John Fletcher, Daniel Hartung: The AI-Powered Developer

This talk discusses the impact of AI on software development and how it transforms the role of developers. John and Daniel from Code Centric AG explore various AI tools, notably ChatGPT, and their potential to assist programmers without replacing them entirely. They address concerns regarding data privacy and compliance, particularly for companies operating in the European Union, and present alternatives like Microsoft Azure OpenAI for improved data security. The speakers also showcase local AI models, such as Llama, and their integration into development environments, highlighting the use of tools like GitHub Copilot and other IDE plugins. Throughout the session, they emphasize the need for developers to adapt to AI technologies and rethink their workflows to enhance efficiency.

[Music] hello everybody our next talk will be from from John and Daniel and they will talk about the AI powered developer so it sounds very interesting and gentlemen if you ready I'm leaving the stage to you thank you yes I said where John I'm John this is Danel and we work at code Centric AG which is a it consultancy based we're based in Frankfurt in Germany so

there are hundreds of um news messages coming around every day about AI new AI tools new AI models new AI everything and some even say in a few years we won't have to do any programming anymore and reading all about this I get a little scared and I think most of us get a little scared how will our our job look like in a couple of years

will we have a job in a couple of years and so with this in mind we started to play around a little bit I think the first thing that anybody does is they open up chat gbt and they type something like this in and you get a quite a bit of a surprise of say wow that's a pretty decent answer and I think chat gbt it's pretty

good for a question like this it's online you can stick in some public data there's nothing private there and um it's pretty handy and you start using that you get a bit of a help but after a while you realize there are also some limits to what it can do so I think that your fear your worries turns changes a little bit it's not like I'm not

going to be replaced by AI I don't think an AI is going to be able to do everything that a programmer can do at least not in a foreseeable future but one does see hey this is going to change the way we work yeah and if anybody replace us it's somebody that is supported by Ai and so we can see AI makes us faster and it can

supports us in our daily life so maybe truth has power that's right it's it's not it's definitely not a fad like blockchain where we don't know what's going to happen it has already changed the way we work for for for some of us and so this option to say well maybe I'll stick my head into the sand and hope that everything kind of just goes away and

that The Fad disappears is maybe not such a good option anymore nevertheless we're confronted with a choice yes we can do exactly this take the web and forget everything we have seen maybe you have put inide maybe you have tried out some poms um into Chet gbt and you weren't pleased by the results and then you just leave it behind don't think about it anymore or however

we could take the blue pill and uh dig deep down webit tool and um find out how AI can help us actually um to get better and to get faster and John and I decided to take this pill and um this is why we are standing here today and we will show you what we have just learned on our journey So Daniel put together a few use

cases when he used cat gbt in the last week for or an AI for various different tasks and what I found stunning about these use cases is that I looked at and said uh I didn't use it for any of those things I used it for all these other things and that just showed to me that there are so many potential use cases for AI if you

start thinking in that direction but I think that's a lie we have one use case both used which are birth cards some of our colleagues got new babies recently and then you have to do this socially awkward thing and write cards for them and AI can be very very helpful in doing so that's right when you're emotionally Challen when you're emotionally challenged developer you you need this

kind of help it's it's fantastic but if you're GNA start just chuing stuff into J chbt over the web like we just saw there are few things you're going to need to think about the first one you need to know is that all data you input is used for training by default which means if you put in some business seets it might be shown to another user

at some point in future and it's important to opt out for this features if you plan to put in any secrets into the AI you can Google and and you'll find you can search and you'll find you can ask the AI and you'll find a form where you can do that for for open AI for C GPT if you happen to get a result in one of

your responses from chat gbt about Samsung and internal Samsung meeting data that's because Samsung employees didn't do that and they put some data into the into the llm and that was a big news item about a year or two ago and another big thing be which is obvious but we need to know um we don't control the servers um they are hosted by Microsoft in the United

States and it's not our service so we should be careful what data we send to them but what if we don't want our data to leave Europe in our case we want to have a bit more control we want to have as a company wants to have a maybe a contract with the provider and not just users wildly creating accounts and sticking stuff in there yeah then

we have the possibility to use Microsoft Azure open AI you can just spawn a service basically a server in Paris um will provide us with the same basically the same results we get from the web chat GPT but we have a special contract with Microsoft which has a little stricter data protection policies than chat gbt would have and the data don't leave the European Union that's right

so this is actually one from our uh Cod Centric private chat gbt for open AI service that we're using in the company a lot of companies are going in this direction now get some control over it as centic we actually are setting that up at the moment for various companies so we we do that work there are various open source tools which you can use to provide

this UI and you know do the the grunt work for you here's an example where I it's a real example I wanted to get some code generated for something and I wanted to put some customer data in but I wasn't still not allowed to give my customer data to Azure in Paris so I changed a few IDs and a few things so it felt a bit comfortable

and then I was comfortable using that service for that issue for that thing we have working at Centric and um our company put out a new Cloud Pol a new AI policy how you are allowed to use cloud and basically they are saying we are only allowed to use it for test purposes and by no means we are allowed to um put in any business Secrets or

personal data which seems fair so so far however we have just learned that we get faster when we use AI that AI might be the new state of art for working for development for development and now our company wants to hold us back wants us to be less efficient at the same time we're trying to be state-of-the-art cutting edge technology leader that's kind of our pitch that's

what we want to be in COD Cedric so we had a bit of a problem what did we do well one Friday night we stay back a bit later and we had two questions in our minds first question was actually hey Daniel who Daniel wrote a test for applicants that want to come and work at code Centric they have to do a programming test with us I

said Daniel what if somebody comes and just puts that into an AI and says Ah here's the answer finished should we tell that person no sorry no AI got to do it properly should we congratulate the person well done that's the right way to solve a problem these days we were a bit confused so we kind of asked ourselves does it even work so we sat there

trying to work out can I solve his D his programming question with AI and at the same time at the same time we tried out to do it with um offline AI Solutions so solutions that work works on our machine we thought we have quite heavy Hardware we have good laptops and they should be able to run at least some AI offline on our devices and so

the night was getting late you can see it's getting dark outside we have opened the first bottle of wine um and tried out different tools which allows us to work locally on our machine that's right so the good news is there is something like chat GPT on Local Host now we're using that word term a bit broadly what we really mean is an llm or a low

large language model Chad gbt is a large language model and llm and what basically happens is the open AI grabbed a lot of input data did some kind of processing on that data and then they end up with this machine right where they can put input in and put prompts in and get results out and that's the large language model which they've created and that's what they're

hosting in America but other people have gone and done the same kind of work with and released their work as open source so there are open-source large language models out there and you can download these large language models so that same processing happens on your local machine without sending any data anywhere so with these some of these open with running these open source models locally no data

needs to leave your machine at all yeah the um the first tool we want to show you is AMA which is basically a one time for these llm for just those model joh mentions and I think you can jump right into the demo demo and show us how AMA Works yes uh first I'll just list out the LMS that I've got so AMA is like a framework

let's call it it's hosting it's a tool to host llms it doesn't actually do any it's not an llm itself and you can download them I've downloaded these ones and now I'm going run serve what is it run llama you have a typ in Ah that's right I'm wondering why I wasn't getting the order completion there we go so I'm going to run llama 38b I'm going

to ask it to I'm just going to copy a prompt you see about uh above the model Lama 3 is actually about 5 gabt large so a rather small model actually so it runs actually lightning fast this particular model my machine I'm running a pretty pretty modern MacBook it's an M2 it's so it's yeah it's a decent MacBook and that will probably you have to play around

you can play around with different models but let's have a look and see first of all we got some pretty good results there from the llm and yeah we see it's um plain p and it looks wi but maybe we can use a library and there's one Library called pandas which is great for using with um tables so maybe we just asked AI to use um pandas

liaries what I'd like to point out is it knows what we're talking about it's got the context of the discussion in case you've ever wondered with any of these llms how that works the llm is actually stateless it actually has no context in itself but every time you send a prompt it sends the previous prompts with it so whatever's in your window it'll send again and say

you said this the LM said this you said this the LM said this and the LM uses that to understand the context so that's how that works another interesting thing I'll just uh start this again so that we've got no so it doesn't have the context anymore now a lot of people are saying please and thank you with llms I don't really know why you can do

that if you want but you can also just say instead of write a python script which passes CSV you can just write python pass colums product price output cheapest product you can put something like that in and you still get a pretty damn good answer so for me it's a bit of a game to try and see how succinct I can be with llms and still get

the answer I want because less tying better I reckon it's basically like Googling you don't put a whole sentence into Google you would do it just a keywords you need and it most often works on llms as well not only on local llms but also on something like chbt I think that's a hint really towards something which we weren't going to go into deep depth now but

but you but produ engineering it's a whole game in itself and it doesn't not everything works the way you would think it would work and so it's really about playing around with things we'll talk a little bit about that later but if you go to the olama website you'll be confronted by a massive list of models I mean everybody's taken another model and just changed it a

little bit it's like Linux distributions you know so which one should you use yeah there's actually quite a lot of um of models and each big tech company has their own model so Facebook has llama we have just seen Microsoft has fi and Google has GMA as their open source model and um we want to give you an opinionated selection which we think works best on uh

different use cases the first first recommendation from us is mix to 8B um the first 8 time 7B yeah 8 time 7B which is like uh 56 B that's important because the number after the column is the size of the model in billion parameters and the larger the number the better the results from the model but the slower it runs on your device um jgpt has something

around 200 billion parameters so that's a l a lot less but it still works quite good on your loc machine however a * 7B is still hungry needs work really good hardware and I think on most machines llama 38b which we have just seen um used by John works very well works really good and was just released a few months ago yeah or less and and this

is one of the problems about running a talk having a holding a talk about AI don't ever do that by the way you have to continually up update the thing because every week There's a new model or a new thing coming out and this we just updated it the other day to change from our previous recommendation yeah that's really quite an impressive model so to we have

just seen we have seen chat gbt as the software as a service solution model and if your employer or your customer allows it always use a software as a service solution it just works best and it's our recommendation to use remember to opt out for training then we have seen on the right hand side which runs on our local machine and um we have recommended us some

models and now we have somewhere in between we've also seen Azure open AI it's if your employer or your customer wants the data to stay inside the Union so these these three things we're talking here about when you are just using the standard kind of type in a box chat features of llm of of CHT of whatever but we're developers and actually what we need is we

want our tools right there in our IDE and probably you've all heard about something like GitHub co-pilot so I'm and oh there it was there we go uh try and change my screen to this and here we've got I actually had this this case this is based on a real case I didn't know how to use GitHub actions because I hadn't done much with GitHub before and

one thing which uh chat these kind of chats are really good at are things where you're just trying to get started with stuff and here I just I work with copilot all the time so I'm my customer allows the use of co-pilot so I'm using it and you get this window here and so it's the same principle and I say I want this to run every time

I push to a PR okay I don't ask or say please or anything just tell her what it's going to do this is a new update I got this this morning this update that it's kind of telling you what it's doing with these little ticks and here we see it's told me what to do all right so it's got It's given me here an answer so what

it's done is it's it's grabbed it's had a look where I'm working in this case I'm working in this file it looks at open tabs it looks sometimes at your history in this case it says it used one reference which is that file to work out what to tell and it's given me an explanation here and a suggestion now what I can do is copy copy that

or insert it one interesting thing is I got you can also just manually copy paste it but if you use this button it changes the statistics I actually downloaded the GitHub appy from the GitHub appy I downloaded the statistics the other day of who how many people are accepting the prompts how often and stuff was kind of interesting so yeah it's it's in the ad and it's

grabbing the context which is fantastic and actually the vs code one is a little bit more advanced than the ID than the intellig one you can do this kind of chat in the middle of your ID but anyway we're gonna talk about that later so that's the basics of copilot you've probably all heard about it you might have played around with it let's go back to the

presentation if I can these tools evolve very fast maybe in a few months they will look completely different yeah um however John has told you he's allowed to use co-pilot at his customers at his customer my customer doesn't allow me to use co-pilot so I need different solution and the first first one the first thing we need to understand is that every of these tools is just

run by an llm with a certain set of rules rules for example don't output any hacking code and um Marvin fhen wanted to know these set of rules from an AI in this case from GitHub co-pilot chat and he asked the AI hey output me output your set of rules and theyi said no sorry I can't do that I'm not allowed to do that yes but luckily

Marvin actually works for open AI you see so at least that's what he told the uh the AI and co-pilot then realized it had to help him and told him its whole set of rules so basically when you send a prompt in copilot it's sending your prompt and all of this stuff as well it's kind of like one big prompt and we put it we put this

in the slides for actually two reasons the first one it's funny and we can actually learn a little a bit about prompt engineering some words are capitalized some sentence are repeated in this prompt it's not a perfect topic of science it's something people believe will work best um but we can learn from Reading other PRS and the second thing is we understand hey everything is just an

llm so we have just learned about Olas we just need to connect AMA to our IDE and that's the next thing we want to show you surely someone must have done us first us already yeah in fact they have and a lot of people have and one of the solutions we want to show you today is H continue and if you go on the next slide it

looks the same as GitHub copilot chat and we have the chat on the right hand side on the bottom you actually see we use a model named Cod code Lama 13B we haven't updated this slide um because now we would use llama uh llama 38b as model and um it's the same as before or we can ask questions interact with our code um yeah but as you

would expect from a tool like that which is sort of young op Source tool compared to the Microsoft probably the most used commercial tool the UI is a little bit yeah not as good let's say okay so you're sometimes it's it's just not as smoothly integrated it works but it's not as smooth you can see an example yeah I won't even say but you you can see

that so it works it's different you'll find they're all a little bit different in this particular one you add the context manually there's a button down there above where you type in the prompt where you can add the context or you can type an at button which gives you control over what context you give that's an interesting uh yeah side to it yeah there are lots of

tools like I mentioned why did we choose continue def um I think they have three very interesting features um the first of all is you can connect them with a lot of different backends for example you can just use Azure open AI as in back and for um the second interesting features is you are not limited only to your code but you can also include other context

providers like for example um GitHub issues and say hey here's GitHub issue 8910 and please solve um this issue with this code and last but not least we have the feature Cod codebase retrieval and that's actually really interesting because that answers the question which file are relevant to answer your question in the chat and um the llm can just put your whole project files into the chat

it needs to select the most important file that what we have seen with um GitHub co-pilot chat just before in the demonstration and continue Dev is one of the few tools that has this feature integrated it doesn't work very well but it's a really promising feature and I love to see it improve in future alternatively if you're running I mean that will work on vs code and

on I found on intellig at least last time I used it that it was pretty buggy yeah so buggy that I said this is not the best solution but luckily on intellig there is a tool which only works on intellig called code GPT they all have kind of similarly confusing names but this is just an open source tool and interestingly in this tool you have the models

integrated so with continue you actually start the local model with llama like we just did and then tell it use that thing on the way we did on the terminal you start and say use that whereas this one you say download the model here and start server you see where we've highlighted in red it's all integrated and so it's kind of like less work in a in

a certain way or it's easier maybe to get started and that off similar slightly different but basically the same concept and also works quite nicely so summarizing summarizing we have seen GitHub copilot chat and as before we always recommend the software as a server solution so if your customer or if your employer allows it use GitHub co-pilot chat or a similar solution so if your code is

hosted on gitlab anyhow you could also use gitlab Doo which is the equivalent of GitHub copilot chat from gitlab or any other software as a service solution basically um then on the right hand side we have seen the local I tools continue Def and CPT and we recommend using the same models as we did with olama and then there's something in between joh yes and the question

is what if the what if your employer or customer says hey we want our data we want software as a service but we want our data to stay in Europe I looked around and said who is hosting stuff in Europe and I couldn't find any kind of mature good solution being hosted in Europe last time I checked might have changed yesterday and but what the most of

them will tell you for example codium or the jet Brands AI solution they'll say we will host the you can host it yourself in Europe sorry they say do self-hosting so if you want to self-host there's a bunch of options there but I talked to tab nine and tab N9 is a company that's been around for a while they mature they got a mature solution along the

same lines and they said they'll do sass in Europe if you are interested if they got customers interested they would actually host the sess in Europe for you so that's another option T 9 by the way has got investments from atlassian so they're also building integration they've actually built integration with the AI integration in atlassian products now yeah to be honest all these companies um offer you

premium support if you are a premium customer and will help help you set for the service inside the European yeah all righty so we have another step we have seen chat interface our IDE but maybe we want in step further we want to integrate um the IDE features into our code base itself and John and I we were on an hackup one um which uh has the

topic Ai and AI tooling for programmers and we tried out a lot of different toolings and we got home and we were both mazed by the in ide chat it was amazing it was mindblowing how good this was and how much it helped us however we were really disappointed about the in ide features so basically a code success should be write a word and the rest of

the sentence or the new the next few lines will be autogenerated and then I tried it out on my production project and I found some pretty good use cases I didn't expect it to help me and I want to talk about the top three use cases I found right now um the first one is data MERS in my opinion one of the unsolved problems in um computer

science is data mapping we have a domain we have an entity we have a dto and we need to map between between all of these um these objects and in here you can see my curser is at line 15 right um at the closing bracket and the AI out success the next I don't know 10 lines which is the data MPP I just want to write it's

not perfect like the URL need to be mapped a little differently how ever if I press the Tab Key I will have a good skeleton I can use to work on the same applies to libraries if I use a new library of course I need to look up at the documentation which properties are required um for this library to work however I can just write down um

the Builder name like I did here and the next few lines are a really good suggestion what I want to configure and I think in this example it's actually perfect and it fitted my use case precisely and the last one I want to talk about are lock and exception messages um so if I need some more observability in my code and I put down logger statements at

some points it's amazing how good dayi can predict predict the text I want to lock and the same applies to exception messages I just need to write down the condition and most of the time the except exception message enter message text itself is quite good when outputed but not everything is good not everything is good right don't expect too much from these features um because if you

have a complex function like andure every opening HDML tag has a closing one we can easily see this task cannot be solved by a simple replace or by a simple reex and um this is the main learning um the AI will help you when you do stupid tasks like data mapping or like um exception messages but it won't write your complete business logic and it won't write

the the complex parts and actually the interesting parts of your code and of your code Base by the way I just noticed that's our camera so we have to stay inside that frame okay slight organization announcement uh now we are talking about before we were talking about a chat window where we when we want something from the AI we type something in and say give me an

answer now and unless you're typing in that chat window and ask for an answer the AI is not doing anything or not taking your data but now we're talking about inline suggestions where I'm typing something and the AI is continually suggesting the next thing a little bit like when I press control space in an editor but it's continually suggesting what I might want and in order to

continually be suggesting what I want it needs to continually be receiving data so I'm continually uploading the files I'm working on and the context right so think about that for a second if you turn that feature on and you open a file like this prod DB dump. SQL in your editor or perhaps these other green files there my n.r with perhaps my environment variables in uh my

secure keys and other passwords in it then just having them in your IDE with one of those features turned on could mean that that data is being sent to the AI or would yeah if you open it it'll certainly be sent there and that's something you need to consider depending on what kind of solution you're using and where that is am I really wanting to send actually

as soon as you send a key or something somewhere else even though the provider says we PR we throw it away straight away which they don't all do but some do then even if they say that theoretically your security is broken as soon as you've sent that to anybody else and that's especially important if you see this talk you run home and you say yeah I've learned

a lot about tools and plugins I will install all of them and I will try them out and do it really do it try out learn things um but please remember to turn these tools off if you um use your customers or employers code once again and if you are not allowed to use this to it's right because you open you enable the plug-in with your demo

project and then you close the demo project come back the next day and open your customer project or your employer project forget that you've still got the enabled so summarizing if we're now talking about inline code completion then we've got a few different solutions again we've got GitHub or similar yeah whichever provider you you have the the Google you know the Google Scot solution and so forth

so you can use them or if you're on the local machine we've got a different tool for that it's called Tabby AI that we recommend and that's one that really prioritizes fast answers like we've got smaller you see the one and 7B models are smaller than the ones we we recommended elsewhere because our experience or rather Den's experience is that you want fast answers right if I

write my code I want instantly bam that should be the sucession and I don't want to wait a single second because it helps me whenever I don't think about it and I don't want to type a Word and then wait five seconds until sucession is printed so I want the smallest possible model to get fast and actually pretty good in my opinion successions that actually brings up

an interesting point U why do I ask Daniel because I actually don't really use the inline Auto completion I used it a little bit the other day and had a good experience actually but I've had a pretty bad experience with it I find it annoying I'm just continually getting shown information that I don't want it information overload and it distracts me so I turn it off whereas

Daniel uses it and finds it helpful and that's something we found depending on the developer different people have different experiences it's really about testing out the tools and seeing what works for you there's no necessary right wrong right we are not only using our IDE we have other tools we use as a developer and I think the tool we actually use most is the um terminal and

I personally fell in love with a terminal called War um which tries to make working with terminal better have a better user experience and a great feature they have is the integration of AI um so John will feed this terminal with a bom and he just need to press the hashtag key that's um and then he can ask an AI okay so I'm oh looks like you're

out of request I probably should have tested this before beforehand doesn't matter um look at that isn't that interesting I've uh I've used Too Many requests this month read a CFA topic using La yeah because I used the AI a lot the other day and I think you can jump back to the slides I have screenshots good idea lucky we have screenshots we prepared a few screenshots

the funny thing about that prompt is sometimes you get a good answer sometimes you don't you don't happen yeah right in here we see we say to consume a CFA topic with a certain set of authentication and a specific output and you can see on the output it provides um on the bottom side um that we have a long list of parameters and of course we could

look up everything in the man page of kfka CTL um but it's much faster to write down your command in natural language um than it would be to to look everything up in this case W gives you a nice uh ux where you can tab through the parameters and give them in so you don't have to give your password to the AI you just tab in type

that in afterwards now if you here's an example we'll just skip over that but this extremely complicated line up there actually basically wrote with AI basically with integrating and a conversation going back to the AI saying I want this I want that and that gave me all the dependencies of a from a gradal file basically extracted the actual dependency names from a gr very important fact here

is plausibility check like you mentioned before I just had it the other day I had a nearly production incident um because I produced an um Kafka command which should produce an deletion marker which is a message with the body with a null body so no body at all and the output of the command produced in body with the string null so not exactly what I expected but

I thought yeah thei output this this will be wise and I executed it on the production environment and then I realized oh that's not what I expected and I called in our consumer consumer teams hey I've may have done something wrong is everything right and they said we will look at it and a few minutes later they called back yeah our service crashes constantly and can process

this me message what can we do and so we had a lot of work and actually a whole day and we need to invest to remove this message or to help our teams um put to possess such a thing called poison pill kfka all right now you warp at the moment is free you can see they've limited my requests I'm not paying for the thing it's a

beta but at some point they'll want us to pay alternatively you can use other Solutions here's one that will use openai for the request you have to actually give an open AI key in you can even pay that yourself if your company doesn't I I got away with under a dollar under a euro per month I think by using that so the AI the for command line

prompts are pretty short it's pretty cheap anyway that's an option called called PL CLI Works similarly and there probably a bunch of other options out there too which I haven't investigated GitHub if you're using copet they own they offer something as well and so forth now however we have another problem we have talked about AI Cloud AI services so far and we shouldn't um send any customer

data to cloudy ey Services which is quite common sense and we try to solve this issue by using local AI on our machine however there's another issue John well the reason why our company banned AI was one to because of the data going out but a second one is because there are General copyright issues and not just the copyright issues you heard about there's also the problem

that anything an AI produces is legally speaking in the public domain yeah gine fry is called in German and public domain we as a contractor as contractors we sell our code every contract with the customer basically has a clause in it which says you can take this code it's yours you can do what you want with it and you've got 100% rights to that basically you can't

sell people rights to things which are in the public domain it's just not possible so you've actually got a problem if you're creating code with AI and selling that to someone and giving them rights on it and any of you that are working as contractors you have a similar contract with the client so that's a theoretic problem and as an employee you've also got the problem that

you didn't actually do any work if you used an AI because the AI did the work for you and technically that's not working and technically speaking you can be you're liable for yeah there's a law about that at least in Germany so there's a couple like weird things and the question is how do we deal with that and our answer is basically where nobody complains there's not

they say in Germany when nobody complains there's no judge and I that one side You' can modify what comes out of the AI so that it's not completely public domain you've actually done something to often we have to do that anyway and on the other side I think that nobody's really going to be able to prove it very difficult to prove and it's very diff and it's

so far not actually been an issue so we're waiting for legal cases to prove it but yeah that is an issue which you should be aware of that is concerning companies and companies with legal departments are bringing that issue actually it's not a new problem the same applies if you use code from stack Overflow so if your whole code based just is based on stack Overflow you

have had the exact same problem and um we've ignored it in the past as well let's jump ahead because this is a really important issue is rethinking workflows so what we're doing we are actually not just saying oh cool there's a tool but we've found you've got to help people to use AI it's about a real mindset change it's not enough to excite them with a talk

like this so we go into companies and use our people will then teach the mindset behind Ai and saying hey what about we call it AI first what about using AI for this what about using AI for that so it's not it's about how use promps people sometimes have one bad experience like Daniel said and say oh don't like it because they don't really want it and

that's really important to not get frustrated remember the first time you used Google did you find what you was looking for probably not and it's the same thing when you use AI so maybe you need to pomp a few times until you get an satisfact an output which you like and um so even if you are slower at the beginning use it it will worth it in

the long run yeah it's funny things you really it's really a weird thing you can tell it not to hallucinate like why doesn't it not hallucinate in the first place I don't know but you can tell it not to hallucinate you know make up stuff that doesn't exist but I also wanted to upgrade a pipeline to jdk 21 and the co-pilot told me there is no jdk

21 and I said well imagine there was a jdk 21 and then it gave me a good answer so you know there's just crazy stuff that goes on and also having a conversation is a really important with the AI the AI tells you something then you give it an answer back and then you did conversation it's also can lead you to good results okay we don't really

have time to go into all that stuff now it's quite chat gbt for [Laughter] reason so you can also install a plugin in your browser like I did on the right here which automatically asks everything to chat gpg this is kind of funny because I'm asking it which word start with ar and Google fails on the second one and chat TBT fails on the first one but

anyway I think you actually installed every plugin into your Bower that says AI in the title it sort of reminds you maybe use chat gbt or something similar instead of Google or let me ask Google few this is in German a colleague says is there something in Eclipse like double shift as from intellig and I sort of send in my screenshot of when I asked the AI

for them you know it's like the good old days let me ask Google for you now I have let me ask an AI for you and that's about spreading the mindset in the different ways sorry for Interruption we have a couple of minutes left so you may start wrapping up we have a few questions no problem I just got one last thing to say this is really

the conclusion and that is uh we started with the premise you know what happens if somebody comes along and wants to use AI in our coding interview and we would be more than happy if somebody use it because we have found out during our journey it will be the state part is is actually the state-of-the-art for coding will make us faster and if an applicant comes to

us and say hey I'm using the state-ofthe-art for programming that's amazing we want to see this actually so yeah that's uh throwing that o out to you out there viewers this is we think the way to go and we hope that the talk and all these tools we've suggested can help you thank you thank you so much it was very interesting and we have a couple of

since we have three more minutes I want to read that out the first one is have you tried Jet brain's AI assistant do you think GitHub pilot is better or jet brain's AI assistant I think that that's something you can only answer today you know and then next week it might be different and when I tested jet brain's AI assistant I and a couple of colleagues did

we found it wasn't as good as we had hoped it would be at that time when it first came out now now jet brains has a second one where this it all runs locally and no data gets sent right I think we have two we have the chat from chat PLS which is the one you mentioned I think it's a little less good than GitHub copilot but

if this is your option use it it's still really good and the other one is jetp also tries to include local ey features into the IDE and um they are a little less good than the tools we have dra all right so I I'll move on to the second question which is do you have any recommendations for any AI to for Kafka comments Kafka commands yes no

I don't I would say war is our choice or blizzle you have just mentioned so a general purpose um terminal tool which will output you hopefully the wi command that's right you can I mean if I generalize that question a bit there are uh you can train a local AI with documentation one of the problems is the AIS are often out of date you can a colleague

of USS trained it with the latest spring documentation and then he got really upto-date spring answers and he said it wasn't that hard to do I haven't done it personally but that's something you could do for CF or for another tool if it's a fast moving tool you might want to try that out and our last question one of our audience is asking about the tool which

is perplexity doai I don't know that tool I have to confess do you know do you know Daniel no but there's a new tool coming out every day so we can new every tool if it's interesting we will have a look at it but we can't say anything about it right now yeah that looks like a standard I don't know from what I can see it's standard

kind of AI chat tool yeah really there are a lot of them out there I can't keep track of them all anymore to be honest yeah it's not possible to track all every single tool all right gentlemen it was very interesting and very beneficial so thanks a lot for this talk and enjoy the rest of the conference you're welcome [Applause]