 Welcome to this session. First of all, we're going to excuse ourselves for the contrast. There is no possibility to dim the lights. So sorry for that. Today we're going to talk about our new colleagues. And these colleagues are not human. They are AI tools. So first, a little bit about us, the company. Sorry. And then the context, the setting. What are we actually talking about? How it started? How we onboarded our new colleagues? Some initial learnings? And then a presentation of our new colleagues. How does it actually work? Our concerns? They are many, but we're going to only go through a few of them. Next steps, what are we doing at the company? And some takeaways. There are some free seats up here if you want to come in. So a few words about us. Once upon a time we were called Node 1. Back in 2002 we were formed. And then in 2012 we became Wunderkraut. And back in 2019 we became Digitalist. And now we're going to have a little bit of a bragging slide. We're really proud to be a great place to work, certified. And we're also certified in quality, environment and information security. That's it for the bragging. So, who am I? I'm Ulrike Mingsol-Vedelin. I'm the CEO of Digitalist. And I have my colleague here. Yes, and my name is David Hornblum. And I'm a Chief AI Officer. And this is how Mid Journey interprets us. So, what are we talking about here? Yes, and to set the context of what we're talking about here. We have, of course, artificial intelligence, which is a very broad and multi-disciplinary field. That basically is all about creating artificial or synthetic intelligence in various ways. And then we have a subfield of that. All right, sorry about that. We have a subfield of that. There is machine learning. That's actually about having systems learn from data without explicit programming. And then we had a further subfield of that. It's called deep learning, which is utilizing neural networks in different layers to have deeper learning from data. And it's over here we start to have a really hard time seeing exactly what's going on and debugging what's happening. And then we come down to what we're going to talk a lot about today. There is generative AI and that is actually being able to create new content based on these learnings. And one of the ways we can do this is we have a very large data set, a lot of data on the internet, a lot of text images and all that stuff, which can be used then to create new content, like a new book or something. That's one of the big things that has led to this generative AI revolution that we see right now. Another major thing was, of course, in architecture we have the transformer architecture and self-attention, which was a big step back in 2017 that led us here. We also have other things like we have a lot more powerful GPUs right now. We have really great matrix calculations and things like that. So there's a lot of things that has led us to where we stand right now. This has happened fairly recent. So we don't claim to be experts in AI research or development for that sake, but we want to become experts in using AI tools and how to incorporate them into our business and to support our client's needs. How did this start? How did we onboard our new employee? So it started back in March this year. We had a management team meeting and we discussed what was happening. ChatGPT4 had just been released and it was this big bus in the tech world. And we said, hey, we really need to do something about this. So what we decided was that everyone in the company should start using AI tools and the company should pay for all the licenses. And what we also did was we created an AI policy with ethical guidelines and also what data must never be shared. And this was really, really important for us in starting using these tools. We appointed a chief AI officer. We started to have meetups around AI in Stockholm where we based. And we also, on every staff meeting, we have staff meetings every Monday. We have an AI section. And this is really part of our awareness training. And we think the awareness training is essential in this. And we also created a home on our internet for our new colleague. So some initial learnings. First of all, we created a motto. And when we start doing something, if it's something new or we're going to create some new content or anything like that, we try, it's not easy, but we try to pause ourselves and say, hey, can AI help me with this? So we try to do this in every moment, every part of our work. And six months later, these are our learnings. Yes, and some real life example. Here we have the happy developer. And this is something that happened. As a developer, you might get pushed into a new project. It's a tight deadline. There's no documentation. There's no colleague or no other part to talk about the project. And you've got to get it done in time. And it might be that this part of this project is written in frameworks or languages you're not familiar with. So it's a problem. But here we've seen how AI can help explain the code, help debug the code, and make it easy to kind of talk to documentation and fill up the gaps from lacking documentation or another part to talk about. So we've seen how that has been a really big help. So we're happy robot for that. And we have the efficient product owner. We're going to give an example later in, but for backlog refinement, for instance, it's really, really good. Happy robot. And we have the lazy CEO. They might not, you know, be confident about writing in English, but instead can use her native language to write and get a starting point on her document and all the, you know, boring bureaucracy documents and policies that need to be created. It's a good starting point. They're not boring. And the confused CFO. This is actually not an example from work life, but this is an anecdote of the first encounter our CFO had with ChatGPT. So it's from her private life. And in, you know, tomato plants. If you want tomato plants to actually grow tomatoes, which is kind of the reason for having them, you need to pluck this thing there in the middle. In Swedish, it's called that we steal them. So the CEO at CFO asked, how do you steal a tomato plant? And ChatGPT answered, I'm sorry, I cannot and will not give advice or guidance about illegal actions, including theft. Okay, so the CFO then tried to say that in Swedish this actually means, and ChatGPT answered, oops, I'm sorry for the misunderstanding. You're completely correct. The term stealing in this context means. So it's, this is just an anecdote where it can go a little bit wrong. And also we have the poor editor that suddenly get bombarded with different texts that looks good, but it doesn't know if, does it come from an AI? Do I need to proofread everything? We've seen there, we've seen many examples of people deliver texts. When you start reading them all through, there's a lot of issues with them. So that might be a sad robot situation. And also back in March, we thought that the client requests would start pouring in. They didn't. But we're seeing now that clients are asking us for support in how they can make all their valuable data into something new, that they can get insights for or support in their organizations. So it's happening now. And that is making us happy. So, we did a survey in the company a few days ago, and we want to understand how we're actually using these AI tools. And what we can see is that ChatGPT is really used, but also co-pilot and mid-journey. And we also asked how often do you use an AI tool? And we can see that about 65% uses it several days per day or several days per week. So we use it quite a lot today. And also the biggest gains, what were those? I saved time. 70% of us thought that we saved time. And it's also a good starting point for creating new content, to learn things, to get new ideas, et cetera, et cetera. So we use it for several different use cases. And if we marked, I saved time, how much time do we actually save per week? Most of us save one to two hours, but we even have one person who saved nine hours. That's a lot. And if we put this into some context for a year, it's about 80 hours per week that we save, about 330 hours per month, and about 3,500 hours per year. And that is significant. And what we measure ahead, does it really make us more efficient? Can we deliver more value to our clients? Has our quality actually improved? Have we gained any new insights that we would not have before? Have we been able to start up projects faster? Are our employees happier? And does it have any effect on the bottom line? We don't know yet, but we will investigate this further. So, let us present our new colleagues. First of all, the 24-7 Assistant. That's a large language model. And I'm going to talk about ChatGPT. So what do we use ChatGPT for? For content creating, creating first drafts to get critique on the content we create. We also create it for getting new ideas for brainstorming or solution scenarios. We use it for getting feedback, pros and cons of what we've written. And we also use it for expert knowledge. Notice that this is Dries. And I think this one is the one that you should be most careful with, because they are known to hallucinate these large language models. So be careful with the expert part. And what we usually do when we start using, start creating new content is we set the persona of ChatGPT. So for here, I'm asking ChatGPT to be a very experienced background product owner, have vast knowledge about unions, because that's the client, and they know exactly how to write really valuable and clear and concise user stories. And finally, I will now give you a number of descriptions that I want you to transform to the best user stories you have ever written. Are you ready? Absolutely. So what we do then is we have these requirements from the clients, and I think all of you know that they can sometimes be a bit fluffy. They are not very precise. It's more of a concept. So we give the description, and for here it's a union, and they want to show salary statistics to their members. And first of all, we get an epic, and it's a pretty good epic, I think. And then we get several different user stories, and they're actually written in a pretty good way as a union member. I want to securely sign into the union platform so that I can access exclusive content. And then we get some acceptance criteria. So we get several of these user stories like this. They're quite short, so what we want to do is we want to work a little bit more with this. We can ask ChatGPT to please elaborate on user story 5, for instance. And we get a much more detailed user story with much more detailed acceptance criteria. Of course, we can nowadays, a few days ago, we could start talking to ChatGPT in our phones and get text replies. And we can also upload images and ask questions about those images. For instance, this one, this is a little France, can I use this socket? And I think it lied because I can use my Swedish socket. But we can also upload really complex images. So here we uploaded a transformer architecture, and we asked ChatGPT to describe it. And it was pretty accurate, so it's quite impressive. And of course, you can do your own settings, the ChatGPT knows a little bit about you, so you can get more precise answers. And I think it's pretty good. Sometimes I turn it off because it's just talking about ISO certifications or our values, or it gets a bit tiresome sometimes. But what you always, always have to do is have the human touch. We can never, ever just take the output and put it out there. I would say we can get like 60% of the content and then 40% we need to actually work on. So the human touch is really, really important. And some examples, Microsoft Bing, OpenAI's ChatGPT, HuggingChat, and Google Bard. And the gains, it's a really quick way to get started. It helps with procrastination. It's always there. It's a good sounding board, and it's a really quick start if you want to create something new. And the pains, stress, and insecurity amongst colleagues. How do you use these tools? The ethics behind it? How are they working? Information security? What kind of data are you putting in? And it still needs human oversight. Yes, and then we have another example, which is the Thales Dev colleague. And one of the things we use a lot is GitHub co-pilot. And we have the first version of co-pilot, which was a tool that helps developers write codes by basically suggesting lines of code based on the context of what they were writing. So that was the original co-pilot. And recently it came out of beta, it's a co-pilot chat. It's basically, you can add it as a plugin for Visual Studio Code, for example, and then chat with your code. I will show an example soon. And then the upcoming is the co-pilot X. There's a vision of the new coding tool. And it's going to include chat and voice interfaces as well as a lot of comprehensive set of tools with specific for devs like creating pull requests and documentation and unit testing and things like that. So here we have a video, you can see this, where we have some very basic PHP code. And we can, with co-pilot chat, ask it to explain the code. And we get a full explanation of the code, what it does and stuff like that. It also finds that there are some issues with the code. So we can ask it to fix the code. And it will do that and suggest some new code. And as we can see here, it fixes both spelling error as well as logical errors in the code. So we can switch to get the new fixed code here. And we can, if you like to, ask it to simplify the code. And we get a simplified version that might not always be better or more readable, but we can get that and also switch that into there. And then we have other things we could do as to help us write the unit tests, for example, for the code. And you will get some examples of unit tests. Of course you need to read through that to make sense, but it's a really good starting point. So this, I feel, is a very powerful tool and a lot more powerful than the original co-pilot that was basically just suggesting codes. Yeah. And besides co-pilot, we have OpenAI's advanced data analytics, which basically enables Python code execution, data analysis, shorting, and math problem solving, which used to be an issue with a lot of large language models. You can execute this in a sandbox environment with file transfer capabilities. And this is only accessible by ChatGPD Plus subscribers. It's restricted to Python. No internet access for external library downloads. But there is an alternative. This is Open Interpreter. It's an open source local extension with full internet, you know, no time file size limits. It's natural language terminal interface for a lot of nice tools for developers. And it's suited for developers to desire some more flexible local environment to leverage. They can use with both advanced data analytics for OpenAI or some other model like Lama 2, for example. So you can run this locally with your whole code base, basically. So very interesting. Some other examples of code assistants is tab nine, pieces co-pilot. We have a code whisper from Amazon. Of course, OpenAI's code interpreter, now called Advanced Data Analytics, and, yeah, GitHub co-pilot, as we talked about. And the gains that we see, I mean, I think it's great as a pair programmer when you don't have a real one available. It's good for initial code reviews when you don't have that available. It's really great for chatting with documentation for your quicker access. It's just as googling for simple syntax and templating. It's really great for creating tests and other boring tasks and also good for explaining code that you might not be familiar with. The pains are, it still requires human oversight and a real code review. So you can't just let it do its work. You need to read it and understand it. It needs context to be really effective in, you know, more complex use cases. And these tools are still very much a work in progress. And basically, in the last months, we've seen a lot of cool things coming out. And our third colleague is the Media Wizard. And this is, we're going to see an example of AI inside another tool. And it's Adobe Generative Fill. And this film that we're going to watch is created by one of our UI designers because David and I are really, really bad at these Adobe stuff. So this is a film, a photo of a car in driving in a beautiful mountain settings. We're going to do a widening of that and fill that with that beautiful landscape. We are going to remove that ugly car and its shadow. And let's remove the road as well. And we're going to give it, going to implement a river instead. So now it's looking better. And the person doing some fishing, that person or that person? Yes, that person. They can take a look at it and yes, yes, yes. Ooh, very, very nice. So that's how easy it is for these, sorry, the people who actually knows what they're doing with photos and videos, et cetera, to fool us. And of course that's with the big fake and fraud and so on. What we're also using Media Wizards with is marketing. So we're creating like new content for us. These are examples from our LinkedIn. So all these images are created with Mid Journey or Dolly or something like that. And also a few words about Dolly and Mid Journey and Humans. I entered a prompt, a woman with short blonde hair, 54 years old, doing a happy dance. Black turtleneck, black baggy pants, black glasses. I have no idea who it is, just a general person. And this is what Dolly too created and as we can see, it's a bit weird looking but still, it's quite impressive. And this is what Mid Journey created. And I think this is one of the problems with Mid Journey because all of these women look like supermodels. It's really hard to get regular looking people, I think, from these tools. But also a few days ago or a week ago, Dolly 3 were launched and we can see how big of a leap it was between Dolly 2 and Dolly 3. And of course, in Media we can also use videos, create new video content, create new music, we can create new voices in different languages with different people talking. So it's moving fast. And some examples, we have Leonardo AI, we have Adobe Firefly, Dolly, Stable Diffusion and Mid Journey. And Gaines, the happy robot, it's really fast to create images suitable for our message we can get the base look and feel. The tools are more and more integrated with these AI tools. It's great for concepts and inspiration. And sometimes you get really, really odd by what the AI is creating. Paints, copyrighted material, pretty bad at human anatomy, for example, hands. And it's still really, really bad at text if you want to have like a text that's fine or something like that. Mid Journey really is really bad. Dolly 3 is pretty good, actually. And it's really hard to get regular looking people. So, how do these tools work? Yes, how do they work? I'm going to be very quick here. So this is fairly complicated. But talk to me after this so we can talk more about it. But let's start with the large language models we call the 24-7 assistance. Well, they are data-driven. They are trained on vast datasets. Trying to capture the nuances and the patterns of human language. They are generative, so they can create coherent, context-relevant text based on these patterns that they have learned. And they are context-aware. So they can use the context to generate meaningful responses. And I will show some examples of that. They are very versatile. They can be adopted to various tasks. As translation, summarization or question answering. One of the most important things on how language models learn in the same way as we do. For instance, the word it, it can mean different things based on the surrounding words in the context. So let's say we have the phrase, the dog chase the cat. And then later on in the text, it says three. We will understand that it means the cat. And that's what the modern language models will understand as well. So it's not only an expert generator. So another example if we give the instructions that the patient has a history of severe allergic reactions to penicillin. Then the language model can then advise that well, it's crucial to avoid medications containing penicillin in this case because we gave those instructions even though it knows also that penicillin is the most useful treatment. So it will understand this long range context. And this is done via the transformer model and self-attention mechanism. That was the 2017 kind of revolution to enable this possibility. And before that it was basically an expert generator and the early ones were like that. And then we have the media wizard or what's called diffusion models. This is a lot more complex. They are probabilistic. They model the data distribution as a diffusion process, representing data as a random work starting from noise. That probably doesn't make sense to most of you, but I will try to explain soon. And they generative as well. They can generate new samples that are similar to the training data. They're time dependent. They have a sequential process. And they are very versatile. They can be applied to different types of data. So it's not only text or images, it's also audio and video as we've seen. So what does it do? So if you think about this process, if you think about a drop of ink in a glass of water, if you start from one point it will spread to an even distribution in the glass, the ink that is. So think about that process in reverse. Basically, training on a lot of images, for example, and then adding random noise to the same point. Then you can reverse the process and then basically create new images or new media based on that. It's a very complex process. If you really want to learn about it, it's easy pixel based coding alternatives so you can follow the whole process. But that's basically how mid-journey for example creates these images. Also it implements large language models to understand the prompt what you want. That's very shortly about how these things work. So our concerns about using these AI tools are new colleagues. As said before, we have many concerns. We are just going to talk about a few of them. So ethics, it's a really vast field. We are not used to actually talk about these things or we could treat it like just a hairy monster and put our heads into the sand or we can get really anal and fall out down in a rabbit hole and almost never come up. But anyways, first of all, bias. These tools are biased. And I did a little experiment. I had the same prompt that I put into mid-journey every Monday for 20 weeks. And that prompt was a happy CEO reporting the last month's financial figures to the board. There are five members of the board. The CEO and the board member sit at the same table. Settings in a modern office, etc. So I didn't have any specifics about these people. But this is the result. Week after week. And as we can see, the CEO should be white. It should be male. It's really good if he has a beard. And it's also very good if it's good looking. And look here. What happened here? All of a sudden we got a stuffed animal. Yeah. I don't know. But we got 80 CEOs in total. None of them were non-white. None of them were non-male. And we got about 300 men in the board. So the training data is everything. And also the people working with that training data. How they're teaching it. And jobs. This is a big one. It will affect all our jobs. That's for sure. Will it actually eliminate some jobs? Probably. Probably some jobs 5 to 10 years from today that will not exist. But it will definitely affect all our jobs. And what we saw in the summer. The strike for the script writers in Hollywood. They were really, really afraid that they were going to lose their jobs. And competence. We need to learn these tools. But also will we be over-reliant on the tools? Will we not develop our own skills? And what about critical thinking? Will we just take the output as is? The law. We are deploying. No, we are not. The big corporations are deploying these AIs into the world. And there are no regulations yet. So please, we welcome the regulations. We welcome the laws. Fraud and fake. We've seen this all over the place. And for instance, there was this on Amazon. Travel guidebooks. All written by AI. And it was completely hallucinating. Also this image from Amnesty International where they actually used the mid-journey generated image because they did not want to actually have real people. But they got critique on that. And I think all of us have heard about poor Tom Hanks being part of a dental plan commercial. And transparency. How do these large language models actually work? How do you debug them? It's really opaque today. And the hallucinations. They can be huge. And also really mansplaining. For certain, this is the case. And of course, corporations. It's the large corporations in the US, in China, India that are deploying these. We don't see any states deploying large AIs. And why are they deploying it? Is it for the public good? Probably not. And energy consumption. It costs a lot to actually write the prompt and send it in to one of these tools. It's a lot of computational powers that takes place there. And personal data, information security. The data we send in. And what we did, we created this AI policy. Data that must never be shared with AI tools. And also the ethical guiding principles that we have at the company. And it's so much more. So, the next steps for us as a company. Yes, so what are we doing and investigating right now? Well, we have our collected knowledge database. We have an intranet with hundreds of documents and policies, routines, guidelines and a big knowledge database. So, basically what we did is we took our own knowledge database and built our own AI service for it. Using Lama 2, which I will get into soon. We can create a self-hosted chat bot that is not sharing data with a company in the US for example. GDPR and Cloud Act compliant in this way. It's fully integrated with our own data and everything new that's coming in. And we can do different tasks and data analysis on this data. We could, for example, search and find personal information that shouldn't be in there. We should try to find factual errors, all kinds of things. So, to do that, we used Lama 2 which is a new area in open AI. It is a model developed by Lama 2, etc. And it's an open, large language model. It's not open source, but it's open with variants of 7, 13 and 70 billion parameters. And its core tasks are basically text and code generation. And these, especially chat models have been fine-tuned using 1 million human annotations. That means that humans have been involved with trying to make sure that it has good responses and stuff like that. And it's released on a very permissive community license. So it's available for both research and commercial use. It's not an open source, but Meta cannot surprise everyone by letting it go freely. So you can use it and build commercial products. They have some rules and cannot use it for training your own data to compete with, let's say, but for our use cases it works great. And hopefully this will also push other big companies to release their models because there's no possibility for normal company to develop these themselves. This costs extreme amounts of money. And this can also be very easily fine-tuned for different use cases. In one way to do this, there's an AI community kind of like the Drupal community. It's called Hugging Face. So it's a collaborative platform for developing, training and deploying machine learning models with OpenTech. So you have different, you have models, you have data sets and applications and examples, fine-tuned models for different use cases. You can find there's so much to be seen on Hugging Face. So if you haven't checked it out and it's embraced by over 50,000 organizations really widely used including some of the big ones like Amazon, Microsoft and Intel. So we're ending the presentation with some takeaways and we think it's quite simple. Set the guidelines. Make sure that everyone understands the guidelines. How are we going to use these tools? Have continuous awareness trainings. Talk about it on staff meetings on one-to-ones. Be open to learn and also give time to learn because these are tools like just any tool. We need time to learn. Be compassionate with insecurities amongst colleagues. Discuss. Discuss and discuss. And start. Allow all employees to start using AI tools and we think also pay for the licenses and evaluate, learn and discuss more. Join us for contribution days on DrupalCon and please give us some feedback on the conference and on this session as well. And David and I say, thank you. Okay, we do have a few questions. This is working. We do have some questions. So let's see. Can you hear me? All right. How often and how do you renew and train the data for the self-hosted chatbot? What's the cost roughly? So this all depends on what you're trying to do. But I can tell you that Lama 2 for chat is really good at understanding text. So if you have depending on what kind of use cases let's say you just want to do something simple like searching for data in your vast knowledge database don't have to do much fine tuning at all. Of course, different languages is a lot more powerful in English than in, for example, Swedish, just speak. So if you have a different language you might have to use a fine-tuned model specific for that. But you can find many examples on Huggingface that probably will do exactly what you're looking for. So hopefully you won't have to spend that much at all. So if you have specific services that hasn't been done before we've been involved with that as well take some more work. But it all depends what you're trying to do. But these new models are so capable they can do a lot already. And the answer is I don't know. We're still fighting with this one. We don't know what data the tools are trained on and it's really hard for us to navigate in this. I think the coming months, the coming years will teach us how to do this. I can just add it's different in different regions. For the U.S. there were court cases where images from mid-journey couldn't be copyrighted that's going back and forth into the courts. So it's probably looking different in different parts of the world and your restrictions. So we've got to keep track on this. For question about developers how do you protect how do you protect that your developers are not just copying, pasting code that might be insecure it might have a secret in it it might have some other issue. Yes, and this is we show that we have policies for this and we got to make sure that all the developers and all employees understand this and that they really understand and are trained on this and how to use data. But I mean it's not really different from the information security that we have. Like if we're using github for example we don't put secrets in the repo. So there's the same kind of routine. So if you're putting something into chat you don't use the client's name so you would do the same even without the A.I. So it's the same kind of training. So this takes training, it needs to be clear for everyone how to do this but it's kind of like the same processes as you normally would have always. Yeah, we always have human code review. So when we're taking code review you might have an initial code review if no one is available and then you send it for real human code review. So we always check the code examples you get you need to understand them and if you don't understand what the A.I. generated for you you shouldn't really use it or publish it. Again I think it's part of the awareness trainings that we have that we talk about these issues that we cannot trust the output at all. We use it for guidance, we use it for getting first drafts and we are humans so the only thing I see that we can do is talk about it have awareness awareness trainings. That's a hard question but I mean anyway and even without A.I. we're using more and more resources. I know some people believe that A.I. might help us solve issues with climate change. I don't know I'm a bit pessimistic about A.I. in general but I mean what we need to do is try to understand it, work with it talk about ethics and try to push it in the right direction because we can't really stop it so we need to do what we can. Yes Yes, we talked about that the human touch becomes even more important but we also need to remember that means that it's going to be the experts that they really see older people that will do those jobs might be harder for juniors people coming up so we also it's going to be like an imbalance here some humans going to get a job like we have today so we really need to try to actively make sure it doesn't just cut away jobs which it will naturally do again like the environment we need to try to push it in the right direction Yeah and I think that human skills will become more and more attractive on the job market so yeah human touch Any more questions? Thank you so much for showing up and I want to give thank you to Cat as well for moderating us Thank you Cat