WEBVTT - Ethics in AI 0:00:12.609 --> 0:00:14.810 Hello, my name is Graham Class and I'm your host 0:00:14.819 --> 0:00:18.099 for this season of technically speaking, an Intel podcast. While 0:00:18.110 --> 0:00:20.940 Intel is at the forefront of so many cutting edge technologies. 0:00:21.000 --> 0:00:23.950 This season is all about artificial intelligence and that's why 0:00:23.959 --> 0:00:27.209 I've been tapped as your host. Having a background in tech, 0:00:27.219 --> 0:00:29.649 as a software engineer. I was always interested in merging 0:00:29.659 --> 0:00:32.810 the advances of artificial intelligence with my love for media 0:00:33.209 --> 0:00:36.259 is culminated in one of my other projects, Daily Dad jokes, 0:00:36.270 --> 0:00:38.949 an A I powered podcast, churning out jokes and humor 0:00:38.959 --> 0:00:40.150 for listeners worldwide. 0:00:40.750 --> 0:00:43.098 But artificial Intelligence can do a lot more than help 0:00:43.110 --> 0:00:46.720 whip up a corny joke. This technology has been revolutionizing 0:00:46.729 --> 0:00:51.049 the way we engage with the world with innovations across healthcare, agriculture, 0:00:51.060 --> 0:00:53.130 business and even the public sector. 0:00:53.750 --> 0:00:56.330 Another way that artificial intelligence is changing the world is 0:00:56.340 --> 0:00:59.689 through philosophy. The term ethical A I is a framework 0:00:59.700 --> 0:01:02.180 on how to use A I, what system should be 0:01:02.189 --> 0:01:04.930 in place to govern its use with business and consumers. 0:01:05.819 --> 0:01:08.220 In this episode, we'll dive into the ethics of artificial 0:01:08.230 --> 0:01:10.760 intelligence with one of the pioneers in the field. 0:01:12.029 --> 0:01:16.349 Joining me for today's conversation is Intel's Ria Chu R 0:01:16.489 --> 0:01:19.149 A can perhaps be described as the moral compass of 0:01:19.160 --> 0:01:22.209 the company's A I as an A I software architect 0:01:22.220 --> 0:01:25.080 and generative A I evangelist. She is charged with finding 0:01:25.089 --> 0:01:29.360 responsible trustworthy solutions for Intel's Internet of Things Engineering group. 0:01:29.949 --> 0:01:33.300 Her role exists at the intersection of hardware and software 0:01:33.389 --> 0:01:35.769 product design and effective consumer use. 0:01:36.750 --> 0:01:39.830 Having studied extensively at Harvard in the subjects of computer 0:01:39.839 --> 0:01:43.069 science and data science. The domains of expertise are solutions 0:01:43.080 --> 0:01:47.010 for security and privacy in machine learning, fairness, explainable and 0:01:47.019 --> 0:01:51.769 responsible A I systems uncertain A I reinforcement learning and 0:01:51.779 --> 0:01:56.349 computational models of intelligence. She is a reoccurring keynote speaker 0:01:56.360 --> 0:01:59.580 on issues in data science and responsible A I. We 0:01:59.589 --> 0:02:01.610 are very excited to have her on the podcast to 0:02:01.620 --> 0:02:04.250 share her expertise on Intel's ethics in their A I 0:02:04.260 --> 0:02:04.769 development 0:02:08.809 --> 0:02:10.789 Ria. Welcome to the show. Thank you, 0:02:10.800 --> 0:02:12.190 Graham. It's awesome to be here. 0:02:12.669 --> 0:02:15.990 I've had a look at your bio and would like 0:02:16.000 --> 0:02:18.309 to know how did you come about to join the 0:02:18.320 --> 0:02:19.329 Intel family? 0:02:19.508 --> 0:02:23.580 Sure, I joined Intel in 2018 when I was 14 0:02:23.589 --> 0:02:26.380 years old as an intern. I had, yes, I had 0:02:26.389 --> 0:02:29.228 an amazing mentor who went through all of the legal 0:02:29.240 --> 0:02:32.448 pages and the review needed to get me to that position. 0:02:32.460 --> 0:02:35.639 So initially, I interviewed with three teams on three different 0:02:35.649 --> 0:02:37.539 areas in the A I space. One of them was 0:02:37.550 --> 0:02:39.699 around A I and healthcare, very theoretic 0:02:40.050 --> 0:02:43.419 and mathematical implications and path finding the other two were 0:02:43.429 --> 0:02:46.860 on software development and profiling. And the next was on 0:02:46.869 --> 0:02:50.478 deep learning optimization specifically. So I did have the opportunity 0:02:50.490 --> 0:02:53.800 to pick the one on optimization for deep learning for hardware. 0:02:53.869 --> 0:02:55.580 And that is how I started off my journey at 0:02:55.589 --> 0:02:59.258 Intel and got introduced to it. The interplay between hardware 0:02:59.270 --> 0:03:02.339 and software is something that always drew my attention. So 0:03:02.350 --> 0:03:04.259 when I was able to work on that as part 0:03:04.270 --> 0:03:06.758 of my first role as an intern, I was really excited. 0:03:07.059 --> 0:03:10.910 OK, great. So now uh I understand that you're a 0:03:10.919 --> 0:03:12.520 software A I architect. 0:03:13.050 --> 0:03:15.880 Can you just give an overview of what that entails 0:03:16.050 --> 0:03:18.570 as a software architect? Today? I have a couple of 0:03:18.580 --> 0:03:21.910 roles and responsibilities corresponding to the latest and greatest, which 0:03:21.919 --> 0:03:24.029 is very exciting to me in my day to day. 0:03:24.119 --> 0:03:26.799 The first is generative A I. So looking at and 0:03:26.809 --> 0:03:30.210 taking into account the different software optimizations that we're planning 0:03:30.220 --> 0:03:33.259 for generative A I, how the workloads are shaping changes 0:03:33.270 --> 0:03:35.839 in the algorithms over time as well as also the 0:03:35.850 --> 0:03:38.759 associated mechanisms that we see that are 0:03:38.854 --> 0:03:41.395 in touch with them as an evangelist. I also get 0:03:41.404 --> 0:03:45.035 to work on top of my software architect role as 0:03:45.044 --> 0:03:48.195 a marketer and an advocate for these technologies. So creating 0:03:48.205 --> 0:03:51.744 very short demos and tutorials for users to quickly grasp 0:03:51.755 --> 0:03:53.895 what exactly is going on with this model. How can 0:03:53.904 --> 0:03:55.434 I use it in my day to day? How can 0:03:55.445 --> 0:03:57.514 I put it to my use case. So a lot 0:03:57.524 --> 0:03:59.475 of the focus today for me is on gender to 0:03:59.514 --> 0:04:01.875 A I, I also look into ethical and explainable A 0:04:01.884 --> 0:04:04.574 I tools and technologies as part of my path finding. 0:04:05.520 --> 0:04:08.380 Yeah, I've been using generative A I apps to do research, 0:04:08.389 --> 0:04:12.899 creating podcast, artwork and experimented with creating music. So this 0:04:12.910 --> 0:04:17.659 leads me into asking you what's your definition of artificial intelligence? 0:04:18.040 --> 0:04:22.390 And maybe some examples of where we're seeing it as 0:04:22.399 --> 0:04:24.660 a central topic in the tech world. 0:04:24.899 --> 0:04:27.469 The way that I like to define it is something 0:04:27.480 --> 0:04:31.000 I copied over actually from our recent regulations on A 0:04:31.010 --> 0:04:35.219 I around how A I models are agents or systems 0:04:35.230 --> 0:04:38.159 that are capable of consuming and producing data in an 0:04:38.170 --> 0:04:41.529 environment and also taking actions that can in turn influence 0:04:41.540 --> 0:04:45.979 our decisions. There's a lot of use cases for them everywhere, healthcare, retail, 0:04:45.988 --> 0:04:46.559 et cetera. 0:04:47.200 --> 0:04:49.950 Yeah. When I talked with uh people even in the 0:04:49.959 --> 0:04:53.789 tech world, there's a lot of confusion around OK, you've 0:04:53.799 --> 0:04:56.678 got algorithms, you've got A I, you've got machine learning 0:04:56.690 --> 0:04:58.880 perhaps if you could start with maybe some of the 0:04:58.890 --> 0:05:02.719 difference between algorithms versus say A I. What do you 0:05:02.730 --> 0:05:04.329 see as the difference between the two 0:05:04.829 --> 0:05:08.239 typical algorithms I'd say are based off of certain schemes 0:05:08.250 --> 0:05:10.600 that we're already aware of with machine learning. You have 0:05:10.609 --> 0:05:13.269 these new paradigms that are coming in and completely spinning 0:05:13.279 --> 0:05:17.488 the narrative, things like continual learning, very large models, different 0:05:17.500 --> 0:05:20.279 types of state machines altogether, depending on the application you 0:05:20.290 --> 0:05:23.299 integrate it into. So I would say there are some 0:05:23.309 --> 0:05:26.630 fundamental differences that are coming in between algorithms and machine 0:05:26.640 --> 0:05:28.470 learning models on that front when it comes to use 0:05:28.480 --> 0:05:31.660 cases application and of course implementation as well. 0:05:32.049 --> 0:05:35.480 And where I see the power is sort of combining 0:05:35.839 --> 0:05:39.709 the traditional sort of if then else algorithms uh with 0:05:39.720 --> 0:05:42.649 A I. And I'm just wondering if you've seen any 0:05:42.660 --> 0:05:46.910 sort of practical applications merging of all these techniques. 0:05:48.260 --> 0:05:50.910 Yes. And I'm very interested in composite A I. It's 0:05:50.920 --> 0:05:53.299 something that I'm getting to work on a lot more 0:05:53.369 --> 0:05:55.179 in my day to day. And something that we're actually 0:05:55.190 --> 0:05:58.190 doing a demo for at intel innovation where we are 0:05:58.200 --> 0:06:01.239 chaining multiple large language models together. The way I see 0:06:01.250 --> 0:06:04.320 composite A I is being able to tie together multiple 0:06:04.329 --> 0:06:05.640 models as part of a 0:06:05.755 --> 0:06:09.255 interface or an application with chaining models. I see it 0:06:09.265 --> 0:06:11.515 as a subset of composite A I where you have 0:06:11.524 --> 0:06:14.184 models that are linked to each other and have dependencies 0:06:14.195 --> 0:06:16.915 on their inputs and outputs. It can be sometimes a 0:06:16.924 --> 0:06:20.294 nightmare to get the dependencies altogether because you have cascading 0:06:20.303 --> 0:06:23.334 models one after the other dependent on each is output, 0:06:23.579 --> 0:06:25.260 but it is possible and it does give you a 0:06:25.269 --> 0:06:28.219 lot of applications and opens up the possibilities where you 0:06:28.230 --> 0:06:31.159 can get to a very nice user interface that users 0:06:31.170 --> 0:06:34.359 can interact with. Developers can build upon businesses and other 0:06:34.369 --> 0:06:37.649 communities can just leverage and adopt that is giving you 0:06:37.660 --> 0:06:40.320 a lot of capabilities at once with ease of deployment. 0:06:40.339 --> 0:06:40.350 Oh, 0:06:40.510 --> 0:06:43.719 that's good. Now, turning to the ethics side of it, 0:06:43.730 --> 0:06:46.470 which you've done quite a lot of thinking and work 0:06:46.480 --> 0:06:50.029 in how would you define ethics in A I 0:06:50.970 --> 0:06:53.390 with ethical A I, the definition that I like to 0:06:53.399 --> 0:06:56.469 adopt is socio technical development of A I systems and 0:06:56.480 --> 0:06:59.779 that involves societal and technical aspects, but really focusing on 0:06:59.790 --> 0:07:01.709 the implications and the intentions with these 0:07:01.720 --> 0:07:04.609 algorithms. In terms of when you're talking with your peers 0:07:04.619 --> 0:07:07.589 and colleagues, it has been a lot of discussion and 0:07:07.600 --> 0:07:10.989 talk about trying to have a uniform ethical framework that 0:07:11.549 --> 0:07:13.899 at least gives a common language into, you know, when 0:07:13.910 --> 0:07:16.200 you're discussing these sorts of things related to ethics in 0:07:16.209 --> 0:07:16.380 A I, 0:07:17.700 --> 0:07:20.570 there are common frameworks that are in place. Most of 0:07:20.579 --> 0:07:23.679 them are centered around implications and intention and how we 0:07:23.690 --> 0:07:27.040 structure that around certain technologies. Right now. It's very popular 0:07:27.049 --> 0:07:30.049 for applications, generative A I where we see these frameworks 0:07:30.059 --> 0:07:32.600 being put into place around, let's look at the inputs, 0:07:32.609 --> 0:07:35.730 the outputs and then the overall modeler framework. And this 0:07:35.739 --> 0:07:36.000 may 0:07:36.105 --> 0:07:38.924 simplistic, but it really is boiled down to these very 0:07:38.934 --> 0:07:41.755 simple elements. Similarly for other A I domains that are 0:07:41.765 --> 0:07:44.524 outside of generative A I like object detection, it's very 0:07:44.535 --> 0:07:47.375 much focused on what is the particular use case? For example, 0:07:47.385 --> 0:07:49.964 is it something that is of high risk like health 0:07:49.975 --> 0:07:52.674 care applications or surveillance or is it something that's a 0:07:52.684 --> 0:07:54.644 bit lower risk like content creation 0:07:54.859 --> 0:07:57.399 and then seeing how exactly our user experience and our 0:07:57.410 --> 0:08:00.359 development of those models is echoing ethical A I principles. 0:08:00.369 --> 0:08:03.619 So I would say like to summarize, there are different 0:08:03.630 --> 0:08:06.440 frameworks and summaries that we apply. But of course, the 0:08:06.450 --> 0:08:08.920 templates need to be flexible when we're talking about ethical 0:08:08.929 --> 0:08:10.660 A I for these new A I models. 0:08:11.119 --> 0:08:14.359 how do you go about ensuring that your staff and 0:08:14.369 --> 0:08:16.850 your engineers and your product managers 0:08:17.510 --> 0:08:22.000 actually embed that ethical framework into its A I development. 0:08:22.290 --> 0:08:25.209 Sure, it's such a challenging problem even to describe as 0:08:25.220 --> 0:08:28.290 well as you're mentioning it, you know, there's so many 0:08:28.299 --> 0:08:30.570 different things that you can actively do, right? Like as 0:08:30.579 --> 0:08:34.489 you mentioned, policies, assessments, et cetera. So at Intel, we 0:08:34.500 --> 0:08:37.809 take a multiple approaches towards it. The one thing that 0:08:37.820 --> 0:08:42.010 we very heavily emphasize on is internal governance. And Laman Nachman, 0:08:42.020 --> 0:08:44.449 who's my mentor and also leading the responsibly 0:08:44.585 --> 0:08:48.726 efforts at Intel very neatly and concisely describes them as 0:08:48.736 --> 0:08:52.074 guardrails that we have internally in place. And these are 0:08:52.085 --> 0:08:56.166 really guidelines that are designed to help our developers, engineers, managers, 0:08:56.255 --> 0:09:00.476 and our communities and marketers, et cetera. Understand the implications 0:09:00.486 --> 0:09:02.676 again of what exactly are we producing in terms of 0:09:02.684 --> 0:09:05.334 the content? What are some technical solutions that we can 0:09:05.346 --> 0:09:09.005 instill mid pipeline or early on before starting the effort 0:09:09.015 --> 0:09:11.406 when we're getting started with A I development efforts 0:09:11.660 --> 0:09:13.562 and I would say that that's the core process that 0:09:13.572 --> 0:09:17.261 we focus on. We're also very heavily invested in technological development, 0:09:17.271 --> 0:09:19.742 whether that's through the deep fake detection work that L 0:09:19.861 --> 0:09:23.221 Deir and team are taking on um explainable A I tools, 0:09:23.231 --> 0:09:25.351 et cetera. So really trying to approach this from a 0:09:25.361 --> 0:09:28.992 governance perspective internally, from a tooling perspective, what we can 0:09:29.002 --> 0:09:32.242 provide to the developer community and our customers and to 0:09:32.252 --> 0:09:35.562 partners and from a third perspective, regulations, how do we 0:09:35.572 --> 0:09:38.642 influence the industry at large and help contribute to discussions 0:09:39.390 --> 0:09:40.189 that's really good. And 0:09:40.840 --> 0:09:43.539 you mentioned the work of Lama Nachman and we're actually 0:09:43.549 --> 0:09:45.539 going to be talking with her in an upcoming episode 0:09:45.549 --> 0:09:48.130 this season. So I'm looking forward to asking her about 0:09:48.140 --> 0:09:50.900 this as well. But I think you've said the key phrase, 0:09:50.909 --> 0:09:53.580 deep fake, so I might switch the to that side 0:09:53.590 --> 0:09:56.150 of things. So in terms of the society and, and 0:09:56.159 --> 0:09:59.489 culture in general, um there's some people that are hesitant 0:09:59.500 --> 0:10:03.539 about A I particularly around A I limiting jobs, you've 0:10:03.549 --> 0:10:07.200 got deep fakes. I've actually created a clone of my voice. 0:10:08.030 --> 0:10:10.940 What do you try and do to reassure people who 0:10:10.950 --> 0:10:12.140 have hesitations? 0:10:12.619 --> 0:10:13.919 I'm definitely not, 0:10:14.659 --> 0:10:18.848 I would say not directly enthusiastic about technologies that are 0:10:18.859 --> 0:10:23.520 allowing for passing off as another person for copying and pasting. 0:10:23.530 --> 0:10:26.718 Essentially in certain cases, we see the development of those 0:10:26.729 --> 0:10:28.890 technologies for a certain use case and then it does 0:10:28.900 --> 0:10:31.520 start to stray away from that into some of these 0:10:31.530 --> 0:10:34.319 newer kind of applications that are scary as you shared. 0:10:34.330 --> 0:10:37.950 So when it comes to reassuring individuals, my family, my 0:10:37.960 --> 0:10:40.900 community as well and the industry at large, I think 0:10:40.909 --> 0:10:43.478 that it's definitely a problem to see in a straightforward way. 0:10:43.580 --> 0:10:47.309 Honestly, without the hype surrounding it, there is a levity 0:10:47.320 --> 0:10:50.320 associated with the disadvantages of the technology that we do 0:10:50.330 --> 0:10:52.950 need to consider. We also do see the benefits of 0:10:52.960 --> 0:10:55.799 them for different things, whether that's improving your ease of 0:10:55.809 --> 0:10:59.059 using it, just being able to communicate with others. From 0:10:59.070 --> 0:11:01.640 my perspective, what I try to do in my space 0:11:01.650 --> 0:11:04.119 is to look at an honest assessment of the technology, 0:11:04.130 --> 0:11:06.159 which is very common in the ethical A I domain. 0:11:06.299 --> 0:11:09.169 And to see what exactly is it really contributing to 0:11:09.210 --> 0:11:11.468 the problem statement and if it isn't contributing to it, 0:11:11.479 --> 0:11:12.358 then do we need it? 0:11:13.099 --> 0:11:18.468 And in terms of intel's I guess method or communication 0:11:18.479 --> 0:11:22.830 with the society and people at large, are they working 0:11:22.840 --> 0:11:24.619 on things to help people? 0:11:25.179 --> 0:11:27.469 I feel a little bit more comfortable about this new 0:11:27.479 --> 0:11:28.579 world we're moving into. 0:11:29.099 --> 0:11:31.659 Yes. And we, we tackle it from a couple of 0:11:31.669 --> 0:11:34.630 different fronts. We've got some amazing teams working on different 0:11:34.640 --> 0:11:38.229 parts of the puzzle. One of them is democratization where 0:11:38.239 --> 0:11:41.359 one of the challenging things about A I from an 0:11:41.369 --> 0:11:43.510 ethical A I perspective, but also in general, 0:11:43.625 --> 0:11:46.655 from a development perspective is being able to give communities 0:11:46.664 --> 0:11:48.885 access to the technology so that they can test it 0:11:48.895 --> 0:11:51.484 and validate it. I've been speaking about ethical A I 0:11:51.494 --> 0:11:54.614 for about two years now or so. Last year, we 0:11:54.625 --> 0:11:56.864 really didn't have the same amount of tools and techniques 0:11:56.875 --> 0:11:59.585 that we have this year and also the popularity of 0:11:59.594 --> 0:12:01.924 testing and validating A I systems, right? 0:12:02.489 --> 0:12:06.179 We always understand and I think many companies and organizations 0:12:06.190 --> 0:12:09.030 understand it's not a one size fits all solution for 0:12:09.039 --> 0:12:12.469 ethical A I. Um you know, many companies and organizations 0:12:12.479 --> 0:12:14.820 are trying to do their best. So I would say 0:12:14.830 --> 0:12:17.530 that again that, that push back that community that we're 0:12:17.539 --> 0:12:19.880 trying to create around ethical A I is critical for 0:12:19.890 --> 0:12:23.140 us going forward to be able to better build solutions. 0:12:23.409 --> 0:12:25.489 Has there been any case studies within intel that you 0:12:25.500 --> 0:12:29.419 could share that maybe there was a real challenging ethical 0:12:29.429 --> 0:12:30.299 conundrum 0:12:30.809 --> 0:12:35.349 uh for producing A I software and you know, how, 0:12:35.359 --> 0:12:36.729 how was it resolved? How did you work 0:12:36.739 --> 0:12:37.250 through it? 0:12:37.280 --> 0:12:39.520 Generative A I is definitely a very big one. So 0:12:39.530 --> 0:12:43.559 we're always actively cautious about the types of implications of 0:12:43.570 --> 0:12:47.030 our technology, whether or not we can incorporate disclaimers or 0:12:47.039 --> 0:12:49.829 clarify on the intent of it as well. And um Graham, 0:12:49.840 --> 0:12:51.440 one of my favorite parts 0:12:51.525 --> 0:12:53.825 of ethical A I from a technical perspective in terms 0:12:53.835 --> 0:12:57.554 of solutions is something called model cards. Model cards, clarify 0:12:57.565 --> 0:13:00.015 a very simple theme around ethical A I which is, 0:13:00.215 --> 0:13:02.744 you know, figure out what exactly is the intention the 0:13:02.755 --> 0:13:05.875 core assumptions and the development that went behind the model 0:13:05.885 --> 0:13:07.585 and what you're going to use it for as part 0:13:07.594 --> 0:13:08.224 of deployment. 0:13:08.510 --> 0:13:10.679 And I think that for me personally, I see that 0:13:10.690 --> 0:13:13.348 that theme is conveyed as part of our efforts in 0:13:13.359 --> 0:13:15.570 generative A I, there's a lot of challenging things out 0:13:15.580 --> 0:13:18.710 there when it comes to image generation, copyright, et cetera 0:13:18.820 --> 0:13:22.630 or even, you know, object detection related technologies for retail. 0:13:22.640 --> 0:13:25.858 If you have solutions like intelligent cue management or automated 0:13:25.869 --> 0:13:28.409 self checkout, it makes sense. But you know, how do 0:13:28.419 --> 0:13:30.330 we keep it from proliferating otherwise? 0:13:30.590 --> 0:13:33.280 And what sort of work is going on with inclusive 0:13:33.289 --> 0:13:33.299 A 0:13:33.309 --> 0:13:34.710 I diversity of state 0:13:35.135 --> 0:13:38.034 is critical for the A I models that we're building today, 0:13:38.044 --> 0:13:41.955 whether that's detection of skin agnostic of skin tone or 0:13:41.965 --> 0:13:45.194 being able to adapt to different folks with different accents. 0:13:45.205 --> 0:13:47.734 So at intel and again, across the industry, I think 0:13:47.744 --> 0:13:49.944 a lot of the efforts are really about making sure 0:13:49.955 --> 0:13:52.275 we have the right people on board, the right experts 0:13:52.284 --> 0:13:55.505 with different backgrounds, we're able to contribute to the technologies. 0:13:55.695 --> 0:13:56.025 One 0:13:56.034 --> 0:13:59.635 thing when I started looking into machine learning very quickly, 0:13:59.645 --> 0:14:01.195 I got a sense of, 0:14:01.559 --> 0:14:04.349 you know, being a traditional engineer, you kind of go OK, 0:14:04.359 --> 0:14:07.010 input output and you kind of know what's in the 0:14:07.020 --> 0:14:10.929 in the black box to transform it. When I started 0:14:10.940 --> 0:14:14.520 working with A I and some machine learning code, I 0:14:14.530 --> 0:14:16.880 couldn't get a sense of that 1 to 1 kind 0:14:16.890 --> 0:14:19.099 of mapping of what the output is to input and 0:14:19.109 --> 0:14:23.190 that comes to the to transparency and uh explainability of 0:14:23.200 --> 0:14:24.140 A I algorithms. 0:14:24.809 --> 0:14:27.070 What are you seeing and also what is intel seeing 0:14:27.080 --> 0:14:30.789 around trying to make that understandable to the end users. 0:14:31.070 --> 0:14:34.169 It's a really interesting question because explainability is one of 0:14:34.179 --> 0:14:36.809 the first topics that we think about when we think 0:14:36.820 --> 0:14:39.780 about responsibly. I and I agree the black box metaphor 0:14:39.789 --> 0:14:43.530 has been used so many times um because it's true. 0:14:43.580 --> 0:14:47.969 But the key idea is about demystifying what exactly is 0:14:47.979 --> 0:14:52.010 going on within the model. Whether that is the internal representation, again, 0:14:52.020 --> 0:14:54.320 the data that it's pulling from how the data is 0:14:54.330 --> 0:14:55.409 being leveraged feature and 0:14:55.502 --> 0:14:59.002 importance, et cetera. There's also an added consideration to explain 0:14:59.013 --> 0:15:01.862 ability around surfacing that to an end user. For them 0:15:01.872 --> 0:15:04.622 to understand why the model made a decision I would 0:15:04.632 --> 0:15:06.783 say with Intel, we're approaching it in a couple of 0:15:06.793 --> 0:15:09.232 different ways. And I'm just, I'm very excited to see 0:15:09.242 --> 0:15:12.393 how can different experts approach our problems. We have a 0:15:12.403 --> 0:15:16.083 dedicated suite of technologies for explainability. I led a team 0:15:16.093 --> 0:15:19.132 that was developing one of these for Intel Cno where again, 0:15:19.143 --> 0:15:19.713 you're getting that 0:15:19.935 --> 0:15:24.455 internal representation analysis, Saliency maps and other technologies for explainability. 0:15:24.585 --> 0:15:28.075 We also incorporate transparency and explainability into our algorithm. So 0:15:28.085 --> 0:15:30.635 whether that's being able to visualize what's going on again, 0:15:30.645 --> 0:15:34.726 saliency maps or you know, really good user experience user 0:15:34.736 --> 0:15:37.526 interface to figure out why am I being surfaced this 0:15:37.536 --> 0:15:40.255 particular prediction or decision from a model? I'd say that's 0:15:40.265 --> 0:15:42.335 a couple of the ways that we are integrating and 0:15:42.346 --> 0:15:44.265 thinking about explainability at Intel. 0:15:46.979 --> 0:15:50.669 You're listening to technically speaking an Intel podcast. We'll be 0:15:50.679 --> 0:15:51.280 right back. 0:16:00.010 --> 0:16:03.309 Welcome back to technically speaking an Intel podcast. 0:16:07.239 --> 0:16:09.179 One of the obviously the big things is around the 0:16:09.190 --> 0:16:13.320 privacy and security of data. Perhaps you could outline some 0:16:13.330 --> 0:16:17.150 of the new techniques and new initiatives out in the 0:16:17.159 --> 0:16:20.059 industry to try and use the power of A I 0:16:20.070 --> 0:16:23.179 but still protect companies, information and and 0:16:23.190 --> 0:16:23.619 data. 0:16:23.869 --> 0:16:27.179 I would say there's mechanisms like differential privacy and many others, 0:16:27.190 --> 0:16:30.440 homomorphic encryption. These were incredibly popular two years ago, you 0:16:30.450 --> 0:16:32.299 kind of don't hear them a lot now. So again, 0:16:32.309 --> 0:16:34.580 the hype is it it depends on the technology of 0:16:34.590 --> 0:16:34.979 the day. 0:16:35.380 --> 0:16:38.520 But yes, localization is a key thing. It's actually something 0:16:38.530 --> 0:16:40.539 I have the opportunity to look at now as part 0:16:40.549 --> 0:16:43.869 of my role around hybrid A I edge versus cloud 0:16:43.880 --> 0:16:47.270 edge and cloud. So there's a number of different parameters 0:16:47.280 --> 0:16:49.809 and assumptions that we can start to make at the 0:16:49.820 --> 0:16:53.599 edge around localization privacy of data, not necessarily having 0:16:53.684 --> 0:16:56.094 to communicate it back to the cloud that are changing 0:16:56.104 --> 0:16:58.205 the way that we think about data privacy and security 0:16:58.215 --> 0:17:01.604 for A I models Federated learning is another paradigm like this. 0:17:01.815 --> 0:17:04.524 So to put it shortly, I'd say there are mechanisms 0:17:04.535 --> 0:17:06.525 that are coming up in place, but there is still 0:17:06.535 --> 0:17:11.125 more needed emphasis on security and privacy, more development for technologies, 0:17:11.135 --> 0:17:11.635 et cetera. 0:17:12.439 --> 0:17:15.280 OK. So just to extend that just a little bit more. 0:17:15.290 --> 0:17:18.079 So say if you're meeting with an executive saying, I've 0:17:18.089 --> 0:17:20.300 been hearing all about large language models and I was 0:17:20.310 --> 0:17:23.379 talking to my colleague uh in another company and they're 0:17:23.390 --> 0:17:26.609 starting to use chatbots with within their organization and using 0:17:26.619 --> 0:17:30.520 the power of that is that related to large language models, 0:17:30.530 --> 0:17:33.438 but fine tuning it to their own corporate data in 0:17:33.449 --> 0:17:34.020 their own 0:17:34.380 --> 0:17:37.339 servers. If you like I sort of on the right track. 0:17:37.520 --> 0:17:40.010 Yes, that is a perfect use case. And thank you 0:17:40.020 --> 0:17:42.910 for bringing that up, you know, centralization of data on 0:17:42.920 --> 0:17:45.839 your server. There's also red teaming um gram that's worth 0:17:45.849 --> 0:17:49.218 mentioning where you're testing your model or your system thoroughly 0:17:49.229 --> 0:17:52.689 with the generative A I space. There's come to life, 0:17:52.699 --> 0:17:55.040 a lot of different types of red teaming approaches including 0:17:55.050 --> 0:17:57.500 prompt injection and many others, which is really a 0:17:57.574 --> 0:18:00.574 being able to test and mock the kinds of inputs 0:18:00.584 --> 0:18:02.994 that adversaries would provide to your model and figure out 0:18:03.005 --> 0:18:05.055 how the model is going to behave. What are its 0:18:05.064 --> 0:18:07.875 strengths and weaknesses, et cetera. Of course, the compute needed 0:18:07.885 --> 0:18:10.755 for that is another story. But in addition to that, 0:18:10.765 --> 0:18:13.334 there's also again, the testing and validation approaches. So red 0:18:13.344 --> 0:18:17.505 teaming is really critical to that validating how susceptible your 0:18:17.515 --> 0:18:20.675 model is to potential attacks, whether it's biased, et cetera. 0:18:20.974 --> 0:18:23.905 So lots of, lots of cool and interesting approaches coming up. 0:18:23.915 --> 0:18:25.954 Exactly as you noted, that's a key example. 0:18:26.290 --> 0:18:29.670 So going back on the ethics side of things, what 0:18:29.680 --> 0:18:33.909 are some of the arguments for a corporation, an organization 0:18:33.920 --> 0:18:37.619 to have a clear set of code of ethics and 0:18:37.630 --> 0:18:42.589 is intel helping companies establish those sorts of guidelines and frameworks. 0:18:43.550 --> 0:18:46.319 There is a number of different best practices that organizations 0:18:46.329 --> 0:18:49.329 can incorporate today for responsible A I. One of them 0:18:49.339 --> 0:18:52.369 is the internal governance assessments that we talked about, which 0:18:52.380 --> 0:18:55.079 is a step by step process to checking where A 0:18:55.089 --> 0:18:57.189 I is used in your organization. How is it being 0:18:57.199 --> 0:19:00.310 shipped outside? What's your go to market strategy? What's your 0:19:00.319 --> 0:19:02.069 change management strategy, etcetera. 0:19:02.349 --> 0:19:05.889 So in terms of Intel's contributions, we're very excited and 0:19:05.900 --> 0:19:10.219 passionate about communication with customers and partners and communities in 0:19:10.229 --> 0:19:13.790 general around. What exactly can we do to help with 0:19:13.800 --> 0:19:16.290 the ethical A I development that can include, you know, 0:19:16.300 --> 0:19:20.810 potential compute platforms that help with running this type of solutions, preprocessing, 0:19:20.819 --> 0:19:22.369 post processing. What exactly 0:19:22.494 --> 0:19:25.275 you need towards that? Or if we have developers working 0:19:25.286 --> 0:19:27.125 with Intel Open Veno and I work in the Open 0:19:27.176 --> 0:19:29.315 Vino team right now. We want to know what makes 0:19:29.326 --> 0:19:31.705 it easier for developers to be able to run these 0:19:31.715 --> 0:19:34.984 models and deploy them their feedback in terms of, you know, hey, 0:19:34.994 --> 0:19:36.955 you know, is this challenging to use? I don't know 0:19:36.965 --> 0:19:39.215 how this is working. Um something that I do as 0:19:39.225 --> 0:19:41.955 part of my evangelism team is again, helping contribute to that. 0:19:41.965 --> 0:19:42.546 So I would 0:19:42.641 --> 0:19:45.131 say that as part of the practices, there's a number 0:19:45.141 --> 0:19:48.592