 Okay, thank you very much a little introduction about myself. Sorry. I think you see me already anyway I'm in the other sessions over there. So but I want to tell you about my work. I work at Consultancy called clinics week. We are working in Germany when we're working on Digital transformation data science consulting. I myself consult on data science on or training people on data science and all python So I'm quite active in the community apart from this conference I'm also there involved in your sci-fi and the German python conference and I really like working with speakers Reading all these submissions because it's awesome awesome stuff in there So I like to talk as well and train on this little scope. We cover as a company but on the I want to a little bit tell you about my story here in deep learning because I've been doing some experiments in deep learning here um and This is actually where the inspiration comes from so if you expect this is like a talk to explain How does deep learning works because you can make shitloads of money from that? You're wrong here. It's more like What you can do in experiments and what works and probably doesn't um in so this is where my inspiration comes from I I visited pie data London last year and there was gene Kogan gene Kogan is an artist And he uses AI for art and he gave a keynote there They're down there. There's a link for the YouTube video. It's totally worth watching and I said a wall awesome Because probably some of you know my first career was not in IT. It wasn't a music industry I'm not as an artist But I was a label manager and I was creating a lot of like techno house records at the time in the 90s And since we ran in the administration problem I was forced to learn what teach myself programming and this is how I ended up here So I said, okay. Wow, great. So you can understand that I'm constantly drawn to creative stuff as well So I thought okay. Wow, great. I always want to do it But you know you see a keynote. Oh, yeah, let's do it But then you get back to work and have stuff to do and forget but lucky me I had a project which was work-related and it involved improving the resolution of images because we had some old images which had like the resolution was not good enough for Current use because it was like images recreated for Digital media 10 years ago and also the whole thing changed. So it was like so I Used this library actually for not Nora and Hans Nora and Hans trains on a data set Train the network and then you can basically do a better job than Photoshop in improving Images here. That's just like a simple example. I've taken from from the git here. So I said, okay. Well So that worked. Well, so it was my so it was my first step into Getting it to work. So I said, well, okay I'm the prototyping went well. So I said, okay now it's time to get a eGPU something I wanted to get all the time anyway because you can plug now and in external a GP graphic card into a MacBook like this use Thunderbolt 3 It's a super fast connection and so hey, well, it's very good for making better prototypes and something you can release later in the cloud for production So so I got this this is basically my experiment setup I have in my home office are and I'm using this work work. So actually cool. So I got this on Some lessons learned here, especially if you're on Mac OS Apple always says hello one more thing and it's all plug-and-play and we have eGPU support Yeah, but Apple has some see eGPU support But it's favoring the other company not in video but for data science and all the AI stuff Nvidia and CUDA is way more enhanced and The thing to use for deep learning and the libraries So actually it was a little bit painful to get the right Nvidia driver because they really depend on the OS build version So it makes a difference whether you would have a fresh install of the system or you did it just like an update So there's a great source. I want to point to its eGPU oil and really a lot of kudos to Gulaka and FR34K They write some great installers scripts and a lot of help to making this work on Mac because I Didn't want to have the whole the cloud overhead So I don't I like to move stuff to the cloud But I wanted to have things close and not just like an extra layer which could be make some other troubles or Something like that. So, okay, I got this to work and and I had a weekend all for myself like a long weekend I was like my partner was not at home. So I could spend 24 hours working on nerdy stuff and This is where my story started. Um, I started with style transfer because this was something I thought it was really nice. Um, who has seen style transfer before in the media? Okay, it's only half because that's something from my perspective And this is also something which you will see this talk a lot from my perspective I expect 100% or 99% have heard about style transfer because it was big in the media But still there's only 50% so let's explain a bit and it's basically style transfer is something you have This is tubing at night tubing is a small town in Germany with very smart people And they came up with this algorithm and basically the idea is you train an image on and Van Gogh or a Picasso and then you apply this style to a photograph and basically you have tubing at night painted digitally by somebody like Van Gogh and This is quite amazing and looks really nice. So I thought, okay, let's experiment on and in the first step I reproduced some stuff. So I just took this down. I was like as a kid. I always liked these comics They're French. Also, they are called in German Valais and Veronique So we already had like in my early childhood in the 2000s. Okay, you know actually in the 80s we had The proof that the main characters was actually Two people and it was a man and a woman. So it's actually quite progressive Not only like one hero and there's like for example these modern flash comics Which I from the style on so much like they're a little bit soft. So I really like these old French style comics So I wondered can I Make the flash comics look like they're in Veronique. So How does this work? What is this? Okay, this is a page from an old Valarien, or it's called Valarien in French comic we take a part of the image on in Apply it on this other image From a modern Marvel comic and this was my very first style transfer ever and I just said, oh my god This is awesome. I mean, it's awesome. It's also like it's also so easy I just like I just took some code from GitHub use deep learning and Hey, I should drop everything else and just do deep learning in the future We also happen to be in holiday in Scotland last year So I had also these amazing Scotland pictures. I could transfer in a Valarien V comic or like the flash style There's some pictures I had taken while traveling in museums like the tiger in afraid of the rain in the jungle here And oh wow and it all worked really well And then of course I tried other styles. For example, I took a more like Psychedelic sequence from their Valarien and we unique on this work pretty well Either so wow This was really nice thing and another thing You can even transfer the style on the whole page So this is a full page of one of the modern ones and now I make it look like One of the style I favor Another nice finding on the way and this is one of the key messages as well AI and this is also another thing which probably draws me to AI and deep learning is a lot of both experiments So it's not there's like not recipes and there's some good practices, baby But still you have to experiment and another thing you have to think as well About stuff. So for example one finding on the way while doing the experiments I used I thought what what if I take an old flash comic like one from the like fifties And I only was able to get hold of like a really bad image of that So basically compared to the picture on the right side the original size was like this It's like really tiny super blurry and I have the most style transfer I could just make like basically brand new and this is like still not an optimized style This is just like try and do stuff So actually this is like really good finding on the way was you can use this for improving images And probably the use case would be if you have old books or like, you know probably like these thermal printed receipts from somewhere Nobody can read anymore To improve the resolution here another nice thing and also like brings Me to okay. Where does I'll start was it in what do you what do you get if you train on something else? and just put white noise in there and I just sent this picture. Hey guys to some friends. Hey guys Take more than I really favor this picture. What do you think and they said? Oh, yeah, it's a great picture And I said no my computer made that and Of course weird your python and we want to do spending stuff. So last night. I thought what about doing a your python comic So I let me tell you about Stan Stan pro cop. I found all the images on Twitter. So This is Stan. He tweeted Stan is his first your python. So he was like super excited So and he treated this picture and Stan. It's basically a little bit made-up story Stan enters your python So he meets some friendly python this is on the way the master of Ceremony Mark Andre enters the stage to greet everyone and says hello welcome your python and and Tells you about all these great keynotes and stuff and then Another magician enters the stage Dave Beasley and tells you about these threats. Yes it's German defects And and yeah, you see all the stuff and it's really easy to do something like this and Make a nice comic or Make it maybe this is something I felt or maybe this is more like some picture is suitable for something more psychedelic Because it's like blade runner read you can also like do another comic and in a different style Like this for some glad I think I owe stuff And all the stuff that most of the work was collecting the images from Twitter Or you can have like a picture like this Or if you ever wanted about quantum computing You can have like this Or you can also like have the cheese shop as a comic. Okay, that was the your patient your python special part So, okay No way to I said, hey, well, this works really great. Um, awesome deep learning So what do we have? Where do we stand here? Because a lot of this talks about the process and having like a proper process when working with stuff like that So I got my local eGPU lab. I got the ability to scale in the crowd in the Clouds not only David Beasley scales to the crowd Linting but I scaled to the cloud Reproduce the ability a documentation of the experiments that's like still like baby steps And this is also something learning on the way which was really important because we have parameters We feed a lot of parameters like learning rates or a random seed number which pictures I do we actually use input output and To you these experiments you have to be able to reproduce them So you also have to document and probably so I wrote myself like a little setup script to say, okay I this is the picture. I'm working on these are the parameters. We're using and all the stuff We probably could also use cookie cutter But at the time I did not think about it and of course there was the big question Oh, which framework do I use because there's TensorFlow there's stuff like carers or there's pytorch And so basically when looking at get what do I sell for I decided to use pytorch? It's because it's research friendly. It's I think very accessible for a Python this sir So I'm also no tensor flow and so forth. Yeah, well, it's a little bit weird so but pytorch you basically look at the code and know Okay, I can I can I can work with this and it also has a great community responsive community And it's it's backed by by Facebook, which was probably now. Let's see. This is a neutral thing It's backed by Facebook research. They had a good guys in New York It's okay. The lessons learned was he were here as well because I want to use my GPU on and and pytorch has to be compiled to work on this machine and this was actually Unexpectedly quite painful. It was not condo install or pip install anything with the GPU I had to compile it and I turned out pytorch does not have something like a master and a dev They always commit to the the new stuff to the master and it's a super complex project so turns out some other Dependencies they use broke the compiler at the quite late stage which was painful because stuff like that can take like half an hour So Actually, so I finally made it so I checked all the branches that of the the commits and found one that works and then I Used condor for keeping a master copy of a work conversion So if you don't know if you have something compiled condor is really good at cloning your environments So I wanted to keep that copy because it's been maybe a wasted one day on that unexpectedly Um Okay, then part two like the vision so you can imagine I feel like this I go. Yes. I transfer everything. I touched deep learning at it works. It's magic. So I feel like this and I got like quite Excited and said, okay, let's propose a talk for pytable and it's like a very prestigious pie data conference. I said, hey, well Deep learning can basically solve everything. Why not even? Do something really fancy and try to synthesize this and who knows this One two or three. Okay, 20% are German Yeah, it's one thing I'll explain so it is like what three question marks and three colors. What is this? Actually, it's called the Freifrage Zeichen it's It's actually an American Juvenile detective stories Books series. It's called the three investigators. They it's basically they investigate mysteries. They're like young Teenager or detectives on they're located in California. It's pretty old So it was very popular, but it's super popular in Germany turns out only in Germany because It's I don't know like and that's the other thing. I'm going to explain a bit. It's called her spieler in Germany Which is something like radio dramas, which is like then there's 200 taped radio drama. So I know I I know for deep learning I need content. I need a lot of content to learn from so I said, okay We'll like 200 ready dumb dramas be enough to get something working with Deep learning so Okay, so What's her spiel actually probably you have heard of awesome wells of all of the worlds not the movie the radio thing Okay, there's definitely some education to do Because this was awesome. I mean here you see they were like in America like when there was just like the age of radio I mean, we are pre-internet we even like pre-television now and there was this guy awesome Welles he also did the best movie of all time Citizen Kane Which is absolutely worth watching and he is like I think he's in the mid 20s And they were like these people like actors in in studios and they were like Speaking and and there was like a narrator and people made sounds so basically it was a movie without pictures at that time Transmitted by radio waves all over the country and the thing which awesome wells as the genius awesome Wells did when Hank he was like in his mid 20s. He made it like a live Documentary about aliens invading the US so and there's also like where the old Tom Cruise and stuff movies come from and and people were really panicking because at that time in the 1920s nobody was aware you have something News or stuff on the radio is not real. It's not news. It's not like and people really panic So this is like a super classic. So a her spiel Now which is still popular in Germany and for kids but also adults or so like you can have also like Harry Potter as it has In voices and it includes a voice recordings noises Stuff around so basically I think a movie without picture Basically brings it to the point. So and it looks basically like this This is one of the cassettes. I used to have as a kid the die Fargatization and the whispering mummy It's like a super is it this is like just like How like a transcript looks like there's somebody speaking and help help help me You know, there's there's some parrot speaking stuff. This is like something and it sounds like a little bit like this Okay, mommy's curse. This is like the pictures look like the covers looks like for like a CD covers look like and So actually there's like multiple dimensions here. You can try it. We will try to synthesize with AI There's a story in plot. What's what's actually happening. Can I make up a story on dialogues? human speech seeming dialogues are like Very dynamic actually a cover. Can I generate a cover for the next radio drama or the radio drama? I produce maybe and the spoken word. Can I basically? Create spoken word and speech sounding like a character and so let's continue your journey and see But how over excitement can also like there's some bumpers on the way So how can I realize this? So I have the transcripts luckily. I found the transcripts on the fan side So at human transcribe text, which was really nice. I had the recordings because I'm a fan So I had basically everything at hand and now let's get the stuff together So technologies are the locally GPU. So if I had something to working I could scale it to the Google Cloud like to because like sometimes deep learning stuff. It takes time. So The workload I sent to the cloud mostly the most recent pytorch within duped a lot and of course a recent Python version And the process for each of the different parts. I'm trying to synthesize now is always the same It's some data acquisition data cleansing get something and a research paper or something Some work from the internet first try to reproduce what is presented there and Then verify it's working and then adapt the solution for my use case So it's always like simple like take something make something to work and always exchange one variable at the time and not try to change too many parameters at the time because this is very confusing and You won't be able to know why did this happen or why it doesn't need happen anymore? So, yeah, and then try to maximize the quality once everything goes into the right direction So the first mission fabricate text to be spoken by a character But I mean remember style transfer and the whole picture thingy worked really well. So I said, okay, that's easy and so Yeah, so in it's also like this 260 296 characters there. I'm just like a little impression about the text corpus it's all in German language 1.5 million words and 56,000 of them are unique one little tip on the side Probably in the wrong room for that But pie charm is actually quite a make great in combined with Synelium to scrape scuff because I had to scrape scuff here And I use the DB the debugger from pie charm and Selenium. So Selenium is the remote on You can test soft browsers or browser software with Selenium. So basically I can automate my browser Combined with the pie charm debugger With all the hall points. I see what's happening. I can see where I'm with the source code and basically It it's it's very easy and fast to scrape stuff from the internet here So if you use beautiful soap, which is also great I used to in the past as well and still are digging through HTML code have a look here because this is like super effective because you actually see and can Interact with what you're doing Who has heard about Andrew Kapathy the Okay, who has read his blog post about RNNs? Okay, three four people. Okay, because Internet is a lot of stuff like this and this is some something really good Andrew Kapathy was still a PhD when he wrote this blog post. He's not a chief data scientist of Tesla And he wrote a great blog post about recurrent neural networks. So and Basically the outcome here. He's playing how is a pretty long blog post who explains really well How everything works with RNNs and what you can do and basically one of the findings or results is okay Feed it a little Shakespeare and you can basically generate new Shakespeare with that using RNNs So awesome. So if somebody can demonstrate in the blog post to Generate somebody like Shakespeare. I mean like it's not just like any it's like Shakespeare I Can do my child's play so I mean I can do these six checklists. I can do this as well So now let's get this going. So first I got something from the internet Produced something which looked Shakespeare-y to me because one thing you always have also have to consider You can show me a lot of stuff in English and can claim. Oh, yeah, this is medieval English or say Yeah, this is like some accent or like a local dialect. Mine cannot judge. I'm not a native speaker So but yeah looked good to me and this is what I produced. Um, this was something produced by a character based RNN so basically you feed all the characters and the distances From the characters into a neural network and it remembers like how likely every Other character in connection with the others and this is like one of the results from my tax corpus. And of course Why are the Germans not laughing? Okay, thank you, but I also Ask Google to translate it for an English. So this is like the English translation Actually, it's for accurate. It's a very good representation of what you see here in German So it really looks in a way like German and there are some real German works in it But Schlo Kuliana is a great word, but it doesn't exist So and also the grammar is like well the grammar we could really work on the grammar So actually this makes no sense at all. And so I said, okay. Well, maybe Let's let's do something word-based because why did I approach character based first because characters? I have to represent everything and as a number when I feed the RNNs Of course if I have a tax corpus I can easily like only work with like 40 characters But if I need to assign a unique number for each and every word I suddenly have like 50 to deal a lot with 50,000 and these things and these matrices we we work with they really blew up I mean it takes a lot more space and computational power and we probably also like break the RAM of my CP My GPU because GPUs don't say okay. I'll do this later I mean if you put too much stuff into your GPU RAM it will just crash and you start from scratch It's not really it's not as mad. It's not as easy to manage like a CPU on your computer So let's do something Yeah, I don't know another finding you can also like send it to your friends like Well, there's also like there's not too bad German Yeah, and then people ask are you speaking wrong and wrong is like a youth language in Germany Which is like really bad German like from people from teenagers who are too lazy to to work with so actually we could also say Yeah, this is like real AI caramba. Yeah as Vincent probably would say This is a picture taken here in Edinburgh and start transfer as well It was the pay data meetup in Edinburgh and Vincent did the great talk about something related It was talking on Also deep learning and AI and thinking is really important about when you do so So let's settle maybe it's just like the character thing. Let's do the word-based thing And I also like the below the courtesy the Google translation So which is also quite accurate. Although I don't know why Google decided to write Kalin Like in the uppercase or title case because Kalin actually above is a verb But I don't know just to give you like a some expression impression if you're not a native German speaker So native German speakers tend to find is like really funny. So I said, well Didn't go as well like sound expected so Another of my ideas and this is the fun stuff again. Hmm. It's about dialogues. What about chatbots? So but I didn't have time to build chatbots now as well So I was doing some research and I stumbled across this just like the Captain Picard like a Star Trek next-generation chatbot And there's also like a paper on archive acts about it And you can probably find some found some blog posts about it and you also said oh, yeah Somebody took all the text from Star Trek next generation and Star Trek next generation is over So basically they build a chatbot and you can while waiting for probably they will come up with 11 season or whatever You can just like Twitter and you get a checkbook. Okay, awesome And we read this stuff and we of course believe all this stuff. So I said, okay, let's try this and I Just treat it to them. I'm Lakuta's actually I tweeted something from the original text corpus there to make it easier to to minimize Anything special and it was basically just like your hoo boo boo boo and it was really hard to get this Yeah, I took like a week. I was got these tweet messages on body wall for plight or whatever But it was not really useless. So it was always like a great idea totally blown up by some blog post but not really working although Still thanks somebody to stuff like on this on Twitter like I don't want to say they doing a bad thing But maybe sometimes people pick up stuff from other people and think they need to exaggerate anything with AI currently which is really bad because people in management or Deciders they they tend to believe this they read only like the executive submarine said, oh, yeah I somebody did Star Trek. Yeah, obviously not at least not from the quality. I expect and I watched all of them Okay speech, how can I now fabricate spoken language and make it sound like one of the characters? So there's some And something really new it's called taco turn true to it's the open source implementation of the most recent Google deep voice they presented for the assistant on there was a really I had a nice natural Speech data set. So basically you always have a piece of audio and a text It's all spoken 24 hours one female speaker and I tried this and it uses male Spectrogram something I haven't stumbled across before a male spectrogram is basically a Representation of the frequency when I do like a pie or say word. Basically you I create a spectrogram. It's not just like Yeah, so it's more like a really nice representation and it learns actually from the spectrograms here Okay, let's do this. So I have have this trained and I trained it for like 10 hours and The data set somebody speaking something and I have just the text and no additional meat information after 10 hours It sounded like this So it's always like saying the sentence above so cut the 10 hours, okay? I cannot understand anything, but this is certainly rhythm. So let's see after 40 now 14 hours. It sounded like this More I Okay, something's going on here some things coming to life and after like five days of training And remember this is the model I own It's like 330 megabytes and I can feed any text into it and read it It's of course not the quality we see with Cyrus Siri or Alexa when they speak something but This is just like this was really really simple to do. So actually this was quite impressive So after nine days, this was pretty awesome Still there's a lot of room to improve I wouldn't really sell this to a customer and send it to a production say a call Apple forget about Siri I have siren but actually this worked really well and Of course, so I said okay proof for me the proof was I did this in English language, so let's okay Let's do this in German and the lessons learned was It's almost impossible to find a comparable data set in German I found some data set in German, but it was not really useful because it was not properly cut So actually I'm really missing data to do a proof of concept Here unfortunately and of course that's another things although German I think probably 100 million people in Europe speak German as mother language It's a big market. So we have everything like we have dubbed television. We have dubbed everything in German, but Turns out the majority of the papers the research the data sources is all in English Well, there's also one reason is in German the things work differently And it's I read it's also like a call to action to change this because yeah, I'm limited So I'm just working on how can I synthesize? How can I create my own data set and scrape radio and maybe I talk to the site which is a newspaper? They gave me a PR access so maybe I can generate some data set to learn from Myself because I don't have resources to cut audio or do like read all this myself So well, let's see. Um, I'll keep you updated keep you posted and of course for the style transfer there was another task and it's just I'm finding because People speak you remember when I played the the radio drama somebody speaks somebody else applies And I need to basically cut that by person if I want to extract How the sound how somebody speak if I want to learn and transform this on some other audio So I said very well great. I have the fan transcripts I only have to ask to good use Google For a transfer script to get these and then start times because I have the audio and with this and the start times I can automate Cutting the audio so if there's only just as Jonas speaking one of the main characters or Bob Andrew speaking one of the main characters And actually that work in a way Actually my finding was I expected because it's a recorded audio. There's sometimes a little noise but most of the stuff is clear language and Basically what Google basically sent me as transcription was sometimes little their the quality was way Worse than I expected I expected something very good And it was like it was okay in average and it also like contain many bad words and Actually, it felt like okay. These are like of course corperses on they are trained Thank you. They're trained on Chats and stuff. So actually I had some I cannot say this because it would be CC where it's like Okay, anyway, um, yeah, some really nasty stuff in there as well, which definitely does not belong into a children's children's Hirshby, so okay. Um, that was it worked in a way. Um, but Not enough so next step is actually to solve the data problem get it the data set and Also the next step because I don't have any characters to learn from I only want to point you to the sound the song long thing here it's speaking like Kate Winslet and somebody trained an audiobook from Kate Limpsack and Made something else read as it was Kate Winslet and it's basically I don't want to play it because we're a little bit running out of time It's basically comparable to the quality we've seen before with my with the speech generation So it's also impressive and I'm not a bad. There's still some potential. We still This is not a fully explored space yet So, um, okay So plot fabricate detective story I started my research and found out this goes way back to the 1960s how to generate stories even like there were stories generated for some western TV shows in the US back in the 60s and 70s I said, okay, long touch it I have to do some more research because this is probably more Not a deep learning problem because for examples, there's some rituals in these Ready stories I want to use for example, they need a potential new client to give them the business card and Their the client always replies. Oh, what is the three question marks are about and just says well It's about mystery blah blah blah blah blah and this is something this has always happened And I don't think an AI will actually pick that up So last step the artwork how can I can I break actually's can at least announce my product on Amazon because I need a cover a Title and hey, I can fix the rest later So there's a convolutional neural networks. Um, it's the basically we know this from style transfer We've seen many examples here. This is just like how the networks looks like So basically you learn style from one image and apply to another image And it's basically all about distances to white noise and combining them. I don't want to go too deep into the tech So I created a Set of pictures to transfer the style on Just to see how how how it works and To get an impression. So and this are some lessons learned and also some some general lessons learned on the way There are limits extracting Style from an image to a model So you cannot really extract anything To anywhere. So there's some some stuff that works really well For example this tunnel view picture I've taken in just emity on in a park in California This works very well on most of the styles. I learned But this is sometimes a style like this This is like an artwork from one of the detective stories one of the original and I thought okay This looks like a decent thing to learn from but if I apply to my One of the picture that works. Well, I get a little filtering thing, but not really a style transfer So it's not everything is working everywhere and the thing is nobody will tell you on the internet because that's all cherry-picking so So there's also like limits applying a style to an image because this is for for example an image I expected oh this should work work really well It's a woman dressed up for Halloween. I guess and if you apply many styles This is something I get pictured none of the stars ever worked on. Yeah, so but Being a consultant and can I write a blog post about this? What will my customers think? Okay, I think they're total failure No, we're not we actually just like okay, not everything works This is the same but probably because she's like a dressed like a skeleton and we learn from a picture data set Which basically are as a sample from the world. We see the network learns from the world We see this basically I use the cocoa data set is open data for Microsoft Which is used for many many research papers you see at the nips conferences and stuff So it's really good and explore data set But it was not able to learn that we can learn something here Mind the cherry-picking because of course I like to show you stuff like this and see your faces and your smiles How awesome my ex is so awesome He does all these great pictures and but Alex is sometimes not so awesome because like all the tech is not so awesome because Probably most people skip showing you like the failures or like the average and it's not circle thingies And of course in the internet. It's about it getting attention and you don't get attention Writing about stuff that doesn't work And I think it's a big problem also like I saw I saw many pictures I was so many code released on GitHub and really I'm really thankful for everybody for putting this stuff up But I saw a lot of repetition. I mean like you have like Style transfer in TensorFlow then you have transfer for in pytorch And if you look at the github descriptions, they all feature the same images They don't even add one custom image. Basically, it's reproducing work But I'm not really sure whether there's some improvement involved as well Or if it's just like hey, I have this nice AI Get and probably recruiter will find me and somebody will hire me for a lot of money So, okay, these are like the images you've seen before and now I have a confession to make three of them are only fake Yeah, you didn't see that coming. So Which of them are like because we're a little short on time These are like the three fake images and I showed it to people who have worked on this before They were able to spot most Mainly like two of them. I showed them to fans from I'm not just like two friends Okay, which are fake and they did not really see any of them. So these are just like made up So they're actually like the titles are real, but I exchanged the picture And Yeah, then the code also we find online. It's awesome. I think point I saw You're a dear Python and we preach Python free for quite some time But it's really I was quite shocked you find there's some super recent thing from released by Facebook Or like a big company and it still uses Python 2 7 and it's released just like a few months ago And if you're not aware Python 2 7 will retire in a bit more than a year. So the clock is really ticking It's like no, so don't use Python 2 7 for anything new actually. So it's like it's really bad and another thing I find for example like Yeah, so another thing I found the code quality It's the I call it sometimes the benefits But also like the curse of Jupiter notebooks very often you see side functions with closures What is what is a closure closure as a reference to a variable which is not within the function? So it looks up in the in the in the namespace above Which it's working as a script in Jupiter notebooks But it's really bad if you are new to the code and you want to understand the function because you have to look Okay, where is this variable actually doing what is it storing and then you have a like one page up in the Jupiter notebook? And also you cannot really Use the function somewhere also use it to a sub module So actually there's like the code quality of course its researchers and researchers I was forgiving and I'm trying to teach researchers to code better all the time But of course something you wear the code could be better Also, like another nice thing is using built-ins for variable names like input or in even equal is equals or list or something like that and So there's some improvement code to be done as well Probably if you're your preference and partner sir, so you're able to fix it. So what are my future? Staff so though not everything was like magic like it felt in the beginning I really want to continue this for this project And improve it and also maybe combine it with other non AI Techniques something maybe more determinedistic or like a decision tree for generating stories and stuff or a nice combination of both of it So these are like my future step so Like key takeaway here is don't get created too crazy about all the stuff we see on Twitter on online Because like it's a lot of hype going on there Though we cannot say it's just like a stupid hype because AI is a thing It's it's something you can do something really useful, but I think we have too many Falls impressions and images in our heads because like every newspaper always like oh, yeah It's a I receive robots or you have these movies with like an eye relationship And I think this is all pointing ourselves to the wrong direction It's not always working and another thing is if you remember old phones They were inspired by Star Trek where you basically have like the Star Trek thing you think I'm but actually no we won't look like this So we have a lot of stuff for media and Hollywood in our heads And also like in AI and robots and stuff And we have to really you should really like try to let go and say this is entertainment And this is something else you can do because you also have to open eyes The experiment in opens new opportunities and possibilities For example while just playing around a finally nice use case for okay I can have like bad scans and I can really improve them so Yeah, so I'll I decided to continue this project and continue and give like constant updates if you want to hear about it To really small announcements. Um, so if you happen to be in Frankfurt or Southern Germany, there's meet-ups We have just launched by data Frankfurt. We're going to launch in in September and and there's a pie data suit vest, which is Heidelberg castle in Mannheim Stuttgart It's already launched and going so if you're around in the area You know, we would be very happy if you reach out to us to contribute a talk Or just come by and have a drink or the chat and Peter's always like happy to show you cars rule on Stuttgart. Isn't it? Yeah, it's Peter. Well, so it was organizers And we also have this great conference coming Picon de in the all in cultural at the center of for medium arts and of October So cultural is down there But it's very well connected to airports and fast speed trains and it's one conference Which makes keynote speakers really happy because they are really treat so first keynote speaker We announced is West McKinney the creator of pandas and West replied yesterday on Peter's lightning talk said if I miss it I'm really happy So thank you very much and if you have any questions, I'm around the conference. Just come talk to me. Thank you very much