1 00:00:04,440 --> 00:00:12,600 Speaker 1: Welcome to Tech Stuff, a production from iHeartRadio. Today, we 2 00:00:12,680 --> 00:00:15,640 Speaker 1: are witnessed to one of those rare moments in history, 3 00:00:16,000 --> 00:00:19,239 Speaker 1: the rise of an innovative technology with the potential to 4 00:00:19,360 --> 00:00:24,080 Speaker 1: radically transform business and society forever. That technology, of course, 5 00:00:24,560 --> 00:00:28,120 Speaker 1: is artificial intelligence, and it's the central focus for this 6 00:00:28,280 --> 00:00:32,320 Speaker 1: new season of Smart Talks with IBM. Join hosts from 7 00:00:32,320 --> 00:00:36,040 Speaker 1: your favorite Pushkin podcasts as they talk with industry experts 8 00:00:36,080 --> 00:00:39,640 Speaker 1: and leaders to explore how businesses can integrate AI into 9 00:00:39,720 --> 00:00:43,040 Speaker 1: their workflows and help drive real change in this new 10 00:00:43,120 --> 00:00:46,800 Speaker 1: era of AI, and of course, host Malcolm Gladwell will 11 00:00:46,840 --> 00:00:49,120 Speaker 1: be there to guide you through the season and throw 12 00:00:49,240 --> 00:00:52,120 Speaker 1: in his two cents as well. Look out for new 13 00:00:52,159 --> 00:00:55,040 Speaker 1: episodes of Smart Talks with IBM every other week on 14 00:00:55,080 --> 00:00:59,320 Speaker 1: the iHeartRadio app, Apple Podcasts, wherever you get your podcasts, 15 00:00:59,520 --> 00:01:03,760 Speaker 1: and learn more at IBM dot com slash smart Talks. 16 00:01:04,840 --> 00:01:08,560 Speaker 2: Hello, Hello, Welcome to Smart Talks with IBM, a podcast 17 00:01:08,560 --> 00:01:14,959 Speaker 2: from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. This season, 18 00:01:15,160 --> 00:01:19,760 Speaker 2: we're continuing our conversation with new creators visionaries who are 19 00:01:19,840 --> 00:01:23,880 Speaker 2: creatively applying technology in business to drive change, but with 20 00:01:23,920 --> 00:01:28,760 Speaker 2: a focus on the transformative power of artificial intelligence and 21 00:01:28,840 --> 00:01:31,840 Speaker 2: what it means to leverage AI as a game changing 22 00:01:31,920 --> 00:01:36,399 Speaker 2: multiplier for your business. Our guest today is Jeff Boutier, 23 00:01:36,840 --> 00:01:40,360 Speaker 2: head of Product and Growth at hugging Face, the leading 24 00:01:40,480 --> 00:01:45,640 Speaker 2: open source and open science artificial intelligence platform. An engineer 25 00:01:45,680 --> 00:01:49,560 Speaker 2: by background, he has a self professed obsession with the 26 00:01:49,600 --> 00:01:54,160 Speaker 2: business of technology. Recently, IBM and hugging Face announced a 27 00:01:54,240 --> 00:01:58,880 Speaker 2: collaboration bringing together hugging faces repositories of open source AI 28 00:01:58,960 --> 00:02:03,800 Speaker 2: models with IBM's Watson X platform. It's a move that 29 00:02:03,840 --> 00:02:08,239 Speaker 2: gives businesses even more access to AI while staying true 30 00:02:08,280 --> 00:02:13,920 Speaker 2: to IBM's long standing philosophy of supporting open source technology. 31 00:02:14,840 --> 00:02:18,720 Speaker 2: With open source, businesses can build better AI models that 32 00:02:18,800 --> 00:02:23,320 Speaker 2: suit their specific needs using their own proprietary data while 33 00:02:23,360 --> 00:02:28,440 Speaker 2: browsing a ready catalog of pre trained models. In today's episode, 34 00:02:28,639 --> 00:02:31,760 Speaker 2: you'll hear why open source is so crucial to the 35 00:02:31,800 --> 00:02:36,400 Speaker 2: advancement of AI, how IBM's Watson X interacts with open 36 00:02:36,440 --> 00:02:40,919 Speaker 2: source AI, and Jeff's thoughts on why this singular omnipotent 37 00:02:41,160 --> 00:02:45,240 Speaker 2: AI model is a myth. Jeff spoke with Tim Harford, 38 00:02:45,440 --> 00:02:49,680 Speaker 2: host of the Pushkin podcast Cautionary Tales, a longtime columnist 39 00:02:49,760 --> 00:02:53,239 Speaker 2: at the Financial Times, where he writes the Undercover Economist. 40 00:02:53,520 --> 00:02:57,320 Speaker 2: Tim is also a BBC broadcaster with his show More 41 00:02:57,480 --> 00:03:01,160 Speaker 2: or Less. Okay, let's get to the interview. 42 00:03:08,600 --> 00:03:11,480 Speaker 3: I am a Jeff Boudier and I'm a product director 43 00:03:11,560 --> 00:03:12,600 Speaker 3: at hugging. 44 00:03:12,280 --> 00:03:16,800 Speaker 4: Face, So I'm immediately intrigue. Hugging Face. Is this a 45 00:03:16,880 --> 00:03:18,840 Speaker 4: reference to the Alien movie or something else? 46 00:03:20,200 --> 00:03:24,040 Speaker 3: It is not, and it may be not obvious to 47 00:03:24,160 --> 00:03:27,800 Speaker 3: a listener, but hugging Face is the name of that 48 00:03:27,960 --> 00:03:30,799 Speaker 3: cute emoji, you know, the one that's smiling with his 49 00:03:30,960 --> 00:03:34,120 Speaker 3: two hands extended like that to give you a big hug. 50 00:03:34,360 --> 00:03:37,640 Speaker 3: That's hugging Face. So basically we name the company after 51 00:03:37,760 --> 00:03:39,120 Speaker 3: an emoji. 52 00:03:40,000 --> 00:03:42,400 Speaker 4: And it is I saw your website and it is 53 00:03:42,440 --> 00:03:45,480 Speaker 4: a very friendly emoji. So that's that's nice. So tell 54 00:03:45,560 --> 00:03:47,880 Speaker 4: us a little bit about hugging Face and about what 55 00:03:47,920 --> 00:03:48,320 Speaker 4: you do that. 56 00:03:48,960 --> 00:03:53,440 Speaker 3: Of course, hugging Face is the leading open platform for 57 00:03:53,800 --> 00:03:59,000 Speaker 3: AI builders, and it's the place that's all the AI 58 00:03:59,080 --> 00:04:04,600 Speaker 3: researchers use to share their work, their new AI models 59 00:04:04,880 --> 00:04:09,600 Speaker 3: and collaborate around them. It's the place where the data 60 00:04:09,840 --> 00:04:15,080 Speaker 3: scientists go and find those pre train models and access 61 00:04:15,160 --> 00:04:19,000 Speaker 3: them and use them and work with them, and increasingly 62 00:04:19,040 --> 00:04:23,360 Speaker 3: it's the place where developers are coming to turn all 63 00:04:23,400 --> 00:04:28,480 Speaker 3: of these AI models and data sets into their own applications, 64 00:04:28,520 --> 00:04:29,680 Speaker 3: their own features. 65 00:04:30,360 --> 00:04:33,039 Speaker 4: So it's like the I don't know, the Facebook group 66 00:04:33,120 --> 00:04:36,000 Speaker 4: or the Reddit or the Twitter for people who are 67 00:04:36,040 --> 00:04:40,280 Speaker 4: interested in particularly generative language AI, or all kinds of 68 00:04:40,360 --> 00:04:41,920 Speaker 4: artificial intelligence. 69 00:04:42,160 --> 00:04:46,400 Speaker 3: All kinds of AI really, and of course generative AIS 70 00:04:46,600 --> 00:04:51,320 Speaker 3: this new wave that has caught the world by storm. 71 00:04:51,680 --> 00:04:55,360 Speaker 3: But on Hiking Face you can find any kind of model, 72 00:04:55,680 --> 00:04:59,560 Speaker 3: the new sort of transformers models to do anything from 73 00:05:00,000 --> 00:05:04,080 Speaker 3: translation or if you wanted to transcribe what I'm saying 74 00:05:04,120 --> 00:05:07,560 Speaker 3: into text, then you would use a transformer model. If 75 00:05:07,560 --> 00:05:10,880 Speaker 3: you wanted to then take that text and make a summary, 76 00:05:11,320 --> 00:05:15,000 Speaker 3: that would be another transformer model. If you wanted to 77 00:05:15,360 --> 00:05:19,400 Speaker 3: create a nice little thumbnail for this podcast by typeing 78 00:05:19,440 --> 00:05:23,280 Speaker 3: a sentence, that would be another type of model. So 79 00:05:23,360 --> 00:05:26,480 Speaker 3: all these models you can find. There's actually three hundred 80 00:05:26,640 --> 00:05:31,119 Speaker 3: thousands that are free and publicly accessible. You can find 81 00:05:31,160 --> 00:05:34,840 Speaker 3: them on our website at Hikingphase dot co and use 82 00:05:34,920 --> 00:05:37,520 Speaker 3: them using our open source libraries. 83 00:05:38,360 --> 00:05:41,160 Speaker 4: And so this is this is fascinating. So there are 84 00:05:41,160 --> 00:05:44,520 Speaker 4: three hundred thousand models. Now when you say model, I'm 85 00:05:44,560 --> 00:05:46,400 Speaker 4: thinking in my head, oh, it's kind of like a 86 00:05:47,360 --> 00:05:50,080 Speaker 4: computer program. There were three hundred thousand computer programs. Is 87 00:05:50,680 --> 00:05:52,359 Speaker 4: that roughly right or it not? 88 00:05:52,440 --> 00:05:57,839 Speaker 3: Really, it's a general idea. A model is a giant 89 00:05:59,680 --> 00:06:05,120 Speaker 3: set of numbers that are working together to sift through 90 00:06:05,760 --> 00:06:08,960 Speaker 3: some inputs that you're going to give it. So think 91 00:06:09,000 --> 00:06:13,480 Speaker 3: of it of a big black box filled with numbers, 92 00:06:14,440 --> 00:06:19,240 Speaker 3: and you give it as an input, maybe some text, 93 00:06:19,960 --> 00:06:23,880 Speaker 3: maybe a prompt, so you're asking, you're giving an instruction 94 00:06:24,120 --> 00:06:26,719 Speaker 3: to the model, or maybe you give it an image 95 00:06:26,800 --> 00:06:31,240 Speaker 3: as an input, and then it will sift through that 96 00:06:31,400 --> 00:06:35,400 Speaker 3: information thanks to all of these numbers, which we call 97 00:06:35,440 --> 00:06:39,880 Speaker 3: in the field parameters, and it will produce an output. 98 00:06:40,480 --> 00:06:43,039 Speaker 3: So when I told you, hey, we can transcribe this 99 00:06:43,279 --> 00:06:47,200 Speaker 3: conversation into text, the input would have been the conversation 100 00:06:47,800 --> 00:06:50,440 Speaker 3: in an audio file, and then the output would have 101 00:06:50,480 --> 00:06:53,479 Speaker 3: been the text of the transcription. If you want to 102 00:06:53,560 --> 00:06:57,599 Speaker 3: create a thumbnail for this podcast episode, then the input 103 00:06:57,640 --> 00:07:00,520 Speaker 3: would be what we call the prompt, which is really 104 00:07:00,520 --> 00:07:05,040 Speaker 3: a text description like a Frenchman in San Francisco talking 105 00:07:05,080 --> 00:07:10,720 Speaker 3: about machine learning, and the output would be completely original image. 106 00:07:11,280 --> 00:07:14,600 Speaker 3: So that's how I think about what an AI model is, 107 00:07:15,080 --> 00:07:19,200 Speaker 3: and I think what we're starting to realize is that 108 00:07:20,080 --> 00:07:24,280 Speaker 3: this is becoming the new way of building technology in 109 00:07:24,320 --> 00:07:28,320 Speaker 3: the world. It has been for the field of dealing 110 00:07:28,440 --> 00:07:32,400 Speaker 3: understanding generating text for quite some time, but now it's 111 00:07:32,520 --> 00:07:36,440 Speaker 3: sort of moving across every field of technology. We have 112 00:07:36,840 --> 00:07:40,960 Speaker 3: models to create images, as I say, but also to 113 00:07:41,120 --> 00:07:46,600 Speaker 3: generate new proteins to make predictions on numerical data. So 114 00:07:46,640 --> 00:07:51,160 Speaker 3: every kind of field of machine learning is now using 115 00:07:52,600 --> 00:07:56,480 Speaker 3: this new type of models. But what's interesting is that 116 00:07:57,080 --> 00:08:00,720 Speaker 3: if you're, say a product manager at a company, and 117 00:08:00,760 --> 00:08:03,840 Speaker 3: you say, hey, I want to build a feature that 118 00:08:03,960 --> 00:08:07,320 Speaker 3: does this. A few years ago, the approach would have 119 00:08:07,360 --> 00:08:11,360 Speaker 3: been to ask a software developer to write a thousand 120 00:08:11,400 --> 00:08:14,840 Speaker 3: lines of code in order to build a prototype. And 121 00:08:14,920 --> 00:08:18,360 Speaker 3: the new way of doing things today is to go 122 00:08:18,520 --> 00:08:23,040 Speaker 3: look for an off the shelf pre train model that 123 00:08:23,080 --> 00:08:27,200 Speaker 3: does a pretty good job at solving exactly that problem, 124 00:08:27,320 --> 00:08:30,400 Speaker 3: so you can create a prototype of that feature fast. 125 00:08:30,440 --> 00:08:33,000 Speaker 3: So it's a new approach of building tech. 126 00:08:33,200 --> 00:08:36,320 Speaker 4: I'm not a programmer, but I'm aware that there was 127 00:08:36,520 --> 00:08:39,080 Speaker 4: this idea of open source code, and now we have 128 00:08:39,160 --> 00:08:42,120 Speaker 4: open source models. So what does it mean for something 129 00:08:42,120 --> 00:08:43,040 Speaker 4: to be open source. 130 00:08:43,640 --> 00:08:49,400 Speaker 3: Open source AI actually means a lot of different specific things. 131 00:08:50,080 --> 00:08:54,280 Speaker 3: It's the open source implementation of the model. So if 132 00:08:54,320 --> 00:08:58,600 Speaker 3: you use the Hugging Phase transformers library to use a model, 133 00:08:58,640 --> 00:09:03,000 Speaker 3: you're using an open source code library to use that model. 134 00:09:03,080 --> 00:09:06,320 Speaker 4: Just to end up on the transformers. These are these 135 00:09:06,400 --> 00:09:09,640 Speaker 4: kind of ways of turning a picture of a dog 136 00:09:09,760 --> 00:09:12,440 Speaker 4: into a text output that says, hey, this is a 137 00:09:12,440 --> 00:09:15,079 Speaker 4: picture of a dog, or this is a French text 138 00:09:15,080 --> 00:09:17,920 Speaker 4: and with the transformers helping you turn it into English text, 139 00:09:18,000 --> 00:09:19,880 Speaker 4: or it's doing all of these things that you've been describing. 140 00:09:19,960 --> 00:09:23,920 Speaker 4: That's the transformer is the kind of the engine at 141 00:09:23,920 --> 00:09:24,760 Speaker 4: the heart of that. 142 00:09:25,559 --> 00:09:29,960 Speaker 3: Yes, exactly. And we call them transformers because they correspond 143 00:09:30,000 --> 00:09:33,920 Speaker 3: to this new way of building machine learning models that 144 00:09:34,080 --> 00:09:38,800 Speaker 3: was introduced by Google actually with a very important paper 145 00:09:39,120 --> 00:09:41,920 Speaker 3: called Attention is All You Need and that was published 146 00:09:41,920 --> 00:09:46,440 Speaker 3: in twenty seventeen by researchers out of Google Deep Mind. 147 00:09:47,400 --> 00:09:50,680 Speaker 4: Well that's just six years so new. 148 00:09:51,960 --> 00:09:55,240 Speaker 3: It is very new, and ever since the piece of 149 00:09:55,480 --> 00:10:00,920 Speaker 3: innovation of like new model architectures has real really accelerated. 150 00:10:01,240 --> 00:10:06,000 Speaker 3: But it really started from this inflection point that came 151 00:10:06,120 --> 00:10:10,400 Speaker 3: from this paper and its implementation in what is now 152 00:10:10,440 --> 00:10:16,240 Speaker 3: called Transformer models, the transformer that has conquered every area 153 00:10:16,360 --> 00:10:18,080 Speaker 3: of machine learning since. 154 00:10:18,280 --> 00:10:21,840 Speaker 4: Okay, so say turned up. So you've got this library 155 00:10:21,840 --> 00:10:26,120 Speaker 4: of Transformer models and that open source, and that means 156 00:10:26,280 --> 00:10:28,480 Speaker 4: that means what anyone can use them for free, or 157 00:10:29,240 --> 00:10:31,320 Speaker 4: that anybody can implement them for free. What does it mean? 158 00:10:32,840 --> 00:10:35,800 Speaker 3: So again, there's lots that go into it, but the 159 00:10:35,840 --> 00:10:40,240 Speaker 3: most important thing is for the model itself to be 160 00:10:40,480 --> 00:10:44,600 Speaker 3: available so that a data scientists or an engineer can 161 00:10:45,000 --> 00:10:49,400 Speaker 3: download them and use them. And also there are a 162 00:10:49,400 --> 00:10:54,240 Speaker 3: lot of considerations about how you make them accessible, and 163 00:10:54,280 --> 00:10:58,240 Speaker 3: a very important one is whether or not you give 164 00:10:58,480 --> 00:11:03,520 Speaker 3: access to the training data, all the information that went 165 00:11:03,679 --> 00:11:07,920 Speaker 3: into training that model and teaching it to do what 166 00:11:08,720 --> 00:11:09,640 Speaker 3: it's trained to do. 167 00:11:09,800 --> 00:11:12,800 Speaker 4: So I might have fed millions of words into a 168 00:11:12,920 --> 00:11:16,040 Speaker 4: into a language transformer, or I might have fed millions 169 00:11:16,040 --> 00:11:18,640 Speaker 4: of photographs into a into a picture transformer. 170 00:11:18,720 --> 00:11:22,160 Speaker 3: Yeah, yes, and now it's trillions and that and the 171 00:11:22,520 --> 00:11:26,160 Speaker 3: accessibility of that training data is very very important. 172 00:11:27,160 --> 00:11:32,960 Speaker 4: What's the relationship between the hugging face libraries and GitHub, which, 173 00:11:34,080 --> 00:11:38,360 Speaker 4: if I understand GitHub correctly, it's this the repository of 174 00:11:38,400 --> 00:11:42,360 Speaker 4: open source code lots and lots of lines of code 175 00:11:42,360 --> 00:11:47,280 Speaker 4: and routines and programs that are shared and updated and tracked, 176 00:11:47,320 --> 00:11:50,480 Speaker 4: and they're all available on GitHub, which sounds similar to 177 00:11:50,520 --> 00:11:52,959 Speaker 4: what you're doing with hugging face for AI. So what 178 00:11:52,960 --> 00:11:55,600 Speaker 4: what what is the interaction or the relationship there? 179 00:11:56,200 --> 00:11:58,640 Speaker 3: Yeah, I think you nailed it on the head there. 180 00:11:58,679 --> 00:12:02,839 Speaker 3: So hugging phase is to AI what GitHub is to code, right, 181 00:12:02,840 --> 00:12:08,959 Speaker 3: It's this central platform where AI builders can go find 182 00:12:09,440 --> 00:12:14,720 Speaker 3: and collaborate around AI artifacts, which are models and data sets. 183 00:12:14,760 --> 00:12:18,719 Speaker 3: So it's quite different than software, but we play this 184 00:12:18,840 --> 00:12:23,079 Speaker 3: central role in the community to share and collaborate and 185 00:12:24,080 --> 00:12:28,880 Speaker 3: access all of those artifacts for AI, like GitHub offers 186 00:12:28,880 --> 00:12:29,839 Speaker 3: for code. 187 00:12:30,679 --> 00:12:33,600 Speaker 4: And that community must be incredibly important. I mean, the 188 00:12:33,640 --> 00:12:36,240 Speaker 4: open source is nothing if you don't have a community 189 00:12:36,280 --> 00:12:38,640 Speaker 4: of people working on it. So how have you been 190 00:12:38,679 --> 00:12:41,800 Speaker 4: able to foster and nurture that community. 191 00:12:42,400 --> 00:12:45,760 Speaker 3: Well, I think it goes to the origins of the 192 00:12:45,840 --> 00:12:49,960 Speaker 3: transformer model and hugging and face role into that. So 193 00:12:50,600 --> 00:12:55,160 Speaker 3: when the first sort of open model came out, it 194 00:12:55,280 --> 00:12:58,440 Speaker 3: was called Bird and it came out of Google. The 195 00:12:58,480 --> 00:13:02,720 Speaker 3: only way you could would access it was to use 196 00:13:02,920 --> 00:13:07,360 Speaker 3: a tool called TensorFlow. But it happened that most of 197 00:13:07,400 --> 00:13:12,840 Speaker 3: the AI community was using a different tool called PyTorch, 198 00:13:13,960 --> 00:13:18,920 Speaker 3: and something that Hugging Face did is to make that 199 00:13:19,000 --> 00:13:25,480 Speaker 3: new model Bert accessible to all PyTorch user and they 200 00:13:25,480 --> 00:13:28,680 Speaker 3: did it in open source. It was a project called 201 00:13:29,200 --> 00:13:32,720 Speaker 3: Bert's pre Trained PyTorch or bird pitworch pre trained. 202 00:13:33,240 --> 00:13:35,360 Speaker 4: So this is like being able to play my Zelda 203 00:13:35,400 --> 00:13:39,440 Speaker 4: game on an Xbox or a PlayStation, right or am 204 00:13:39,480 --> 00:13:41,120 Speaker 4: I not really understanding what's going on? 205 00:13:41,559 --> 00:13:43,920 Speaker 3: No, That's exactly what it is. And the thing is 206 00:13:44,120 --> 00:13:48,080 Speaker 3: everybody was using the game Boy and so it became 207 00:13:48,440 --> 00:13:53,200 Speaker 3: a very popular and from there the community sort of 208 00:13:53,280 --> 00:13:56,839 Speaker 3: gathered to make all the other models that were then 209 00:13:56,960 --> 00:14:01,360 Speaker 3: published by AI researchers available with that library, which was 210 00:14:01,440 --> 00:14:07,000 Speaker 3: quickly renamed from bird bretrain Bytorch into Transformers to welcome 211 00:14:07,120 --> 00:14:12,280 Speaker 3: like all of these different new models, and today that's 212 00:14:12,440 --> 00:14:17,440 Speaker 3: open source library. Transformers is what all AI builders are 213 00:14:17,559 --> 00:14:20,880 Speaker 3: using when they want to access those models, see how 214 00:14:20,920 --> 00:14:22,400 Speaker 3: they work, and build upon them. 215 00:14:23,720 --> 00:14:26,880 Speaker 4: What's striking about this field is that it's changing so fast, 216 00:14:26,920 --> 00:14:30,720 Speaker 4: it's improving so quickly. So how do open source models 217 00:14:31,440 --> 00:14:35,320 Speaker 4: keep up with that? How do they get iterated and improved? 218 00:14:35,440 --> 00:14:38,400 Speaker 3: Actually? It's not so much that open source is keeping 219 00:14:38,480 --> 00:14:41,440 Speaker 3: up with it. It's actually open source that is driving 220 00:14:42,160 --> 00:14:45,600 Speaker 3: that is driving this piece of change. And that's because 221 00:14:46,320 --> 00:14:51,680 Speaker 3: with open source and open research data, scientists researchers can 222 00:14:51,800 --> 00:14:55,480 Speaker 3: build upon each other's work, they can reproduce each other's work, 223 00:14:55,760 --> 00:14:59,760 Speaker 3: they can access each other's work using our open source library, 224 00:15:00,000 --> 00:15:02,320 Speaker 3: et cetera. So in a sense, it's not really that 225 00:15:02,720 --> 00:15:07,320 Speaker 3: open source AI is a new idea. It's rather the opposite. 226 00:15:07,480 --> 00:15:11,600 Speaker 3: There's been a blip of time in which closed source 227 00:15:11,840 --> 00:15:15,560 Speaker 3: AI seemed to be the dominant way, but it's really 228 00:15:16,120 --> 00:15:19,840 Speaker 3: a blip. In fact, you know, none of the incredible 229 00:15:19,880 --> 00:15:24,480 Speaker 3: advances that we're marvel about today would be possible without 230 00:15:24,680 --> 00:15:27,680 Speaker 3: open source. We're standing upon the shoulders of fifty years 231 00:15:27,680 --> 00:15:32,120 Speaker 3: of research and open source software. So I think that 232 00:15:32,120 --> 00:15:35,000 Speaker 3: that's really important. If it wasn't for that, we'll probably 233 00:15:35,000 --> 00:15:39,880 Speaker 3: be fifty years away from having these amazing experiences like 234 00:15:40,040 --> 00:15:45,840 Speaker 3: JGBT or stable diffusion, et cetera. So it's really open 235 00:15:45,880 --> 00:15:50,240 Speaker 3: source that is fueling this pace of change, all these 236 00:15:50,280 --> 00:15:53,800 Speaker 3: new models, all these new capabilities. To give you an example, 237 00:15:54,120 --> 00:15:58,640 Speaker 3: so Meta released the Lama large language model just a 238 00:15:58,680 --> 00:16:02,960 Speaker 3: few months ago, and ever since, there's been this Cambrian 239 00:16:03,120 --> 00:16:07,520 Speaker 3: explosion of variations and improvements upon the original models, and 240 00:16:07,560 --> 00:16:10,600 Speaker 3: today there are over a thousands of them that we 241 00:16:11,160 --> 00:16:16,560 Speaker 3: host and track and evaluate. So yeah, open source is 242 00:16:16,600 --> 00:16:20,280 Speaker 3: really the gas and the engine for that. 243 00:16:21,560 --> 00:16:24,400 Speaker 2: Jeff just made it clear that it is open source, 244 00:16:24,640 --> 00:16:28,640 Speaker 2: not closed that sets the pace for AI innovation. If 245 00:16:28,680 --> 00:16:33,240 Speaker 2: that's true, then forward thinking businesses shouldn't shy from leveraging 246 00:16:33,320 --> 00:16:37,680 Speaker 2: open source AI to solve their own proprietary challenges. But 247 00:16:37,880 --> 00:16:42,800 Speaker 2: how businesses can face serious obstacles when trying to adopt 248 00:16:43,040 --> 00:16:47,600 Speaker 2: open source technologies, like complying with government regulation or making 249 00:16:47,640 --> 00:16:51,880 Speaker 2: sure their customers data stays protected. In the next part 250 00:16:51,920 --> 00:16:56,200 Speaker 2: of their conversation, Jeff and Tim discuss how IBM's collaboration 251 00:16:56,360 --> 00:17:00,520 Speaker 2: with hugging Face empowers businesses to tap into the open 252 00:17:00,560 --> 00:17:04,879 Speaker 2: source AI community and how the watsonex platform can enable 253 00:17:04,920 --> 00:17:08,720 Speaker 2: them to customize those AI models to their needs. 254 00:17:09,400 --> 00:17:11,920 Speaker 4: Just want to ask about the partnership between hugging Face 255 00:17:11,960 --> 00:17:14,720 Speaker 4: and an IBM. How did that come about? 256 00:17:16,680 --> 00:17:23,280 Speaker 3: Well, it came through a conversation, a conversation between our CEO, 257 00:17:24,080 --> 00:17:29,320 Speaker 3: Clement de Lange and Bill Higgins IBM, who's really really 258 00:17:29,400 --> 00:17:34,280 Speaker 3: close to all the amazing research work and open source 259 00:17:34,400 --> 00:17:39,399 Speaker 3: work that's happening at IBM, and that conversation sort of 260 00:17:39,680 --> 00:17:44,240 Speaker 3: sparked the evidence that we needed to do something together. 261 00:17:44,840 --> 00:17:48,840 Speaker 3: We share a lot of values in terms of the 262 00:17:48,880 --> 00:17:53,600 Speaker 3: importance of open source, which is fundamental to us, with 263 00:17:54,000 --> 00:17:58,800 Speaker 3: the importance of doing things in an ethics first way 264 00:17:58,920 --> 00:18:04,040 Speaker 3: to enable the commune to incorporate ethical considerations in how 265 00:18:04,520 --> 00:18:09,760 Speaker 3: they're building AI. And we sort of have a different 266 00:18:10,040 --> 00:18:14,119 Speaker 3: audience to start with, which is all the AI builders 267 00:18:14,240 --> 00:18:18,840 Speaker 3: use hiking phase today to access all the models we 268 00:18:18,960 --> 00:18:22,879 Speaker 3: talked about, to use them using our open source and 269 00:18:22,920 --> 00:18:27,320 Speaker 3: build with them. And IBM has this incredible history of 270 00:18:27,440 --> 00:18:32,920 Speaker 3: working with enterprise companies and enabling them to make use 271 00:18:32,960 --> 00:18:37,000 Speaker 3: of that technology in a way that's compliant with everything 272 00:18:37,040 --> 00:18:40,800 Speaker 3: that an enterprise requires, and so being able to marry 273 00:18:40,840 --> 00:18:45,000 Speaker 3: these two things together is an amazing opportunity. And now 274 00:18:45,040 --> 00:18:49,280 Speaker 3: we can enable the largest corporations that have sort of 275 00:18:49,520 --> 00:18:54,920 Speaker 3: complex requirements in order to deploy machine learning systems and 276 00:18:55,720 --> 00:18:59,080 Speaker 3: give them an easy experience to take advantage of all 277 00:18:59,119 --> 00:19:01,600 Speaker 3: the latest and great is that AA has to offer 278 00:19:02,119 --> 00:19:02,920 Speaker 3: through our platform. 279 00:19:04,480 --> 00:19:08,040 Speaker 4: Let's talk about this idea of a single model or 280 00:19:08,080 --> 00:19:11,600 Speaker 4: a variety of models, because what I've been hearing you say. 281 00:19:12,160 --> 00:19:14,000 Speaker 4: You've been saying, oh, there are lots of models, there 282 00:19:14,040 --> 00:19:18,119 Speaker 4: are hundreds of thousands of models available on hugging Face. 283 00:19:18,280 --> 00:19:21,640 Speaker 4: But you've also said there's this single thing, the transformer, 284 00:19:22,280 --> 00:19:26,720 Speaker 4: and they're all transformers. So if they're all basically the 285 00:19:26,760 --> 00:19:31,480 Speaker 4: same thing, why can't you just build one super clever 286 00:19:31,560 --> 00:19:32,640 Speaker 4: model that can do everything. 287 00:19:34,760 --> 00:19:39,679 Speaker 3: That's a really interesting idea and very much a new idea. 288 00:19:40,520 --> 00:19:44,400 Speaker 3: The reason we have over a million repositories three hundred 289 00:19:44,480 --> 00:19:48,119 Speaker 3: thousand free and accessible models on a hiking Face platform 290 00:19:48,560 --> 00:19:52,320 Speaker 3: is that models are typically trained to do one thing, 291 00:19:52,680 --> 00:19:55,920 Speaker 3: and they're typically trained to do one thing with specific 292 00:19:55,960 --> 00:20:02,439 Speaker 3: types of data. And what became new and evidence in 293 00:20:02,480 --> 00:20:04,920 Speaker 3: the research that came out over the last couple of 294 00:20:05,000 --> 00:20:09,120 Speaker 3: years is that if you train a big enough model 295 00:20:09,600 --> 00:20:14,680 Speaker 3: with enough data, then those models start to have sort 296 00:20:14,680 --> 00:20:18,720 Speaker 3: of general capabilities. You can ask them to do different things. 297 00:20:19,000 --> 00:20:22,480 Speaker 3: You can even train them to respond to instructions. So 298 00:20:22,600 --> 00:20:26,840 Speaker 3: with the same model, you can say, hey, summarize this paragraph, 299 00:20:27,240 --> 00:20:30,960 Speaker 3: translate this into English, start a conversation in French, and 300 00:20:30,960 --> 00:20:34,560 Speaker 3: pivot to German. And so these are general sort of 301 00:20:34,680 --> 00:20:42,000 Speaker 3: language capabilities. And I think when CHGBT came online and 302 00:20:42,320 --> 00:20:47,000 Speaker 3: the world sort of discovered these new capabilities, there was, 303 00:20:47,560 --> 00:20:50,480 Speaker 3: at least for a short period, this sort of idea, 304 00:20:50,600 --> 00:20:54,480 Speaker 3: this sort of myth that the endgame of all this 305 00:20:55,440 --> 00:20:59,199 Speaker 3: is maybe one or a handful of models there are 306 00:20:59,400 --> 00:21:03,640 Speaker 3: so much better than anything else than exists, that they 307 00:21:03,640 --> 00:21:06,280 Speaker 3: can do anything that we can ask them to do, 308 00:21:07,080 --> 00:21:10,560 Speaker 3: and that's the only model that we will need. And I, 309 00:21:10,800 --> 00:21:15,080 Speaker 3: for one, think it is a myth. I don't think 310 00:21:15,119 --> 00:21:19,200 Speaker 3: it is practical for a variety of reasons. Say you're 311 00:21:19,600 --> 00:21:23,760 Speaker 3: writing an email and you have like this great suggestion 312 00:21:23,920 --> 00:21:28,199 Speaker 3: of text to sort of complete your sentence, Well, that's AI. 313 00:21:28,640 --> 00:21:31,159 Speaker 3: That's a large language model, that's a transformer model that 314 00:21:31,200 --> 00:21:33,840 Speaker 3: does that. So there are a ton of existing use 315 00:21:33,880 --> 00:21:37,520 Speaker 3: cases like this, and these use cases are powered by 316 00:21:38,320 --> 00:21:41,280 Speaker 3: specific models that have been trained to do one thing 317 00:21:41,400 --> 00:21:44,479 Speaker 3: well and to do it fast. If you wanted to 318 00:21:44,600 --> 00:21:51,200 Speaker 3: apply these sort of all knowing, powerful oracle type of model, 319 00:21:51,600 --> 00:21:55,639 Speaker 3: you would not be able to serve millions of customers 320 00:21:55,680 --> 00:21:58,359 Speaker 3: through a search engine. You will not be able to 321 00:22:00,080 --> 00:22:04,119 Speaker 3: complete people's sentences because the amount of money that you 322 00:22:04,160 --> 00:22:07,400 Speaker 3: would need, the number of computers that you would need 323 00:22:07,640 --> 00:22:13,240 Speaker 3: to run such of service just exceeds what is available 324 00:22:13,359 --> 00:22:18,760 Speaker 3: on the planet. So one reason for which it's not 325 00:22:18,880 --> 00:22:24,359 Speaker 3: a practical scenario is that it's just very expensive to 326 00:22:24,600 --> 00:22:27,440 Speaker 3: run those very very large models. 327 00:22:27,760 --> 00:22:29,920 Speaker 4: What I'm hearing is it's like, look, if you want 328 00:22:29,920 --> 00:22:33,679 Speaker 4: to screw in a screw you need a screwdriver. You 329 00:22:33,720 --> 00:22:37,720 Speaker 4: don't want an entire tool shed full of tools if 330 00:22:37,800 --> 00:22:39,960 Speaker 4: the task is to screw in a screwdriver, and sure 331 00:22:40,040 --> 00:22:43,240 Speaker 4: you could bring the toolshed that are all the tools. 332 00:22:43,280 --> 00:22:47,320 Speaker 4: There's a screwdriver there, but it's not necessary. It's incredibly expensive, 333 00:22:47,320 --> 00:22:52,119 Speaker 4: it's incredibly cumbersome, and that cost exists even though maybe 334 00:22:52,200 --> 00:22:54,879 Speaker 4: is the user who's just typing in a into a 335 00:22:54,920 --> 00:22:57,680 Speaker 4: prompt box. The user may not see it, but it's 336 00:22:57,680 --> 00:22:58,720 Speaker 4: still very real. 337 00:23:00,040 --> 00:23:03,480 Speaker 3: That's right. And then another one is performance. So taking 338 00:23:03,520 --> 00:23:06,760 Speaker 3: the screwdriver example, so and by the way, like we're 339 00:23:06,800 --> 00:23:09,560 Speaker 3: not quite there at this moment where we have this 340 00:23:09,720 --> 00:23:13,240 Speaker 3: all knowing, powerful oracle that is still sort of a 341 00:23:13,320 --> 00:23:16,919 Speaker 3: sci fi scenario, but we have screw drivers, but we 342 00:23:17,040 --> 00:23:21,680 Speaker 3: also have the leatherman, right, the multitol Swiss army knife. 343 00:23:21,920 --> 00:23:24,919 Speaker 3: And that's sort of the moment that we are in today. 344 00:23:24,960 --> 00:23:28,600 Speaker 3: But now if I'm trying to open up my computer, 345 00:23:29,200 --> 00:23:32,439 Speaker 3: turns out that it requires a specific kind of screw 346 00:23:32,600 --> 00:23:36,760 Speaker 3: like these tiny little tork screws, and having a torqu 347 00:23:36,800 --> 00:23:40,520 Speaker 3: screwdriver will get me much further than trying to use 348 00:23:40,760 --> 00:23:43,399 Speaker 3: my leather man, where maybe I'll get the knife blade 349 00:23:43,440 --> 00:23:46,520 Speaker 3: and it will mess up the screw and maybe eventually 350 00:23:46,520 --> 00:23:49,160 Speaker 3: I'll get to what I need. But my point is 351 00:23:49,280 --> 00:23:54,399 Speaker 3: that if you take a very specifically trained model for 352 00:23:54,480 --> 00:23:57,960 Speaker 3: a particular problem, it will work much better. It will 353 00:23:57,960 --> 00:24:02,760 Speaker 3: give you better results than a very very generalistic, big 354 00:24:02,840 --> 00:24:05,800 Speaker 3: model that can do a lot of things. And so 355 00:24:05,880 --> 00:24:10,119 Speaker 3: for things like search engines or things like translation, for 356 00:24:10,280 --> 00:24:15,000 Speaker 3: things that are very specific, companies are much better off 357 00:24:15,119 --> 00:24:19,680 Speaker 3: using smaller, more efficient models that produce better results. 358 00:24:19,480 --> 00:24:24,000 Speaker 4: That's really interesting. And presumably then being able to know 359 00:24:24,040 --> 00:24:26,800 Speaker 4: which model to use, or being able to know who 360 00:24:26,840 --> 00:24:30,640 Speaker 4: to ask which model to use, becomes a very important capability. 361 00:24:31,480 --> 00:24:35,000 Speaker 3: Yes, and that's what we're trying to make easy through 362 00:24:35,040 --> 00:24:35,800 Speaker 3: our platform. 363 00:24:37,160 --> 00:24:41,160 Speaker 4: So tell me about how this works with IBM's what's 364 00:24:41,160 --> 00:24:44,760 Speaker 4: an X platform? How do you see hugging faces customers 365 00:24:44,800 --> 00:24:45,640 Speaker 4: benefiting from that? 366 00:24:47,560 --> 00:24:51,640 Speaker 3: The end goal is to make it really easy for 367 00:24:51,760 --> 00:24:56,240 Speaker 3: what's an X customers to make use of all the 368 00:24:56,320 --> 00:25:00,600 Speaker 3: great models and libraries that we talked about, all the 369 00:25:00,600 --> 00:25:03,320 Speaker 3: the three hundred thousand models are today on hugging face 370 00:25:04,160 --> 00:25:08,440 Speaker 3: and to do this we need to really collaborate deeply 371 00:25:08,520 --> 00:25:12,080 Speaker 3: with the IBM teams that build the What's and X 372 00:25:12,160 --> 00:25:17,360 Speaker 3: platform so that our libraries, our open source our models 373 00:25:17,760 --> 00:25:21,480 Speaker 3: are well integrated into the platform. If you are a 374 00:25:21,640 --> 00:25:24,560 Speaker 3: single user, if you are a data science student and 375 00:25:24,600 --> 00:25:26,680 Speaker 3: you want to use a model, is we make it 376 00:25:26,720 --> 00:25:29,399 Speaker 3: super easy, right. We have our open source library. You 377 00:25:29,440 --> 00:25:32,159 Speaker 3: can download the model on your computer and run with 378 00:25:32,240 --> 00:25:37,320 Speaker 3: it then. But in enterprises there is a vast complexity 379 00:25:37,560 --> 00:25:42,800 Speaker 3: of infrastructure and rules around what people can do and 380 00:25:43,400 --> 00:25:47,600 Speaker 3: how the data can be accessed, and all this complexity 381 00:25:48,280 --> 00:25:52,879 Speaker 3: is sort of solved by the Watson X platform. 382 00:25:53,560 --> 00:25:57,520 Speaker 4: This season of the Smart Talks podcast features what we're 383 00:25:57,520 --> 00:26:00,399 Speaker 4: calling new creators. Do you see yourself as being a 384 00:26:00,440 --> 00:26:01,440 Speaker 4: creative person? 385 00:26:02,359 --> 00:26:05,960 Speaker 3: Ah, I think it's a requirement for the job. I mean, 386 00:26:05,960 --> 00:26:10,720 Speaker 3: we're in such a new and rapidly evolving industry that 387 00:26:11,000 --> 00:26:15,000 Speaker 3: we have to be creative in order to invent the 388 00:26:15,080 --> 00:26:19,640 Speaker 3: business models the use cases of tomorrow. My role within 389 00:26:19,680 --> 00:26:24,680 Speaker 3: the company is really to create the business around all 390 00:26:24,840 --> 00:26:28,840 Speaker 3: the great work of our science and open source and 391 00:26:28,960 --> 00:26:33,080 Speaker 3: product team, and by and large, the business model of 392 00:26:33,240 --> 00:26:38,200 Speaker 3: AI within the whole ecosystem is still something that companies 393 00:26:38,240 --> 00:26:43,200 Speaker 3: are trying to figure out. So creativity is really important 394 00:26:43,320 --> 00:26:47,120 Speaker 3: to really have the conversation with companies, understand what they're 395 00:26:47,160 --> 00:26:49,240 Speaker 3: trying to do, and then build the right kind of solution. 396 00:26:49,840 --> 00:26:54,080 Speaker 3: So that's like where creativity comes into play. 397 00:26:54,800 --> 00:26:59,000 Speaker 4: And one of the things that you've you've been talking 398 00:26:59,040 --> 00:27:02,520 Speaker 4: about is just this growing number of models, this growing 399 00:27:02,600 --> 00:27:09,040 Speaker 4: number of capabilities, this growing number of use cases enormously 400 00:27:09,080 --> 00:27:15,000 Speaker 4: exciting but also I think completely bewildering for most people 401 00:27:16,000 --> 00:27:20,640 Speaker 4: who are trying to navigate their way through this maze 402 00:27:20,680 --> 00:27:23,960 Speaker 4: of possibilities that is growing faster than they can even 403 00:27:24,200 --> 00:27:28,200 Speaker 4: learn about it. So how are you helping people navigate 404 00:27:28,400 --> 00:27:30,879 Speaker 4: and make choices in that environment? And how does the 405 00:27:30,920 --> 00:27:32,840 Speaker 4: partnership with IBM help with that? 406 00:27:35,640 --> 00:27:39,520 Speaker 3: Well? As I said, our vision is that AI machine 407 00:27:39,600 --> 00:27:44,639 Speaker 3: learning is becoming the default way of creating technology and 408 00:27:44,680 --> 00:27:48,520 Speaker 3: that means like every product, app, service that you're going 409 00:27:48,600 --> 00:27:52,159 Speaker 3: to be using is going to be using AI to 410 00:27:52,280 --> 00:27:57,280 Speaker 3: do whatever it is better faster, And I guess there 411 00:27:57,280 --> 00:28:01,400 Speaker 3: are two competing visions of doing world coming from that. 412 00:28:01,480 --> 00:28:07,639 Speaker 3: There is this vision of the oracle, all powerful model 413 00:28:07,720 --> 00:28:12,159 Speaker 3: that can do everything, and our vision is different. Our 414 00:28:12,240 --> 00:28:17,640 Speaker 3: vision is that every single company will be able to 415 00:28:17,720 --> 00:28:22,680 Speaker 3: create their own models that they own, that they can use, 416 00:28:22,760 --> 00:28:27,560 Speaker 3: that they control, and that's the vision that we're trying 417 00:28:27,600 --> 00:28:31,440 Speaker 3: to bring to life through our open source tools that 418 00:28:31,760 --> 00:28:35,560 Speaker 3: make this work easy. Through our platform where you can 419 00:28:35,600 --> 00:28:38,640 Speaker 3: find all those pre train models are shared by the community. 420 00:28:39,080 --> 00:28:41,840 Speaker 3: So we really want to empower companies to build their 421 00:28:41,880 --> 00:28:45,640 Speaker 3: own stuff, not to outsource all the intelligence to a 422 00:28:45,720 --> 00:28:51,120 Speaker 3: third party. And the What's on next platform from IBM 423 00:28:51,920 --> 00:28:56,920 Speaker 3: gives those tools to enterprise companies, So that's you can 424 00:28:57,600 --> 00:29:02,680 Speaker 3: use the open source models hiking Face offers, then you 425 00:29:02,760 --> 00:29:07,480 Speaker 3: can improve them with your own data without sharing that 426 00:29:07,600 --> 00:29:10,520 Speaker 3: data to a third party, and then you could do 427 00:29:11,160 --> 00:29:16,680 Speaker 3: all of this work in compliance with whatever governance requirements 428 00:29:17,080 --> 00:29:20,800 Speaker 3: that you have for your company, maybe your finance services 429 00:29:20,800 --> 00:29:24,680 Speaker 3: company and you have a specific set of rules, maybe 430 00:29:25,000 --> 00:29:30,120 Speaker 3: your healthcare company and you have very strong privacy requirements 431 00:29:30,320 --> 00:29:35,480 Speaker 3: for patients data. Maybe your tech company, and you have 432 00:29:35,600 --> 00:29:40,560 Speaker 3: your customers, your users personal information, so you need to 433 00:29:40,560 --> 00:29:43,320 Speaker 3: be able to do this work respecting all of that. 434 00:29:44,360 --> 00:29:46,280 Speaker 4: Jeff Bridier, thank you very much. 435 00:29:46,960 --> 00:29:48,640 Speaker 3: Thanks so much to it's fun. 436 00:29:50,320 --> 00:29:53,280 Speaker 2: To create the AI models of the future. We're going 437 00:29:53,280 --> 00:29:55,800 Speaker 2: to need open source. That means as a place for 438 00:29:55,960 --> 00:29:58,960 Speaker 2: business in the open source community to harness the game 439 00:29:59,080 --> 00:30:04,600 Speaker 2: changing potential of AI innovation. Like Jeff said, businesses face 440 00:30:04,840 --> 00:30:08,800 Speaker 2: unique challenges they need to solve at scale without proper 441 00:30:08,840 --> 00:30:13,000 Speaker 2: support systems. Tapping into open source AI at enterprise level 442 00:30:13,320 --> 00:30:16,600 Speaker 2: is daunting finding the right size model for the job, 443 00:30:16,920 --> 00:30:21,480 Speaker 2: fine tuning its purpose, all while addressing governance requirements around 444 00:30:21,560 --> 00:30:27,520 Speaker 2: data privacy and ethics. So for businesses, IBM's collaboration with 445 00:30:27,640 --> 00:30:31,440 Speaker 2: hugging Face is a market progress because it signifies that 446 00:30:31,600 --> 00:30:36,040 Speaker 2: business can tap into open source AI while preserving enterprise 447 00:30:36,120 --> 00:30:41,280 Speaker 2: level integrity. Businesses should embrace the open source community and 448 00:30:41,360 --> 00:30:45,120 Speaker 2: the AI future, much like hugging Face and its emoji 449 00:30:45,200 --> 00:30:49,720 Speaker 2: namesake suggests. I'm Malcolm Gladwell. This is a paid advertisement 450 00:30:49,840 --> 00:30:54,479 Speaker 2: from IBM. Smart Talks with IBM is produced by Matt Romano, 451 00:30:54,960 --> 00:30:59,200 Speaker 2: David jaw Nisha Nkat and Royston Deserve with Jacob Goldstein 452 00:31:00,320 --> 00:31:04,240 Speaker 2: by Lydia gene Kott. Our engineers are Jason Gambrel, Sarah 453 00:31:04,280 --> 00:31:09,720 Speaker 2: Bruger and Ben Tolliday. Theme song by Gramoscope. Special thanks 454 00:31:09,720 --> 00:31:13,400 Speaker 2: to Carlei Migliori, Andy Kelly, Kathy Callahan, and the eight 455 00:31:13,440 --> 00:31:17,440 Speaker 2: Bar and IBM teams, as well as the Pushkin marketing team. 456 00:31:17,640 --> 00:31:20,600 Speaker 2: Smart Talks with IBM is a production of Pushkin Industries 457 00:31:20,960 --> 00:31:25,720 Speaker 2: and Ruby Studio at iHeartMedia. To find more Pushkin podcasts, 458 00:31:25,920 --> 00:31:30,560 Speaker 2: listen on the iHeartRadio app, Apple Podcasts, or wherever you 459 00:31:30,720 --> 00:31:42,360 Speaker 2: listen to podcasts.