WEBVTT - Scaling AI With Purpose 0:00:03.120 --> 0:00:06.600 Hello, Hello, Welcome to Smart Talks with IBM, a podcast 0:00:06.680 --> 0:00:13.000 from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. This season, 0:00:13.000 --> 0:00:16.159 we're diving back into the world of artificial intelligence, but 0:00:16.239 --> 0:00:22.240 with a focus on the powerful concept of open its possibilities, implications, 0:00:22.280 --> 0:00:26.320 and misconceptions. We'll look at openness from a variety of 0:00:26.360 --> 0:00:30.200 angles and explore how the concept is already reshaping industries, 0:00:30.640 --> 0:00:34.199 ways of doing business and our very notion of what's possible. 0:00:35.320 --> 0:00:39.239 In today's episode, Jacob Goldstein sits down with Rebecca Finley, 0:00:39.520 --> 0:00:43.720 the CEO of the Partnership on Ai, a nonprofit group 0:00:44.000 --> 0:00:48.040 grappling with important questions around the future of AI. Their 0:00:48.080 --> 0:00:52.120 conversation focuses on Rebecca's work bringing together a community of 0:00:52.240 --> 0:00:58.000 diverse stakeholders to help shape the conversation around accountable AI governance. 0:00:58.880 --> 0:01:02.960 Rebecca explains why transparency is so crucial for scaling the 0:01:03.000 --> 0:01:07.280 technology responsibly, and she highlights how working with groups like 0:01:07.319 --> 0:01:10.960 the AI Alliance can provide valuable insights in order to 0:01:10.959 --> 0:01:17.560 build the resources, infrastructure, and community around releasing open source models. So, 0:01:17.760 --> 0:01:21.000 without further ado, let's get to that conversation. 0:01:28.120 --> 0:01:29.800 Can you just say your name? And your job. 0:01:30.280 --> 0:01:33.480 My name is Rebecca Finley. I am the CEO of 0:01:33.520 --> 0:01:37.240 the Partnership on AI to Benefit People and Society, often 0:01:37.280 --> 0:01:39.160 referred to as PAI. 0:01:40.080 --> 0:01:42.640 How did you get here? What was your job before 0:01:42.680 --> 0:01:44.559 you have the job that you have now? 0:01:45.480 --> 0:01:50.400 I came to PAI about three years ago, having had 0:01:50.440 --> 0:01:55.640 the opportunity to work for the Canadian Institute for Advance Research, 0:01:56.200 --> 0:02:00.920 developing and deploying all of their programs related to the 0:02:00.960 --> 0:02:06.480 intersection of technology and society. And one of the areas 0:02:06.720 --> 0:02:11.040 that the Canadian Institute had been funding since nineteen eighty 0:02:11.080 --> 0:02:14.919 two was research into artificial intelligence. 0:02:15.160 --> 0:02:17.160 Wow early, they were early. 0:02:18.240 --> 0:02:22.480 It was a very early commitment and an ongoing commitment 0:02:22.680 --> 0:02:28.040 at the Institute to fund long term fundamental questions of 0:02:28.120 --> 0:02:37.040 scientific importance in interdisciplinary research programs that were often committed 0:02:37.080 --> 0:02:40.600 and funded to for well over a decade. The AI 0:02:40.800 --> 0:02:43.840 Robotics and Society program that kicked off the work at 0:02:43.919 --> 0:02:49.640 the Institute eventually became a program very much focused on 0:02:50.240 --> 0:02:55.360 deep learning and reinforcement learning, neural networks. All of the 0:02:56.040 --> 0:03:00.560 current iteration of AI, or certainly the pregenerative of AI 0:03:00.800 --> 0:03:05.640 iteration of AI that led to this transformation that we've 0:03:05.680 --> 0:03:09.040 seen in terms of online search and all sorts of 0:03:09.040 --> 0:03:12.040 ways in which predictive AI has been deployed. So I 0:03:12.080 --> 0:03:15.400 had the opportunity to see the very early days of 0:03:15.480 --> 0:03:19.919 that research coming together, and when in the early sort 0:03:19.960 --> 0:03:25.360 of two thousand, twenty and tens, when compute capability came 0:03:25.400 --> 0:03:29.960 together with data capability through some of the Internet companies 0:03:29.960 --> 0:03:33.960 and otherwise, and we really saw this technology start to 0:03:34.000 --> 0:03:37.000 take off. I had the opportunity to start up a 0:03:37.040 --> 0:03:42.160 program specifically focused on the impacts of AI in society. 0:03:42.840 --> 0:03:45.600 There was, as you know, at that time, some concerns 0:03:45.640 --> 0:03:50.120 both about the potential for the technology, but also in 0:03:50.240 --> 0:03:52.800 terms of what we were seeing around data sets and 0:03:53.080 --> 0:03:57.360 bias and discrimination and potential impact on future jobs. And 0:03:57.440 --> 0:04:01.600 so bringing a whole group of experts, whether they were 0:04:01.640 --> 0:04:07.400 ethicists or lawyers or economists sociologists into the discussion about 0:04:07.440 --> 0:04:10.880 AI was core to that new program and continues to 0:04:10.920 --> 0:04:14.600 be core to my commitment to bringing diverse perspectives together 0:04:14.680 --> 0:04:18.520 to solve the challenges and opportunities that AI offers today. 0:04:19.600 --> 0:04:22.159 So specifically, what is your job now? What is the 0:04:22.200 --> 0:04:24.240 work you do? What is the work that PAI does? 0:04:25.320 --> 0:04:29.160 I like to answer that question by asking two questions, 0:04:29.640 --> 0:04:33.000 First and foremost, do you believe that the world is 0:04:33.240 --> 0:04:37.159 more divided today than it ever has been in recent history? 0:04:38.040 --> 0:04:41.839 And do you believe that if we don't create spaces 0:04:42.440 --> 0:04:45.760 for very different perspectives to come together, we won't be 0:04:45.839 --> 0:04:48.599 able to solve the challenges that are in front of 0:04:48.600 --> 0:04:52.560 the world today. My answer to both of those questions is, yes, 0:04:53.040 --> 0:04:56.599 we're more divided, and two, we need to seek out 0:04:56.760 --> 0:05:01.479 those spaces where those very different perspectives can come together 0:05:02.000 --> 0:05:05.200 to solve those great challenges. And that's what I get 0:05:05.240 --> 0:05:08.760 to do as CEO of the Partnership on AI. We 0:05:08.760 --> 0:05:13.479 were begun in twenty sixteen with a fundamental commitment to 0:05:13.720 --> 0:05:20.280 bringing together experts, whether they were in industry, academia, civil society, 0:05:20.360 --> 0:05:24.600 or philanthropy, coming together to identify what are the most 0:05:24.640 --> 0:05:28.359 important questions when we think about developing AI centered on 0:05:28.440 --> 0:05:31.800 people and communities, and then how do we begin to 0:05:31.920 --> 0:05:34.760 develop the solutions to make sure we benefit appropriately. 0:05:35.760 --> 0:05:41.039 So that's a very big picture set of ideas. I'm 0:05:41.080 --> 0:05:43.400 curious on a sort of more day to day level. 0:05:43.400 --> 0:05:45.719 I mean, you talk about collaborating with all these different 0:05:45.839 --> 0:05:48.160 kinds of people, all these different groups, what does that 0:05:48.200 --> 0:05:50.920 actually look like, what are some specific examples of how 0:05:50.960 --> 0:05:51.839 you do this work? 0:05:52.360 --> 0:05:55.920 So right now we have about one hundred and twenty 0:05:56.040 --> 0:06:01.640 partners in sixteen countries. They come together through working groups 0:06:01.720 --> 0:06:04.640 that we look at through a variety of different perspectives. 0:06:04.680 --> 0:06:08.520 It could be AI, labor and the economy. It could 0:06:08.600 --> 0:06:13.280 be how do you build a healthy information ecosystem. It 0:06:13.320 --> 0:06:16.520 could be how do you bring more diverse perspectives into 0:06:16.640 --> 0:06:21.039 the inclusive and equitable development of AI. It could be 0:06:21.200 --> 0:06:25.480 what are the emerging opportunities with these very very large 0:06:25.520 --> 0:06:29.440 foundation model applications and how do you deploy those safely? 0:06:29.920 --> 0:06:33.479 And these groups come together most importantly to say what 0:06:33.600 --> 0:06:37.040 are the questions we need to answer collectively, So they 0:06:37.040 --> 0:06:39.840 come together in working groups. I have an amazing staff 0:06:39.920 --> 0:06:43.440 team who hold the pen on synthesizing research and data 0:06:43.800 --> 0:06:49.360 and evidence, developing frameworks, best practices, resources, all sorts of 0:06:49.400 --> 0:06:51.839 things that we can offer up to the community, be 0:06:51.960 --> 0:06:55.599 they in industry or in policy, to say this is 0:06:55.680 --> 0:06:58.360 how we can well, this is what good looks like, 0:06:58.440 --> 0:06:59.720 and this is how we can do it on a 0:06:59.800 --> 0:07:01.880 day to day basis. So that's what we do, and 0:07:01.920 --> 0:07:05.240 then we publish our materials. It's all open. We make 0:07:05.320 --> 0:07:07.520 sure that we get them into the hands of those 0:07:07.560 --> 0:07:10.240 communities that can use them, and then we drive and 0:07:10.320 --> 0:07:12.840 work with those communities to put them into practice. 0:07:13.480 --> 0:07:16.480 You used the word to open there and describing your publications. 0:07:17.400 --> 0:07:20.120 I know, in the world of AI, on the sort 0:07:20.160 --> 0:07:24.880 of technical side, there's a debate, say, or discussion about 0:07:24.920 --> 0:07:29.640 kind of open versus closed AI, And I'm curious how 0:07:29.760 --> 0:07:33.560 you kind of encounter that particular discussion. What is your 0:07:33.640 --> 0:07:35.360 view on open versus closed AI. 0:07:36.440 --> 0:07:42.240 So the current discussion between open and closed release of 0:07:42.400 --> 0:07:47.640 AI models came once we saw chat, GPT and other 0:07:47.920 --> 0:07:52.280 very large generative AI systems being deployed out into the 0:07:52.360 --> 0:07:58.040 hands of consumers around the world, and there emerged some 0:07:58.200 --> 0:08:02.960 fear about the potential of these models to act in 0:08:03.040 --> 0:08:06.560 all sorts of catastrophic ways. So there were concerns that 0:08:06.600 --> 0:08:11.640 the models could be deployed with regard to different development 0:08:11.680 --> 0:08:17.000 of viruses or biomedical weapons or even nuclear weapons, or 0:08:17.040 --> 0:08:21.120 through manipulation or otherwise. So this are emerged about over 0:08:21.160 --> 0:08:26.960 the last eighteen months, this real concern that these models, 0:08:27.080 --> 0:08:31.400 if deployed openly, could lead to some level of truly 0:08:31.520 --> 0:08:37.160 catastrophic risk. And what emerged is actually that we discovered 0:08:37.720 --> 0:08:39.559 that through a whole bunch of work that's been done 0:08:39.600 --> 0:08:42.960 over the last little while, that releasing them openly has 0:08:43.000 --> 0:08:45.360 not led and doesn't appear to be leading in any 0:08:45.400 --> 0:08:50.640 way to catastrophic risk. In facts, releasing them openly allows 0:08:50.679 --> 0:08:55.439 for much more greater scrutiny and understanding of the safety 0:08:55.480 --> 0:08:58.080 measures that have been put into place, And so what 0:08:58.240 --> 0:09:01.880 happened was sort of the pendulum swamp very much towards 0:09:01.920 --> 0:09:05.480 concerned about really catastrophic risk and safety over the last year, 0:09:05.520 --> 0:09:08.120 and over the last year we've seen it swing back 0:09:08.200 --> 0:09:10.880 as we learn more and more about how these models 0:09:11.200 --> 0:09:13.760 are being used and how they are being deployed into 0:09:13.800 --> 0:09:19.760 the world. My feeling is we must approach this work openly, 0:09:20.160 --> 0:09:23.439 and it's not just open release of models or what 0:09:23.480 --> 0:09:27.400 we think of as traditional open source forms of model 0:09:27.480 --> 0:09:30.600 development or otherwise, but we really need to think about 0:09:30.679 --> 0:09:35.400 how do we build an open innovation ecosystem that fundamentally 0:09:35.440 --> 0:09:39.520 allows both for the innovation to be shared with many people, 0:09:39.600 --> 0:09:43.320 but also for safety and security to be rigorously upheld. 0:09:43.760 --> 0:09:47.199 So when you talk about this kind of broader idea 0:09:47.240 --> 0:09:51.840 of open innovation beyond open source or you know, transparency 0:09:51.840 --> 0:09:55.119 in models like what do you mean sort of specifically, 0:09:55.160 --> 0:09:56.680 how does that look in the world. 0:09:57.040 --> 0:10:00.959 So I have three particular points view when it comes 0:10:01.000 --> 0:10:03.520 to open innovation, because I think we need to think 0:10:03.720 --> 0:10:07.400 both upstream around the research that is driving these models, 0:10:07.679 --> 0:10:10.559 and downstream in terms of the benefits of these models 0:10:10.559 --> 0:10:14.200 to others. So first and foremost, what we have known 0:10:14.400 --> 0:10:16.720 in terms of how AI has been developed, and yes, 0:10:16.800 --> 0:10:18.680 I had an opportunity to see it when I was 0:10:18.720 --> 0:10:22.080 at the Canadian Institute for Advanced Research is a very 0:10:22.200 --> 0:10:28.120 open form of scientific publication and rigorous peer review. And 0:10:28.240 --> 0:10:31.240 what happens when we release openly is you have an 0:10:31.240 --> 0:10:35.440 opportunity for the research to be interrogated to determine the 0:10:35.559 --> 0:10:38.840 quality and significance of that, but then also for it 0:10:38.880 --> 0:10:41.559 to be picked up by many others. And then secondly, 0:10:42.040 --> 0:10:46.120 openness for me is about transparency. We released a set 0:10:46.160 --> 0:10:49.720 of very strong recommendations last year around the way in 0:10:49.760 --> 0:10:54.040 which these very large foundation models could be deployed safely. 0:10:54.679 --> 0:10:58.760 They're all about disclosure. They're all about disclosure and documentation 0:10:58.920 --> 0:11:02.199 right from the early days pre R and D development 0:11:02.240 --> 0:11:04.720 of these systems, right in terms of thinking about what's 0:11:04.760 --> 0:11:07.319 in the training data and how is it being used 0:11:07.480 --> 0:11:11.480 all the way through to post deployment monitoring and disclosure. 0:11:12.040 --> 0:11:15.280 So I really think that this is important transparency through it. 0:11:15.320 --> 0:11:18.240 And then the third piece is openness in terms of 0:11:18.280 --> 0:11:21.640 who is around the table to benefit from this technology. 0:11:22.080 --> 0:11:24.200 We know that if we're really going to see these 0:11:24.240 --> 0:11:29.280 new models having being successful deployed into education or healthcare 0:11:29.440 --> 0:11:33.319 or climate and sustainability, we need to have those experts 0:11:33.320 --> 0:11:36.480 in those communities at the table charting this and making 0:11:36.520 --> 0:11:39.079 sure that the technology is working for them. So those 0:11:39.080 --> 0:11:40.880 are the three ways I think about openness. 0:11:41.960 --> 0:11:45.400 Is there like a particular project that you've worked on 0:11:45.480 --> 0:11:49.440 that you feel like you know reflects your approach to 0:11:49.600 --> 0:11:50.600 responsible AI. 0:11:51.640 --> 0:11:54.600 So there's a really interesting project that we have underway 0:11:54.600 --> 0:11:58.800 at PAI that is looking at responsible practices squarely when 0:11:58.840 --> 0:12:02.360 it comes to the use of synthetic media. And what 0:12:02.400 --> 0:12:06.000 we heard from our community was that they were looking 0:12:06.040 --> 0:12:09.600 for a clear code of conduct about what does it 0:12:09.679 --> 0:12:12.600 mean to be responsible in this space. And so what 0:12:12.800 --> 0:12:15.800 happened is we pulled together a number of working groups 0:12:15.840 --> 0:12:19.679 to come together. They included industry representatives. They also included 0:12:20.000 --> 0:12:25.760 civil society organizations like WITNESS, a number of academic institutions 0:12:25.760 --> 0:12:28.840 and otherwise, And what we heard was that there were 0:12:29.120 --> 0:12:35.000 clear requirements that creators could take, that developers of the 0:12:35.080 --> 0:12:38.160 technology could take, and then also distributors. So when we 0:12:38.200 --> 0:12:42.719 think about those generative AI systems being deployed across platforms 0:12:42.720 --> 0:12:45.920 and otherwise, and we came up with a framework for 0:12:45.960 --> 0:12:49.360 what responsibility looks like. What does it mean to have consent, 0:12:49.520 --> 0:12:53.120 what does it mean to disclose responsibly, what does it 0:12:53.200 --> 0:12:57.440 mean to embed technology into it? So, for example, we've 0:12:57.440 --> 0:13:00.520 heard many people talk about the importance of water marking 0:13:00.600 --> 0:13:02.920 systems right and making sure that we have a way 0:13:03.000 --> 0:13:05.240 to water mark them. But what we know from the 0:13:05.320 --> 0:13:09.560 technology is that is a very very complex and complicated problem, 0:13:09.800 --> 0:13:13.040 and what might work on a technical level certainly hits 0:13:13.040 --> 0:13:16.040 a whole new set of complications when we start labeling 0:13:16.080 --> 0:13:19.200 and disclosing out to the public about what that technology 0:13:19.240 --> 0:13:22.840 actually means. All of these, I believe are solvable problems, 0:13:22.840 --> 0:13:26.480 but they all needed to have a clear code underneath 0:13:26.520 --> 0:13:28.720 them that was saying this is what we will commit to. 0:13:29.040 --> 0:13:32.280 And we now have a number of organizations, many many 0:13:32.360 --> 0:13:35.600 of the large technology companies, but also many of the 0:13:36.080 --> 0:13:39.200 small startups who are operating in this based civil society 0:13:39.200 --> 0:13:43.040 and media organizations like the BBC and the CBC who's 0:13:43.080 --> 0:13:46.559 have signed on. And one of the really exciting pieces 0:13:46.679 --> 0:13:50.559 of that is that we're now seeing how it's changing practice. 0:13:50.800 --> 0:13:53.559 So a year in we asked each of our partners 0:13:53.600 --> 0:13:56.800 to come up with a clear case study about how 0:13:56.840 --> 0:14:00.000 that work has changed the way they are making decisions, 0:14:00.480 --> 0:14:05.079 deploying technology and ensuring that they're being responsible in their use. 0:14:05.120 --> 0:14:08.040 And that is creating now a whole resource online that 0:14:08.080 --> 0:14:10.440 we're able to share with others about what does it 0:14:10.520 --> 0:14:13.840 mean to be responsible in this place. There's so much 0:14:13.880 --> 0:14:16.040 more work to be done, and the exciting thing is 0:14:16.080 --> 0:14:18.240 once you have a foundation like this in place, we 0:14:18.320 --> 0:14:21.880 can continue to build on it. So much interest now 0:14:21.920 --> 0:14:24.880 in the policy space, for example, about this work as well. 0:14:25.960 --> 0:14:29.760 Are there any specific examples of those sort of case 0:14:29.800 --> 0:14:34.520 studies or the real world experiences that say media organizations 0:14:34.560 --> 0:14:37.520 had that are interesting that are illuminating. Yes. 0:14:37.840 --> 0:14:43.040 So, for example, what we saw with the BBC is 0:14:43.080 --> 0:14:47.000 that they're developing a lot of content as a public broadcaster, 0:14:47.160 --> 0:14:50.240 both in terms of their news coverage but also in 0:14:50.320 --> 0:14:52.960 terms of some of the resources that they are developing 0:14:53.720 --> 0:14:56.480 for the British public as well. And what they talked 0:14:56.480 --> 0:14:59.720 about was the way in which they had used synthetic 0:14:59.760 --> 0:15:04.600 meat in a very very sensitive environment where they were 0:15:04.840 --> 0:15:09.440 hearing from individuals talk about personal experiences, but wanted to 0:15:09.480 --> 0:15:13.040 have some way to change the face entirely in terms 0:15:13.080 --> 0:15:16.240 of the individuals who were speaking. So that's a very 0:15:16.280 --> 0:15:19.920 complicated ethical question, right, how do you do that responsibily 0:15:20.080 --> 0:15:23.160 and what is the way in which you use that technology, 0:15:23.520 --> 0:15:26.640 and most importantly, how do you disclose it? So their 0:15:26.720 --> 0:15:29.800 case study looked at that in some real detail about 0:15:29.840 --> 0:15:33.800 the process they went through to make the decision responsibly 0:15:33.920 --> 0:15:36.840 to do what they chose, how they intended to use 0:15:36.880 --> 0:15:38.240 the technology in that space. 0:15:39.000 --> 0:15:41.960 As you describe your work in some of these studies, 0:15:42.160 --> 0:15:47.160 the idea of transparency seems to be a theme. Talk 0:15:47.200 --> 0:15:49.560 about the importance of transparency in this kind of work. 0:15:50.680 --> 0:15:55.640 Yeah, transparency is fundamental to responsibility. I always like to 0:15:55.680 --> 0:15:59.480 say it's not accountability in a complete sense, but it 0:15:59.560 --> 0:16:03.520 is a first step to driving accountability more fully, so, 0:16:04.040 --> 0:16:07.320 when we think about how these systems are developed, they're 0:16:07.320 --> 0:16:12.560 often developed behind closed doors inside companies who are making 0:16:12.640 --> 0:16:16.680 decisions about what and how these products will work from 0:16:16.680 --> 0:16:21.920 a business perspective, and what disclosure and transparency can provide 0:16:22.000 --> 0:16:25.360 is some sense of the decisions that were made leading 0:16:25.440 --> 0:16:28.239 up to the way in which those models were deployed. 0:16:28.320 --> 0:16:33.720 So this could be ensuring that individual's private information was 0:16:33.760 --> 0:16:38.080 protected through the process and won't be inadvertently disclosed, or otherwise, 0:16:38.520 --> 0:16:41.560 it could be providing some sense of how well the 0:16:41.600 --> 0:16:45.320 system performs against a whole level of quality measures. So 0:16:45.400 --> 0:16:48.160 we have all of these different types of evaluations and 0:16:48.240 --> 0:16:51.520 a measures that are emerging about the quality of these 0:16:51.560 --> 0:16:55.400 systems as they're deployed. Being transparent about how they perform 0:16:55.480 --> 0:16:58.560 against these systems is really crucial to that as well. 0:16:58.840 --> 0:17:01.920 We have a whole ecosis that's starting to emerge around 0:17:02.000 --> 0:17:04.960 auditing of these systems. So what does that look like 0:17:05.080 --> 0:17:07.480 we think about auditors and all sorts of other sectors 0:17:07.480 --> 0:17:10.000 of the economy. What does it look like to be 0:17:10.080 --> 0:17:13.240 auditing these systems to ensure that they're meeting all of 0:17:13.280 --> 0:17:16.960 those both legal but additional ethical requirements that we want 0:17:17.040 --> 0:17:18.200 to make sure that are in place. 0:17:19.520 --> 0:17:24.040 What are some of the hardest ethical dilemmas you've come 0:17:24.119 --> 0:17:26.240 up against in AI policy. 0:17:27.480 --> 0:17:30.840 Well, the interesting thing about AI policy right is what 0:17:30.880 --> 0:17:35.119 it works very simply in one setting can be highly 0:17:35.200 --> 0:17:38.520 complicated in another setting. And so, for example, I have 0:17:38.600 --> 0:17:41.439 an app that I adore. It's an app on my 0:17:41.600 --> 0:17:44.199 phone that allows me to take a photo of a 0:17:44.240 --> 0:17:47.560 bird and it will help me to better understand what 0:17:47.640 --> 0:17:50.280 that bird is and give me all sorts of information 0:17:50.359 --> 0:17:54.520 about that bird. Now, it's probably right most of the time, 0:17:54.600 --> 0:17:56.880 and it's certainly right enough of the time to give 0:17:56.920 --> 0:18:00.159 me great pleasure and delight when I'm out walking. You 0:18:00.200 --> 0:18:04.520 could think about that exact same technology applied. So for example, 0:18:04.600 --> 0:18:07.840 now you're a security guard and you're working in a 0:18:07.960 --> 0:18:12.000 shopping plaza, and you're able to take photos of individuals 0:18:12.040 --> 0:18:15.040 who you may think are acting suspiciously in some way 0:18:15.119 --> 0:18:17.840 and match that photo up with some sort of a 0:18:18.040 --> 0:18:21.679 database of individuals that may have been found, you know, 0:18:21.800 --> 0:18:25.119 to have some sort of connection to other criminal behavior 0:18:25.119 --> 0:18:27.639 in the past. Right, So what goes from being a 0:18:27.680 --> 0:18:30.840 delightful Oh, isn't this an interesting bird? To a very 0:18:31.000 --> 0:18:35.600 very creepy What does this say about surveillance and privacy 0:18:35.720 --> 0:18:39.000 and access to public spaces? And that is the nature 0:18:39.160 --> 0:18:42.520 of AI. So much of the concern about the ethical 0:18:42.720 --> 0:18:48.520 use and deployment of AI is how an organization is 0:18:48.600 --> 0:18:53.960 making the choices within the social and systemic structure they sit. 0:18:54.200 --> 0:18:57.960 So so much about the ethics of AI is understanding 0:18:58.000 --> 0:19:00.959 what is the use case, how is it being used, 0:19:01.080 --> 0:19:04.560 how is it being constrained? How does it start to 0:19:04.760 --> 0:19:08.159 infringe upon what we think of as the human rights 0:19:08.160 --> 0:19:12.240 of an individual to privacy? And so you have to 0:19:12.359 --> 0:19:15.640 constantly be thinking about ethics. What could work very well 0:19:15.640 --> 0:19:19.240 in one situation absolutely doesn't work in another. We often 0:19:19.280 --> 0:19:23.200 talk about these as socio technical questions. Right, just because 0:19:23.240 --> 0:19:26.640 the technology works doesn't actually mean that it should be 0:19:27.040 --> 0:19:28.080 used and deployed. 0:19:28.840 --> 0:19:33.840 What's an example of where the partnership on AI influence 0:19:34.080 --> 0:19:38.399 changes either in policy or in industry practice. 0:19:38.720 --> 0:19:41.399 We talked a little bit about the Framework for Synthetic 0:19:41.480 --> 0:19:45.880 Media and how that has allowed companies and media organizations 0:19:45.880 --> 0:19:49.080 and civil society organizations to really think deeply about the 0:19:49.080 --> 0:19:51.960 way in which they're using this. Another area that we 0:19:52.119 --> 0:19:58.239 focused on has been around responsible deployment of foundation on 0:19:58.320 --> 0:20:01.479 large scale models. I said, we issued a set of 0:20:01.560 --> 0:20:06.359 recommendations last year that really laid out for these very 0:20:06.480 --> 0:20:11.000 large developers and deployers of foundation and frontier models were 0:20:11.520 --> 0:20:14.760 what does good look like? Right from R and D 0:20:15.000 --> 0:20:18.880 through to deployment monitoring, and it has been very encouraging 0:20:19.040 --> 0:20:21.840 to see that that work has been picked up by 0:20:22.240 --> 0:20:26.359 companies and really articulated as part of the fabric of 0:20:26.400 --> 0:20:31.200 the deployment of their foundation models and systems moving forward. 0:20:31.600 --> 0:20:34.680 So much of this work is around creating clear definitions 0:20:34.680 --> 0:20:37.960 of what we're meaning as the technology evolves and clear 0:20:38.000 --> 0:20:40.520 sets of responsibilities. So it's great to see that work 0:20:40.520 --> 0:20:44.000 getting picked up. The NTIA in the United States just 0:20:44.080 --> 0:20:48.439 released a report on open models and the release of 0:20:48.480 --> 0:20:51.159 open models. Great to see our work cited there as 0:20:51.240 --> 0:20:54.560 contributing to that analysis. Great to see some of our 0:20:54.560 --> 0:20:58.520 definitions and synthetic media getting picked up by legislators in 0:20:58.560 --> 0:21:03.159 different countries. It's important, i think, for us to build capacity, 0:21:03.200 --> 0:21:06.199 knowledge and understanding and our policy makers in this moment 0:21:06.680 --> 0:21:10.960 as the technology is evolving and accelerating in its development. 0:21:12.080 --> 0:21:15.680 What's the AI Alliance and why did Partnership on AI 0:21:15.760 --> 0:21:16.520 decide to join? 0:21:17.080 --> 0:21:20.600 So you had asked about the debate between open versus 0:21:20.720 --> 0:21:25.200 closed models and how that has evolved over the last year, 0:21:25.560 --> 0:21:30.040 and the AI Alliance was a community of organizations that 0:21:30.200 --> 0:21:34.480 came together to really think about, okay, if we support 0:21:34.960 --> 0:21:38.479 open release of models what does that look like and 0:21:38.520 --> 0:21:41.000 what does the community need? And so that's about one 0:21:41.119 --> 0:21:45.680 hundred organizations. IBM, one of our founding partners, is also 0:21:45.760 --> 0:21:48.960 one of the founding partners of the AI Alliance. It's 0:21:49.000 --> 0:21:53.000 a community that brings together a number of academic institutions 0:21:53.400 --> 0:21:56.800 many countries around the world, and they're really focused on 0:21:57.400 --> 0:22:01.800 how do you build the resource is an infrastructure and 0:22:01.920 --> 0:22:06.000 community around what open source in these large scale models 0:22:06.040 --> 0:22:09.600 really mean. So that could be open data sets, that 0:22:09.720 --> 0:22:14.520 could be open technology development. Really building on that understanding 0:22:14.560 --> 0:22:17.639 that we need an infrastructure in place and a community 0:22:17.680 --> 0:22:22.959 engaged in thinking about safety and innovation through the open lens. 0:22:23.720 --> 0:22:27.879 This approach brings together organizations and experts from around the 0:22:27.880 --> 0:22:33.840 globe with different backgrounds, experiences, and perspectives to transparently and 0:22:34.000 --> 0:22:38.800 openly address the challenges and opportunities today. I poses the 0:22:38.840 --> 0:22:43.639 collaborative nature of the AI Alliance encourages discussion, debate, and innovation. 0:22:44.440 --> 0:22:47.520 Through these efforts, IBM is helping to build a community 0:22:47.840 --> 0:22:51.480 around transparent open technology. 0:22:52.160 --> 0:22:55.240 So I want to talk about the future for a minute. 0:22:55.480 --> 0:22:58.720 I'm true, is what you see as the biggest obstacles 0:22:58.800 --> 0:23:03.240 to why spread adoption of responsible AI practices. 0:23:03.960 --> 0:23:09.280 One of the biggest obstacles today is an inability and 0:23:09.480 --> 0:23:13.560 really a lack of understanding about how to use these 0:23:13.680 --> 0:23:18.159 models and how they can most effectively drive forward a 0:23:18.240 --> 0:23:22.840 company's commitment to whatever products and services it might be deploying. 0:23:23.240 --> 0:23:26.520 So I always recommend a couple of things for companies 0:23:26.760 --> 0:23:30.080 really to think about this and to get started. One 0:23:30.280 --> 0:23:34.639 is think about how you are already using AI across 0:23:34.680 --> 0:23:38.439 all of your business products and services, because already AI 0:23:38.760 --> 0:23:42.680 is integrated into our workforces and into our workstreams, and 0:23:42.720 --> 0:23:45.680 into the way in which companies are communicating with their 0:23:45.760 --> 0:23:49.280 clients every day. So understand how you are already using 0:23:49.320 --> 0:23:53.800 it and understand how you are integrating oversight and monitoring 0:23:53.840 --> 0:23:56.600 into those One of the best and clearest ways in 0:23:56.640 --> 0:23:59.560 which a company can really understand how to use this 0:23:59.640 --> 0:24:02.960 response is through documentation. It's one of the areas where 0:24:02.960 --> 0:24:06.199 there's a clear consensus in the community. So how do 0:24:06.240 --> 0:24:09.120 you document the models that you are using, making sure 0:24:09.160 --> 0:24:11.240 that you've got a registry in place. How do you 0:24:11.359 --> 0:24:13.879 document the data that you are using and where that 0:24:14.000 --> 0:24:16.439 data comes from. This is sort of the first system, 0:24:16.560 --> 0:24:19.879 first line of defense in terms of understanding both what 0:24:20.080 --> 0:24:22.119 is in place and what you need to do in 0:24:22.200 --> 0:24:25.600 order to monitor it moving forward. And then secondly, once 0:24:25.600 --> 0:24:28.480 you've got an understanding of how you're already using the system, 0:24:28.880 --> 0:24:31.120 look at ways in which you could begin to pilot 0:24:31.280 --> 0:24:34.000 or iterate in a low risk way using these systems 0:24:34.040 --> 0:24:36.960 to really begin to see how and what structures you 0:24:37.040 --> 0:24:39.560 need to have in place to use it moving forward. 0:24:39.920 --> 0:24:43.840 And then thirdly, make sure that you structure a team 0:24:44.000 --> 0:24:47.040 in place internally that's able to do some of this 0:24:47.200 --> 0:24:52.520 cross departmental monitoring, knowledge sharing and learning boards are very 0:24:52.680 --> 0:24:55.679 very interested in this technology, So thinking about how you 0:24:55.720 --> 0:24:58.080 can have a system or a team in place internally 0:24:58.119 --> 0:25:01.000 that's reporting to your board, giving them a sense of 0:25:01.040 --> 0:25:04.920 both the opportunities that it identifies for you and the 0:25:04.960 --> 0:25:08.240 additional risk mitigation and management you might be putting into place. 0:25:08.600 --> 0:25:11.439 And then you know, once you have those things into place, 0:25:11.800 --> 0:25:15.399 you're really going to need to understand how you work 0:25:15.480 --> 0:25:18.760 with the most valuable asset you have, which is your people. 0:25:19.400 --> 0:25:22.400 How do you make sure that AI systems are working 0:25:22.760 --> 0:25:25.240 for the workers, making sure that they're going into place. 0:25:25.280 --> 0:25:28.960 The most important and impressive implementations we see are those 0:25:29.000 --> 0:25:31.439