ones that take the months and terabytes of data and 231 00:12:06,480 --> 00:12:09,360 Speaker 2: all the GPUs to train. So IBM has created one 232 00:12:09,400 --> 00:12:12,520 Speaker 2: of those, and they have listed out and linked to 233 00:12:12,559 --> 00:12:14,760 Speaker 2: the data sets that they used, and they talk about 234 00:12:14,760 --> 00:12:17,960 Speaker 2: the relative proportions they used when pre training, so it's 235 00:12:17,960 --> 00:12:20,079 Speaker 2: not just the black box. You know where the data 236 00:12:20,120 --> 00:12:22,959 Speaker 2: came from, which is a pretty open position to take. 237 00:12:23,360 --> 00:12:25,400 Speaker 2: That is what we recommend as the base. So you 238 00:12:25,480 --> 00:12:28,559 Speaker 2: use the instruct lab tuning. You take this base granite 239 00:12:28,600 --> 00:12:31,280 Speaker 2: model that IBM has provided, and you use the instruct 240 00:12:31,320 --> 00:12:33,640 Speaker 2: lab tooling that red Hat works on, and you use 241 00:12:33,679 --> 00:12:36,760 Speaker 2: that to fine tune the model to make it whatever 242 00:12:36,800 --> 00:12:37,280 Speaker 2: you want. 243 00:12:37,840 --> 00:12:40,760 Speaker 1: I want to go back to the partnership between IBM 244 00:12:40,800 --> 00:12:44,800 Speaker 1: and red Hat here with them providing the granite model 245 00:12:45,400 --> 00:12:48,680 Speaker 1: to your instruct lab. Is this the first time red 246 00:12:48,679 --> 00:12:50,680 Speaker 1: Hat and IBM have collaborated like this. 247 00:12:51,480 --> 00:12:54,040 Speaker 2: I think it's something that's been going on. Like another 248 00:12:54,320 --> 00:12:57,560 Speaker 2: a product within the red Hat family would be OpenShift AI, 249 00:12:57,679 --> 00:13:00,679 Speaker 2: where they collaborate a lot with IBM Research team, Like 250 00:13:00,920 --> 00:13:03,280 Speaker 2: bolm is one of the components of that product that 251 00:13:03,320 --> 00:13:07,280 Speaker 2: there's a nice kind of exchange and collaboration between the 252 00:13:07,320 --> 00:13:08,480 Speaker 2: two companies. 253 00:13:09,240 --> 00:13:11,920 Speaker 1: How large is the potential community of people who might 254 00:13:11,960 --> 00:13:13,560 Speaker 1: contribute to instruct lab. 255 00:13:14,480 --> 00:13:17,040 Speaker 2: It could be thousands of people. I mean, we'll see. 256 00:13:17,080 --> 00:13:20,840 Speaker 2: It's early days. This is early technology that was invented 257 00:13:20,840 --> 00:13:23,320 Speaker 2: at IBM Research that they partnered with us at red 258 00:13:23,320 --> 00:13:26,200 Speaker 2: Hat to kind of build the software around it. There's 259 00:13:26,200 --> 00:13:28,480 Speaker 2: still more to go, Like right now, we have a 260 00:13:28,559 --> 00:13:30,559 Speaker 2: team in the community that's actually trying to build a 261 00:13:30,600 --> 00:13:34,120 Speaker 2: web interface to make it easier for anybody to contribute. 262 00:13:34,280 --> 00:13:35,800 Speaker 2: So we have a lot of those sort of user 263 00:13:35,880 --> 00:13:39,320 Speaker 2: experience for the contributor to the model stuff to work 264 00:13:39,320 --> 00:13:42,040 Speaker 2: out that we're still actively building on. But like my 265 00:13:42,200 --> 00:13:44,840 Speaker 2: vision for it even is I like going back to 266 00:13:44,840 --> 00:13:47,960 Speaker 2: that academic model of learning from what others and building 267 00:13:48,000 --> 00:13:50,880 Speaker 2: upon it over time. It would be very good for 268 00:13:51,040 --> 00:13:53,880 Speaker 2: us to sort of go out and try to collaborate 269 00:13:54,040 --> 00:13:56,520 Speaker 2: with academics of all fields, like, hey, you know, the 270 00:13:56,559 --> 00:14:00,079 Speaker 2: model doesn't know about your field. Would you like to 271 00:14:00,120 --> 00:14:02,480 Speaker 2: put something into the model about your field so it 272 00:14:02,520 --> 00:14:05,880 Speaker 2: knows about it? Or even you know, talk to the model. 273 00:14:06,080 --> 00:14:08,440 Speaker 2: It got it wrong, let's correct it. Can we lean 274 00:14:08,520 --> 00:14:10,640 Speaker 2: on your expertise to correct it and make sure it 275 00:14:10,640 --> 00:14:13,760 Speaker 2: gets it right and sort of use that community model 276 00:14:13,800 --> 00:14:17,839 Speaker 2: as a way for everybody to collaborate because before instruct lab, 277 00:14:18,600 --> 00:14:22,040 Speaker 2: my understanding is if you wanted to take a model 278 00:14:22,040 --> 00:14:24,120 Speaker 2: that's open source license and play with it, you could 279 00:14:24,160 --> 00:14:25,840 Speaker 2: do that. You could take a model kind of off 280 00:14:25,840 --> 00:14:28,880 Speaker 2: the shelf from Hugging Face and fine tune it yourself. 281 00:14:29,280 --> 00:14:30,840 Speaker 2: But it's a bit of a dead end because you 282 00:14:30,920 --> 00:14:33,440 Speaker 2: made your contributions, but there's no way for other people 283 00:14:33,920 --> 00:14:36,840 Speaker 2: to collaborate with you. So the way that we've built 284 00:14:36,880 --> 00:14:41,000 Speaker 2: this is based on how the technology works. Everybody can 285 00:14:41,000 --> 00:14:43,360 Speaker 2: contribute to it. This is something that it can keep 286 00:14:43,400 --> 00:14:44,840 Speaker 2: growing and growing and growing over time. 287 00:14:45,160 --> 00:14:48,720 Speaker 1: Yeah. Yeah, what's the level of expertise necessary to be 288 00:14:48,760 --> 00:14:49,480 Speaker 1: a contributor? 289 00:14:50,080 --> 00:14:51,960 Speaker 2: You don't need to be a data scientist, and you 290 00:14:52,000 --> 00:14:54,960 Speaker 2: don't need to have exotic hardware. Honestly, if you don't 291 00:14:54,960 --> 00:14:57,400 Speaker 2: even have laptop hardware that meets the spec for doing 292 00:14:57,400 --> 00:15:01,080 Speaker 2: instruct labs laptop version, you can minutes to the community 293 00:15:01,160 --> 00:15:03,640 Speaker 2: and then we'll actually build it for you. We have 294 00:15:03,760 --> 00:15:06,120 Speaker 2: bots and stuff that do that, and we're hoping over 295 00:15:06,160 --> 00:15:08,680 Speaker 2: time to make that more accessible, first by having a 296 00:15:08,760 --> 00:15:11,360 Speaker 2: user interface and then maybe later on having a web service. 297 00:15:11,680 --> 00:15:14,880 Speaker 1: Yeah, so give me an example of how a business 298 00:15:15,000 --> 00:15:17,040 Speaker 1: might make use of instruct lab. 299 00:15:17,600 --> 00:15:20,160 Speaker 2: One of the things that businesses are doing with AI 300 00:15:20,240 --> 00:15:23,920 Speaker 2: right now is using hosted API services. You're quite expensive, 301 00:15:24,200 --> 00:15:27,040 Speaker 2: but they're finding value. But it's hard given the amount 302 00:15:27,040 --> 00:15:29,320 Speaker 2: of money they're spending. And one of the things that's 303 00:15:29,320 --> 00:15:31,200 Speaker 2: a little scary about it too, is like you have 304 00:15:31,520 --> 00:15:35,960 Speaker 2: very sensitive internal documents and you have employees maybe not 305 00:15:36,080 --> 00:15:38,960 Speaker 2: understanding what they're actually doing because you know, how would 306 00:15:39,000 --> 00:15:42,400 Speaker 2: you if you're not technical enough when you're asking said 307 00:15:43,080 --> 00:15:48,760 Speaker 2: public web service AI model information about your copy pasting 308 00:15:48,880 --> 00:15:52,920 Speaker 2: internal company documents. It's going across the Internet into another 309 00:15:52,960 --> 00:15:56,160 Speaker 2: company's hands, and that company probably shouldn't have access to that. 310 00:15:56,600 --> 00:15:59,560 Speaker 2: So what both RedHat and IBM and the space are 311 00:15:59,600 --> 00:16:02,960 Speaker 2: looking at, like, the instruct lab model is very modest. 312 00:16:03,000 --> 00:16:06,480 Speaker 2: It's seven billion parameter model, very small. It's very cheap 313 00:16:06,520 --> 00:16:09,920 Speaker 2: to serve inference on a seven billion parameter model. It's 314 00:16:09,920 --> 00:16:13,480 Speaker 2: competing with trillion parameter models that are hosted. You take 315 00:16:13,480 --> 00:16:16,320 Speaker 2: this small model that is cheap to run inference on, 316 00:16:16,960 --> 00:16:20,880 Speaker 2: you train it with your own company's proprietary data inside 317 00:16:20,920 --> 00:16:23,480 Speaker 2: the walls of your company, on your own hardware. You 318 00:16:23,520 --> 00:16:26,800 Speaker 2: can do all sorts of actual data analysis on your 319 00:16:26,840 --> 00:16:29,440 Speaker 2: most sensitive data and have the confidence that has not 320 00:16:29,520 --> 00:16:30,400 Speaker 2: left the premises. 321 00:16:31,240 --> 00:16:34,280 Speaker 1: In that use case, you're not actually training the model 322 00:16:34,320 --> 00:16:37,560 Speaker 1: for everyone. You're just taking it and doing some private 323 00:16:37,600 --> 00:16:40,080 Speaker 1: stuff on it exactly, which doesn't leave the building. But 324 00:16:40,120 --> 00:16:44,640 Speaker 1: that's separate from an interaction where you're doing something that 325 00:16:45,360 --> 00:16:46,880 Speaker 1: contributes overall. 326 00:16:46,960 --> 00:16:49,520 Speaker 2: Right, And that's something maybe that I should be more 327 00:16:49,560 --> 00:16:51,800 Speaker 2: clear about. Is there's sort of two tracks here, and 328 00:16:51,920 --> 00:16:55,160 Speaker 2: this is very red hat classic. You have your upstream 329 00:16:55,280 --> 00:16:58,320 Speaker 2: community track and you have your business product track. So 330 00:16:58,360 --> 00:17:02,360 Speaker 2: the upstream community track is just enabling anybody to contribute 331 00:17:02,400 --> 00:17:04,320 Speaker 2: to a model in a collaborative way and play with it. 332 00:17:04,760 --> 00:17:08,600 Speaker 2: The downstream product business oriented track is now take that 333 00:17:08,720 --> 00:17:13,359 Speaker 2: tech that we've honed and developed in the open community, 334 00:17:14,000 --> 00:17:16,480 Speaker 2: and apply it to your business knowledge and skills. 335 00:17:17,520 --> 00:17:21,360 Speaker 1: This community driven approach marks a pivotal shift towards more 336 00:17:21,400 --> 00:17:27,159 Speaker 1: accessible AI solutions. The contrast between externally hosted AI services 337 00:17:27,440 --> 00:17:30,639 Speaker 1: and the open model enhanced by instruct lab underscores the 338 00:17:30,680 --> 00:17:35,760 Speaker 1: potential for broader adoption of AI in diverse business contexts. 339 00:17:36,280 --> 00:17:39,680 Speaker 1: She envisions a future in which technological innovation is more 340 00:17:39,680 --> 00:17:43,960 Speaker 1: tailored to individual business needs, guided by principles of openness 341 00:17:44,119 --> 00:17:49,000 Speaker 1: and security. To an imaginary case study, Sure, I'm a 342 00:17:49,080 --> 00:17:53,879 Speaker 1: law firm. I'm an entertainment law I have one hundred 343 00:17:53,880 --> 00:17:58,360 Speaker 1: clients who are big stars. They all have incredibly complicated contracts. 344 00:17:59,160 --> 00:18:03,600 Speaker 1: I feed a thousand of my company's contracts from the 345 00:18:03,680 --> 00:18:07,520 Speaker 1: last ten years into the model, and then every time 346 00:18:07,560 --> 00:18:10,160 Speaker 1: I have a new contract, I ask the model, am 347 00:18:10,160 --> 00:18:12,840 Speaker 1: I missing something? Can you go back and look through 348 00:18:12,880 --> 00:18:15,240 Speaker 1: all our own contracts and show me a contract that 349 00:18:15,520 --> 00:18:18,800 Speaker 1: is missing key components or exposes us to some liability. 350 00:18:19,680 --> 00:18:23,280 Speaker 1: In that case, the model would know my law firm 351 00:18:23,520 --> 00:18:26,880 Speaker 1: contracts really really well. It's as if they've been working 352 00:18:27,400 --> 00:18:30,800 Speaker 1: out my law firm. They're not distracted by other people's 353 00:18:30,840 --> 00:18:36,600 Speaker 1: particular styles, or a bunch of contracts from the utility industry, 354 00:18:36,720 --> 00:18:41,360 Speaker 1: or they know entertainment law contracts exactly. 355 00:18:41,480 --> 00:18:43,520 Speaker 2: Yeah, you can train it in your own image, your 356 00:18:43,600 --> 00:18:47,400 Speaker 2: style of doing things. It's something that your company can 357 00:18:47,440 --> 00:18:50,879 Speaker 2: produce that is uniquely helpful to you. No third party 358 00:18:50,880 --> 00:18:53,159 Speaker 2: could do that because no third party understands how you 359 00:18:53,200 --> 00:18:56,919 Speaker 2: do business and understands your history and your documents. So 360 00:18:56,960 --> 00:18:59,560 Speaker 2: it's sort of a way of getting value out of 361 00:18:59,600 --> 00:19:01,320 Speaker 2: the stuff if you already have sitting in a file 362 00:19:01,359 --> 00:19:03,520 Speaker 2: cabinet somewhere, it's very cool. 363 00:19:03,800 --> 00:19:06,800 Speaker 1: Yeah, give me a sort of a real world case 364 00:19:06,800 --> 00:19:10,000 Speaker 1: study where you think the business use case would be 365 00:19:10,080 --> 00:19:14,680 Speaker 1: really powerful. What's a business that really could see an 366 00:19:14,680 --> 00:19:18,680 Speaker 1: advantage to using instruct lab in its way. 367 00:19:19,160 --> 00:19:21,399 Speaker 2: The demo that I've given a couple of times at 368 00:19:21,440 --> 00:19:25,040 Speaker 2: different events used an imaginary insurance company. So you say, 369 00:19:25,040 --> 00:19:28,960 Speaker 2: you have this company, you have to recommend repairs for 370 00:19:29,080 --> 00:19:32,480 Speaker 2: various types of claims. You've been doing this for years, 371 00:19:32,560 --> 00:19:35,400 Speaker 2: you know. If you know the windshield's broken and you've 372 00:19:35,400 --> 00:19:38,240 Speaker 2: gotten this type of accident and it's this model car, 373 00:19:38,480 --> 00:19:40,200 Speaker 2: these are the kinds of things you want to look at. 374 00:19:40,920 --> 00:19:43,679 Speaker 2: So you could talk to any insurance agent in the 375 00:19:43,720 --> 00:19:46,400 Speaker 2: field and be like, oh, you know, it's a Tesla. 376 00:19:46,520 --> 00:19:48,639 Speaker 2: You might want to look at the battery or something like. 377 00:19:48,720 --> 00:19:52,120 Speaker 2: They'll have some latent knowledge just so you can take 378 00:19:52,160 --> 00:19:54,880 Speaker 2: that and train it into a model. Honestly, I think 379 00:19:54,920 --> 00:19:58,160 Speaker 2: these kind of new technologies are better when they're less visible. 380 00:19:58,760 --> 00:20:01,199 Speaker 2: So say you have the as agents in the field 381 00:20:01,280 --> 00:20:03,280 Speaker 2: and they have this tool and they're kind of entering 382 00:20:03,280 --> 00:20:06,240 Speaker 2: the claim data. They're on the scene at the car, 383 00:20:06,840 --> 00:20:09,480 Speaker 2: and it might say, oh, look, I see this is 384 00:20:09,520 --> 00:20:12,080 Speaker 2: a Ford Fiesta. These are things you want to look 385 00:20:12,080 --> 00:20:15,280 Speaker 2: at for this type of accident. As you're entering the data, 386 00:20:15,720 --> 00:20:17,600 Speaker 2: it could be going through the knowledge you had loaded 387 00:20:17,600 --> 00:20:20,080 Speaker 2: into the model and be making these suggestions based on 388 00:20:20,119 --> 00:20:23,080 Speaker 2: your company's background, and hey, you know, let's not make 389 00:20:23,080 --> 00:20:25,640 Speaker 2: the same mistake twice. Let's make new mistakes and let's 390 00:20:25,680 --> 00:20:28,560 Speaker 2: learn from the stuff we already did. So that's one example, 391 00:20:28,600 --> 00:20:31,159 Speaker 2: but there's so many different industries in ways that this 392 00:20:31,240 --> 00:20:34,000 Speaker 2: could help, and it could make those agents in the 393 00:20:34,040 --> 00:20:35,520 Speaker 2: field more efficient. 394 00:20:36,240 --> 00:20:38,680 Speaker 1: Have you had anyone talk to you about using instruct 395 00:20:38,720 --> 00:20:40,240 Speaker 1: lab in a way that surprised you. 396 00:20:42,280 --> 00:20:46,960 Speaker 2: I mean, some people have done funky things, but sort 397 00:20:46,960 --> 00:20:49,360 Speaker 2: of playing with the skills stuff. That's where I see 398 00:20:49,359 --> 00:20:52,360 Speaker 2: a lot of creativity. The difference between knowledge and skills 399 00:20:52,400 --> 00:20:55,720 Speaker 2: is that knowledge is pretty pretty understandable, right, like oh, 400 00:20:55,800 --> 00:20:59,679 Speaker 2: historical insurance claims or you know, legal contracts. Skills are 401 00:20:59,680 --> 00:21:02,600 Speaker 2: a little different. So whenever somebody submits a skill, sometimes 402 00:21:02,720 --> 00:21:04,680 Speaker 2: it tends to be really creative because it's not something 403 00:21:04,720 --> 00:21:08,040 Speaker 2: that's super intuitive. Somebody submitted a skill. I don't know 404 00:21:08,040 --> 00:21:11,040 Speaker 2: how well it worked, but it was like making ASKI art, 405 00:21:11,240 --> 00:21:13,640 Speaker 2: like draw me a I don't know, draw me a dog, 406 00:21:13,680 --> 00:21:15,359 Speaker 2: and would do like an ASKI art dog. I mean, 407 00:21:15,400 --> 00:21:17,960 Speaker 2: it's stuff that you can do programmatically. One that was 408 00:21:17,960 --> 00:21:21,879 Speaker 2: actually very very helpful was you know, take this table 409 00:21:21,920 --> 00:21:25,120 Speaker 2: of data and convert it to this format like, ooh, 410 00:21:25,160 --> 00:21:26,800 Speaker 2: that's nice. That actually saves me time. 411 00:21:27,359 --> 00:21:29,880 Speaker 1: How far away are we from the day when I, 412 00:21:29,960 --> 00:21:34,679 Speaker 1: Malcolm Globwell technology ignore Amus can go home and easily 413 00:21:34,720 --> 00:21:36,480 Speaker 1: interact with instruct lab. 414 00:21:37,840 --> 00:21:39,600 Speaker 2: Maybe a few months, a few. 415 00:21:39,400 --> 00:21:42,080 Speaker 1: Months, you're gonna say a few years. 416 00:21:42,720 --> 00:21:45,879 Speaker 2: No, I think it'd be a few months. Wow, I 417 00:21:45,880 --> 00:21:48,600 Speaker 2: hope it's power open source innovation. 418 00:21:49,040 --> 00:21:53,000 Speaker 1: Yeah, oh that's really interesting. Yeah, I'm always taken by surprise. 419 00:21:53,320 --> 00:21:55,960 Speaker 1: I'm still thinking in twentieth century terms about how long 420 00:21:56,040 --> 00:21:59,200 Speaker 1: things take, and you're in the twenty second century as 421 00:21:59,200 --> 00:21:59,880 Speaker 1: well as I could tell. 422 00:22:00,000 --> 00:22:04,560 Speaker 2: The instruct lab core invention was invented in a hotel 423 00:22:04,680 --> 00:22:07,720 Speaker 2: room at an AI conference in December with an amazing 424 00:22:07,760 --> 00:22:10,880 Speaker 2: group of IBM research guys December of twenty twenty three. 425 00:22:11,200 --> 00:22:13,879 Speaker 1: Wait back up, you have to tell the story. 426 00:22:14,119 --> 00:22:17,080 Speaker 2: This group of guys we've been working with, they were 427 00:22:17,119 --> 00:22:19,479 Speaker 2: at this conference together, and it's a really funny story 428 00:22:19,560 --> 00:22:22,399 Speaker 2: because you know, it's hard to get access to GPUs 429 00:22:22,720 --> 00:22:24,520 Speaker 2: and like even you know, you're at IBM and it's 430 00:22:24,520 --> 00:22:27,280 Speaker 2: hard to get access because everybody wants access. They did 431 00:22:27,320 --> 00:22:30,240 Speaker 2: it over Christmas break because nobody was using the cluster 432 00:22:30,320 --> 00:22:32,399 Speaker 2: at the time, and they ran all of these experiments 433 00:22:32,480 --> 00:22:34,280 Speaker 2: and I'm like, whoa, this is really cool. 434 00:22:34,680 --> 00:22:38,640 Speaker 1: And wait. Their idea was we can do a stripped 435 00:22:38,640 --> 00:22:44,000 Speaker 1: down AI model, and was the idea and even back 436 00:22:44,040 --> 00:22:45,280 Speaker 1: then combine it with grantite. 437 00:22:45,320 --> 00:22:48,080 Speaker 2: What was the original the original idea. It's sort of 438 00:22:48,200 --> 00:22:51,800 Speaker 2: multi there's like multiple aspects to it. So like one 439 00:22:51,800 --> 00:22:54,280 Speaker 2: of the aspects it actually came on later, but it 440 00:22:54,320 --> 00:22:56,600 Speaker 2: starts at the beginning of the workflow. Is you're using 441 00:22:56,680 --> 00:23:00,639 Speaker 2: a taxonomy to organize how you're fine too the model. 442 00:23:00,680 --> 00:23:03,240 Speaker 2: So the old approach they call it the blender approach, 443 00:23:03,640 --> 00:23:06,000 Speaker 2: to just take a bunch of data of roughly the 444 00:23:06,040 --> 00:23:07,800 Speaker 2: type of data that you'd like and you kind of 445 00:23:07,840 --> 00:23:10,399 Speaker 2: throw it in and then see what comes out. Don't 446 00:23:10,480 --> 00:23:12,879 Speaker 2: like it, Okay, throw in more, try again, see what 447 00:23:12,960 --> 00:23:16,399 Speaker 2: comes out. They had used this taxonomy technique, so you 448 00:23:16,440 --> 00:23:20,560 Speaker 2: actually build like a taxonomy of like categories and subfolders 449 00:23:20,600 --> 00:23:22,959 Speaker 2: of like this is the knowledge and skills that we 450 00:23:23,000 --> 00:23:25,680 Speaker 2: want to train into the model. And that way you're 451 00:23:25,720 --> 00:23:29,000 Speaker 2: sort of systematic about what you're adding, and you can 452 00:23:29,040 --> 00:23:31,560 Speaker 2: also identify gaps pretty easily. Oh, I don't have a 453 00:23:31,600 --> 00:23:33,720 Speaker 2: category for that, let me add that. So that's like 454 00:23:33,920 --> 00:23:36,200 Speaker 2: one of the parts of the invention here. 455 00:23:37,000 --> 00:23:41,919 Speaker 1: Point number one is let's be intentional and deliberate in 456 00:23:42,000 --> 00:23:43,240 Speaker 1: how we build and train this thing. 457 00:23:43,440 --> 00:23:46,719 Speaker 2: Yeah, and then the next component would be Okay, So 458 00:23:47,000 --> 00:23:49,560 Speaker 2: it is actually quite expensive. Part of the expense of 459 00:23:49,600 --> 00:23:53,119 Speaker 2: like tuning models and just training models in general is 460 00:23:53,160 --> 00:23:56,359 Speaker 2: coming up with the data. And what they wanted to 461 00:23:56,400 --> 00:23:58,520 Speaker 2: do is have a technique where you could have just 462 00:23:58,560 --> 00:24:01,679 Speaker 2: a little bit of data and expand it with something 463 00:24:01,680 --> 00:24:05,080 Speaker 2: they're calling synthetic data generation. And this is where it's 464 00:24:05,080 --> 00:24:09,040 Speaker 2: sort of like you have this student and teacher workflow, 465 00:24:09,640 --> 00:24:14,359 Speaker 2: so you have your taxonomy. The taxonomy has sort of 466 00:24:14,359 --> 00:24:17,320 Speaker 2: the knowledge like a business's knowledge documents, their insurance claims, 467 00:24:17,600 --> 00:24:20,840 Speaker 2: and it has these quizzes that you write and that's 468 00:24:20,880 --> 00:24:22,800 Speaker 2: to teach the model. So I'm writing a quiz based 469 00:24:22,920 --> 00:24:24,520 Speaker 2: just like you do in school. You read the chapter 470 00:24:24,720 --> 00:24:26,800 Speaker 2: on the American Revolution, and then you answer a ten 471 00:24:26,880 --> 00:24:30,160 Speaker 2: question quiz where you're giving the model quiz. You need 472 00:24:30,200 --> 00:24:33,360 Speaker 2: at least five questions and answers, and the answers need 473 00:24:33,400 --> 00:24:36,000 Speaker 2: to be taken from the context of the document, and 474 00:24:36,119 --> 00:24:39,600 Speaker 2: then you run it through a process called synthetic data generation, 475 00:24:39,880 --> 00:24:41,919 Speaker 2: and it looks at the documents, so we'll look at 476 00:24:41,920 --> 00:24:44,680 Speaker 2: the history chapter, it'll look at the questions and answers, 477 00:24:45,040 --> 00:24:47,680 Speaker 2: and then it'll look to that original document and come 478 00:24:47,760 --> 00:24:50,280 Speaker 2: up with more questions and answers based on the format 479 00:24:50,320 --> 00:24:52,680 Speaker 2: of the questions and answers you made. So you can 480 00:24:52,760 --> 00:24:56,080 Speaker 2: take five questions of answers amplify them into one hundred 481 00:24:56,359 --> 00:24:59,120 Speaker 2: questions and answers two hundred questions and answers. And it's 482 00:24:59,440 --> 00:25:02,560 Speaker 2: a second model that is making the questions and answers, 483 00:25:02,680 --> 00:25:05,639 Speaker 2: so it's synthetic data generation using an AI model to 484 00:25:05,680 --> 00:25:08,480 Speaker 2: make the questions. We use an open source model to 485 00:25:08,520 --> 00:25:11,480 Speaker 2: do that. So that's the second part. And then the 486 00:25:11,520 --> 00:25:14,520 Speaker 2: third part is we have a multi phase tuning technique 487 00:25:14,520 --> 00:25:18,360 Speaker 2: to actually take the synthetic data and then basically bake 488 00:25:18,440 --> 00:25:20,760 Speaker 2: it into the model. So sort of that's the approach. 489 00:25:21,440 --> 00:25:24,520 Speaker 2: A general philosophy of the approach is using granite because 490 00:25:24,560 --> 00:25:27,280 Speaker 2: we know where the data came from. Another approach is 491 00:25:27,280 --> 00:25:29,640 Speaker 2: the fact that we're using small models that are cheap 492 00:25:29,640 --> 00:25:32,160 Speaker 2: to run inference on. They're small enough that you can 493 00:25:32,160 --> 00:25:34,520 Speaker 2: tune them on laptop hardware. You don't need all the 494 00:25:34,520 --> 00:25:38,679 Speaker 2: fancy expensive GPU mania you're good. So sort of like 495 00:25:38,720 --> 00:25:41,879 Speaker 2: a whole system, it's like not any one component. But 496 00:25:41,960 --> 00:25:44,480 Speaker 2: it's sort of the approach they took with somewhat novel, 497 00:25:44,760 --> 00:25:47,440 Speaker 2: and they were very excited when they saw the experimental results. 498 00:25:47,840 --> 00:25:50,600 Speaker 2: There was a meeting between red hat and IBM. It 499 00:25:50,640 --> 00:25:52,760 Speaker 2: was actually an IBM research meeting that red hatters were 500 00:25:52,800 --> 00:25:56,040 Speaker 2: invited to, and I think the red Hatter's involves sort 501 00:25:56,040 --> 00:25:59,640 Speaker 2: of saw the potential, WHOA, we can make models open 502 00:25:59,680 --> 00:26:03,680 Speaker 2: source finally, rather than them just being these endless dead forks, 503 00:26:04,600 --> 00:26:06,760 Speaker 2: we could make it so people could contribute back and 504 00:26:06,800 --> 00:26:09,560 Speaker 2: collaborate around it. So that's when red Hat became interested 505 00:26:09,560 --> 00:26:12,719 Speaker 2: in it and we sort of worked together, and the 506 00:26:12,800 --> 00:26:15,840 Speaker 2: research engineers from IBM Research who came up with the technique, 507 00:26:15,960 --> 00:26:18,520 Speaker 2: and then my team, the software engineers who know how 508 00:26:18,600 --> 00:26:23,560 Speaker 2: to take research code and productize it into actually runnable, 509 00:26:23,640 --> 00:26:28,520 Speaker 2: supportable software, kind of got together. We've been hanging out 510 00:26:28,520 --> 00:26:31,200 Speaker 2: in the Boston office at red Hat and building it out. 511 00:26:31,560 --> 00:26:34,800 Speaker 2: April eighteenth was when we went open source and we 512 00:26:34,840 --> 00:26:37,320 Speaker 2: made all of our repositories with all of the code public, 513 00:26:37,359 --> 00:26:39,600 Speaker 2: and right now we're working towards a product release, so 514 00:26:39,640 --> 00:26:40,600 Speaker 2: a supported product. 515 00:26:40,720 --> 00:26:42,800 Speaker 1: How long did it take you to be convinced of 516 00:26:43,760 --> 00:26:47,040 Speaker 1: the value of this idea? I mean, so people get 517 00:26:47,040 --> 00:26:51,240 Speaker 1: together in this hotel room. They're running these experiments over Christmas. 518 00:26:51,480 --> 00:26:53,439 Speaker 1: Are you aware of the experiments as they're running them? 519 00:26:54,800 --> 00:26:57,240 Speaker 2: I didn't find out till February. 520 00:26:57,280 --> 00:26:59,399 Speaker 1: So they come to you in February and they say, MO, 521 00:27:00,720 --> 00:27:02,800 Speaker 1: can you recreate that conversation? 522 00:27:03,840 --> 00:27:08,280 Speaker 2: Well, our CEO, Matt Hicks, and then Jeremy Eater, who's 523 00:27:08,320 --> 00:27:10,960 Speaker 2: one of our distinguished engineers, and Steve Watt, who's a VP, 524 00:27:11,160 --> 00:27:13,720 Speaker 2: were present, I think at that meeting. So they kind 525 00:27:13,720 --> 00:27:16,000 Speaker 2: of brought it back to us and said, listen, we've 526 00:27:16,000 --> 00:27:20,440 Speaker 2: invited these IBM research folks to come visit in Boston, 527 00:27:21,160 --> 00:27:23,600 Speaker 2: you know, work with them, like, see, does this have 528 00:27:23,640 --> 00:27:25,880 Speaker 2: any merit? Could we build something from it? And so 529 00:27:25,920 --> 00:27:29,040 Speaker 2: they gave us some presentations. We were very excited when 530 00:27:29,040 --> 00:27:32,520 Speaker 2: they came to us. It only had support for Mac laptops. 531 00:27:33,119 --> 00:27:35,240 Speaker 2: Of course, you know, Red Hat were Linux people, so 532 00:27:35,320 --> 00:27:37,120 Speaker 2: we're like, all right, we've got to fix that. So 533 00:27:37,280 --> 00:27:40,000 Speaker 2: a bunch of the junior engineers around the office kind 534 00:27:40,000 --> 00:27:41,600 Speaker 2: of came and they're like, okay, we're going to build 535 00:27:41,640 --> 00:27:43,720 Speaker 2: Linux support. And they had it within like a couple 536 00:27:43,760 --> 00:27:46,600 Speaker 2: of days. It was crazy because this was just meant 537 00:27:46,640 --> 00:27:49,200 Speaker 2: to be, Hey, guys, you know what, these are invited 538 00:27:49,240 --> 00:27:52,760 Speaker 2: guests visiting our office. See what happens. And we ended 539 00:27:52,840 --> 00:27:56,240 Speaker 2: up doing like weeks of hack fests and late night 540 00:27:56,280 --> 00:27:58,919 Speaker 2: pizzas in the conference room and like playing around with 541 00:27:58,960 --> 00:28:01,880 Speaker 2: it and learning, and it was it was very fun. 542 00:28:01,960 --> 00:28:02,680 Speaker 2: It's very cool. 543 00:28:02,800 --> 00:28:05,800 Speaker 1: Anyone else do anything like this is. 544 00:28:05,800 --> 00:28:08,359 Speaker 2: Not my understanding that anybody else is doing it. Yet 545 00:28:08,960 --> 00:28:11,920 Speaker 2: maybe others will try. A lot of the focus has 546 00:28:11,960 --> 00:28:15,600 Speaker 2: been on that pre training phase. But for us, again 547 00:28:15,680 --> 00:28:19,280 Speaker 2: that fine tuning, it's more accessible because you don't need 548 00:28:19,320 --> 00:28:21,920 Speaker 2: all the exactic hardware. It doesn't take months. You can 549 00:28:21,960 --> 00:28:23,840 Speaker 2: do it on a laptop. You can do a smoke 550 00:28:23,920 --> 00:28:26,240 Speaker 2: test version of it in less than an hour. 551 00:28:26,760 --> 00:28:27,920 Speaker 1: What is the word smoke test. 552 00:28:28,080 --> 00:28:30,480 Speaker 2: Smoke test means you're not doing a full fine tuning 553 00:28:30,520 --> 00:28:33,600 Speaker 2: on the model. It's a different tuning process. It's like 554 00:28:33,640 --> 00:28:35,680 Speaker 2: kind of lower quality so to run on lower grade 555 00:28:35,680 --> 00:28:37,880 Speaker 2: hardware so you can kind of see them didn't move 556 00:28:37,920 --> 00:28:39,440 Speaker 2: the model or not, but it's not going to give 557 00:28:39,440 --> 00:28:42,040 Speaker 2: you like the full picture. You need higher end hardware 558 00:28:42,080 --> 00:28:43,880 Speaker 2: to actually do the full thing. So that's what the 559 00:28:43,880 --> 00:28:46,160 Speaker 2: product will enable you to do once it's launched. Is 560 00:28:46,880 --> 00:28:48,840 Speaker 2: you're going to need the GPUs, but when you have them, 561 00:28:48,880 --> 00:28:50,480 Speaker 2: will help you make the best usage of them. 562 00:28:50,760 --> 00:28:53,560 Speaker 1: Yeah, yeah, and no, there's all the detail. I want 563 00:28:53,600 --> 00:28:55,960 Speaker 1: to go back to sure in order to run the 564 00:28:56,040 --> 00:29:01,920 Speaker 1: tests on this idea way back when they needed time 565 00:29:02,000 --> 00:29:05,240 Speaker 1: on the GPUs, So this will be the in house 566 00:29:05,640 --> 00:29:09,800 Speaker 1: IBM and they were quiet at Christmas, So how much 567 00:29:09,920 --> 00:29:12,920 Speaker 1: time would you need on the GPUs to kind of 568 00:29:13,040 --> 00:29:14,040 Speaker 1: get proof of concept. 569 00:29:14,440 --> 00:29:16,760 Speaker 2: Well, what happens is and it's sort of like a 570 00:29:16,800 --> 00:29:19,120 Speaker 2: lot of trial and error, right, And there's a lot 571 00:29:19,120 --> 00:29:22,720 Speaker 2: about this stuff that like you come up with the hypothesis, 572 00:29:22,800 --> 00:29:24,760 Speaker 2: you test it out, did it work or not? Okay, 573 00:29:24,920 --> 00:29:26,920 Speaker 2: it's just like you know in the lab, but you know, 574 00:29:27,120 --> 00:29:30,920 Speaker 2: buns and burners and beakers and whatever. So it really depends. 575 00:29:31,000 --> 00:29:34,360 Speaker 2: But it can be hours, it can be days. It 576 00:29:34,400 --> 00:29:36,520 Speaker 2: really depends on what they're trying to do. And then 577 00:29:36,520 --> 00:29:38,880 Speaker 2: sometimes they can cut the time down, you know, with 578 00:29:38,920 --> 00:29:40,560 Speaker 2: the number of GPUs you have. So like I have 579 00:29:40,600 --> 00:29:43,400 Speaker 2: a cluster of agpus, Okay, it might take a day, 580 00:29:43,480 --> 00:29:45,440 Speaker 2: but then if I can get thirty two, I can 581 00:29:45,480 --> 00:29:47,240 Speaker 2: pipeline it and make it go faster and get it 582 00:29:47,280 --> 00:29:49,280 Speaker 2: down to a few hours. So it really depends, you know. 583 00:29:49,360 --> 00:29:52,440 Speaker 2: But it's like everybody's home for the holidays. It's a 584 00:29:52,480 --> 00:29:55,040 Speaker 2: lovely playground to kind of get that stuff going fast. 585 00:29:55,800 --> 00:29:59,360 Speaker 1: Let's jump forward one year. Tell me the status of 586 00:29:59,400 --> 00:30:02,880 Speaker 1: this project, tell me who's using it, tell me how 587 00:30:02,920 --> 00:30:08,960 Speaker 1: big is it. Give me your optimistic but plausible prediction 588 00:30:09,240 --> 00:30:12,960 Speaker 1: about what instruct lab looks like a year from now. 589 00:30:13,880 --> 00:30:17,280 Speaker 2: A year from now, I would like to see kind 590 00:30:17,320 --> 00:30:23,720 Speaker 2: of a vibrant community around not just building knowledge and 591 00:30:23,760 --> 00:30:27,440 Speaker 2: skills into a model, but coming up with better techniques 592 00:30:27,480 --> 00:30:30,040 Speaker 2: and innovation around how we do it. So I'd like 593 00:30:30,080 --> 00:30:33,200 Speaker 2: to see the contributor experience as we grow more and 594 00:30:33,240 --> 00:30:35,960 Speaker 2: more contributors to be refined. So like a year from now, 595 00:30:36,160 --> 00:30:39,280 Speaker 2: Malcolm Gladwell could come impart some of his wisdom into 596 00:30:39,280 --> 00:30:41,640 Speaker 2: the model and it wouldn't be difficult, it wouldn't be 597 00:30:41,640 --> 00:30:44,560 Speaker 2: a big lift. I would love to see the user 598 00:30:44,600 --> 00:30:48,680 Speaker 2: interface tooling for doing that to be more sophisticated. I 599 00:30:48,720 --> 00:30:52,240 Speaker 2: would love to see more people taking this and even 600 00:30:52,360 --> 00:30:54,600 Speaker 2: using it. Maybe you're not sharing it with the community, 601 00:30:54,640 --> 00:30:57,560 Speaker 2: but you're using it for some private usage. Like I'll 602 00:30:57,560 --> 00:31:01,080 Speaker 2: give you an example. I'm in contact with a fellow 603 00:31:01,160 --> 00:31:03,880 Speaker 2: who is doing AI research and he's working with doctors. 604 00:31:03,920 --> 00:31:06,880 Speaker 2: They're GPS in an area of Canada where there's not 605 00:31:07,000 --> 00:31:09,719 Speaker 2: enough GPS for the number of patients. So you know, 606 00:31:09,800 --> 00:31:13,640 Speaker 2: anything you can do to save doctors time to get 607 00:31:13,680 --> 00:31:15,960 Speaker 2: to the next patient. It's like one of the things 608 00:31:15,960 --> 00:31:18,840 Speaker 2: that he has been doing experiments with is can we 609 00:31:18,960 --> 00:31:22,719 Speaker 2: use an open source, licensed model that the doctor can 610 00:31:22,800 --> 00:31:24,720 Speaker 2: run on their laptop so they don't have to worry 611 00:31:24,760 --> 00:31:27,280 Speaker 2: about all of the different privacy rules, Like it's privates 612 00:31:27,280 --> 00:31:31,360 Speaker 2: on the laptop right there, take his live transcription of 613 00:31:31,360 --> 00:31:35,040 Speaker 2: his conversation with the patient and then convert it automatically 614 00:31:35,080 --> 00:31:37,440 Speaker 2: to a soap format that can be entered in the database. 615 00:31:37,680 --> 00:31:40,240 Speaker 2: Typically this will take a doctor fifteen to twenty minutes 616 00:31:40,320 --> 00:31:44,080 Speaker 2: of paperwork. Why not save them the paperwork at least 617 00:31:44,080 --> 00:31:45,320 Speaker 2: have the model take a stab. 618 00:31:45,520 --> 00:31:48,120 Speaker 1: Does the model then hold on to that information and 619 00:31:47,520 --> 00:31:50,880 Speaker 1: he interacts with the model again when well, that's. 620 00:31:50,760 --> 00:31:53,040 Speaker 2: The thing not within struct lab. Maybe that could be 621 00:31:53,040 --> 00:31:56,640 Speaker 2: a future development. It doesn't once you're doing inference, it's 622 00:31:56,680 --> 00:31:59,160 Speaker 2: not ingesting that what you're saying to it back in. 623 00:31:59,480 --> 00:32:01,760 Speaker 2: It's only the fine tuning phase. So the idea would 624 00:32:01,760 --> 00:32:05,360 Speaker 2: be the doctor could maybe load in past patient data 625 00:32:05,640 --> 00:32:08,320 Speaker 2: as knowledge and then when he's trying to diagnose maybe 626 00:32:08,480 --> 00:32:10,960 Speaker 2: you know what I'm saying. Like, But the main idea 627 00:32:11,040 --> 00:32:13,440 Speaker 2: is somebody might have some private users. I would love 628 00:32:13,520 --> 00:32:17,720 Speaker 2: to see more usage of this tool to enable people 629 00:32:17,720 --> 00:32:20,080 Speaker 2: who otherwise never would have had access to this type 630 00:32:20,080 --> 00:32:22,840 Speaker 2: of technology who never like you know, a small country 631 00:32:22,960 --> 00:32:27,080 Speaker 2: GP doctors, it doesn't have GPUs. They're not going to 632 00:32:27,160 --> 00:32:29,320 Speaker 2: hire some company to custom build them a model. But 633 00:32:29,400 --> 00:32:31,160 Speaker 2: maybe on the weekend, if he's a techie guy, he 634 00:32:31,200 --> 00:32:32,200 Speaker 2: could say with that. 635 00:32:32,280 --> 00:32:34,560 Speaker 1: Sim well, I mean the more you talk, the more 636 00:32:34,600 --> 00:32:38,480 Speaker 1: I'm realizing that the simplicity of this model is the 637 00:32:38,640 --> 00:32:41,400 Speaker 1: killer app here. Once you know you can run it 638 00:32:41,400 --> 00:32:45,160 Speaker 1: on a laptop, you have democratized use in a way 639 00:32:45,200 --> 00:32:48,680 Speaker 1: that's inconceivable with some of these other much more complex. 640 00:32:49,560 --> 00:32:52,720 Speaker 1: But that's interesting because one would have thought intuitively that 641 00:32:53,200 --> 00:32:55,720 Speaker 1: at the beginning that the winner is going to be 642 00:32:55,760 --> 00:33:01,400 Speaker 1: the one with the biggest, most complex version, saying actually, no, 643 00:33:01,600 --> 00:33:07,000 Speaker 1: there's a whole series of uses where being lean and focused, 644 00:33:07,280 --> 00:33:11,120 Speaker 1: focused is actually you know, it enables a whole class 645 00:33:11,120 --> 00:33:14,480 Speaker 1: of uses. Maybe another way of saying this is who 646 00:33:14,520 --> 00:33:16,960 Speaker 1: wouldn't be a potential instruct lab customer. 647 00:33:17,360 --> 00:33:20,160 Speaker 2: We don't know yet. It's it's so new, like we 648 00:33:20,200 --> 00:33:22,680 Speaker 2: haven't really had enough people experimenting and playing with it 649 00:33:22,720 --> 00:33:25,400 Speaker 2: and finding out all the things yet. But that's that's 650 00:33:25,400 --> 00:33:27,120 Speaker 2: the thing that's so exciting about it. It's like, I 651 00:33:27,160 --> 00:33:28,640 Speaker 2: can't wait to see what people do. 652 00:33:29,080 --> 00:33:30,840 Speaker 1: Is this the most exciting thing you've worked on in 653 00:33:30,880 --> 00:33:31,320 Speaker 1: your career? 654 00:33:31,640 --> 00:33:32,080 Speaker 2: I think so? 655 00:33:33,320 --> 00:33:37,160 Speaker 1: I think so, Yeah, Well, we are reaching the end 656 00:33:37,160 --> 00:33:39,960 Speaker 1: of our time. But before we finished, we can do 657 00:33:40,000 --> 00:33:44,280 Speaker 1: a little speed round. Sure, all right, complete the following sentence. 658 00:33:44,920 --> 00:33:47,560 Speaker 1: In five years, AI will. 659 00:33:47,640 --> 00:33:52,120 Speaker 2: Be boring, it will be integrated, It'll just work, and 660 00:33:52,160 --> 00:33:54,400 Speaker 2: there will be no now with AI thing. It'll just 661 00:33:54,440 --> 00:33:55,040 Speaker 2: be normal. 662 00:33:56,680 --> 00:33:59,840 Speaker 1: What's the number one thing that people misunderstand about AI? 663 00:34:00,440 --> 00:34:03,960 Speaker 2: It's just matrix algebra. It's just numbers. It's not sentient. 664 00:34:04,240 --> 00:34:07,560 Speaker 2: It's not coming to take us over. It's just numbers. 665 00:34:07,760 --> 00:34:10,799 Speaker 1: You're on this side of you're on the team humanity. Yeah, 666 00:34:10,880 --> 00:34:15,760 Speaker 1: you're good. What advice would you give yourself ten years 667 00:34:15,760 --> 00:34:17,720 Speaker 1: ago to better prepare for today? 668 00:34:18,280 --> 00:34:22,120 Speaker 2: Learn Python for real. It's a programming language that's extensively 669 00:34:22,239 --> 00:34:25,000 Speaker 2: used in the community. I've always dabbled in it, but 670 00:34:25,160 --> 00:34:26,760 Speaker 2: I wish I had taken it more seriously. 671 00:34:27,040 --> 00:34:28,960 Speaker 1: Yeah, did you say, who had a daughter? 672 00:34:29,560 --> 00:34:30,520 Speaker 2: I have three daughters? 673 00:34:30,600 --> 00:34:33,320 Speaker 1: You have three daughters. I have two. You're if you 674 00:34:33,440 --> 00:34:37,319 Speaker 1: got three year you're you're on your own. Are you 675 00:34:37,360 --> 00:34:38,600 Speaker 1: making them study Python? 676 00:34:39,719 --> 00:34:42,759 Speaker 2: I am actually trying to do that. We're using a 677 00:34:42,800 --> 00:34:45,919 Speaker 2: microbit micro controller tool to do like a custom video 678 00:34:45,960 --> 00:34:49,280 Speaker 2: game controller. They prefer Scratch because it's a visual programming language, 679 00:34:49,280 --> 00:34:51,080 Speaker 2: but it has a Python interface too, and I'm like 680 00:34:51,200 --> 00:34:52,360 Speaker 2: pushing them towards Python. 681 00:34:52,719 --> 00:34:57,040 Speaker 1: Good chat box and image generators are the biggest things 682 00:34:57,040 --> 00:34:59,520 Speaker 1: in consumer AI right now. What do you think is 683 00:34:59,560 --> 00:35:01,520 Speaker 1: the next business application? 684 00:35:03,000 --> 00:35:08,360 Speaker 2: Private models, small models, fine tuned on your company's data 685 00:35:08,960 --> 00:35:10,640 Speaker 2: for you to use exclusively. 686 00:35:11,360 --> 00:35:14,720 Speaker 1: Are you using AI in your own personal life these days? 687 00:35:14,960 --> 00:35:16,759 Speaker 2: Honestly, I think a lot of us are using it 688 00:35:16,800 --> 00:35:18,000 Speaker 2: and we don't even realize it. 689 00:35:18,440 --> 00:35:18,680 Speaker 1: Yeah. 690 00:35:18,800 --> 00:35:22,360 Speaker 2: I mean, I'm a ficiano of foreign languages. There's translation 691 00:35:22,480 --> 00:35:25,680 Speaker 2: programs that are built using machine learning underneath. One of 692 00:35:25,680 --> 00:35:28,040 Speaker 2: the things I've been dabbling with lately is using tech 693 00:35:28,080 --> 00:35:31,120 Speaker 2: summarizations because I tend to be very loquacious in my 694 00:35:31,200 --> 00:35:33,759 Speaker 2: note taking and that is not so useful for other 695 00:35:33,800 --> 00:35:36,399 Speaker 2: people who would just like a paragraph. So that's something 696 00:35:36,440 --> 00:35:39,239 Speaker 2: I've been experimenting with myself just to help my everyday work. 697 00:35:39,400 --> 00:35:43,640 Speaker 1: Yeah. We hear many definitions of open related to technology. 698 00:35:44,239 --> 00:35:47,480 Speaker 1: What's your definition of open and how does it help 699 00:35:47,520 --> 00:35:48,000 Speaker 1: you innovate? 700 00:35:48,400 --> 00:35:54,239 Speaker 2: My definition of open is basically sharing and being vulnerable, 701 00:35:54,360 --> 00:35:57,239 Speaker 2: like not just sharing in a have a cookie way, 702 00:35:57,320 --> 00:35:59,560 Speaker 2: but in a you know what, I don't actually know 703 00:35:59,600 --> 00:36:02,360 Speaker 2: how this works? Could you help me? And being open 704 00:36:02,600 --> 00:36:06,040 Speaker 2: to being wrong, being open to somebody helping you and 705 00:36:06,120 --> 00:36:08,400 Speaker 2: making that collaboration work. So it's not just about like 706 00:36:08,440 --> 00:36:12,040 Speaker 2: the artifactor opening, it's your approach, like how you do 707 00:36:12,080 --> 00:36:12,880 Speaker 2: things being open. 708 00:36:13,120 --> 00:36:16,680 Speaker 1: Yeah yeah, well I think that wraps us up. How 709 00:36:16,680 --> 00:36:20,000 Speaker 1: can listeners follow your work and learn more about granted 710 00:36:20,120 --> 00:36:21,040 Speaker 1: and instruct lab. 711 00:36:21,320 --> 00:36:24,200 Speaker 2: Sure, you can visit our project webpage at instruct lab 712 00:36:24,320 --> 00:36:27,440 Speaker 2: dot ai, or you can visit our GitHub at GitHub 713 00:36:27,520 --> 00:36:30,839 Speaker 2: dot com slash instruct lab. We have lots of instructions 714 00:36:30,920 --> 00:36:33,600 Speaker 2: on how to get involved in an instruct lab wonderful. 715 00:36:33,960 --> 00:36:40,000 Speaker 1: Thank you so much, Thank you, Malcolm. A big thank 716 00:36:40,040 --> 00:36:43,840 Speaker 1: you to MO for the engaging discussion on the groundbreaking 717 00:36:44,160 --> 00:36:49,040 Speaker 1: possibilities of instruct lab. We've explored how this platform has 718 00:36:49,080 --> 00:36:53,439 Speaker 1: the potential to revolutionize industries from insurance to entertainment law 719 00:36:53,760 --> 00:36:56,640 Speaker 1: by using an open source community approach that makes it 720 00:36:56,680 --> 00:36:59,759 Speaker 1: easier for more people from all backgrounds to fine tune 721 00:36:59,800 --> 00:37:06,360 Speaker 1: my for specific purposes, ultimately making AI more accessible and 722 00:37:06,480 --> 00:37:11,640 Speaker 1: impactful than ever. Looking ahead, the future of AI isn't 723 00:37:11,680 --> 00:37:16,760 Speaker 1: just about technological efficiency. It's about enhancing our everyday experiences 724 00:37:16,800 --> 00:37:20,799 Speaker 1: in ways that were never possible before, like streamlining work 725 00:37:20,880 --> 00:37:25,040 Speaker 1: for doctors to improve the patient experience or assisting insurance 726 00:37:25,080 --> 00:37:30,480 Speaker 1: agents to improve the claims experience. Instruct Lab is paving 727 00:37:30,520 --> 00:37:34,839 Speaker 1: the way for more open, accessible AI future, one that's 728 00:37:34,880 --> 00:37:41,360 Speaker 1: built on collaboration and humanity. Smart Talks with IBM is 729 00:37:41,400 --> 00:37:46,080 Speaker 1: produced by Matt Romano, Joey Fishground and Jacob Goldstein. We're 730 00:37:46,280 --> 00:37:50,040 Speaker 1: edited by Lydia jen Kott. Our engineers are Sarah Bruger 731 00:37:50,360 --> 00:37:54,840 Speaker 1: and Ben Tolliday. Theme song by Gramscope Special thanks to 732 00:37:54,880 --> 00:37:57,400 Speaker 1: the eight Bar and IBM teams, as well as the 733 00:37:57,400 --> 00:38:00,840 Speaker 1: Pushkin marketing team. Smart Talks with ib VBM is a 734 00:38:00,880 --> 00:38:05,640 Speaker 1: production of Pushkin Industries and Ruby Studio at iHeartMedia. To 735 00:38:05,719 --> 00:38:11,360 Speaker 1: find more Pushkin podcasts, listen on the iHeartRadio app, Apple Podcasts, 736 00:38:11,440 --> 00:38:16,200 Speaker 1: or wherever you listen to podcasts. I'm Malcolm Gladwell. This 737 00:38:16,320 --> 00:38:19,960 Speaker 1: is a paid advertisement from IBM. The conversations on this 738 00:38:20,080 --> 00:38:35,960 Speaker 1: podcast don't necessarily represent IBM's positions, strategies or opinions.