WEBVTT 00:00.000 --> 00:14.160 Next up, we are going to stream a video from a speaker, speaker, that couldn't make it 00:14.160 --> 00:26.680 because of visa consideration, what's the person from Africa, the continent, I will 00:26.760 --> 00:36.680 discuss the specifics, but we intended the diffusion of the video as some sort of maybe 00:36.680 --> 00:43.240 that's better not recording. I was going to say that we intended the diffusion of her video 00:44.520 --> 00:51.880 more or less against the first-time policies as an act of resistance because it was 00:52.440 --> 01:00.280 tough hearing that she didn't get her visa. We will have her video for 20 minutes and then 01:00.280 --> 01:10.520 we are going to have a live Q&A session with her on the screen, if it works. We did test it beforehand, 01:10.760 --> 01:21.800 but we are going to have a live Q&A session, we are going to have a live Q&A session with her on the screen, 01:21.800 --> 01:28.840 that's true, go ahead and then, that's it. 01:51.800 --> 02:19.080 Can you hear anything? 02:19.080 --> 02:26.080 Oh really? 02:26.080 --> 02:34.080 Okay, maybe we should test that instead of... 02:34.080 --> 02:37.080 Oh, you mean a centred? 02:37.080 --> 02:39.080 Yeah, but actually if you... 02:39.080 --> 02:41.080 Or is that just... 02:41.080 --> 02:43.080 You might be the first light? 02:43.080 --> 02:46.080 Yeah, it's just the first light. 02:46.080 --> 02:48.080 That's good. 02:48.080 --> 02:52.080 Okay, perfect. 02:52.080 --> 03:01.080 Okay, perfect. 03:01.080 --> 03:04.080 Tell her, tell her, we are launching the video. 03:04.080 --> 03:09.080 I'm just gonna write after her. 03:09.080 --> 03:23.080 I guess we should be in advance. 03:23.080 --> 03:38.080 So we have time for switching from the video to the queue. 03:38.080 --> 03:48.080 So let's do this. 03:48.080 --> 04:02.080 Hi everybody. 04:02.080 --> 04:05.080 I hope that you're doing well and you are enjoying it. 04:05.080 --> 04:06.080 I'm a first-end. 04:06.080 --> 04:09.080 He's telling me that I am not able to leave here physically. 04:09.080 --> 04:12.080 Everybody bought something for technology. 04:12.080 --> 04:16.080 And the amazing first-end thing for making space for me. 04:16.080 --> 04:19.080 And maybe we need to send you a recording of my talk. 04:19.080 --> 04:23.080 I'm also thankful that you have decided to give me the next few minutes. 04:23.080 --> 04:25.080 Please, let's do what I have to say. 04:25.080 --> 04:30.080 Now, without further ado, let's get into it. 04:30.080 --> 04:35.080 So, I'm going to talk to you in a reproducivalier practices. 04:35.080 --> 04:41.080 And then I will be focusing on L.M.I.C.S. 04:41.080 --> 04:42.080 Good morning. 04:42.080 --> 04:44.080 My name is pressure seniority. 04:44.080 --> 04:47.080 I am an open source manager from Disaster Science Program Policy Project. 04:47.080 --> 04:49.080 Glyst-Ostronal team. 04:49.080 --> 04:52.080 I am a software source and ability in 2006. 04:52.080 --> 04:53.080 Hello? 04:53.080 --> 04:54.080 Yeah. 04:55.080 --> 04:57.280 And I sometimes just memorize the question, I have a question, I have a question, 04:57.280 --> 05:03.080 I do it on a podcast adaptation, I'm speaking, I love to read and I absolutely love to believe. 05:06.080 --> 05:11.080 From my talk today, I have a few things that I really like it to take from it. 05:11.080 --> 05:15.080 Because this is a foundation and what I have to say. 05:15.080 --> 05:19.080 So, to definition of the key words, reproducibility AI. 05:19.080 --> 05:23.080 And then open as a blanket statement for what opens up and opens times right? 05:23.080 --> 05:30.080 I would also be explaining why which is ability and open is important. 05:30.080 --> 05:36.080 I will be talking about the challenges with non-feeting for which is ability and openness. 05:36.080 --> 05:41.080 I would also be talking about how to build for which is ability and openness in mind. 05:41.080 --> 05:44.080 And sharing some case studies from a global context. 05:44.080 --> 05:52.080 Then I would also be talking about how to build for which is ability and openness in mind for and in my sees. 05:53.080 --> 05:59.080 Now, to definitions, we can't talk about which is ability, without talking about research. 05:59.080 --> 06:01.080 So what exactly is your religious ability? 06:01.080 --> 06:09.080 Your religious ability research is work that can be independently created from the same data from the same code that the original scene used. 06:09.080 --> 06:15.080 So if anybody decides that they want to make use of that in material and they get the code of data and the code, 06:15.080 --> 06:17.080 it should give them the same results. 06:18.080 --> 06:29.080 This definition is that that pair from the turnery and amazing community and resource that typically tries to make which is a bit reset to easy not to do. 06:29.080 --> 06:34.080 Now, explaining why which is ability is essentially almost the same thing. 06:34.080 --> 06:44.080 But now like which is ability now when we start to work in practice and independently created from the same data and same code at the original scene used. 06:44.080 --> 06:48.080 Right, as you can see, this same definition is also that pair from the turnery. 06:48.080 --> 06:54.080 So there are some words that are usually using the same context and sometimes can be in touch games. 06:54.080 --> 06:58.080 They don't mean the same thing and this image here tries to explain that. 06:58.080 --> 07:03.080 So if the analysis is the same and data is the same, then it should be which is support. 07:03.080 --> 07:08.080 If the analysis is different or the data is same, then it is robust. 07:08.080 --> 07:16.080 If analysis is the same, but the data being used is different and the research work is applicable. 07:16.080 --> 07:23.080 And then if the data is different and the analysis is also different and the research work is generalizable. 07:23.080 --> 07:28.080 This image shows the whole process of building of like a research work. 07:28.080 --> 07:31.080 So data collection and data processing. 07:31.080 --> 07:36.080 And analysis, data publishing and access, data visualization and data reuse. 07:36.080 --> 07:40.080 And then research ideas can come up again and become the whole process. 07:40.080 --> 07:43.080 We said data planning and design collection. 07:43.080 --> 07:46.080 Like it's a whole cycle. 07:46.080 --> 07:49.080 Now, we come to explain what open source is. 07:49.080 --> 07:51.080 This is generally an open source compressive. 07:51.080 --> 07:54.080 It might be the assumption that everybody knows what it is. 07:54.080 --> 07:57.080 But for every fresher, I will just explain what it is. 07:58.080 --> 08:04.080 The open source.com describes open source as something people can modify and share because it design is probably accessible. 08:04.080 --> 08:12.080 Something is quite important because as we've seen, the concept of open source is now up to the extent that beyond the context of original 08:12.080 --> 08:14.080 Chinese, which was software, right? 08:14.080 --> 08:22.080 So now other forms of cultural, creative outputs, such as hardware, education, resources, and even now sciences, right? 08:22.080 --> 08:29.080 And then refer into something that's open source, generating means that it's available under an open license. 08:29.080 --> 08:34.080 This definition, this general definition has got to come to an egosry. 08:34.080 --> 08:39.080 And then explaining what an AI system is because that's what we're talking about here. 08:39.080 --> 08:42.080 And AI system is in machine-based system. 08:42.080 --> 08:55.080 It will exist or implicit objectives within the fairs from the impetus we see is how to generate output such as predictions, content, recommendations, or decisions are coming from physical or virtual environments. 08:55.080 --> 09:01.080 Different AI systems vary in their levels of autonomy and adaptiveness, are part of the program. 09:01.080 --> 09:03.080 This is the OECD definition. 09:03.080 --> 09:11.080 And it's quite important because it's the one that the open source initiative uses to explain what an open source AI is. 09:11.080 --> 09:13.080 That's where the adapts the definition from. 09:13.080 --> 09:19.080 And now we explain what an open AI system means according to the open source initiative. 09:19.080 --> 09:27.080 So an open source AI is an AI system that is made available on that terms. 09:27.080 --> 09:33.080 And in a way that grant freedom to study, modify, and share. 09:33.080 --> 09:38.080 It has to have these four freedoms to be considered open source AI. 09:38.080 --> 09:43.080 This freedom is available to fully functional system and to display elements of these systems. 09:43.080 --> 09:46.080 So put in models, we wait to train in data. 09:46.080 --> 09:55.080 And then if the condition to exercise in these freedoms is to have access to the preferred form to make modifications to the system. 09:58.080 --> 10:06.080 It's very easy when we make mention of open our reproducibility to sometimes think that they mean the same thing. 10:06.080 --> 10:08.080 But they do not, right? 10:08.080 --> 10:15.080 It work can be reproducible and open, but doesn't have to claim that all reproducible works are open, right? 10:15.080 --> 10:21.080 So reproducible doesn't always mean open, but the central intersection, where they meet is the answer to the question, 10:21.080 --> 10:27.080 can others reproduce understand and then build on this work without insight on knowledge? 10:27.080 --> 10:30.080 Can I see material and just work with what has been created there? 10:30.080 --> 10:32.080 What has been provided there? 10:32.080 --> 10:37.080 With having to know the actor researchers or having to find out whoever it is, right? 10:37.080 --> 10:43.080 If I can't do that, it means it work has been considered for open and which is all. 10:44.080 --> 10:47.080 What sticks to this intersection for this presentation? 10:47.080 --> 10:54.080 As I'll be speaking about AI systems and creating them within this context. 10:54.080 --> 10:58.080 So what is the potential open our reproducibility? 10:58.080 --> 11:02.080 It helps make results very viable and trustworthy. 11:02.080 --> 11:08.080 When a material is created and publicly available, it's easy for people to trust what they've seen. 11:08.080 --> 11:15.080 If it's made, if it's reproducible, if you can't find the data or the analysis, that was made from it. 11:15.080 --> 11:21.080 As opposed to a research, I just making claims and everybody can't check if it works or if it makes any sense, 11:21.080 --> 11:24.080 I would have to read those what it has been said to do, right? 11:24.080 --> 11:33.080 It also enables adaptation across different contexts because if person B in the global north is able to apply this to the demographic, 11:33.080 --> 11:40.080 passing C in the global south, you'll be able to reproduce that same result by what has been provided. 11:40.080 --> 11:46.080 So it can work in any part of the world that's applied or any situation depending. 11:46.080 --> 11:52.080 And then it enables the support of a responsible and SQL AI. 11:52.080 --> 11:59.080 Because of how AI is very prevalent right now, and everybody's making use of it, and there's a lot of probing to how they use this. 11:59.080 --> 12:05.080 This is an important tool, but we also need to make that this shift and also respect basic human rights. 12:05.080 --> 12:10.080 There is a lot of questions into how is the data being sourced and the ethics of it. 12:10.080 --> 12:18.080 So when the data is created and open as a reproducibility is covered in, it supports the use of it to be responsible. 12:18.080 --> 12:27.080 And it also enables research as a user to be able to trace where it's come from, and everyone was actively sourced and also probing to make ensure that. 12:27.080 --> 12:32.080 And if you nice using feature is you, or it's got some activity. 12:32.080 --> 12:39.080 It also makes the research as a cognitive because you know that somebody's going to look into your work and they're going to question where this has come from. 12:39.080 --> 12:40.080 Right. 12:40.080 --> 12:48.080 So it also reduces waste and the vacation because when you see that something has happened or has been done before, it's easy for you to build on it as opposed to. 12:48.080 --> 12:55.080 The research from someone else doing the same work, knowing that somebody else has worked on it because that's material was not made publicly. 12:55.080 --> 13:01.080 I really wanted to probably found out when they had to submit it to a journal. 13:01.080 --> 13:07.080 Publication and also reduces the prevents the concentration of power. 13:07.080 --> 13:15.080 Most of us or some of us might have seen the current news of how much our AS systems are taking. 13:15.080 --> 13:24.080 The way it's affecting certain environments when we are able to spread the use or the knowledge and study of AS systems. 13:24.080 --> 13:33.080 It enables power to not be concentrated at one place because now different environments and different researches in particular about what are trying out the same thing. 13:33.080 --> 13:38.080 So it makes the use of energy to not be one place. 13:38.080 --> 13:49.080 It also enables the strengthening of global research capacity of mention how it work can be adapted in different contexts or in different environments or different demographics. 13:49.080 --> 13:56.080 By doing that and creating a material or a research article that is. 13:56.080 --> 14:05.080 Global, so that's reproducible, given from different parts of the world can try it and because now they have something to do don't, it makes work. 14:05.080 --> 14:13.080 It makes innovation very easy to work with or to work on and now it's like widespread and you have people from different parts of the world are trying out the same thing. 14:13.080 --> 14:17.080 So it's very great that the strengthening of global research capacity. 14:17.080 --> 14:21.080 So now that is the importance. What are the challenges? 14:21.080 --> 14:27.080 Interestingly, the inverse of the importance of all these things has challenges right now. 14:27.080 --> 14:30.080 So we see that results cannot be trusted. 14:30.080 --> 14:38.080 There are models out there in the world that people are making use of or that researchers are making claims of and saying, this is open or this is just a problem. 14:38.080 --> 14:44.080 Up to practice or in the up to theory of it, which is not, which is why there are. 14:44.080 --> 14:49.080 Which is why there is open source air definition to be able to help use that in the bio key. 14:49.080 --> 14:55.080 If someone says this is open source, this is what you use as a checkmark to see. 14:55.080 --> 15:00.080 Okay, this works. So this is not a good answer to the definition. 15:00.080 --> 15:06.080 And there is also the difficulty for we using or adapt an existing research material. 15:06.080 --> 15:15.080 It is what's exist and what is publicly available that people can use and build on and can foster innovation. 15:15.080 --> 15:26.080 When something is not publicly available, when it is not accessible, it is harder for researchers to be able to build on the right. 15:26.080 --> 15:32.080 And then you now find out there is also the consistency on in the end twist, which I will come to. 15:32.080 --> 15:41.080 Another challenge that we see is our resources have been concentrated because material is not being shared or publicly available and kind of be produced. 15:41.080 --> 15:50.080 So you find out that it is most likely the same set of maybe people, demographic or organizations that make. 15:50.080 --> 16:05.080 Or work on these systems because the most likely are ones who have access to the first set of resources. 16:05.080 --> 16:17.080 We find that another challenge is power resources being concentrated because, again, this materials are not publicly available or they are not able to be produced. 16:17.080 --> 16:23.080 And so it's just the same set of people that keep on turning out materials or findings. 16:23.080 --> 16:33.080 And people with low resources or lower access to these tools are not able to work on something that is groundbreaking because. 16:33.080 --> 16:36.080 The resources are not available for them. 16:36.080 --> 16:41.080 It's cool because also increase because now nobody can probe into how something is being built. 16:41.080 --> 16:48.080 There's no way to see, okay, let me try this test results that you said brought out this and to be able to check. 16:48.080 --> 16:53.080 It's correct or it's right or it's not right. So when. 16:53.080 --> 16:56.080 It's not easy for people to look into the work that's been done. 16:56.080 --> 17:01.080 Then the question of ethics comes up and there's less happening to question it. 17:01.080 --> 17:12.080 And then we see that time energy on funding our police said, I can mention previously when we put the difficulty in using or adoption and existing research material. 17:12.080 --> 17:20.080 And you can also happen that people now work on the same team and you cannot tell until you've probably submitted it to be approved. 17:20.080 --> 17:29.080 And then you can also go over participation strings which is really sad because when work cannot be reproduced. 17:29.080 --> 17:44.080 It's hard for people who are not able to do like first set of research or innovation to now be able to do don't plan and that's to their own context. 17:44.080 --> 17:54.080 Building open our business meeting mind, we wanted first of all start with transferring training. So log and ten more than a detector, lost functions optimizers. 17:54.080 --> 18:00.080 I have a permit as I want them seeds and separate data, pre-processing training and money. 18:00.080 --> 18:04.080 You want to report the results across multiple rounds to capture variability. 18:04.080 --> 18:13.080 This is important because research has once we would see different methods that we tried and different results that came out from them. 18:13.080 --> 18:22.080 If you want to document data pipelines, try to describe all the resources, filtering the very non-sleeves and book these relevant statistics. 18:22.080 --> 18:32.080 Then in the case when data cannot be shared, it's just like if you're working on something sensitive, which you must like do you have some points if in that's the case then. 18:32.080 --> 18:38.080 Just let the researchers know that we made use of the data but it's currently sensing what we cannot be shared. 18:38.080 --> 18:42.080 We have the documents, computational and environmental requirements. 18:42.080 --> 18:47.080 We report hardware, memory, training time, batch sizes and energy or cost requirements. 18:47.080 --> 18:50.080 Again, it seems we could use what we've done. 18:50.080 --> 18:56.080 We need to have the same analysis and then same code or the data set. 18:56.080 --> 19:02.080 So share with you the variables between libraries and class setup instructions. 19:02.080 --> 19:04.080 We want to model behavior as possible. 19:04.080 --> 19:16.080 We also want to design for other patients. 19:16.080 --> 19:21.080 They need to reproduce the whole evaluation by achieving them and doing other patients. 19:21.080 --> 19:32.080 So AI can be meaningless across different contexts, including resource constraints, constraint and LMIC environments. 19:32.080 --> 19:48.080 These are some resources and commences that exist to enable education and common practice to use openness and reproducibility. 19:48.080 --> 19:54.080 There's a two-in-way which has made an image on to make reproducibility as to either not to do. 19:54.080 --> 19:57.080 We should try it and go check it out. 19:57.080 --> 20:01.080 It's like whole found called resources. 20:01.080 --> 20:09.080 There's open data and then there's get from the data or the images to collect data anywhere. 20:09.080 --> 20:16.080 The next point is now with building an open array which is being made for LMICs. 20:16.080 --> 20:24.080 So the previous points that we've considered, which are transparent training, document head data pipelines, 20:24.080 --> 20:26.080 of heat and environment clarity. 20:26.080 --> 20:31.080 One of the behavior is constantly released and designed by the patient. 20:31.080 --> 20:36.080 Considering them and an active active concentration for lower research settings, 20:36.080 --> 20:40.080 we have a design for reproducibility from the start. 20:40.080 --> 20:43.080 It's not something that should happen so many things we do. 20:43.080 --> 20:49.080 Because again, this is something that should be applied, but regardless of the context. 20:49.080 --> 20:58.080 But again, if it's applied from the beginning, it helps lower research like context. 20:58.080 --> 21:02.080 So define who can reproduce your work on the level of reproducibility. 21:02.080 --> 21:12.080 The exact results, similar performance or conclusions, is also going to be where you now specify tools that were used. 21:12.080 --> 21:18.080 So whoever is going to reproduce the work, nobody needs to have a very good view to do that. 21:18.080 --> 21:23.080 Right, we'll talk about the constitution for technology to somewhere in Lita. 21:23.080 --> 21:25.080 But that's for point one. 21:25.080 --> 21:27.080 You want to call it a use local data. 21:27.080 --> 21:32.080 A availability of data sets within the African context is constantly improving. 21:32.080 --> 21:39.080 Training AI models on data sets from here improves the chances of creating AI systems with local contact. 21:39.080 --> 21:41.080 Here means Africa. 21:41.080 --> 21:44.080 In general, like other LMICs. 21:44.080 --> 21:50.080 Citizen client, you want to design models to run a limited hardware and encourage local communities and contributions. 21:50.080 --> 22:01.080 This is important because first of all, when you're doing a research work, trying to implement or reproduce some of these. 22:01.080 --> 22:02.080 Reset copy. 22:02.080 --> 22:03.080 Sorry. 22:11.080 --> 22:13.080 Let me start that again. 22:13.080 --> 22:18.080 Citizen science, design models on limited hardware and encourage local community collaboration. 22:18.080 --> 22:24.080 This is an important because from the beginning, when you decipher which is basically from the start, 22:24.080 --> 22:31.080 you want to factor in using tools that people from lower resource settings will have access to. 22:31.080 --> 22:34.080 Obviously, that happens when you cannot help it. 22:34.080 --> 22:36.080 You have to use something very, very high end. 22:36.080 --> 22:44.080 When the case is where you can, it is encouraged to try out into the actual property accessible. 22:44.080 --> 22:50.080 And then also encouraging local community contributions means that individuals, normal citizens, 22:50.080 --> 22:53.080 I mean, not be really strategized by involved in the process. 22:53.080 --> 22:57.080 That way, it also gives you a great larger data sets. 22:57.080 --> 23:00.080 And it shares more knowledge about what you're building. 23:00.080 --> 23:04.080 And this is just create interesting to what is being built. 23:05.080 --> 23:15.080 Enable adaptation on local reuse, provide guidance for high end journey and reproduce and result on smaller or local data sets. 23:15.080 --> 23:20.080 So let's whatever resource that is created to guide documentation. 23:20.080 --> 23:25.080 Documentations should be clear and easy to use for future readers. 23:25.080 --> 23:30.080 And sometimes also it's so really important because even for the team that has done the work, 23:30.080 --> 23:35.080 you want to be able to come back to two, three weeks data or one year, two years data. 23:35.080 --> 23:37.080 You say you want to be done that work, are you done before? 23:37.080 --> 23:43.080 They want to understand what you, what you were previously. 23:43.080 --> 23:51.080 I have listed some existing communities as the teachers and tools that exist to help our open area, 23:51.080 --> 23:57.080 which is ability to focus on the analyses as DHIS to, 23:57.080 --> 24:01.080 then as the African reproducible network community, 24:01.080 --> 24:06.080 then as DHIS has done, what are the projects which is focused on making up the data, 24:06.080 --> 24:08.080 accountable, accessible and interpretable. 24:08.080 --> 24:17.080 And then open science, Africa, every system, as a list of what are open science communities or projects or funders, 24:17.080 --> 24:26.080 just basically policy makers that focus on building the open area which is easy in mind. 24:27.080 --> 24:33.080 So to end my talk, I've mostly done this as a way to share knowledge for books, 24:33.080 --> 24:39.080 core researches, not so core researches, like the technical and non-technical, 24:39.080 --> 24:41.080 mostly as general knowledge. 24:41.080 --> 24:46.080 And at least understand what this means and how to apply it and what our work is to study the current. 24:46.080 --> 24:51.080 The open area is how we ensure communities and the main cities and not just these other way, 24:51.080 --> 24:56.080 or contributors and it's showing me how it's built and applied. 24:56.080 --> 25:00.080 And it means science that truly knows what that is. 25:00.080 --> 25:02.080 And that is one of the key tenets that we focus on, 25:02.080 --> 25:07.080 such as the ability to transparency, collaboration, and sharing. 25:07.080 --> 25:10.080 Thank you so much for your time. 25:10.080 --> 25:12.080 So this is a some research that I can tell you, 25:12.080 --> 25:19.080 how I have understood from tools that helped me in making this presentation. 25:19.080 --> 25:21.080 Yeah. 25:21.080 --> 25:26.080 I've shared some resources that you found in the class to get a better idea. 25:26.080 --> 25:30.080 So this is an exhaustive, so I just listed up a bunch of them. 25:30.080 --> 25:37.080 If you like to find out more of it, please connect me and I'll do that. 25:37.080 --> 25:38.080 Yeah. 25:38.080 --> 25:44.080 If you kind of really was actually back, because I like to learn what you learn 25:44.080 --> 25:48.080 or you think I can improve on, because I know and I look forward to it. 25:48.080 --> 25:55.080 And I'm going to continue on some time and we'll see you in some months. 25:59.080 --> 26:00.080 Thanks, sir. 26:00.080 --> 26:06.080 So this works really glad we could host that talk even through video. 26:06.080 --> 26:13.080 I especially liked the part where precious said enabling science that truly knows no border, 26:13.080 --> 26:16.080 because that's really the problem here. 26:16.080 --> 26:21.080 We would like to try to have a live Q&A session. 26:21.080 --> 26:25.080 It looks like it works, and I hope it will. 26:25.080 --> 26:28.080 Can you hear us, precious? 26:28.080 --> 26:31.080 Yes, I can. Can you hear me? 26:31.080 --> 26:34.080 Yes, we can. 26:34.080 --> 26:37.080 Okay. 26:37.080 --> 26:42.080 So how do we do that? Do we have questions? 26:42.080 --> 26:45.080 I may repeat them myself. 26:45.080 --> 26:47.080 Yeah, so we have one right here. 26:47.080 --> 26:48.080 Okay. 26:48.080 --> 26:50.080 I want to know. 26:50.080 --> 26:53.080 Maybe, maybe you can take the mic. 26:53.080 --> 26:59.080 I remember I participate on project from the Wikmedia Foundation called Global Development. 26:59.080 --> 27:06.080 And there was a documentary called People are Knowledge focused on people from India and South Africa 27:06.080 --> 27:10.080 where the oral citations or the main source of information. 27:10.080 --> 27:16.080 So I'm wondering if you are thinking about training some of the LLMs in the African context. 27:16.080 --> 27:23.080 Now that's easier to make the transcription from voice to train this model and improve the African context. 27:23.080 --> 27:29.080 Just I don't know if you know something about it. 27:29.080 --> 27:32.080 So precious, did you hear the question? 27:32.080 --> 27:33.080 No. 27:33.080 --> 27:35.080 That's terrible. 27:35.080 --> 27:36.080 Sorry. 27:36.080 --> 27:37.080 It was a great question. 27:37.080 --> 27:40.080 And you, do you hear me? 27:40.080 --> 27:42.080 Yes, I hear you well. 27:42.080 --> 27:46.080 I heard some part about like contributing to projects Wikmedia. 27:46.080 --> 27:54.080 And then the question about not how we can easily chant and chant text to speech. 27:54.080 --> 27:57.080 But in 19 here, after I've got the question. 27:57.080 --> 28:01.080 Okay, so you got most of the important part of the question. 28:01.080 --> 28:11.080 The question was, do you plan or anyone around you plan on training LLMs on transcribed oral data? 28:11.080 --> 28:16.080 On data that's mainly oral and then because it's getting easier to transcribe, 28:16.080 --> 28:23.080 having specifically oral purposes, corporas of data for training models. 28:23.080 --> 28:25.080 Did you hear me? 28:25.080 --> 28:27.080 Yes, I did. 28:27.080 --> 28:29.080 Thank you. 28:29.080 --> 28:33.080 So I do know some projects that have stats in that. 28:33.080 --> 28:36.080 Honestly, I don't know about it in the science context. 28:36.080 --> 28:44.080 That's why I do know of a musla and funded projects where people from the Hebrew region have encouraged to send video records. 28:44.080 --> 28:45.080 Sorry. 28:45.080 --> 28:51.080 What do the recordings of the language transcribed in some of the things that have written in text for longevity purposes? 28:51.080 --> 28:57.080 So the hope is that since we have these things beginning, 28:57.080 --> 29:02.080 then a lot of other sectors can also apply it and then if it's applied in one part, 29:02.080 --> 29:06.080 which is a vision of language, then it can be applied to other sectors. 29:06.080 --> 29:08.080 So it's a great way to start. 29:08.080 --> 29:13.080 So I do think that in the future we can see more things like that. 29:13.080 --> 29:16.080 Great. Great. 29:16.080 --> 29:20.080 Any other questions? 29:21.080 --> 29:25.080 Okay. 29:25.080 --> 29:30.080 Shout it out and then I repeat it. 29:41.080 --> 29:42.080 Great question. 29:42.080 --> 29:44.080 So precious the question was, 29:45.080 --> 30:02.080 can you give us an idea of the constraints and restrictions for training and using AI and LLMs in LMIC's context? 30:02.080 --> 30:03.080 Okay. 30:03.080 --> 30:08.080 So this is considered a general ability of data, right? 30:08.080 --> 30:10.080 So obviously there's more improvement, 30:10.080 --> 30:12.080 conversations like this are happening, 30:12.080 --> 30:15.080 and you have institutions that are making it their goals, 30:15.080 --> 30:17.080 and we have more data available. 30:17.080 --> 30:19.080 Because that's in from there. 30:19.080 --> 30:23.080 We have less data than it should be available currently out there. 30:23.080 --> 30:28.080 So one is not particularly proper according to data. 30:28.080 --> 30:31.080 You get things happen and then, 30:31.080 --> 30:34.080 but time you want to check, 30:34.080 --> 30:36.080 can we find where this happened? 30:36.080 --> 30:37.080 What is the data saying? 30:37.080 --> 30:39.080 It's hard to see that they were documented. 30:39.080 --> 30:41.080 Sometimes that it's not digitized, 30:41.080 --> 30:44.080 and then you find out that probably from some hospitals, 30:44.080 --> 30:46.080 where I need to have direct codes, 30:46.080 --> 30:48.080 taking by pen and paper, 30:48.080 --> 30:50.080 and then if maybe a large fire happens, 30:50.080 --> 30:52.080 or that's a thing, 30:52.080 --> 30:55.080 then the information is lost, right? 30:55.080 --> 30:59.080 So one of the things that projects like DSWB's 30:59.080 --> 31:03.080 and DSWB is to first of all train people 31:03.080 --> 31:06.080 to become more digitally trained, right? 31:06.080 --> 31:07.080 So by doing that, 31:07.080 --> 31:10.080 then we have more data that's digitized, 31:10.080 --> 31:12.080 and that's like one constraint. 31:12.080 --> 31:15.080 Another is sometimes for internet and power, 31:15.080 --> 31:17.080 because you obviously need like a little power 31:17.080 --> 31:19.080 to use a run system. 31:19.080 --> 31:26.080 So some places don't have as good access to the internet, 31:26.080 --> 31:29.080 as there is the water as it should be, 31:29.080 --> 31:32.080 and then now it's like less availability of power. 31:32.080 --> 31:34.080 And sometimes it's also like computational resources, 31:34.080 --> 31:37.080 because it takes a lot of memory and RAM to be able 31:37.080 --> 31:39.080 to run some of the LLMs. 31:39.080 --> 31:41.080 And from the first part of the world, 31:41.080 --> 31:44.080 it's generally a little bit more expensive 31:44.080 --> 31:47.080 to procure these devices. 31:47.080 --> 31:49.080 So, yeah. 31:49.080 --> 31:52.080 These are thoughts or conversations that have been had, 31:52.080 --> 31:54.080 but some of the constraints like listening them out 31:54.080 --> 31:56.080 would be an ability to see a data, 31:56.080 --> 31:58.080 lack of digital literacy to large extent, 31:58.080 --> 32:02.080 but that's just been with respects to storing data 32:02.080 --> 32:05.080 and then level power and internet constraints. 32:05.080 --> 32:08.080 And so sometimes lack of education too, 32:08.080 --> 32:11.080 but thankfully more and more people are talking about this, 32:11.080 --> 32:13.080 and then we have more organizations 32:13.080 --> 32:16.080 that are coming up to educate members of the public, 32:16.080 --> 32:20.080 and as many people as I will need to listen. 32:20.080 --> 32:22.080 Cool. 32:22.080 --> 32:23.080 Thanks. 32:24.080 --> 32:27.080 Questions? 32:27.080 --> 32:28.080 It's good. 32:28.080 --> 32:30.080 We are on time. 32:30.080 --> 32:31.080 We have. 32:31.080 --> 32:32.080 Yeah. 32:32.080 --> 32:35.080 It's good. 32:35.080 --> 32:37.080 So. 32:37.080 --> 32:42.080 Well, there's no questions. 32:42.080 --> 32:44.080 It's fine. 32:44.080 --> 32:45.080 Yeah? 32:45.080 --> 32:46.080 We can wrap up. 32:46.080 --> 32:49.080 Brush is you want to add anything. 32:49.080 --> 32:51.080 If you hear me. 32:52.080 --> 32:55.080 I want to say that this is part of my SSI. 32:55.080 --> 32:57.080 This is my SSI topic. 32:57.080 --> 33:00.080 So, next year, I'm going to be working intensively 33:00.080 --> 33:04.080 on trying to actually make the practices easier to reproduce 33:04.080 --> 33:06.080 and educate people and how to do that. 33:06.080 --> 33:09.080 So, hopefully, I lend better. 33:09.080 --> 33:12.080 But next time, I'm probably speaking on something like this. 33:12.080 --> 33:14.080 I would have like a larger resource. 33:14.080 --> 33:16.080 I sent out like my links in profiles. 33:16.080 --> 33:18.080 If you have maybe more information on something, 33:18.080 --> 33:21.080 you think that might be what they can need to please reach out to me. 33:21.080 --> 33:22.080 And let me know. 33:22.080 --> 33:25.080 And thank you so much for attending my talk. 33:25.080 --> 33:26.080 I'm for the full of them. 33:26.080 --> 33:27.080 I'm giving me space. 33:27.080 --> 33:29.080 I'm making a possible for this to happen. 33:29.080 --> 33:31.080 Thank you so much. 33:31.080 --> 33:32.080 Thank you. 33:38.080 --> 33:39.080 Thanks. 33:39.080 --> 33:40.080 See you soon.