WEBVTT - Smart Talks with IBM: Hugging Face and watsonx: Why Open Source Is the Future of AI in Business 0:00:00.120 --> 0:00:02.840 Hey everyone, it's Robert and Joe here. Today we've got 0:00:02.840 --> 0:00:04.720 something a little bit different to share with you. It 0:00:04.840 --> 0:00:08.000 is a new season of the Smart Talks with IBM 0:00:08.119 --> 0:00:09.119 podcast series. 0:00:09.600 --> 0:00:11.680 Today we are witnessed to one of those rare moments 0:00:11.680 --> 0:00:14.360 in history, the rise of an innovative technology with the 0:00:14.360 --> 0:00:18.680 potential to radically transform business and society forever. The technology, 0:00:18.760 --> 0:00:22.200 of course, is artificial intelligence, and it's the central focus 0:00:22.239 --> 0:00:24.800 for this new season of Smart Talks with IBM. 0:00:25.320 --> 0:00:28.400 Join hosts from your favorite Pushkin podcasts as they talk 0:00:28.480 --> 0:00:31.640 with industry experts and leaders to explore how businesses can 0:00:31.680 --> 0:00:35.360 integrate AI into their workflows and help drive real change 0:00:35.400 --> 0:00:38.160 in this new era of AI. And of course, host 0:00:38.280 --> 0:00:40.440 Malcolm Gladwell will be there to guide you through the 0:00:40.479 --> 0:00:42.640 season and throw in his two cents as well. 0:00:43.120 --> 0:00:46.120 Look out for new episodes of Smart Talks with IBM 0:00:46.400 --> 0:00:49.519 every other week on the iHeartRadio app, Apple Podcasts, or 0:00:49.560 --> 0:00:53.360 wherever you get your podcasts, and learn more at IBM 0:00:53.479 --> 0:00:55.480 dot com slash smart talks. 0:00:56.200 --> 0:00:59.800 Hello, Hello, Welcome to Smart Talks with IBM, a podcast 0:00:59.880 --> 0:01:05.840 from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. This 0:01:05.959 --> 0:01:10.920 season we're continuing our conversation with new creators visionaries who 0:01:10.920 --> 0:01:15.080 are creatively applying technology in business to drive change, but 0:01:15.160 --> 0:01:19.720 with a focus on the transformative power of artificial intelligence 0:01:20.040 --> 0:01:22.760 and what it means to leverage AI as a game 0:01:22.880 --> 0:01:27.759 changing multiplier for your business. Our guest today is Jeff Boutier, 0:01:28.200 --> 0:01:31.720 head of Product and Growth at hugging Face, the leading 0:01:31.840 --> 0:01:37.039 open source and open science artificial intelligence platform. An engineer 0:01:37.040 --> 0:01:40.920 by background, he has a self professed obsession with the 0:01:40.959 --> 0:01:45.560 business of technology. Recently, IBM and hugging Face announced the 0:01:45.600 --> 0:01:50.200 collaboration bringing together hugging faces repositories of open source AI 0:01:50.320 --> 0:01:55.160 models with IBM's Watson X platform. It's a move that 0:01:55.200 --> 0:01:59.600 gives businesses even more access to AI while staying true 0:01:59.600 --> 0:02:05.320 to IBA's long standing philosophy of supporting open source technology. 0:02:06.200 --> 0:02:10.080 With open source, businesses can build better AI models that 0:02:10.160 --> 0:02:14.680 suit their specific needs using their own proprietary data while 0:02:14.720 --> 0:02:19.799 browsing a ready catalog of pre trained models. In today's episode, 0:02:20.000 --> 0:02:23.120 you'll hear why open source is so crucial to the 0:02:23.200 --> 0:02:27.800 advancement of AI, how IBM's Watson X interacts with open 0:02:27.800 --> 0:02:32.280 source AI, and Jeff's thoughts on why this singular omnipotent 0:02:32.560 --> 0:02:36.600 AI model is a myth. Jeff spoke with Tim Harford, 0:02:36.840 --> 0:02:41.119 host of the Pushkin podcast Cautionary Tales, a longtime columnist 0:02:41.120 --> 0:02:44.600 at The Financial Times, where he writes the Undercover Economist. 0:02:44.880 --> 0:02:48.680 Tim is also a BBC broadcaster with his show More 0:02:48.840 --> 0:02:52.520 or Less. Okay, let's get to the interview. 0:03:00.040 --> 0:03:02.840 I am a Jeff Boudier and I'm a product director 0:03:02.919 --> 0:03:04.279 at Hugging Face. 0:03:05.120 --> 0:03:08.640 So I'm immediately intrigue. Hugging Face. Is this a reference 0:03:08.639 --> 0:03:10.240 to the Alien movie or something else. 0:03:11.560 --> 0:03:15.400 It is not, and it may be not obvious to 0:03:15.520 --> 0:03:19.200 a listener, but hugging Face is the name of that 0:03:19.320 --> 0:03:22.160 cute emoji, you know, the one that's smiling with his 0:03:22.320 --> 0:03:25.520 two hands extended like that to give you a big hug. 0:03:25.760 --> 0:03:29.000 That's hugging Face. So basically, we name the company after 0:03:29.120 --> 0:03:30.480 an emoji. 0:03:31.360 --> 0:03:33.760 And it is I saw your website and it is 0:03:33.800 --> 0:03:36.880 a very friendly emoji. So that's that's nice. So tell 0:03:36.960 --> 0:03:39.240 us a little bit about hugging Face and about what 0:03:39.320 --> 0:03:39.880 you do there. 0:03:40.320 --> 0:03:44.800 Of course, hugging Face is the leading open platform for 0:03:45.160 --> 0:03:49.720 AI builders and it's the place that all of the 0:03:50.000 --> 0:03:55.440 AI researchers use to share their work, their new AI 0:03:55.520 --> 0:04:00.560 models and collaborate around them. It's the place where the 0:04:00.720 --> 0:04:05.880 data scientists go and find those pre train models and 0:04:06.040 --> 0:04:09.680 access them and use them and work with them. And 0:04:09.760 --> 0:04:14.440 increasingly it's the place where developers are coming to turn 0:04:14.560 --> 0:04:19.840 all of these AI models and datasets into their own applications, 0:04:19.880 --> 0:04:21.039 their own features. 0:04:21.720 --> 0:04:24.440 So it's like the I don't know, the Facebook group 0:04:24.520 --> 0:04:27.360 or the Reddit or the Twitter for people who are 0:04:27.400 --> 0:04:31.680 interested in particularly generative language AI or all kinds of 0:04:31.760 --> 0:04:33.280 artificial intelligence. 0:04:33.520 --> 0:04:37.520 All kinds of AI really, and of course generative AI 0:04:37.720 --> 0:04:42.679 is this new wave that has caught the world by storm. 0:04:43.040 --> 0:04:45.679 But on a hiking face you can find any kind 0:04:45.800 --> 0:04:50.120 of model. The new sort of transformers models to do 0:04:50.200 --> 0:04:54.920 anything from translation or if you wanted to transcribe what 0:04:54.960 --> 0:04:58.760 I'm saying into text, then you would use a transformer model. 0:04:58.800 --> 0:05:01.600 If you wanted to then take that text and make 0:05:01.640 --> 0:05:05.560 a summary, that would be another transformer model. If you 0:05:05.640 --> 0:05:10.080 wanted to create a nice little thumbnail for this podcast 0:05:10.120 --> 0:05:14.000 by typing a sentence, that would be another type of model. 0:05:14.560 --> 0:05:17.400 So all these models you can find, there's actually three 0:05:17.480 --> 0:05:22.120 hundred thousands that are free and publicly accessible. You can 0:05:22.200 --> 0:05:25.920 find them on our website at Hikingphase dot co and 0:05:26.040 --> 0:05:28.880 use them using our open source libraries. 0:05:29.720 --> 0:05:32.680 And so this this is fascinating. So there are three 0:05:32.800 --> 0:05:36.200 hundred thousand models. Now when you say model, I'm thinking 0:05:36.200 --> 0:05:39.560 in my head, Oh, it's kind of like a computer program. 0:05:39.600 --> 0:05:42.640 There were three hundred thousand computer programs. Is that roughly 0:05:42.720 --> 0:05:43.719 right or it not? 0:05:43.839 --> 0:05:49.920 Really, it's a general idea. A model is a giant 0:05:51.080 --> 0:05:56.480 set of numbers that are working together to sift through 0:05:57.120 --> 0:06:00.320 some inputs that you're going to give it. So think 0:06:00.400 --> 0:06:04.880 of it of a big black box filled with numbers, 0:06:05.800 --> 0:06:10.960 and you give it as an input, maybe some text, 0:06:11.320 --> 0:06:15.240 maybe a prompt, so you're asking you're giving an instruction 0:06:15.480 --> 0:06:18.120 to the model, or maybe you give it an image 0:06:18.160 --> 0:06:22.640 as an input, and then it will sift through that 0:06:22.760 --> 0:06:26.800 information thanks to all of these numbers, which we call 0:06:26.839 --> 0:06:31.320 in the field parameters, and it will produce an output. 0:06:31.839 --> 0:06:34.400 So when I told you, hey, we can transcribe this 0:06:34.640 --> 0:06:38.560 conversation into text, the input would have been the conversation 0:06:39.160 --> 0:06:41.839 in an audio file, and then the output would have 0:06:41.880 --> 0:06:44.880 been the text of the transcription. If you want to 0:06:44.920 --> 0:06:48.960 create a thumbnail for this podcast episode, then the input 0:06:49.000 --> 0:06:51.880 would be what we call the prompt, which is really 0:06:51.880 --> 0:06:56.400 a text description like a Frenchman in San Francisco talking 0:06:56.440 --> 0:07:02.080 about machine learning, and the output would be completely original image. 0:07:02.640 --> 0:07:05.960 So that's how I think about what an AI model is, 0:07:06.440 --> 0:07:10.600 and I think what we're starting to realize is that 0:07:11.480 --> 0:07:15.640 this is becoming the new way of building technology in 0:07:15.720 --> 0:07:19.679 the world. It has been for the field of dealing, 0:07:19.840 --> 0:07:23.760 understanding generating text for quite some time, but now it's 0:07:23.880 --> 0:07:27.800 sort of moving across every field of technology. We have 0:07:28.200 --> 0:07:32.360 models to create images, as I say, but also to 0:07:32.480 --> 0:07:37.960 generate new proteins to make predictions on numerical data. So 0:07:38.040 --> 0:07:42.520 every kind of field of machine learning is now using 0:07:43.960 --> 0:07:47.840 this new type of models. But what's interesting is that 0:07:48.440 --> 0:07:51.960 if you're say a product manager at a tech company 0:07:52.000 --> 0:07:54.440 and you say, hey, I want to build a feature 0:07:55.080 --> 0:07:58.560 that does this. A few years ago, the approach would 0:07:58.560 --> 0:08:02.680 have been to ask software developer to write a thousand 0:08:02.760 --> 0:08:06.200 lines of code in order to build a prototype. And 0:08:06.280 --> 0:08:09.720 the new way of doing things today is to go 0:08:09.880 --> 0:08:14.400 look for an off the shelf pre train model that 0:08:14.480 --> 0:08:18.560 does a pretty good job at solving exactly that problem, 0:08:18.680 --> 0:08:21.800 so you can create a prototype of that feature fast. 0:08:21.800 --> 0:08:24.360 So it's a new approach of building tech. 0:08:24.560 --> 0:08:27.640 I'm not a programmer. But I'm aware that there was 0:08:27.880 --> 0:08:30.440 this idea of open source code and now we have 0:08:30.520 --> 0:08:33.480 open source models. So what does it mean for something 0:08:33.520 --> 0:08:34.400 to be open source? 0:08:35.040 --> 0:08:40.760 Open source AI actually means a lot of different specific things. 0:08:41.480 --> 0:08:45.640 It's the open source implementation of the model. So if 0:08:45.679 --> 0:08:49.959 you use the Hugging Phase transformers library to use a model, 0:08:50.000 --> 0:08:54.560 you're using an open source code library to use that model, just. 0:08:55.040 --> 0:08:57.680 To end up on that. The transformers, these are these 0:08:57.760 --> 0:09:01.000 kind of ways of turning a picture of a dog 0:09:01.160 --> 0:09:03.800 into a text output that says, hey, this is a 0:09:03.840 --> 0:09:06.400 picture of a dog, or this is a French text 0:09:06.440 --> 0:09:09.280 and with the transformers helping you turn it into English text, 0:09:09.360 --> 0:09:11.280 or it's doing all of these things that you've been describing. 0:09:11.320 --> 0:09:15.280 That's the transformer is the kind of the engine at 0:09:15.320 --> 0:09:15.880 the heart of. 0:09:15.840 --> 0:09:20.719 That, yes, exactly. And we call them transformers because they 0:09:20.720 --> 0:09:24.800 correspond to this new way of building machine learning models 0:09:25.200 --> 0:09:29.720 that was introduced by Google actually with a very important 0:09:29.800 --> 0:09:32.800 paper called Attention Is All You Need and that was 0:09:32.840 --> 0:09:37.839 published in twenty seventeen by researchers out of Google Deep Mind. 0:09:38.760 --> 0:09:42.560 Well that's just six years so new. 0:09:43.320 --> 0:09:46.600 It is very new, and ever since the pace of 0:09:46.840 --> 0:09:52.280 innovation of like new model architectures has really really accelerated, 0:09:52.600 --> 0:09:57.360 but it really started from this inflection point that came 0:09:57.480 --> 0:10:01.760 from this paper, and it's implement in what is now 0:10:01.800 --> 0:10:07.600 called transformer models, the transformer that has conquered every area 0:10:07.720 --> 0:10:09.440 of machine learning since. 0:10:09.640 --> 0:10:12.680 Okay, so say churned you up. So you've got this 0:10:12.800 --> 0:10:17.160 library of Transformer models and that open source and that 0:10:17.200 --> 0:10:19.679 means that means what anyone can use them for free, 0:10:20.480 --> 0:10:22.360 or that anybody can implement them for free. What does 0:10:22.360 --> 0:10:24.040 it mean? 0:10:24.200 --> 0:10:27.160 So again, there's lots that go into it, but the 0:10:27.200 --> 0:10:31.600 most important thing is for the model itself to be 0:10:31.840 --> 0:10:35.960 available so that a data scientists or an engineer can 0:10:36.360 --> 0:10:40.719 download them and use them. And also there are a 0:10:40.760 --> 0:10:45.600 lot of considerations about how you make them accessible, and 0:10:45.679 --> 0:10:49.600 a very important one is whether or not you give 0:10:49.840 --> 0:10:54.880 access to the training data, all the information that went 0:10:55.080 --> 0:10:59.800 into training that model and teaching it to do what 0:11:00.080 --> 0:11:01.000 it's trained to do. 0:11:01.200 --> 0:11:04.160 So I might have fed millions of words into a 0:11:04.280 --> 0:11:07.360 into a language transformer, or I might have fed millions 0:11:07.400 --> 0:11:10.000 of photographs into a into a picture transformer. 0:11:10.080 --> 0:11:13.559 Yeah, yes, and now it's trillions and that and the 0:11:13.880 --> 0:11:17.600 accessibility of that training data is very very important. 0:11:18.520 --> 0:11:24.319 What's the relationship between the hugging Face libraries and GitHub, which, 0:11:25.440 --> 0:11:29.720 if I understand GitHub correctly, it's this the repository of 0:11:29.760 --> 0:11:33.720 open source code, lots and lots of lines of code 0:11:33.720 --> 0:11:37.440 and routines and programs that are shared and updated and 0:11:38.360 --> 0:11:41.760 tracked and they're all available on GitHub, which sounds similar 0:11:41.800 --> 0:11:44.000 to what you're doing with hugging Face for AI. So 0:11:44.040 --> 0:11:46.959 what what what is the interaction or the relationship there? 0:11:47.559 --> 0:11:50.000 Yeah, I think you nailed it on the head there. 0:11:50.040 --> 0:11:54.199 So hugging phase is to AI where GitHub is to code. Right, 0:11:54.200 --> 0:12:00.920 there's this central platform where AI builders can go and 0:12:01.000 --> 0:12:06.160 collaborate around AI artifacts, which are models and datasets. So 0:12:06.200 --> 0:12:10.640 it's quite different than software, but we play this central 0:12:10.920 --> 0:12:15.920 role in the community to share and collaborate and access 0:12:16.040 --> 0:12:20.240 all of those artifacts for AI like it had offers 0:12:20.240 --> 0:12:20.800 for code. 0:12:22.040 --> 0:12:24.960 And that community must be incredibly important. I mean, the 0:12:25.000 --> 0:12:27.600 open sources is nothing if you don't have a community 0:12:27.640 --> 0:12:30.000 of people working on it. So how have you been 0:12:30.040 --> 0:12:33.160 able to foster and nurture that community? 0:12:33.760 --> 0:12:37.120 Well, I think it goes to the origins of the 0:12:37.200 --> 0:12:41.320 transformer model and hugging and face role into that. So 0:12:41.960 --> 0:12:46.560 when the first sort of open model came out. It 0:12:46.640 --> 0:12:49.800 was called Bird and it came out of Google. The 0:12:49.840 --> 0:12:54.400 only way you could access it was to use a 0:12:54.440 --> 0:12:58.880 tool called TensorFlow. But it happened that most of the 0:12:59.320 --> 0:13:05.400 AI community was using a different tool called PyTorch. And 0:13:05.480 --> 0:13:10.560 something that Hugging Face did is to make that new 0:13:10.640 --> 0:13:17.040 model Bert accessible to all PyTorch user and they did 0:13:17.040 --> 0:13:21.040 it in open source. It was a project called Bert's 0:13:21.240 --> 0:13:24.079 pre Trained PyTorch or Burt PyTorch pre Trained. 0:13:24.600 --> 0:13:26.720 So this is like being able to play my Zelda 0:13:26.760 --> 0:13:30.800 game on an Xbox or a PlayStation, right or am 0:13:30.840 --> 0:13:32.480 I not really understanding what's going on? 0:13:32.920 --> 0:13:35.280 No, That's exactly what it is. And the thing is 0:13:35.480 --> 0:13:39.440 everybody was using the game Boy and so it became 0:13:39.800 --> 0:13:44.560 a very popular and from there the community sort of 0:13:44.640 --> 0:13:48.000 gathered to make all of the other models that were 0:13:48.040 --> 0:13:52.520 then published by AI researchers available through that library, which 0:13:52.679 --> 0:13:57.640 was quickly renamed from bert pre Trained PyTorch into transformers 0:13:57.679 --> 0:14:02.160 to welcome like all of these different new models, and 0:14:02.520 --> 0:14:07.960 today that's open source library. Transformers is what all AI 0:14:08.000 --> 0:14:11.840 builders are using when they want to access those models, 0:14:11.880 --> 0:14:14.239 see how they work, and build upon them. 0:14:15.120 --> 0:14:18.280 What's striking about this field is that it's changing so fast, 0:14:18.280 --> 0:14:22.080 it's improving so quickly, So how do open source models 0:14:22.800 --> 0:14:26.680 keep up with that? How do they get iterated and improved? 0:14:26.800 --> 0:14:29.760 Actually, it's not so much that open source is keeping 0:14:29.840 --> 0:14:32.800 up with it. It's actually open source that is driving, 0:14:33.520 --> 0:14:37.000 that is driving this piece of change. And that's because 0:14:37.680 --> 0:14:43.040 with open source and open research data, scientists researchers can 0:14:43.160 --> 0:14:46.840 build upon each other's work, they can reproduce each other's work, 0:14:47.160 --> 0:14:51.280 they can access each other's work using our open source libraries, 0:14:51.280 --> 0:14:53.600 et cetera. So, in a sense, it's not really that 0:14:54.080 --> 0:14:58.680 open source AI is a new idea. It's rather the opposite. 0:14:58.840 --> 0:15:02.960 There's been a blip of time in which closed source 0:15:03.200 --> 0:15:06.920 AI seemed to be the dominant way, but it's really 0:15:07.480 --> 0:15:11.200 a blip. In fact, you know, none of the incredible 0:15:11.240 --> 0:15:15.840 advances that we're marvel about today would be possible without 0:15:16.040 --> 0:15:19.040 open source. We're standing upon the shoulders of fifty years 0:15:19.040 --> 0:15:23.360 of research and open source software. So I think that 0:15:23.480 --> 0:15:26.360 that's really important. If it wasn't for that, we'll probably 0:15:26.360 --> 0:15:31.240 be fifty years away from having these amazing experiences like 0:15:31.400 --> 0:15:37.200 CHTGBT or stable diffusion, et cetera. So it's really open 0:15:37.280 --> 0:15:41.600 source that is fueling this pace of change, all these 0:15:41.640 --> 0:15:45.160 new models, all these new capabilities. To give you an example, 0:15:45.480 --> 0:15:49.960 so Meta released the Lama large language model just a 0:15:50.040 --> 0:15:54.360 few months ago, and ever since there's been this Cambrian 0:15:54.480 --> 0:15:58.920 explosion of variations and improvements upon the original models, and 0:15:58.960 --> 0:16:02.080 today there are there are thousands of them that we 0:16:02.520 --> 0:16:07.960 host and track and evaluate. So yeah, open source is 0:16:08.000 --> 0:16:11.640 really the gas and the engine for that. 0:16:12.920 --> 0:16:15.760 Jeff just made it clear that it is open source, 0:16:16.000 --> 0:16:20.040 not closed, that sets the pace for AI innovation. If 0:16:20.040 --> 0:16:24.600 that's true, then forward thinking businesses shouldn't shy from leveraging 0:16:24.680 --> 0:16:29.040 open source AI to solve their own proprietary challenges. But 0:16:29.240 --> 0:16:34.160 how businesses can face serious obstacles when trying to adopt 0:16:34.400 --> 0:16:38.960 open source technologies, like complying with government regulation or making 0:16:39.000 --> 0:16:43.240 sure their customers data stays protected. In the next part 0:16:43.280 --> 0:16:47.560 of their conversation, Jeff and Tim discuss how IBM's collaboration 0:16:47.760 --> 0:16:51.880 with hugging Face empowers businesses to tap into the open 0:16:51.920 --> 0:16:56.240 source AI community and how the watsonex platform can enable 0:16:56.280 --> 0:17:00.160 them to customize those AI models to their needs. 0:17:00.760 --> 0:17:03.280 Just wants to ask about the partnership between hugging Face 0:17:03.360 --> 0:17:05.919 and an IBM, how did that come. 0:17:05.800 --> 0:17:13.560 About, Well, it came through a conversation, a conversation between 0:17:13.920 --> 0:17:20.359 our CEO, Clement Delangue and Bill Higgins IBM, who's really 0:17:20.400 --> 0:17:25.280 really close to all the amazing research work and open 0:17:25.320 --> 0:17:30.680 source work that's happening at IBM, and that conversation sort 0:17:30.680 --> 0:17:35.600 of sparked the evidence that we needed to do something together. 0:17:36.200 --> 0:17:40.200 We share a lot of values in terms of the 0:17:40.240 --> 0:17:45.000 importance of open source, which is fundamental to us, with 0:17:45.359 --> 0:17:50.200 the importance of doing things in an ethics first way 0:17:50.280 --> 0:17:55.399 to enable the community to incorporate ethical considerations in how 0:17:55.880 --> 0:18:01.120 they're building AI. And we sort of have a different 0:18:01.440 --> 0:18:05.480 audience to start with, which is all the AI builders 0:18:05.600 --> 0:18:10.240 use hiking phase today to access all the models we 0:18:10.320 --> 0:18:14.239 talked about, to use them using our open source and 0:18:14.320 --> 0:18:18.679 build with them. And IBM has this incredible history of 0:18:18.800 --> 0:18:24.320 working with enterprise companies and enabling them to make use 0:18:24.359 --> 0:18:28.360 of that technology in a way that's compliant with everything 0:18:28.440 --> 0:18:32.159 that an enterprise requires, and so being able to marry 0:18:32.200 --> 0:18:36.360 these two things together is an amazing opportunity. And now 0:18:36.400 --> 0:18:40.679 we can enable the largest corporations that have sort of 0:18:40.880 --> 0:18:46.280 complex requirements in order to deploy machine learning systems and 0:18:47.080 --> 0:18:50.480 give them an easy experience to take advantage of all 0:18:50.520 --> 0:18:53.639 the latest and greatest that AI has to offer through 0:18:53.680 --> 0:18:55.760 our platform. 0:18:55.840 --> 0:18:59.439 Let's talk about this idea of a single model or 0:18:59.440 --> 0:19:02.960 a variety of models, because what I've been hearing you say. 0:19:03.520 --> 0:19:05.360 You've been saying, oh, there are lots of models. There 0:19:05.400 --> 0:19:09.480 are hundreds of thousands of models available on hugging Face. 0:19:09.640 --> 0:19:13.000 But you've also said there's a single thing, the transformer, 0:19:13.640 --> 0:19:18.080 and they're all transformers. So if they're all basically the 0:19:18.160 --> 0:19:22.879 same thing, why can't you just build one super clever 0:19:22.920 --> 0:19:24.000 model that can do everything. 0:19:26.119 --> 0:19:31.080 That's a really interesting idea and very much a new idea. 0:19:31.880 --> 0:19:35.800 The reason we have over a million repositories three hundred 0:19:35.840 --> 0:19:40.000 thousand free inaccessible models on a hiking Face platform is 0:19:40.040 --> 0:19:44.159 that models are typically trained to do one thing, and 0:19:44.200 --> 0:19:47.719 they're typically trained to do one thing with specific types 0:19:47.800 --> 0:19:53.960 of data. And what became new and evident in the 0:19:54.040 --> 0:19:56.720 research that came out over the last couple of years 0:19:57.320 --> 0:20:01.080 is that if you train a big enough model with 0:20:01.680 --> 0:20:06.119 enough data, then those models start to have sort of 0:20:06.320 --> 0:20:10.080 general capabilities. You can ask them to do different things, 0:20:10.359 --> 0:20:13.840 you can even train them to respond to instructions. So 0:20:13.960 --> 0:20:18.280 with the same model you can say, hey, summarize this paragraph, 0:20:18.600 --> 0:20:22.119 translate this into English, start a conversation in French, and 0:20:22.359 --> 0:20:25.919 pivot to German. And so these are general sort of 0:20:26.080 --> 0:20:32.600 language capabilities. And I think when ch GBT came online 0:20:33.200 --> 0:20:38.359 and the world sort of discovered these new capabilities, there was, 0:20:38.920 --> 0:20:41.840 at least for a short period, this sort of idea, 0:20:42.000 --> 0:20:45.840 this sort of myth that the endgame of all this 0:20:46.840 --> 0:20:50.600 is maybe one or a handful of models that are 0:20:50.800 --> 0:20:55.000 so much better than anything else than exists, that they 0:20:55.000 --> 0:20:57.640 can do anything that we can ask them to do, 0:20:58.440 --> 0:21:01.919 and that's the only model that we will need. And I, 0:21:02.160 --> 0:21:06.440 for one, think it is a myth. I don't think 0:21:06.480 --> 0:21:10.560 it is practical for a variety of reasons. Say you're 0:21:11.000 --> 0:21:15.159 writing an email and you have like this great suggestion 0:21:15.280 --> 0:21:19.560 of text to sort of complete your sentence. Well, that's AI. 0:21:19.720 --> 0:21:22.520 That's a large language model, that's a transformer model that 0:21:22.560 --> 0:21:25.240 does that. So there are a ton of existing use 0:21:25.240 --> 0:21:28.880 cases like this, and these use cases are powered by 0:21:29.680 --> 0:21:32.640 specific models that have been trained to do one thing 0:21:32.760 --> 0:21:35.840 well and to do it fast. If you wanted to 0:21:36.000 --> 0:21:42.440 apply these sort of all knowing, powerful oracle type of model, 0:21:42.960 --> 0:21:47.000 you would not be able to serve millions of customers 0:21:47.040 --> 0:21:49.720 through a search engine. You will not be able to 0:21:51.359 --> 0:21:55.480 complete people's sentences because the amount of money that you 0:21:55.520 --> 0:21:58.800 would need, the number of computers that you would need 0:21:59.000 --> 0:22:04.000 to run such of of service just exceeds what is 0:22:04.040 --> 0:22:09.960 available on the planet. So one reason for which it's 0:22:10.040 --> 0:22:14.960 not a practical scenario is that it's just very expensive 0:22:15.640 --> 0:22:18.800 to run those very very large models. 0:22:19.160 --> 0:22:21.280 What I'm hearing is it's like, look, if you want 0:22:21.320 --> 0:22:25.040 to screw in a screw you need a screwdriver. You 0:22:25.080 --> 0:22:29.080 don't want an entire tool shed full of tools if 0:22:29.160 --> 0:22:31.359 the task is to screw in a screwdriver, and sure 0:22:31.400 --> 0:22:34.600 you could bring the tool shed that are all the tools. 0:22:34.640 --> 0:22:38.680 There's a screwdriver there, but it's not necessary. It's incredibly expensive, 0:22:38.680 --> 0:22:43.480 it's incredibly cumbersome, and that cost exists even though maybe 0:22:43.560 --> 0:22:46.280 is the user who's just typing in a into a 0:22:46.280 --> 0:22:49.040 prompt box. The user may not see it, but it's 0:22:49.080 --> 0:22:50.040 still very real. 0:22:51.320 --> 0:22:54.840 That's right. And then another one is performance. So taking 0:22:54.920 --> 0:22:58.120 the screwdriver example, so and by the way, like we're 0:22:58.160 --> 0:23:00.919 not quite there at this moment where we have this 0:23:01.080 --> 0:23:04.600 all knowing, powerful oracle that is still sort of a 0:23:04.720 --> 0:23:08.720 sci fi scenario. But we have screwdrivers, but we also 0:23:08.840 --> 0:23:13.439 have the leatherman, right, the multitole Swiss army knife, And 0:23:13.480 --> 0:23:16.280 that's sort of the moment that we are in today. 0:23:16.359 --> 0:23:19.960 But now, if I'm trying to open up my computer, 0:23:20.560 --> 0:23:23.840 turns out that it requires a specific kind of screw 0:23:24.000 --> 0:23:28.159 like these tiny little tork screws, and having a torque 0:23:28.160 --> 0:23:31.560 screw driver will get me much further than trying to 0:23:31.680 --> 0:23:34.760 use my leatherman, where maybe I'll get the knife blade 0:23:34.800 --> 0:23:37.880 and it will mess up the screw and maybe eventually 0:23:37.880 --> 0:23:40.520 I'll get to what I need. But my point is 0:23:40.640 --> 0:23:45.760 that if you take a very specifically trained model for 0:23:45.840 --> 0:23:49.320 a particular problem, it will work much better. It will 0:23:49.320 --> 0:23:54.120 give you better results than a very very generalistic, big 0:23:54.200 --> 0:23:57.159 model that can do a lot of things. And so 0:23:57.240 --> 0:24:01.080 for things like search engines or things things like translation, 0:24:01.440 --> 0:24:06.080 for things that are very specific, companies are much better 0:24:06.160 --> 0:24:11.040 off using smaller, more efficient models that produce better results. 0:24:10.840 --> 0:24:15.359 That's really interesting. And presumably then being able to know 0:24:15.400 --> 0:24:18.160 which model to use, or being able to know who 0:24:18.200 --> 0:24:22.000 to ask which model to use, becomes a very important capability. 0:24:22.840 --> 0:24:26.400 Yes, and that's what we're trying to make easy through 0:24:26.400 --> 0:24:27.159 our platform. 0:24:28.520 --> 0:24:32.520 So tell me about how this works with IBM's what's 0:24:32.520 --> 0:24:36.120 an X platform. How do you see Hugging faces customers 0:24:36.160 --> 0:24:37.040 benefiting from that. 0:24:38.920 --> 0:24:43.000 The end goal is to make it really easy for 0:24:43.119 --> 0:24:47.600 what's and ex customers to make use of all the 0:24:47.680 --> 0:24:51.080 great models and libraries that we talked about, all the 0:24:52.080 --> 0:24:55.600 three hundred thousand models are today on Hugging Face. And 0:24:55.800 --> 0:25:00.040 to do this we need to really collaborate deeply with 0:25:00.040 --> 0:25:03.480 with the IBM teams that build the What's and X 0:25:03.520 --> 0:25:08.720 platform so that our libraries, our open source our models 0:25:09.119 --> 0:25:13.760 are well integrated into the platform. If you're a single user, 0:25:13.760 --> 0:25:16.240 if you are a data science student and you want 0:25:16.280 --> 0:25:19.040 to use a model, is we make it super easy? Right. 0:25:19.080 --> 0:25:21.480 We have our open source library, you can download the 0:25:21.520 --> 0:25:24.119 model on your computer and run with it then. But 0:25:24.640 --> 0:25:31.080 in enterprises there is a vast complexity of infrastructure and 0:25:31.280 --> 0:25:35.560 rules around what people can do and how the data 0:25:35.680 --> 0:25:40.159 can be accessed, and all this complexity is sort of 0:25:40.560 --> 0:25:44.239 solved by the Whatson X platform. 0:25:44.920 --> 0:25:48.840 This season of the Smart Talks podcast features what we're 0:25:48.880 --> 0:25:51.760 calling new creators. Do you see yourself as being a 0:25:51.800 --> 0:25:52.800 creative person. 0:25:54.840 --> 0:25:57.320 I think it's a requirement for the job. I mean, 0:25:57.320 --> 0:26:02.080 we're in such a new and rapidly evolving industry that 0:26:02.359 --> 0:26:06.359 we have to be creative in order to invent the 0:26:06.440 --> 0:26:11.000 business models the use cases of tomorrow. My role within 0:26:11.040 --> 0:26:16.040 the company is really to create the business around all 0:26:16.240 --> 0:26:20.199 the great work of our science and open source and 0:26:20.320 --> 0:26:24.439 product team, and by and large, the business model of 0:26:24.600 --> 0:26:29.560 AI within the whole ecosystem is still something that companies 0:26:29.600 --> 0:26:34.600 are trying to figure out. So creativity is really important 0:26:34.720 --> 0:26:38.520 to really have the conversation with companies, understand what they're 0:26:38.520 --> 0:26:40.600 trying to do, and then build the right kind of solution. 0:26:41.200 --> 0:26:45.400