1 00:00:00,160 --> 00:00:02,840 Speaker 1: Hey everyone, it's Robert and Joe here. Today we've got 2 00:00:02,880 --> 00:00:04,760 Speaker 1: something a little bit different to share with you. It 3 00:00:04,840 --> 00:00:08,039 Speaker 1: is a new season of the Smart Talks with IBM 4 00:00:08,160 --> 00:00:09,160 Speaker 1: podcast series. 5 00:00:09,640 --> 00:00:11,639 Speaker 2: Today we are witnessed to one of those rare moments 6 00:00:11,680 --> 00:00:14,400 Speaker 2: in history, the rise of an innovative technology with the 7 00:00:14,400 --> 00:00:18,680 Speaker 2: potential to radically transform business and society forever. The technology, 8 00:00:18,760 --> 00:00:22,240 Speaker 2: of course, is artificial intelligence, and it's the central focus 9 00:00:22,239 --> 00:00:24,840 Speaker 2: for this new season of Smart Talks with IBM. 10 00:00:25,320 --> 00:00:28,400 Speaker 1: Join hosts from your favorite Pushkin podcasts as they talk 11 00:00:28,480 --> 00:00:31,680 Speaker 1: with industry experts and leaders to explore how businesses can 12 00:00:31,720 --> 00:00:35,400 Speaker 1: integrate AI into their workflows and help drive real change 13 00:00:35,400 --> 00:00:38,240 Speaker 1: in this new era of AI. And of course, host 14 00:00:38,280 --> 00:00:40,479 Speaker 1: Malcolm Gladwell will be there to guide you through the 15 00:00:40,479 --> 00:00:42,640 Speaker 1: season and throw in his two cents as well. 16 00:00:43,120 --> 00:00:46,200 Speaker 2: Look out for new episodes of Smart Talks with IBM 17 00:00:46,400 --> 00:00:49,559 Speaker 2: every other week on the iHeartRadio app, Apple Podcasts, or 18 00:00:49,600 --> 00:00:53,360 Speaker 2: wherever you get your podcasts. And learn more at IBM 19 00:00:53,479 --> 00:01:03,360 Speaker 2: dot com, slash smart Talks, Pushkin. 20 00:01:08,959 --> 00:01:12,600 Speaker 3: Hello, Hello, welcome to Smart Talks with IBM, a podcast 21 00:01:12,600 --> 00:01:17,840 Speaker 3: from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. This 22 00:01:17,920 --> 00:01:21,240 Speaker 3: season We're diving back into the world of artificial intelligence, 23 00:01:21,520 --> 00:01:24,920 Speaker 3: but with a focus on the powerful concept of open 24 00:01:25,280 --> 00:01:30,440 Speaker 3: its possibilities, implications, and misconceptions. We'll look at openness from 25 00:01:30,520 --> 00:01:33,520 Speaker 3: a variety of angles and explore how the concept is 26 00:01:33,600 --> 00:01:37,200 Speaker 3: already reshaping industries, ways of doing business, and a very 27 00:01:37,280 --> 00:01:41,039 Speaker 3: notion of what's possible. In today's episode, I sat down 28 00:01:41,040 --> 00:01:44,880 Speaker 3: with mo Duffy, software engineering manager at red Hat, who 29 00:01:44,920 --> 00:01:48,600 Speaker 3: works on instruct Lab, a project co developed by red 30 00:01:48,600 --> 00:01:52,360 Speaker 3: Hat and IBM. Most shared with me how this new 31 00:01:52,440 --> 00:01:57,680 Speaker 3: initiative is revolutionizing AI training, making it not only more accessible, 32 00:01:57,880 --> 00:02:02,200 Speaker 3: but also more inclusive. In this project, unique in the industry, 33 00:02:02,400 --> 00:02:07,760 Speaker 3: allows developers to submit incremental contributions to one base AI model, 34 00:02:08,040 --> 00:02:12,240 Speaker 3: creating a continuous loop of development, much like normal open 35 00:02:12,280 --> 00:02:17,639 Speaker 3: source software. By leveraging community contributions and IBM's cutting edge 36 00:02:17,680 --> 00:02:20,840 Speaker 3: granite models, Mo in the team of ibmrs and red 37 00:02:20,840 --> 00:02:24,359 Speaker 3: hatters are paving the way for a future where AI 38 00:02:24,480 --> 00:02:28,880 Speaker 3: development is a communal endeavor. Our insights into open source 39 00:02:28,919 --> 00:02:33,959 Speaker 3: software extend beyond technical proficiency to the profound impact of 40 00:02:34,120 --> 00:02:37,720 Speaker 3: collaborative effort. At the heart of Moe's work is a 41 00:02:37,720 --> 00:02:42,519 Speaker 3: belief in democratizing technology, ensuring that AI becomes a tool 42 00:02:42,919 --> 00:02:47,320 Speaker 3: accessible to all. So let's explore how MOE, red Hat 43 00:02:47,360 --> 00:02:51,880 Speaker 3: and IBM are empowering individuals and businesses alike to reshape 44 00:02:51,880 --> 00:03:01,600 Speaker 3: the future of technology through collaboration and innovation. MO, thank 45 00:03:01,639 --> 00:03:03,800 Speaker 3: you for joining me today. Thank you so much, for 46 00:03:04,040 --> 00:03:08,480 Speaker 3: I have just about the most Irish name ever. I 47 00:03:08,560 --> 00:03:12,040 Speaker 3: do very proudure you weren't born in Ireland. 48 00:03:12,240 --> 00:03:13,239 Speaker 4: No, my grandparents. 49 00:03:13,320 --> 00:03:15,200 Speaker 3: Oh your grandparents, So I see, where did you grow up? 50 00:03:15,600 --> 00:03:16,639 Speaker 4: New York Queens? 51 00:03:16,960 --> 00:03:20,200 Speaker 3: Oh you're la see. So tell me a little bit 52 00:03:20,200 --> 00:03:22,960 Speaker 3: about how how you got to red hat. What was 53 00:03:22,960 --> 00:03:23,480 Speaker 3: your path? 54 00:03:23,960 --> 00:03:26,480 Speaker 4: When I was in high school, it was a chatty girl, 55 00:03:26,800 --> 00:03:29,280 Speaker 4: teenage girl on the phone. We had one phone line. 56 00:03:29,639 --> 00:03:32,440 Speaker 4: My older brother was studying at the local state college 57 00:03:32,440 --> 00:03:34,720 Speaker 4: computer science, and he had to tell that end to 58 00:03:35,040 --> 00:03:37,920 Speaker 4: compile his homework. One phone line, and I'm on it 59 00:03:37,960 --> 00:03:40,920 Speaker 4: all the time. He got very frustrated and he needed 60 00:03:40,920 --> 00:03:43,480 Speaker 4: a compiler to do his homework. So he bought red 61 00:03:43,480 --> 00:03:47,920 Speaker 4: Hat Linux from a CompUSA, brought it home and that 62 00:03:48,040 --> 00:03:49,960 Speaker 4: was on the family computer. So I learned Linux and 63 00:03:50,000 --> 00:03:52,720 Speaker 4: I started playing around with it. I really liked it 64 00:03:52,760 --> 00:03:56,680 Speaker 4: because you could customize everything, like the entire user interface. 65 00:03:57,040 --> 00:03:59,200 Speaker 4: You could actually modify the code of the programs you 66 00:03:59,240 --> 00:04:01,480 Speaker 4: were using to do what you wanted. And for me, 67 00:04:01,560 --> 00:04:03,760 Speaker 4: it was really cool because especially when you're a kid 68 00:04:03,800 --> 00:04:05,560 Speaker 4: and like people tell you this is the way things 69 00:04:05,560 --> 00:04:07,320 Speaker 4: are and you just have to deal with it, it's 70 00:04:07,400 --> 00:04:08,920 Speaker 4: nice to be like, I'm going to make things the 71 00:04:09,000 --> 00:04:12,280 Speaker 4: way I want, modify the code and playing. Yeah, it 72 00:04:12,360 --> 00:04:14,200 Speaker 4: was amazing and it was just such a time and 73 00:04:14,240 --> 00:04:17,240 Speaker 4: like before it was cool, I was doing it and 74 00:04:18,040 --> 00:04:20,000 Speaker 4: what I saw on that is sort of the potential 75 00:04:20,040 --> 00:04:23,080 Speaker 4: like number one of like a community of people working together. 76 00:04:23,080 --> 00:04:26,160 Speaker 4: And like the Internet existed, it was slow, it involved modems, 77 00:04:26,560 --> 00:04:28,480 Speaker 4: but there were people that you could talk to who 78 00:04:28,520 --> 00:04:32,279 Speaker 4: would give you tips and you'd share information, and this 79 00:04:32,400 --> 00:04:35,320 Speaker 4: collaborative building something together is really something special. 80 00:04:35,440 --> 00:04:35,640 Speaker 2: Right. 81 00:04:36,000 --> 00:04:39,640 Speaker 4: I could file a complaint to whatever large software company 82 00:04:39,680 --> 00:04:42,080 Speaker 4: made whatever software I was into, or I could go 83 00:04:42,080 --> 00:04:44,279 Speaker 4: to an open source software community and be like, hey, guys, 84 00:04:44,320 --> 00:04:46,719 Speaker 4: I think we should do this. I'm like, yeah, okay, 85 00:04:46,720 --> 00:04:48,880 Speaker 4: I'll help. I'll pitch in. So you don't feel powerless, 86 00:04:48,880 --> 00:04:50,240 Speaker 4: you feel like you can have an impact, and that 87 00:04:50,320 --> 00:04:54,200 Speaker 4: was really exciting to me. However, open source software has 88 00:04:54,360 --> 00:04:57,599 Speaker 4: a reputation for not having the best user interface, not 89 00:04:57,640 --> 00:05:02,159 Speaker 4: the best user experience. So I ended up studying Computer 90 00:05:02,200 --> 00:05:05,480 Speaker 4: science and Electronic media dual major, and then I did 91 00:05:05,560 --> 00:05:09,440 Speaker 4: human computeraction as my master's And my thought was, wouldn't 92 00:05:09,440 --> 00:05:13,120 Speaker 4: it be nice if this free software accessible to anybody, 93 00:05:13,440 --> 00:05:15,440 Speaker 4: if it was easier to use, some more people could 94 00:05:15,520 --> 00:05:18,160 Speaker 4: use it and take advantage of it. And so, long 95 00:05:18,200 --> 00:05:21,479 Speaker 4: story short, I ended up going to Red Hat saying, Hey, 96 00:05:21,480 --> 00:05:22,880 Speaker 4: I want to learn how you guys work. Let me 97 00:05:22,920 --> 00:05:25,719 Speaker 4: embed in your team draft out of my graduate program. 98 00:05:25,760 --> 00:05:27,520 Speaker 4: And I'm like, I want to do this for a living. 99 00:05:27,640 --> 00:05:30,240 Speaker 4: This is cooler. So I thought this is the way 100 00:05:30,279 --> 00:05:31,920 Speaker 4: to go, and I've been there ever since. They haven't 101 00:05:31,920 --> 00:05:32,760 Speaker 4: been able to get rid of me. 102 00:05:34,839 --> 00:05:37,560 Speaker 3: To backtrack just a little bit, you were talking about 103 00:05:37,600 --> 00:05:41,280 Speaker 3: the sense of community that surrounds this way of thinking 104 00:05:41,320 --> 00:05:45,039 Speaker 3: about software. Talk a little bit more about what that 105 00:05:45,080 --> 00:05:47,840 Speaker 3: community is like, the benefits of that community, why it 106 00:05:47,920 --> 00:05:48,720 Speaker 3: appeals to you. 107 00:05:49,200 --> 00:05:51,520 Speaker 4: Sure, well, you know part of the reason I actually 108 00:05:51,640 --> 00:05:54,720 Speaker 4: ended up going to the graduate school track. Suddenly you're 109 00:05:54,720 --> 00:05:57,920 Speaker 4: a peer of your professors and you're working side by 110 00:05:58,000 --> 00:06:01,040 Speaker 4: side with them. At some point they retire and you're 111 00:06:01,040 --> 00:06:04,599 Speaker 4: in the next generation. So it's sharing information, building on 112 00:06:04,640 --> 00:06:07,320 Speaker 4: the work of others in sort of this cycle that 113 00:06:07,520 --> 00:06:11,880 Speaker 4: extends past human lifespan and in the same way, like 114 00:06:11,960 --> 00:06:15,359 Speaker 4: the open source model is very similar, but you're actually 115 00:06:15,440 --> 00:06:18,080 Speaker 4: you're building something, and it's something in me. I'm just 116 00:06:18,440 --> 00:06:20,840 Speaker 4: really attracted, Like I don't like talking about stuff. I 117 00:06:21,000 --> 00:06:24,960 Speaker 4: like doing stuff with open source software. The software doesn't 118 00:06:24,960 --> 00:06:28,599 Speaker 4: cost anything, the code is out there, generally uses open 119 00:06:28,640 --> 00:06:32,039 Speaker 4: standards for the file formats. I can open up files 120 00:06:32,080 --> 00:06:34,680 Speaker 4: that I created and open source tools as a high 121 00:06:34,720 --> 00:06:38,279 Speaker 4: school student today because they were using open formats and 122 00:06:38,320 --> 00:06:40,800 Speaker 4: that software still exists. I can still compile the code 123 00:06:40,880 --> 00:06:44,080 Speaker 4: and it's an active community project. Like these things can 124 00:06:44,120 --> 00:06:46,800 Speaker 4: outlast any single company in the same way that the 125 00:06:46,839 --> 00:06:49,400 Speaker 4: academic community has been going on for so many years 126 00:06:49,440 --> 00:06:51,960 Speaker 4: and hopefully we'll continue moving on. So it's sort of 127 00:06:52,040 --> 00:06:54,919 Speaker 4: like not just the community around it, but just the 128 00:06:55,000 --> 00:06:58,119 Speaker 4: knowledge sharing and also bringing up the next generation as well. 129 00:06:58,160 --> 00:06:59,680 Speaker 4: Like all of that stuff really appealed to me. And 130 00:06:59,760 --> 00:07:02,440 Speaker 4: also so at the center of it the fact that 131 00:07:02,880 --> 00:07:05,960 Speaker 4: we could democratize it by following this open source process 132 00:07:06,320 --> 00:07:08,320 Speaker 4: and feel like we have some control. We're not at 133 00:07:08,360 --> 00:07:11,400 Speaker 4: the mercy of some faceless corporation making changes and we 134 00:07:11,440 --> 00:07:13,520 Speaker 4: have no impact. Like that really appealed to me too. 135 00:07:14,320 --> 00:07:19,400 Speaker 3: For those of us who are not software phisionados, take 136 00:07:19,400 --> 00:07:23,800 Speaker 3: a step backwards and give me a kind of description 137 00:07:23,920 --> 00:07:26,800 Speaker 3: of terms. What's the opposite of open source proprietary? 138 00:07:26,880 --> 00:07:28,360 Speaker 4: Proprietary is what we say. 139 00:07:28,280 --> 00:07:32,679 Speaker 3: So specifically and practically, the difference would be what between 140 00:07:32,720 --> 00:07:34,840 Speaker 3: something that was opens us in something that was proprietary. 141 00:07:34,920 --> 00:07:37,520 Speaker 4: Sure, so there's a lot of difference. So with open 142 00:07:37,560 --> 00:07:41,520 Speaker 4: source software you get these rights when you're given the software, 143 00:07:41,600 --> 00:07:44,000 Speaker 4: you get the right to be able to share it. 144 00:07:44,080 --> 00:07:46,520 Speaker 4: And depending on the lot, different licenses that are considered 145 00:07:46,520 --> 00:07:49,520 Speaker 4: open source have different little things that you have to 146 00:07:49,520 --> 00:07:55,320 Speaker 4: be aware of. With proprietary code, it's one copyright the company. 147 00:07:55,360 --> 00:07:57,720 Speaker 4: Even a lot of times, when you sign your employment 148 00:07:57,720 --> 00:08:00,160 Speaker 4: contract for a software company and you write code for them, 149 00:08:00,360 --> 00:08:02,000 Speaker 4: you don't own it. You sign over your rights to 150 00:08:02,040 --> 00:08:03,960 Speaker 4: the company. So if you leave the company, the code 151 00:08:03,960 --> 00:08:06,280 Speaker 4: doesn't go with you. It stays in the ownership of 152 00:08:06,280 --> 00:08:08,920 Speaker 4: that company. So then one like one company buys out 153 00:08:08,960 --> 00:08:11,080 Speaker 4: another and kills a product, that code's gone. 154 00:08:11,240 --> 00:08:15,160 Speaker 3: It's gone for a business, Why would a business want 155 00:08:15,240 --> 00:08:19,040 Speaker 3: to be have open source code as opposed to a proprietary. 156 00:08:18,400 --> 00:08:20,480 Speaker 4: Well for the same reasons. Like, say you're a business. 157 00:08:20,960 --> 00:08:24,840 Speaker 4: You've invested all this money into this software platform, right 158 00:08:25,320 --> 00:08:28,440 Speaker 4: and you've upskilled your employees on it, and it's a 159 00:08:28,480 --> 00:08:31,320 Speaker 4: core part of your business, and then a few years 160 00:08:31,360 --> 00:08:34,679 Speaker 4: later that company goes out of business or something happens, 161 00:08:34,840 --> 00:08:38,199 Speaker 4: or even something less drastic. You really need this future, 162 00:08:38,520 --> 00:08:41,640 Speaker 4: But for the company that makes the software, it's not 163 00:08:41,640 --> 00:08:44,480 Speaker 4: in their best interests. It's not worth the investment. They're 164 00:08:44,520 --> 00:08:46,440 Speaker 4: not going to do it. How do you get that future? 165 00:08:46,800 --> 00:08:49,320 Speaker 4: You either have to completely migrate to another solution, and 166 00:08:49,360 --> 00:08:51,280 Speaker 4: this is something it's core at your business, that's going 167 00:08:51,320 --> 00:08:53,439 Speaker 4: to be a big deal to migrate. But if it's 168 00:08:53,480 --> 00:08:57,240 Speaker 4: open source, you could either hire a team of experts. 169 00:08:57,440 --> 00:08:59,880 Speaker 4: You could hire software engineers who are able to go 170 00:09:00,120 --> 00:09:03,319 Speaker 4: do this for you. Go in the upstream software community, 171 00:09:03,800 --> 00:09:06,800 Speaker 4: implement the feature that you want, and it'll be rolled 172 00:09:06,840 --> 00:09:09,400 Speaker 4: into the next version of that company software. So even 173 00:09:09,440 --> 00:09:12,559 Speaker 4: if that company didn't want to implement the feature, if 174 00:09:12,600 --> 00:09:15,640 Speaker 4: they did it open source, they would inherit that feature 175 00:09:15,679 --> 00:09:18,160 Speaker 4: from the upstream community, is what we call it. So 176 00:09:18,240 --> 00:09:20,880 Speaker 4: you have some control over the situation. If it's open source, 177 00:09:20,960 --> 00:09:25,120 Speaker 4: you have an opportunity to actually affect change in the product, 178 00:09:25,400 --> 00:09:27,400 Speaker 4: and you could then pick it up or pay somebody 179 00:09:27,400 --> 00:09:29,520 Speaker 4: else to pick it up, or another company could form 180 00:09:29,559 --> 00:09:32,160 Speaker 4: and pick it up and keep it going. So there's 181 00:09:32,160 --> 00:09:34,640 Speaker 4: more possibilities. If it's open source, it's more like it's 182 00:09:34,679 --> 00:09:36,240 Speaker 4: like an insurance policy almost. 183 00:09:36,280 --> 00:09:39,840 Speaker 3: So innovation from the standpoint of the customer, innovation is 184 00:09:39,880 --> 00:09:42,880 Speaker 3: a lot easier when you're working in an open source environment. 185 00:09:43,040 --> 00:09:43,679 Speaker 4: Absolutely. 186 00:09:44,080 --> 00:09:47,840 Speaker 3: Yeah. So now at RedHat you're working with something called 187 00:09:47,840 --> 00:09:50,560 Speaker 3: instruct lab. Tell us a little bit about what that is. 188 00:09:51,040 --> 00:09:53,280 Speaker 4: So the thing that really excites me about getting to 189 00:09:53,320 --> 00:09:55,680 Speaker 4: work on this project is AI is sort of that 190 00:09:55,760 --> 00:09:58,920 Speaker 4: has been this scary thing for me because it's one 191 00:09:58,920 --> 00:10:02,040 Speaker 4: of those things like in order to be able to 192 00:10:02,120 --> 00:10:07,239 Speaker 4: pre train a model, you have to have unobtainium GPS, 193 00:10:07,920 --> 00:10:12,200 Speaker 4: you have to have rich resources, It takes months, it 194 00:10:12,240 --> 00:10:17,200 Speaker 4: takes expertise. There's a small handful of companies that can 195 00:10:17,240 --> 00:10:21,240 Speaker 4: build a model from pre train to something usable, and 196 00:10:21,480 --> 00:10:23,800 Speaker 4: it kind of feels like those early days when I 197 00:10:23,920 --> 00:10:26,640 Speaker 4: was kind of delving in software in the same way. 198 00:10:26,679 --> 00:10:30,000 Speaker 4: I think if more people could contribute to AI models, 199 00:10:30,800 --> 00:10:34,400 Speaker 4: then it wouldn't be just influenced by whichever company had 200 00:10:34,440 --> 00:10:37,679 Speaker 4: the resources to build it. And there's been a lot 201 00:10:37,679 --> 00:10:41,040 Speaker 4: of emphasis on pre training models, so taking massive terabytes 202 00:10:41,160 --> 00:10:45,120 Speaker 4: data sets, throwing them through masses of GPS over months 203 00:10:45,120 --> 00:10:48,440 Speaker 4: of time, spending hundreds of millions of dollars to build 204 00:10:48,480 --> 00:10:51,320 Speaker 4: a base model. But when instruct lab does is say okay, 205 00:10:51,559 --> 00:10:54,440 Speaker 4: you have a base model. We're going to fine tune in. 206 00:10:54,480 --> 00:10:57,920 Speaker 4: On the other end, it takes less compute resources. The 207 00:10:57,960 --> 00:11:00,240 Speaker 4: way we've built in struck lab, you can play around 208 00:11:00,280 --> 00:11:02,760 Speaker 4: with the technology and learn it on it off the 209 00:11:02,760 --> 00:11:06,200 Speaker 4: shelf laptop that you can actually buy. So in this 210 00:11:06,320 --> 00:11:10,040 Speaker 4: way we're enabling a much broader set of people to 211 00:11:10,160 --> 00:11:12,839 Speaker 4: play with AI, to contribute it, to modify it. And 212 00:11:12,880 --> 00:11:15,840 Speaker 4: I'll tell you one story from red Hat Succi, who 213 00:11:15,920 --> 00:11:20,360 Speaker 4: is our chief diversity officer, very interested in inclusive language 214 00:11:20,360 --> 00:11:23,280 Speaker 4: and open source software, doesn't have any experience with AI. 215 00:11:24,040 --> 00:11:26,079 Speaker 4: We have a community model that we have an upstream 216 00:11:26,080 --> 00:11:28,680 Speaker 4: project around for people to contribute knowledge and skills to 217 00:11:28,720 --> 00:11:30,840 Speaker 4: the model. She's like, I want to teach the model 218 00:11:31,160 --> 00:11:34,120 Speaker 4: how to use inclusive language, like replace this word with 219 00:11:34,160 --> 00:11:35,959 Speaker 4: this word or this word with this word. OHM Like, oh, 220 00:11:36,000 --> 00:11:38,880 Speaker 4: that's so cool. So She paired up with Nicholas who 221 00:11:38,960 --> 00:11:41,760 Speaker 4: is a technical guy at red Hat, and they built 222 00:11:41,880 --> 00:11:44,920 Speaker 4: and submitted a skill to the model that you can 223 00:11:45,000 --> 00:11:47,160 Speaker 4: just tell the model, can you please take this document 224 00:11:47,240 --> 00:11:49,679 Speaker 4: and translate this language to more inclusive language and it 225 00:11:49,720 --> 00:11:52,120 Speaker 4: will do it. And they submitted it to the community. 226 00:11:52,200 --> 00:11:53,920 Speaker 4: They were so proud. It was like, that's the kind 227 00:11:53,960 --> 00:11:56,360 Speaker 4: of thing that, like, you know, maybe a company would 228 00:11:56,400 --> 00:11:58,840 Speaker 4: be incentivized to do that, but if you have some 229 00:11:58,960 --> 00:12:02,560 Speaker 4: tooling that's open source and something that anybody could access, 230 00:12:02,720 --> 00:12:05,320 Speaker 4: than those communities could actually get together and build that 231 00:12:05,400 --> 00:12:06,640 Speaker 4: knowledge into AI models. 232 00:12:06,880 --> 00:12:11,400 Speaker 3: Just so understand, what you guys have is the structure 233 00:12:11,480 --> 00:12:15,760 Speaker 3: for an AI system, And in other cases, individual companies 234 00:12:15,920 --> 00:12:19,880 Speaker 3: own and train their own AI systems. It takes enormous 235 00:12:19,880 --> 00:12:22,920 Speaker 3: amount of resources. They hoover up all kinds of information, 236 00:12:23,480 --> 00:12:26,640 Speaker 3: train it according to their own hidden set of rules, 237 00:12:26,720 --> 00:12:31,160 Speaker 3: and then a customer might use that for some price. 238 00:12:31,520 --> 00:12:33,440 Speaker 3: What you're saying is, in the same way that we 239 00:12:33,520 --> 00:12:37,360 Speaker 3: democratize the writing of software before, let's democratize the training 240 00:12:37,400 --> 00:12:41,040 Speaker 3: of an AI system. So anyone can contribute here and 241 00:12:41,480 --> 00:12:45,160 Speaker 3: teach the model the things that they're interested in teaching 242 00:12:45,160 --> 00:12:48,079 Speaker 3: the model. I'm guessing correct me. On the one hand, 243 00:12:48,559 --> 00:12:50,680 Speaker 3: this model, at least in the beginning, is going to 244 00:12:50,679 --> 00:12:53,840 Speaker 3: have a lot fewer resources available to it. But on 245 00:12:53,880 --> 00:12:55,640 Speaker 3: the other hand, it's going to have a much more 246 00:12:56,000 --> 00:12:58,000 Speaker 3: diverse set of inputs. 247 00:12:58,440 --> 00:13:01,520 Speaker 4: That's right. And the other thing is that IBM, basically 248 00:13:01,600 --> 00:13:04,280 Speaker 4: is part of this project, has something called the Granite 249 00:13:04,320 --> 00:13:07,480 Speaker 4: Model family, and they've donated some granite models. So these 250 00:13:07,520 --> 00:13:10,200 Speaker 4: are the ones that take the months and terabytes of 251 00:13:10,280 --> 00:13:13,440 Speaker 4: data and all the GPUs to train. So IBM has 252 00:13:13,480 --> 00:13:16,760 Speaker 4: created one of those, and they have listed out and 253 00:13:16,800 --> 00:13:18,959 Speaker 4: linked to the data sets that they used, and they 254 00:13:19,000 --> 00:13:21,920 Speaker 4: talk about the relative proportions they used when pre training, 255 00:13:22,280 --> 00:13:24,280 Speaker 4: so it's not just the black box. You know where 256 00:13:24,320 --> 00:13:27,160 Speaker 4: the data came from, which is a pretty open position 257 00:13:27,200 --> 00:13:29,760 Speaker 4: to take. That is what we recommend as the base. 258 00:13:29,840 --> 00:13:32,440 Speaker 4: So you use the instruct lab tuning. You take this 259 00:13:32,520 --> 00:13:35,480 Speaker 4: base granite model that IBM has provided, and you use 260 00:13:35,480 --> 00:13:37,920 Speaker 4: the instruct lab tooling that red Hat works on, and 261 00:13:37,960 --> 00:13:40,280 Speaker 4: you use that to fine tune the model to make 262 00:13:40,320 --> 00:13:41,960 Speaker 4: it whatever you want. 263 00:13:42,480 --> 00:13:45,400 Speaker 3: I want to go back to the partnership between IBM 264 00:13:45,480 --> 00:13:49,480 Speaker 3: and red Hat here with them providing the granite model 265 00:13:50,080 --> 00:13:53,320 Speaker 3: to your instruct lab Is this the first time red 266 00:13:53,360 --> 00:13:56,760 Speaker 3: hat and IBM have collaborated like this, I think it's. 267 00:13:56,600 --> 00:13:59,920 Speaker 4: Something that's been going on. Like another a product within 268 00:14:00,080 --> 00:14:02,560 Speaker 4: the red hat family would be open Shift AI, where 269 00:14:02,559 --> 00:14:06,199 Speaker 4: they collaborate a lot with IBM Research team, Like BLM 270 00:14:06,280 --> 00:14:08,240 Speaker 4: is one of the components of that product that there's 271 00:14:08,440 --> 00:14:13,160 Speaker 4: a nice kind of exchange and collaboration between the two companies. 272 00:14:13,920 --> 00:14:16,559 Speaker 3: How large is the potential community of people who might 273 00:14:16,640 --> 00:14:18,240 Speaker 3: contribute to instruct lab. 274 00:14:19,160 --> 00:14:21,720 Speaker 4: It could be thousands of people. I mean, we'll see. 275 00:14:21,760 --> 00:14:25,520 Speaker 4: It's early days. This is early technology that was invented 276 00:14:25,520 --> 00:14:28,000 Speaker 4: at IBM Research that they partnered with us at red 277 00:14:28,000 --> 00:14:30,840 Speaker 4: Hat to kind of build the software around it. There's 278 00:14:30,880 --> 00:14:33,160 Speaker 4: still more to go, Like right now, we have a 279 00:14:33,240 --> 00:14:35,240 Speaker 4: team in the community that's actually trying to build a 280 00:14:35,280 --> 00:14:38,800 Speaker 4: web interface to make it easier for anybody to contribute. 281 00:14:38,960 --> 00:14:40,480 Speaker 4: So we have a lot of those sort of user 282 00:14:40,520 --> 00:14:43,960 Speaker 4: experience for the contributor to the model stuff to work 283 00:14:44,000 --> 00:14:46,680 Speaker 4: out that we're still actively building on. But like my 284 00:14:46,880 --> 00:14:49,520 Speaker 4: vision for it even is I like going back to 285 00:14:49,520 --> 00:14:52,600 Speaker 4: that academic model of learning from what others and building 286 00:14:52,680 --> 00:14:55,560 Speaker 4: upon it over time. It would be very good for 287 00:14:55,720 --> 00:14:58,560 Speaker 4: us to sort of go out and try to collaborate 288 00:14:58,720 --> 00:15:01,520 Speaker 4: with academics of all, like, hey, you know, the model 289 00:15:01,520 --> 00:15:05,000 Speaker 4: doesn't know about your field, would you like to put 290 00:15:05,040 --> 00:15:07,440 Speaker 4: something into the model about your field so it knows 291 00:15:07,480 --> 00:15:10,520 Speaker 4: about it, or even you know, talk to the model 292 00:15:10,760 --> 00:15:13,160 Speaker 4: it got it wrong, let's correct it. Can we lean 293 00:15:13,200 --> 00:15:15,320 Speaker 4: on your expertise to correct it and make sure it 294 00:15:15,320 --> 00:15:18,400 Speaker 4: gets it right and sort of use that community model 295 00:15:18,440 --> 00:15:22,480 Speaker 4: as a way for everybody to collaborate because before instruct lab, 296 00:15:23,240 --> 00:15:26,680 Speaker 4: my understanding is if you wanted to take a model 297 00:15:26,720 --> 00:15:28,800 Speaker 4: that's open source license and play with it, you could 298 00:15:28,840 --> 00:15:30,520 Speaker 4: do that. You could take a model kind of off 299 00:15:30,520 --> 00:15:33,560 Speaker 4: the shelf from Hugging Face and fine tune it yourself. 300 00:15:33,960 --> 00:15:35,520 Speaker 4: But it's a bit of a dead end because you 301 00:15:35,560 --> 00:15:38,120 Speaker 4: made your contributions, but there's no way for other people 302 00:15:38,600 --> 00:15:41,520 Speaker 4: to collaborate with you. So the way that we've built 303 00:15:41,560 --> 00:15:45,640 Speaker 4: this is based on how the technology works. Everybody can 304 00:15:45,680 --> 00:15:48,040 Speaker 4: contribute to it. This is something that you can keep 305 00:15:48,080 --> 00:15:49,520 Speaker 4: growing and growing and growing over time. 306 00:15:49,840 --> 00:15:53,400 Speaker 3: Yeah. Yeah, what's the level of expertise necessary to be 307 00:15:53,440 --> 00:15:54,160 Speaker 3: a contributor? 308 00:15:54,760 --> 00:15:56,640 Speaker 4: You don't need to be a data scientist and you 309 00:15:56,680 --> 00:15:59,600 Speaker 4: don't need to have exotic hardware. Honestly, if you don't 310 00:15:59,600 --> 00:16:02,520 Speaker 4: even have laptop hardware that meets SUSPEC for doing instruct 311 00:16:02,560 --> 00:16:05,720 Speaker 4: labs laptop version. You can submit it to the community 312 00:16:05,800 --> 00:16:08,320 Speaker 4: and then we'll actually build it for you. We have 313 00:16:08,400 --> 00:16:10,800 Speaker 4: bots and stuff that do that, and we're hoping over 314 00:16:10,840 --> 00:16:13,320 Speaker 4: time to make that more accessible, first by having a 315 00:16:13,440 --> 00:16:16,040 Speaker 4: user interface and then maybe later on having a web service. 316 00:16:16,360 --> 00:16:19,560 Speaker 3: Yeah, so give me an example of how a business 317 00:16:19,680 --> 00:16:21,720 Speaker 3: might make use of instruct lab. 318 00:16:22,280 --> 00:16:24,800 Speaker 4: One of the things that businesses are doing with AI 319 00:16:24,920 --> 00:16:28,600 Speaker 4: right now is using hosted API services. They're quite expensive, 320 00:16:28,880 --> 00:16:31,680 Speaker 4: but they're finding value, but it's hard given the amount 321 00:16:31,720 --> 00:16:34,000 Speaker 4: of money they're spending. And one of the things that's 322 00:16:34,000 --> 00:16:35,840 Speaker 4: a little scary about it too, is like you have 323 00:16:36,200 --> 00:16:40,560 Speaker 4: very sensitive internal documents and you have employees maybe not 324 00:16:40,760 --> 00:16:43,640 Speaker 4: understanding what they're actually doing because you know, how would 325 00:16:43,680 --> 00:16:47,080 Speaker 4: you if you're not technical enough when you're asking said 326 00:16:47,720 --> 00:16:53,440 Speaker 4: public web service AI model information about your copy pasting 327 00:16:53,560 --> 00:16:57,600 Speaker 4: internal company documents. It's going across the Internet into another 328 00:16:57,640 --> 00:17:00,800 Speaker 4: company's hands, and that company probably shouldn't have access to that. 329 00:17:01,280 --> 00:17:04,280 Speaker 4: So what both RedHat and IBM in the space are 330 00:17:04,280 --> 00:17:07,640 Speaker 4: looking at, like the instruct lab model is very modest. 331 00:17:07,640 --> 00:17:11,160 Speaker 4: It's seven billion parameter model very small. It's very cheap 332 00:17:11,200 --> 00:17:14,600 Speaker 4: to serve inference on a seven billion parameter model. It's 333 00:17:14,600 --> 00:17:18,120 Speaker 4: competing with trillion parameter models that are hosted. You take 334 00:17:18,160 --> 00:17:21,000 Speaker 4: this small model that is cheap to run inference on, 335 00:17:21,640 --> 00:17:25,560 Speaker 4: you train it with your own company's proprietary data inside 336 00:17:25,600 --> 00:17:28,159 Speaker 4: the walls of your company, on your own hardware. You 337 00:17:28,200 --> 00:17:31,480 Speaker 4: can do all sorts of actual data analysis on your 338 00:17:31,480 --> 00:17:34,080 Speaker 4: most sensitive data and have the confidence that has not 339 00:17:34,160 --> 00:17:35,080 Speaker 4: left the premises. 340 00:17:35,920 --> 00:17:38,960 Speaker 3: In that use case, you're not actually training the model 341 00:17:39,000 --> 00:17:42,240 Speaker 3: for everyone. You're just taking it and doing some private 342 00:17:42,280 --> 00:17:44,720 Speaker 3: stuff on it exactly, which doesn't leave the building. But 343 00:17:44,800 --> 00:17:49,320 Speaker 3: that's separate from an interaction where you're doing something that 344 00:17:50,000 --> 00:17:51,200 Speaker 3: contributes overall. 345 00:17:51,600 --> 00:17:54,200 Speaker 4: Right, And that's something maybe that I should be more 346 00:17:54,240 --> 00:17:56,480 Speaker 4: clear about is there's sort of two tracks here, and 347 00:17:56,560 --> 00:17:59,840 Speaker 4: this is very red hat classic. You have your upstream 348 00:18:00,040 --> 00:18:02,960 Speaker 4: community track and you have your business product tract. So 349 00:18:03,040 --> 00:18:07,040 Speaker 4: the upstream community track is just enabling anybody to contribute 350 00:18:07,080 --> 00:18:09,000 Speaker 4: to a model in a collaborative way and play with it. 351 00:18:09,440 --> 00:18:13,280 Speaker 4: The downstream product business oriented track is now take that 352 00:18:13,400 --> 00:18:18,000 Speaker 4: tech that we've honed and developed in the open community 353 00:18:18,680 --> 00:18:21,160 Speaker 4: and apply it to your business knowledge and skills. 354 00:18:22,200 --> 00:18:26,040 Speaker 3: This community driven approach marks a pivotal shift towards more 355 00:18:26,040 --> 00:18:31,840 Speaker 3: accessible AI solutions. The contrast between externally hosted AI services 356 00:18:32,119 --> 00:18:35,320 Speaker 3: and the open model enhanced by instruct lab underscores the 357 00:18:35,320 --> 00:18:40,440 Speaker 3: potential for broader adoption of AI in diverse business contexts. 358 00:18:40,920 --> 00:18:44,320 Speaker 3: She envisions a future in which technological innovation is more 359 00:18:44,359 --> 00:18:48,679 Speaker 3: tailored to individual business needs, guided by principles of openness 360 00:18:48,800 --> 00:18:53,679 Speaker 3: and security. To an imaginary case study, Sure, I'm a 361 00:18:53,760 --> 00:18:58,520 Speaker 3: law firm, I'm an entertainment law I have one hundred 362 00:18:58,560 --> 00:19:02,879 Speaker 3: clients who are big stars. They all have incredibly complicated contracts. 363 00:19:03,840 --> 00:19:08,280 Speaker 3: I feed a thousand of my company's contracts from the 364 00:19:08,359 --> 00:19:12,200 Speaker 3: last ten years into the model, and then every time 365 00:19:12,240 --> 00:19:14,800 Speaker 3: I have a new contract, I ask the model, am 366 00:19:14,840 --> 00:19:17,520 Speaker 3: I missing something? Can you go back and look through 367 00:19:17,560 --> 00:19:19,919 Speaker 3: all our own contracts and show me a contract that 368 00:19:20,200 --> 00:19:23,439 Speaker 3: is missing key components or exposes us to some liability. 369 00:19:24,320 --> 00:19:27,960 Speaker 3: In that case, the model would know my law firm 370 00:19:28,200 --> 00:19:31,560 Speaker 3: contracts really, really well. It's as if they've been working 371 00:19:32,080 --> 00:19:35,480 Speaker 3: out my law firm. They're not distracted by other people's 372 00:19:35,520 --> 00:19:41,240 Speaker 3: particular styles or a bunch of contracts from the utility industry, 373 00:19:41,400 --> 00:19:46,040 Speaker 3: or the They know entertainment law contracts exactly. 374 00:19:46,160 --> 00:19:48,000 Speaker 4: Yeah, and you can train it in your own image, 375 00:19:48,040 --> 00:19:51,800 Speaker 4: your style of doing things. It's something that your company 376 00:19:51,880 --> 00:19:55,200 Speaker 4: can produce that is uniquely helpful to you. No third 377 00:19:55,240 --> 00:19:57,720 Speaker 4: party could do that because no third party understands how 378 00:19:57,720 --> 00:20:01,200 Speaker 4: you do business and understands your his street in your documents. 379 00:20:01,520 --> 00:20:04,120 Speaker 4: So it's sort of a way of getting value out 380 00:20:04,119 --> 00:20:06,000 Speaker 4: of the stuff you already have sitting in a file 381 00:20:06,040 --> 00:20:08,200 Speaker 4: cabinet somewhere. It's very cool. 382 00:20:08,480 --> 00:20:11,480 Speaker 3: Yeah, give me a sort of a real world case 383 00:20:11,480 --> 00:20:14,720 Speaker 3: study where you think the business use case would be 384 00:20:14,720 --> 00:20:19,320 Speaker 3: really powerful. What's a business that really could see an 385 00:20:19,359 --> 00:20:23,360 Speaker 3: advantage to using instruct lab in its way. 386 00:20:23,840 --> 00:20:26,119 Speaker 4: The demo that I've given a couple of times at 387 00:20:26,119 --> 00:20:29,680 Speaker 4: different events used an imaginary insurance company. So you say, 388 00:20:29,720 --> 00:20:33,639 Speaker 4: you have this company, you have to recommend repairs for 389 00:20:33,720 --> 00:20:37,159 Speaker 4: various types of claims. You've been doing this for years, 390 00:20:37,200 --> 00:20:40,040 Speaker 4: you know. If you know the windshield's broken and you've 391 00:20:40,080 --> 00:20:42,920 Speaker 4: gotten this type of accident and it's this model car, 392 00:20:43,119 --> 00:20:44,880 Speaker 4: these are the kinds of things you want to look at. 393 00:20:45,560 --> 00:20:48,359 Speaker 4: So you could talk to any insurance agent in the 394 00:20:48,359 --> 00:20:51,080 Speaker 4: field and be like, oh, you know, it's a Tesla. 395 00:20:51,160 --> 00:20:53,280 Speaker 4: You might want to look at the battery or something like. 396 00:20:53,400 --> 00:20:56,800 Speaker 4: They'll have some latent knowledge just so you can take 397 00:20:56,840 --> 00:20:59,520 Speaker 4: that and train it into a model. Honestly, I think 398 00:20:59,560 --> 00:21:02,840 Speaker 4: these kind of new technologies are better when they're less visible. 399 00:21:03,440 --> 00:21:05,879 Speaker 4: So say you have the claims agents in the field 400 00:21:05,960 --> 00:21:07,960 Speaker 4: and they have this tool and they're kind of entering 401 00:21:07,960 --> 00:21:10,920 Speaker 4: the claim data. They're on the scene at the car, 402 00:21:11,520 --> 00:21:14,159 Speaker 4: and it might say, oh, look, I see this is 403 00:21:14,200 --> 00:21:16,720 Speaker 4: a Ford fiesta. These are things you want to look 404 00:21:16,760 --> 00:21:19,960 Speaker 4: at for this type of accident. As you're entering the data, 405 00:21:20,400 --> 00:21:22,280 Speaker 4: it could be going through the knowledge you had loaded 406 00:21:22,280 --> 00:21:24,760 Speaker 4: into the model and be making these suggestions based on 407 00:21:24,800 --> 00:21:27,719 Speaker 4: your company's background, and hey, you know, let's not make 408 00:21:27,760 --> 00:21:30,280 Speaker 4: the same mistake twice. Let's make new mistakes and let's 409 00:21:30,359 --> 00:21:33,240 Speaker 4: learn from the stuff we already did. So that's one example, 410 00:21:33,280 --> 00:21:35,840 Speaker 4: but there's so many different industries in ways that this 411 00:21:35,920 --> 00:21:38,679 Speaker 4: could help, and it could make those agents in the 412 00:21:38,720 --> 00:21:40,200 Speaker 4: field more efficient. 413 00:21:40,920 --> 00:21:43,320 Speaker 3: Have you had anyone talk to you about using instruct 414 00:21:43,400 --> 00:21:44,919 Speaker 3: lab in a way that surprised you. 415 00:21:46,960 --> 00:21:51,600 Speaker 4: I mean, some people have done funky things, but sort 416 00:21:51,640 --> 00:21:54,000 Speaker 4: of playing with the skills stuff, that's where I see 417 00:21:54,040 --> 00:21:57,040 Speaker 4: a lot of creativity. The difference between knowledge and skills 418 00:21:57,080 --> 00:22:00,960 Speaker 4: is that knowledge is pretty pretty understandable, right, oh, historical 419 00:22:00,960 --> 00:22:04,480 Speaker 4: insurance claims or you know, legal contracts. Skills are a 420 00:22:04,520 --> 00:22:07,439 Speaker 4: little different. So whenever somebody submits a skill, sometimes it 421 00:22:07,480 --> 00:22:09,520 Speaker 4: tends to be really creative because it's not something that's 422 00:22:09,560 --> 00:22:12,880 Speaker 4: super intuitive. Somebody submitted a skill. I don't know how 423 00:22:12,920 --> 00:22:15,720 Speaker 4: well it worked, but it was like making ASKI art, 424 00:22:15,920 --> 00:22:18,119 Speaker 4: like draw me a I don't know, draw me a 425 00:22:18,160 --> 00:22:20,040 Speaker 4: dog I would do like an ASKI art dog. I mean, 426 00:22:20,080 --> 00:22:22,600 Speaker 4: it's stuff that you can do programmatically. One that was 427 00:22:22,640 --> 00:22:26,560 Speaker 4: actually very very helpful was, you know, take this table 428 00:22:26,600 --> 00:22:29,720 Speaker 4: of data and convert it to this format, like, ooh, 429 00:22:29,840 --> 00:22:31,440 Speaker 4: that's nice. That actually saves me time. 430 00:22:32,000 --> 00:22:34,520 Speaker 3: How far away are we from the day when I 431 00:22:34,640 --> 00:22:39,320 Speaker 3: Malcolm Globwell technology ignore Amus can go home and easily 432 00:22:39,400 --> 00:22:44,440 Speaker 3: interact with instruct lab Maybe a few months, a few months, 433 00:22:45,560 --> 00:22:46,760 Speaker 3: you're gonna say a few years. 434 00:22:47,400 --> 00:22:49,080 Speaker 4: No, I think it'd be a few months. 435 00:22:49,680 --> 00:22:50,879 Speaker 3: Wow, I hope. 436 00:22:51,560 --> 00:22:53,280 Speaker 4: Hey it's power open source innovation. 437 00:22:53,680 --> 00:22:57,679 Speaker 3: Yeah, oh that's really interesting. Yeah, I'm always taken by surprise. 438 00:22:58,000 --> 00:23:00,640 Speaker 3: I'm still thinking in twentieth century terms about how long 439 00:23:00,680 --> 00:23:03,880 Speaker 3: things take, and you're in the twenty second century as 440 00:23:03,880 --> 00:23:04,240 Speaker 3: well as. 441 00:23:04,119 --> 00:23:04,719 Speaker 1: I could tell. 442 00:23:04,960 --> 00:23:09,240 Speaker 4: The instruct lab core invention was invented in a hotel 443 00:23:09,320 --> 00:23:12,400 Speaker 4: room at an AI conference in December with an amazing 444 00:23:12,440 --> 00:23:15,560 Speaker 4: group of IBM research guys December of twenty twenty three. 445 00:23:15,840 --> 00:23:18,560 Speaker 3: Wait, back up, you have to tell the story. 446 00:23:18,760 --> 00:23:21,760 Speaker 4: This group of guys we've been working with, they were 447 00:23:21,800 --> 00:23:24,159 Speaker 4: at this conference together, and it's a really funny story 448 00:23:24,200 --> 00:23:27,080 Speaker 4: because you know, it's hard to get access to GPUs 449 00:23:27,359 --> 00:23:29,200 Speaker 4: and like even you know, you're at IBM and it's 450 00:23:29,200 --> 00:23:31,960 Speaker 4: hard to get access because everybody wants access. They did 451 00:23:32,000 --> 00:23:34,919 Speaker 4: it over Christmas break because nobody was using the cluster 452 00:23:35,000 --> 00:23:37,119 Speaker 4: at the time, and they ran all of these experiments 453 00:23:37,119 --> 00:23:38,960 Speaker 4: and I'm like, whoa, this is really cool. 454 00:23:39,359 --> 00:23:43,320 Speaker 3: And wait. Their idea was we can do a stripped 455 00:23:43,320 --> 00:23:48,639 Speaker 3: down AI model, and was the idea and even back 456 00:23:48,680 --> 00:23:51,399 Speaker 3: then combine it with grantede, what was the original idea? 457 00:23:51,440 --> 00:23:54,880 Speaker 4: The original idea, it's sort of multi there's like multiple 458 00:23:54,920 --> 00:23:57,520 Speaker 4: aspects to it. So like one of the aspects it 459 00:23:57,560 --> 00:23:59,720 Speaker 4: actually came on later, but it starts at the beginning 460 00:23:59,760 --> 00:24:03,879 Speaker 4: of the workflow. Is you're using a taxonomy to organize 461 00:24:03,960 --> 00:24:06,440 Speaker 4: how you're fine tuning the model. So the old approach 462 00:24:06,560 --> 00:24:08,800 Speaker 4: they call it the blender approach, to just take a 463 00:24:08,800 --> 00:24:11,439 Speaker 4: bunch of data of roughly the type of data that 464 00:24:11,520 --> 00:24:13,320 Speaker 4: you'd like, and you kind of throw it in and 465 00:24:13,359 --> 00:24:16,080 Speaker 4: then see what comes out. Don't like it, Okay, throw 466 00:24:16,160 --> 00:24:19,160 Speaker 4: in more, try again, see what comes out. They had 467 00:24:19,400 --> 00:24:22,520 Speaker 4: used this taxonomy technique, so you actually build like a 468 00:24:22,560 --> 00:24:25,919 Speaker 4: taxonomy of like categories and subfolders of like this is 469 00:24:25,960 --> 00:24:28,400 Speaker 4: the knowledge and skills that we want to train into 470 00:24:28,440 --> 00:24:32,320 Speaker 4: the model. And that way you're sort of systematic about 471 00:24:32,440 --> 00:24:35,520 Speaker 4: what you're adding, and you can also identify gaps pretty easily. 472 00:24:35,560 --> 00:24:37,280 Speaker 4: Oh I don't have a category for that, let me 473 00:24:37,280 --> 00:24:40,080 Speaker 4: add that. So that's like one of the parts of 474 00:24:40,119 --> 00:24:40,840 Speaker 4: the invention here. 475 00:24:41,680 --> 00:24:46,600 Speaker 3: Point number one is let's be intentional and deliberate in 476 00:24:46,640 --> 00:24:47,880 Speaker 3: how we build and train this thing. 477 00:24:48,119 --> 00:24:51,359 Speaker 4: Yeah, and then the next component would be okay, so 478 00:24:51,680 --> 00:24:54,240 Speaker 4: it is actually quite expensive. Part of the expense of 479 00:24:54,280 --> 00:24:57,800 Speaker 4: like tuning models and just training models in general is 480 00:24:57,840 --> 00:25:01,040 Speaker 4: coming up with the data. And what they wanted to 481 00:25:01,080 --> 00:25:03,199 Speaker 4: do is have a technique where you could have just 482 00:25:03,240 --> 00:25:06,320 Speaker 4: a little bit of data and expand it with something 483 00:25:06,359 --> 00:25:09,760 Speaker 4: they're calling synthetic data generation. And this is where it's 484 00:25:09,760 --> 00:25:13,680 Speaker 4: sort of like you have this student and teacher workflow, 485 00:25:14,320 --> 00:25:19,040 Speaker 4: so you have your taxonomy. The taxonomy has sort of 486 00:25:19,040 --> 00:25:21,959 Speaker 4: the knowledge like a business's knowledge documents, their insurance claims, 487 00:25:22,240 --> 00:25:25,440 Speaker 4: and it has these quizzes that you write and that's 488 00:25:25,520 --> 00:25:27,480 Speaker 4: to teach the model. So I'm writing a quiz based 489 00:25:27,600 --> 00:25:29,200 Speaker 4: just like you do in school. You read the chapter 490 00:25:29,400 --> 00:25:31,480 Speaker 4: on the American Revolution, and then you answer a ten 491 00:25:31,560 --> 00:25:34,840 Speaker 4: question quiz where you're giving the model quiz. You need 492 00:25:34,840 --> 00:25:38,040 Speaker 4: at least five questions and answers, and the answers need 493 00:25:38,080 --> 00:25:40,679 Speaker 4: to be taken from the context of the document, and 494 00:25:40,800 --> 00:25:44,280 Speaker 4: then you run it through a process called synthetic data generation, 495 00:25:44,560 --> 00:25:46,679 Speaker 4: and it looks at the documents or look at the 496 00:25:46,720 --> 00:25:49,840 Speaker 4: history chapter. It'll look at the questions and answers, and 497 00:25:49,880 --> 00:25:52,600 Speaker 4: then it'll look to that original document and come up 498 00:25:52,600 --> 00:25:55,000 Speaker 4: with more questions and answers based on the format of 499 00:25:55,040 --> 00:25:57,640 Speaker 4: the questions and answers you made. So you can take 500 00:25:57,760 --> 00:26:01,320 Speaker 4: five questions of answers amplify them into one hundred questions 501 00:26:01,320 --> 00:26:04,199 Speaker 4: and answers, two hundred questions and answers, and it's a 502 00:26:04,280 --> 00:26:07,439 Speaker 4: second model that is making the questions and answers. So 503 00:26:07,440 --> 00:26:10,479 Speaker 4: it's synthetic data generation using an AI model to make 504 00:26:10,560 --> 00:26:13,520 Speaker 4: the questions. We use an open source model to do that. 505 00:26:14,119 --> 00:26:16,760 Speaker 4: So that's the second part. And then the third part 506 00:26:16,840 --> 00:26:19,760 Speaker 4: is we have a multi phase tuning technique to actually 507 00:26:19,920 --> 00:26:23,440 Speaker 4: take the synthetic data and then basically bake it into 508 00:26:23,480 --> 00:26:26,680 Speaker 4: the model. So sort of that's the approach. A general 509 00:26:26,720 --> 00:26:29,439 Speaker 4: philosophy of the approach is using granite because we know 510 00:26:29,480 --> 00:26:32,240 Speaker 4: where the data came from. Another approach is the fact 511 00:26:32,280 --> 00:26:34,640 Speaker 4: that we're using small models that are cheap to run 512 00:26:34,640 --> 00:26:37,199 Speaker 4: inference on. They're small enough that you can tune them 513 00:26:37,240 --> 00:26:40,040 Speaker 4: on laptop hardware. You don't need all the fancy expensive 514 00:26:40,080 --> 00:26:44,280 Speaker 4: GPU menia. You're good. So sort of like a whole system. 515 00:26:44,359 --> 00:26:47,159 Speaker 4: It's like not any one component. But it's sort of 516 00:26:47,280 --> 00:26:49,800 Speaker 4: the approach they took with somewhat novel, and they were 517 00:26:49,880 --> 00:26:52,800 Speaker 4: very excited when they saw the experimental results. There was 518 00:26:52,840 --> 00:26:55,639 Speaker 4: a meeting between red hat and IBM. It was actually 519 00:26:55,640 --> 00:26:57,960 Speaker 4: an IBM research meeting that red hatters were invited to, 520 00:26:58,720 --> 00:27:00,800 Speaker 4: and I think the red Hatter and Voves sort of 521 00:27:00,840 --> 00:27:05,640 Speaker 4: saw the potential, WHOA, we can make models open source finally, 522 00:27:05,760 --> 00:27:09,320 Speaker 4: rather than them just being these endless dead forks, we 523 00:27:09,359 --> 00:27:12,000 Speaker 4: could make it so people could contribute back and collaborate 524 00:27:12,040 --> 00:27:14,320 Speaker 4: around it. So that's when red Hat became interested in 525 00:27:14,359 --> 00:27:17,840 Speaker 4: it and we sort of worked together, and the research 526 00:27:17,880 --> 00:27:20,560 Speaker 4: engineers from IBM Research who came up with the technique, 527 00:27:20,640 --> 00:27:23,239 Speaker 4: and then my team, the software engineers who know how 528 00:27:23,280 --> 00:27:28,240 Speaker 4: to take research code and productize it into actually runnable, 529 00:27:28,280 --> 00:27:33,200 Speaker 4: supportable software, kind of got together. We've been hanging out 530 00:27:33,200 --> 00:27:35,880 Speaker 4: in the Boston office at red Hat and building it out. 531 00:27:36,240 --> 00:27:39,479 Speaker 4: April eighteenth was when we went open source and we 532 00:27:39,520 --> 00:27:41,959 Speaker 4: made all of our repositories with all of the code public, 533 00:27:42,000 --> 00:27:44,280 Speaker 4: and right now we're working towards a product release, so 534 00:27:44,320 --> 00:27:45,280 Speaker 4: a supported product. 535 00:27:45,400 --> 00:27:47,479 Speaker 3: How long did it take you to be convinced of 536 00:27:48,440 --> 00:27:51,720 Speaker 3: the value of this idea? I mean, so people get 537 00:27:51,720 --> 00:27:55,919 Speaker 3: together in this hotel room they're running these experiments over Christmas. 538 00:27:56,160 --> 00:27:58,440 Speaker 3: Are you aware of the experiments as they're running them? 539 00:27:59,080 --> 00:27:59,199 Speaker 2: They? 540 00:27:59,240 --> 00:28:00,760 Speaker 4: Oh, I didn't find out to February. 541 00:28:02,040 --> 00:28:05,560 Speaker 3: They come to you February and they say, MO, can 542 00:28:05,640 --> 00:28:07,480 Speaker 3: you recreate that conversation? 543 00:28:08,520 --> 00:28:12,960 Speaker 4: Well, our CEO, Matt Hicks, and then Jeremy Eater, who's 544 00:28:12,960 --> 00:28:15,640 Speaker 4: one of our distinguished engineers, and Steve Watt, who's a VP, 545 00:28:15,840 --> 00:28:18,360 Speaker 4: were present I think at that meeting. So they kind 546 00:28:18,400 --> 00:28:20,640 Speaker 4: of brought it back to us and said, listen, we've 547 00:28:20,680 --> 00:28:25,080 Speaker 4: invited these IBM research folks to come visit in Boston, 548 00:28:25,840 --> 00:28:28,280 Speaker 4: you know, work with them, like, see, does this have 549 00:28:28,320 --> 00:28:30,560 Speaker 4: any merit? Could we build something from it? And so 550 00:28:30,600 --> 00:28:33,840 Speaker 4: they gave us some presentations. We're very excited. When they 551 00:28:33,840 --> 00:28:37,200 Speaker 4: came to us. It only had support for Mac laptops. 552 00:28:37,800 --> 00:28:39,880 Speaker 4: Of course, you know Red Hat were Linux people, So 553 00:28:39,960 --> 00:28:41,800 Speaker 4: we're like, all right, we've got to fix that. So 554 00:28:41,960 --> 00:28:44,640 Speaker 4: a bunch of the junior engineers around the office kind 555 00:28:44,680 --> 00:28:46,240 Speaker 4: of came and they're like, okay, we're going to build 556 00:28:46,280 --> 00:28:48,400 Speaker 4: Linux support. And they had it within like a couple 557 00:28:48,400 --> 00:28:51,280 Speaker 4: of days. It was crazy because this was just meant 558 00:28:51,320 --> 00:28:53,840 Speaker 4: to be Hey, guys, you know what, these are invited 559 00:28:53,920 --> 00:28:57,440 Speaker 4: guests visiting our office. See what happens. And we ended 560 00:28:57,480 --> 00:29:00,920 Speaker 4: up doing like weeks of hack fe and late night 561 00:29:00,960 --> 00:29:03,600 Speaker 4: pizzas in the conference room and like playing around with 562 00:29:03,640 --> 00:29:06,560 Speaker 4: it and learning and it was it was very fun. 563 00:29:06,640 --> 00:29:07,360 Speaker 4: It's very cool. 564 00:29:07,480 --> 00:29:08,920 Speaker 3: Anyone else do anything like this. 565 00:29:10,320 --> 00:29:12,560 Speaker 4: Is not my understanding that anybody else is doing it, 566 00:29:12,800 --> 00:29:16,360 Speaker 4: yet maybe others will try a lot of the focus 567 00:29:16,400 --> 00:29:19,960 Speaker 4: has been on that pre training phase. But for us, 568 00:29:20,040 --> 00:29:23,200 Speaker 4: again that fine tuning. It's more accessible because you don't 569 00:29:23,760 --> 00:29:26,360 Speaker 4: need all the exotic hardware. It doesn't take months. You 570 00:29:26,400 --> 00:29:28,240 Speaker 4: can do it on a laptop. You can do a 571 00:29:28,280 --> 00:29:30,880 Speaker 4: smoke test version of it in less than an hour. 572 00:29:31,440 --> 00:29:32,560 Speaker 3: What is the word smoke test. 573 00:29:32,760 --> 00:29:35,160 Speaker 4: Smoke test means you're not doing a full fine tuning 574 00:29:35,200 --> 00:29:38,240 Speaker 4: on the model. It's a different tuning process. It's like 575 00:29:38,320 --> 00:29:40,880 Speaker 4: kind of lower quality, so to run on lower grade hardware, 576 00:29:41,040 --> 00:29:42,640 Speaker 4: so you can kind of see them didn't move the 577 00:29:42,680 --> 00:29:44,200 Speaker 4: model or not, but it's not going to give you, 578 00:29:44,200 --> 00:29:46,800 Speaker 4: like the full picture. You need higher end hardware to 579 00:29:46,840 --> 00:29:48,880 Speaker 4: actually do the full thing. So that's what the product 580 00:29:48,880 --> 00:29:51,720 Speaker 4: will enable you to do once it's launched, is you're 581 00:29:51,720 --> 00:29:53,520 Speaker 4: going to need the GPUs, but when you have them, 582 00:29:53,520 --> 00:29:55,160 Speaker 4: will help you make the best usage of them. 583 00:29:55,440 --> 00:29:58,239 Speaker 3: Yeah, yeah, and no, there's all the detail. I want 584 00:29:58,280 --> 00:30:01,320 Speaker 3: to go back to. Sure to run the tests on 585 00:30:01,360 --> 00:30:07,800 Speaker 3: this idea way back when they needed time on the GPUs, 586 00:30:08,240 --> 00:30:12,360 Speaker 3: So this will be the in house IBM and they 587 00:30:12,400 --> 00:30:15,320 Speaker 3: were quiet at Christmas, So how much time would you 588 00:30:15,400 --> 00:30:18,719 Speaker 3: need on the GPUs to kind of get proof of concept? 589 00:30:19,120 --> 00:30:21,480 Speaker 4: Well what happens is and it's sort of like a 590 00:30:21,480 --> 00:30:23,760 Speaker 4: lot of trial and error, right, And there's a lot 591 00:30:23,800 --> 00:30:27,400 Speaker 4: about this stuff that like you come up with the hypothesis, 592 00:30:27,480 --> 00:30:29,440 Speaker 4: you test it out, did it work or not? Okay, 593 00:30:29,560 --> 00:30:31,600 Speaker 4: it's just like you know in the lab, but you know, 594 00:30:31,760 --> 00:30:35,640 Speaker 4: buns and burners and beakers and whatever. So it really depends. 595 00:30:35,680 --> 00:30:39,040 Speaker 4: But it can be hours, it can be days. It 596 00:30:39,080 --> 00:30:41,120 Speaker 4: really depends on what they're trying to do. And then 597 00:30:41,200 --> 00:30:43,560 Speaker 4: sometimes they can cut the time down, you know, with 598 00:30:43,600 --> 00:30:45,240 Speaker 4: the number of GPUs you have. So like I have 599 00:30:45,240 --> 00:30:48,080 Speaker 4: a cluster of agpus, Okay, it might take a day, 600 00:30:48,160 --> 00:30:50,120 Speaker 4: but then if I can get thirty two, I can 601 00:30:50,120 --> 00:30:51,920 Speaker 4: pipeline it and make it go faster and get it 602 00:30:51,920 --> 00:30:53,959 Speaker 4: down to a few hours. So it really depends, you know. 603 00:30:54,040 --> 00:30:57,120 Speaker 4: But it's like everybody's home for the holidays. It's a 604 00:30:57,160 --> 00:30:59,719 Speaker 4: lovely playground to kind of get that stuff going fast. 605 00:31:00,480 --> 00:31:04,040 Speaker 3: Let's jump forward one year. Tell me the status of 606 00:31:04,080 --> 00:31:07,560 Speaker 3: this project, tell me who's using it, tell me how 607 00:31:07,600 --> 00:31:13,600 Speaker 3: big is it. Give me your optimistic but plausible prediction 608 00:31:13,920 --> 00:31:17,640 Speaker 3: about what instruct lab looks like a year from now. 609 00:31:18,560 --> 00:31:21,960 Speaker 4: A year from now, I would like to see kind 610 00:31:21,960 --> 00:31:28,360 Speaker 4: of a vibrant community around not just building knowledge and 611 00:31:28,400 --> 00:31:32,120 Speaker 4: skills into a model, but coming up with better techniques 612 00:31:32,160 --> 00:31:34,720 Speaker 4: and innovation around how we do it. So I'd like 613 00:31:34,760 --> 00:31:37,880 Speaker 4: to see the contributor experience as we grow more and 614 00:31:37,920 --> 00:31:40,640 Speaker 4: more contributors to be refined. So like a year from now, 615 00:31:40,840 --> 00:31:43,960 Speaker 4: Malcolm Gladwell could come impart some of his wisdom into 616 00:31:43,960 --> 00:31:46,320 Speaker 4: the model and it wouldn't be difficult, it wouldn't be 617 00:31:46,320 --> 00:31:49,240 Speaker 4: a big lift. I would love to see the user 618 00:31:49,240 --> 00:31:53,360 Speaker 4: interface tooling for doing that to be more sophisticated. I 619 00:31:53,360 --> 00:31:56,920 Speaker 4: would love to see more people taking this and even 620 00:31:57,040 --> 00:31:59,240 Speaker 4: using it. Maybe you're not sharing it with the community, 621 00:31:59,280 --> 00:32:02,240 Speaker 4: but you're using it for some private usage. Like I'll 622 00:32:02,240 --> 00:32:05,720 Speaker 4: give you an example. I'm in contact with a fellow 623 00:32:05,840 --> 00:32:08,560 Speaker 4: who is doing AI research and he's working with doctors. 624 00:32:08,600 --> 00:32:11,560 Speaker 4: They're GPS in an area of Canada where there's not 625 00:32:11,680 --> 00:32:14,360 Speaker 4: enough GPS for the number of patients, So you know, 626 00:32:14,480 --> 00:32:18,280 Speaker 4: anything you can do to save doctors time to get 627 00:32:18,360 --> 00:32:20,640 Speaker 4: to the next patient. It's like one of the things 628 00:32:20,640 --> 00:32:23,480 Speaker 4: that he has been doing experiments with is can we 629 00:32:23,640 --> 00:32:27,400 Speaker 4: use an open source, licensed model that the doctor can 630 00:32:27,480 --> 00:32:29,440 Speaker 4: run on their laptop so they don't have to worry 631 00:32:29,440 --> 00:32:31,960 Speaker 4: about all of the different privacy rules, Like it's privates 632 00:32:31,960 --> 00:32:36,040 Speaker 4: on the laptop right there, take his live transcription of 633 00:32:36,040 --> 00:32:39,720 Speaker 4: his conversation with the patient, and then convert it automatically 634 00:32:39,760 --> 00:32:42,120 Speaker 4: to a soap format that can be entered in the database. 635 00:32:42,360 --> 00:32:44,959 Speaker 4: Typically this will take a doctor fifteen to twenty minutes 636 00:32:45,000 --> 00:32:48,720 Speaker 4: of paperwork. Why not save them the paperwork at least 637 00:32:48,760 --> 00:32:50,000 Speaker 4: have the model take a stab. 638 00:32:50,200 --> 00:32:52,800 Speaker 3: Does the model then hold on to that information and 639 00:32:52,200 --> 00:32:54,760 Speaker 3: he interacts with the model again when. 640 00:32:55,040 --> 00:32:57,400 Speaker 4: Well, that's the thing not within struct lab. Maybe that 641 00:32:57,440 --> 00:33:00,440 Speaker 4: could be a future development. It doesn't once you're doing it, diference, 642 00:33:01,120 --> 00:33:03,840 Speaker 4: it's not ingesting that what you're saying to it back in. 643 00:33:04,160 --> 00:33:06,400 Speaker 4: It's only the fine tuning phase. So the idea would 644 00:33:06,440 --> 00:33:10,040 Speaker 4: be the doctor could maybe load in past patient data 645 00:33:10,320 --> 00:33:13,000 Speaker 4: as knowledge and then when he's trying to diagnose maybe 646 00:33:13,160 --> 00:33:15,640 Speaker 4: you know what I'm saying. Like, But the main idea 647 00:33:15,720 --> 00:33:18,120 Speaker 4: is somebody might have some private usage. I would love 648 00:33:18,200 --> 00:33:22,400 Speaker 4: to see more usage of this tool to enable people 649 00:33:22,400 --> 00:33:24,720 Speaker 4: who otherwise never would have had access to this type 650 00:33:24,760 --> 00:33:27,520 Speaker 4: of technology who never like you know, a small country 651 00:33:27,600 --> 00:33:31,760 Speaker 4: GP doctors, it doesn't have GPUs. They're not going to 652 00:33:31,840 --> 00:33:34,000 Speaker 4: hire some company to custom build them a model. But 653 00:33:34,040 --> 00:33:35,840 Speaker 4: maybe on the weekend, if he's a techie guy he 654 00:33:35,880 --> 00:33:37,000 Speaker 4: can say with us. 655 00:33:37,160 --> 00:33:39,440 Speaker 3: Well, I mean, the more you talk, the more I'm 656 00:33:39,480 --> 00:33:43,600 Speaker 3: realizing that the simplicity of this model is the killer 657 00:33:43,640 --> 00:33:46,160 Speaker 3: app here. Once you know you can run it on 658 00:33:46,200 --> 00:33:50,080 Speaker 3: a laptop, you have democratized use in a way that's 659 00:33:50,120 --> 00:33:54,360 Speaker 3: inconceivable with some of these other much more complex. But 660 00:33:54,400 --> 00:33:58,000 Speaker 3: that's interesting because one would have thought intuitively that at 661 00:33:58,040 --> 00:34:00,560 Speaker 3: the beginning that the winner is going to be the 662 00:34:00,560 --> 00:34:06,080 Speaker 3: one with the biggest, most complex version, And you're saying, actually, no, 663 00:34:06,280 --> 00:34:11,680 Speaker 3: there's a whole series of uses where being lean and focused, 664 00:34:11,960 --> 00:34:15,800 Speaker 3: focused is actually you know, it enables a whole class 665 00:34:15,800 --> 00:34:19,160 Speaker 3: of uses. Maybe another way of saying this is who 666 00:34:19,200 --> 00:34:21,640 Speaker 3: wouldn't be a potential instruct lab customer. 667 00:34:22,000 --> 00:34:25,160 Speaker 4: We don't know yet. It's so new, like we haven't 668 00:34:25,160 --> 00:34:27,480 Speaker 4: really had enough people experimenting and playing with it and 669 00:34:27,520 --> 00:34:30,160 Speaker 4: finding out all the things yet. But that's that's the 670 00:34:30,200 --> 00:34:32,080 Speaker 4: thing that's so exciting about it. It's like, I can't 671 00:34:32,080 --> 00:34:33,319 Speaker 4: wait to see what people do. 672 00:34:33,760 --> 00:34:35,520 Speaker 3: Is this the most exciting thing you've worked on in 673 00:34:35,560 --> 00:34:36,000 Speaker 3: your career? 674 00:34:36,320 --> 00:34:38,440 Speaker 4: I think so. I think so. 675 00:34:39,040 --> 00:34:42,360 Speaker 3: Yeah, Well, we are reaching the end of our time, 676 00:34:42,880 --> 00:34:46,080 Speaker 3: but before we finished, we can do a little speed round. Sure, 677 00:34:46,560 --> 00:34:50,840 Speaker 3: all right, complete the following sentence. In five years, AI 678 00:34:51,120 --> 00:34:52,400 Speaker 3: will be. 679 00:34:52,520 --> 00:34:56,920 Speaker 4: Boring, it will be integrated, It'll just work, and there 680 00:34:56,960 --> 00:34:59,720 Speaker 4: will be no now with AI thing. It'll just be normal. 681 00:35:01,360 --> 00:35:04,520 Speaker 3: What's the number one thing that people misunderstand about AI? 682 00:35:05,120 --> 00:35:08,640 Speaker 4: It's just matrix algebra. It's just numbers. It's not sentient. 683 00:35:08,880 --> 00:35:12,240 Speaker 4: It's not coming to take us over. It's just numbers. 684 00:35:12,440 --> 00:35:15,479 Speaker 3: You're on this side of You're on the team humanity. Yeah, 685 00:35:15,560 --> 00:35:20,080 Speaker 3: you're one good. What advice would you give yourself ten 686 00:35:20,160 --> 00:35:22,360 Speaker 3: years ago to better prepare for today? 687 00:35:22,960 --> 00:35:26,799 Speaker 4: Learn Python for real. It's a programming language that's extensively 688 00:35:26,880 --> 00:35:29,680 Speaker 4: used in the community. I've always dabbled in it, but 689 00:35:29,840 --> 00:35:31,440 Speaker 4: I wish I had taken it more seriously. 690 00:35:31,680 --> 00:35:33,640 Speaker 3: Yeah, did you say, who had a daughter? 691 00:35:34,200 --> 00:35:35,200 Speaker 4: I have three daughters? 692 00:35:35,280 --> 00:35:38,000 Speaker 3: You have three daughters. I have two. You're if you 693 00:35:38,080 --> 00:35:41,879 Speaker 3: got three year you're you're on your own. What are 694 00:35:41,880 --> 00:35:43,280 Speaker 3: you making them study Python? 695 00:35:44,400 --> 00:35:47,440 Speaker 4: I am actually trying to do that. We're using a 696 00:35:47,480 --> 00:35:50,560 Speaker 4: microbit micro controller tool to do like a custom video 697 00:35:50,600 --> 00:35:53,960 Speaker 4: game controller. They prefer Scratch because it's a visual programming language, 698 00:35:53,960 --> 00:35:55,759 Speaker 4: but it has a Python interface too, and I'm like 699 00:35:55,880 --> 00:35:57,040 Speaker 4: pushing them towards Python. 700 00:35:57,400 --> 00:36:01,680 Speaker 3: Good chat, bock and image generators are the biggest things 701 00:36:01,680 --> 00:36:04,200 Speaker 3: in consumer AI right now. What do you think is 702 00:36:04,239 --> 00:36:06,200 Speaker 3: the next big business application? 703 00:36:07,680 --> 00:36:13,040 Speaker 4: Private models, small models fine tuned on your company's data 704 00:36:13,640 --> 00:36:15,319 Speaker 4: for you to use exclusively. 705 00:36:16,040 --> 00:36:19,400 Speaker 3: Are you using AI in your own personal life these days? 706 00:36:19,600 --> 00:36:21,440 Speaker 4: Honestly, I think a lot of us are using it 707 00:36:21,480 --> 00:36:23,879 Speaker 4: and we don't even realize it. Yeah, I mean, I'm 708 00:36:23,880 --> 00:36:27,840 Speaker 4: a ficiano of foreign languages. There's translation programs that are 709 00:36:27,880 --> 00:36:30,920 Speaker 4: built using machine learning underneath. One of the things I've 710 00:36:30,960 --> 00:36:33,960 Speaker 4: been dabbling with lately is using tech summarizations because I 711 00:36:34,040 --> 00:36:36,719 Speaker 4: tend to be very loquacious in my note taking and 712 00:36:36,760 --> 00:36:39,120 Speaker 4: that is not so useful for other people who would 713 00:36:39,160 --> 00:36:42,080 Speaker 4: just like a paragraph. So that's something I've been experimenting 714 00:36:42,080 --> 00:36:43,919 Speaker 4: with myself just to help my everyday work. 715 00:36:44,040 --> 00:36:48,319 Speaker 3: Yeah. We hear many definitions of open related to technology. 716 00:36:48,880 --> 00:36:52,160 Speaker 3: What's your definition of open and how does it help 717 00:36:52,200 --> 00:36:52,680 Speaker 3: you innovate? 718 00:36:53,040 --> 00:36:58,920 Speaker 4: My definition of open is basically sharing and being vulnerable, 719 00:36:59,040 --> 00:37:02,080 Speaker 4: like not just sharing gonna have a cookie way, but 720 00:37:02,200 --> 00:37:04,520 Speaker 4: in a you know what, I don't actually know how 721 00:37:04,560 --> 00:37:07,360 Speaker 4: this works? Could you help me? And being open to 722 00:37:07,440 --> 00:37:11,080 Speaker 4: being wrong, being open to somebody helping you, and making 723 00:37:11,080 --> 00:37:13,200 Speaker 4: that collaboration work. So it's not just about like the 724 00:37:13,320 --> 00:37:16,680 Speaker 4: artifact or opening, it's your approach, like how you do 725 00:37:16,760 --> 00:37:17,520 Speaker 4: things being open. 726 00:37:17,800 --> 00:37:21,319 Speaker 3: Yeah yeah, well I think that wraps us up. How 727 00:37:21,360 --> 00:37:24,680 Speaker 3: can listeners follow your work and learn more about granted 728 00:37:24,760 --> 00:37:25,719 Speaker 3: and instruct lab. 729 00:37:26,000 --> 00:37:28,600 Speaker 4: Sure, you can visit our project web page at instruct 730 00:37:28,680 --> 00:37:31,600 Speaker 4: lab dot ai, or you can visit our GitHub at 731 00:37:31,680 --> 00:37:34,759 Speaker 4: GitHub dot com slash instruct lab. We have lots of 732 00:37:34,800 --> 00:37:38,280 Speaker 4: instructions on how to get involved in an instruct lab wonderful. 733 00:37:38,600 --> 00:37:44,600 Speaker 3: Thank you so much, Thank you, Malcolm. A big thank 734 00:37:44,680 --> 00:37:48,520 Speaker 3: you to Mow for the engaging discussion on the groundbreaking 735 00:37:48,840 --> 00:37:53,719 Speaker 3: possibilities of instruct lab. We've explored how this platform has 736 00:37:53,760 --> 00:37:58,120 Speaker 3: the potential to revolutionize industries from insurance to entertainment law 737 00:37:58,400 --> 00:38:01,200 Speaker 3: by using an open source community, the approach that makes 738 00:38:01,200 --> 00:38:04,200 Speaker 3: it easier for more people from all backgrounds to fine 739 00:38:04,239 --> 00:38:10,319 Speaker 3: tune models for specific purposes, ultimately making AI more accessible 740 00:38:10,920 --> 00:38:15,600 Speaker 3: and impactful than ever. Looking ahead, the future of AI 741 00:38:15,880 --> 00:38:20,440 Speaker 3: isn't just about technological efficiency. It's about enhancing our everyday 742 00:38:20,480 --> 00:38:25,279 Speaker 3: experiences in ways that were never possible before, like streamlining 743 00:38:25,320 --> 00:38:29,160 Speaker 3: work for doctors to improve the patient experience, or assisting 744 00:38:29,200 --> 00:38:34,680 Speaker 3: insurance agents to improve the claims experience. Instruct Lab is 745 00:38:34,800 --> 00:38:39,319 Speaker 3: paving the way for more open, accessible AI future, one 746 00:38:39,320 --> 00:38:45,840 Speaker 3: that's built on collaboration and humanity. Smart Talks with IBM 747 00:38:45,960 --> 00:38:50,120 Speaker 3: is produced by Matt Romano, Joey Fishground and Jacob Goldstein. 748 00:38:50,520 --> 00:38:54,320 Speaker 3: We're edited by Lydia jen Kott. Our engineers are Sarah 749 00:38:54,320 --> 00:38:59,440 Speaker 3: Bruger and Ben Tolliday. Theme song by Gramoscope Special thanks 750 00:38:59,440 --> 00:39:01,960 Speaker 3: to the Eight Bars and IBM teams, as well as 751 00:39:02,000 --> 00:39:05,520 Speaker 3: the Pushkin marketing team. Smart Talks with IBM is a 752 00:39:05,560 --> 00:39:10,319 Speaker 3: production of Pushkin Industries and Ruby Studio at iHeartMedia. To 753 00:39:10,400 --> 00:39:15,759 Speaker 3: find more Pushkin podcasts, listen on the iHeartRadio app, Apple Podcasts, 754 00:39:16,120 --> 00:39:20,880 Speaker 3: or wherever you listen to podcasts. I'm Malcolm Gladwell. This 755 00:39:21,000 --> 00:39:24,640 Speaker 3: is a paid advertisement from IBM. The conversations on this 756 00:39:24,719 --> 00:39:40,360 Speaker 3: podcast don't necessarily represent IBM's positions, strategies or opinions.