AI Expert: Here’s how AI Agents 10x my productivity

so what's the most promising use case for AI agents you've seen so far being a software engineer AI agents able to run code and able to execute it find errors so recently I've been trying with Devon so it is able to execute code and able to if any errors are there it is able to fix that and able to make it work it's is much more advanced generally it will take like a week for me to do and he doing it in like few hours which is really good so and how did you get how did you get access to dein cuz like I tried to message them and they still didn't give me access yeah I think same same I did I did the same thing but probably I think they chose few members luckily I'm one among them and still creating videos and testing it I'm doing a lot of testing like day and night so if I start today it takes few hours so I need to come back after 2 three hours to test if it has completed or not and do you think like you know because there has been a lot of controversy around Devon do you think it's overhyped or do you think it really has potential it has the potential I think every tool when it starts generally we don't see I mean it's two perspectives some people see only the positive and the vision where it's going towards some people see only the negative from my perspective I think they're going in the right direction I can see a lot of things is going slow and also few bugs are there but I think people are working towards fixing that error and then making it a good product so I believe it will be like a good one and you touched on this phenomenon that there is like no balanced view right it's either people that are thinking like agents are completely useless or people thinking that you can automate everything right now which is not the case why is there not no like rational opinions in AI Agents from my perspective I think those who haven't tried agents probably they might be against that kind of Agents but those who really tried it and have a vision towards the future definitely they will like it that's what my perspective is yeah I mean when I interview people that are you know at The Cutting Edge that are actually building stuff with agents that are designing these Frameworks they're always the most like positive it's always like the people that are almost lack the skills to build good agents that think they're useless and stuff like that but like they you know it's much easier to blame the technology than to blame yourself and feel like okay maybe I need to learn something I think once you get your hands on like uh to any tool I think first when you see from outside it looks like okay this could be not good but once you start working on it and also if you got some previous experience of doing the same job without agents then you will realize the power yeah and I think you're actually lowering the barrier to entry right when it comes to agents with your own project prais and AI so can you talk about that bit one of my main goal to praise and AI that generally I create a lot of videos in regards to autogen crew AI task weer Lang grth so a lot of tutorials in regards that on my Marin PR channel so when I do that on my comments if you see there's a lot of questions like how can I do this how can I do that and I feel like it's people are finding more complicated sometime because most of my subscribers I believe they are they want to get transitioning into AI because that's how I focus on so I try to make it easy as possible so that they can understand and Implement based on the questions I see like okay the entry point is kind of uh the programming so only programming people can create agents so I want to break that barrier so I just thought like okay what if we have a text file they just need to fill those details in the text file and also triggering those agents is just one small bit praise in AI That's the command so as soon as they type the agents will automatically get created and they work together to give you the final output make it more easy for any beginners that's the ultimate goal yeah I've looked at the documentation in your website and it's like super simple but I think like we still we still need to somehow I don't know I feel still there is still a need for a simple UI because a lot of people are just like intimidated with when you say like vs code or like you know if you out stab from let's say like downloading olama if you alt tab to vs code a lot of people instantly get intimidated just by seeing like the programming IDE right so do you think we need like a really simple UI to for agents to become mainstream I definitely agree with that we are going towards having a user interface for anything for any framework or anything just because this is a starting phase even for prais and AI this is just starting to start with we need to make sure the back end is working fine because whatever the attractive thing that goes in the front end ultimately is what it's in the behind the scenes so what in the back end so making sure the strong Foundation or the programming side is strong then if we create a front ENT I believe then it'll be more effective yeah I mean if CH GPT you know was the same UI but the llms were like gpt2 level then yeah fail so obviously the underlying technology has to be solid exactly I think they they spending millions millions billions but still the user interface is literally similar from past one year to now but still they are working behind the scenes so amount of work the millions going to build the foundation rather than the front user interface yeah but like U you know to make a great product you obviously need both but I don't know do you think like because like I I would say like your framework praise and AI is like right between you know programming and UI because it's like super are youning your own UI or do you want to do this low code solution both I think currently even it has a feature if you type PR hyph iPhone UI it automatically creates the user interface but for now it's just a gradi your interface but eventually the ultimate goal is that anyone can like use that like a dashboard and then automatically even without any touching any lines of code then doing it everything by themselves and what's the long what's like the long-term vision for the project long-term vision is to integrate multiple framework for example currently I've integrated autogen and crew AI by just changing one uh the name of the framework you are able to use the same bit of information and create agents on both creaa and autogen I want to make it simple so that we can by just changing to task Beaver they should be able to implement task Beaver similarly by changing it to swarm agents or even any other new framework like lra just by changing the name they should use the same information because ultimately that is a centralized thing like okay this these are the agents and they are going to work sequentially or hierarchical then use that across various framework having all those Frameworks connected if you see Jenkins or gitlab pipeline it's like uh it's like step-by-step process so for example one group of team one group of task will complete then it goes to the next group crew AI team work together to form this answer then you go the autogen team who is powerful to perform the code creation and code execution task so that will be passed to the next it's like a pipeline so Tim goal is to create a pipeline and all these group of Frameworks we can use all the power and then integrate that together yeah that's like fascinating I'm already getting like ideas when you describe that because like we could just U you know if you have this like almost like a rapper right a simple way to use all of the Frameworks you could just create a task and you don't really care like how you solve it right you just need it solved so maybe there is like let's say in a year there's like seven main Frameworks and let's say in praison you have all of them right so you just describe the task and spin up seven different teams all using a different agent Frameworks and whichever solves it the best or the fastest or the cheapest you go with that and build on to of that that's the end goal that's I'm planning to achieve because every framework has its own benefits for example autogen have its own benefits cre has it own so putting it all together making it as effective as possible amazing okay so speaking of different AI agent projects what would you say are the top three projects in the AI agent space right now if you're serious about AI then I have an offer for you inside of our community you get access to free weekly calls where you can talk to me directly and of course you get all the trainings and workshops we have which will put you ahead of everyone else so if you want to be among the people who are building the future then make sure to join it's the first link in the description autogen I say crew Ai and then langro I think these three I feel like these are the one of most popular as of now but there are eventually there are more coming in but from my experience and working these are the popular okay I've tried autogenic crew AI but I haven't tried Lang graph what is the difference in that what makes a difference Lang graph I think langra L chain for example if anyone is really good at L chain if they have already learned L chain and it is able to integrate multiple apis multiple things together and if they want to build the agent system then with that Lang chain experience they can integrate that easily with the lro adding agents with more conditions and logic in a more structured way uh similarly I think autogen recently published another update which which also involves that kind of structured approach but if you're new then you can go to autogen if you already know Lang chain and if you want to choose lra so you would say like langra is almost more advanced because that's what I heard from other people like it's more complicated right exactly because L Chain by itself is a big ecosystem which contains lot of tools lot of Integrations and autogen not as early stage as L chain so you have more customization option yeah and it's crazy that uh you know there is like one is from a big tech company you know Microsoft then one is from xia Mora single developer so it's like you know it's it's amazing seeing open source like single developer team or like I think he has like four people right now but like U you know small team of like super new startup versus a multi- trillion dollar company exactly ex I think it's like the fight yeah I think it also depends on how the simplicity so let's see how it goes I think eventually we'll know how this is going to be turn probably after a year we'll exactly know how it's going to be like and are you betting on the open source or do you think it's going to be like you know same with the llms where the closed Source companies have a lead open source have a good area of like uh potential it it is capable of going to near close Source but the the aim for the closed Source model is that they have to be ahead of the game if not they have to close the company so I believe they have the more pressure to make sure that they are ahead of Open Source it's a good competition but open source models are still catching up with close and do you think like the big tech companies have their own in-house agent teams that are way more advanced than what we have or no yeah I think if they fine tune the model they and also based on their own requirement out perform a general model so I believe they will be having some models like that so it's all about like fine-tuning the llm to for a specific task right I believe that's going to be the future like for now we got all General models so next step is to fine unit based on your industry because that means for that industry that model will be the best one talk yeah the only issue like when I hear that is like what comes to mind is immediately like the average person right cuz like we talk about agents and they can you know learn crei and stuff like that but like fine tuning I feel like that's almost like the next level where most people have not even tried fine tuning a model ever and you know it's just sounds even more intimidating than python or Creer orogen so I mean very true very true but like I feel like that that's where the real like unlock is right because if you find you let's say like llama 3 70b or even 8B you could probably get better performance than gbd4 turbo on a very specific task right definely exactly that's that's where I think whole AI area is going to evolve further I think that's why custom GPT came by I mean released by open a they believe that GPT 4 con do everything at once so they need to customize it based on their requirements so for now they have the rack system but when they Implement fine-tuning as well in those custom GPT I think that's going to take next but like opening isn't really um doesn't really seem to want to go that way because they like offer gbd 3.5 fine tuning and gbd4 I think it's like you have to go through their sales team and it's like millions of dollars so like why doesn't open AI make fine tuning more available maybe like because the like the model like you know like Facebook Google they have other business models right but for open AI the llm the language model that's like the core value ad so maybe they're trying to protect that exactly I think that's a one good point now I got it so basically they want to focus on main important areas rather than then focusing on something else which they are not much interested in same like Google and perplexity so perplexity mainly focusing on llm Google if they make the search engine llm they will lose all their ad revenu so that kind of mod oh yeah that's true because you know Google AdWords is their main product so if they transfer everything to Gemini exactly then then they have to close close all the AdWords things so yeah yeah that's kind of crazy that like this is almost preventing more Innovation right or like strategically like CU like if if there was no um Google AdWords they would just go all in on AI obviously but like yeah they just want to keep the cash cow definitely need to make a definition I believe that's why they have the gmany so in case something happens if now complexity already is growing fast so Google need to make some decisions soon in regards to how they're going to tackle that and uh maybe propose a new Google Okay so let's say someone is new to the field of AI agents why should they care about agents right why should they learn how to build them I think even a lar language model on its own is it on one level but with agents we are taking that to the next level so with one llm competing with like three llms working together to finish a task is going to be of high quality definitely it's going to be high quality and we can effectively rely on that compared to a single large language model so definitely they need to use agents or that is the power of Agents so it's more like you know unlocking extra capabilities and really being ahead of the wave because what I Tred to tell in my videos is like you know maybe current agents are not perfect but it's more about like seeing where this is going and preparing for the future right so like even like your point if you have one llm well if you do that three times and each one is prompted differently or maybe even different running different llm right with different agents with different llms like that alone could help you solve any task much faster definitely it's going to complete the task more efficiently that's the one thing and the quality of the output is going to be much better so even if there are errors in agents now if you see in like 2 years now in the future or or 3 years now in the future I think in future everything is going to be agents because of the power or the quality of the output is going to generate and how far do you think like we are from getting the CH GPT moment for agent I think we not very far probably in in a year I think most of the system like even for chatbot I believe that if someone ask question even a normal chatbot will have a few agents working to And discussing with them and G giving us a reply so it's not far and what do you think are like the technical things that we need do you think like we need the next generation of llms do you think we need better memory reasoning or just UI I think system I believe like a proper system which is capable of handling that kind of uh agents and easy integration with an application is that the Gap is there so if that is filled definitely it's going to be much easier for any of the websites or any of the softwares to integrate those agents and get their work completed quickly right so like basically when somebody builds something with let's say cre Ai and they have a python file um when somebody's new to this right they don't know like what's the next step how how do they integrate it into their service right so like I think that's definitely a huge thing that needs to be solved because I've heard that so many times like obviously um as the agents become more popular you will have Cloud providers like Azure AWS Google Cloud all of them will have you know crei autogen in them still there needs to be I don't know whether it's like a library or something but like something where you just easily deploy it on your website or on somebody else's website because right now it still requires a good amount of programming knowledge to do that definitely definitely I think now it's much more complicated I think it's supposed to be much more easier and simpler to implement it on the system or integrate it I think that is a gap currently I see at the moment it's almost like open air has this perfect opportunity right because like if they somehow I mean obviously you can say that it's with assistance API you can do that but like why don't they just uh allow the custom gpts which are much more approachable for people why don't they allow some like easy integration for for like for your own GPT to somebody's website I think like that could be a huge product definitely I think because they already got the user interface they just need to allow website owners to integrate that or even on their own application if they could integrate that I think that could take that to the next level I think that's big move for them many of startups are relying on this so if they Implement that probably all these startups might go down yeah that's something open a has been doing you know just destroying startups like every time they announce a huge update it's like you know you see on Twitter multiple start startups just been like outco competed basically like we've seen this all these like PDF and when they announced um the plugins for the first time it was like why why would you go to some website when you can just do it in chbt right so definitely definitely then after that all the plugins is now closed oh it everything came to the custom GPT then before mid Journey was like really popular I was like subscribing to Mid journey and then using it regularly then di 3 came then mid journey is going slowly down but still every Everything has its own place I mean I I would say mid journey is still better from my experience because the issue with I have with di 3E is that like the model itself it's really good but you cannot actually prompt the model you can only prompt chbt and chbt rewrites the prompt and feeds that into D so I agree with that as well for customization but for me personally generally inside chat GPD I just do a lot of work and then immediately if I want to create an image I just ask can you based on this context or content can you please create an image so it immediately creates so yeah it has both yeah are just like low friction right cuz like you're already there you might as well just create it there ex definitely yeah that's low friction and like why pay an extra subscription even though I would say m journey is better for most people it's like Dolly 3 is good enough right yeah ex yeah exactly Dolly 3 more than enough I mean imag be better yeah imagine exactly yeah do you think like if they release D 4 or Sora it will be part of the jbt Plus subscription or will it be like extra I think they will keep it extra those things are like like much more advanced and also it could be if everyone have access to that kind of video production thing then you can't even believe the video is coming on YouTube is it a real person or someone else creating but eventually it'll be there but for now at least they need to like restrict further because they need to add the alignment making sure these models are not used in the wrong manner so and why why do you think like the GPD store was like failure I mean obviously it's still is running so it's not like complete failure but like if you go back to like November right or December there was so much hype around it they were promising that we're going to pay everybody right and like I don't know if even now anybody got paid like I don't know what do you think happened there I think it's just one of the thing which they thought that would work and also I think it's not being same as you mentioned like if they had an option where you can integrate that interface in on a website I think that's kind of a new area to tap into but for now they thought it would go and it didn't go because they didn't go to the full extent of supporting the other software so that they can actually use their interface on their system which didn't happen so I believe that would be the drawback and I think they also like failed to support the builders the developers because I've seen a lot of people excited uh to building gpts even people that have like really promising projects going on right but when they saw that like a month went by and the payment support still wasn't there two months went by three months went by it's like okay seemed like open a kind of gave up on it so I think like they kind of lost a lot of the most promising people and even if they launch now I mean yeah maybe if their the program is really good the people will come back but like I feel like um they messed up by promising the payments and then not delivering on that exactly the payments is the key so if they have delivered on payments I think at least same like Apple Store if you develop some app on App Store then definitely you're get going to get paid if they are going to purchase that because a lot of people are there and you can try ads things like that but here you can't even push your own ads there so even if there's an option where you can publish your own ads oh if area where you can publish your own ads it's not option is not there so they're failing there yeah and then you cannot even like take it out so if you spend like 100 hours making a really great GPT with all these custom knowledge custom functions you know super like good prom engineering and then you cannot even like export it it's like you know yeah really weird definitely exactly so it's you do and then time waste I think mainly is because of even I created like five six plugins for chat gbt and one day they I mean I received an email saying like Okay now we are going to close it soon we are moving everything to GPT custom GPT so now I need to go and work it out to transfer it across to custom GPT which I thought that okay let me focus on something else which is more important you know what's crazy that actually if you switch from CH gbt plus to teams or if you switch from teams to back you lose access to all your gbds so they're like in the void and I think I was the only one that happened to this but I got confirmed from another business um owner that for her it happened the same so she had like a GPD and then she switched to cgbd teams and she lost access to it even though she was the developer so this is like a super big issue because like you even let's say you have GPD with like 100,000 chats and it's growing popularity but if you lose access to it you cannot edit it you cannot you know go back to it so this is like so did you actually upgrade your package to teams and then later downgrade it yeah yeah wow up until now have you got the access or still you're waiting for Access no I don't have the access I mean I I mean I haven't emailed them or anything but like um I mean sure I think they can probably give it back somehow but uh right now my gpts are I cannot edit them so they're like in the void and they don't have any owner any Builder any developer so yeah it's kind oh you missing payments then yeah probably I mean I don't know my most popular ones only like 4,000 chats probably not not anything substantial but you mentioned that people are being fooled by you know Advanced video like Sora but even the current generation of AI images we've seen like there's a pictures on Facebook that have like half a million likes from especially targeting like the older audience especially like older women like in their 50s 60s that people just believe it's usually like some you know these kids build a you know massive robot with pet bottles like plastic bottles or like some crazy stuff that seems possible but you know to anybody who's seen M Journey or Dyre images is like obviously AI but those who haven't it's like it's easy to be fooled right yeah definitely I think this is going to so it's it's always uh whenever the new technology comes there always two things one is to make it like in the right way and there are people who try to make it in the wrong way so I think we need to focus more on using it in the right way and also making sure the model is prepared and also when someone post it on the social media I think there should be a proper filtration or system in place so that it marks okay this is AI generated clearly so people know that okay this is a fake so that's one of the key thing which I believe the social media need to bring I absolutely agree I think it's actually the responsibility of the platforms themselves to figure it out because they have the most data they have the most resources so you know to expect like the users themselves to be able to identify everything that's kind of unrealistic right so if the platform has some um you know image model just analyzing every post it probably can over time especially with like human feedback right reinforcement learning it can tune itself to be pretty accurate maybe even more accurate than people who have generated 50,000 Maj images and just like can spot AI instantly I think honestly like this responsibility follows like on Twitter Facebook and like that yeah and also I think it also depends on the model generators for example chat GPT or mid Journey when they generate I think they should have a standard in place saying like Okay any model any image generator using mid Journey or di 3 or any other image generator they should have some kind of a mock or something that only the language models or only this system is able to track that I think Facebook and Twitter can integrate that kind of system on there to identify those things maybe there should be a standard in place so like almost like open AI collaborating with Facebook Twitter and all the platforms where they have some because they do already have Watermark all the images it's just that it's not really implemented in any way it's not really utilize right so you think it would be good if like open AI made the protocol or the encryption what whatever it is available to Facebook and only then they could see like okay this is and flag the post right this is AI generated do you think that would be good or do you think that's like too much um centralized control no I think that that would be good like there are many standards in place similarly they should have standard on this so even like a normal human being can't identify the marks because we can even uh for example if you upload a white paper in chat GPD which contains a letter written in white letters CH GPD can understand but a normal human can't read that similarly there should be a lot of approaches where these models and even these top companies can uh come together and create a standard so that even a normal person if they generate an image they can't identify okay I need to edit this to make sure that it's it's to prove that it's not a genate and that kind of thing if it comes then it'll be really good and do you think uh it's also possible to Watermark text or no yeah real text is like really complicated I think as of now there are system in place to find that but still people can use chat GPT and then use they can use Quil bolt quillbot to make sure that uh it's not even identified by any AI personally I would say it's probably well I mean obviously it's like possible for like the level one but like if anybody wants to you know make it unrecognizable they can do that probably very easily right it it's just like even probably without leaving CH gbt you can just say write it as a college student with bad grammar and it accidentally makes a few mistakes that seem plausible and then you probably pass most Checkers right I think yesterday I saw a video in regards to that where these uh IM videos got generated and then they were edited further and also it could be the reason like even from one dog there were multiple dogs coming one of the video which came through so they they also agree with that so the initial model is not as good without any errors so I believe they're still training on it so that's why they are delaying the release and do you think they'll wait like December until the election is over or do you think it will be like dropped in a few months I think it should be in the few months I don't I don't think they will wait till then if they wait many other open source model will try to compete that they will try to understand how they did and probably they will try to go more than that them which will be negative for them even I believe for chat gbt 5 if they don't release on time probably these open source model even the um Mark zucker's uh the next level of llama is going to beat gbd fire also soon they need release gbd fire otherwise it's going to be like they're going to be left behind we've already seen I think uh few days ago the model from Alibaba right was a yeah very promising and it seemed like was trained on actually soda data there was like some similarities there kind of crazy yeah yeah true true the working thing everything I think they try to remate the same kind of thing and saying that okay we are much better than open AI that kind of thing so what do you think like is the solution to not people not getting fooled right because we've talked about one possibility is the platforms adding like a warning this is AI generated or something like that but like do you do you think we need um just everybody to kind of know about this like the same way everybody knows about like phone calls being a scam or like at most people know do you think we need that kind of education yeah definitely I think uh that's a good point because even for phone or even like using mobile phones even for uh checking if a email is a spam email or real email uh if it's spoofing email or non-s spoofing so that kind of things still we've been trained in our workplace or even in our schools to know whether is a right one or wrong one similarly I think definitely there should be a training for everyone even from the company or schools or anyone to know that okay this exists so not to believe everything yeah I mean everybody knows like the Nigerian prince email by this point right so it almost need to be like I think maybe in a year or two years from now by default if something is crazy we'll just think oh yeah that's AI like it'll be like it needs to be super proofed there needs to be tons of evidence different video angles and stuff like that until we we believe something actually happened for real because of how good AI will be right yeah I think even it's scary to see sometimes it's scary to see how advanced it's going to to be because uh even at this initial stage it is so powerful imagine after 5 years every engineer is trying to work to make sure that it is more original rather than a duplicate one because that's their goal because that's exactly what they're working towards to make sure that no one can identify there's a duplicate it'll be interesting after five years to see and why do you think like music has been lagging behind I mean obviously we've seen a huge improvements with udio and the other one waso yeah so these two have been huge huge leaps but like if you look at images right they've been way further ahead from music and it seems like all all of the different media types music has been the last why do you think that is I think most of the musicians they are probably uh not into more programming because I think that's kind of much more deviated from a computer normal computer they just have instruments and they don't normally work but composers comparatively they work on computers but if image generation or image editors they need to work on computers and for any video creators they work on computers so anything in related to computer generally uh people have that kind of knowledge to involve into machine learning and TR that model but music is like slightly comparatively uh away from computer so I believe that could be one of the reason it's slightly delaying yeah and I also think like the yeah go yeah I myself play music so I know like sometime even I see that kind of Gap so whenever I see uh music model or music thing coming definitely I will go and try even uh recently I tried music Lang which is kind of I don't know if people understand the things what it does because previously when a music was generated it was one layered but now the current reason tools like music Lang things like that it creates multi-layers so trumpet is on a different layer you can change the trumpet if you want you can remove the trumpet if you want and every single instrument in every single layer so that is more advanced compared to how it was before like 6 months or one year yeah I mean that's definitely needed because like any producer you know if you give him a complete file that's just one file he's like okay what should I do with that I me sure they can cut it into samples and stuff like that and there are tools that can like isolate the vocals and remove the vocals but if you have the all the stem separated into like you know Melody base pick it's like all different layers I think that's where the real adaptation by the music industry might begin definitely definitely I think that's that's the dividing that into tracks and also identifying different tracks from the existing music is more advanced uh I would say that would take things to the next level and why do you think there's such a resistance by artist yeah I think that's because of the revenue share so if AI comes into place even any non- musicians like can create a music without any skills so they just need to say can you create a music uh which is similar to this and shouldn't exactly look like that then definitely the music is going to generated and also the large language model it is going to make sure that it's not duplicate or it doesn't look like that and definitely more AI artists rather than a real artist AI music artists are going to like come up in the near future who doesn't know music who doesn't know how to sing who doesn't know who haven't played music who is going to probably top the list in the music industry but that's like the same as like before photography was invented or popular you know you would say like oh we need sketch artists in everything right but now that we can just take a photo um like making sketches or like drawing some moment that's uh ancient right it's obsolete I mean obviously it still happens in courts where cameras are not allowed but like other than that it's not really happening so I think for me I would say uh it's all about the taste and if somebody doesn't know anything about music they probably don't have like good taste in terms of selecting so even if you can create 1,000 tracks with AI you don't know which one is a banger and which one is trash right so I think the current artists somebody who's like has that musical touch will still have a huge Advantage with AI tools because they can select like okay this would be huge on the radio and they can select that and can see that but while somebody who is completely new who just listens to you know garbage let's say they will not be able to select a masterpiece right yeah definitely I think the taste I think you got the correct Point The Taste person who has good taste in music and able to even if they could understand the technical behind it for example if the list of instruments being made at this time they can actually prompt okay I need this instrument at this place I need the notes to vary this area so if they know a little bit then they can take it to the next level similarly if an music artist use AI they can take this to the next level even I saw a few music artists they already created music with AI so even they actually took the voice of a dead person and they use that voice with permission from their family and they use that voice to S make them to sing a song so that's really powerful I mean that's like has so many amazing opportunities because you can make any singer Immortal basically yeah true yeah and I feel like if you don't have any experience and you hear AI generated song and you feel it's not quite there but it's almost there you are screwed because you don't know how to you know look at the tracks and which track needs adjusting you don't know how to mix it if somebody knows how to do that they can hear like something okay maybe the base is a bit too loud and they can fix that while if somebody is completely new to this and just you know has a few subscription on these AI music tools they'll not be able to identify what is wrong even if they can sport it right this is going to be issue for the next few years so next few years the musicians who are got the knowledge with AI will be to the next level but probably even that can be fixed probably in the near future probably after 5 years there will be tools okay can you identify the instrument which is creating more noise so they're like Okay this not so remove that that kind of thing so definitely it's going to change a lot in the near future I think actually one idea that I think definitely will happen is we will have ai DJs who are like real people but they use mostly AI music right so if you go to like most DJs at at their concerts they play maybe 20 30% of their songs and then popular stuff right I think it's going to be the same I think it's going to be still some human like you know nostalgic music hey that's made by humans but then uh it's going to be people who are experts with these AI tools that are going to make lots of bangers and then there's going to be concerts where all of the music basically with a few exceptions will be AI generated but because of the person the DJ will be the one who has that taste or that you know selection again it goes to the taste so if the B on The Taste DJ person can like modify things like based on how the crowd is responding they can actually play that for them even in real time maybe you know like real time generation yeah even real time even real time they can ask like to the crowd like what kind of music you want okay this kind kind of music so what kind of feel in which flavor so they can immediately on the live real time they can generate music and then make them dance I think you don't even have to ask you can just hear the response right because if if let's say a DJ has a set right and it's it's finished before the before he starts DJing it's basically impossible to start adjusting it I mean it's it's not impossible but it's hard right to adjust it in real time so if if you hear like okay maybe the the crowd is responding to melodic track but not one that's like more dub step or whatever then you can in real time adjust the rest of your playlist to be like more of that track that the crown is responding to well now doing that is like really hard in the future it be just be like oh wow they're really like this type of people all really like this type of song let's just uh generate in real time uh new music for them in 5 minutes and it's going to be like every concert will be like filled with original music it's kind of amazing yeah yeah I think in this case they can use multimodal for example they can hear the sound of the crowd what they're doing and also the vision how they responding so they can actually merge that together create a DJ based on what they see and what they hear so that'll be much more powerful and you know this might be like too intrusive but let's say if there was a camera and it was like doing facial recognition based on like the races or the nationalities you it would be like insane if it adjusted to like what's popular in that country right that would be like super nice L as long as it's not too much intrusion then that's fine okay so shifting topics a bit AMA just like me you used AMA in a lot of your videos what made you fall in love with AMA I think it's the simplicity so you install on your Mac or Windows it's there and if you are creating an application the first place to go to test any AI application is to activate the endpoint making sure have the model there and it's quick and easy download a model and run it is quick and then integrating that with any application is easy and you got quantized version everything there and you are able to upload even your own model to AMA which is also super easy so and it's like a Docker file you can customize based on your requirement which is also super easy that makes it my favorite yeah I mean like when Lama fre was released it was on all Lama within like few hours I think was kind of crazy so I mean even the mix right I think mix like a X 22b took a bit longer but the Llama 3 was super fast that was very very I mean like yeah even I now before I was using mystal on my local computer now I move that to llama 3 so by default I'm using llama 3 compared to what it was before and what was your initial reaction to seeing llama I seen the benchmarks and I was like wow this is really good and also it's from llama 2 already was popular at that time so when llama 3 came I definitely know that this is going to be like groundbreaking and then I saw all the benchmarks I tested it myself and then started creating videos helping people to install that on their own computer and integrating that on their own application and what was the first thing you've built with llama three the first thing is to run uh autogen and crew a that's the most uh two things which I tried yeah and how how good do you think the 405 billion model will be because we still only have the 70 billion I was like mentioning this before that uh if it comes sooner I think probably it could be even like competing really with gp5 uh when they when it comes so I think open a need to take a step quickly to releasee GPT 5 because if you see llama 3 with 70 billion it's it's still on the first page not very close but still it's in the first place a open source model and it's still in the first place with 70 billion parameter model imagine how it's going to be like 4. 45 billion so it's going to definitely outperform most of the model and uh yeah that's going to be a favorite for everyone so you you are confident that the 405 billion will be better than even the best gbt 4 Turbo yeah that's what I believe yeah cuz a lot of people think like there's this like asymptot that you I don't know cannot be broken but like I don't think that's happening I think he been improving ultimate goal for all those Engineers who are working in meta is like okay I need to make this better rather than having training it for like next one month and then keeping the same stage so I think even they should have their own benchmarks to hit so they will hit uh those benchmarks yeah the crazy thing is that they actually stopped training llama 3 70 billion even though it was still improving because they were like okay let's start working on Lama 4 it's like that's kind of crazy uh because now the AI hype is there and uh if meta delay things further if any of the company delayed things further definitely some other company is going to come and overtake them and uh that that will result in their own model coming down so ultimately in this AI race the company who is doing regular updates and regular release are going to top the list but I it still feels like all the companies are doing these like once a year once every two years like huge releases even like even llama 3 right llama 2 was released last year well it's still like has been a year and obviously GPT gpt3 was 2020 like 3.5 was 2022 and then gbt 4 was 2023 so it's still not like I I think the best would be like every quarter or maybe six months right but why do you think the companies are still doing these big like once year once every two years I think fun I think the pace is increasing for example mistl they are like hyper pace so you see mistal one day then you see M medium m large Mixel then next mix so recent like a new startup I think from France I think they're doing really good I think by regularly pushing a new model setting up a new Benchmark the more releases and more work you do in a short space of time because the hype is there so whenever some something new comes people are like okay that's really good let me go and try it I think the mistal kind of company they are using that kind of momentum and they are growing quickly so now open AI probably they might even they are also releasing but as a small small version like uh 3.5 3.5 3.5 turbo 4 4.5 turbo four four then similarly you saw Germany from Google Germany 1 gin 1.5 then Gemma then even the open source models even I think I see like uh they uh even Google they releasing models every once or two months so they I think they're also in Fast Pace now so in general you think the trend is going to speed up exactly because otherwise uh it's difficult to compete within this a race yeah because if you wait two years like by that point it might be just outdated right exactly yeah I think it's not going to be like before because uh they need to focus on releasing it quickly yeah okay so speaking of Agents how do you personally use AI agents like what saves you the most time I think one of the main use cases like AI research AI research and programming these are the two main areas which I currently work work on so for example if I need to like go go through like 10 different pages okay come summarize on that find the gaps in the research paper and able to improve that so for example I give a list of papers okay can you identify the knowledge gaps in on all these papers then it's going to give me all the knowledge gaps and then I'll be asking the same language model what are the areas where I can like improve on these gaps can you please go and search the internet then it use the search and uh try to fill those gaps by researching on the internet so I think in the area of research I feel I I personally I use more in regards to using a PDF or even like through URL search that kind of thing so that's one of my main and secondary is using programming so agents uh for example autogen if I give a task and I give like a description okay can you do this for me and it goes there it executes if any error it's try to fix it on itself so that kind of most two things which I try to use it yeah I mean I definitely think like I don't know I feel like we're so close because even with the research imagine if we had like 5 million contact window and you could keep all of the papers loaded in the context window and then like cross reference between them I feel like we're so close to it being like Ultra useful because right now it's useful but it's not super useful right I feel like we're like almost there right when it's like mindblowing yeah I think there are still there are new models coming which is with RNN with RNN there are new new models which trying to have no context length that kind of models are still uh they're building it so hopefully unlimited context will come soon I believe so at that time we can uh use unlimited context and aable to perform a lot or research going to the next level yeah because I think people don't realize that a big part of why llms or agents are not perfect or why they suck in some instances is because you don't give it enough context and people don't realize how much context we as people have like from the time we are babies so much information has been entered you know visually audio or like reading it's like so much uh training data basically and people don't realize that they have so many life experiences but like when they type a single prompt it's just one short sentence and they they expect the AI to be you know just as good at that super Niche specific task as you even though you have like 20 25 30 years of training exactly exactly that's true I think definitely I think the context is one of the key thing for any large language model and also the training data also very important the quality of the data also important and the way it got trained generally for a model is like full knowledge by default by having focus with context I think that's going to improve the performance of the model and also the application which we are trying to use so when did you first realize that AI agents are going to change the world mainly when the when chbd got released I was uh Amazed by the performance like because I was doing machine learning before that but when chbd got released I saw the power in text generation models and then the AI agents mainly when the autogen got released I saw the power of AI agents because when I started running those things even the task which will take a lot of time it actually reduce that further so at that time I thought like okay this agent Frameworks or anything regarding to agent is going to take everything to the next level and for those like who are new to this because you know like you've you've been in machine learning for multiple years most people are still finding about AI right now for those who are like new to AI agents and want to get into the field what would you recommend them to build I think to start with uh small things just basic knowing of what the LW language model law and how it's being behind each agent and just start with a new like a simple framework even praise and AI I mean it just uh go with any text and just write it down if they if they want to do it on a like a coding like basic text file and then creating agents running it just the basic enough and then if they want to take it further they can actually customize that probably they can choose crew Ai and autogen then taking it more advaned level like Lang chain and then if they want to go in the direction of developing those agent customizing those agents then they can go in that route but if they want to just use those agents they can find a user interface where you'll have all the agents create all agents and then interact with them it's amazing because the way you described it is like step by step you know as you improve your skills but people are not patient they're always want to jump into the last level and build something that's going to automate half of their workday when in fact they don't even know how to save a p python file right so like people just need to learn the fundamentals and learn them deeply as naal ravik says and then build on top of that right exactly I think knowing the fundamentals actually keeps you ahead of the game like if you buy a computer and if you don't know use it you can still use it like going to the browser and use it but if you know more about the computer it's like using into the full potential you have a big music instrument if you know one card or two card you can play only two cards but if you you know 10 cards then you'll be like uh really Rockstar that kind of thing yeah and the computer example is brilliant because like you can see somebody who's like not teex heavy just you know looking where the key are and like being confused and then you have somebody who's like completely mastered their operating system knows all keyboard shortcuts and those two people even if they have same intelligence and everything the person a will be like super slow and the person B can probably do five times more or 10 times more in the same time just because they know the operating system exactly they'll be more productive and they'll finish any task much more quicker yeah I think the issue with agents is that people are not willing to like be humbled they're not willing to admit that okay I have a lot to learn I'm just going to build agents uh that suck my first agent team is not going to be this breakthrough that's going to change my company or whatever you know like I don't know I think people need to be more patient basically and accept that the first agent team you built is not going to be it second one is still going to be bad and just focus on the skill itself and getting better at understanding how agents work how llms work how to build tools and like developing that understanding and when maybe GPT 5 comes or llama 4 you already have such a strong Foundation that you can just defeat everybody else who just like trying to scramble at the last possible moment exactly I think the same same approach say starting from one just try to same thing happened to me as well so when I started creating agents I was creating agents then it didn't work first time then it work didn't work second time then I understood that okay I need to change the prompt a bit to make it work then I changed the prompt then it worked then after that I tried open source log language model for the same code but it didn't work so I need to make sure that the prompting is in the level of the language model which we are using and gave more details to make it work so yeah it's going step by step and even if there's failures just learning from the failure and fixing those error will make you grow and learn about agents yeah I think this is going to be super helpful to people because you are a machine learning engineer so like people that have no background in AI or ml expect to get it on the right on the first try but hearing that even you with your machine learning background still didn't get it right on the first try I think that's going to be really like eye opening for people and hopefully make them a bit more patient the concept could be same but the way we interact with that application is different so in that way uh definitely there's a learning curve there and how did you overcome that learning curve was it by reading papers by watching tutorials reading documentation or simply by building more agent not only learning utilizing I mean whatever you learn just put it into practice that's the ultimate goal so that's one of the things reason why I have a YouTube channel 80% of the video is like practical things so I don't want to show things okay this is good this is good and then go I want to make sure that okay this is how you you need to run it then in that way I'm also learning myself like okay I'm running it so you can also run rather than learning myself you're I'm running it I was thinking okay why why I can't just create a video videos for that and that's how everything came forth yeah that's almost like the biggest hack of having a channel because the if you want the way you if you need to teach something you need to understand it even better so that's like force it's a great way to force yourself to learn something new because in order to teach it you have to actually understand it so I mean I would say that's a super under underestimated part of having a YouTube channel like this yeah definitely I think even for like the questions they asked like I mean I have tons of questions and I want to like try to respond because the pace the AI is going the time I read like I have 10 20 comments I like seeing 20 first comment second comment third comment by the time I read the third comment there'll like another related video a latest update or any tweet okay you got the latest advance or then I'll be like okay I need to learn on that so go through that kind of phase and trying to keep up with the pace uh going through and implementing and learning and implementing that is a key for anyone to like grow in that area yeah because a lot of people are like almost um using that as a productive procrastination where they're just learning learning learning but they never actually download crei autogen and try something build with it and like I always try to push people to like to actually take action because if they learn and take no action then they've learned nothing really they just wasted wasted their time yeah I think especially in regards to programming and and also agents learning only is not important because even if you learn if you if you know all the commands in the Python Programming and you have never implemented that's like learning 10% or 20% because only when you try the first program then you'll know the list of Errors then you try to fix it that is the main learning curve in regards to programming and I feel it's similar to running and creating any AA agents yeah and the people who are actually successful in life it's not the people who know all the syntax and all the documentation it's people who take the software and build something with it right take python or AI agents and make something useful that others want and nobody cares how they build it nobody cares how much they used Google how much they used to the help of C gbt stack Overflow it's just like if you made a great product and people want to use it that's the goal exactly I mean like you have knowledge and as long as you you use that knowledge that's where the power is just having the knowledge is not no use okay so let's give the people listening here a action step what would you say uh they should do in terms of AI agents first they need to learn how AI agents work and then start implementing those agents probably through framework or any other thing and then automate your work the things what you do on a daily basis try to automate that that make all your work more productive yeah because I like to say the best use case is your own use case right so it's easy to get like distracted by flashy demos on Twitter or whatever but like if you solve a task for yourself that takes 2 minutes every day or 5 minutes every other day that's already a huge win exactly it's a huge win for example even I was thinking of a concept where I mean it's it's already there like uh the open interpreter so you got the device there so each time I want to open my email I was like okay I need to go to Gmail open that and everything so imagine you have a button and you click and email is on and you click and these kind of small small things it's going to take off so if you want to open Discord just click and Discord is open that is that is taking to the next level yeah I mean I can't wait for that future it's going to be insane like we're we're going to look back in 2 3 years at our current productivity be like oh my God like these guys have got nothing done in our day like in the future we'll get in one hour done the same we get in a day right now exactly so I feel like okay one small slot same like uh the open interpreted device uh Gmail click open then uh respond to this person everything done that's it kind of yeah just respond to all emails based on my previous patterns and based on my goals and my slack messages and you just like maybe oversee the most important ones or like the one that is like confidence score is low you're just going to double check those and rest of it will be like automatic man definitely that'll be like amazing yeah okay I think we can wrap it up here thank you so much mvin for coming on the podcast I think it was a great conversation um and yeah if you want to plug your network praise and AI I'm I was happy having a conversation with you it was really nice so everybody if you're watching this definitely check out praise and AI it's super simple it's like low code uh so if you aren't a programmer it's probably like the best way to start I would say yeah definitely yeah try that and uh if anyone can support developing that they are more than welcome okay again thank you for coming on the show and have a nice day Che see

Channel: David Ondrej

Views: 17,382

Keywords: David Ondrej, david ondrej, AI, ChatGPT, artificial intelligence, ai, Artificial Intelligence, OpenAI, chatgpt, chat gpt, Chat GPT, Elon Musk, Sam Altman, sam altman, AGI, midjourney, david ondrej podcast, Emad Mostaque, GPT-4, GPT, AI revolution, new society, david ondrej new society, david ondrej community, praison ai, mervin praison, ai agents, ai agent, llama 3, crewai, crew ai, autogen, autogen studio, agency swarm, ollama, langgraph, langchain, meta ai, devin, devin ai agent

Id: CtOGupq0r7M

Length: 54min 9sec (3249 seconds)

Published: Wed May 01 2024