1 00:00:00,160 --> 00:00:02,920 Speaker 1: Hey everyone, it's Robert and Joe here. Today we've got 2 00:00:02,920 --> 00:00:05,160 Speaker 1: something a little different to share with you. It's a 3 00:00:05,200 --> 00:00:08,799 Speaker 1: new season of the Smart Talks with IBM podcast series. 4 00:00:09,280 --> 00:00:12,080 Speaker 2: This season, on smart Talks, Malcolm Gladwell and team are 5 00:00:12,119 --> 00:00:15,320 Speaker 2: diving into the transformative world of artificial intelligence with a 6 00:00:15,360 --> 00:00:18,680 Speaker 2: fresh perspective on the concept of open What does open 7 00:00:18,760 --> 00:00:21,960 Speaker 2: really mean in the context of AI. It can mean 8 00:00:22,079 --> 00:00:25,680 Speaker 2: open source code or open data, but it also encompasses 9 00:00:25,760 --> 00:00:30,840 Speaker 2: fostering an ecosystem of ideas, ensuring diverse perspectives are heard, 10 00:00:31,200 --> 00:00:33,599 Speaker 2: and enabling new levels of transparency. 11 00:00:33,920 --> 00:00:37,159 Speaker 1: Join hosts from your favorite pushkin podcasts as they explore 12 00:00:37,159 --> 00:00:41,000 Speaker 1: how openness and AI is reshaping industries, driving innovation, and 13 00:00:41,040 --> 00:00:44,920 Speaker 1: redefining what's possible. You'll hear from industry experts and leaders 14 00:00:44,920 --> 00:00:48,400 Speaker 1: about the implication and possibilities of open AI, and of course, 15 00:00:48,760 --> 00:00:50,960 Speaker 1: Malcolm Gladwell will be there to guide you through the 16 00:00:51,000 --> 00:00:52,760 Speaker 1: season with his unique insights. 17 00:00:53,040 --> 00:00:55,600 Speaker 2: Look out for new episodes of Smart Talks every other 18 00:00:55,680 --> 00:00:59,240 Speaker 2: week on the iHeartRadio app, Apple Podcasts, or wherever you 19 00:00:59,280 --> 00:01:02,560 Speaker 2: get your podcast and learn more at IBM dot com, 20 00:01:02,600 --> 00:01:13,240 Speaker 2: Slash smart Talks. 21 00:01:10,760 --> 00:01:11,720 Speaker 3: Pushkin. 22 00:01:15,959 --> 00:01:19,200 Speaker 4: Hello, Hello, Welcome to Smart Talks with IBM, a podcast 23 00:01:19,200 --> 00:01:25,160 Speaker 4: from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Glabo. This season, 24 00:01:25,400 --> 00:01:28,560 Speaker 4: we're diving back into the world of artificial intelligence, but 25 00:01:28,640 --> 00:01:34,560 Speaker 4: with a focus on the powerful concept of open its possibilities, implications, 26 00:01:34,600 --> 00:01:38,080 Speaker 4: and misconceptions. We'll look at openness from a variety of 27 00:01:38,120 --> 00:01:41,880 Speaker 4: angles and explore how the concept is already reshaping industries, 28 00:01:42,400 --> 00:01:46,240 Speaker 4: ways of doing business and our very notion of what's possible. 29 00:01:47,040 --> 00:01:51,360 Speaker 4: In today's episode, Jacob Goldstein sat down with maryam Ashuri, 30 00:01:51,760 --> 00:01:54,680 Speaker 4: the Director of Product Management and a Head of Product 31 00:01:54,920 --> 00:01:59,160 Speaker 4: for IBM's Watson x dot AI, where she spearheads the 32 00:01:59,160 --> 00:02:04,960 Speaker 4: product strategy and delivery of IBM's watsonex foundation models. She 33 00:02:05,040 --> 00:02:08,480 Speaker 4: is a technologist with more than fifteen years of experience 34 00:02:08,840 --> 00:02:14,320 Speaker 4: developing data driven technologies. The conversation focused on how enterprises 35 00:02:14,520 --> 00:02:19,079 Speaker 4: can use technology to build and deliver greater transparency in AI. 36 00:02:19,639 --> 00:02:24,320 Speaker 4: With Granite. Mariam explained how Grantite can be utilized to 37 00:02:24,360 --> 00:02:29,640 Speaker 4: improve efficiency across various domains. She discussed how these models 38 00:02:29,639 --> 00:02:33,640 Speaker 4: are being used in real world business applications, particularly in 39 00:02:33,720 --> 00:02:38,080 Speaker 4: areas like customer care, where AI can help enable quick, 40 00:02:38,560 --> 00:02:44,320 Speaker 4: accurate responses based on internal company data. Mariam provided a 41 00:02:44,360 --> 00:02:49,120 Speaker 4: fascinating look into how enterprises have moved from mere experimentation 42 00:02:49,360 --> 00:02:55,120 Speaker 4: with generative AI to actual production, navigating challenges such as 43 00:02:55,160 --> 00:03:00,480 Speaker 4: increased latency, cost, and energy consumption. She highlighted how the 44 00:03:00,560 --> 00:03:05,119 Speaker 4: emerging trend of smaller models customized with proprietary data can 45 00:03:05,160 --> 00:03:08,840 Speaker 4: potentially deliver high performance at a fraction of the cost, 46 00:03:09,520 --> 00:03:14,760 Speaker 4: marking a significant shift in how enterprises leverage AI. Whether 47 00:03:14,800 --> 00:03:17,560 Speaker 4: you're an AI enthusiast, we're a business leader looking to 48 00:03:17,639 --> 00:03:22,360 Speaker 4: harness the power of artificial intelligence, this episode is packed 49 00:03:22,880 --> 00:03:26,560 Speaker 4: with valuable insights and forward thinking strategies. 50 00:03:30,880 --> 00:03:32,840 Speaker 3: Let's just start with your background. How did you come 51 00:03:32,880 --> 00:03:34,040 Speaker 3: to work at IBM. 52 00:03:34,720 --> 00:03:38,200 Speaker 5: I join IBM right after I graduated. I have an 53 00:03:38,200 --> 00:03:44,280 Speaker 5: AI background, and throughout the years, I've held many roles 54 00:03:44,320 --> 00:03:49,760 Speaker 5: in design, engineering, development, research, mostly focused on AI application 55 00:03:49,920 --> 00:03:54,040 Speaker 5: development and design. In my current job, I'm the product 56 00:03:54,040 --> 00:03:58,520 Speaker 5: owner for What's the Next DAYI, which is the IBM 57 00:03:58,560 --> 00:04:02,640 Speaker 5: platform for enterprise AI. What excites me about this job, 58 00:04:02,680 --> 00:04:06,480 Speaker 5: I would say, is the technology advancements over the last 59 00:04:06,480 --> 00:04:09,920 Speaker 5: eighteen months in the market. We've been witnessing how GENERATIVELI 60 00:04:10,000 --> 00:04:12,680 Speaker 5: has been changing the market. The way that I see 61 00:04:12,720 --> 00:04:16,360 Speaker 5: that is JENNYI has been perhaps one of the largest 62 00:04:16,400 --> 00:04:19,960 Speaker 5: paradigm shifts when we think about productivity. The same way 63 00:04:20,000 --> 00:04:25,160 Speaker 5: that Internet and personal computers impacted the productivity of workforce, 64 00:04:25,320 --> 00:04:30,200 Speaker 5: now we are witnessing another wave of all those opportunities 65 00:04:30,240 --> 00:04:33,480 Speaker 5: that it can unlock for especially enterprise AI when it 66 00:04:33,520 --> 00:04:37,600 Speaker 5: comes to enhancing the productivity of the workforce and releasing 67 00:04:37,680 --> 00:04:42,320 Speaker 5: some time that can potentially be put into creating more 68 00:04:42,440 --> 00:04:46,840 Speaker 5: value work for enterprise. So that's the major part that 69 00:04:47,000 --> 00:04:50,840 Speaker 5: I picked this team to have an impact on the 70 00:04:50,880 --> 00:04:54,880 Speaker 5: market and the community, but also of course using the 71 00:04:55,520 --> 00:04:58,440 Speaker 5: skills that I gain through all these years through IBM 72 00:04:58,640 --> 00:05:02,599 Speaker 5: to help to establish IBM as the market leader for 73 00:05:02,760 --> 00:05:03,520 Speaker 5: enterprise AI. 74 00:05:04,080 --> 00:05:08,159 Speaker 3: So you talked about JENAI as this sort of generational, 75 00:05:08,320 --> 00:05:13,360 Speaker 3: transformational technological force, and I'm curious just in terms of 76 00:05:13,800 --> 00:05:16,280 Speaker 3: how it's going to come into the world, Like, how 77 00:05:16,279 --> 00:05:20,360 Speaker 3: do you see market adoption of GENAI sort of evolving 78 00:05:20,400 --> 00:05:20,880 Speaker 3: from here? 79 00:05:21,680 --> 00:05:25,239 Speaker 5: Well, last year was the year of excitement about generative AI. 80 00:05:25,400 --> 00:05:28,520 Speaker 5: Most of the companies were experimenting and exploring with GENI. 81 00:05:29,240 --> 00:05:32,800 Speaker 5: We see that energy shifted towards how to best monetize 82 00:05:32,839 --> 00:05:35,719 Speaker 5: that technology. Almost half of the market has moved from 83 00:05:35,960 --> 00:05:41,200 Speaker 5: investigation to pilots. Ten percent has moved to production. When 84 00:05:41,240 --> 00:05:44,880 Speaker 5: you're exploring with this technology, you're looking for a valve factor, 85 00:05:45,160 --> 00:05:48,679 Speaker 5: You're looking for an AHA moment. That's why very large 86 00:05:48,680 --> 00:05:53,279 Speaker 5: general purpose models shine. But as companies move toward production 87 00:05:53,400 --> 00:05:56,320 Speaker 5: and scale, they soon realized the past success is not 88 00:05:56,360 --> 00:06:01,440 Speaker 5: that straightforward. For example, they're larger the model, the larger 89 00:06:01,480 --> 00:06:06,039 Speaker 5: computer resources it requires. That translates to increased latency that's 90 00:06:06,040 --> 00:06:10,440 Speaker 5: your response time. That translates to increased cost. That translates 91 00:06:10,480 --> 00:06:14,039 Speaker 5: to increase carbon food print, and energy consumption. So think 92 00:06:14,040 --> 00:06:17,720 Speaker 5: about that. At the scale of enterprise in production, some 93 00:06:17,800 --> 00:06:19,680 Speaker 5: of them can be a showstopper. 94 00:06:20,040 --> 00:06:20,760 Speaker 6: Because of this. 95 00:06:20,920 --> 00:06:25,240 Speaker 5: Reason, what actually c is emerging in the market is 96 00:06:25,520 --> 00:06:31,520 Speaker 5: instead of focusing on very large general purpose models, coming 97 00:06:31,600 --> 00:06:37,320 Speaker 5: back to very small, trustworthy models that they can customize 98 00:06:37,480 --> 00:06:41,280 Speaker 5: on their own proprietary data that's the data about their customers, 99 00:06:41,320 --> 00:06:45,080 Speaker 5: that the data about their specific domains to create something 100 00:06:45,160 --> 00:06:49,760 Speaker 5: differentiated that is much smaller and delivers the performance that 101 00:06:49,800 --> 00:06:53,120 Speaker 5: they want on a target use case for a fraction 102 00:06:53,200 --> 00:06:53,760 Speaker 5: of the cost. 103 00:06:54,080 --> 00:06:58,000 Speaker 3: Uh huh. So let's talk a little bit more specifically 104 00:06:58,040 --> 00:07:02,440 Speaker 3: about what you're working on. Talk about Granite. First of all, 105 00:07:02,480 --> 00:07:03,799 Speaker 3: tell me what is Granite. 106 00:07:04,400 --> 00:07:09,960 Speaker 5: Granite is our industrial leading family of models, flagship IBM models. 107 00:07:10,680 --> 00:07:14,520 Speaker 5: These are the models that we train from scratch. When 108 00:07:14,600 --> 00:07:18,000 Speaker 5: offered to our platform, we offer indemnification and we stand 109 00:07:18,000 --> 00:07:23,720 Speaker 5: behind them today. It comes in four flavors, language, code, 110 00:07:24,400 --> 00:07:31,880 Speaker 5: time series, and geospecial models. Granite Language series is covering English, Spanish, German, 111 00:07:32,320 --> 00:07:37,120 Speaker 5: Portuguese and Japanese. We have a combination of commercial and 112 00:07:37,360 --> 00:07:41,320 Speaker 5: open source language models on Granite. For example, we recently 113 00:07:41,520 --> 00:07:46,680 Speaker 5: released the Granite seven B language model, small powerful English model. 114 00:07:47,400 --> 00:07:50,720 Speaker 5: On the code front, our models are state of the 115 00:07:50,840 --> 00:07:55,120 Speaker 5: art models ranging from three billion to thirty four billion parameters. 116 00:07:55,760 --> 00:08:00,960 Speaker 5: These are very powerful models that performs or outperforms in 117 00:08:00,960 --> 00:08:04,800 Speaker 5: some cases the popular open source models in their weight class. 118 00:08:04,840 --> 00:08:06,200 Speaker 5: So very powerful models. 119 00:08:06,400 --> 00:08:09,080 Speaker 3: So I get the idea a big picture about these models, 120 00:08:09,120 --> 00:08:10,800 Speaker 3: but it would be helpful to just get a sense 121 00:08:10,840 --> 00:08:12,960 Speaker 3: specifically of what they're doing, Like, can you give me 122 00:08:13,000 --> 00:08:16,640 Speaker 3: any specific examples of how these models are being used 123 00:08:17,440 --> 00:08:19,720 Speaker 3: in businesses in the real world right now? 124 00:08:20,880 --> 00:08:24,119 Speaker 5: Well, the top use cases for generative AI are really 125 00:08:24,240 --> 00:08:31,120 Speaker 5: content generation, summarization, information extraction. Perhaps the most popular use 126 00:08:31,160 --> 00:08:34,840 Speaker 5: case that we are seeing in enterprise is content grounded 127 00:08:34,920 --> 00:08:39,160 Speaker 5: question and answering. So using these models as a base 128 00:08:39,440 --> 00:08:42,320 Speaker 5: to connect them to a body of information let's say, 129 00:08:42,360 --> 00:08:46,680 Speaker 5: their policies, their documents that is internal to the enterprise, 130 00:08:46,960 --> 00:08:51,160 Speaker 5: and get the model to provide answers based on that question. 131 00:08:51,559 --> 00:08:55,520 Speaker 5: One example of that is for customer agents customer care, 132 00:08:55,920 --> 00:09:00,520 Speaker 5: when a customer is asking a question. Previously, the agent 133 00:09:00,559 --> 00:09:04,080 Speaker 5: that responds to the customer had to answer the question 134 00:09:04,200 --> 00:09:06,959 Speaker 5: and if they don't know the answer escalated to the product. 135 00:09:07,120 --> 00:09:10,600 Speaker 5: Especially is keeping people on hold on the line to 136 00:09:10,760 --> 00:09:13,720 Speaker 5: go figure out the answer for that and then come back. 137 00:09:13,800 --> 00:09:16,160 Speaker 5: You can think of the time it takes to resolve 138 00:09:16,200 --> 00:09:19,880 Speaker 5: an issue. But now we llms, we have an opportunity 139 00:09:19,920 --> 00:09:23,960 Speaker 5: to automatically retrieve the information based on the internal documents 140 00:09:24,000 --> 00:09:27,000 Speaker 5: of the company, formulate an answer, show it to the 141 00:09:27,080 --> 00:09:30,400 Speaker 5: human agent, and then if they verify with the sources 142 00:09:30,440 --> 00:09:33,160 Speaker 5: of varies coming from, they can just translate it directly 143 00:09:33,200 --> 00:09:33,880 Speaker 5: to the customer. 144 00:09:34,800 --> 00:09:35,360 Speaker 6: This is a. 145 00:09:35,360 --> 00:09:39,200 Speaker 5: Very simple example of how it's impacting the customer care. 146 00:09:39,679 --> 00:09:44,160 Speaker 3: So one big theme of this season is this idea 147 00:09:44,160 --> 00:09:47,560 Speaker 3: of open and one of the things that's interesting to 148 00:09:47,679 --> 00:09:51,480 Speaker 3: me about the work you're doing is you are using 149 00:09:51,559 --> 00:09:55,200 Speaker 3: not only granted this model IBM developed, but you're also 150 00:09:55,360 --> 00:09:58,960 Speaker 3: using third party models right from other places. So tell 151 00:09:59,000 --> 00:10:01,000 Speaker 3: me about that work and how that is sort of 152 00:10:01,040 --> 00:10:05,120 Speaker 3: fitting into your kind of real world typically enterprise Jenai work. 153 00:10:05,760 --> 00:10:08,600 Speaker 5: When it comes to a model strategy, our strategy is 154 00:10:08,800 --> 00:10:13,160 Speaker 5: really focused on two pillars, multimodel and multi deployment. It 155 00:10:13,280 --> 00:10:16,559 Speaker 5: means that we don't believe one single model rules all 156 00:10:16,559 --> 00:10:19,000 Speaker 5: the use cases. And I think at this point the 157 00:10:19,040 --> 00:10:22,520 Speaker 5: market has also realized the enterprise markets in average today 158 00:10:22,559 --> 00:10:27,040 Speaker 5: are using five to ten different models for different use cases. 159 00:10:27,200 --> 00:10:28,199 Speaker 3: Oh interesting. 160 00:10:28,520 --> 00:10:30,800 Speaker 5: So in our portfolio, if you look into what's on 161 00:10:30,840 --> 00:10:33,640 Speaker 5: Extra DAYI today, we are offering a large sets of 162 00:10:33,880 --> 00:10:36,760 Speaker 5: high performing, state of the art models coming from open 163 00:10:36,800 --> 00:10:41,000 Speaker 5: source commercial models that we are bringing through our partners 164 00:10:41,320 --> 00:10:45,200 Speaker 5: and also IBM developed models. In addition to all of these, 165 00:10:45,400 --> 00:10:48,400 Speaker 5: we also have an option for bring your own model 166 00:10:48,720 --> 00:10:51,680 Speaker 5: from outside the platform. Let's say you have a custom 167 00:10:51,760 --> 00:10:54,440 Speaker 5: model that you made it yourself, you can bring it 168 00:10:54,480 --> 00:10:59,360 Speaker 5: to the platform and really helping the customers to navigate 169 00:10:59,400 --> 00:11:03,040 Speaker 5: through aid range of models and pick the right model 170 00:11:03,320 --> 00:11:06,960 Speaker 5: for their target use case. Throughout that we've been heavily 171 00:11:07,000 --> 00:11:10,200 Speaker 5: working with our partners, and you know, this is the 172 00:11:10,240 --> 00:11:13,720 Speaker 5: market that is evolving rapidly. We've been at the forefront 173 00:11:13,720 --> 00:11:15,880 Speaker 5: of a spit to delivery. One example that I like 174 00:11:15,960 --> 00:11:21,400 Speaker 5: to highlight is recently Metal released Lama four or five billion, 175 00:11:21,720 --> 00:11:24,240 Speaker 5: such a powerful model. On the same day that it 176 00:11:24,440 --> 00:11:27,400 Speaker 5: was released to the market, we made it available in 177 00:11:27,480 --> 00:11:30,520 Speaker 5: our platform to our customers the same day. And not 178 00:11:30,600 --> 00:11:33,040 Speaker 5: only we delivered it on the same day. We are 179 00:11:33,040 --> 00:11:37,520 Speaker 5: offering competitive pricing but also for flexibility in where to deploy. 180 00:11:37,640 --> 00:11:40,559 Speaker 5: So we are giving an option to enterprise to deploy 181 00:11:40,640 --> 00:11:44,880 Speaker 5: these models on the platform of dage choice, either multi 182 00:11:44,880 --> 00:11:48,760 Speaker 5: cloud it can be gcpaws as youre IBM cloud, or 183 00:11:48,800 --> 00:11:54,160 Speaker 5: on premises. The same for mistrall Ai. Mistrall Ai recently 184 00:11:54,320 --> 00:11:57,320 Speaker 5: released the model misroll launch too on the same day 185 00:11:57,600 --> 00:12:00,600 Speaker 5: we delivered that through the platform. That's an example of 186 00:12:00,640 --> 00:12:04,960 Speaker 5: a commercial model. Lama as open source, but MS large 187 00:12:04,960 --> 00:12:08,000 Speaker 5: two is a commercial model that we made available through 188 00:12:08,040 --> 00:12:08,800 Speaker 5: the platform. 189 00:12:09,320 --> 00:12:14,920 Speaker 3: Great, So I want to talk about enterprise grade foundation models. 190 00:12:15,640 --> 00:12:18,520 Speaker 3: Just to get into it briefly, what's a foundation model. 191 00:12:19,000 --> 00:12:22,719 Speaker 5: People associate foundation models with a large language model, but 192 00:12:22,840 --> 00:12:26,160 Speaker 5: large language models are really a subset of foundation models. 193 00:12:26,320 --> 00:12:30,240 Speaker 5: Large language models are focused on language, but foundation models 194 00:12:30,280 --> 00:12:34,120 Speaker 5: can be code generators, can be focused on time series 195 00:12:34,120 --> 00:12:36,720 Speaker 5: model we talked about, they can be images, it can 196 00:12:36,760 --> 00:12:41,200 Speaker 5: be jew special models. So foundation model, as the term 197 00:12:41,320 --> 00:12:46,439 Speaker 5: suggests that your foundations to create a series of subsequent 198 00:12:46,640 --> 00:12:51,000 Speaker 5: models that can be customized for a downstream use case. 199 00:12:51,040 --> 00:12:54,439 Speaker 5: And that's why they are calling them foundation models. Lm 200 00:12:54,480 --> 00:12:56,480 Speaker 5: ME is a good example of that as a subset 201 00:12:56,520 --> 00:13:00,240 Speaker 5: for language that you can further customize on your space. 202 00:13:00,480 --> 00:13:04,040 Speaker 6: Data to get the model to do other works. 203 00:13:04,080 --> 00:13:07,280 Speaker 5: So the core of these foundation models, they are basically 204 00:13:07,920 --> 00:13:11,680 Speaker 5: trained on an ab third amount of data data sets 205 00:13:11,880 --> 00:13:15,120 Speaker 5: that most of the institutions today are sourcing them from 206 00:13:15,120 --> 00:13:18,080 Speaker 5: the internet. So you can imagine what can potentially go 207 00:13:18,120 --> 00:13:20,880 Speaker 5: to those models and then it comes to the enterprise 208 00:13:21,000 --> 00:13:25,200 Speaker 5: and they start using it. So for us also, when 209 00:13:25,240 --> 00:13:29,440 Speaker 5: we started looking into in particular, it was triggered by 210 00:13:29,720 --> 00:13:33,880 Speaker 5: customers asking us to provide client protections on these models, 211 00:13:33,920 --> 00:13:36,440 Speaker 5: and we started thinking about, let's look into how the 212 00:13:36,520 --> 00:13:40,120 Speaker 5: models are trained and if you are comfortable of fering 213 00:13:40,200 --> 00:13:43,680 Speaker 5: client protections on the models that are available in the market. 214 00:13:43,800 --> 00:13:45,199 Speaker 6: And guess what, for a. 215 00:13:45,200 --> 00:13:49,280 Speaker 5: Majority of these models there is absolutely no visibility into 216 00:13:49,360 --> 00:13:52,760 Speaker 5: what data vent into those models, not much transparency into 217 00:13:52,840 --> 00:13:56,880 Speaker 5: how the model trains, and the responsibility lies on you 218 00:13:56,960 --> 00:13:59,240 Speaker 5: as the customers we start using those models. 219 00:13:59,240 --> 00:14:03,080 Speaker 3: So just to be that is presenting like potential risk, 220 00:14:03,200 --> 00:14:06,480 Speaker 3: real potential risk to a company that is using these models, 221 00:14:06,720 --> 00:14:07,120 Speaker 3: it is. 222 00:14:07,240 --> 00:14:10,720 Speaker 5: It is a potential risk in particular for the customers 223 00:14:10,760 --> 00:14:15,319 Speaker 5: in highly regulated industries. So what we did for Granite 224 00:14:15,880 --> 00:14:19,120 Speaker 5: was when we started training these models from scratch, Basically 225 00:14:19,160 --> 00:14:22,280 Speaker 5: we went to the corpus of data that was available 226 00:14:22,320 --> 00:14:25,440 Speaker 5: to us. So, for example, the very first version of 227 00:14:25,800 --> 00:14:29,800 Speaker 5: Granite was exposed to twenty percent of its data from 228 00:14:29,880 --> 00:14:33,680 Speaker 5: finance and legal because we have a lot of financial 229 00:14:33,680 --> 00:14:38,120 Speaker 5: institutions as our clients. We worked directly with our IBM 230 00:14:38,160 --> 00:14:43,080 Speaker 5: research to identify detectors for harmful information like haytyp use 231 00:14:43,160 --> 00:14:44,600 Speaker 5: and profanity detectors. 232 00:14:45,160 --> 00:14:47,480 Speaker 3: Okay, so we're talking about Granted, we're talking about this 233 00:14:47,680 --> 00:14:51,000 Speaker 3: set of models IBM has developed. Let's talk about using 234 00:14:51,000 --> 00:14:55,840 Speaker 3: Granite on Watson X compared to downloading open source models, 235 00:14:55,960 --> 00:14:56,880 Speaker 3: Like how do those differ? 236 00:14:57,520 --> 00:15:01,160 Speaker 5: By using Granite and what's on ex you get two things. 237 00:15:01,520 --> 00:15:05,280 Speaker 5: The first one is the client protection and thementification that 238 00:15:05,320 --> 00:15:07,520 Speaker 5: we talked about. You get that if the model is 239 00:15:07,560 --> 00:15:08,960 Speaker 5: consumed through our platform. 240 00:15:09,440 --> 00:15:10,440 Speaker 6: And the second. 241 00:15:10,120 --> 00:15:14,600 Speaker 5: One is really the ecosystem of platform capabilities that we 242 00:15:14,640 --> 00:15:17,760 Speaker 5: are offering to help you create value on top of 243 00:15:17,800 --> 00:15:21,960 Speaker 5: those data. So for example, bringing your data to customize 244 00:15:22,000 --> 00:15:25,720 Speaker 5: granted for your own specific use case. But also one 245 00:15:25,720 --> 00:15:28,520 Speaker 5: thing that I like to highlight in particular is the 246 00:15:28,560 --> 00:15:31,800 Speaker 5: AI governance. So when you get one of these pre 247 00:15:31,880 --> 00:15:35,040 Speaker 5: train models, you put it in front of your own users. 248 00:15:35,840 --> 00:15:39,600 Speaker 5: Through the input and instructions that the user provides for 249 00:15:39,760 --> 00:15:44,080 Speaker 5: the model, they can notdge the model to potentially create 250 00:15:44,400 --> 00:15:48,000 Speaker 5: undesired behavior and change the behavior of the model. And 251 00:15:48,040 --> 00:15:52,120 Speaker 5: because of this is extremely important to automatically document the 252 00:15:52,240 --> 00:15:56,400 Speaker 5: lineage of who touched the model at one point, so 253 00:15:56,480 --> 00:15:58,880 Speaker 5: if something happens, you can trace it back and see 254 00:15:58,920 --> 00:16:02,920 Speaker 5: where it's coming from. And that's what's an extra governance 255 00:16:03,040 --> 00:16:07,160 Speaker 5: is offering automatically documenting the lineage. When you use the 256 00:16:07,200 --> 00:16:10,200 Speaker 5: granite within the platform, you get all of those you 257 00:16:10,240 --> 00:16:13,320 Speaker 5: can have the end to end governance, you can have 258 00:16:13,640 --> 00:16:17,720 Speaker 5: access to all these scalable deployment opportunities that is available 259 00:16:17,760 --> 00:16:20,560 Speaker 5: for you, like to allow you deploy them on the 260 00:16:20,600 --> 00:16:23,320 Speaker 5: platform of your choice that we talked about, either multiple 261 00:16:23,960 --> 00:16:27,440 Speaker 5: cloud or on prem and it also helps you to 262 00:16:27,520 --> 00:16:32,080 Speaker 5: have access to avoid range of model customizations, approaches, prompt tuning, 263 00:16:32,160 --> 00:16:36,080 Speaker 5: fine tuning, retrival augmented generations agents. There is a series 264 00:16:36,120 --> 00:16:38,960 Speaker 5: of them available to use an apply to your model. 265 00:16:39,760 --> 00:16:44,240 Speaker 4: This distinction between large language models and foundation models is 266 00:16:44,280 --> 00:16:48,760 Speaker 4: eye opening. Mariam emphasized that foundation models can be tailored 267 00:16:48,760 --> 00:16:53,760 Speaker 4: to specific tasks, but with that versatility comes a significant 268 00:16:53,840 --> 00:16:58,200 Speaker 4: challenge the lack of transparency and how these models are trained. 269 00:16:59,040 --> 00:17:05,280 Speaker 4: This composed a real especially in highly regulated industries like finance. Essentially, 270 00:17:05,359 --> 00:17:10,160 Speaker 4: by using Granite and watsonex together, enterprises get powerful and 271 00:17:10,200 --> 00:17:11,560 Speaker 4: customizable tools. 272 00:17:12,760 --> 00:17:14,960 Speaker 3: So let's talk about the future a little bit. What 273 00:17:15,040 --> 00:17:17,120 Speaker 3: do you think are some of the big developments were 274 00:17:17,200 --> 00:17:20,040 Speaker 3: likely to see in the realm of AI models? 275 00:17:20,400 --> 00:17:21,280 Speaker 6: Very good question. 276 00:17:22,040 --> 00:17:26,199 Speaker 5: I feel like the generative AI of the past was 277 00:17:26,400 --> 00:17:30,800 Speaker 5: powered by large language models. The generative AI of the 278 00:17:30,840 --> 00:17:35,439 Speaker 5: future is going to reason, plan, act and reflect. 279 00:17:35,960 --> 00:17:39,359 Speaker 3: Huh, and so I mean in the context of Granite 280 00:17:39,560 --> 00:17:43,000 Speaker 3: in particular, like, what are we likely to see both 281 00:17:43,160 --> 00:17:45,040 Speaker 3: you know, in the near term and in the sort 282 00:17:45,080 --> 00:17:46,320 Speaker 3: of medium to long term. 283 00:17:46,920 --> 00:17:51,239 Speaker 5: There are multiple elements to implement an agentic workflow that 284 00:17:51,280 --> 00:17:54,800 Speaker 5: I just mentioned. One element of that is the LLM 285 00:17:54,880 --> 00:17:59,000 Speaker 5: itself to be able to do the planning and reasoning 286 00:17:59,080 --> 00:18:03,439 Speaker 5: and acting and doing something that we call tool calling. 287 00:18:03,840 --> 00:18:07,439 Speaker 5: So basically, a series of tools are available to the model. 288 00:18:08,000 --> 00:18:10,480 Speaker 5: You ask the model to call those and. 289 00:18:10,400 --> 00:18:10,880 Speaker 6: Make a call. 290 00:18:11,040 --> 00:18:14,199 Speaker 5: For example, we can say, hey, Granted, what is the 291 00:18:14,200 --> 00:18:19,960 Speaker 5: weather like where Jacob lives. It's connect to web search API, 292 00:18:20,520 --> 00:18:23,280 Speaker 5: look up your location. Then it's going to connect to 293 00:18:23,720 --> 00:18:28,080 Speaker 5: weather API, calculate the weather and come back and formulate 294 00:18:28,119 --> 00:18:31,680 Speaker 5: an answer and respond to that. So during this process, 295 00:18:32,240 --> 00:18:34,720 Speaker 5: it first has to plan the task of how to 296 00:18:34,760 --> 00:18:37,639 Speaker 5: answer that question, look into what are the tools that 297 00:18:37,680 --> 00:18:40,360 Speaker 5: are available to it, and call them, and that's an 298 00:18:40,359 --> 00:18:43,040 Speaker 5: ability of the model to do that. What we did 299 00:18:43,080 --> 00:18:47,159 Speaker 5: with Granted was we expanded the Granite capabilities to be 300 00:18:47,240 --> 00:18:50,880 Speaker 5: able to do function calling. So for example, today we 301 00:18:51,240 --> 00:18:54,320 Speaker 5: have an open source granted to an eb function calling 302 00:18:54,400 --> 00:18:57,320 Speaker 5: that is available on hugging face to try on and 303 00:18:57,400 --> 00:18:59,960 Speaker 5: you can grab the model and the model has capability 304 00:19:00,080 --> 00:19:03,359 Speaker 5: to do the tool callings. I'm anticipating that in the 305 00:19:03,400 --> 00:19:07,639 Speaker 5: near future the planning and reasoning and acting and reflecting 306 00:19:07,680 --> 00:19:10,760 Speaker 5: capabilities of the large language models are going to continue 307 00:19:10,800 --> 00:19:11,280 Speaker 5: to evolve. 308 00:19:12,680 --> 00:19:16,720 Speaker 3: So thinking now from the point of view of buyers 309 00:19:16,760 --> 00:19:20,400 Speaker 3: and users of AIS, really people who are listening from 310 00:19:20,400 --> 00:19:26,840 Speaker 3: that perspective, as people are evaluating AI tools and solutions, 311 00:19:27,480 --> 00:19:30,359 Speaker 3: what is the most important thing they should be thinking about? 312 00:19:30,440 --> 00:19:32,879 Speaker 3: How do you think about kind of that process? 313 00:19:33,920 --> 00:19:37,240 Speaker 5: I think they should always start with the area at 314 00:19:37,320 --> 00:19:41,400 Speaker 5: which they think it would benefit from AI, and then 315 00:19:41,720 --> 00:19:45,720 Speaker 5: within that area, look into what data they have available 316 00:19:45,880 --> 00:19:50,080 Speaker 5: to potentially fit into those AI service architects do they 317 00:19:50,080 --> 00:19:53,639 Speaker 5: have access to quality data? And the second question that 318 00:19:53,680 --> 00:19:55,560 Speaker 5: they have to ask themselves is do I have a 319 00:19:55,600 --> 00:19:59,520 Speaker 5: trusted partner that can supply what I need to be 320 00:19:59,560 --> 00:20:03,320 Speaker 5: able to implement AI. That can be a collection of 321 00:20:03,359 --> 00:20:05,920 Speaker 5: the foundation models that you're going to need, that can 322 00:20:05,960 --> 00:20:10,000 Speaker 5: be a collection of the platform capabilities that the trusted 323 00:20:10,040 --> 00:20:13,399 Speaker 5: partner can offer you to implement such a thing. The 324 00:20:13,480 --> 00:20:18,600 Speaker 5: third thing is go and evaluate the regulations. Does regulation 325 00:20:19,000 --> 00:20:23,240 Speaker 5: allow you to apploy AI to the specific area that 326 00:20:23,760 --> 00:20:27,159 Speaker 5: you are investigating and you're targeting for AI? And the 327 00:20:27,280 --> 00:20:30,520 Speaker 5: last part, but not least, is back to the principles 328 00:20:30,560 --> 00:20:34,200 Speaker 5: of design, thinking, what is the problem in that area? 329 00:20:34,680 --> 00:20:39,120 Speaker 5: I'm solving with AI, and if AI is even appropriate, 330 00:20:39,640 --> 00:20:41,639 Speaker 5: because we want to make sure that you use AI 331 00:20:41,800 --> 00:20:44,680 Speaker 5: not just because it's a cool, hot toy in the market, 332 00:20:44,720 --> 00:20:48,600 Speaker 5: but you are convinced that it can significantly enhance the 333 00:20:49,119 --> 00:20:52,960 Speaker 5: user experience of your customers in that area. And once 334 00:20:53,000 --> 00:20:55,520 Speaker 5: you have an answer to those all these four questions, 335 00:20:55,600 --> 00:20:58,840 Speaker 5: then maybe you have a good candidates to start applying AI. 336 00:21:00,720 --> 00:21:03,880 Speaker 3: What about from the side of project managers who are 337 00:21:04,040 --> 00:21:07,400 Speaker 3: trying to just keep up with how fast things are changing, 338 00:21:07,440 --> 00:21:11,359 Speaker 3: how fast innovation is happening, Like, what advice would you 339 00:21:11,440 --> 00:21:12,280 Speaker 3: give those people? 340 00:21:12,880 --> 00:21:17,159 Speaker 5: My advice would be focused on agility. This is a 341 00:21:17,160 --> 00:21:20,879 Speaker 5: market that is evolving rapidly and the winners of the 342 00:21:20,960 --> 00:21:24,439 Speaker 5: market would be those that are able to take advantage 343 00:21:24,440 --> 00:21:27,680 Speaker 5: of the best the market can offer at any point 344 00:21:27,680 --> 00:21:30,680 Speaker 5: of time. So in order to do that, they need 345 00:21:30,720 --> 00:21:39,000 Speaker 5: to be open to experimentation, continuous learning, and to rapidly 346 00:21:39,320 --> 00:21:40,880 Speaker 5: adopting the new ideas. 347 00:21:42,080 --> 00:21:45,520 Speaker 3: And when you think about the future and GENAI, is 348 00:21:45,600 --> 00:21:49,480 Speaker 3: there a particular, say problem that you are most excited 349 00:21:49,520 --> 00:21:50,040 Speaker 3: to solve. 350 00:21:50,720 --> 00:21:53,600 Speaker 5: I think that would be productivity. If you look into 351 00:21:53,640 --> 00:21:57,040 Speaker 5: the stats that are out there, there are surveys that 352 00:21:57,320 --> 00:22:01,000 Speaker 5: confirm that sixty to seventy persons of the time of 353 00:22:01,000 --> 00:22:07,000 Speaker 5: our employees can be potentially enhanced to the productivity gains 354 00:22:07,000 --> 00:22:10,440 Speaker 5: of generative I For example, I personally myself use my 355 00:22:10,520 --> 00:22:14,040 Speaker 5: product for content generation a lot, so the time that 356 00:22:14,080 --> 00:22:19,080 Speaker 5: it frees up can be potentially put into generating a 357 00:22:19,160 --> 00:22:23,359 Speaker 5: higher value work. And because of that, I'm super excited 358 00:22:23,480 --> 00:22:27,919 Speaker 5: with all the opportunities that it represents for enterprises to 359 00:22:28,359 --> 00:22:31,480 Speaker 5: go and dedicate the time of the employees to higher 360 00:22:31,560 --> 00:22:32,640 Speaker 5: value items. 361 00:22:32,880 --> 00:22:37,399 Speaker 3: Great. Okay, a couple of Granite specific questions. So what 362 00:22:37,520 --> 00:22:40,280 Speaker 3: are like the key things you want the world to 363 00:22:40,400 --> 00:22:41,760 Speaker 3: know about Granite. 364 00:22:42,320 --> 00:22:48,320 Speaker 5: Granite is open, trusted, and targeted. Two ways to think 365 00:22:48,359 --> 00:22:52,840 Speaker 5: about openness. One open as open weights it's available for 366 00:22:52,880 --> 00:22:57,120 Speaker 5: public to download, and the second one is open as 367 00:22:57,200 --> 00:23:02,080 Speaker 5: in there is less restrictions on how the customers can 368 00:23:02,200 --> 00:23:05,280 Speaker 5: legally use these models for a range of use cases. 369 00:23:05,400 --> 00:23:08,760 Speaker 5: We have released Grantite open source models on their Apache 370 00:23:08,960 --> 00:23:12,760 Speaker 5: license that is enabling a large range of use cases. 371 00:23:13,240 --> 00:23:16,399 Speaker 5: The second one was trusted. We talked about that like 372 00:23:16,520 --> 00:23:20,720 Speaker 5: it's rooted in the trustworthy governance process that we established 373 00:23:20,760 --> 00:23:24,760 Speaker 5: thereund how we are training these models and the responsibility 374 00:23:24,800 --> 00:23:27,280 Speaker 5: that we take for these models, and the third one 375 00:23:27,320 --> 00:23:31,800 Speaker 5: is targeted, targeted for enterprise. We talked about like exposing 376 00:23:31,800 --> 00:23:36,159 Speaker 5: Granted to enterprise data or the domain specific Granted some 377 00:23:36,240 --> 00:23:39,600 Speaker 5: of them like Cobalt Java Translation that is targeting to 378 00:23:39,760 --> 00:23:44,840 Speaker 5: solve the specific enterprise needs. And that's granite, so open, trusted, 379 00:23:44,920 --> 00:23:45,560 Speaker 5: and targeted. 380 00:23:46,280 --> 00:23:48,080 Speaker 3: So there are a lot of models out in the 381 00:23:48,119 --> 00:23:51,240 Speaker 3: world all of a sudden, right, it's a crowded market. 382 00:23:51,840 --> 00:23:54,679 Speaker 3: Where does granted fit in that universe? What is the 383 00:23:54,720 --> 00:23:55,600 Speaker 3: market for granted? 384 00:23:56,600 --> 00:24:00,480 Speaker 5: We talked about the enterprise market shifting away from very 385 00:24:00,600 --> 00:24:05,560 Speaker 5: large general purpose models to target a smaller models, and 386 00:24:05,680 --> 00:24:10,400 Speaker 5: Granted is a small model that enterprise can pick up 387 00:24:10,680 --> 00:24:15,600 Speaker 5: and customize on their proprietary data to create something that 388 00:24:15,720 --> 00:24:19,720 Speaker 5: is differentiated for a target use case. So Granted is 389 00:24:19,760 --> 00:24:24,520 Speaker 5: well suited as a small, domain specific business, ready tailored 390 00:24:24,520 --> 00:24:29,800 Speaker 5: for business and trained on enterprise data to solve enterprise questions. 391 00:24:30,200 --> 00:24:33,080 Speaker 3: You mentioned small as one of the things that granted 392 00:24:33,200 --> 00:24:38,240 Speaker 3: is why is that useful in some contexts for enterprise 393 00:24:38,320 --> 00:24:39,360 Speaker 3: for businesses. 394 00:24:40,160 --> 00:24:44,640 Speaker 5: The larger the model, the larger computer resources it requires, 395 00:24:45,320 --> 00:24:50,560 Speaker 5: it translates to increased latency that's your response time. It 396 00:24:50,600 --> 00:24:57,240 Speaker 5: translates to increased cost and in translates to increased carbon 397 00:24:57,240 --> 00:25:01,760 Speaker 5: footprint and energy consumption. So at this case of enterprise transactions, 398 00:25:01,800 --> 00:25:04,160 Speaker 5: when you move to production and you want to scale, 399 00:25:05,000 --> 00:25:10,159 Speaker 5: some of these challenges can be multiple times stronger. Like 400 00:25:10,280 --> 00:25:13,560 Speaker 5: costs can add up, the energy consumption can be a 401 00:25:13,640 --> 00:25:17,240 Speaker 5: serious thing, and the latency is depending on the application, 402 00:25:17,920 --> 00:25:24,200 Speaker 5: can be a showstopper and blocker because for longer, larger models, 403 00:25:24,200 --> 00:25:27,720 Speaker 5: more powerful models, it just takes the way longer time 404 00:25:27,920 --> 00:25:29,800 Speaker 5: to process and calculate the output. 405 00:25:29,880 --> 00:25:33,719 Speaker 3: For you, we are going to finish up with a 406 00:25:33,760 --> 00:25:38,240 Speaker 3: speed round and I want you to just answer with 407 00:25:38,280 --> 00:25:40,720 Speaker 3: the first thing that comes to mind. Don't overthink this, Okay, 408 00:25:41,000 --> 00:25:43,920 Speaker 3: complete this sentence. In five years, AI. 409 00:25:43,800 --> 00:25:46,240 Speaker 6: Will be invisible. 410 00:25:46,560 --> 00:25:48,840 Speaker 3: Ah, I like that. What do you mean by that? 411 00:25:49,320 --> 00:25:49,639 Speaker 6: Today? 412 00:25:49,720 --> 00:25:54,320 Speaker 5: AI is everywhere. But if you ask my kids at home, 413 00:25:55,240 --> 00:25:57,760 Speaker 5: they know AI. But if you say very like how 414 00:25:57,800 --> 00:26:00,600 Speaker 5: do you use AI, they don't know the answer because 415 00:26:01,160 --> 00:26:04,720 Speaker 5: it's so blended in their life that they don't feel 416 00:26:04,720 --> 00:26:08,520 Speaker 5: like it's something that they are using. They are getting 417 00:26:08,600 --> 00:26:11,120 Speaker 5: used to that. So when I think of next generation 418 00:26:11,800 --> 00:26:15,840 Speaker 5: and the years to come, that generation is so used 419 00:26:15,840 --> 00:26:19,520 Speaker 5: to AI being part of their life that they feel 420 00:26:19,520 --> 00:26:22,600 Speaker 5: like it's just there. That's one, and the second one 421 00:26:22,680 --> 00:26:25,760 Speaker 5: is the simplicity of interaction with AI that you don't 422 00:26:25,800 --> 00:26:28,920 Speaker 5: feel like you're interacting with the system. It's just there, 423 00:26:29,000 --> 00:26:32,159 Speaker 5: like you talk to AI. Everything is automated. So I 424 00:26:32,200 --> 00:26:37,000 Speaker 5: would say the simplicity and being blended to solve the 425 00:26:37,160 --> 00:26:41,520 Speaker 5: right problems is the part that I'm referring to as invisible. 426 00:26:41,640 --> 00:26:44,960 Speaker 5: Like Internet is everywhere and it's invisible. But we used 427 00:26:44,960 --> 00:26:48,119 Speaker 5: to dial in, like you remember the dialing zone to 428 00:26:48,320 --> 00:26:49,359 Speaker 5: connect the Internet. 429 00:26:49,920 --> 00:26:53,160 Speaker 6: It's gone. The Internet is completely invisible today. 430 00:26:53,000 --> 00:26:55,720 Speaker 3: Right, Like we used to talk about logging on, right, 431 00:26:55,760 --> 00:26:58,800 Speaker 3: and you don't log on anymore because you're always logged on. 432 00:26:59,359 --> 00:27:00,720 Speaker 6: Yeah, always connected. 433 00:27:00,840 --> 00:27:05,399 Speaker 3: Yeah. What's the number one thing that people misunderstand about AI? 434 00:27:06,000 --> 00:27:10,800 Speaker 5: AI is anivitable but should not be feared. 435 00:27:11,800 --> 00:27:14,560 Speaker 3: What advice would you give yourself ten years ago to 436 00:27:14,800 --> 00:27:16,680 Speaker 3: better prepare you for today? 437 00:27:17,640 --> 00:27:21,160 Speaker 5: I would say, develop a broad range of skills. Even 438 00:27:21,400 --> 00:27:25,120 Speaker 5: if you think they will not help you today, they 439 00:27:25,160 --> 00:27:26,679 Speaker 5: may be valuable in the future. 440 00:27:27,280 --> 00:27:30,439 Speaker 3: So on the consumer side, right now, we hear a 441 00:27:30,480 --> 00:27:35,800 Speaker 3: lot about chatbots and image generators. But on the business side, 442 00:27:35,840 --> 00:27:38,359 Speaker 3: what do you think is the next big business application? 443 00:27:38,920 --> 00:27:41,480 Speaker 6: AI? Influencers generating content. 444 00:27:41,920 --> 00:27:44,440 Speaker 3: Huh how do you use AI in your day to 445 00:27:44,520 --> 00:27:45,240 Speaker 3: day life today? 446 00:27:46,119 --> 00:27:50,159 Speaker 5: One simple example is LinkedIn posts. I love it to 447 00:27:50,320 --> 00:27:52,640 Speaker 5: just go to my product. I'll give you an example, 448 00:27:52,680 --> 00:27:55,600 Speaker 5: which is my favorite one. Lama three point one four 449 00:27:55,720 --> 00:27:59,160 Speaker 5: or five b the post that I announced on LinkedIn 450 00:27:59,359 --> 00:28:02,360 Speaker 5: on Hey, IBM is releasing the model on the same 451 00:28:02,440 --> 00:28:05,520 Speaker 5: day it was generated by lamatory point one four or 452 00:28:05,520 --> 00:28:08,680 Speaker 5: five billion. So using the same model to post the 453 00:28:09,160 --> 00:28:12,400 Speaker 5: generate the announcement note very elegant. 454 00:28:13,000 --> 00:28:14,560 Speaker 3: Is there anything else I should ask you? 455 00:28:15,000 --> 00:28:17,919 Speaker 5: Oh, we didn't talk about instruct lab. So when you 456 00:28:18,040 --> 00:28:20,919 Speaker 5: grab a model, you start from the model, but you 457 00:28:21,040 --> 00:28:26,040 Speaker 5: need to then customize it on your proprietary data to 458 00:28:26,080 --> 00:28:29,720 Speaker 5: create value on top of that. So instruct lab is 459 00:28:29,760 --> 00:28:36,240 Speaker 5: giving you a method based on open source contributions to 460 00:28:36,280 --> 00:28:42,520 Speaker 5: collectively contribute to improve the base model. So if you're 461 00:28:42,560 --> 00:28:48,720 Speaker 5: an enterprise, you can leverage your internal employees to collectively 462 00:28:48,840 --> 00:28:52,600 Speaker 5: all contribute to improve the model. And I'll give you 463 00:28:52,600 --> 00:28:54,840 Speaker 5: an example of why it matters. Like if you go 464 00:28:54,880 --> 00:28:57,800 Speaker 5: to hugging Pace today and look for Lama, there are 465 00:28:58,000 --> 00:29:01,719 Speaker 5: about fifty thousand different lama us coming up. And the 466 00:29:01,760 --> 00:29:04,760 Speaker 5: reason is because there is no way to contribute to 467 00:29:04,800 --> 00:29:07,800 Speaker 5: the base model. If you're a developer, you have to 468 00:29:07,840 --> 00:29:09,960 Speaker 5: make a colon of the copy of the model and 469 00:29:10,080 --> 00:29:13,360 Speaker 5: finding need for your own purpose. We figure the method 470 00:29:13,400 --> 00:29:17,320 Speaker 5: that we call instruct lab to be able to collectively 471 00:29:17,560 --> 00:29:20,960 Speaker 5: collect all that information and contribute to the base model 472 00:29:21,000 --> 00:29:21,440 Speaker 5: and enhance. 473 00:29:21,880 --> 00:29:22,960 Speaker 6: So that's instruct lab. 474 00:29:24,080 --> 00:29:26,520 Speaker 5: I just wanted to highlight the value of being open 475 00:29:27,680 --> 00:29:30,280 Speaker 5: because that's another topic that has been emerging in the 476 00:29:30,360 --> 00:29:33,960 Speaker 5: market over the past eighteen months. In particular, I believe 477 00:29:34,000 --> 00:29:37,120 Speaker 5: the future of AI is open, and we've been seeing 478 00:29:37,200 --> 00:29:42,720 Speaker 5: how the open source markets has been changing, how the 479 00:29:42,800 --> 00:29:47,080 Speaker 5: models are accessible to a wider audience, and good things 480 00:29:47,120 --> 00:29:51,120 Speaker 5: typically happen when you make technology pieces accessible to a 481 00:29:51,160 --> 00:29:55,400 Speaker 5: broader range of community to stress test that, and that's 482 00:29:55,440 --> 00:29:58,080 Speaker 5: the direction that we've been adopting with granted, and I 483 00:29:58,120 --> 00:30:00,160 Speaker 5: felt like that's really the adoption that the market kit 484 00:30:00,280 --> 00:30:02,600 Speaker 5: is gonna emerge to moving forward. 485 00:30:02,800 --> 00:30:07,160 Speaker 3: Yeah, there's this interesting I think, maybe naively unintuitive, but 486 00:30:07,240 --> 00:30:09,640 Speaker 3: it makes sense once you think about it, thing that 487 00:30:10,400 --> 00:30:13,440 Speaker 3: open source things are safer. You might naively think, oh no, 488 00:30:13,600 --> 00:30:15,320 Speaker 3: put it in a box so nobody can see it 489 00:30:15,360 --> 00:30:17,520 Speaker 3: and that'll be safer, But like it turns out of 490 00:30:17,520 --> 00:30:19,920 Speaker 3: the world. If you let everybody poke at it, the 491 00:30:19,960 --> 00:30:22,480 Speaker 3: world will find the vulnerabilities for you and you can 492 00:30:22,600 --> 00:30:23,400 Speaker 3: fix them. Right. 493 00:30:23,800 --> 00:30:25,240 Speaker 6: That's exactly what's going to happen. 494 00:30:25,480 --> 00:30:28,840 Speaker 3: Yeah, great, it was lovely to talk with you. Thank 495 00:30:28,880 --> 00:30:29,880 Speaker 3: you so much for your time. 496 00:30:30,360 --> 00:30:34,200 Speaker 6: The same here, thanks Jacob, and. 497 00:30:34,120 --> 00:30:36,880 Speaker 4: That wraps up this episode. A huge thanks to Mariam 498 00:30:36,920 --> 00:30:40,600 Speaker 4: and Jacob. Today's conversation open my eyes as to how 499 00:30:40,720 --> 00:30:45,360 Speaker 4: open technology and AI are intersecting to create more transparent 500 00:30:45,680 --> 00:30:50,000 Speaker 4: and efficient systems for enterprises. From the power of smaller, 501 00:30:50,080 --> 00:30:53,160 Speaker 4: more targeted models like granted to the importance of trust 502 00:30:53,280 --> 00:30:58,120 Speaker 4: and governance in AI, these developments are reshaping how businesses 503 00:30:58,200 --> 00:31:02,200 Speaker 4: operate at their core. As we continue to unpack the 504 00:31:02,280 --> 00:31:07,480 Speaker 4: complexities of artificial intelligence, it's clear that openness, whether in data, 505 00:31:07,920 --> 00:31:12,160 Speaker 4: technology or collaboration, is not just a concept, but a 506 00:31:12,240 --> 00:31:18,640 Speaker 4: driving force that can unlock new possibilities. Smart Talks with 507 00:31:18,680 --> 00:31:21,960 Speaker 4: IBM is produced by Matt Romano, Joey fish Ground, Amy 508 00:31:22,000 --> 00:31:26,400 Speaker 4: Gains McQuaid, and Jacob Goldstein. We're edited by Lydia Jane 509 00:31:26,440 --> 00:31:30,600 Speaker 4: kott Or. Engineers are Sarah Brugerier and Ben Tolliday. Theme 510 00:31:30,680 --> 00:31:33,440 Speaker 4: song by Gramoscope special thanks to the eight Bar and 511 00:31:33,520 --> 00:31:38,000 Speaker 4: IBM teams, as well as the Pushkin marketing team. Smart 512 00:31:38,000 --> 00:31:40,640 Speaker 4: Talks with IBM is a production of Pushkin Industries and 513 00:31:40,760 --> 00:31:45,480 Speaker 4: Ruby Studio at iHeartMedia. To find more Pushkin podcasts, listen 514 00:31:45,560 --> 00:31:49,120 Speaker 4: on the iHeartRadio app, Apple Podcasts, or wherever you listen 515 00:31:49,440 --> 00:31:55,960 Speaker 4: to podcasts. I'm Malcolm Glauwell. This is a paid advertisement 516 00:31:56,200 --> 00:32:00,560 Speaker 4: from IBM. The conversations on this podcast don't necessarily represent 517 00:32:00,840 --> 00:32:10,920 Speaker 4: IBM's positions, strategies, or opinions.