WEBVTT 00:00.009 --> 00:24.121 [Automatic captions by Autotekst using NB Whisper. May contain recognition errors.] This webinar will be taken up and posted on our websites afterwards. In addition, we ask that people do not have words without raising their hands. There is a chat here. 00:24.138 --> 00:52.602 We can look a little more at those questions partly during the webinar. We are happy to receive feedback on that online form, if anyone wants anything in the application. Then I thought we could have a little round of presentations by those gathered here from the university. 00:52.619 --> 01:22.568 My name is Dagfinn Bergsager, and I am then the head of the department called research, communication and education here at the HIT department, which is responsible for GPT. Yes, hi. My name is Anne Bergsaker. I am the section manager for a section in Dagfinn's department called advanced user contact. I am Magnus Nybård, and work as a developer here. I am Omid, and I am the first names on the Institute for Information. 01:23.541 --> 01:46.510 I work as a designer with GPT-UiO. So can we just take a little recap on why we made GPT-UiO? 01:46.527 --> 02:05.623 GPT Ujo was developed with us a year ago, and some of the background was that the University of Oslo said it was forbidden to use ChatGPT in teaching and ... Or to order students to use it. It is not allowed to share personal data or ... 02:05.623 --> 02:27.500 We saw some challenges with ChatGPT, which we said we don't want people to use. The biggest challenge was that we didn't get a proper agreement at the time. We uncovered a number of errors and universal design in the handle that was then outside. 02:27.517 --> 02:57.244 And the biggest problem was that we had no control over our data. We didn't know where our data was stored or processed. That's why we said it wasn't okay. Then we set up GPTUIO, where we have more control over the data. GPTUIO is a front end to other GPT models. For now, we have only used the models that we get from Asher. 02:57.841 --> 03:23.336 Our service is available to anyone who has an agreement with Microsoft and Asher to buy this front end. In our GPT solution, we store all the sites that are sent in. It is stored in our database, and we identify these so that our subcontractors cannot know who has put what. 03:23.353 --> 03:53.251 In addition, we can control that part with Microsoft, which is where data is stored, how long they are stored, etc. We probably have today how many others use GPTUJO. Is it ten or twelve external institutions? Then they also have to set up GPTUJO to go towards their Asher, so that you manage yourself that back end. 03:53.660 --> 04:18.046 We will tell a little about the pieces we have made of further development in 2024 at GPTU. It was quite similar to GPT at the beginning, now it is starting to divide something. I wonder if you would like to start showing something, Pål? 04:20.333 --> 04:47.227 Microsoft has been in contact with us too, because they say we are the only ones they know of in Europe who have created such a solution. We offer long-term storage of the discussions. It is planned that the discussions should have a value and can be downloaded and shared. It is a bit unique, what we have done here. 04:48.166 --> 05:13.353 Was one of the first higher education institutions in Europe to offer a ChatGPT-like application to all employees and students which is GDPR-friendly. We are proud of that. Shall we say a little about our first year of use? 05:13.370 --> 05:42.176 Do we have any statistics on the number of users with us? We have. Now as of today, we have almost a ... almost ten million messages distributed over almost a million conversations. 05:43.439 --> 05:50.862 And ... yes, exactly how many users there are. 05:51.357 --> 06:17.039 But I can find that out in a short time. Yes. We saw that this has been one of the fastest growing Vale publications that we have had in the IT department ever. It went from 0 to 12,000-14,000 users in a few months. Or a few weeks, really. So it was a lot of fun, then. Then we have had a steady increase in use. 06:17.415 --> 06:44.633 So this must have a sensible load balancing has been a significant part of the service. And when we talk to employees and such, we see that this is what people use. And people use it for a little different things. Preferably cleaning up data or finding some commands and codes in the financial system. It's a lot of strange things, but people use it at work every day. And we have also received feedback that people ... 06:44.650 --> 07:14.087 Some projects with us indicate that they have increased productivity by up to four times. So when we see that some key people in our organization experiencing that type of productivity, then we think that only that is worth the cost for us, then. 07:14.104 --> 07:34.906 Some literature has come about applied KI, and one of the regulars is to see this as a generative AI, as an assistant. So an assistant to boost productivity in learning, coding, creativity etc. So more like an assistant, then. 07:37.244 --> 07:53.865 Early in the autumn season, we launched GP4 Omni, i.e. GP4O. So it is significantly faster and more responsive than its predecessors, GP4. 07:53.882 --> 08:11.442 And then we probably started with GPT-5, and then we moved to GPT-35 Turbo, and then we took GPT-4. That's right. So GPT-4O is the newest, most advanced model we offer, then. 08:11.459 --> 08:34.514 It is important for those who use our solution from other institutions, they must activate this themselves in Microsoft's Ash. We have seen that some have been a bit slow. But we should be a bit better at informing others about how they should set up and follow up more on the new things that come. 08:34.531 --> 09:04.241 Omni has also recommended a model in the service, so we have set it as standard. When you choose a new call, it is KPT4 Omni that is ... But we still have 35 Turbo for them and 4 ? Yes, we have it. So it is still possible to choose previous models. The idea is that we should phase out KPT4 Turbo and replace it with 4 Mini. 09:05.964 --> 09:34.787 Good for small everyday tasks. It is fast and responsive. We also see that the model 4 Omni is significantly more cost-effective than GPT-4. There is also good reason to have it as standard. We can show a bit around. 09:34.804 --> 10:00.896 We have seen a small increase too. We increased the limit on how much data you can glue in. How much data do we allow you to glue in now? Now we allow on GPT-4 Omni, that is the difference between the various models, to the maximum limit of 128,000 tokens. 10:01.169 --> 10:12.739 Which corresponds, not completely, but as, very roughly, 128,000 characters, i.e. text. 10:12.756 --> 10:39.445 That's enough. It's typically an article, but not a book. And what we saw was that some of the increase we got in consumption was due to that we knew a bit a lot of data when people continued to chat. So now we made a small change and are trying to make it more efficient. But it is still not this we spend all our money on, to put it that way. 10:40.998 --> 11:09.701 Then we have light and dark mode. There are preferences. Right? Yes, that is the most popular feature we have created, dark mode. It very quickly became known that the students wanted dark mode, and we have started to implement that on other services as well, because it is quite comfortable with dark mode. 11:10.179 --> 11:36.630 A new function this autumn is adapted instructions that can be shared. The new thing is that they can be shared. 'Custom instruction', or so-called system prompts, can now be shared with other users of GPTUO. This is a way to give the model a specific behaviour. 11:36.647 --> 11:59.036 You want it to give you a type of answer, or answer in a special way, then. What we saw was that teachers used the sharing of instructions in Canvas and other e-learning solutions we had. In the beginning we allowed quite small instructions, now we have increased the number of signs there. Because what we saw were lecturers were specifically for teaching. 11:59.036 --> 12:25.947 I found an instruction on UV that was about teachers training to talk to angry parents about a problematic child. I tried to start a conversation subject. Then I got a wash bucket back, and very bad atmosphere. 12:25.947 --> 12:55.009 Yes. Can you switch to light mode now while showing? I think it's easier to see. 12:55.026 --> 13:23.234 It was a question at least. Then we will do it. There we have a long instruction that you have created or that others have found. Yes, I have found this, just as a simple example, then. Teachers now get the opportunity to create their own subject-specific simulations or otherwise, which they can then share with students. So it is done quite simply by copying a link. 13:24.155 --> 13:52.449 And then other users can choose to subscribe to this instruction. When the teacher updates it, will all the students get it updated? Yes, right. Then you also see in the menu here that those I subscribe to have their own icon. And those I have shared have another icon. 13:54.172 --> 14:22.415 Have I shared, I think. Yes, right. And then I also subscribe to NA, which is from Magnus. You can show how it looks when you ... You can open the laters formulation. 14:23.575 --> 14:53.148 Yes. So then you have the opportunity to copy it and make your own and edit that instruction. Because otherwise you won't be able to do anything about it since you're someone else who owns it, then. Yes. Can you see the content of the instruction? Yes. But we should try to make that hidden, yes. Because it's a lesson. Yes, it's very nice teaching so the students don't know what we've asked not to answer, for example. 14:53.166 --> 15:21.476 Do you have an example of how he used it? Or intend to use it? Often we would ask it to be an e.g. a Socratic approach to not giving the answer. If the students know you are not giving an answer, then you don't bother to use it. So it's a bit of a thing. Or don't go through it and that is not the syllabus. Which is useful for the students, because then they know how you can delimit. So it is appropriate to have parts of it a bit hidden, then. 15:21.988 --> 15:48.660 We know that for medicine they want to use it to find out what is wrong with people. How do you feel, do you cough? I think you have mycoplasma. And then actually came the answer afterwards. So of course you have to give it. But it doesn't work when you can read the instruction. And this sharing instructions is also relevant for some of the environments that we collaborate with at UiO, who want to use this for learning purposes. 15:48.677 --> 16:18.660 One of the needs there is to be able to hide the instruction for the student. So that it comes and becomes a choice, then. The way we have financed the further development of GPTU is mainly that environments at the university have applied for funds, and then we have taken and further developed GPTU. That is the way we will finance the further development of it. 16:18.677 --> 16:48.575 Yes, and Simsambot wants to use LLM, i.e. with GPT-UiO as a limit average, for training in e.g. patient communication. Where you should be able to practice making important conversations with practicing being a clinic. And e.g. be able to convey a difficult course of disease to a patient. Make it a safe environment. 16:49.701 --> 17:10.947 There are a number of things you cannot practice properly. Among other things there was a case around giving death messages. It is very difficult to use students to watch and practice it. But then there was, among other things, a lot like you have to go through this in that order, you have to first say who you are, etc. Then we'll test it out a bit. Yes. Correct. 17:11.527 --> 17:36.084 Instructions have been focused on a lot in the past year. Because there has been so much concrete feedback about the use of instructions. And the technology behind it is then instructions and rag. So retriever argued generation. 17:37.705 --> 18:05.981 Where it is about optimizing the output of the model by using authoritative sources. So that you can upload an authoritative source that the model takes into account when it submits an answer. So that you get better quality of the answer. So you have a knowledge base. 18:07.739 --> 18:33.831 Did you find out how many users we had? Yes, approx. 37,000. That's a lot. But then it's probably also with the other institutions that use ... Yes, it's across everyone. But with 37,000 active users, we must be able to claim to be a much-used GPD service. 18:34.804 --> 18:55.606 We have received a couple of questions. Do you get a better answer if you write in English, or is it approx. equal to Norwegian? That is what you should define as a good answer. Linguistically, it is better, that. 18:56.988 --> 19:16.664 But the nuances in the language are better in English. So I use English a lot. And then when it becomes a final product ... To say it in English ... It becomes translated into English. No, in Norwegian, I mean. 19:18.234 --> 19:46.442 But it is subjective. But much of what I have practiced is also English. But there will also be good Norwegian language models across the country. Yes, which will be on, absolutely, yes. I use it a bit to write English, because I'm not too fond of it. If I'm going to write English and mail, then I write it in Norwegian. What I'm not good at are the Norwegian culture and social things to understand based on Norwegian context is my job. 19:46.988 --> 20:13.729 If you make him write a complete letter, it will be completely unfair and pompous. It is like a slightly different way of writing on behalf of the University of Oslo. I remember how you trampled him, then. It could be that you did it then. Yes, but I just wrote it, yes. Shall we see. Is it appropriate to increase the number of signs on the instruction? Absolutely, we have done that several times. 20:13.746 --> 20:42.841 It's just for playing. Follow the online form. There are some environments, especially in medicine, which use instructions that are longer than what is supported now. We will do it. Then it was a good answer. We assume 7-8000 signs. 20:43.251 --> 21:00.179 Yes, because now there are two. So then we think we might go up to 10,000. 21:03.302 --> 21:29.582 It has emerged that ChatCupT and similar KM models use a lot of energy. What kind of relationship do we have with the use of energy at GPT Unio? Again, this with having a sensible load balance. We tune the load, so that it doesn't be too intense. So we have a conscious relationship with that, I could say. 21:29.599 --> 21:36.783 Do you place ethical restrictions on what one can ask? Are you allowed to e.g. simulate adolf Hitler? 21:37.363 --> 22:03.422 We have entered the setup ... We have no control in our web front. But there are some limitations at Microsoft on the models we use there. There we have turned off everything available of filter, really. And taken the most liberal opportunity we have. And it has been following feedback that at the university we should be ... 22:03.951 --> 22:25.350 We follow ... Microsoft has its limitations, but in the beginning we had moderate, and then it was like a lot was stopped. It was actually quite troublesome. He stopped weird things, then. You are not allowed to do this, sort of. If you make nuclear bombs, you are not allowed to do that. But there were so many strange things he didn't want to ... 22:25.350 --> 22:46.988 He wasn't allowed to simulate a child, e.g. if we were going to use it as training. Then it didn't work, then. Then we've really just turned it off. In that sense, we are quite liberal, at least with us. But other institutions may have done it differently in their back end in Asher. So we're just a step further? Yes, so we take a lot of time. 22:47.602 --> 23:16.049 We can also mention that we have tested local PLMs, which run on our infrastructure. Then it has been LAMA31, as a proof of concept. It is therefore Large Language Model Meta-AI, i.e. MetaSINN, then. It should really be called LAMAI. 23:16.749 --> 23:45.657 But this is the best model we have seen so far. It is very interesting. The interface to GPTVO is now adapted to be able to choose alternative models. So the work hypothesis is that it is possible to set up the largest model for the llama on our calculation cluster. 23:47.466 --> 24:17.261 And research data. We see that there are a number of researchers who are skeptical about sending data out of the country. Because you may not fully know the value of the data yet. Although it is not sensitive data, this may not be the data you should give up. It runs on our DevService, so we can only short. 24:17.824 --> 24:46.971 This is Lama with 405 billion parameters. We will see. See if we can choose another model. Yes, so it's up and running now, yes. It's fun. It's still going somewhat slowly, it's in full swing, I would say. But it works. 24:49.480 --> 25:17.807 The biggest Lama 3 model has done very well in performance tests. It should be very good. If it can be used up to red dates, e.g. to run on our local infrastructure, then it could give added value or more value for the service. 25:19.480 --> 25:46.049 We envision that we can then allow more or less infinite amounts of data. That gives a bit in relation to being able to upload entire books and Ibsen's collected works. Correct, correct. Yes, then I think we can jump a little further to how we envisage further development of the service going forward. So a road map. 25:46.971 --> 26:08.558 We have had a lot of demand for filing. We have held back a bit here. This has been available at IR Ashore in the past. But we are unsure of how much it will cost. 26:08.814 --> 26:34.036 Another thing is that files are a bit scary. One is the content of the files, but we also have metadata in the files. So we are a bit skeptical about giving up files and what happens to the files afterwards. So we will now tell how we think we have solved this in a rather elegant way. Among other things to avoid us giving up. 26:34.053 --> 27:02.329 Metadata is stuff that is stored on the files, and it often contains who has edited those files, who owns and created them, and dates. And there may be things here that are a bit scary, and the users probably don't know that either. So it was still a bit important that this was something we didn't hand over to Microsoft. And this is a very exciting technology with machine vision, with Computer Vision i.a. 27:02.995 --> 27:25.623 One can imagine that a student takes a picture of a whiteboard with notes and a figure, for example, and gets it transcribed and interpreted, then. Or that a researcher loads up a data set, so you want to get an overview of the content and ... 27:25.640 --> 27:53.985 Take a picture of the whiteboard and then put it in a report, yes. It's great fun. I'm excited to see how it will work, then, but at least some demo here. I thought we'd show that too. This is something we are working on now, and will be the focus for the next sprint with development, which is in December. And this is then a visual presentation of a method for gene editing, called CRISPR-CAS9. 27:55.196 --> 28:23.234 Is it a Nobel Prize? It may happen, yes. Now we will see how the machine, with its vision, interprets and processes this, then. Then we upload a file, and now there is a picture, then. We already have this on the test servers, and do we think we will launch this before Christmas? 28:23.933 --> 28:51.988 Explain the method in a simple way. For a medical student in his first year, refer to relevant literature. That is the front together with the picture. But this will come in the next sprint? Yes, we are working on this next sprint. Around Christmas Eve, approx. You can do this well. You recognize what it is. That gives a good explanation. 28:52.500 --> 29:21.613 And refers to literature. It seems cool. This is quite cool. We are looking forward to getting that in place. It could also be a transcribed interview from Autotex, for example, which you upload as a file. And can get it summarized, interpreted, you can ask for themes and trends in the interview. 29:22.159 --> 29:47.858 Does it seem that people talk more if they have short questions? So GBT4 Omni with machine vision can also interpret handwritten text. If you upload handwritten text. Currently we have set a limit of 100 MB for file. Isn't that right? 29:47.875 --> 30:12.551 That's right. Yes, 100 MB. So we want to calibrate this a bit when we see more experience. And then pure text files and CSV format are supported. So simple text files and CSV. So it's reasonably complicated to make this happen. But we already have the basics in place. Have you used this, Ahmed? 30:13.319 --> 30:34.036 Yes, so experimenting, yes. I think it's very, very good. And then the fact that you can ... I really liked the notepad, so I've taken pictures of the sheet. I have drawn on and asked for help in structuring it. It is very nice. 30:34.053 --> 31:02.227 We have also tested by taking pictures of data on a whiteboard, we have set up numerical columns and such, and then say that okay, I just want to have this in a table, and then I want to have it processed, yes. Team code to create this interface with signs on the board? No. So maybe we can just show a little how we have thought the architecture. We have a ... 31:02.244 --> 31:29.497 We are afraid of the metadata and the cost picture if we only have a file that is known back and forth to Ashram with 100 MB. So we can perhaps talk about the architecture here. I can do that. This is what we are talking about when we refer to Retrieval Augmented Generation. 31:31.647 --> 31:58.899 The user uploads a source, e.g. a picture, a car, something like that. And then we complement the prompt to the user with what it has uploaded, to improve the generation. Because it is so unopened in our model. 31:59.275 --> 32:27.142 It does not have access to the Internet, and the size of the prompt is limited. So we have to limit what it gets in, or choose what it gets in. That is what Retrieval Management Patient is all about. Then the user uploads a file. And if it is a picture, then we send it to ... In this case, we probably use ... 32:27.534 --> 32:50.469 Which are multimodal and can take in that file. We send it in as a BAS64 string. If there is text, we can chunk it up. We only send the content of the file, not the meta data. 32:51.084 --> 33:13.302 We send a string. It represents the file, i.e. the content of the file. And what happens then is that we run it through one of these embedding models and get a vector. And we could in principle run it on a local model and get a similar vector. 33:13.319 --> 33:39.036 Do we get the vector on our own software, or is it from Azure? Today we use it on Azure, the same as we use others. But there is nothing technical other than perhaps capacity which means that we could use it on our own. E.g. with a Lama embedding or another multimodal embedding. Then we get a vector. 33:39.599 --> 34:06.084 We store that vector in a database, together with the actual content. When the user asks about something later, this question is also turned into a vector of the same type of embedding. And we can use that vector to search in our database for vectors that are close to what the user is asking about. And these embeddings, these vectors ... 34:06.101 --> 34:35.845 They concentrate the meaning in the content in a way that the models understand. Then we can extract content that is similar to what the user is asking for. At the end, when we ask the model or ask it to complete the chat, we then add e.g. the user asked with ... 34:36.032 --> 35:05.418 And based on the weight of our search in the database, we have these results. Incorporate it into your answer. So, use what ensures information. We already know this. Take it into account when creating your chat. So in a way it works in exactly the same way as if you had used an instruction, or if you had pasted in all the data in the same way. 35:05.828 --> 35:34.582 But by loading, and then we keep the arrows on our infrastructure, we probably use S3 today, which is locally lost, then you can e.g. upload a file and then use it across conversations in principle. And you can collect different information from different files through a chat. 35:35.060 --> 36:01.510 So that means that if you upload a hexable of 100 MB, or 99.9, then all the data is not sent to ... Exactly. Only ... We can decide how much we will send up that file ourselves, but then you only send with what is useful from that file. And that is how it will work. So then we have to first ... 36:01.527 --> 36:30.179 We need to process that file first. So that after loading the file, it will take a little while, not very long. We have tested with an open biology book on 1000 pages. And it probably didn't take very long, but maybe minutes at least, then. You feel well through all that, and before it would be available for you to search for 36:32.005 --> 37:00.503 This provides many possibilities. It's very exciting technology. I have a question about whether it will be possible to e.g. upload a video and get out some information about what it contains. Right. And that ... In theory it could work, but it depends on, firstly, that the video is below the limit we set, and that the embedding we use copes with video. 37:00.520 --> 37:28.422 Today there are several multimodal embeddings that can bring in image data, PDFs, files, etc. There are probably some who also take videos, but there aren't that many of them yet. But over time it might work. VideoEye has ... The answer is really yes. It will probably come. 37:29.445 --> 37:56.510 So then you can first make the film, and then we can automatically make the book afterwards. And then the question is whether you can also search in my chat history? It could also be possible, yes. Then we have to go through all the chats and vectorize them in the same way. But it could work. Do we have any plans to do that? It's probably a bit like that... 37:56.527 --> 38:23.148 After all, prize time. Yes, we look at that type of opportunity. It's very good. This is going to be awesome. Then I can get the overside. So ... we can go ahead and say something about another feature that has been much wanted, especially by students, and that is the rendering of mathematical formulas. 38:23.729 --> 38:49.548 We almost get questions about that every day from students. They want to see the forms visually instead of expressing them in latex, syntax. Because some have difficulty reading it. So it works already. It's out on GPTU, yes. Yes, it's out. 38:51.340 --> 39:21.323 For now, you have to give an initial prompt for it to work. But we will add it as a system prompt, so that users don't have to ask about it every time. We can start a new instruction ... We start a new conversation. Then I choose instruction. Now I have already entered the initial prompt here. 39:21.340 --> 39:49.735 Like that. And here we have also listened to the users, then. They don't just get darkmob, they get formulas too? We work centrally at the user's side, so we are very happy with that type of feedback, then. Then we can ask it e.g. about ... can you ... 39:50.879 --> 40:18.609 Yes. It is that bit there that is the new one? That bit there is the new one. 40:18.626 --> 40:41.920 So the alternative was that this came up in latex syntax, then. So can you glue it in, right? But can you glue that formula in too, and then ask what the answer is, like? I think so. Oi, then there will be a small fix tomorrow. 40:49.633 --> 41:12.381 Yes, that's fun. 41:13.456 --> 41:29.991 Could you say a few words about my experiences around the teaching you work with? 41:30.572 --> 41:56.084 And also how good these models are at giving good answers. We are also sitting in KI across the board, if you can pick up the page that has been created? What is KI across? It is a network that is across Holocity, both study administrative and technical, but also Link, which is also ... 41:56.101 --> 42:22.756 Where you try to explore how KI affects and is. If you go up a bit, 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 kind of sentence it is. Then it's a little further down, then it's ... Yes. If you go a little further down, then it's like how to use KI as a teacher, and that's very interesting. 42:22.773 --> 42:50.708 Ethics and social responsibility are a bit around this. And also some legal guidelines for how to use it. And then the idea is also a bit in collaboration with KI across the board. And Link is also ... we have plans to have workshops at the institutes here about how to use it. And you are involved in this network as both researcher and teacher? Yes. 42:51.578 --> 43:14.121 As you mentioned, there are a number of funds that the project has not applied for. There are a number of amounts there. 43:14.599 --> 43:26.527 And there, cooperation with you and you will be in the middle of the development of UiO's GPT. 43:27.551 --> 43:55.691 That is the advantage of having control over the GPT we did, and can adapt it. We can tell about another project that also had funds, and which we will do. It was at our finance department. Then they posted and created their adapted tenant, that's probably what it's called in Asher, by uploading data to it. 43:55.708 --> 44:07.944 We have opened up so that employees at the university can get access directly to Asher. By filling in a form, we give them access to ours. 44:07.961 --> 44:32.517 They can use it for a couple of thousand kroner, and if they spend more than that, they get a bill. There we have 20-30 employees who have tested this. There you can e.g. upload data and train your own tenant in Asher. At the finance department at UiO, they have done this with documentation for our financial system, which they want to reduce the number of than 40%. 44:33.336 --> 44:52.517 And this was someone who had made a small front end on his computer, and then someone came and tested it on his computer. And then they wondered what it would take to get this on the web. Then they said that if you want to have it on the web, it has to go into GPTU. Then we have agreed and have just made an agreement with you that here ... 44:52.517 --> 45:09.019 It will be a way for them to choose this economy chatbot, which is specially trained in Ujord's economic system and knows all accounts and that they will see if this gets any number of support inquiries. 45:09.036 --> 45:33.865 We are now creating a pre-scaling. Then everyone can get their own specially trained chatbot. We offer this in Asher, that you can train it up there. I think he had trained this and spent NOK 10,000 on training it. So it was somehow not that expensive. He has uploaded a lot of documentation and such. 45:33.882 --> 45:53.063 But of course we want to have this on local GPU and on our supercomputers. So it must be one of the first things. Then we can imagine that we train one per subject. Then we can change it in society. Folks, I now have a specially created model. 45:53.063 --> 46:15.862 Here you will be able to get tasks that fit, you can just ask for it etc. It is not too far off, and it is a bit of fun that we see that we can achieve this now. We will see how it will be in the limit average at GPT. 46:15.879 --> 46:44.906 There are several good things happening in the future. One of them is alternative login to GPTUO for institutions that do not have feud. We have received a number of requests about it. It will be possible to log in with Educloud according to plan. 46:45.128 --> 47:14.138 It is our research solution, where there is a separate user database that can easily create users. We have our second KI service called Autotekst, which is used for transcription. Which runs whispers models locally, which quite a few have bought access to. It was only available for those we say it to. But there were so many people who wanted it, so we have almost 60 institutions that have bought it. 47:14.138 --> 47:32.978 And where we have used EDU, so they can control their own users if they want it instead. So now we can offer GPTU to ministries which have wanted to use it. 47:32.978 --> 48:02.125 I look forward to offering this to everyone in Norway who wants to use it. It is a cheap service, but you pay for the consumption yourself. There are also research environments in the health sector, i.a. at the hospitals, who have wanted an alternative loan to use the epithelium. It's something that's coming. It is estimated to be early next year. 48:03.575 --> 48:33.285 And organizing the conversation history is something we want to further develop. So get a refresher of the design on it to be able to more easily e.g. delete more messages at the same time. Or more conversations, then. And we are also following new models from OpenAI and Google GameMeny. So we want to be able to offer the best, newest models. 48:34.462 --> 48:43.677 How do those who have bought UiO GPT out of us proceed to gain access to new models and new functionality? 48:44.599 --> 49:06.288 We have taken some self-criticism for it. That we have not informed the other customers well enough about how we have ... GPCVO has been much better than many others' models, because they have not made changes in Ashore. When we make changes to our setup in Ashore, then we will send an explanation to all our customers about what we recommend now. 49:06.288 --> 49:35.213 We will become a little better at informing about that. Those who use GPTU must have someone who has admin in Asher for that institution. 49:35.725 --> 50:03.251 Resonate, i.e. think through before it gives an answer. So it can reason over more complicated tasks before it answers. I know many people who cannot think before they speak. That is to say that I can solve more difficult tasks in science and technology. 50:04.309 --> 50:33.712 And math and coding. So we look forward to being able to try the latest models. Do you have any more questions that were good to take here, or? Someone asked if there will be a phone app? Yes, so we are responsive here now, so you can finely use it on a smart phone. It's not a native app, it's a ... 50:33.729 --> 50:56.749 So ... well, we can say that, possibly. We'll see if it has ... If you can talk straight into the PPT view. It might work on a mobile if you press ... You can do it on the mobile apps for part of them, then you can talk to it. And then you can say back. 50:57.039 --> 51:25.623 We will certainly not rule out ... Whatever this person says in Norwegian, then you have suddenly created an app for it? It was very difficult to tell me about society when I was young. After all, we have universal design and want to get even better at it. So being able to offer speech, text and the other way around, it can be interesting. 51:26.169 --> 51:43.609 Are we going to offer API access to the local models we set up? Yes, I think we will get that done. We want to do that. We love to offer these models. We hope we can achieve that. 51:46.681 --> 52:11.459 We are now running local models for some research projects inside Edu Cloud. This research system we have here at the university. But then it will only be run for one project. But then there is only one person who can use it, when he is logged in and uses resources and resources on it. So we have some testing going on around it, and it's a bit of fun. We have, among other things, some at HF who had ... 52:11.459 --> 52:22.210 Images of hieroglyphs or petroglyphs. Something old written language which they now saw they were able to interpret in a completely new way. 52:22.227 --> 52:48.166 It has never been transcebeded before, and you don't want to give this gold. We don't want to give research data to cloud solutions. It would be expensive to do that in Eskil. They had a 10,000 photo of a mass print that they were supposed to interpret, and they had managed to do it. And now we help them do it in the spring solution with a separate GPT model. So it's also a bit of a direct result. 52:50.299 --> 53:16.271 Shall we say a little about how external institutions that want to adopt BTUO proceed? The prerequisite is that you must have an agreement with Microsoft, with Asher. I think everyone in Norway has ended up having that. Only that there are very few who use it for so much. You have to know someone who has access to do the exchange of keys that you have to do with us. 53:16.783 --> 53:45.179 In addition, it is easiest for those with fades, because then you can turn on access for an entire institution. We think it is very nice that we have been very clear about that, that we will not distinguish between employees and students here at the university. Everyone should get the same access. But then it is now becoming impossible to manage users more one by one user by getting an educal user database for those who do not have errors. 53:46.374 --> 54:10.043 You can go to our website for information on how you contact us. We see this as a joint service in the sector, more like a splicing scheme. It is important to emphasize that we are not a commercial player, but want to cover self-cost. 54:10.742 --> 54:21.254 Yes, we guarantee that when we sell it today, we will not earn any profits from it. But we hope that it can help cover some of the development that we expect to do in the future. 54:21.271 --> 54:48.848 And you can also bring funds to us to further develop GPTU. If other research environments outside UiO want to do such as the migraine here or other adaptations, now we will do a lot around instructions for a concrete project at UiO that has funding. But we also have the opportunity to do this for external use of GPTU. 54:48.865 --> 55:18.780 BI has been involved and partially funded this by being able to release it to other institutions. So they were one of the first on the solution. And the premise is that we can then give it back to others who want this. 55:20.009 --> 55:32.432 We hope that people have benefited from this and thank everyone who has been there. 55:32.892 --> 55:59.940 From my side, it is interesting if anyone thinks about pedagogical use. Then it is just a matter of getting in touch. You write 'Omid interim EFI' in Google, then you will get straight to Omid. Then you will get straight to him, not worse. Then I am very positive about taking in, both from UiO and externally, where we can contribute to a boost in the knowledge of the use of Way. 56:04.360 --> 56:21.203 It's a lot of fun. I think we'll say it like that and say thank you. Thanks for us.