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