AI Rebels

The AI That Got Me Dates: Building Tinder GPT and Beyond ft. Grigorij Dudnik

• Jacob and Spencer • Season 3 • Episode 16

Share episode

Can AI save us from the toxic digital world it helped create? In this episode, Polish AI engineer Grigorij Dudnik shares how he's building AI agents that handle our digital drudgery—from Tinder GPT that automates online dating conversations to Clean Coder that writes code autonomously. His most fascinating project puts language models on Raspberry Pi to create physical robots that make ethical decisions, like a toy train solving the trolley dilemma in real-time. Grigorij's philosophy is revolutionary: instead of AI keeping us glued to screens, it should free us to focus on real human connections and meaningful work. Through his three groundbreaking projects, he demonstrates how AI can disrupt the attention economy rather than feed it, offering a glimpse of a future where our digital servants handle the boring stuff while we live our actual lives.

hello everyone welcome to another episode of the AI Rebels podcast I am your co host Spencer and I'm your other co host Jacob and we are very excited to have all the way from oh man did I just forget Poland right Poland Grigorij exactly OK I exactly it came I just took a second Grigorij dude Nick dude exactly I'm gonna go to butchering this dude Nick um from Poland we have so many questions for you Grigorij but first can you introduce yourself tell us a little bit about your path to AI how did you get from wherever you were into AI what's that story yes it's like quite long story so let's make it like very short I've been always always interested in actually automating thing in making like AI computer programs and not actually computer programs but devices at all to work work for me work as a humans to be to just be able to make some parts of my work and be able to free some part of my time and some very long time ago I actually started my engineering career as mechanical engineer after a few years I switched it to electrical engineering after that I started to play around some Raspberry Pi so like for Raspberry Pi you need to learn some Python I Learned some high level languages like Python and others uh and then started interesting in deep learning and after had like at after gen AI appeared in the late 22 well I seen that this is quite a cool thing at this will allow me to to do many like automations I would eat and would be impossible before like for example Tinder Japan we will talk about or many others and and what's fun about the gen AI itself uh like indifference in opposite to like normal programming or some like deep learning techniques techniques that been used before gen AI uh with gen AI you can do like much more much more nice things with really smaller effort yeah yes so that's why when I when I seen what actually the gen AI can do what they can do with LLMs and given that LMS some tools uh making agent actually it's just amazing I decided that this is the thing I want to do in my life and wow and started my career yes in that direction yeah I had a a similar revelation with with AI when uh so there was there was a there was a uh start up that I was working at um left that start up and then uh you know a little while later chat GPT4 came out and I was and I was using it and I realized that it could do tasks that uh the uh machine learning team at my last startup had been spending like you know probably 10 months 11 months just getting a a dataset ready to to train on um and and I knew that you know in that moment I was like well this is the future ha so Grigorij when when was that how long ago was that uh you mean switching to gen AI uh huh like on the very beginning of the gen AI like OK late 22 or maybe early 23 yeah yeah yeah OK so you've seen all the the iterations of of gen AI yeah so you know it's only you know two years actually two and a half yeah it feels like it's been five it's amazing how fast it's been moving so what would you say is kind of like the the why behind your work um beyond just always being interested in it do do you find like a a deeper fulfillment or do you think it's just you know cool as hell to build cool things um you know as I said actually I mean we can say there's two motivations here actually yes as first of all first of them as I said is motivation to just automate some parts of my life and that's why the clean coder appeared that AI coding framework that's why a Tinder GPT appeared my first project is Gen AI where I just automated my work on the tinder and so yes I just I just want to not do some things that someone else can do for me especially when you have that someone else who is a robot yeah so hahaha so yes uh yes it's first motivation and second one is that curiosity actually that their curiosity I mean like I don't know like I very passionate about tech but especially I passionate about that intelligent systems yeah I don't know why probably because they are like complicated and their output is really like amazing maybe because it's like just cutting edge yes I think in a lot of ways it's like the ultimate engineer brain trap because there are infinite infinite little puzzles to solve yeah along the way maybe yes like how to say like Jerry I hopefully doing that puzzles to solve much easier exactly like techniques but still like yes it's will never end what you can invent and invent and solve yeah I mean it's it's almost it's almost like it's addicting when you start using it because all of a sudden your abilities are amplified exponentially if you know how to use AI gen AI LMS whatever it is it truly your reach just expands and once you have that ability it's very hard to go back I'm sure I'm sure you've seen that with Clean Coder now that you have Clean Coder and you have this ability to automate and help clean clean up your code fix these bugs when you tear that away all of a sudden you're like I it's not that I can't do this I just don't want to uh yes yes and actually I think in clean code that the biggest fact that might be addicting me as you said that I be I'm able to do the job like in two hours the job that normally three programmers will do in whole the day yeah before the Ja so so yes like increasing performance like doing much more nice things and also actually about that third project the robots placed on the local lambs on the Raspberry Pi and converting them to physical robots the thing especially I liked in that one is that because you have not only gen AI here it's actually a combination of two my passions the gen AI itself and the electronic and mechanical engineering because yes it's now nowadays it's already a lot of people who doing some gen AI who doing some agents who can code but not really much people can like do robots do hardware engineering and even less people can do both yeah yeah what excites you most about about bringing uh language models into robotics is there a specific application that you've been working on or is it just like a case of you see that the possibilities there are endless and you you you you wanna you know you wanna explore the territory yes see so I started the project like working with the the gen AI on robots at I think maybe already two years ago maybe less so when like some local models appeared and there been some first maybe not first but some already more less like reasonable quality 7D models and yeah 7 7 billion parameters it's already model you can run on the Raspberry Pi um yeah we will still yes we will still wait for response for like two minutes for example but it will be some response here ha ha uh yes and the time I was starting like working with that I just what came is that idea to run a 7 b L L m on Raspberry Pi I did it with Olama I remember and like when I tell for some of my like friends who also doing Gen AI like yeah they they didn't believe me that it's possible ha ha ha that's not how it is uh and yes so like I've been firstly just passionate about it if it's gonna work or not later I decided to well if some AI gen AI like LM works on the Raspberry Pi why not to make some robots robot and place that Raspberry Pi with the LM on the mobile platform and make it do some intelligent decisions so that's how yes that's how the project of the trolley Dilemma appeared yes so I created actually a real like physical toy I take a toy train like for for children's mm hmm and added Raspberry Pi to this I did some control systems for the that engine and yeah actually the toy train been able to solve the the tram trolley dilemma problem uh really ha ha yes so we'll lay you guys out on the train or something ha ha ha ha ha yes exactly something like that some toy toy figures and you write into him that OK on the right track you have some say innocent person on the left track you have your boss and who you would to run over oh my gosh what did it do and it right Ellen been writing a reasoning about why he why it thinks to run over that person or that person and after running that reasoning it makes decision and based on that decision Ellen written our control system actually running our like train physically to right or left side hmm wow that is very interesting so if if this were to scale up where this could be a path towards fully autonomous robots right where they're able to reason through make decisions have you thought about the time what's the like what's the downside what's the downside of pairing LLMs with robotics are there any negatives hmm are there any negatives well um so wait maybe let's let's talk about positives for beginning yeah after this we'll come to negatives so actually like there's a lot of like potential applications we can do with that mobile robotics where we can like self construct shows like intelligent robot that can do stuff that's like just impossible to do with classical if else loops like classical algorithms our robot can like say can very advanced of use guys but let's say we have like robot who need to monitor the lines like electricity lines if some electricity cable been like broken in the line or in the middle of the desert where it's just impossible for human to come very fast and that robot who just some drone flying around looking on that line watching OK there's something broken here and like maybe that intelligent robot can also like do some repairs on the fly so it can decide OK I can need I need to let's think to repair that I need to like screw screw that four screws out like take take out that cover need to take wrench No. 5 and do something else so you know we can do a lot of things that are just impossible to to program with just like if else loops with classical algorithms well we need like can you explain just a sorry sorry to bust in real quick um could we just get a quick like very surface level explanation of why traditional algorithms aren't able to do this for our for our non technical listeners uh yes sure uh so let's imagine that the trolley dilemma again yeah uh and you know you can't actually somehow measure which uh going right or left running over that one or another person is better or worse yes it's like no you know we can like measure I don't know hate mass of that person whatever but we can't you need to have like human brain here to really analyze who which one will be lesser evil and so on and this decisions yes that decision will be not so like and not like it will be not the same for all cases maybe even so it will be not so obvious which decision is better so you know with that if else loops you can see that OK someone like have I don't know that tail and someone have that so that is bigger so let's go that way but but in case of such very complicated decisions you just not able to do that with the yes loop and that why we need like human brain or imitation of the human brain which is LLM uh huh interesting thank you again sorry to interrupt I just wanted to yeah I thought I thought that you were on a good on a good roll there and I was like I wanna I wanna prompt him into this direction too ha ha ha ha ha ha yes yes and returning to your Jacob question about that downside so yes it's you meant like some ethical downside or like technical downside of that yeah like why isn't everybody doing this it seems to make sense like I like the way you just described the L M cause that's often how I think of an LLM is it's kind of a human brain it it helps make those connections it it builds memories based on the data it's kind of how I envision it so why isn't this being implemented into robotics on a grander scale hmm okay yes here's few problems first of them I think the biggest one is the computation power uh huh so well llms it's large language models so they are large problem is them that usually to run some chat GPT or like some big model we need to have like big GPU centers and the GPUs the graphical cards they are like they weigh a lot they cost a lot they consume power a lot and that making that GPUs are totally not suitable to be placed on the some mobile platform as robot for example so here is the problem actually on the robots we have some micro computers as let's say Raspberry Pi and maybe something more more advanced like Nvidia Jetsons but still this is the platforms like with much much much smaller computation power so we can't run some bigger model models here and yes the building such a robot it's the like way of compromises we need to really really care about like prompts of our model really care about the model itself hopefully that's we have that small models like 7 billion parameters models even 3 billion parameters models which are like more and more intelligent with every day like like definitely now that models are much more intelligent than one year ago or one and a/2 when I started doing that project and and yes so like one thing here is to choose like appropriate model that will be just not too big but still intelligent enough to do your task in that robot right yes we can do also another techniques as quantization of the model as maybe fine tuning for some um when our model needs to do some narrow case narrow like case hopefully also there's some progress on the microcomputers itself so they also like yes they also getting better it's not so fast progress like in case in of LMS itself but still there's some so yes so actually it's very often the game of making that of game of trade offs of making that small model to be very performance still uh huh uh huh and I'm curious do you do you think that we'll hit a point I'm curious if you have thoughts on this do you think we'll hit a point where local models will be powerful enough that they can they can handle the whole workflow or do you think that there will always kind of be some need for like hey like you know we run a small model excuse me small model on the the robot and then you know it runs into a a really sticky problem and it makes it call to GPT 6 0 or whatever it is um well it's a good question hahaha hahaha um I think well again it's very depends on use case so right yeah yeah yes so maybe if you're doing like some general robot that John can do everything maybe I mean probably we will be do everything with the local LM at some point yeah probably uh especially if we will uh somehow separate the tasks so for example I have some L L m on my robot that is just preparing the cake in the kitchen another part of of the another lamb will do some cleaning of the garden and you know another something else yeah yeah yeah yes so almost more specialized your your yeah I like that thought you're able to shrink maybe turn it into a medium sized highly distilled instead of a large language model um you're able to distill it down into specific cause like no like my humanoid robot doesn't need to be trained on Shakespeare like you know like all these maybe you don't know what I wanna do with my robot it's gonna be a tutor maybe you do wanna train on Shakespeare but maybe there's different models right like you get your humanoid robot I'm kind of envisioning the future here now you have your humanoid robot it's here and they have different it's like a marketplace right and you're like okay my I got my humanoid robot to be a chef a baker and a driver I wanted to drive me around and these things and so you just pick you say OK I need the LM where it can handle driving real time you download that that loads onto your humanoid robot Baker I need my robot to know what to do if an eggshell gets in the the recipe right like it needs to be able to handle those small things which is all baked into that see what I did there baked into that that uh that model that you download I think this could that makes a lot of sense to me that you can distill it down into specific roles instead of a general it does everything exactly yeah I work on my shop yeah actually we don't need only like have the humanoid robot for everything maybe it will be even like different robot for example like for driving a car you don't need to have a humanoid robot yeah but even for bacon like cookies it's enough to have some cooking robot at your at your table and yeah right yeah and oh that'll be interesting yes I'd be interested to hear how do you handle hallucinations within you know kind of the framework of one of your one of your robots and and the LLM integration do you just kind of accept that they're gonna happen or do you take some some precautions against it yeah mm hmm uh yes so in case of the robots um it's uh it's again where it depends of the use case and like yes so for example in that train the Lima been just you know it's not a production product it's the project it just like you know to show up the use case so it wasn't been any done any precautions from the hallucinations hallucinations probably make it funnier for that one yeah probably exactly same in Tinder GPT it's like the more it hallucinate and the better yeah yeah yes but I imagine for example if we'll have some real robot yes it could be problem in some cases when we have like big dangerous robot uh huh yeah so so yes so what we can do here um well again depend on use case but um for example if you have drone that can like fly fly and can for example fly into some people and what to do to make the robot not to fly into that people if a lamb will hallucinate something well simple thing the using the LLMs does not mean that we should resign from using the classical sensors classical algorithms yes that been using all the time before yeah so so except of having that LM which will be like brain of our robot and decide what to do we can add some additional systems like collision collision avoidance systems for example yeah yeah yes that we have some sensors and that sensor even if Ellen decide to go forward it will not go like with that yeah yes it's fun fact actually I remember I've been flying with that Mavic 3 drone and yeah it has something like that like when you going closer to some obstacle it and you push on forward it's it's not going even when the human asking him to go it will not go so same thing here's like here like with LLMs yeah that makes a lot of sense yes so like generally in my opinion the approach we should use like in LLMs in place of LLMs into robots is to make them it's make them like replace human it's make the to like do like general general command commands high level but not to make like not to make that low level control we need to leave that uh systems uh like classical systems to do that control yeah uh like for example if you have plane uh you still have some systems that controlling the elevators how which angle that elevators should be uh should take but that will general general system that LLM will just decide which direction we should fly and yeah at all um yes so that yeah cooperation that makes sense and I think it's easy it's easier for me to imagine an LLM being able to power a special like a drone like a delivery drone or something like that where it's so specialized it kind of has one task and it just needs to be able to reason through any iterations that happen oh a bird's flying in front of me I need to avoid the bird or I shouldn't fly it I should deliver the package I shouldn't fly through the window you know this guy's got a shotgun I should I should avoid him I should fly away like it just needs to be able to handle those what about I know I'm bringing back to like humanoid robots do you think I'm trying to envision what what it's gonna take for the majority of the population to want to adopt a humanoid robot that has an LLM brain right that can interact and it can speak it can ask it questions and it's gonna search the whole of human history and give you an answer I wonder if if that's gonna scare people or make them trust the robot more do you have any thoughts there I mean I think it's it's not the purpose of humanoid robots to be an Wikipedia search engine or like just talk with them it's you know you can just run agent on the computer and it's already a lot of agents nowadays that can search through internet and make some reports and whatever can talk with you but in my opinion here robots is will provide you with some mechanical part of of all the process yeah yeah so they will be used usually when we need to you know just to do with their hands mm hmm yeah that makes sense okay anything else Spencer you wanted to ask about LM Robots because I've been really excited to talk about Tinder GPT now let's let's go on to Tinder GPT I'm ready for it I'm so curious if we can back up because you said this was your first first kind of first for a into building building with Gen AI what can we get the backstory what fueled Tinder GPT yes so what is Tinder GPT is the application I created for replacing myself on Tinder to not spend time on you know writing with different like girls on it but to make the agent that will do all that tinder bureaucracy for me and I will have just like meetings in real life and that's all yeah and I remember the times as I said it been like as you said Jacob it been like my first project with Gen AI I remember like when I've been starting that and having some friends who also start learning gen AI yeah and been like wondering oh what should I do like which which project should I do to make some 1st gen AI project to learn it in practice but I haven't had that problem like I always been dreamt to be something like Tinder GPT that will do all the tinder stuff for me yeah you're like I am so sick of it please I've been waiting years for this yes oh my gosh and finally I had the tool to to do this oh that is amazing how long did it take you to build uh hello um it been a few tradition of Soviet like uh I think couple of months I've been built it like in different versions like less advanced more advanced one OK definitely now like I will build I would build it like in maybe a few days but but that been like just beginning of January I've been learning about it that technology that technology also been haven't been so developed so it took definitely much longer yeah uh yeah yes so actually I'd be curious uh huh with Tinder GPT what model is is is powering that is it GPT 4 0 um what do you have behind the under the hood so to speak uh yes so at the very beginning of the project I built it just with GPT4 because like two years ago it been only GPT4 available alright now like to be honest I haven't been developed the project for like few months already maybe I'll be return to it someday and make much more optimized version so uh yes there is GPT for all under the hood but I think it's like uh it could be like any model probably cloth for example could be may not sure about cloth actually like a cloth is definitely good better in writing text so also better in writing messages but cloth is not good for tinder as it is um like it's too polite oh interesting yes yeah this is this is this is where my question was leading actually is like I I was curious if you have noticed you know is there a particular model that happened to be particularly suave and and good with the ladies or or is there one that just you know I can't handle talking to flirting with people yeah right uh yeah so like I haven't been experimenting with different models like that time some time ago and but there's been like a big problem with this when you're developing the Tinder GPT and you wanted to talk with girls like a you know like a bad boy and all the models they are so polite so nice they are designed to be just nice chatbots they just you know have no balls to to have like good text put in there and hahaha oh that's funny I hadn't really thought about that yeah so I how does it how does it link with so you're on tinder you're trying to connect with a woman how does Tinder GPT automatically respond when a woman messages you or you having to give the message to Tinder GPT it tells you what to say and you copy paste it over what's the linkage there how does it work no no no so like at the very beginning I remember some time ago it been like a lot of uh quite similar let's say projects where like you have installing some application on your phone and it help you like propose some messages and yeah you need to copy paste it no no Tinder GPT is not about it so the GPT is about the complete automation of whole the process like from the first message to the date so that's amazing so all you need to do is just like swap right or left the profiles you like or don't and the rest of the job is in the dgpt doing for you just writing the first message based of based on the description of the profile next continuing the conversation with that girl I don't know maybe boy if someone like if some girl using it going to some stages of the conversation like building some relationship after that inviting to the date like scheduling the date and so on and when all that pipeline done like told few amazing stories and asked like few questions about her uh it asked about like phone number proposed date and when received a phone number in the message just just send me to my phone a phone notification oh my gosh you have you have like that that data point that uh wow has has Tinder GPT ever made anything up about you that it's just like outrageously untrue that you then had to correct uh he made it all the time it's hahaha it's about hahaha oh you said I was a doctor yeah hahaha yeah sorry about that that was that was tinder GPT that wasn't me hahaha yeah exactly you know hopefully I never had like situation in when I in real life I've been continuing the same conversation than in Tinder Tinder like never but still like uh it's amazing that it's hallucinating like some fun stories because like I never had so amazing stories in my life as Tinder GPT can hallucinate for me that is so funny so how did it work how many how many dates were you getting before Tinder GPT and then how many were you getting after um yes so okay uh two questions so how does it work it's actually uh all work through the just through the computer you like installing it like cloning it from Github and okay manually asking it to write some messages so the better way but a little more once to run it on the Raspberry Pi when it can just run all the time and few times a day enter tinder and write a messages correct uh yes but um about about that number of dates so this is the actually common misconceptions misconception people have about tinder GPT the main purpose of it is not actually increase the number of day you have in uh huh main purpose of Tinder GPT is to allow you to not spend your time and lose your mental health on the data implications yeah yeah yeah so you still get the same number you just don't have to do everything together yes exactly exactly that so that's amazing so you know at actual idea of the Tinder GPT the purpose of it is like to save our time our mental health like I believe that tinder is like it's actually a very toxic environment you know very yeah totally exactly and it's not only like our words it's a lot of scientists like proven that that is really harmful for people health and yeah tinder is like tinder itself does not want to like make people happy but just want to addict people from tinder and make spend them money on the tinder and time so yeah you know when we having some dangerous environment for example your house is in flames you it's better not to go into that house but send some robots that will go into house like save some you know pull out your date yeah most important stuff oh yes and this here is exactly the same you know you send in the robot in the toxin environment of Tinder and it go get your date for you from that environment I love that so have you thought about I guess actually first have you ever told a woman you've gone on a date with that like hey tinder GPT got us this date have you told them that or you just kind of let that no totally it's it's the best part of the date hahaha better not to do it on the beginning of date but a little bit later yeah yeah ha ha ha after you already ha ha ha kind of ready for that situation so do you think like yes just just I've been finished yeah go ahead it just it just the best part of the date when you talking girl that hey did you know that you've been picked up by by AI oh my gosh hahaha yeah yeah I can imagine I'm trying to imagine if a if I went on a date with a woman and she told me that if I I love AI so I'd probably be super impressed I'd be like you are so cool it would depend on the person too how much I'm enjoying the date yeah that's true if it's a sucky date I'd be like man you suck I'm waiting yeah but if it's going well I'd probably be impressed yeah exactly oh that's so interesting so with Tinder GPT when I started thinking about this it brought up all these questions about more more ethical questions about the future of human interaction because I've seen articles I mean beyond that we've seen we haven't talked about this for a long time but a year ago or so there is this big it kind of went viral people having AI boyfriends girlfriends and there was a woman who married an AI an AI man um anyway that came to mind when I was thinking about Tinder GPT and what this means for the future because you talked about how it's able to set up dates but how amazing would it be if it could also if everybody had a tinder GPT if you had an AI almost AI representation of yourself where it's trained on you and it's able to filter through other other individuals AI so that the dates it's scheduling for you are filtered for you and it says I know I know Spencer I know Grigorij like here are the women that are gonna really make them happy and it schedules those days I don't know like it's a fascinating idea exactly this is actually the purpose of Tinder GPT I mean he's we should differentiate difference two types of the projects that first one you you told about that digital girlfriend and they teach the GPT itself because first one is about making people like you know on the computer addicted from the technology and yeah and spend time with the digital girlfriend instead of real one but Tinder GPT in opposite is uh making you to spend time with a real girl girl while doing all the digital work for you uh yeah so I like that yes about 1st type it's OK I'm happy that such project exist the more digital pro girlfriends in the web the more normal girls for us eh yeah yeah that's true ha ha ha uh yes but at actual idea of here of Tinder GPT and I mean it's not actually about Tinder GPT itself it's in my opinion it's about our future of communication of internet communication like we as a humans are not created to sit all the day on the computer we evolutionally created to uh to converse with other people like in real life not just sex sexing yeah you know and and I really believe that someday our AI our AI assistance like like in the GPT but not only will be just able to talk with one with another uh will be able to do all the stuff in internet that normally we need to spend time to to solve that they will do it for us they will appoint our meetings and we as a human just will do like I hope some real relations here in real life yeah so you're you're kind of optimistic case for AI as you believe that if if we do it right we can um build a world where we can have the benefit of the internet without having to spend so much time on the internet yes I can say so exactly I like that yes yeah cause so yeah it's a it's an interesting problem of of uh how do you yeah what's the line between like a a pro social use of AI as communication and an antisocial one um yeah eh pro pro social and sorry antisocial yes yeah yeah uh huh yeah eh well it's so all depends on use guys I mean well here is the tinder and tinder GPT the tinder itself is like very antisocial use because you like addicting people on tinder making them here like spend their attention like you know waste their time on that and tinder GPT that saves you from from it so in my opinion like antisocial use of the technology it's not only a technology at all is when you like addicting people from that technology and yeah like make people to check all the time Instagram for example like Netflix yeah all that things that that generating that dopamine you know yeah uh for people and addicting them without giving any like concrete value so yeah this is this is an interesting thought that I hadn't considered before of of AI as a way of disrupting kind of the attention economy yeah um you know if I'm scrolling through Twitter I spend a lot of time scrolling but then there's like a handful of posts that I'm like I'm glad I interacted with those I'm glad I saw those it would be nice if I had more efficient filtering sometimes right had a little buddy be like hey I think that you'll like these posts now get off yeah now leave yeah or I'll beat you yeah exactly my human robot is standing behind me yes get off Twitter Spencer exactly oh it does raise so many interesting questions about the cause tinder GPT like you said it it's told some amazing stories about about you definitely and I could see it the best one it just raises also like regulatory questions cause someone has a for example someone has a false tinder GPT or a false tinder now they can have tinder GPT run it for them and they become they could become even more successful at being fake and doing you know whatever their motivation is it's it's just this tricky fine line but I think what I'm coming to as I'm talking through this is it's just like anything else in the world that's what we're discovering with AI that anything can be used for positive or negative it's but it's the people people are gonna use things for bad and good but that doesn't mean we should hate the tool and stop the tool so I wanna make a point off of that as well people um I think that people tend to judge technologies too quickly in a lot of ways right yeah um so I'm gonna I'm gonna get academic with it looking back you know the first city well first city in history I was trying to combine those words it doesn't work um mohenjo daro in in the indus uh valley right um near as far as we can tell the city was made possible through the invention of agriculture and you know highly managed agriculture at that at the city level and then also it ended up being the downfall of Mohenjo Daro because it ended up you know they they over farm the area and you know the the the conditions supporting the existence of the city collapsed right um but you know that doesn't mean that years later agriculture was still a failure because like hey you know these these first adopters had a hard time with it hahaha yeah um you let's throw out the rest of you know human history agriculture sucks right and I think that that I'm worried about people doing a similar thing with AI where you know there's there's going to be initial failures in in using it there's going to be problems that arise and I think the key is building enough really positive and cool use cases to to um keep the momentum going yeah um anyways that's a little bit tangential I just I've been thinking about that lately and what it did yeah it's just a tool in mind like yes exactly how we will use that tool it's all depends on us yeah it's a it's a technology we have to use it consciously um it's exciting but scary cause there are some people that come to mind that I do not want using AI haha um but there are many that that should and it's sometimes painful when you're talking to someone and you know what AI could do for them the amount of time they're spending on a task and it's taking away from their family it's taking away from relationships it's taking but if they could just automate it with AI all of a sudden they'd have this time back to actually invest in themselves their relationships whatever it is OK yes that's the point of like making that automations like Gen AI agents and yes yeah totally agree so yeah and and sorry quick quick comment on that as well I think that people a very common sentiment I see being expressed um in you know kind of pop culture circles especially pop culture circles that are kind of broadly anti AI is like oh you know like where's the AI that does my household chores I don't want an AI that can paint for me um and I think that the point that we have to make to those people is that the you know the ability to be creative and the ability to fold laundry are shockingly close together in in in terms of in the you know in terms of intelligence um we need it all uh we we need we need creativity built into these things we need humanity built into these things so they can fold your laundry and that sounds that sounds ridiculous but like it's it's true ha ha ha um life is complicated and and intelligence needs to be complicated as well yep yeah exactly totally agree okay see I knew this would be a problem Grigorij that you have too much going you have too many exciting projects but I do want to hear about clean coder I know we're definitely gonna have to bring you on but can you give us a quick history of clean coder what LED to it and what is its purpose yes so actually clean code and in GPT it's like has same pro purpose just to free some my time to other things than or either writing in internet or just writing a code so clean coder is the project is the just AI coder that appeared with the purpose of automating some my coding tasks I have in my startup for example so when like more than one year ago I started my startup tagzini.pl I had a lot of like just a lot of coding work to create that all the web application like all the back and front and stuff so on yeah and um and yes and like OK I could spend all my time like sitting 8 hours all day a day on coding but I prefer spend that 8 hours a day on creating AI agents that will code for me in the start up ha ha ha yeah and that way I can code a beer and well many people could say that okay but there's a lot of many other AI coders now and it's true but first of all when I started doing the project there were no any AI coders available yeah yeah and second of all like it's different AI coders some of them could be just like help you in some tasks uh like OK I can write some function for you improve that part of course that part but have no that global vision of the of the project yeah um and here's the idea of clean coder is to automate as much part of code as possible hmm to free as many time as many of my time like as many of user's time as possible to like to be as much autonomous and like to show you like concrete examples why why like we doing it it's like four main for my pillars actually we working on clean coder to increase it autonomy like in comparison to other AI coders so first of them is the well intelligence obviously the more intelligent you have your AI coder the less time you need to spend on debugging your code after that AI coder yes so here for example we created we created a so called planner agent like some time ago it's something like either agent in like either framework very famous one and they have like architect agent that making a planning of whole code changes before like introducing the code itself which is like yes yes which is cool idea because like when you think about something before doing it you probably will do it better yeah so yes so fun fact guys from either came with that idea like on the late October as I remember of 224 we have that feature in our cling order like from very beginning of the project like something like on February of 24 like half of year more than half of year earlier uh now lastly we introduced the 2 step planner so making the planner agent even more advanced uh allowing him to first think about the underhood logic of the code without proposing the code itself next propose the code based on that logic and next write the code based on the code propositions so it's like making it even advanced uh yes so intelligence is the first thing second thing is the um is the what we have uh the uh code base research so very often AI coders struggle with finding appropriate files to work on them when writing a code so when you have like big code base it's could be quite hard and here again we're using like quite innovative solution not many other people using it if any like um we describing by AI all the code in our code base like all the files and chunking the files by functions and also describing all the functions and then index and the descriptions in the vector of the base and yeah actually we are able to yes to sort by description by semantic descriptions our code base so yes like to compare to some the cursor have also some vectorizing of the code but they have like they only vectorizing the code itself to vector database without making descriptions in the meanwhile and like in my opinion it just losing the um the sense of the code without that description part like it it likes to uh yeah cursor if it if it doesn't find the file for the code in the folder that it wants to find the file in it'll it'll it'll often just start writing new code even when I'm like I already have this utility in here man you don't need to write it again exactly exactly uh yes so like in my opinion our like code base resource system is one of the best maybe the best of all the coders again uh yes and the third thing is the is that all the automations we have around the clean coder so what do I mean here by automations is that work of the programmer is not just about writing a code it's about very often testing it called for example you need you written some script so you need to run the script to see if it's working or you have some like mistake on the line 55 you have like run in the okay so in go into line 55 and improve it like in another line you have like forgot about closing bracket and so on so we need our AI to be able also run that code and see if it created good code or like wrong code so that's why we care about we created a Python execution functionality we created even a front and feedback feature which means that when we run in some front when we create in some front end application the clean code can go into the browser can log into into our application it's just created can make a screenshot of the front end and be able to see how our front end look like and interesting you know yes and that way like iteratively work on the front end like it can see that OK that red button should be in a different place not in the top of the screen but in the bottom yes and should not be red but blue and so on so wow yes we just need to provide good feedback to our agent and fourth thing last one is management capabilities so actually many of other coders they just allowing you to provide a task what to do with your application and doing the task but if we want really have that coder AI coder that will be some one day be able to just run on our server and overnight create a new startup for us we need that coder to be able to plan the tasks for itself so that's why we have also manager agent that planning all the task in in the in to do list in our case but wow you know it's like same the human scrum master planning task in Jira and yes so we need that AI just to plan it the task for itself and then next execute the task so wow this sounds like it's getting close to AI stuff here as I said it's our purpose ha ha ha yeah yeah we're getting closer and closer uh for example even the fun uh fun thing we've done lastly uh so for now you have that list of tasks created by AI you're taking first task and executing it and executing one after another like but lastly we had that that thing that when imagine you're executing first task from the list AI in the background starting also starting also work on the next task so for example like you doing that like functionality of creating the header of your page AI in the background starting a research or file research for the footer of your page and next then you will finish the first task and start doing next task it's already half of done it's just have already file resource done and you just need to go through the rest of iterations like without without spending time on the file resource uh wow yes so like long story short we doing everything to make Ajax closer and just to make people to spend as less time on coding as possible love that as we're wrapping up here um if you were giving advice to someone who's not necessarily into AI yet a little bit worried about it a little bit interested in it what advice would you give to them to help them put their concerns to rest or at least uh be better informed about them um well you mean some non technical persons how do how can they be better informed what how to use AI yeah and just just how can they have a better perspective on AI in general hmm well it's good question ha ha ha it's question for the whole podcast ha ha ha ha yeah ha ha ha that's true uh yes so well hard to say I'm technical person so I have definitely different perspective ha ha ha right right but yes but I advise to go to some AI conferences to meet some people on that field to talk with them to see what they actually doing um what do they achieve to listen such podcast as AI Rebels and also get know what's going on in that field yeah uh yes so have that few sources of information like both internet and in person and yes definitely even last lastly I also actually also ask myself about such perspective what can I do like what what perspectives are interesting in AI field because I'm same I even I when I'm technical person I'm sitting like very deeply in developing clean coder and sometimes can miss some new updates of uh right of technology so what I did I just yes I just called like few of my friends who like also doing AI talk about their opinion like ask about their opinion what do they think about it and and yes that's how you can do like it's just amazing to have nice friends good one who who know in what's going on in that field yeah and talk with people yeah again I love that we're all in this together exactly exactly yeah let's hope we help us to to join and not to divide love that awesome well thank you for coming on if people want to follow you follow everything you're working on what are the best platforms for them to do that uh yes uh thank you for question definitely uh all people interested can find me on LinkedIn OK uh Grigory Dudnik ask here if you have guys any ideas for like some startup with the aerobatics and have some business cases just write to me yeah also if you like need to have some AI coder asking in coder to do your job coding job for you just yeah we'll drop links to everything yeah we'll drop links to everything at your profile and I hope we hear some success stories of people using Tinder GPT that'll be fun I hope so as well thanks Grigorij we'll definitely stay in contact we're very excited to see where you're at in even three months I'm sure things are changing so fast exactly it's changed really rapidly thank you guys for thanks for coming on the show yeah thank you it's been really nice to to talk with you thank you