This is the best explanation I ever heard of a neural network people started using it and so on and right now you can't do anything anymore without the internet and that's the same with AI if inside the system in the Deep learning model it kind of switches okay the most optimal way to avoid that there are coming Rockets to us is firing our Rockets welcome to Apple Finch pudding Your gateway into the world of science today's scientist is Bruce akan a PhD researcher in Agricultural and food Sciences at ilvo and ugent welcome Bruce hey hey
good morning thanks for inviting me thanks for being here my first question is always the same do you have a fun science fact for our listeners well I knew that a question was would be coming but uh yeah I didn't know anyone or any any fact but uh yesterday I uh I was reading um an article and apparently Trips so the insect with which I'm I'm working um apparently they prefer virgin females compared to other females so yeah I had quite a laugh that's a bit of a weird fact but that's nice so the title
this experience males recognize and avoid mating with non- virgin females in Western Flower trips so yeah they're a slippy bastards yeah okay so you work in Agricultural and food Sciences I can imagine that most listeners are just imagining Farmers on a field that's not completely what you do so could you explain the the research field in general um so the the whole research field so Agriculture and food Sciences are any science related to Agriculture and food you could say makes sense yeah yeah true but not only I mean um new uh harvesting techniques and so
on but could also be um for example ILO the research institute on which I'm working they also develop high-tech drones which Which help at sewing the the fields uh but also alternative food processing methods um also breeding for climate resilient crops um so it's not only like very farmer related I mean it's related to to Agriculture and food but like very very broad for example my research it's also um part of it but I'm focused on ornamental crops so more flowers and nice looking plants which only actually have the function of being beautiful and they
don't need to produce food for Example um so yeah that's basically my research field the ornamental plant Sciences or floriculture all plants are nice but but I get your point because yeah some plants plants have a better aesthetic aspect or more a greater aesthetic aspect than others yeah so you have the electrician plants and the the princess plant uh okay yeah we can mention it like that so and and and you're working on the princess plants or something yeah Yeah just need to look nice okay and then what is the goal so you're working with
the princess plants what do you want to do with the princess plants so my PhD is part of a bigger uh project which actually is looking if it would be possible for the Flemish breeders so we're Belgian both of us um and yeah I don't know why I mention I because of the Flemish breeders or Growers and currently they grow the flowers the the night looking flowers Inside green houses but actually green houses it works quite well but it of course uses quite some energy uh to heat the whole greenhouse and so on um and
also space is quite limited here in in Belgium um so that's why they're looking from the government uh if be possible to switch to ornamental vertical farming so vertical farming instead of growing just plants um next to each other above each other so stacking them vertically Because in that way you could uh use the space more efficiently of course it can also be pests and diseases occurring inside the vertical farm so I'm also Scouting For That using the cameras or at least trying to uh and we're also looking at can we uh um treat pests
and um disease infections using lights so UVC light RGB lights and so yeah and I can imagine like you said that temperature like if you have to heat a greenhouse that's a major cost so By implying or applying vertical farming you can actually reduce that cost because you have more plants in the same space is that one of the ideas that you try to tackle it's not actually a problem that I'm trying to tackle that's a A major advantage of vertical Farm because it's also insulated and so on but I'm trying to make some kind
of crop development model and look if it's possible to steer it using uh other climates so you adjust the climate based On how the plant develops uh real time so you know uh if you plan for example okay my plant will be ready within eight weeks normally but if you're six week far and you see okay my plant is not uh growing that well my model would um suggest some changes in climate early on to make sure that you get that uh Harvest State and I think one of or a few of the bigger differences
compared to normal Greenhouse farming is that yeah because of the vertical layers the Humidity will be different but also light because you have to add more lights right and it's that then part of the inputs that you're trying to control to have an ideal output yep so the parameters which I'm focusing on right now is light intensity which is very important for flower plants of course um temperature humidity and CO2 concentrations one of the things that you focus on is also pest detection but on what bests do you focus and why are They a problem
yeah so um my system will focus on drips and whitefly drips are actually very small uh insects um and the the flamish name of it is actually um yeah dund which is like thunder fly because when it's it's when you have a thunderstorm just the minutes before or the hours before you have like very small flies that are or insects that are everywhere and apparently those are also trips um I didn't analyze uh we don't have that many thunderstorms over here So I didn't analyze or check the fact um but yeah it's a very small
insect um I think it's 1 mm big so very very small very tiny um and white fly it's 1 to 2 mm it's actually a small fly with white Wings you can yeah see them very well because they they're white um but um yeah they cause major damage actually to crops if you have it in a greenhouse for example they spread very fast so their reprod reproduction cycle is very um short so they can produce yeah their Amount can increase very rapidly and spread everywhere they cause direct damage to the plant so they eat parts
of it um but also indirect because most of them can also transmit viruses and you really don't want to have that in your plants as a grower um so that's why we focus on it because especially for ornamentals like I said they're princess plant so every small um dot on top of it or or every small blister or I don't know what really causes Financial loss You you mention that a short life Cy life cycle but what is short is it an hour or is it 10 days 10 days is also relatively short they're not
bacteria no yeah but uh but for example trips depend it also depends on uh climate conditions temperature humidity and so on but for trips I think around 25 degrees and let's say 60% humidity it's only two weeks to go from an EG to an adult but every adult on their adult I think the females lay a few hundred eggs per day Uh and they live yeah one and a half month or something like that yeah yeah so if you see them you have to act very fast because otherwise you have a big problem yeah yeah
that's why um normally right now they use sticky plates or sticky traps it's actually um just a fluorescent plastic uh yeah of plastic which they cover with very sticky glue and which You Love by the way I I no when we start before we started you said some of that glue on my Headset okay it's really pain has to get it off but yeah it it works it's like it's just a piece of plastic plastic card um covered with glue and they get attracted by the fluorescent color of it most of the time they're yellow
because the color also depends on which uh species they attract uh so we have different trip species subspecies um and some of them are more attracted to Yellow some of them are more attracted to green and so on and that really Depends on their so they have light receptors inside their their eyes so for UV blue and green light but um it really also depends from population to population um some of them are more like their sensors are more towards the green uh part or more towards the blue so yeah but yellow is quite commonly
used because they attract everything um so they just yeah they fly around they think that it's a flower they fly towards it and then they get stick to it And actually they use that to monitor for trips but they just put the sticky traps there and then you have people walking around continuously in the in the greenhouse looking at those card do we already have trips no okay that's fine but once they see we have trips they yeah really try to act as fast as possible to get the plants out to uh spray pesticides but
at the moment that's not allowed anymore to use every product also you have a lot of um Pesticide resistance so yeah and those traps also of course catch some trips I mean they're like covered in trips but you won't um get your Greenhouse Trips free just by using those sticky traps it's really for monitoring ah okay so that's that's something I didn't know actually two things I didn't know because I didn't know they were attracted really to the color so that that's interesting to me so because yeah I I understand Could they could be attracted
to flowers in the same color but it it doesn't look like the shape of a of a flower most of the time it's like a big square or something yeah yeah so and and also it's like you said it's used for monitoring but not for actually removing the pests so how do you remove the pest then well it's a a combination so if you put a lot of C traps you can kind of reduce the the trips population but you won't kill all of them at least if not if there's a Big population uh but
then yeah it's combination with spraying pesticide uh they also have biological um counter measures which are just like other trips which eat the bad trips or other problem with with the with the new trips yeah yeah for example biowest is a Belgian I think yeah I think it's originally a Belgian company and they sell um like beneficials they call them it's are other insects which eat the bad insects for example uh everybody knows um effit But um is eff like the the common English name or is like I think so yeah for example they they
use them um ladybug ladybug yeah thanks and ladybugs aren't bad for your plants but they eat the the effs which is nice so for trips it's kind of the same yeah yeah they also had some yeah that's also sometimes a problem because like they imported in the past like Asian ladybugs but now those Asian ladybugs are a problem here cuz they out compete the L the native Ladybugs yeah yeah so it's good to have a pest control that is local if possible yeah yeah true but I think biobat is also looking for that because they
can't just do whatever they want they have strict legislation by the way I think I said that the trips think that the yellow plate is a a flower but they just think it's a plant because the yellow color triggers their uh green receptor okay okay one of the bigger parts of your research is Actually the detection of these pests mhm you use AR for that uh so can can you tell tell us a little bit about that workflow what you're doing exactly yeah so AI is a very broad term first of all because at the
moment you hear it everywhere every company says yeah we're going to solve it with AI and I kind of hate it when they tell it um because yeah it's such a broad term but yeah what I'm doing I'm using deep learning which is a um a sub field of AI um and Actually I just feed images to a neural network probably everybody has heard of neural networks so far um I feed images to it then train the neural network and then the system actually learns how to recognize um my pest so I feed images I
feed the solution so like the labeled image so just an image with a rectangle like rectangular shape around the pest and I say okay over there there is a pest it's a trips for example um and then the system kind of Uh yeah trains itself actually one of the things that you mentioned and I think it's really important and I love because you already explained it a few times to me he said deep learning as a type of AI but there are different types of AI could you explain like the different levels because I love
your explanation of that okay I'll try to redo it then um but yeah first I don't know if anybody's listening who also has uh Knowledge about artificial intelligence because right now I'm going to give an a Very simplified explanation because you have a lot of definitions a lot of interpretations levels and so on but first of all you can see AI as just a system which performs a task and that task normally requires human intelligence I know it's very broad so you actually have three parts a system a task and human intelligence that's the three
things you have in it and then AI Is yeah the broad term but that can for example be a deep learning model that I'm using like a best detection but it can also be a vending machine it's also AI if you look at it because it's also it performs a task it decides am I giving you that soda bottle yes or no if you put the correct amount of money inside of it that's rule based but that can also be seen in theory as AI for example then but basically for AI you have data which
you put inside you have Rules that you put inside the system and it outputs answers solutions to it um for example if you would I think I explain it to you using a plant sorting uh machine so you have a machine which needs to sort three types of plants what you do you kind of measure a lot of plants get the features out of it you phenotype them you get the height the weight uh the shape all those those parameters you get out of it and you look at it and you say okay if the
shape Is cylindrical let's say then it's a cactus and you put it inside the system and then the next time you feed data to it so just features and rules then the system will know okay I look at this feature okay if it's bigger than this or that then okay uh I can classify it in that class that's generally speaking also AI but then a more a sub field of it is um machine learning and machine learning is very good at pattern recognition for example but the Difference here is that for machine learning you still
provide data input data but you don't provide the rules most commonly most of the time you don't give them the rules but only the solutions so you kind of say um yeah you Fe the yeah the input is data the other input is Solutions and the system learns the the rules um for example for the plant sorting machine you give it a big Excel table filled with all features and then Um you add an additional column with the answer so the class to which um plant it belongs to which plant category it belongs and then
the system will look at all those features look for patterns inside uh how it can classify and get to the answer that you also gave as an input um that's machine learning and then deep learning is a St field of machine learning yeah you can say it's quite the same it's also not giving uh it learns the the rules itself you don't Give it but it uses more complex data but as for machine learning I kind of gave it tabular data so in Excel sheet with all the features nicely put in columns for deep learning
I just give it a very very very big data set very big Excel table like all kind of features some of them are useful some of them are not like I can mention the daytime uh also the I don't know how many people that are walking around the field that doesn't make sense to make your Classification based on that but that's something that the system also needs to learn so it kind of looks at all those features and looks okay this this and this is useful this not it just it's weight so it trains itself
and then it gets to the output also uh so it learns the rules itself um that uses neural networks deep learning most of the time so yeah and they call it deep learning because it has like neural network with very deep layers so lot of layers uh Behind each other that's why it's called Deep learning so AI is the broadest term a subfield of it is machine learning which is good at pattern recognition machine learning uh for example support Vector machines case near nearest neighbors and so on uh are also machine learning and then the
subfield is deep learning which is more like object detection chat GPT large language models um image generation and so on that's then more deep learning yeah I hope that Was clear sorry you no no it's slower than the last time I explained it to me it was clear so AI in general is when you replace some human intelligence let's say then you have machine learning where you don't actually say these are the rules and you just give it data and deep learning is when there is a lot of layers in between so it's similar to
machine learning but there's a a lot of different layers in between yeah I think simplify you can Say it like that yeah uh because like the the the borders between all those sub fields are like very vague it also depends on what you classify as as what but I think um I would pass the exam if I if I would explain it like that okay um and maybe also because so you use deep learning but there are a lot of stories of biases uh based on deep learning for example there was an algorithm that was
trained to recognize the difference between dogs and wolves And work perfectly and they recognized that the machine learning or the Deep learning was actually just seeing if there's snow then it's a wolf if there no snow there's a dog that's obviously not what you want so yeah yeah how do you distinguish between that yeah it's um deep learning is very powerful because it can um find pattern which are too complex for us to to see for example um right now we're focusing on image data like you say but you you also have In other types
of data like audio um yeah genetic data and so on it's very good at finding patterns and you need to be aware that if you beat your um training set so we is it okay if we stick to images I think that's the most easy type of data for people to understand um if you feed those images of dogs and wolves uh for example yeah you need to make sure that there aren't any other differences it's really only the thing that's on the image that Changes only the dog and only the wolf um yeah you
could solve it by by kind of cutting out the the dog and the the wolf yeah uh and put it on another background that everybody everything has the same background or just a random background doesn't matter that the system only learns to classify based on the the the dog or the the W features and the feature that you're actually interested in yeah yeah yeah because yeah it's also be when I'm taking my Pictures for for trips for example I also need to make sure that the lighting conditions um are the same are always the same
um because if one day I only photograph best trip species a and the next day I photograph best species B but um on dayb also the Light intensity change then the system will just look ah okay if it's a high light intensity then it's this strips and if it's low intensy it dist strips It's a lazy system I mean It's just optimized for yeah Energy Efficiency and and and resarch resource efficiency so it just like looks at what's the easiest way to classify and indeed dogs and wolf for example that example if you put snow
on the background or all on all images training images you have snow on the background and on the uh in one clause and the other Clause not yeah the system will just look at that and that's actually a very um uh tricky part about deep Learning like you think oh it's working nicely and then you give it another set training set test set and then you see ah it's not working um and that's a problem with deep learning them it's a blackbox they also call it because it learns the rules itself but what are the
rules that are just the weight actually of the neural network and maybe I can quickly explain what the neural Network yes please okay I try to find a analogy with something you can imagine so I'll Try to explain it like this and if it's not good you can just cut it out um but for example if you have a prehistoric civilization uh and they kind of live next to Forest where there are a lot of wolves or other bad animals living and they kind of want to create a warning system um so they just one
guy so the the mayor of the the city say like that and he's sitting inside the city at his ter of the big uh Houser he's living in and he just looks uh to the east let's Say where the uh Forest is but that's let's say 50 kilometers further the forest so he he really can't see if there's a wolf or not present so because of that he kind of every kilometer he puts a small Hut with a guy inside and the only thing they can do to communicate to each other is using um smoke
signals um and they kind of put them in a in a GD so every kilometer we have um uh yeah one of those guys sitting next to each other and I mean Also in X and Y axis they're sitting yeah okay yeah so both to the left and right and in the distance between you and the forest yeah indeed and uh but not all of those people are very competent let's say like that um so of course who's the most trustworthy the guys that are sitting sitting at the forest so if they hear something they
can say oh I hear something I'll just Light My Fire I give a smoke signal but then that smoke signal Rises okay And then the next guy who's sitting quite closely one or two kilometers further he can see it and he can decide okay is that only like a smoke signal because he's uh making food or is it really an alarming signal and then he if he's warranty can say okay I also Light My Fire he gives also a smoke um signal and it goes to the next one and like this it can travel to
the the mayor and then in the end he needs to decide on all because I think for example he can See 20 smoke signals and he needs to decide on all those 20 smoke signals is there a wolf present yes or no so let's say um that's the the the general explanation but as I said not all people are um as trustworthy as possible some are like very strict uh or like very scared and they immediately light their their system some are like very lazy they kind of get to sleep so what he can do
he learns over time so like the the first time a wolf appears okay the wolf Gets through and just attacks the city 20 people die think okay this didn't work I didn't look at the right signals so the next day he's looking again and he sees okay this signal this signal and this signal okay I can see it and then a wolf appears he knows okay I should look at these three signals uh to warn myself if there are wolf present or not and then this happens let's say over three months and over those three
months he also gives more fuel to the Huts that he Thinks okay that's a good guy I really need to see that smoke signal uh better because it's let's say he's 10 kilometers far I really need to see that smoke signal better I give him a stronger weight let's say a stronger um more fuel like or something to put on the fire like a very big signal and that's actually what the system is doing so this whole warning system you can see that a neural network uh which are also like a Hut represents here an
artificial Neuron or a node or yeah they have different names for it and it kind of gives a signal to the next uh node then um but the intensity of that signal is also weighted and we can see the or you can say that it's the the strength of the smoke signal here um and then that whole warning system gets trained over time so the weights of all the different Huts they change over time so the end the mayor he is the the guy who needs to decide is there a wolf present yes or no
Um he the that's the output so we have the input layer that's next to the forest then you have all layers in between which need to be weed and then you have the output note which is just the the mirror I think this is the best explanation I ever heard of a neural network it really is it really is maybe I should write it down to explain it you did prepare this right just in my head I thought about a a situation but for me this is completely clear because I Totally understand now okay so
we have to check this and then he sees those are better so we give them more fuel I I I could completely follow your description okay well but to come back um to what I said before the explanation of neural networks so the trained model is just a big Matrix filled with weight if you would look at the train model and open it you would just get a Excel sheet with 0.5 0.3 0.35 like millions millions of Weights which correspond to all those Notes and but if I'm looking at those weights I'm thinking yeah okay
nice numbers but I don't know anything about it so that's a black box we can't we are too St not stupid but we don't have the intelligence to interpret that um that system um but they kind of they are developing right now like semi blackbox systems which it's like explainable Ai and there they can for example the system in the end kind of give some rule or or or give some notion about how it Reasons um you could also for example uh because the the example you gave with the wolf and the and the dog
um yeah in that example they kind of because they saw it didn't work on the new set so they think hm something's is going wrong what is it and then they kind of um for every pixel that they put inside the model they looked at okay what does it predict for this pixel and then at the end they found found out that the snow Pixel it's like 100% sure that it's a wolf um so you kind of back propagate and see what it is but that's quite a lot of work how it reasons to to
back propagate but they kind of developing like more reasonable systems right now also because yeah people are getting more and more scared about blackbox systems and that's also counter countermeasure yeah cuz CU I've talked about this in another episode about where they use actually hydrological Models so on a full Continental scale and there's also part of AI in there and they're developing like you said actually those systems for example the water flow is combined by five equations but you only know three of the five and AI just makes the connection but they're now working on
systems that actually make the connection afterwards say that's how I made the connection so you actually get mechanistic models out of your AI which is insane and you can get Equations you would never thought of yeah true I mean um I think I don't know which equation was had something to do with astronomy um and I don't know they they kind of developed a an AI system quite recently and it's um named to an a mathematician I I think I don't know the name anymore but they kind of fed him all the data for which
were needed uh by a scientist who did I think 60 or 100 years about thinking about an equation it has Something to do with astronomy I don't know the equation anymore um and that AI system learned it in three hours um it also came with the same equation and that's like what 60 to 100 Years of of human civilization and you just fixed in three hours the same with um with chess um I think it's dblue the the chess the the the AI chess computer which beat the world champion in uh I think 1997 or
something like that um but then they just kind of trained the system but Right now they rebuild that system and you mean I mean chess is like okay it's a game but all the strategies like if you go to a Chess Championship or a chess club they kind of De choose certain strategies and that's like given from generation to generation to Generation Um so all those special strategies yeah we know it because it's also some part of human history right now but they kind of with reinforcement learning it's scan uh they kind of have A
a system which doesn't know anything about anything it's just a stupid model um and then they let it play chess to another system so it let it kind of plays chest to to to itself um and then they just say yeah you can't you can't do that or you can't do this and and so on and I I think also in five hours of training of course it simulated I mean millions and millions of games but it got better than the best chess champion at this moment which has like yeah human Civilization and all those
strategies from yeah I don't know how long they played and that's insane yeah but once again you can get very scared right now because I just think okay and in five hours it get better than us but that's only in chess if you give it another simple task like I don't know um classify if this is a a computer mouse or if it's a keyboard yeah it it can't do it can't do that because it's only Trained for chess it's not something you originally said but you mentioned it to me that AI is not going
to steal your job the person who's able to use AI is going to steal your job and I think that that really struck home because it's like you said we should not be afraid of AI although maybe there are some reasons to be cautious of course caution is never a bad thing I think yeah um but at at this point it it's more important to understand it than to Be afraid of it yeah but to be clear and to be honest I'm not a computer scientist um I'm a biomedical engineer and environmental engineer uh but
I I use AI um so I'm not the the right person or the guy who you should listen to to decide oh AI is good or bad and I also can't say if it's good or bad I just people most of the time compare it to the invention of the internet um because when the internet was invented a lot of people were Pro a Lot of people were con but right now you just can't imagine the world anymore without the internet look at all the the benefits that internet brought it's crazy but coupled to it
you also have bad things like uh illegal um weapon markets on the the dark web you also have child born and on really bad things also caused by the internet but I think the beneficial the benefits of um the internet clearly outweigh the the bad things which doesn't mean that Everything is good about the internet and we shouldn't control the bad things of course we should uh but that's the same with AI so when was the internet invented 2000 something like that invented like clearly out yeah outside it might be but the early signs were
a lot earlier well let's say yeah if you're let's say let's say 2000 as the the year 2000 as the cut off as a benchmark yeah yeah as a benchmark if you were living Back then and you're like let's say 40 and you say oh that new technology I'm not using that I would never use internet well I think you won't have the best 20 years of your life until now or 24 years of your life because yeah the first years afterwards people started using it and so on and right now you can't do anything
anymore without the internet and that's the same with AI because um for example if you use chpt um I think chpt everybody knows Right now uh very nice uh but you should use it um wisely um you can't just be a student and let your master thesis be written by um chipt it will your thesis will probably suck but it will tell also um bad things and and things that aren't true and so on but for example I'm I'm a PhD stud at Gant University and there right now they changed their policy that you can
use CH GPT but you should mention where you use it for because they know that they it's it's here AI is Here it's good it works there are a lot of like bad things or like not bad things like things that aren't working properly at this moment but it kind of works it's in most tasks better than us if it's really trained on that so it's here to stay and that's why they also changed their policy because they couldn't stop it and they couldn't check if it's written by chip or not but what this at
this moment is good at CH for for using chip te for example and all Those thesises not everybody was writing nicely um you have good scientists and they're um just the writing skills aren't optimal are there bad scientists no of course not I mean their scientific research is the most important to me but you need to write it on paper anyway so they kind of write it in their writing style very short sentences or very long senten let's say and then they just put it into chat GPT and say let's rewrite this for me and
that's what they allow They say okay you can just use it and then what comes out is very very nice and that's like the collaboration between human and an AI system because that's also the the best I also listened to a podcast I heard it yesterday uh the nerd L podcast yeah yeah and in there they also mentioned that um well Kasparov like the the the chess champion that was defeated by the AI system right now again beat the AI system by competing by collaborating with another AI system so it's like the human it's collaboration
yeah it's um collaborative intelligence I think they call it is it an also sort of reinforcement learning but with a human that he learn from the system or was really he was working together with the system I I don't know the exact details they didn't mention it um it's like uh yeah I just know that they collaborated but they explained that Ani against Ani will always choose let's say the most optimal solution but Then if a human comes in humans are also not stupid have their gut feelings and so on and that kind of can
trick a whole AI system if it's not designed properly uh if you do like a very stupid move it can make the system fail because it thinks what I think it's a very important point because you said like we can use Ai and we chat GPT is here to stay I assume but we have to be careful as well cuz I I that was from where When J GPT Was quite new that there were uh lawyers I think it was in New York it was definitely in us but I think it was New York that
used um chat GPT because they said like okay we have this case we need some presidents so they asked chat GPT for presidents and then it gave them several uh suggestions like this happened this happened this happened and when they were in the court turns out those things didn't happen it J GPD is a language model it make it can Make stuff up it just makes that the sentences are correct it's also not designed actually to um to be used like that it's actually designed as yeah like like you say a language model it's just
a statistical like on your cell phone you have like um if you type a sentence it predicts the next word yeah yeah word suggestion stuff like that that's actually chipd works the same but only it looks it doesn't look only at the word just before but at the thousand Words before and style and things so on very complex but that's like a word prediction so if I if it would be trained on a data set where I um like if I say I go walking um with my dog in the then it will predict park
because that it it has seen in all those training sets that dog and park are like coupled very clearly to to each other it doesn't understand what it's doing it just makes like what's the most possible or most logical Next Step true True because there was also stories that they fed all books of Harry Potter to to language model and it predicted or it wrote the next book but it also included uh characters that already died because it can't interpret di no no yeah and it it can't interpret anything if I said like I'm going
with my dog to the park it doesn't know the concept dog it doesn't know the concept uh park it just yeah knows those words it doesn't know anything more uh than that so to say Like uh to come back to your question so like if you would lose your job it's just those systems I I think humanity and also the the you say it the community doesn't expect everybody to become an AI researcher or expert but like if for example chpt exists and it can help you in your research in your daily life it's better
to use it if it's more efficient then use it uh I'm not saying AI is always better no if you can for example in research then like we do If you can fit a linear model yeah doesn't make any sense to train at the Deep learning model train for days use a lot of electricity to train it and then come to the same conclusion if a linear model is the most optimal that's actually a linear model is actually a um deep learning model with only one layer because it say also the the summation of the
weight um input signals and so it doesn't make sense then so use it wisely for example if you're using um if you Want to know who's the yeah if if you want to ask anything to check GPT uh for stupid questions uh let's say um who invented the cell phone for example or or you invented the phone um if you beat it to Google I think that uses around 130th to 140th of one cell phone charge uh but if I ask it to chat PT it's kind of yeah it goes through the whole cycle of
predicting the next words then you get a sentence like okay it was Bell I think he was called invented the the Phone um but then it uses depending on how long the output is one1 to one full charge of your cell phone so it's like 30 to yeah 15 to 25 or 30 times more electricity that it uses and so use it wisely because yeah doesn't make sense just for Googling like you can Google 40 times 30 to 40 times uh or the same amount of energy yeah so um but I think it's it's here
to stay and it will only like yeah I think we will move to Collaborative system yeah collaborative AI system in the whole society which is not bad I think it can help us a lot um like anyone who was mad about the invention of the the GPS systems yeah I I I get them but right now everybody uses GPS and all the major things that are like uh coupled to it they it's so nice that it was invented so I think we should look like that at it but yeah you also mentioned we should be
careful um and That's true like in Europe they're very very careful uh and it's good that they're writing legislations right now because they waited too much uh to my feeling to waited for too long and right now yeah the problem is where they are that it's developing right now so fast that legislation is lacking behind um but then you have the let's say yeah conservative Europe but Europe is not that conservative but like on AI field they're like very yeah uh you can't say Conservative yeah conservative or yeah and then you have the USA and
then the USA like everything where they are basically you can do anything there's also legislation coming but they're like more like if something bad happens then we're going to invent the legislation for that which is also an opinion I'm not saying if it's better or not um who am I to say that but yeah that way in Europe a lot of uh Innovation maybe is is going slower in that case uh but on The other hand yeah maybe it's a yeah in dut is a war principle I don't know it's a it's better caution something
like that yeah it's better to be cautious but yeah I'm not a believer of that AI will take over the world that we will invent a system um that kind of strives to yeah get over the world and and and take over the world because like I said for example the trips detection system that I'm developing it's very good at one task detecting trips JT is very good at One task uh predicting text and so on and yeah okay they try to make a AI system which kind of knows the concept wolf knows the concept
dog and so on and kind of integrate everything so it's a system that kind of it's like a seron stero steroids which can help you with anything knows anything but yeah we're quite far from that from there and also because like if I shut down my computer Compu which uh is is um doing the model and like I train the system to um detect As many pests as possible that's its goal detect as many post past as possible for example the moment I shut down my computer the system isn't mad because it can't work anymore
it's not like trying to find a um a way that I can't boot shut it off anymore um that's not how it works but the problem here again is the system itself isn't that but it's the human applications and the things we do with it same with the internet like I said yeah you have the Dark web yeah that's because we did bad things with uh with with with the internet if you use it for military for example military applications yeah I'm I'm I'm a peaceful guy uh but yeah if your enemy is using it
yeah you should also use it but then it's like kind of your thing with the atomic bombs at the moment a nuclear the Cold War yeah so I think yeah you should have legislations about that like you know Russia and USA and other countries have kind of a System which continuously monitors if there aren't coming any nuclear Rockets to your country because it's like the first responder of I don't know yeah first strike and second strike and so on so the moment you see it coming you should also fire your Rockets otherwise that you won't
have the mutual destruction yeah but right now this is done by yous they continuously check Radars and then they think okay is it useful or not should we do it yes or no And then in the Gold War you really had very um yeah close calls like that there was some yeah some noise and they thought that America shot a nuclear rocket and then Russia really wanted to to or the USSR wanted to to S their Rockets but then I think they were with three and only one guy said no the other two said yes
and then in the end they waited and it wasn't it was just a bird or I don't know what it was um but then it was good that they didn't fire it but If you give um add the decision to an AI system which is way better than humans like an AI won't sleep won't fall asleep and so on but yeah if it's a deep Learning System and you say that it optimizes in or you ask the system optimize um that there won't be coming any Rockets here okay the system will train and check Radars
and so on but then if inside the system in the Deep learning model it kind of switches okay the most optimal way to avoid that there Are coming Rockets to to us is firing our rockets and that's what you don't want to happen like you said it's often an optimization and we have optimizations but we have like our rules how how would you call it our ethical yeah okay that's something we do or we don't do but the the computer or the AI doesn't have those ethical rules it doesn't have any notion about Society Yeah
no cuz that's also actually what a lot of social media does the optimization is keep people's eyballs at my social media and like there has been multiple reports that um AI actually found out that uh people will stick longer to social media when they're angry so they're actually stimulated to be angry which is also maybe not the goal of social media getting people angry why do you think that that Facebook uh went from just the like Button to six buttons or more like the the heart and the the Angry button and so on it's also
yeah you're helping it yeah yeah no that's true but yeah it's it's just an an optimization and sometimes you get optimizations that are very deliberate but sometimes like you said in thei system it's an optimization that we just don't think about because we don't conceive it as a possibility that just will'll fire our nukes and everything is Solved that's not something we think about but that's also a limitation that the AI doesn't have yeah true and that's why you should I think the future is in collaborative systems like if you ask for example now to
DOL or stable diffusion to generate an image yeah not every image that it generate is fine um sometimes you need to tweak it a little bit ask three three to four times to to generate it um so um yeah you still have that human interaction that's needed the Same with they have a in a song festival contest for for AI songs and so on like the pure AI songs are quite bad or they used to be quite bad but you Al always have that human tweaking like okay this is what we think is nice and
so on so I think the future is in that but of course yeah we should be cautious but I think the governments are cautious and so on and also like there was a letter from I think musk and a lot of other top AI guys who kind of wrote a letter in Which they ask okay can we just have a general um treaty or or act in which we all promise to stop AI development for a moment because we don't know where it's going to H because right now they don't have the rules because it's
going so fast and if yeah if there are 10 companies and all of them say Okay ethically it's not fine if we would do this but then suddenly one of them does it yeah the other ones need to follow um so that's why they ask with that act Like please can we put a a break on the development for a moment until we know what for sure what will the the consequences of it or what is good or what is bad but I think I think like you said like do or or other image um
creation AI you also have the the highly how you say creative part of it and that's very subjective like art is very subjective and and what is nice is very hard to get capture for an AI I think cuz yeah yeah true yeah even for people There's there might you might have an image that 10 people like and 10 people don't but it's so subjective and I think that's hard to capture by AI stuff like that yeah true let's say if you want to generate uh an image of me for example in the style of
um f inent f for example and if as long as you feed it enough images it will generate something that looks like it and okay the models will become better and so on like in the beginning like it's it's very bad that Dev making fingers and so on you always have six fingers or seven fingers or something weird to be honest a lot of painters are bad at that as well so yeah yeah true true yeah but then um yeah it will generate something but it still it doesn't have emotion that system we want emotion
uh to something to to a painting and like it's sometimes I don't know a small imperfection ction or whatever that triggers the emotional thing and that AI is developed based on a lot of Images and it tries to mimic it but yeah it's it's yeah I don't know anything about R to be honest but um yeah it doesn't have emotions so it's not I think we should be coupled to it it's like yeah your car is also designed to get you for from point A to point B and it's very good at doing that but
the car also doesn't have any other thing it doesn't live it's it's long as far as I know it's as far as I Know as well I mean so yeah yeah true you have that emotional that subjective thing and yeah maybe objectively the AI image would be better but subjectively we prefer the human thing it's um yeah well actually when you go back to the cars you mean you have so many different cars and in essence we just want to get go from point A to B as economically as possible but still sometimes I I
prefer it a little less economical but it's Nicer but what is nicer that's yeah yeah true and if you don't put it into the data set on which the system is trained it won't take that into account yeah it will just be as fast as possible yeah yeah true just bind yourself onto a rocket and fly it will be faster if it bind yourself to a rocket go to space first and then come back to earth on yeah it's like the uh the latest ship of Elon Musk I forget the name now um that's not
the Falcon right Starship Starship yeah they also want to use it actually for um InterContinental traveling once it's working because then you would go from yeah I don't know anymore but like in 12 minutes you would be on the other side of the world I don't know the from point A and point B anymore but it's actually just goes to space then flies around and then comes back again that's insane right yeah true I don't know the the CO2 equivalents that come to it but yeah Well um we're actually getting to the end of the
episode but before we close do you have a take- home message for our listeners yeah I would say um don't be afraid of artificial intelligence um I mean the term is very broad everybody uses a vending machine so everybody uses AIS from time to time uh you just take that very broad definition but um yeah use it wisely if it can help you just use it because it's like the new internet um it have a lot of good Applications but also be skeptic for example for chat GPT what gives it as output yeah just always
check if it's right yes or no and also with all the other applications because I think with all the fake news and so on like with the Deep fakes and so on yeah that's a bad application of AI and can get like uh yeah on country or Nation levels bad consequences um if you give the example of social media and I'm not focusing on any platform right Now um yeah you can kind of steer the elections inside the country uh that way so I think it doesn't make sense to say we won't use AI uh
because there there are risks yeah there are risk couple to everything and I think I'm I'm really um yeah convinced that the benefits are way bigger uh than like only like for example only in the biomedical applications it's crazy also in agriculture and in in any field it's Crazy what it can do but yeah you can always use it for bad things as well and I think there we should be cautious and have you can also have ai systems which track other AI systems or for example at the University first they had a system which
kind of detected CH GPT uh I don't know how exactly they did it but uh because jipt always kind of generates some something else if you ask the same query but um yeah they made a new system then JD updated they had to update their System and then in the end they said okay we won't do it anymore it's allowed um but you should use it wisely be skeptic and yeah it doesn't make sense if it would help you and save you time yeah just use it doesn't make sense not to use it but be
aware of all the free AI systems like for um JD and so on just know everything you put in into the system in the free version not the paid version yeah you give their your data to them so if it's a personal thing they Also list it there yeah if it's very personal don't put it into the system um because it's uh yeah and also they can't get it out of it like open AI for example CHD they won't yeah get your password or whatever or yeah they it's it's like a big amount of data
they won't clearly get it out but for example if uh I put my um password for my uh email account in it yeah there is a chance a very small chance but if I type into the system okay my email address is This what would be the perfect password that it generates a password that's true because it has learned it has linked that email address to that password I mean the chances are very small and like um they are continuously building in um like nice behavior things like for example Di and so on uh can't
generate nud uh images but yeah they kind of introduce rules but then people are clever and they kind of get fast the rules so yeah just be skeptic be Analytical and and but yeah don't be afraid of using it um it's good it's here to stay and it will stay AI I'm very sure of that uh if you're interested in AI yeah take just take classes you have a lot of free courses online uh if you have want to have a very deep uh course I think from Andrew in it's called very extensive course uh
it's also for free I think yeah true I also listen to A podcast uh machine learning guide I think by Tyler relli or something like that uh I also learned a lot of it um he's also um he learned it by itself by himself and I think that's very nice because um you start from the same background knowledge and then he starts building um so yeah sorry very long answer again yeah don't be too afraid of it uh be cautious but I think AI will help the world in a very good way um but yeah
be Cautious and use it wisely yeah and I think it's a very good point actually that you just said to be cautious about what you put into the system as well because um there have been documentations that some company information was leaked because they used at GPT to help them but then a competing company asked similar questions and got the answers from the other company yeah true yeah let's just use it for nice stuff and don't troll other people and Just earn your money in an a in a in a right way I think that's
a great take home message the longest me take home message we had but I think no no no that's perfect that's perfect uh so but we're going to round up here this was Apple Finch pudding I want to thank Bruce lakman for the information and let's meet again for the next episode of Apple Finch pudding [Music] [Applause] [Music]
