DevHub Live - One Trillion Agents Debugging Session

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   give me two minutes yeah yeah I am

2. 1:41

   okay w

3. 3:14

   hello hello hey everybody we are here today for a very special episode with my friend and colleague GMO how are you hey how are you doing Joe good good so today we're doing something fun we're going to do some AI debugging uh all around the one trillion agents hackathon that ends on the second uh so if you still plenty of time to kind of get building try some stuff out but you've been working really

4. 3:42

   closely with a lot of this you've actually helped write most of the docks which is really exciting so uh you're a very great resource for all of this uh and I figured we could spend some time and kind of go through some just basics of getting up and running troubleshooting and hopefully people ask questions you can ask questions wherever you're watching this uh put a question in the chat and it will show up and we will be happy to uh help you

5. 4:07

   out so gameo when people are getting started if they're they're building uh what's the best way to just kind of get introduced to the concept of building an agent um for something like this well in the case of near AI the best resource right now is docs. near.

6. 4:27

   where we literally we have a whole section on agents and we guide you step by step on how to set up your environment even if you don't have python we're like you don't have python don't worry we can help you to set up your Python and then it's just like one command near AI to create the agent and that's it you're you're off to go yeah and the rest of the documentation will be guiding you like do you want to use your personal documentation to create a an agent that can answer

7. 4:59

   questions sure we can we can help you do you want to talk with your own documents we have information for that and you know we are adding abilities B literally every single week uh there's some very interesting stuff that is even coming like next week and the following and the following one so yeah just follow dogs.

8. 5:18

   near. and maybe the near AI Twitter account just to get you know like the information on when things are coming out but that's for sure the best resource that you can that you can find right now yeah absolutely so I I thought it might be good to walk through some of just the the basics of kicking off and getting started a little bit and hopefully people obviously come and ask some questions but I do know one of the

9. 5:44

   things that we have heard is sometimes just like getting started and getting the environment set up can be a little bit tricky yeah that's because you actually need to use a specific version of python which is like 3.11 maximum so if anyone is having trouble with you know just running near AI most probably what you need to do is to create a python environment which uh that will

10. 6:09

    depend mostly on what your python interpreter is and how you install it I am a huge fan of something called miniconda which is not something that most people love because it's closed Source it's not open source it's closed Source but it just have so many optimizations and it just makes your life so much easier so in my case I use just cond and creating a new environment is as simple as doing like cond create

11. 6:33

    and then you name the environment and then you put the python version but chances are that you can use anything else so just check what your interpreter is how did you install it if you don't remember how you install it then it might be also a great time to you know maybe find a wrapper for your python uh either using PM or mini or whatever you would like there and just make sure that you have python 3.11

12. 7:00

    because I think the current version of python right now is like 3.12 or something like that and make like changes yes yeah yeah yeah yeah actually it's 313 so yeah we're there's a few versions ahead yeah um I think that that's great initial advice uh so we we have a first question here is it possible to do image recognition that's that's so fun because uh so I have you in this screen and literally in that screen I'm doing image

13. 7:32

    recognition so I don't know if it is is it possible to present here let me see if I can yeah let me see if I can actually manage to share my screen in which I promise I have only python code uh share my screen allow and then I can just share this entire screen which I'm just GNA there so let me see if my yeah I don't know if you can read or not but actually it's very fun because what I was working until before joining here I

14. 8:04

    was also uh so V6 i m l n uh I was also having trouble uh doing some image recognition and the reality is that we have like a small there's a small issue on how we handle H messages on the near AI server which I'm hoping that we can f push a fix for today or not but uh so this is right now not going to be working if you try it uh I'm it's working for me because I'm running it locally uh there's a way also to run uh for those of you that don't know there's a way to

15. 8:41

    run the whole near AI uh Hub locally and there's a tutorial for that I can I will share it now after explaining this uh but basically something that you will do is a you just open your image like literally in this case I was trying with different things so for example like I have that dog right there uh and you just open the image as I hope that the code is legible otherwise I can just make it bigger you just open the image and you encode it as a base 64 so you

16. 9:13

    can see open uh reading and bites reading the image completely like just as a file encoding it as b64 and then decoding it as UTF this is just so it can be a string at the end uh and then this is the kind of message that you will want to be sending to the to the to the platform which is like the user is sending the content of an image which is an image URL uh this is just how the models interpretate images this is not something that we necessarily choose it's just like how models have been

17. 9:45

    trained where you put the data of the image uh and you explain like this is a data this is an image which is like a jpig encoded as Bas 64 and then all of the data and then you can add another message or another content which is like hey what do you see in the image tell me about it at length and then if you just run that uh and I here just I think I have it here running locally which is like what do you see in the image it

18. 10:18

    says like the image provide let me see again it says that it has nothing on the image and that's not true little model making a difficult point is it not working ah this is why I don't like ah there like so the the third is charm and we can see that this image is actually like a dog and now the model after failing two times uh I'm using here yeah your mileage will vary uh your mileage will vary depending on which uh model

19. 10:50

    you use here I am using uh yeah 53 Vision uh yeah and it says like this you have provided ver to feature a dog especially a Labrador Retriever blah blah blah blah let me see but for example now if we give it another image I think yeah here we have cows so if I just use uh I think I have to do let me open uh image.png I don't know if this one because I was also trying to see if what happens if you put like a very large image uh and let's see if it can do it

20. 11:30

    or if it just tells me that the model does not accept it enough CU otherwise there's another model uh yeah the image captures a Serene pastoral scene dominated by blah blah blah blah there are approximately 30 cows which is actually true I count them they're 28 so but I trust them to be honest with you I trust the all more yeah I don't know uh and there is there is uh this model of 53 and then

21. 12:01

    there is also the one of meta which is called uh I will have to Google it but I'm quite sure that is uh meta 3 29 something Vision let me I'm I'm quite sure that I put it somewhere in the in the documentation but anyway it will be exactly the same it's just that that model is better but again like just just to recapitulate and and summarize the the answer you will need to open the image as uh as just a stream

22. 12:36

    of bytes encode it as Bas 64 decode it as utf8 pass it with this exactly uh what so yeah what is exactly that I need for the hackathon entry what is the issue that needs to be fixed I can literally tell you like I which is need to merge a that I created uh yesterday yes so so the answer is we need to just create this PR which again like the pr is there created it has been accepted we just now need a second review from one of our uh senior

23. 13:21

    engineers and then it should be good to go so I'll make sure to to you know ping people today just to see it get through awesome stuff so what are some of the things that you're seeing of like just capabilities that people could actually start building here I mean obviously vision is really pretty neat U what are some other things that are we can start to do uh right now since you can actually run arbitrary python python code and this is something in in that that I

24. 13:53

    think that is important to be said and I try to hammer that idea through the documentation I'm I'm very sorry if if we have failed to to you know like to repeat it enough until it gets into everyone's brain you can run literally arbitrary python code so you know something that you can do and that is actually uh a way that I personally use agents is that you can use agents to as a way to interpret to to to to know what your user is saying and as a way to communicate in a

25. 14:25

    different way with your user right because since you can literally again arbitrary python code um you can do something that is called like using tools which is a very fancy way to just saying like you have a bunch of functions on your on your python code and those function can do whatever a python you know like whatever python can do uh analyze images do Neuroscience check sound whatever you want like call a external apis yeah and then you can just

26. 14:56

    have your AI say like hey uh you know I'm an agent that is specialized on medical imaging what is the user asking me for uh the user is asking to uh actually I don't know check their MRI ah okay then I'm going to call the function check MRI and I'm going to use it to pass this this image that they are passing me and etc etc or you know you can control uh now they have presented like a a very nice flow that you can use to to buy things in Amazon but imagine

27. 15:28

    that you have not seen it like you you did not see what Eric presented like I think yesterday on the near AI Channel which if you have not seen it go and check it h it's in the near AI Twitter you could literally just hardcode or code the API of Amazon and be like my user is asking like the user I know that he wants to make a search what is it searching for right like and the AI can take the input of the user and say he's searching for he wants to they want to

28. 16:00

    search for like a product okay then call the API to search for a product in Amazon and now what do they want to do ah actually they decided that this kind of products better for them and the AI can really be a very a new way to to to interact with with with the software right like instead of like having buttons and which I have nothing against buttons I want to say but like you know

29. 16:23

    instead of having like buttons and like things that like the the UI or the ux that you expected for the user you can just have natural language acting for you like yeah and and this opens like infinite possibilities uh one that I really like personally is uh the chat B that can answer questions about a spe specific documentation I you know like I spend most of my day writing documentation and then you know helping people like by giving support like this

30. 16:55

    yeah and I was always reluctant to to U AI to to help with answering questions about documentation because I was not trusting them but with the newest models that they out there is incredible what what what you can achieve and I could actually wait like I I think I have a bot for that somewhere around here uh and let me just I can share my screen again and yeah I was not expecting but uh yeah wait let me okay here I have it

31. 17:27

    so for example uh here is a bot uh I have to allow and then I have to allow here again here's a bot that is literally the the bot has breed let's say that that's just not happening right like BS do not think remember agents do not think I'm a neuroscientist and I like to stress this out they are not thinking it's just an algorithm but you know this B has like has a lot of knowledge about

32. 18:01

    or or near documentation and it's very interesting that you can actually go and ask like hey what is a wallet selector or you can say I am writing a smart contract how can I do a cross contract call and it what what it will what is happening now is that there's a database where all of our documentation is a store and the agent is trying to search into that database coming out with a document and then summarizing that

33. 18:32

    document for you and now you can see that it's like yeah so to make a cross contract call you can actually either you create the promise and then you make the function call uh and this one is for rust and if you're actually doing typescript you can do it like this and this code is actually right this this this code is is is correct um and if you think about it like now I actually trust this kind of for like you have any company or now you

34. 19:03

    can just go and be like okay here's all the you know documentation that we use for all support you can give all of that information to the vot and now the vot can has a way to go search for it and come back and and you know like and tell you something and when you start to pile these things one on top of the other right like very interesting stuff emerged like for example one little side project that uh and the reason why I was

35. 19:29

    doing like the image recognition and why I also found the same error that this user found is that I would like to create an an an application that uh actually we will use here in the house like lately we have been into counting calories at the house and then you know like a fun thing to do will be like okay so you have like a package of food you take picture of like the the table of yeah the nutrition table and then you just say like hey model that can

36. 19:59

    recognize images can you extract that into an actual table right right and now you have a table and with that table you can write python code which tells you like okay if I have 100 grams of this then 40 grams is this amount 120 grams is that amount right you you can do math you don't don't rely on models to do math for you if you can write python code just write it right but then we can

37. 20:24

    say like okay so today we ate this amount of this and that amount of that and this amount of this how many calories how much sugar and how many calories do we have left for like our diet of like this amount of calories right and and that's the part where you just use Python code and uh like image recognition but then at the vector stores which is like this way of like you know the database with information

38. 20:49

    where now you we can also start saving like ah okay on Monday we ate this goes to a database on Tuesday we ate this goes to a database and then you know three weeks from now you can actually query like hey by the way like how are we doing with or when was the last time that we ate blah and it will go look into the database come back and say like it was the 5th of February right right

39. 21:13

    right right and and and suddenly from like just having image recognition and from having like a a Sprinkle of python code and a small database here and there you can start to develop very interesting software that like me on my day-to-day basis I will want to use and so will do my wife right right and it's stuff that you probably would not do otherwise right so because it's doing a lot of the heavy lifting for you but what I like about this is like you said

40. 21:44

    you can essentially build python based tools for the LM to use so where now if you use like chat GPT you are reliant on Chad jpt to do everything for you you don't have external tooling tied into that so this is a definitely different setup yeah completely and and something that uh um I have to say that I like very much about near Ai and I like very much about like the current state of AI in general right is that if you have a bit of a knowledge of like programming

41. 22:16

    it has really come like we are really now living in a world where running an LM is just writing one line of code which is like in in the case of like you know near AI is like client. completion environment do completions and then dlm will run I don't know how I don't care how and it will give me like a result and the the models right like have come to a point where you can trust most of the time those results and yeah then you know

42. 22:48

    like if you're like me and even when the models are amazing you still don't trust them because you see them right like now I as like uh 10 minutes picture it's like there's no picture there's no picture actually that's a dog right right right and I'm very aware of that so you can just you know as a developer you can just uh adapt to that and be like okay perfect uh you know what uh then I'm goingon to use the llm for the

43. 23:15

    things that will be like you know impossible for me to code like yeah like language recognition I can I cannot make an algorithm that understands what the user is saying but the LM can help right right right but you know for mathematics I don't trust DM I can write mathematics on python right it's very simple I don't trust dlm coming up with the with exactly you know like all the

44. 23:39

    documentation of near protocol but I know the documentation of near protocol I don't know how I know how to find it I know how to put it in a database that's it I trust the llm with querying that database and I do trust the llm with summarizing the document once it found it and what what is going to happen is that right like in in in in the the toy example I would want to create of like

45. 24:00

    counting calories I would trust the the llm to take the image and say like this is the nutritional table which if I had to write an algorithm from scratch to do that I have no idea like I don't even know how to start right and I have a PhD on processing medical images and I you know like and I have an idea of OCR and machine learning to but like making the algorithm that will take that table I have no I don't even know where to start meanwhile now you can just do

46. 24:31

    m. completions and call this image library and that's it like it will give you the table and the rest I can do it so we live been very interesting times for for AI and I'm very happy about that and if you have a a bit like if you know how to code like a little bit like you don't need to be an expert that's also something else that I like you don't need to be you know like proficient in Python eight years of experience like if you just can make a couple of functions

47. 24:55

    sure the AI will help you a lot with in the the real world yeah no I mean it it's you need enough just to be dangerous right like where you can you know how to ask the LM to write you you can even ask the LM write me some python code that I'm going to then use as my Tooling in my next project you know you can just get into this circular thing that's that's so true even more not today even more nowadays with all of like you cannot avoid if you use vs code

48. 25:26

    AI is somewhere around there if you use uh and like now there's there's other one that is very much in in in fashion now uh that everyone's using to code like an ID that everyone's using to code yeah I use cursor that's yeah cursor that one I never tried it right like so then then is it's true it's fun because you're writing your AI while another AI is actually there like being like yeah

49. 25:50

    maybe you want to call this function maybe you want to call that other function no for sure it's definitely a fun time so uh it's interesting stuff yeah and I'm very excited for like what the future departs because uh like we can talk about this near AI has just shown like a demo where people can actually ask an agent to buy stuff for them on Amazon and those things actually will arrive to your house like or Engineers have tested it and like everyone has been buying headphones and like now we have excess of headphones in

50. 26:29

    the company but but it's very interesting like we are coming to a point where where you just you provide tools and you provide help to the AI and the AI just thrives by itself and more importantly a year and or like a year and a half ago it felt like you know only open AI was able to do this right like and only you know their million dollar per day engine years where they once able to you know like create a

51. 27:01

    remotely good Ai and now I'm here sitting in in in in my house I'm you know and I'm just like coding AI for me that that I'm going to be using like uh can you answer questions about documentation can you help me count calories can you you know like help me detect things on an image and and and that's it like it's it's literally for everyone I I really like it and I'm very excited to see what people is building in in this agent because right like

52. 27:32

    there's always someone that will come up with an idea that you're like that's a good idea like it it is it's so much based off of like life experience and some of those things you start to see some different possibilities uh and and also of like how you start to understand like you said like how you understand what the LM is actually capable of and when you start to understand all those pieces

53. 27:55

    really I mean the creativity really is balanced right now there there's no you really can try so many different things yeah and I I'm personally huge fan of like simple ideas that just work well I really love you know like when you have this small project that looks like one of those project that make you go like how is that I did not feel that like how is that I did not you know thought of that before right like like do that's

54. 28:27

    just it's just like maybe 10 lines of code 20 but it's just it works perfectly and solves like a real problem and it just goes well and yeah I love that yeah absolutely absolutely so uh one of the things that definitely would like to talk to you about as someone who's been writing a lot of documentation is is what has that process been like to go into something that is new it has there is no Baseline necessarily for you to go off of you are

55. 28:59

    essentially uh you are trying things out and writing doation at the same time um how has that been for you as an experience well to be honest with you that's one of my favorite experiences for for for like of being like a developer relations of being a Dev right because uh it it's very hard like people people generally don't realize how interesting how fun and how also like challenging it is to you just go there there's like half half bake script here

56. 29:33

    like or like you know an idea of an API around there like by the way like now it's much better right like H we we're not at that point this was like you know when you when we just onboarded ourselves in t i maybe like four five six months ago right like but not not anymore uh and it's very challenging but rewarding at the same time because you get to play with a lot of things that you know like no one has played with before and I'm an engineer myself so

57. 30:03

    then you know like whenever something is failing I can get the intuition of like where could it be failing and then I can jump into the rabbit hole of like okay this line actually is going this function actually something okay it's here because like we are encoding the API with a specific Json when we should be making blah blah blah uh and and it's very nice to do the thing of like create the tutorial that you think will be useful for people and run yourself the tutorial and then meanwhile go fixing

58. 30:33

    everything that does not allow that tutorial to exist and when you finish you not only have a tutorial but now you know the product is stronger like the product is more solid and and I I really I really like it like uh it's it's one of my the things that I like the most yeah yeah yeah no for sure uh one of the questions just came in is about asking about is there any Dev support and there absolutely is

59. 30:58

    yeah yes and actually there's a whole telegram Channel where if you go and ask many of us are there like we are distributed across multiple time zones so if you go uh there will always be someone so let me see if I can find it just to share it it's it's uh shared it's actually in the scroll at the bottom so it's scrolling on the screen right now yeah so as it comes back around you'll see it uh yeah so it's at near AI Alpha it's a telegram

60. 31:29

    Channel it will eventually come ah there it is I don't know if you can see it but I can try to follow it like there so just try to enter there and there's always people in there uh so ask questions we will be there to to to to to help you for sure for sure we will uh awesome let's see what other questions are coming in here see I know there's some uh now I do know that um one of the things that definitely comes up is uh and maybe you might be able to explain I'm sorry I'm putting you on the spot a little bit but what is

61. 32:17

    the difference between using like a tool versus like rag something like that uh yeah no no the these These are different concepts so using tools again it's just a very fancy way of saying you have some python functions and now you just tell your AI hey uh tell me which function I should be calling and with which parameters we have a whole tutorial uh uh and I don't know wait we have documentation in do.ai

62. 32:49

    uh loading agents and then there's a whole section about tools and commands I'll send the link here I don't know how we can share it but I will just pass it here so there's a whole um the there's a whole part where we explain about tools and it's just again like it's a fancy way to saying your AI is going to be able to call some specific functions uh which is not to you know like it's not to downplay it it's super important it's just that you know sometimes we use names and like forget what what it actually means it's

63. 33:31

    super important to have like AIS be able to call like random functions uh so basically what you will do is like there's a specific way where you can say like I have this function and I can register it into the agent and now the agent from time the agent from time to time will be able to say like ah actually I want to call that function with these parameters and and that's it if anyone is going to be using tools something that I would recommend you

64. 33:56

    strongly is that make your function like you're going to be have you have a function that is going to be be it's going to be called by a AI agent which means that hallucinations could happen so just make sure that your function your tool right like your function is ready for like checking the input saying like Okay right uh this AI agent is asking me to you know like call this this and this does that make any sense something that I like to personally do

65. 34:27

    sometimes is depending on the tool right like but sometimes I like to inject into the tool that I create also the the the history like the the whole prompt just so you know my tool can be like did anyone ever said I don't know it's joe.

66. 34:42

    near right like if you're gonna be handling like I don't know sending near tokens to joe. near did anyone in the conversation actually said Jo near because you know if yes then it's something if no then it's something else of course you know your your mileage will vary like depend maybe I just say yeah send it to Joe and the AI is taking that is Joe because he's accessing some database from somewhere but anyway always create the the input yeah and rag

67. 35:14

    that you actually I see that you have your screen there if you go on the bottom it says documents documentation uh tutorials on the yeah exactly so rag is the idea that you give uh you create a special kind of database with documents and now your uh you augment your agent with that database right so the interesting thing about this database actually there if you go to Vector stores there is the second link on the tutorial yeah and if you go scroll a little bit down we made a

68. 35:49

    nice yes exactly so the whole idea is that and maybe you can zoom in a little bit so people can see it yeah perfect so the whole idea is that the nice thing about uh a specific this this type of databases which we call Vector stores is that you put documents in and they transform those documents into vectors into some you know vectorial space so in this case we have two coordinates X and Y and each one of like these documents front end MD API

69. 36:22

    SDK smart contracts all of them are like somewhere living in this space and the nice thing about rag is that in the same way that it transforms documents into this vectorial space it can also transform user queries into the vectorial space so now when someone asks which is here in yellow in the screen how can I create an account then you can in Access this database that query is

70. 36:46

    going to end up in some point like in some part of this space and you can say Okay so this person is asking how they can create an account and the closest files that I see here are the CLI and the API so you know what I'm going to use those two files to answer so now you take those two files and you inject them into the the llm like into the prompt and now you create a very simple prompt which is like here's all the document of the API here's all the document of the

71. 37:15

    CLI here's the user query answer it and the llm will make an amazing job at answering that question using the the two documents uh yeah so those are the two main things things that everyone like that or documentation explains and that I I I invite everyone to make yourself familiarized with one is again tools which is fancy way to say hey how to make your llm called random uh not random but like a set of predefined python functions and rock which is like

72. 37:49

    hey how can I augment my agent with documents literally text documents right yep yeah no it's it's a it's a good distinction and some obviously certain people getting getting used to it um and I know one of the things too if you're using your AI that might be a little bit different that maybe you're not used to is the tooling feature and adding tools so I think if you a lot of other agents are mostly based on uh using some version of rag where you are just

73. 38:22

    essentially dumping more data to it um there are some other agent Builders like bit agent which allows you to actually build in some kind of end points that is like kind of Tooling in here you know with near AI you all your code is right there in the agent file um and then you can actually you have to specifically register th those agents it's not enough just to put the code in the python file for your agent you actually have to then tell it this is an available tool for

74. 38:53

    you to use as well yeah um I'll actually I'll give an example on that because I made that mistake myself so I can show how I trouble shout this here so I'll bring up I'll actually bring up the agent that I worked on and we can kind of talk and again people if you're if you're listening you have questions whatever else uh we can certainly do that so uh I made an agent and I I got stuck for a while where I couldn't get the LM to

75. 39:28

    uh Polar tools and I kept actually putting it in the prompt and that's a mistake that I've seen other people make too and I I made it for a while until I caught on to my mistake but I was I was telling the prompt like you know hey this is what I want you to go do I want you to go like use these functions and that's the incorrect way of doing it um actually so I've built out for example

76. 39:51

    one of my my tools here is is minting an nft I called it minting a treat because my product is called Shard dog it's dog themed right so I'm going take a treat uh so from here uh you can see it's a fairly standard thing of just like hey here's the arguments that we need uh I I provided the endpoint to go hit and so again you have to when you're adding these tools you can't just assume that the a is going to do everything for you you have to be able to give it resources

77. 40:20

    to go do these actions uh and that but this is just a straightforward like function there's nothing really overly complicated here um you can see it's pretty small bit of code but when it comes to run um I had to actually register the actual tool so it knows to use it and then in the initial prompt I said these are things that are available to you and so if you do not register it

78. 40:48

    here uh even though you put it in the main system prompt you you say here's what you need here's where you need to go to here are the endpoints it won't do you want it to do you need to make sure that you're actually telling it that hey we have these things here as well so uh lesson learned uh and if other people running into that definitely definitely something to to be aware of as

79. 41:13

    well nice and I would strongly like since we are seeing your code I will strongly recommend you to double check like on like the username that you're sending to personally in my case for AIS I have to to to notice that if you use English and well you know like wellknown names everything works perfectly but as soon as you know like your name is uh whatever like Willer instead of you know

80. 41:45

    William then most probably you're gonna end up like like William is gonna end up receiving your money I can model you use right that's a great Point yeah it does try to correct certain things like that abely absolutely uh yeah and that's where this again if this is just regular python we could build logic in here easily to just go in and do that um and again and you can see again if you haven't built this

81. 42:11

    one of these before like these are all standard you know python Imports there's nothing really overly fancy here um you know and it's it's almost easier than I think some people actually assume you think of like this you hear AI it sounds maybe more complicated whatever else I mean this is less than 150 lines of code there's a lot of white space so I mean like this is barely anything and it actually goes and it lets you create a

82. 42:41

    new nft collection in mint nft and it does all this stuff and it prompts you and it walks you through it and it's a super simple idea it's not like it's not groundbreaking but it just shows that like you don't have to get a really complicated to do something you know and then a lot of my programming was then okay just the end points you know and uh I'd be lying if AI didn't help me write those too

83. 43:04

    so and and what you said about like line counting is so is so true like um today I was playing again with what you saw about like image recognition and my my my my small script to count cows it was like 20 lines of code because of course like most of the heavy lifting is being done by the AI for sure and for any anyone cuz you know there there's a lot of people that is I noticed that there's a lot of people that is starting of course their career now and starting now you know with the whole like the AI is given and

84. 43:37

    it's something that exists I remember you know eight years ago when I was doing my PhD like how how are you even gonna you know like image recognition was like a whole you know like that's a lot of like that's a whole PhD that you have to do just to maybe you know be able to create that final script that will count house uh and it's incredible that we have arrived to this point where

85. 44:01

    it's just like okay you know what anyone can do it like just write do this one line of code hit this open you know like model which might not be you know open source or completely open but at least you know like the weights are out there like you just use it and now it's like just like magic you get an answer in a couple of seconds that yeah like for anyone that has work like you know like

86. 44:24

    in in image processing that like it's so interesting that we can do this now it it really feels like magic and I love it it it does does it's really funny and and uh it's funny that the two of us are on a call about this because we both actually our backgrounds are not in computer science it's actually you're in neuroscience and mine is actually in Psychology so we're both like uh so our

87. 44:49

    perspective is definitely a little bit different but it is like it does feel like magic I mean I have a similar experience when I was in my undergraduate I worked on an fmri project and I was manually marking uh changes in brain waves I was sitting through and going through each one as the intern having to mark that off now you would just dump it in the thing and it would just probably just return all

88. 45:13

    that back to you so I mean it's it's amazing what's what it's doing now it's it's very interesting and by the way well I mean since we are you know two people with the background of like neuroscience and psychology here then I think that it's it's important for us to say and remember AIS are not thinking there is actually a beautiful website I don't know Joy if you have seen it it's called The the machines.com no I have not seen it no no which like I

89. 45:39

    like the the name is clearly clickbait like because it's very strong uh yeah but it makes it explains beautifully about like you know what the llms are actually doing and like the actual limitations which is you know it's a they're amazing ah we use them we love them just remember they are not thinking so as a programmer you know like if you want to convince your users that they are

90. 46:05

    thinking that's great right because that's good marketing but as a programmer remember that is just an algorithm so it's it's not you don't have a personal thingy for you so act accordingly sanitize your input you know like make know that it might fail and you just need to try one or two times if it gives you an answer something that you can do if you know depending the the algorithm right because sometimes it takes too depending the model because sometimes it takes too long to run but

91. 46:31

    sometimes you can just do majority voting like hey I'm gonna be counting cows how many are they 40 42 40 38 40 well there must be around 40 so act accordingly the machine you don't have someone else being like one two three four five six it's just like an algorith so yeah and I would encourage everyone to read the is very interesting like uh it's a very nice read and it's like maybe 10 minutes so it's completely your time have to

92. 47:05

    check it check that out that's always it's kind of been like my sticking point when I talk to people too is I I understand why these companies are using the words like thinking and thought and I get it because trying to make those parallels but like yeah it drives me it's kind of like a pet peeve too of just yeah but I also but I also like that you know we are not on the other side of the hating because you can totally find like if you if you're a

93. 47:29

    neuroscientist and you have one neuroscientist in in Twitter you will see how so many of them like just hate and like you know and like how like no this machine is not thinking because in order to think you need the cognition and for the cognition you need front it's like yeah okay it's just chea it's just right right for sure uh we don't have any more questions uh you know I I definitely want to emphasize that if you are watching this on replay you can always go to the uh

94. 48:01

    telegram group and ask questions anytime um again like said like this time people in all different time zones so you know you and I are in very different time zones so uh like just there's always someone around who can kind of jump in and help the near AI team is around and available uh they'll help as well um so definitely jump into there and the docs um you know and I'm not saying this cuz you're on but the doc have really gotten really great over the past few weeks uh it's really in a great spot where uh you can certainly

95. 48:34

    jump in start building get going from there and just try things you know yeah yeah and I would like uh to encourage everyone to actually try it the varer to create AI agents is so much lower if you you know like where maybe like me a year ago that you're like I just you know like I don't have the back ground or I don't know enough about machine learning that that's just not the case anymore like right now using an AI model is as simple as like running one line of code it's not like before

96. 49:07

    like if you also were like me that you know where you have been reading about AI for a while and you were always seeing like how much you had to just fight with the right promp in order to get the right answer that has got gotten so much better with the latest like Lama 3.2 like everything like the the barrier is very very very very low right now like for entrance so go in try it have fun right and you know participate in

97. 49:32

    the hackathon you never know what could happen yeah it's low stakes it's low stakes um definitely just try stuff uh yeah don't be dissuaded by like oh I don't know anything about like you say anything about AI or machine learning or any of that stuff uh you can certainly just try different ideas and and if you get stuck ask for help nothing WR with asking for help completely cool uh yeah I I don't really have much anything else today uh anything you want

98. 50:01

    to end on Yo No that's it please come and participate and hope to see you on the tegram channel yeah absolutely perfect thanks so much see you lat have a great day byebye bye