{
  "video_id": "AVl3k7u0X0U",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "Product Hunt's \"Maker Of The Year\" on how to build AI",
  "duration_seconds": 2664.0,
  "url": "https://www.youtube.com/watch?v=AVl3k7u0X0U",
  "upload_date": "",
  "transcript": "all the people that think this today I\nthink they are so wrong too like there's\nso many things to be left model that\njust got out a few weeks ago it's much\nbetter at creative thinking and it's 10\ntimes cheaper replicate is an incredible\nexample so from day Zero I want to save\ntime by use uh super high level apis it\nwas way better compared to what a human\ncan do to an alive summarize cluster and\ncategorize stuff it honestly Saved Me 1\nhour by creating like the bones of the\n[Laughter]\n[Music]\ncomponent meet TBO his first product had\n0 in sales his second product hardly\nmade any cash at all but on his 15th\nproduct TBO sold tweet Hunter and Tapo\nfor $8 million in 2022 he was awarded\nmaker of the Year by product hunt in\nthis interview TBO walks us through\nwhich AI tools he has bookmarked in his\nbrowser which techniques he's exploring\nfor character consistency in video\ngeneration and his thoughts about when\nto use 01 and Claud Sonic together or\nmaybe not together when generating blog\nposts let's get into AI showand tell\nwith tibo tibo I tell you what so I was\nlooking at your Twitter during my\nresearch here and in 2021 you had a very\ninteresting tweet and the Tweet said do\nyou want me to have GPT three analyze\nyour tweets for you and what blew my\nmind was this is 2021 so when did you\nstart with AI how would you walk me\nthrough that I I think that's with that\nI think it was like one of my first who\ngot some decent engagements that's night\nand I think it was the very very\nbeginning of my successful startup the\nthing is I think at the time I already\nhad a few uh project failures with gpt3\nbecause like I think like everyone else\nI was trying to generate stuff like I\nwas trying to use gpt3 to generate\ntweets generate copyright copy and at\nthat time with these tweets I realized\nthat jpg3 was super super good at\nanalyzing stuff like it was it was way\nbetter compared to what a human can do\nto analiz summarize cluster and\ncategorize stuff and so I I did this\ntweet where I would just use the AP the\nTwitter API behind the SC uh getting\nsome\ninformation like dumping everything into\na GB3 prompts and just just tell it like\nanalyze stuff tell me what's what's\ninteresting about this profile what's\nwhat's the main topic that's uh generate\nengagement for him and it worked and\njust I I was I remember that I was\nmanually uh like launching my function\nmy gpt3 function uh like copy pasting\nthe handle of the guy putting putting\nthe handle on my code launching the\nfunction uh and copy pasting the results\nback to Twitter as a reply yeah I did\nlike 20 or 30 and every time I had a\nlike I had a mistake in my Cod like I\nwas seeing something uh bad I was I was\nactually\nfixing the prompts on the go and just\ntrying to improve it it was so fun like\nI was I was seeing the likes going up\nand my prompts getting better and better\nit was it was so fun that's wild so your\nlikes go up the prompt and performance\ngoes up but then also probably your\nmotivation and your excitement goes up\nat the same time yeah exactly yeah yeah\nyeah and that was back when it was still\nmagic CU like 2021 that was pre-chat GPT\nso this stuff was still like relatively\nnew all over the place I feel like it\nwasn't a big crowd that even knew what\ngpt3 was at that point you know what\nlike when we joined I think I think we\nneeded like 6 month to get a gpt3 key\nlike an an open AI key because it was so\nrestricted at the time and when we\nfinally got it it was like everything\nwas done like all the cool use cases was\nuh buils and it was too late like we\nwere we missed the train yeah and it was\nso wrong and all the people that think\nthis today I think there are so wrong\ntoo like there's so many things to be\nleft well so let let's keep on going\ndown that path and so I love that\nmetaphor that that was your mindset back\nthen in\n2021 and you still kind of feel the same\nway today so like what's your advice for\npeople who say we're too late like we've\nbuilt everything we need to so far I\nwould like I think my the thing that\nworked for me and that I want to repeat\nat every new thing that I that I build\nis I want to build something for myself\nlike I I want I want to be\nmaybe not the main user but I want to be\nthe user of my own product because I\ndon't have the discipline or I don't\nhave the skills to I think to manage a\nbig team and to build the right\nprocesses so that the correct uh product\nmanager can learn from the client and\ncategorize the feedback in a good way so\nthat another team can build the thing I\ndon't have these skills uh what I know\nis that when I when I use a tool\nI know how to fix my problems I know how\nto take my uh daily workflow and to make\nit simpler thanks to some cod that I\nwill just damp somewhere in an app yeah\nyeah yeah absolutely um well I'd love to\nhear more about the tool stack that\nyou're building with and so as you're\nbuilding these tools for yourself um you\nhad a tweet that said or it it kind of\nshowed your p l so it showed with your\nexpenses and you were talking about deep\ngram open AI uh fall Luma seeve 11 labs\nall this where is your tool stack\nsettled from an AI perspective right\nnow I think I have a are you are you a\nArc user uh no I no I I do c like in Arc\nis my brother and in Arc I have like\nfolders like like it's a little bit like\na bookmark but uh it's it's better\norganized and have F could you show us\nyeah yeah sure that'd be really cool to\nsee um so this is like my all my uh cool\nAI API that I'm going to use like it's\nif if you combine the power of file uh\nthe power of like\nreplicat uh and and save too where where\nlike this one you have like you have\ntons of AI models that are ready to use\nand super\npowerful I think like most people they\nare trying to like I see tons of people\nthat are trying to get the server up up\nand running uh like set up a few gpus uh\nand scroll hacking face finding cool\nmodels to use and spend so much time\nsetting them up fixing crash thinking\nconfiguration issues and and I really\ndon't see the point of that like those\nproviders they're going to be more\nexpensive but they're going to save you\nso much much time like yeah replicate\nreplicate is an incredible example\nbecause they have they have so many\nmodels there are so many that you can't\nyou cannot even like explore them you\nhave to search for something because you\nhave so many other\nthings so from pretty much day Zero uh\nbecause I want to have a super lean team\nI've always wanted to like uh use like\nready to use things like I I want I want\nto save time by using\nsuper highle apis plag them\ntogether with code and and make a\nproduct out of that yeah and I when you\ndo when you are able to chain a few apis\nlike chain CH GPT to then generate an AI\nimage to to then put that in the video\nit's like so satisfying you know yeah\nyeah yeah yeah um well I tell so I tell\nyou what I would love to hear how do\nthese tools chain together so what do\nyou replicate for what do you use C for\nwhat do you use Luma for and like how\nhow does your stack look\nlike I would try to like I think I think\nfile is very good as so you have for\nexample like a I'm using flux a lot uh\nin in F I think it's like one of the\nbest provider for for Flex um I'm using\nI'm just generating images there uh then\nI'm going to\num like let's say\nuh Minimax this one or Runway they have\na few um image to video models and so\nyou are like that you are able to like\nuh have a prompt with gpt3 or sunet or\nanything else like that would generate\nthe image then you use flux to generate\nthe image uh with the prompts then you\ntake this image and you put it into a\nRunway and just like that you have\na three-step workflow where you're able\nto generate a video a short video uh\nthroughout like basically anything yeah\none thing that is pretty cool with file\nand I haven't explored that but they\nthey make it quite easy to do comy UI uh\nworkflows which then U let you build\nsuper specific and complex um\nAI workflows by just like chaining\nmodels between them like sure something\nlike one thing that I haven't nailed yet\nand I'm working a lot on that is I want\nto be able to like use this kind of\nstuff like train a flux Laura meaning\nthat I would be like training training\nan actual uh model from a few images of\nlike someone or something mhm uh and\nthen then use that model to then\ngenerate um generate a video of the same\nperson um with a strong consistency in\nthe video itself so like a a video is\nlike one or two minutes long right now\nit's incredibly hard to have character\nconsistency in a video because you have\nto generate many many images and and\nshort videos sure and well what what's\nthe secret to consistency for those\nthings I don't so there's there's one\nsecret that I'm like I'm just scatching\nthe surface of that but\num\nyou I I I I think there is like a a\nstrong like a huge value in uh using\nmask and in painting like basically you\nwill generate an image you will with a\ncharacter and you will just try to\ncreate a mask uh from about this\ncharacter you will extract this\ncharacter and generates the correct\nimage with by using this custom model uh\nwith a tech called in painting where you\njust you just to tell the the model to\njust generates the missing piece of of\nthe image does that make sense to you\nyeah for sure it's like let the model\nrun wild on everything in the background\nand the body and the sky and everything\nblah blah blah because that's not what\nyou want the consistency for but then\nmask and thenay maybe the face because\nthen that's what people really care\nabout and that's what you really need to\nget right is the face piece of it and\nand this step is super important if you\nwant uh multiple characters if you want\nthe the issue with um fine-tuned models\nwhen you're trying to generate images is\nthat if you are using if you combine\nmultiple fine tune\nmodels where each one has been trained\non a specific character MH the image\ngeneration model will just merge all the\ncharacters into one and so all the\ncharacters on the image will look the\nsame and will be a merg version of all\nthe fine tunes models sure which is\ninteresting Harbor yeah yeah yeah for\nsure it's interesting it's like you want\na model for every subject that you want\nconsistent and so it's almost like you\nhave the parent image and then you have\na tree of all the subjects in there that\nyou need to have remain consistent yeah\nand so that's that's some I I'm just I'm\njust I'm Blown Away by by the fact that\nthis has not been uh properly handled by\nlike this is a super important problem\nin my opinion and there's Big Value in\nfixing that well I tell you what I mean\nlook at where text was even just six\nmonths ago you couldn't do it and now\nit's like a non-issue so I think that\nthese things will just come around yeah\nsure so you had another tweet that was\nsuper interesting and you said that AI\nis eating your\nmargins so basically I that I read that\nas there it's super expensive paying a\nlot of money for it how do you manage\ncost or how do you track cost across all\nthese tools I have a big issue right now\nit's that's like um with my projects\nlike rev. a it's it's the first time\nthat I'm I'm uh working with videos and\nAI videos is the most expensive stuff\nthat I have ever seen if you like if you\nare if you're trying to use um if you're\ntrying to do like the the top level\nmodels like Runway Luma or cling they\ncost you between\nbetween 20 to 50 cents per 5sec video\nwow yeah and so when you have a user\nthat is trying to create like a a one\nminute video it's it's easily going to\ncost you like more than a dollar $2 do\nuh maybe three if you add some some like\nvoice generation on top of that and so\nthere is a huge disconnects between the\nprice of those API and the expectation\nfrom the user like the the users they\ndon't understand why it would cost that\nmuch because it's just one minute video\nand so they are constantly expecting us\nto lower the price or they're just\nscreaming at us because of the price of\nthe products but at the same time our AI\ncosts are so high like they are I\nhaven't like more than more than 60% of\nthe projects that's wild that's wild you\nknow one of my uh Favorite Things I\nDon't Love everything about Elon but one\nof my Frameworks I took from him was his\nidiot index which is the price of\nsomething versus the price of the raw\nmaterials to actually do it and then the\nDelta between the two says how much of\nlike you know buffer there is from that\num do you know how much it costs\nHardware providers to generate those\nvideos versus what they're actually\ncharging you like what are their margins\nI have no idea like this is not public\nit's it might be that they are not even\nprofitable like like CH gbt is not like\nopen AI is not profitable maybe because\nsure I think I think they want to like\nuh they want to like get the markets\nthey want to get the market share uh and\nso they\nare uh like lowering their price to\ntheir bare minimum they can it's it's\nlikely that they are not even making\nprofit out of that sure well it's like\nthe rumor about uh grock having\nunprofitable tokens as well the the um\nnot the xai one but the uh the hardware\nprovider yeah yeah I see like it's it's\nit's possible the generation is so fast\nlike it's it's likely that's it's the\ncase yeah yeah yeah yeah for sure for\nthis project that I talked about like\nrit. the issue is like\nit's there is something that you uh\nlearned at business school and uh for\nlike for 10 years I thought that it was\nnot relevant to me uh but um seems like\nit is like the thing is our AI cost are\nmore than half of what we are getting\nevery month and uh we are growing so\nfast like 50 60 60% month over month wow\nand and our Revenue we we get the payout\nof our Revenue one month later compared\nto what when we are getting it yeah all\nthat some deps makes us very hard to\nsustain the growth because we are paying\nright now the expenses of this month but\nwe are getting right now the revenue of\nuh the month before and so since the\ngrowth is so hard we are just barely\ngetting enough money from the sales to\ncover the expenses of this month because\nof this\ndelay yeah that's wild um it's funny how\nthere's no better motivator to learn\nbusiness topics than having too many\nexpenses and too much money to try to\ndeal with it that's that's a good\nproblem to have I guess but still\nsurprising yeah yeah yeah absolutely you\nknow one thing that you were describing\nas you were coming through you were\ntalking about fall with to UI then you\ncan do complex workflows for that was\nthe mask example that you gave an\nexample of one of those complex\nworkflows or what's complex to you that\nyou would be you know coding up in there\nyes it is like this this process where\nuh like uh meta just released some an\nawesome model called called like Sam 2\nthe object detection one yeah exactly it\nit's it's able to detect um the very\nspecific uh a of an object on an image\nand even a video and it's able to label\nit and so using that plus uh some in\npainting Tech uh and some other Advanced\nuh labeling technology like all that\ntogether yes using confy UI is is a nice\nuh thing to use I'm not unfortunately\nbecause I'm not comfortable with that\nyet mhm um I'm too I guess I'm too\nScrappy to to use it so I need to like\nsee my code and being able to like fully\ncontrol it but I'm pretty sure I would\nbe able to do great things with this\nyeah well I tell you what you you strike\nme as somebody who when he wants to\nlearn something he has a pretty good way\na determinist like a very determined way\nof going into it so say you're going to\napproach learning compi what's your\npersonal learning plan to go do\nso I would try to be like the very\nminimum thing like the just just I I\nwould try to like get the first ID that\nI have and start working on it right\naway and try to ship something uh to the\npublic as soon as possible and I would I\nwould treat about that even if it's like\neven if it's shitty and even if it's not\nnot that useful I would just tweet about\nit and expect a few of my friends to\ntest it and give me\nfeedback so two years ago I was living\nin body for like 6 months nice and at\nthe time I was working on tweet Hunter\none of like my most um successful\nbusiness MH and what was super cool is\nthat I thanks thanks to a small group of\nhackers that I met there I was able to\nlike all the week work on tweet Hunter\nmy main projects but one day per week on\nThursday I would meet the group of\nhacker and during one day I would try to\nbuild something totally different and it\nit can be related to tweet like many\ntimes I build like some mini tool that\nwas like different codebase uh different\nsetup different uh server compared to\nthe main projects but could serve the\ngrowth of the M products but just being\na f thing aiming at going viral and and\nI found this I found this to be quite\nhealthy like to have like four days per\nweek dedicated to my main thing and one\nday per week where I would try to bre\nand ship something in just one day yeah\nand I I did like many times and I think\nmy most viral products my my most viral\ntool I bre it that way like during one\none day Sprints uh and that's it that's\nso cool it just what stands out to me\nabout that is when Sam Alman says that\nthe entrepreneur's way is to basically\nhave a lot of shots on goal because you\njust need to be right for one of them\nand so having one day a week where you\ncan I I think the important part about\nthat is is two things one you have the\nexecution to actually do it and to\nactually commit to it which is really\nnice and the consistency but two is you\npackage up the entire thing that you can\nship in one day so if you can't do it in\none day then it's it's really tough to\ndo I think I think you're so right\nthat's when we started like I think in\nin 2021 like a few weeks after the\ninitial tweet that you\nmentioned we we we were in this process\nof building one project per week for and\nwe did that for like 4 month during four\nmonth we try to like ship one product\nper week because nine of nine product\nout of 10 fails and it was like the 10th\none who eventually got successful yeah\nand that's like I think that's how you\nbuild something truly valuable like by\nby creating 10 things without knowing\nthe one that will be\nvaluable if if you're up for that I I\nwant to talk about this one thing\nbecause because I I kind of uh changed\nmy mind oh yeah that's great so let's\ntalk about 01 what what are your\nthoughts on it so the thing is I'm I'm\nworking for like a few weeks on this new\nSEO projects uh where basically um we\nhave an interesting approach where we\ntrying to find the super relevant\nkeywords to to rank for for your\nbusiness uh it's not launched yet but\nI'm onboarding people right now in a\nbeta phase and it's it's working very\nwell and at the very beginning of the\nproject what unlocked the project was\nthe re of o1 because on this project we\nwere we willing to spend a lot of money\non prompts and LMS and uh to to generate\nhigh quality blog posts and 0an was\nquite good at that like it it's uh we\nthought that it would be the the the the\npiece of tag that would unlock the high\nquality\ncontents and the thing is\nit's very good at math it's very good at\nlogic uh it's surprisingly good at uh\ncomplex problem\nsolving uh but the thing is the new\nrelease of like the new\nSonet um model that just got out a few\nweeks ago MH it's much better at\ncreative\nthinking and it's I think it's 10 times\ncheaper and so\nwe reverted to this one\nand and and yeah we just by by prompting\nby getting more uh in depth about what\nwe want uh by providing some example\nlike I think I think this is super\nimportant people don't do this enough\nbut by by providing good context and\nexample about what we wanted uh I think\nwe are able to work with\nSonet um and to get better result than\n01 that's specific to our workflow by\nthe way but I think creative thinking is\na very very interesting way where 01 is\nnot very good for sure so I would\nimagine when you're doing uh AI generate\nblog post there's many steps in the\npipeline from candidate ideas pick the\nidea pick the title write the actual\nthing come up with the outline write the\nactual paper etc etc like there's a lot\nthat goes on to it will you use son it\nfor every single one of those steps will\nyou oneshot it how how will you break it\ndown like what what's your thought\nprocess with that\nnot not entirely like there is one thing\nwhere when you're trying to generate\nvery long stuff like very accurate and\nin-depth\ncontent uh Sonet can sometimes try to\nlimit the length of the generation the\none model that I found that is not\ntrying to do that is Gemini J one gini\n1.5 Pro is very good at generating long\ncontent\ncool beautiful so will you have Gemini\nactually do the writing for you or will\nit write the outline for you or how's\nthat work I cannot go into that okay\nthat's fine no I understand tot\nunderstand um okay so that's the 01 side\nbeautiful um the maybe the one thing\nthat's uh I think it's interesting and\nuh would be cool to show up is um I\nwould I would highly\nencourage I I will share my screen yes\nfabulous I love this the I I think think\nlike you said that you were working on\nmany projects and I am too and I think\nthat's the norm like I think I think\nright now uh most people are working on\nlike many many projects because they\nhave unfinished stuff um they have the\nnew stuff and one thing that I do which\nI think is quite useful is I'm using\ntyping mind it's from it's basically\nlike another chat GPT uh\nUI but it's very convenient for me to to\nmake it um to work on multiple project\nbecause I have those folders and those\nfolders they are that like my um all my\nproject that I'm running right now and\nif we take this one for example I would\nbe able to like set a specific set of\ninstruction for these specific projects\nH and so then every time that I will run\na new\nprompt uh on on on this project it will\nhave this like uh this context about\nthis project I will not have to put\neverything back here mhm I think this is\na very simple Stu like it's it's not\ncrazy workflow that you will spend weeks\nto build but sure just spend spend the\ntime to work on these Mega context\nprompts for projects do it for each of\nyour projects and and then you have\nsomething super specific super useful\nwhere you can ask anything\nanytime yeah that's very cool did you\ndidn't build that project did you no I\ndidn't I would I would love to if I did\nbut I'm not it's it's a guy called Tony\nDean okay and he's he's pretty big on\nTwitter that's very cool um that'd be\ncool if after each chat session that you\nhave with it it's like hey here's the\nnew stuff that I extracted that's not in\nyour system prompt already do you want\nme to add it to it and then it just go\nbe amazing yeah that would be amazing I\nI I think tgbt is trying to do that but\nyeah it's it's so it's way too b2c to do\nit in a in a good way for us sure it's\nalmost it B to see yeah you almost need\nyou want the power user use case is\nexactly yeah yeah yeah yeah that's the\nside that you're on for it that's cool\num nice well do you think that the Rex\ndemo is interesting like would that be\ncool to walk through is that kind of\nsmall no I don't think so but okay\nthat's fine basically basically I'm\nusing\nCur do do you know about cursor oh yeah\nbig time I'm a I'm a huge cursor user\nand basically every time I have to every\ntime I have to use a crown I I triple\ncheck my my crown uh with with cursor\nlike I just I have these short where I\njust press uh common t uh by selecting\nthe the text the the reject or the the\ncrown and and just tell me if it's good\nor not something that I do a lot too\nafter testing I select my function or my\ncode and I I do that common T again and\nI would ask it if is there a mistake\nhere like is it is there an age case\nthat I didn't think about and that's it\nand and just doing that allow me to go\nto be to be way lighter on the on the\ntesting like on the manual testing and\npush the production very fasts H\ninteresting and um I was just speaking\nwith silly Omar who says that when he\nuses cursor and he writes with it he\nactually has it write the test first and\nthen it's a lot easier to actually write\nthe function to execute against that\ntest because you have a validated test\nthat comes with it yeah that's amazing\nthis this is this is a super interesting\nwork for the thing is I've never I've\nnever written any test in my lives and I\ndon't really I don't really plan in\nstarting but when you do when you Ed to\ndo that I think that's very powerful\nyeah yeah yeah that's cool that's\nbeautiful um I want to jump back over to\nthe prompt side real quick and so you\nsaid you're using a lot of Sonet and I\nknow that Sonet is um works better with\nXML so what sort of special formatting\ndo you put into your prompts with\nregards to headers or XML or anything\nlike that I'm doing Json every time Json\nokay I'm basically\nI think I think the the sentence that I\ntype the most is output in jasm do not\noutput anything\nelse so it's basically structured\noutputs only that come out from it yeah\nand it's super reliable like I I don't\nreally understand why people say that\nit's not but it is like it's very\nreliable in in like being following\nthese format that I'm going to I'm\nsharing it to you yeah do you have a\nvalidator on the other end or do you\njust like accept the output no matter\nwhat and see what hopefully comes out\nall right I have a custom made function\nwhere I validate the just output and I\nretry if it failed to do so but it it\nwas super useful two years ago when uh I\nwas using an like previous version of\ngpt3 and gbt 4 but right now it's not\nthat much useful like honestly it's it's\nvery R cool that's beautiful back in\n2022 you said you're never going to\nwrite you're never going to write\nanother commit message again my question\nare you writing commit messages still or\nno yeah because I'm working back with\npeople at the time when I wrote that I\nthink I was I was using that so I was I\nthink I was I was being a douche but let\nme tell you why when I W when I when I\nwrote that I was working with with a\nteam of people that I hired and so\nI um I made them uh agree to me using AI\nin my comic message meaning that every\nsingle one of my comit message it was it\nwas basically me pressing common enter\nand when I was pressing common enter it\nwould Auto commit and auto push with an\nAI generated um message uh and right now\nI'm working with people who are uh not\nnot not employees they are partner and\nthey really didn't like\nit and so I had to change and and try to\nlike uh try to be nice with them and\nyeah write my own comage message yeah\nyeah yeah for sure well that leans into\none of the philosophies I saw that I\nreally resonate that you had mentioned\nwhich was generated by AI Polished by\nhumans and really I think that that\nsimple sentence alludes to a whole lot\nmore about where AI fits or should fit\ninto folks workflow where you have it do\nthe first 95% and then humans are on the\nother end doing the validation to make\nsure it's good um how does that saying\nfit into your\nworkflow that's that's how I cut right\nnow that that's that's how I I program\nevery day and it's I think it's it's the\nbeauty of cursor is when I want to\ncreate a new feature I would just uh\ntell it what to do what I want to do how\nI want it to work and it will create the\nfirst draft of the components um and if\nif you set it up correctly if you have\nthese Master prompts if you if the\nproject is somewhat well structured it\nwill create decent things and and like I\njust I just did it like a few hours ago\nand it honestly saved me one hour by\ncreating like the like the bones of the\ncomponent and I really didn't have to\nchange like a few lines sure and I then\nshipped it in production it worked very\nwell and then just ship it there are you\nusing vzer or anything like that for\nfront end components I'm not but I think\nI should yeah yeah I should it's it's\nsuper powerful and from what I'm seeing\non Twitter it's getting better and\nbetter every day so yeah I I should\nreally do that well what's amazing is\nhaving seen vzer from the beginning they\nweren't investing it that much I think\nit blew up it started going nuts and now\nthey're throwing like a real team and\nreal money behind it which is really\ncool to see yeah yeah and that that's\nall right like it's it's super valuable\nyeah yeah I I should I should give it a\nmore like a a fair try I guess yeah yeah\nyeah yeah for sure um we just need to\nget cursor to import vzer and have it\njust right there actually like yeah I I\nhave this so I'm not using it but I have\nin my project\nMH so here I'm in uh cursor and if I\nwant to like uh create a new comp\nusing zero I would reference my master\nprompts V zero oh cool\num is that prompt public Have you shared\nthat yes I think I I got it from Twitter\nso it should be nice yeah uh multi\nselect for\ncontri poof and and by doing\nthat I think it should be able to like\nyeah it it creates a link and this\nlink it's it's a v link with the prompt\nembedded in the query oh oh maybe it's\nnot working anymore wow either way I was\ngoing to say I did not know that and I\nlike that a lot yeah but maybe like it\nit seems like it's not work anymore\nso maybe they they remov the support for\nuh the the query like the interesting\nwell call to Builders out there this is\nneeded we want\nthis somebody please go do it just a few\nmore questions here as we start to wrap\nup what um that typewrite or the the the\nthe chat gbt alternative that you showed\nme that was really interesting what\nother Mini tools like that are you\ngrabbing for that's a good question not\nnot that much actually uh I'm I'm using\nso I'm using uh typing mind the tool\nthat I showed you earlier I'm using a\nglip to like U to handle customer\nsupport to handle tickets it it's\nnot that much deep into AI but it has\njust the minimum stuff that uh is useful\nto me and I want to start experimenting\nwith uh side GPT side GPT yes it's it's\nthe it's a completely new uh s that is\nsupposed to totally like remove the need\nfor a support team H so that's by by\nlike by training it with the the correct\num documents helps Center it should be\nable to\nfully um fully answer to your users with\nthese unique capabilities that it it can\ntake action like if if a user want to\nconsel it will not only show the user\nhow to cancel but it will actually be\nable to cancel the subscription for the\nuser and the way it it it's doing that\nis pretty simple like it's it's quite\neasy to set up and so I really want to\ndive into that and see if it can just\nhelp me save time that's cool and this\nis site gp. right\nyes cool well it's it's quite it's quite\nold but the maker completely revamps the\nsoftware and it's like it's the new\nversion is completely new and mind\nblowing that's very cool um tell me\nabout what else is on your to learn\nlist uh let me check I have\noh I love this this is so cool it's very\nlong it's like it's a mix\nof uh things to explore uh can you show\nus too no because it's it's it's like\npersonal so I I've seen that uh seeve\nlike uh the the AI provider that we\nmentioned earlier it has some new cool\nmodels and I want to explore them cool I\ntry to it's it's it's a very cool\nprovider and uh it has some things that\nI'm already using but things like eye\ncontact correction for example it's it's\nquite powerful and this is just a way to\nlike when like you take this video\nwhere the you see the eyes sing around\nand it's to correct the video and to\nmake the eye very static and always\nlooking at the camera\nyeah this is the kind of mod that it it\nmight not change the game for me mhm but\nit's it's still quite useful to improve\nthe quality of some generated video for\nravit\nyeah they have a lots that I haven't\nexplored yet and I really want to dig\nthat yeah that's beautiful you had\nanother tweet that caught my eye in July\nof 24 you said I still haven't found a\nsuitable use case where AI agents could\nhelp me for real have you found use\ncases where agents help you for real no\nand I so maybe maybe it can work now\nwith so I haven't I haven't tried with a\nnew Sonet but um there was a few times\nwhere I buil an entire agent\ninfrastructure meaning a set of\nactions like this this kind of infinite\nLoop where an agent would try to achieve\na goal by splitting this goal into\nmultiple uh like smaller tasks and the\nthing is um because because those models\nthat we have right now they are like 95%\naccurate if you if you like if you take\nthe those 5% errors that you have and\nyou multiply them by the number of steps\nthat the agent will take to achieve the\ntask uhuh it's it it makes them\nunsustainable able like it it doesn't\nwork because like most of the time one\nerror in the chain will mess up with the\nentire\nthing and I understand that it's super\ninteresting but I just I just found that\nright now I was not able to plug the\ncorrect context and set of actions to\ngenerate value but for sure it's\nhappening and probably in 2025 we'll\nhave very interesting stuff related to\nagent like\nSam Sam samman himself said that going\nto be like the the the time for agent is\ncoming so yeah it's something to watch\ndefinitely yeah that's beautiful well I\ntell you what on that one um one\nquestion I'd like to ask at the end is\nyou're on Twitter a lot you see a lot of\nAI talk you see a you see a lot of smart\npeople talking about AI what are the\ntopics that smart people in AI are\ntalking about right\nnow I I've seen a lot of things about\ngaming it's it was surprising because\nthe I I kind of think that there are\ntons of people working on text for like\nfor like four years right now uh pretty\nmuch the same for images you have like\nvery incredible models with two images\num videos is kind of worked like we have\nwe have things like we have very good\nproviders it's expensive and probably\nit's going to get descripted but we\nalready know some of the main players\nand they are out there and gaming it it\nit was like it was pretty silence for\nlike\nfor since the AI Revolution started we\nhaven't seen any good good games and\ngood game engines and being distrupted\nand maybe maybe it's going to come uh in\n2025 like um I've seen those demos about\nlike Counter Strike and Minecraft\nbeing run by a a neuronet and not game\nengine and I found it super impressive\nlike it's honestly it's blowing my mind\nbecause thinking that from just a prompt\nyou could create an entire game without\nall the game engine that would support\nit it's it's such a huge uh change in\nand the thing is people playing video\ngames they already have gpus so it could\nbe very easy to like to to them if if it\nworks you can just get the games on\nquite powerful PC from Gamers and they\ncould just run it without many things\nthat makes game complicated Tot so again\nwe could see an entire industry being\ndisrupted uh tons of people becoming\nlike quite\nuseless and huge companies like uh unre\nengine uh or Unity 3D be being\ncompletely\nuseless yeah uh you know I love in life\nwhen you have kind of like a frame\nbreaker moment where you think one way\nbut then you hear something it's like\ncompletely different it completely shift\nyour frame and when I thought of games I\nthought of sure you're going to walk\ninto a virtual bar and the per the\nbartender is going to have like a chat\nbot behind it and you're going to have\nlike an engaging conversation just some\nlike a chat bot basically in a game but\nthen when I saw those Minecraft demos I\nthought holy cow I haven't even thought\nabout this beforehand exact exactly just\nthe neural net generating the game so\nit's like whole it's like this is\ndifferent wait a minute there's\nsomething interesting going on here yeah\nwe are so we're so like we were thinking\nabout how text based edms would descript\nvideo games and it's really not\nhappening that way yeah yeah yeah yeah\nit's always fun in life when that\nhappens well I tell you what I want to\nend on that one tibo thank you very much\nfor the conversation today thank you\nGreg was awesome",
  "transcript_chars": 35860,
  "ingested_at": "2026-05-15T04:40:32.190204+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 4359,
    "like_count": 128,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}