WEBVTT 00:00.289 --> 00:11.135 [Automatic captions by Autotekst using OpenAI Whisper V3. May contain recognition errors.] This webinar will be uploaded and posted on our website in the future. 00:11.155 --> 00:20.099 In addition, we ask that people do not take words without raising their hands. 00:20.359 --> 00:22.720 There is a chat here, and in addition, there is also 00:24.438 --> 00:28.542 a link to a web-scheme where you can ask questions. 00:29.282 --> 00:35.788 You can see more about the questions during the webinar. 00:37.149 --> 00:43.674 The web-scheme we will gladly receive feedback if anyone wants to send something in the application. 00:46.076 --> 00:51.861 So then I thought we could have a little presentation round of us gathered here from the university. 00:52.719 --> 01:05.263 My name is Dagfinn Werkshager, and I am the head of the research, communication and education department here at the IT department, which is responsible for GPT-view. 01:05.603 --> 01:09.304 Hi, my name is Anne Werkshager. 01:09.384 --> 01:14.446 I am the head of the department in Dagfinn's department, which is called Advanced User Contact. 01:15.606 --> 01:19.567 My name is Magnus Nygaard, and I work as a developer here. 01:19.587 --> 01:22.408 My name is Omid, and I am the head of the Institute for Informatics. 01:23.701 --> 01:27.082 I'm Tor Magne Kipflund, and I'm also a developer, along with Magnus. 01:28.303 --> 01:32.705 I'm Katrine Nordheide-Keipa, and I worked as a designer with GPT-UI. 01:33.845 --> 01:36.946 And I'm Fugli, section leader in app development. 01:38.947 --> 01:44.809 So, we can just take a little recap of why we created GPT-UIO. 01:46.607 --> 02:02.240 GPT-UIO was developed a year ago, and the University of Oslo said that it was forbidden to use chat GPT in teaching, or to assign students to use it. 02:02.260 --> 02:05.103 It was not allowed to share personal data, or... 02:05.723 --> 02:12.989 We saw some challenges with ChatGPT, which we said we didn't want people to use. 02:13.369 --> 02:16.011 The biggest challenge was that we didn't get a proper agreement at the time. 02:16.051 --> 02:21.996 We discovered some errors in the universal format in the content that was released. 02:27.857 --> 02:32.738 And the biggest problem was that we didn't have control over our data. 02:32.758 --> 02:36.159 We didn't know where our data was stored or processed. 02:36.539 --> 02:38.079 That's why we thought it wasn't okay. 02:39.180 --> 02:45.121 So we put up GPT-UIO, where you have more control over the data. 02:45.141 --> 02:51.063 GPT-UIO is a web frontend for other GPT models. 02:51.083 --> 02:54.303 So far, we've only used the models we get from Azure. 02:58.141 --> 03:05.726 Our service is available to everyone who has an agreement with Microsoft Azure and buys this frontend. 03:06.486 --> 03:12.249 In our GPT solution, we store all the prompts that are sent in. 03:12.370 --> 03:19.714 They are stored in our database, and we identify these so that Microsoft or the subcontractors cannot know who has sent what. 03:19.794 --> 03:21.215 And in that way, we have 03:23.433 --> 03:27.536 control over the large amount of data we have. 03:27.836 --> 03:36.581 In addition, we can control the part of Microsoft that is where the data is stored, how long they are stored and so on. 03:36.621 --> 03:39.563 Today, how many others use GPT-UIO? 03:39.663 --> 03:41.504 Is it 10 institutions? 03:41.784 --> 03:43.805 Yes, it is 10-12 external institutions. 03:43.845 --> 03:52.771 Then they also have to set up GPT-UIO to go towards their Azure, so that it controls the backend itself. 03:53.760 --> 04:01.602 And we would like to tell you a little about the pieces that we have now further developed in 2024 at GPT-UIO. 04:01.862 --> 04:07.163 It was quite similar to GPT in the beginning, and it is now starting to be divided. 04:09.344 --> 04:15.885 So, I wonder if you want to start showing something, Paul? 04:16.746 --> 04:17.626 Yes, I can. 04:20.353 --> 04:27.257 I can also say that Microsoft has been in contact with us too, because they say that we are the only ones they know about in Europe who have made such a solution. 04:27.317 --> 04:36.001 That is, a solution where we have such a rich borderline, and we provide long-term storage of the discussions we have most of the time. 04:36.021 --> 04:44.966 And it is designed so that the discussions should have a value and can be downloaded and shared, so that we know that it is a little unique what we have done here. 04:45.366 --> 04:47.027 So we can also mention that we 04:48.226 --> 05:03.390 was one of the first higher education institutions in Europe- to offer a chat GPT-like application- to all employees and students who are GDPR-friendly. 05:04.871 --> 05:08.112 We are proud of that. 05:08.152 --> 05:11.332 Let's talk about our first year of use. 05:13.430 --> 05:19.431 Do we have any statistics on how many users we have? 05:19.491 --> 05:21.632 Yes, we do. 05:21.712 --> 05:38.375 Now, per day, we have almost ten million messages, divided by almost one million calls. 05:38.415 --> 05:38.875 That's a lot. 05:43.720 --> 05:50.200 And... Yes, exactly how many people use it. 05:51.617 --> 05:56.081 I can't remember, but I can find out. 05:56.441 --> 06:03.067 We saw that this has been one of the fastest growing publications we have had on IT-avd. 06:03.167 --> 06:07.251 It never went from 0 to 12-14,000 users in a few months. 06:07.271 --> 06:08.372 Or a few weeks, actually. 06:08.412 --> 06:09.493 So that was very fun. 06:09.533 --> 06:12.315 We have had a uniform increase in use. 06:17.515 --> 06:25.720 So having a reasonable load balance has been an essential part of the service. 06:25.821 --> 06:28.983 When we talk to employees, we see that this is what people use. 06:29.003 --> 06:37.989 People use it in different ways, such as cleaning data, or finding commands and codes in economic systems. 06:38.069 --> 06:41.831 It's very different, but people use it in their work every day. 06:41.851 --> 06:43.232 We have received feedback that people 06:44.770 --> 06:47.231 who think this is very effective. 06:47.272 --> 06:50.814 Was it someone who said it was four times faster? 06:51.094 --> 06:59.058 There are some projects that indicate that they have increased productivity by four times. 07:00.239 --> 07:07.583 When we see that some key people in our organization experience that type of productivity, we think that is the cost for us. 07:07.643 --> 07:08.244 And then... 07:14.364 --> 07:22.373 There was a lot of literature about using AI, and one of the sources is to see this as a generative AI, as an assistant. 07:23.334 --> 07:32.523 An assistant for boosting productivity, learning, coding, creativity and so on. 07:33.364 --> 07:34.546 More like an assistant. 07:37.344 --> 07:43.872 Early in the autumn semester, we launched GPT-4 Omni, or GPT-4O. 07:43.973 --> 07:51.322 It is much faster and more responsive than the previous GPT-4. 07:54.002 --> 07:58.707 And then we started with GPT-3-5, and then we went to GPT-3-5 Turbo, and then we took GPT-4. 07:58.728 --> 07:59.308 That's right. 07:59.348 --> 08:02.952 So GPT-4 is the newest, most advanced model we offer. 08:11.499 --> 08:16.602 And that's important for those who use our solutions from other institutions. 08:16.622 --> 08:20.465 They have to activate this themselves in Microsoft's Azure. 08:20.645 --> 08:30.712 We have seen that some have been a bit slow, but we have to be a bit better at informing others about how they should set up and follow up on the new things that are coming. 08:34.691 --> 08:38.692 Omni has also recommended models in the service, so we have set it as standard. 08:38.892 --> 08:46.435 When you choose a new conversation, then it is KBT 4 Omni that is... But Marfonel is 3.5 Turbo for the 4? 08:47.315 --> 08:53.737 Yes, we have that, so it is still possible to choose earlier models. 08:54.978 --> 09:01.300 The idea is that we will phase out KBT 4 Turbo and replace it with 4 Mini. 09:01.780 --> 09:02.080 Oh yes. 09:02.901 --> 09:03.321 It is... 09:05.964 --> 09:09.367 It's good for small and everyday tasks. 09:09.387 --> 09:10.748 It's fast and responsive. 09:13.490 --> 09:21.576 We also see that the 4 Omni model is much more cost-effective than GPT-4. 09:21.616 --> 09:25.620 That's a good reason to have it as a standard. 09:26.580 --> 09:31.224 We can show you around. 09:34.864 --> 09:44.188 Since we are talking about costs, we have seen an increase in the amount of data you can line up. 09:44.669 --> 09:47.090 How much data do you allow to be lined up now? 09:48.030 --> 09:55.514 Now we allow for GPT-4 Omni, the difference in the different models, to the maximum limit of 128,000 tokens. 10:01.229 --> 10:07.154 som tillsvarar, icke helt, men som väldigt grått sett 180 000 tegn, altså texttegn. 10:12.876 --> 10:13.376 It's enough. 10:13.796 --> 10:15.997 It's typically an article, but not a book. 10:16.697 --> 10:26.961 And what we saw was that some of the increase we got in consumption was due to the fact that we sent a lot of data when people kept chatting. 10:26.981 --> 10:29.542 So we made a small change to try to make it more effective. 10:29.562 --> 10:31.942 But that's not what we're going to spend all our money on. 10:41.038 --> 10:45.041 Then we have light and dark mode, preferences. 10:46.823 --> 10:52.687 That is the most popular feature we have made, dark mode. 10:53.028 --> 11:06.018 It came very quickly that students wanted dark mode, and we have started to implement it on other services as well, because it is quite comfortable with dark mode. 11:10.299 --> 11:17.342 A new function this autumn is adapted instructions that can be shared. 11:17.362 --> 11:20.883 The new thing is that they can be shared. 11:21.724 --> 11:30.768 Custom Instruction, or so-called System Prompt, can now be shared with other users of GPT-VO. 11:32.748 --> 11:35.810 This is a way to give the model a specific approach. 11:36.787 --> 11:41.429 Do you want it to give you a type of answer, or answer in a special way? 11:42.109 --> 11:48.812 What we saw was that lecturers used a division of instructions in Canvas, the other learning solutions we had. 11:48.832 --> 11:52.593 In the beginning, we allowed quite small instructions. 11:52.613 --> 11:58.256 Now we have increased the number of signs there, so we saw that lecturers, especially in teaching, this was... 11:59.116 --> 12:01.217 quite actively used. 12:01.237 --> 12:13.542 And I remember I found at the university, the Faculty of Engineering, there was an instruction that said that teachers should train themselves to talk to shy parents about a problematic child. 12:13.562 --> 12:17.123 I remember I glued it in, and then I tried to start a conversation. 12:17.183 --> 12:20.145 I said, I have something I want to talk about your child. 12:20.285 --> 12:25.107 And then I got a shiver back, and a very bad mood, because it was 12:25.967 --> 12:32.072 There was a pregnant parent in the GPT on that instruction, and they separated out the school system. 12:32.252 --> 12:34.453 I tried a little, but I didn't find this. 12:34.473 --> 12:37.956 It's a bit funny that they can use it for that kind of training. 12:37.976 --> 12:40.718 Then you know that it is used for medicine, among other things, for studies. 12:40.738 --> 12:43.720 Paul, can you switch to light mode now while you show? 12:43.740 --> 12:44.181 I can. 12:44.381 --> 12:45.462 Is it better? 12:45.482 --> 12:48.264 I think it's easier to see. 12:55.386 --> 12:56.366 That was a question. 12:56.486 --> 12:57.827 Yes, let's do that. 12:57.847 --> 13:05.249 Here we have a long instruction that you have made, or that others have found. 13:05.749 --> 13:08.570 Yes, I have found this one, just as a simple example. 13:08.610 --> 13:19.173 Teachers now have the opportunity to create their own specialized simulations, or something else, that they can then share with students. 13:19.853 --> 13:22.794 So that is done quite simply by copying a link, 13:24.195 --> 13:29.238 And then other users can choose to subscribe to this instruction. 13:29.278 --> 13:34.920 So when it's updated here, when the teacher updates it, all students will get it updated? 13:34.940 --> 13:35.240 Yes, correct. 13:35.260 --> 13:45.386 And then you also see in the menu here that those I subscribe to have their own icon, right? 13:45.426 --> 13:48.267 And those I have shared have another icon. 13:49.688 --> 13:52.049 So this one called Latex Matte, 13:54.212 --> 13:56.194 Have I shared it? 13:56.214 --> 13:57.034 I think so. 13:57.615 --> 13:58.335 Yes, correct. 14:00.197 --> 14:02.519 And I also subscribe to one from Magnus. 14:02.819 --> 14:11.766 You can just show how it looks when you're going to... You can go to the later format. 14:11.786 --> 14:15.489 You can just show how it looks. 14:15.529 --> 14:17.571 When someone has made it for you. 14:17.631 --> 14:22.395 It's under... It's under, because then you've got a new conversation with the instructions, yes. 14:23.635 --> 14:23.855 Yes. 14:24.155 --> 14:37.721 So then you have the opportunity to copy it and make your own, and edit that instruction, because otherwise you wouldn't be able to do anything with it, since you are someone else who owns it. 14:37.741 --> 14:37.921 Yes. 14:37.941 --> 14:38.241 That's right. 14:38.281 --> 14:40.162 Can you see the content of the instruction? 14:40.182 --> 14:40.242 Yes. 14:40.803 --> 14:46.105 Yes, but we were supposed to try to make sure that it was hidden, yes, because it's a kind of teaching. 14:46.865 --> 14:52.488 Yes, it's very good teaching, so the students know what we have asked them not to answer, for example. 14:52.608 --> 14:52.688 Yes. 14:53.246 --> 14:55.607 Do you have an example of how to use it? 14:55.667 --> 14:57.708 Or have you thought of using it? 14:57.788 --> 15:04.470 Yes, often we would ask them to have a Socratic approach to not giving the answer. 15:04.530 --> 15:09.292 If the students know that they don't give the answer, they don't want to use it. 15:09.312 --> 15:10.972 So it's a bit like that. 15:11.012 --> 15:17.154 Or not to go into it, and that's not the point, but it's useful for the students, because then they know how to limit themselves. 15:17.955 --> 15:18.915 But it's a bit, yes. 15:18.935 --> 15:20.896 So in terms of purpose, I've divided it a bit. 15:22.028 --> 15:22.248 Yes. 15:22.568 --> 15:28.131 And we know that in medicine, they want to use it to find out what is wrong with people. 15:28.171 --> 15:29.251 How do you feel? 15:29.271 --> 15:31.612 Do you have a cough? 15:31.672 --> 15:33.113 I think you have mycoplasma. 15:33.373 --> 15:34.854 And then the answer comes afterwards. 15:35.134 --> 15:36.314 So of course he has to give it. 15:36.755 --> 15:38.756 But it doesn't work when you can read the instructions. 15:39.156 --> 15:48.000 And this has to share the instructions, so it's relevant for some of the environments that we work with at WHO, who want to use this for learning purposes. 15:48.717 --> 15:51.779 such as SimSum Bot, which is one of the largest. 15:51.819 --> 16:01.147 One of the desired conditions is to be able to hide the instruction for the student, so that it becomes a choice. 16:01.587 --> 16:14.557 The way we have financed the development of GPT-UO is mainly that the environment at the university has applied for funds, and we have developed GPT-UO. 16:14.597 --> 16:17.279 That is the way we have financed the development of GPT-UO. 16:18.937 --> 16:32.926 Yes, and SimSomBot wants to use LLM and GPT-UVO as a threshold for training in patient communication, for example. 16:33.686 --> 16:41.611 Where you should be able to practice important conversations and train to be a clinician. 16:42.531 --> 16:46.554 And, for example, be able to convey a difficult disease course to a patient. 16:46.934 --> 16:47.995 So make it a safe environment. 16:49.741 --> 16:52.402 And there are a lot of things you can't practice properly. 16:52.622 --> 16:55.623 For example, there was a case about giving death messages. 16:56.143 --> 17:00.084 And it's very difficult to use students to see and practice that. 17:00.124 --> 17:03.945 But then there was, among other things, a part where you have to go through this in that order. 17:03.965 --> 17:06.206 You have to first say who you are, and so on. 17:06.306 --> 17:07.626 And then we have to test it a little. 17:07.666 --> 17:10.567 Yes, correct. 17:11.687 --> 17:23.575 You have focused a lot on instructions over the past year, because there have been so many concrete feedbacks about the use of instructions. 17:24.036 --> 17:34.963 Yes, and the technology behind it is instructions and RAG, Retriever Augmented Generation. 17:38.666 --> 17:59.858 where it is about optimizing the output to the model by using authoritative sources, so that you can load an authoritative source that the model takes into account when it gives a response, so that you get better quality of the response, so that you have a knowledge base. 17:59.898 --> 17:59.998 Yes. 18:07.779 --> 18:11.701 Did you find out how many users there were? 18:11.741 --> 18:12.041 Yes, about 37,000. 18:12.081 --> 18:12.041 37,000. 18:12.121 --> 18:13.121 That's a lot. 18:13.141 --> 18:22.506 So it's clear that... But you have assumed that it is also with the other institutions that use... Yes, it is on the opposite side of everything. 18:22.566 --> 18:27.728 But it is clear that with 37,000 active users, we have to be able to assume that it has been used a lot in the GPD. 18:27.748 --> 18:28.228 We will see. 18:34.844 --> 18:36.926 We have received a couple of questions. 18:37.166 --> 18:46.116 One is asking, do you get better answers when you write in English, or is it more or less the same as Norwegian? 18:46.556 --> 18:50.060 I think you define it as a good answer. 18:50.120 --> 18:52.603 In terms of language, it is better. 18:53.744 --> 18:55.466 And ChatCup is also very... 18:57.048 --> 19:01.712 But the nuances in the language are better in English. 19:02.653 --> 19:05.535 So I use English a lot. 19:05.695 --> 19:12.361 And when it becomes a final product, you can say it in English, it will be translated into English. 19:12.381 --> 19:13.261 No, into Norwegian, I mean. 19:18.294 --> 19:19.794 But it's a little subjective, I think. 19:19.834 --> 19:23.295 But a lot of what I've trained on is also English, so that's that. 19:23.335 --> 19:26.616 But there are also good Norwegian language models for me. 19:26.756 --> 19:28.777 Yes, which I would be on, absolutely. 19:29.297 --> 19:32.038 I use it a little to write in English, because I'm not so fond of it. 19:32.538 --> 19:34.779 So if I have to write an English email, I write it in Norwegian. 19:34.799 --> 19:45.222 But what I'm not good at is the Norwegian, both cultural and social things, to understand from a Norwegian context, I'm still on it. 19:45.242 --> 19:45.322 Yes. 19:47.128 --> 19:52.993 I see that if you get him to write a whole letter, it will be completely unethical and pompous. 19:53.013 --> 19:57.736 It's like a different way of writing. 19:57.857 --> 20:00.539 I write on the way from the university in Oslo. 20:00.599 --> 20:02.120 I don't know how you do it. 20:02.140 --> 20:03.881 It could be that you do it. 20:03.941 --> 20:05.382 Yes, but I just write. 20:05.442 --> 20:07.584 Let's see. 20:07.644 --> 20:12.028 Is it relevant to increase the number of signs in the instruction? 20:12.248 --> 20:12.648 Absolutely. 20:12.668 --> 20:13.649 We have done that several times. 20:14.927 --> 20:17.948 So it's not something that's just to play with. 20:17.968 --> 20:19.649 Follow up on the net, if you want to. 20:19.669 --> 20:20.810 Then we can look at it. 20:20.830 --> 20:32.115 We also know that there are some environments, especially in medicine, that use instructions that are longer than what's being supported now. 20:32.135 --> 20:35.937 So we'll be able to... We'll be able to do it, absolutely. 20:35.957 --> 20:37.198 We'll be able to do it. 20:37.278 --> 20:38.739 Then it was a good answer. 20:38.759 --> 20:39.759 We'll do it. 20:39.779 --> 20:42.701 We assume it's about 8-7-8000 signs. 20:43.431 --> 20:44.492 As a work hypothesis. 20:44.552 --> 20:46.052 Yes, because now there are two. 20:46.132 --> 20:48.894 Okay, so now we might go up to 10,000 people. 20:48.914 --> 20:53.036 Yes, we have to keep track of how much it costs and things like that. 20:53.076 --> 20:54.937 But we have a certain control over the budget. 21:03.642 --> 21:08.786 It has come up that chatGPT and similar KM models use a lot of energy. 21:08.806 --> 21:13.370 What kind of relationship do we have to the use of energy in GPT-Union? 21:15.691 --> 21:19.915 Again, this is about having a reasonable load balance. 21:20.135 --> 21:24.298 We tune the load so that it should not be too intense. 21:26.119 --> 21:27.781 We have a certain relationship to that. 21:29.679 --> 21:32.910 Do you have ethical limitations on what you can ask about? 21:32.930 --> 21:36.402 Are you allowed to, for example, simulate Adolf Hitler? 21:37.483 --> 21:41.506 Vi har gått inn i oppsettet. 21:41.526 --> 21:49.912 Vi har jo ikke noe styring i vår webfronten, men det er jo noen begrensninger hos Microsoft på de modeller vi bruker. 21:50.352 --> 21:58.879 Der har vi gått og skrudd av alt som finnes av filter, og tatt den mest liberale muligheten vi har. 21:58.899 --> 22:03.142 Det har vært etter tilbakemeldinger, og på universitetet skal vi på en måte være... 22:04.071 --> 22:11.077 Microsoft has very nice restrictions, but in the beginning we had moderate, and it was like there was a lot that we stopped. 22:11.117 --> 22:16.202 It was actually quite annoying, because they stopped strange things. 22:16.302 --> 22:17.083 You can't do this. 22:17.163 --> 22:20.786 If you write that you're making a nuclear bomb, you can't do that. 22:20.866 --> 22:24.169 But there were so many strange things they didn't want... 22:25.510 --> 22:31.795 They weren't allowed to simulate a child, for example, when they were supposed to use it for training. 22:32.256 --> 22:33.257 That didn't work. 22:33.297 --> 22:35.278 They just turned it off. 22:35.498 --> 22:41.283 We are quite liberal, but other institutions may have done it differently in their background. 22:41.443 --> 22:44.926 Shall we go one step further? 22:45.146 --> 22:46.227 Yes, shall we take our time? 22:47.682 --> 22:57.404 Yes, we can also mention that we have tested local LLMs that run on our infrastructure. 22:58.385 --> 23:02.766 And then it has been LAMA 3.1 as a proof of concept. 23:04.846 --> 23:09.167 It is Large Language Model Meta AI. 23:09.287 --> 23:13.968 It should actually be called LAMAi. 23:15.229 --> 23:15.409 Yes. 23:16.789 --> 23:19.030 But it's not that fun, so it's just called Lama. 23:20.270 --> 23:28.032 But this is probably the best model we've seen so far, in terms of how we can set up our servers. 23:28.232 --> 23:30.333 It's very interesting. 23:30.373 --> 23:35.834 And the threshold for GPT-VO is now adapted to be able to choose alternative models. 23:37.435 --> 23:45.397 So the workaround is that it is possible to set up the largest model for Lama on our computer. 23:47.486 --> 23:48.507 For those with special needs. 23:49.607 --> 23:53.029 And if you, for example, have person-sensitive data. 23:53.249 --> 23:55.030 And research data. 23:55.090 --> 24:06.856 We see that there are a number of researchers who are skeptical about sending data out of the country, because they may not know the whole value of the data. 24:06.876 --> 24:12.599 Because even if it is not sensitive data, it may not be the data they want to give away. 24:12.859 --> 24:16.200 So it runs on our dev-server, so we can just briefly... 24:17.864 --> 24:20.065 to show that it is possible. 24:20.145 --> 24:22.425 This is Lama with 405 billion parameters. 24:22.445 --> 24:22.885 Let's see. 24:22.905 --> 24:27.566 Let's see if we can choose another model. 24:27.606 --> 24:29.087 Yes, so it is up and running now. 24:29.127 --> 24:33.328 That's fun. 24:33.368 --> 24:35.308 It is still running a bit slow. 24:35.328 --> 24:38.629 It is in... In the control, I would say. 24:38.649 --> 24:39.489 But it works. 24:49.800 --> 24:51.000 Is it better in English? 24:51.220 --> 24:54.501 Or is there a bigger difference in Norwegian and English? 24:54.521 --> 25:00.223 The largest Laman 3 model has done very well in performance tests, for example on Hugging Face. 25:01.603 --> 25:04.144 So it should be very good. 25:04.164 --> 25:16.867 So we think that if it can be used with up to red data, for example, running on our local infrastructure, it could give Padded Value more value for the service. 25:19.600 --> 25:24.122 I imagine that we can allow more or less infinite amounts of data. 25:25.542 --> 25:31.344 That would be a bit in order to be able to load up entire books and Ibsen's collected works. 25:31.364 --> 25:32.764 Correct, correct. 25:36.366 --> 25:43.568 Yes, then I think we can jump a little further to how we see the further development of the service in the future. 25:43.828 --> 25:44.729 So a roadmap. 25:47.131 --> 25:51.152 One thing we have had a lot of questions about is the loading of files. 25:51.192 --> 25:55.074 And here we have held back a little. 25:55.174 --> 26:05.537 This has been something that has been available in Azure before, but we are a bit unsure of how much it will cost. 26:05.597 --> 26:07.578 We see now that we have great control over the costs. 26:08.914 --> 26:12.817 Another thing is that files are a bit scary. 26:12.878 --> 26:23.867 One is the content of the files, but we also have metadata in the files, so we are a bit skeptical about giving away the files and what happens to the files afterwards. 26:24.488 --> 26:33.395 We are now going to tell you how we have solved this in a rather elegant way, among other things to avoid giving away 26:34.233 --> 26:47.240 metadata, and metadata are things that are stored on the files, and it's in the brain of those who have edited the files, those who own and arrange the data, and it can be things that are a bit scary, and they probably don't use it either. 26:48.321 --> 26:52.964 So it was important for us that this was something we didn't deliver to Microsoft. 26:54.284 --> 27:02.189 Yes, and this is a very exciting technology, with machine vision, or computer vision, among other things, 27:03.115 --> 27:11.278 You can imagine that a student takes a picture of a whiteboard with a notator and a figure, for example, and transcribes it. 27:11.518 --> 27:18.400 Or that a researcher loads up a dataset, so you want to get an overview of the content. 27:25.700 --> 27:25.940 And so on. 27:25.960 --> 27:28.921 Take a picture of the whiteboard and put it in the repository. 27:28.961 --> 27:30.361 That's really fun. 27:30.461 --> 27:32.921 I'm excited to see how it will work. 27:33.041 --> 27:34.542 I have a demo here. 27:35.142 --> 27:38.342 This is something we are working on now. 27:38.502 --> 27:45.824 It will be the focus for the next sprint with development in December. 27:47.164 --> 27:50.924 This is a visual presentation of a method for gene editing. 27:50.964 --> 27:53.425 It's called CRISPR-Cas9. 27:55.276 --> 27:56.237 Is it a Nobel Prize? 27:57.217 --> 27:58.898 Yes, it might be. 27:59.359 --> 28:11.546 Now we will see how the machine, with its vision, interprets and processes this. 28:11.747 --> 28:15.829 Now we upload a file, and this is now a picture. 28:17.470 --> 28:23.154 We already have this on the test servers, and we think we launched this before Christmas. 28:23.973 --> 28:25.174 What do you see here? 28:25.374 --> 28:26.835 Explain the method in a simple way. 28:26.855 --> 28:29.896 For example, if you see it in the first year, refer to relevant literature. 28:30.637 --> 28:32.458 That's the prompt, together with the picture. 28:36.520 --> 28:37.580 But this will be in the next sprint? 28:37.620 --> 28:40.422 Yes, this is what we're working on in the next sprint. 28:40.762 --> 28:44.004 On Christmas Eve, approximately. 28:44.024 --> 28:48.646 You see, we're doing fine with this. 28:49.287 --> 28:50.927 You can recognize what it is. 28:50.947 --> 28:51.828 Give a good explanation. 28:52.600 --> 28:54.381 and references to literature. 28:59.683 --> 29:01.424 That seems cool. 29:01.444 --> 29:02.144 It's pretty cool. 29:02.184 --> 29:04.725 We're looking forward to getting that in place. 29:04.846 --> 29:16.351 It could also have been a transcribed interview from Autotex, for example, which you upload as a file, and can get it summarized, interpreted. 29:16.451 --> 29:20.613 You can ask about themes and trends in the interview. 29:22.459 --> 29:33.347 Does it seem like people talk more if they have short questions? 29:33.387 --> 29:33.587 Yes. 29:33.607 --> 29:39.932 GPT-4 Omni with machine vision can also interpret handwritten text, if you load up handwritten text. 29:41.673 --> 29:42.914 We have set a limit of 100 MB for file. 29:42.934 --> 29:44.175 Isn't that right? 29:48.135 --> 29:49.816 Yes, 100 MB. 29:50.036 --> 29:55.740 So we will calibrate this when we get more experience. 29:55.760 --> 30:01.264 And then pure text files and CSV format are supported. 30:01.284 --> 30:02.945 So simple text files and CSV. 30:02.965 --> 30:10.109 So it's pretty complicated to get this done, but we already have the basics in place. 30:11.410 --> 30:12.411 Have you used this, Hommel? 30:15.541 --> 30:16.802 Yes, with experimentation. 30:16.862 --> 30:20.545 I think it's very good. 30:21.105 --> 30:27.511 I really liked the notepad block, so I took pictures of the arc and drew on it. 30:27.531 --> 30:30.653 So I asked him to help me structure it. 30:30.673 --> 30:31.154 It's very cool. 30:34.193 --> 30:45.679 We have also tried to take pictures of data on a whiteboard, so we have put up numerical columns and such, and then said, okay, I just want to have this in a table, and then I want to have it processed. 30:45.719 --> 30:50.541 Make a code to make this border section, you have to think on the board. 30:51.061 --> 30:51.262 No. 30:53.423 --> 30:58.845 So, maybe we can just show a little bit of how we have thought the architecture. 30:59.246 --> 31:01.247 We have a... And I'm going back to the... 31:02.284 --> 31:14.850 We talked about how we are a bit afraid of metadata, among other things, and that we are also a bit afraid of the cost picture if we only have a file that is sent back and forth to Azure with 100 megabytes, and that it is an ineffective way. 31:14.870 --> 31:19.832 So can Tor maybe tell us about the architecture here? 31:20.733 --> 31:22.534 I can do that. 31:22.554 --> 31:27.616 This is what we are talking about when we refer to Retrieval Augmented Generation. 31:31.807 --> 31:54.097 What we do is that the user loads up a source, for example a picture, a file, something like that, and then we augment, that is, we fill out the prompt to the user with what it has loaded up to improve the generation. 31:54.577 --> 31:58.519 Because Azure is open to our model, 31:59.315 --> 32:13.018 doesn't have access to the internet or anything like that, and the prompt size is limited, so we have to limit what it gets in, or we have to choose what it should get in, and that's what retrieval augmentation is all about. 32:13.579 --> 32:25.202 So what happens is that the user loads up a file, and since it is a picture, we send it to, in this case, we probably use 32:27.874 --> 32:34.839 OpenAI's embeddings through Azure, which are multi-modal and which can take in the file. 32:35.099 --> 32:37.481 We send it in as a base 64 string, I mean. 32:38.521 --> 32:41.463 If there is text, we can chunk it up. 32:41.903 --> 32:43.925 We don't send the file, is that what you're saying? 32:43.945 --> 32:47.127 We don't send the file, we just send the content in it. 32:47.167 --> 32:49.148 So we don't send, for example, metadata, as you say. 32:51.124 --> 32:52.385 We send a string. 32:52.525 --> 32:53.226 What is that? 32:53.486 --> 32:54.887 We send a string. 32:54.927 --> 32:57.569 It represents the file, the content of the file. 32:57.589 --> 33:05.756 And what happens then is that we run it through one of these embedding models and get a vector. 33:07.717 --> 33:13.062 And we could basically run it on a local model and get a similar vector. 33:13.539 --> 33:18.482 Is the vector on our own software, or is it from Azure? 33:18.823 --> 33:24.867 Today we use it on Azure, just like we use other chats and code computations and such. 33:25.647 --> 33:32.932 But there is nothing technical other than maybe capacity that makes us be able to use it on our own. 33:32.992 --> 33:36.895 For example, with a LANA embedding or another multimodal embedding. 33:38.115 --> 33:38.916 Then we get a vector. 33:39.679 --> 33:44.760 This vector is important, because we store it in a database, along with the actual content. 33:45.540 --> 33:52.581 When the user asks about something later, this question is also transferred to a vector of the same type of embedding. 33:52.601 --> 34:02.423 We can use this vector to search in our database for vectors that are close to what the user is asking about. 34:02.463 --> 34:03.664 These embeddings, these vectors, 34:06.261 --> 34:16.269 They actually concentrate the meaning in the content in a way that the model understands. 34:16.309 --> 34:21.193 And then we can get content that looks like what the user is asking for. 34:21.213 --> 34:32.622 And then at the end, when we ask the model, or ask it to complete the chat completion, we put a V. For example, the user asked with 34:36.112 --> 34:41.993 And based on vector search in our database, we have these results. 34:43.173 --> 34:44.954 Incorporate it in your answer. 34:45.514 --> 34:47.934 Use it as safe information. 34:48.615 --> 34:50.475 We know this from before. 34:50.555 --> 34:53.756 Take care of it when you make your chat. 34:54.796 --> 35:05.158 So in a way, it works exactly the same way as if you had used an instruction, or if you had pasted the data in the same way. 35:05.968 --> 35:13.172 But by loading up, and then we keep the files, these vectors and content files, on our infrastructure. 35:13.192 --> 35:16.653 We use S3 today, I think, which is locally hosted. 35:16.673 --> 35:25.057 So you can, for example, load up a file and then use it across conversations, in principle. 35:25.858 --> 35:33.161 And you can use, you can pick up different information from different files through a chat by 35:35.140 --> 35:41.502 So that means that if they made a text file of 100 megabytes, or 99.9, then not all the data will be sent to them? 35:41.522 --> 35:41.783 That's right. 35:41.803 --> 35:49.145 We can decide how much we want to chunk up the file itself, but then you only send what is useful from that file. 35:49.225 --> 35:50.726 And that's how it will work. 35:50.746 --> 35:52.146 So then we have to first... 36:01.587 --> 36:03.488 To make it work, we have to process the file first. 36:03.508 --> 36:12.191 We have to make the embedding, so that after loading the file, it would take a short time, not very long. 36:12.652 --> 36:23.256 We have tested it with an open source biology book on a thousand pages, and it didn't take a long time, but at least a few minutes. 36:23.276 --> 36:29.839 You feel you've gone through all that, and you feel it would be available for you to search for. 36:32.185 --> 36:34.246 This gives a lot of possibilities. 36:34.326 --> 36:35.487 It's very exciting technology. 36:36.808 --> 36:37.709 I have a question. 36:37.729 --> 36:42.251 Would it be possible to, for example, upload a video and get information about what it contains? 36:42.291 --> 36:42.492 Right. 36:42.532 --> 36:52.138 In theory, it could work, but it depends on, first of all, that the video is within the limits we set, and that the embedding we use solves the video. 37:00.920 --> 37:11.529 So today there are quite a few multi-modal embeddings that can both take in image data, PDFs, files, text and so on. 37:12.489 --> 37:17.734 But there are probably some who also take video, but there are not so many of them yet. 37:18.394 --> 37:20.015 But over time, it could work. 37:20.035 --> 37:23.198 VideoEye has... The answer is actually yes, it can work. 37:23.238 --> 37:24.219 It will probably work. 37:24.319 --> 37:25.079 It could work. 37:29.985 --> 37:33.706 So you can first make the film, and then we can automatically make the book afterwards. 37:34.086 --> 37:40.147 And then the question is, can you also search in your chat history? 37:40.227 --> 37:42.487 That could also be possible. 37:42.567 --> 37:49.529 Then we have to process, go through all the chats and vectorize them in the same way, but that could be possible. 37:49.569 --> 37:52.109 Do we have any plans to do that? 37:52.489 --> 37:54.810 That is probably a bit... 37:56.587 --> 37:57.508 At any given time. 37:57.568 --> 38:01.651 Yes, at any given time, we look at that type of roulette. 38:01.671 --> 38:03.112 It's very good. 38:04.413 --> 38:06.595 This was awesome. 38:08.176 --> 38:09.757 I can't help myself. 38:11.519 --> 38:22.768 So, we can move on and say something about another feature that has been much wanted, especially by students, and that is rendering of mathematical formulas. 38:23.889 --> 38:33.776 We get almost daily questions from students, because they want to see the forms visually, instead of expressing them in latex syntax. 38:34.357 --> 38:38.240 Because some have difficulty reading it. 38:39.481 --> 38:41.042 So it works already. 38:41.062 --> 38:43.904 Let's see. 38:44.584 --> 38:46.826 It's out on GPTV, right? 38:48.047 --> 38:49.228 Yes, it's out. 38:51.500 --> 39:02.208 For the time being, you have to give an introductory prompt so that it will work, but we will add it as a system prompt so that the users do not have to ask about it every time. 39:03.649 --> 39:09.594 Let's see, we can start a new conversation. 39:09.634 --> 39:19.181 So I choose the instruction, and I have already put in the introductory prompt here. 39:20.703 --> 39:21.183 Let's see. 39:26.083 --> 39:29.024 So, and here we have also listened to the users. 39:30.565 --> 39:33.827 They didn't just get dark mode, they get formulas too. 39:33.867 --> 39:36.968 We work user-centric, so we are very happy with that type of feedback. 39:37.008 --> 39:38.329 Let's see. 39:38.369 --> 39:49.215 Then we can ask them, for example, if you can... 39:51.059 --> 39:54.520 One unsolved... Or one unsolved, shall we say. 39:58.602 --> 40:04.924 Mathematical... Theoremic... Let's see. 40:05.144 --> 40:05.304 Yes. 40:05.364 --> 40:13.927 Is that the new part? 40:13.987 --> 40:15.248 That part is the new one. 40:18.746 --> 40:22.648 So the alternative was that this came up in the latex syntax. 40:22.668 --> 40:24.750 Can you glue it in? 40:25.270 --> 40:25.550 Yes. 40:25.570 --> 40:28.952 Show me more. 40:29.132 --> 40:37.517 But can you glue in the formula O and then ask what the answer is? 40:37.537 --> 40:38.098 I think so. 40:38.878 --> 40:40.699 Oh, there's a little problem this morning. 40:49.653 --> 40:49.753 So. 41:13.496 --> 41:18.200 Could you tell us about some of the experiences you've had working in this field? 41:18.220 --> 41:19.942 Yes, there are a lot of exciting things going on. 41:20.102 --> 41:25.767 I also work with different disciplines, and it depends on what you do. 41:30.632 --> 41:36.775 can use it and how you can use it, and also how good these models are to give good answers. 41:37.635 --> 41:42.358 Now we are also in KI on the other side, if you can take a look at the page that is made. 41:42.398 --> 41:44.839 What is KI on the other side? 41:44.939 --> 41:54.503 It is a network that is on the other side of VIO, where there is both study administrative and technical, but also LINK, which also 41:56.161 --> 42:01.644 where you try to see how the challenge with how KI affects and is. 42:01.664 --> 42:21.475 If you go a little higher, then it is first, if you know how to use it as a student, then the students can go in there and see how they can use it and what their position is, and then a little further down, if you go a little further down, then it is how you can use KI as a teacher, and that is very interesting, and then there are questions around 42:22.913 --> 42:30.257 Ethics and social responsibility is a bit around this, and also some legal guidance on how to use it. 42:30.297 --> 42:45.545 And the idea is also a bit in collaboration with KI på Tvers, and Link is also, we have plans to have workshops around the institutes here on how to use it. 42:45.905 --> 42:49.727 And you are involved in this network as both researcher and journalist? 42:50.008 --> 42:50.128 Yes. 42:51.738 --> 43:06.854 You mentioned that there are a number of funds that the project has applied for, from UiO Central, from SKIT, to work and incorporate AI into business training, and there is a lot going on. 43:14.779 --> 43:17.761 and research and try to find new ways to use it. 43:17.841 --> 43:21.864 And then it will come both in cooperation with you, but also with you in the development of UiO's GPT. 43:21.904 --> 43:22.424 Absolutely. 43:27.611 --> 43:33.255 And that's the advantage of having control over this GPT-VO and being able to adapt it. 43:33.275 --> 43:40.220 I can tell you about another project that also had means and that we have come to do. 43:40.560 --> 43:43.783 And that was at the economy department with us. 43:43.823 --> 43:55.071 They have tested out and made their adapted tenant, that's what it's called in Azure, by loading data to it. 43:55.788 --> 44:07.063 Vi har jo åpnet opp for at ansatte på universitetet kan få tilgang rett i Azure med å fylle ut et skjema, så gir vi dem tilgang til våre, som har sagt at det er de 44:08.081 --> 44:13.004 I think I have said that they can use it for a couple of thousand kroner, and if they use more than that, they get a discount. 44:13.024 --> 44:16.126 And there we have a 20-30 employee who has tested this out. 44:16.166 --> 44:22.430 And there you can, among other things, load up data and train your own tenant in Azure. 44:22.490 --> 44:31.877 And at the economic department at UiO, they have done this with documentation for our economic system, which they have a desire to get a number of support tenants for. 44:33.496 --> 44:46.150 And this was someone who had made a little frontend on his data machine, and then someone came and tested it on his data machine, and then they wondered, what am I supposed to do to get this on the web? 44:46.190 --> 44:50.074 Then he said, if you want to have it on the web, then it has to be in GPT-UIO. 44:50.094 --> 44:52.377 Then we agreed, we made an agreement with them, that this 44:53.218 --> 45:08.679 It will be a way for them to go in and choose the economy chatbot, which is specialised in creating an economy system, and know all accounts, and that they will see if this will get a number of support messages. 45:09.516 --> 45:14.538 And we have also said that when we do this, we make it for scaling. 45:14.558 --> 45:19.420 That means that everyone can get their own special trained chatbot. 45:20.120 --> 45:24.722 And in the first document, it's in Azure, we offer this, that you can train it up there. 45:25.382 --> 45:29.984 I think he trained this one for about 10,000 kroner, so it wasn't half as expensive. 45:30.044 --> 45:32.384 He had made a lot of documentation and such. 45:34.002 --> 45:41.350 But of course we want to have this on local GPUs and on our supercomputers. 45:41.550 --> 45:43.653 So that has to be one of the first things. 45:43.713 --> 45:47.897 And then we can imagine that we train one per subject. 45:48.278 --> 45:52.843 Then we can give the society, folks, I now have a special model. 45:53.283 --> 45:59.328 Here you will be able to get tasks that fit, you can just ask for it, and so on. 45:59.348 --> 46:04.773 It's not that far away, and it's a little fun to see that this can happen. 46:04.873 --> 46:09.717 It will come soon that we can connect to special pentamodels. 46:09.737 --> 46:13.900 We'll see how it goes on GPTV, but I'm not saying that everyone should get this. 46:15.979 --> 46:23.001 a bit boring economy chatbot in the face, because there are 39,000 users and perhaps not more than a few hundred who want to use it. 46:23.021 --> 46:30.022 So there are several good things that are happening in the future. 46:30.102 --> 46:36.044 One of them is alternative login to GPT-VO for institutions that do not have FIDE. 46:36.684 --> 46:38.344 We have received a number of requests for that. 46:38.364 --> 46:44.606 It will be possible to log in with EduCloud after the plan. 46:45.248 --> 46:53.951 EdoCloud is our research solution, where there is a user database, among other things, that can easily make users. 46:53.971 --> 47:01.454 We have this other KI service called Autotext, which uses transcription, which runs whisper models locally. 47:01.734 --> 47:05.055 There are quite a few who have bought access to it. 47:05.075 --> 47:11.237 It was only available for those with FADE, but there were so many who wanted to have it. 47:11.277 --> 47:13.598 We have almost 60 institutions that have bought it. 47:15.199 --> 47:20.024 And there we have used EduCloud, so they can control their own users if they want that instead. 47:20.285 --> 47:25.850 So now we can then add GPT-UIO to departments, among other things, who want to use it. 47:25.890 --> 47:30.856 Because it's so problematic that when you write in data, you don't know where they end up. 47:30.896 --> 47:32.718 You don't know how long they will be made. 47:33.398 --> 47:39.640 And I look forward to being able to offer this to everyone in Norway who wants to have this in use. 47:39.700 --> 47:45.321 And this is actually a cheap service, but you pay for the use yourself directly through Microsoft. 47:45.401 --> 47:55.603 So it is also the research environment in the health sector, among other things in hospitals, who have wanted an alternative application to take the Epitego in use? 47:56.404 --> 47:58.144 That is something that will come. 47:58.164 --> 48:01.685 It is estimated to come early next year, in January. 48:03.795 --> 48:11.218 And the organization of the conversation history, something that we want to further develop. 48:11.258 --> 48:20.441 To get a refresh of the design on it, for example, to be able to delete several messages at the same time. 48:20.541 --> 48:21.441 Or several conversations. 48:23.902 --> 48:29.144 We are also following up on new models from OpenAI and Google Gemini. 48:30.804 --> 48:32.865 So we want to be able to offer the best, newest models. 48:34.502 --> 48:43.377 Hvordan går de som har kjøpt UiO GPT av oss frem for å få tilgang til nye modeller og ny funktionalitet? 48:44.659 --> 48:57.004 Yes, we have taken some self-criticism for that, that we have not informed our other customers well enough about how we have... GPCVO has been much better than many other models, because they have not made any changes in Azure. 48:57.364 --> 49:06.048 So when we make changes in our Azure setup, we will send an explanation to all our customers about what we recommend now. 49:06.448 --> 49:10.229 We recommend Merck tokens, we recommend this model. 49:10.249 --> 49:21.151 So we should be a little smarter to inform about it, because those who use GPT-VO must have someone who has admin in Azure for that institution. 49:21.591 --> 49:32.172 And the latest model for OpenAI that is being developed now, Orion, is also very interesting, because it has the opportunity to 49:36.705 --> 49:42.947 to think through before it gives an answer. 49:43.007 --> 49:49.768 It can reason over more complex tasks before it answers. 49:51.368 --> 49:54.549 I know many people who can't do that, think before they speak. 49:54.569 --> 49:58.470 That means it can solve more difficult tasks within science and technology. 50:04.329 --> 50:05.710 and math, and coding. 50:09.693 --> 50:15.318 So that's what we're looking forward to, to be able to try the latest models. 50:15.378 --> 50:20.822 Are there any other questions that are good to ask here? 50:20.902 --> 50:23.464 Someone asked if there will be a phone app? 50:25.045 --> 50:30.169 Yes, we are responsive here, so you can use it on a smart phone. 50:30.189 --> 50:32.771 It's not a native app, it's a... 50:34.310 --> 50:34.570 Yes. 50:34.930 --> 50:35.390 The website. 50:36.591 --> 50:39.094 So yes, we can say that. 50:39.274 --> 50:39.754 Possibly. 50:39.774 --> 50:44.919 We'll have to see if it has... I was wondering if you could talk directly into PPTVO. 50:44.999 --> 50:51.104 It might work differently on a mobile phone if you press... You can do it on the mobile app. 50:51.124 --> 50:52.245 You can talk to it. 50:52.305 --> 50:53.486 And then you can talk back. 50:53.646 --> 50:54.187 So you can... 50:57.459 --> 51:03.785 And we absolutely do not exclude... In any case, everything that this person says in Norwegian. 51:03.985 --> 51:06.127 So they suddenly created an app for that. 51:06.747 --> 51:09.690 It's very difficult to tell a five-year-old, so they do it. 51:09.930 --> 51:13.993 We have a universal challenge, and we want to get even better at it. 51:14.013 --> 51:19.978 So being able to offer speech, the text and the other way around, can be interesting. 51:26.790 --> 51:30.594 Will we be able to offer API access to the local models? 51:30.634 --> 51:33.077 Yes, I think we will. 51:33.297 --> 51:35.179 We would like to offer these models. 51:35.219 --> 51:35.920 We hope we will. 51:46.721 --> 51:56.788 We now run local models for some research projects in EduCloud, this research system that we have here at the university. 51:56.868 --> 51:58.670 But then we only run for that one project. 51:59.070 --> 52:04.454 So there is only one person who can use it when he logs in and uses resources on it. 52:04.514 --> 52:07.236 So we have some testing around it, and that's a little fun. 52:07.256 --> 52:10.779 We have, among other things, some at HF who had 52:11.799 --> 52:20.208 images of hieroglyphs or helleristings, some kind of old script language that they now manage to interpret in a completely new way. 52:20.248 --> 52:21.309 And this is also data that is 52:22.587 --> 52:26.970 Secret, in the sense that it has never been transcribed before, and we would not give this gold. 52:27.051 --> 52:30.853 The research data we would not give to cloud solutions. 52:31.174 --> 52:44.183 First of all, it would have been expensive to do it in Eschøl, because they had a 10,000 photo of a mass script, which they then interpreted, and they had managed to do it, and now we help them to do it in our solution with their own GPT model. 52:44.483 --> 52:47.726 So that is also a bit of a direct result. 52:50.339 --> 52:56.362 Can you tell us a little bit about how external institutions that want to take advantage of MTU go about it? 52:56.382 --> 53:02.625 The prerequisite is that you have to have an agreement with Microsoft, with Azure. 53:02.645 --> 53:06.006 I think everyone in Norway has ended up having that. 53:06.426 --> 53:08.868 It's just that there are very few who use it that much. 53:08.888 --> 53:15.371 Because you also have to know someone who has access to doing the development of the keys that you have to do with us. 53:17.143 --> 53:22.886 In addition, it is the easiest for those with FADA, because then you can turn on access to a whole institution. 53:24.027 --> 53:32.092 We have been very clear about the fact that we should not differentiate between employees and students here at the university. 53:32.172 --> 53:35.674 Everyone should get the same access. 53:35.694 --> 53:44.339 But now it is possible to control users more one by one by getting an educated user database for those who do not have FADA. 53:46.694 --> 53:56.218 And then you just have to go to our website, and there is information about how to contact us, and then we can set this up for you. 53:56.238 --> 54:02.020 Yes, and we see this as a joint service in the sector, more as a place investment. 54:02.120 --> 54:09.343 It is also important to note that we are not a commercial actor, but we want to cover sales costs. 54:10.822 --> 54:14.987 Yes, we guarantee that when we sell this one today, we have no benefit from it. 54:15.027 --> 54:20.573 But we hope that it can cover some of the development that we expect to do in the future. 54:21.531 --> 54:27.115 And it can also come with means of delivery, as they do at WHO, to further develop the GPT-V. 54:27.255 --> 54:42.404 So if other research environments outside of WHO want to do things like what I've talked about, or other adaptations, now we're going to do a lot around instructions for a specific project at WHO that has funding. 54:42.744 --> 54:47.127 But we also have the opportunity to do this for external uses of the GPT-V. 54:49.605 --> 55:00.091 BEI has been part of the funding for the release of other institutions, so they were one of the first in the solution. 55:00.131 --> 55:04.613 And the premise is that we can give it back to others who wish for this. 55:04.693 --> 55:05.934 Yes, now the time is approaching 10.57, 55:20.069 --> 55:21.271 En god time. 55:22.113 --> 55:29.066 Vi håper at folk har hatt nytta av dette, og takk til alle som har vært her. 55:29.086 --> 55:31.711 Nå tror jeg rett og slett bare vi sier... 55:32.952 --> 55:40.974 I think from my point of view, it is very interesting if someone is thinking about pedagogical use or judo, then it is just to make contact. 55:41.094 --> 55:46.416 You write OMID between the words ifi in Google, then you come straight to OMID. 55:46.436 --> 55:48.397 Then you come straight to the other, right? 55:48.417 --> 55:58.560 So it is very positive to take both from the EU and also externally, where we can contribute to a lift in terms of the knowledge and use of AI. 56:04.360 --> 56:07.564 I think we'll say the same thing, and then we'll say thank you. 56:07.925 --> 56:08.265 Thank you.