{
  "video_id": "pjZI2MuLWWw",
  "channel_slug": "dataindependent",
  "channel_handle": "dataindependent",
  "title": "$100M AI Investor Reveals What is Coming in 2025",
  "duration_seconds": 3243.0,
  "url": "https://www.youtube.com/watch?v=pjZI2MuLWWw",
  "upload_date": "",
  "transcript": "we are in the middle of a technical\nRevolution there was a time when people\nmade machines to make manufacturing\nclothes better and we're in that\nRevolution for knowledge work I felt\nlike there needed to be more investors\nwho just nerded out on Tech I think the\nbiggest unlock for me personally is it's\nreally simple I I use for like half of\nthe stuff and it buzzes and it says\nsomething something El want tweeted some\nbut then my phone just starts blowing up\nwith like notification after\nnotification I'm like what is going on\nwhen I do play with Claude I noticed\nlike this one small one weird trick that\nmakes it so much\n[Laughter]\n[Music]\nbetter hey crew what is going on today\nwe have DD Doss with us today very\nexcited to have him on board today we\nare going to be talking not only about\nhis Builder Hat but also his investor\nhat how he's evaluating companies but\nthe coolest part which I love and I'm\nexcited to share with you is he's\nactually going to screen share and he's\ngoing to show us how he builds with CLA\ncomputer use and mCP so let's jump into\nit so ADI you're a builder and you're an\ninvestor what are the cool people or\nwhat are the smart people in AI talking\nabout right now I think the one question\non everybody's mind is how far does this\ngo like scaling or just I whether I\nthink scaling is more of a tactical\nthing it's just how good can these\nmodels be come um can it do we get to a\nworld where this solves problems we\ncould never solve ourselves yeah uh I\nthink that's the frontier because when\npeople talk about AGI it's this is\namorphous thing I I love to tell exactly\nwhat blows my mind is that us as an\nindustry we don't have a solid\ndefinition of intelligence that we\nformally rely on it's wild that we're\nstill arguing about what is intelligence\nit's like a weird side product of this\nwhole uh kind of Revolution I was\ntelling a buddy yesterday was like if we\nasked somebody four years ago what AGI\nwas by all practical purposes we're\nthere we would probably we'd probably\nagree that you open up literally GPT 40\nstart chatting and you're like well yep\nwe're here yeah yeah okay so um how far\ndoes this go so you're it sounds like\nyou're talking about the application\nside like value providing right no well\nthere is the value side I'm less\nconcerned about that obviously as an\ninvestor I concerned myself about that\nas more of a technical nerd I just think\nabout how good can these models get like\nthey still have flaws today right like\nthere's still limits you have you have\nthe arc prize and\nare liit do\nit the AL zero alphao analogy will see a\nworld hopefully 2025 where these models\ngive it a general task it'll figure out\nthe rules it'll figure out what it needs\nto do it'll take some time it'll take as\nmuch time as it wants and it just gets\nsuperum level Intelligence on arbitrary\ntasks and in some ways depending on how\nfar you want to take that I think that\nthat is reason close to AGI yeah um it's\npretty wild so that sounds like that's\nyour Builder Hat definition of AGI which\nis pretty cool so while we're on this\ntopic I got to ask about your investor\nhat of AGI so how do you approach what\nthe next five years are going to look\nlike from an investor side I've only\nbeen investing for a year the the reason\nI wanted to get into investing after a\n10year career and and engineering and\nand and product is because I felt like\nthere needed to be more investors who\njust nerded out on Tech\num there's a lot of investors who are\nreally good at what they do analyzing\nbusinesses but I wanted to be that voice\nof like just show me what's cool and\nshow me what's possible uh Yohi comes to\nmind yeah like Builder like it's so cool\nto be like here's my pipeline here are\nthe tools I built for myself and here's\nhow it's actually providing value\nexactly I love that exactly so as an\ninvestor I also take like at least 50%\nis that view which is just show me\nwhat's possible yeah and we'll figure\nout a way to make it work we'll figure\nout a way to make money sure sure that's\nthat's much easier than groundbreaking\ntechnology um and then the other 50% is\nyeah how do you take what exists today\nand apply it and and useful and\nmeaningful yeah it's also nice cuz you\ncan call on a lot of the tech\nwhen you need to oh man you can just\nlike call them out like ask real\nquestions I was literally I don't want\nto name the technology maybe I can't\nname the technology uh I was talking to\na couple of uh other investors the other\nday and they were like you know I had\nthis pitch and these guys said you know\nthey do like rag but better and they\nthrew all these fancy terms at me like\nwhat do you think and I'm like look I\nfeel like that's every pitch almost\nevery every other pitch I've had for the\nlast three months and when it comes to\nthings especially that are hot like AI\nthere's going to be a lot of snake oil\nsales and it is pretty important to call\nBS on a lot of these things well so I'm\nI get a lot of pitches myself too and\none of the easiest ways to go through it\nis don't tell me what your Tech does\nshow me the value or show me the\ncustomer that's paying for it or like\njust show me something that's an output\nthat comes from it because it's hard to\nvalue like that's proof of work\nalmost right there true I would argue\nthat it is still it's it's the reason\ninvesting is especially hard with AI is\nyou can value like if I've seen\nso many startups 1 million AR because\nyou know cios and CTO are saying go buy\ngen tools does that mean they created\nany value yet the the revenue is there\nin the beginning year one but the r\nrenewals may not be there and even\nfurther than that dig down into the user\nmetrics the intention engagement and all\nin activation it's like well if that's\nnot even there either then you're going\nto get yourself in trouble okay um so I\nlove the question how far does this go\nwhat are the obvious gaps that are\nlimiting us going further from a um Tech\nuh AGI perspective there's a couple of\ndifferent ways to to break that down I\nthink let's talk just about text models\nyeah um I I look at text models today\nand I'm like okay well this is good for\na lot um I use Claude almost all the\ntime uh and what I and I played with all\nthe other models too like all like1 and\nand llamas and everything and and what I\nreally want to know is okay I I've seen\npeople have all had this comment is when\nyou use it for a very specific task like\na task that has a correct outcome an\nobjective task it goes pretty far and it\ncan amaze people but it doesn't go 100%\nof the way and I I love to call this\nlike the full self-driving problem where\nit's kind of easy to get a demo of full\nself-driving to kind of work in most\nsituations it's really really hard to\nget an actual full self-driving vehicle\nthat can do everything in real roads\neverywhere um and let's just take uh say\ncoding for example that's one example of\na very objective thing you can also take\nmath you can also take any other form of\nlogical reasoning biology which combines\nmemory and reasoning um and I think\nabout those things and I and I wonder\nlike okay we're clearly at a gap people\nlove cursor but there's also some people\nwho are like okay cursor is awesome\nexcept I hit a wall I can't do Beyond a\ncertain thing and I don't know what what\nwent wrong imagine how um how entitled\nor like selfish those people must be to\nthink that oh it's this AI is not good\nenough to do my job for me you know what\nI mean it's just funny that they're\ncritical of AI when it's like literally\nso helpful for so many things but no\nbecause it's not throwing you a\ntouchdown at the very end of the end\nzone that you're giving it a hard time\noh yeah where well humans are Infamous I\nthink Louis CK of all people has this\ngreat joke about how people complain in\nflights and they're like guys flying in\nthe middle of the sky this is the most\nincredible thing that could ever happen\nlike have some have some you know be\nthankful be grateful and so I think\nabout that too obviously but you know we\nalso want to think about the frontier\nand what's possible and even when I've\nbeen coding with cursor and things like\nthis there does come a time where I'm\nlike damn it did a lot for me but now I\nkind of Wonder there's clearly something\nit can't do no matter how much I prompt\nit and it goes like five cycles of it's\njust not working and then I'm in this\nweird spot where I don't even know where\nto look generated like 10 files 500\nlines each like if I was writing it\nmyself I would know how tog this needs\nto be like a vocabulary word for a\nshared feeling of I just had the AI\ngenerate a ton of stuff and now there's\na bug and I have no idea where it is\nexactly I don't know what that word is\nbut I I hear people talk about that all\nthe time we need to come up with a word\nfor it\num I have a hypothesis that GPT 4 can do\nalmost any task as long as it's\nsufficiently scoped down absolutely and\nwe'll maybe get to this later on but I\nlove this like that that anecdote\nbecause I play with computer use all the\ntime and um I wish I could do it with\nleg gb4 and all the o1 stuff as well but\nI just don't like they don't have the\ncomputer Ed version yet but when I do\nplay with Claude I noticed like this one\nsmall like one weird trick that makes it\nso much better\nand that really it was a really simple\ntrick and it was exactly what you said\nwhich is I just went to CLA first and\nsaid here's my high L three promps you\nClaude have to now give instructions to\na computer to break this down into tasks\nand then I feed that huge prompt into\ncomputer use and say now you do it huge\nImprovement one of my um uh when I work\nwith clients and we talk about prompting\none of the most common problems I see is\nthat they stuff too much freaking stuff\ninside the same prompt they want five\nasks I wanted to make a mini little tool\nand it sounds like we have the same\nthought it's like feed them a prompt and\noutput what are the individual prompts\nyou should actually be getting from your\nbig one just scope it down even further\num because then that makes evals a lot\neasier too then you know you know where\nit's messing up does it does and they\ncall it meta prompting and I think it it\nkind of alludes to this like fundamental\nCS thing which is everything is solved\nby one additional layer of abstraction\nand in this case it's just on you to\nbuild that layer of abstraction and then\nand then you know give a prompt that\ngives you more prompts and then do it\nthat's exactly it on the investor side\nyou're a lot of startup pitches come\nthrough what are the trends in the past\nlet's call it three months not like the\nbeginning stop not the 2023 but what's\nthe last three months look like from\ninvestment side well so so a lot of\npeople are talking about agents right\nand agents are one of those other words\nthat people are like what does that\nreally mean when you say agent and um\nand so when when it comes to AI\ncompanies let me let me sort of break\nbreak down the field there's\nfoundational companies foundational\ncompanies can be tough cuz sometimes\nyou're really asking the question are\nyou really trying to compete with people\nwith thousand and what are some examples\nof these uh well a good example is\nactually this company dickart um they\nbold brilliant founder it's just one of\nthose guys and as an investor now I've\nlearned this like just trust your\ninstinct you meet this guy uh Dean I\nthink his name is just whiz like you're\nlike oh my God like this guy's so smart\nit's almost like you're back in you're\nback in the person go get the IDE yeah\nyou're almost always actually back in\nthe person um but some people just have\nthis energy where they're like this guy\nis a genius and he was one of those\ndecart taking a problem way out of their\nleague which is this idea of generative\ngaming so you know it's hard enough to\ngenerate good images it's hard to do it\nreal time and then consistently so what\nthe question to ask is you have a game\nwhere people which involves thousands\nand thousands of game designers\nengineering different parts of the\ngraphics how things behave in that game\nwhat if you could just throw gen at it\nand just have each pixel just show up\nand travel in the game um he has no\nbusiness doing this get that he's like\nkind of a kid and with a small team and\nthese are things that like Google can't\ndo and open now Google has a thing that\ndoes this opening ey can't do anthropic\ncan't do but that was a beautiful one\nthat's foundational and that's great um\nyou have other companies there's a\ncompany called Cartia Cartia is a\ncompany doing a foundational model for\naudio they came definitely after 11 Labs\num and you know they had a different\napproach to it and so good reasonable\noutcome Labs is crushing it right now\ntoo they're freaking shipping and\nbuilding right now a ton of stuff yeah\nokay so foundational companies all right\nwhat are the other Trends you're seeing\nhere too then you have the companies\nthat are very clearly on the application\nside that are doing workflow automation\nI like to call it but that's a really\nboring word the way I would think about\nit is there we are in the middle of a\ntechnical Revolution there was a time\nwhen people made machines to make like\nmanufacturing clothes better and other\nthings better we're in that Revolution\nfor knowledge work and the question is\nwhat machines can we make and and we're\ncalling those machines agents but this\nis everything and and people love to\nlike take the extreme side and go like\nwell well my job is not automatable\nbecause it's so smart and my but like\nthere's clearly jobs on the other side\nwhich are dumb and boring and no one\nwants to do them and a lot of them\nreally involve look at this Google sheet\nlook at this Excel look at this PDF get\nthis data write this data here send this\ndata to this person look at a response\nthis person sends you then do this\nthat's most jobs um moving stuff around\nand if you have a job that's regular\nenough you can kind of automate that\ntoday so a lot of people doing\nautomations there and that can mean\nvoice stuff um that can mean pure you\nknow text and we talking about vertical\nAI or you talking about like the Lindy\nand the zappers like the horizontal in\nthis case I think most of the ones in\nthis category are are pretty vertical\nbecause at least at first because you\nhave to start with a way to say hey\ncompany I'm know you have a 100 people\ndoing this job we think you don't need\nthat and we can do it for you that's\nkind of the the high level pitch it\nfeels like there's an Arbitrage\nopportunity right now for just going and\nselling AI plus your industry like\npeople will get over that eventually and\nthat's going to get diluted and already\nis starting to get diluted but if you go\nto a law firm and you say hey I'm get an\nAI lawyer they're like great I need one\nof those without even knowing what the\nheck it is yeah uh there is the that\nthere is this idea of or this uh\nphenomena of people buy tools that over\npromised and then under deliver totally\ntotally and it doesn't work a lot of the\ntimes but on the other side you know\nI've seen Insurance Health Care fi\nFinance there are um a lot of pretty\ntedious Industries a lot of Industry\nthat have back offices especially where\nuh you it's just labor that you don't\nwant to pay for basically and so you're\nlike automate that away and let's move\nforward as basically to be Gros as a\nspecies and not have to do dumb work we\ncan move on we can move on to other dumb\nwork yeah want other dumb work every\ngeneration has dumb work one of my\nfavorite Consulting books I for um I\nforget what it's called but basically\none of the main lessons is there's\nalways a number one problem once number\none is solved well number two gets a\npromotion to number one right so there's\nalways going to be dumb work there's\nalways going to be interesting work it's\njust it's a big long cycle that comes\nfrom there true but that you know to\nyour point like this is this is there\nthere's a whole thing of workflow\nautomation like how do I automate\npeople's jobs assist or automate\nwhatever people's jobs and make it a\nlittle bit easier to do um\nthe third sort of category I will touch\non is it's very hard because there's so\nthere big long tale of ai ai startups\nbut especially you asked the last three\nmonths what I'm seeing more of um I'm\nseeing a lot more companies really talk\nabout this idea\nof let me let me I I'll take two yeah\ntotally one is the idea of what they\ncall AI sres is a common bucket of\nthings that people want to solve and\nI'll I'll get to what that is and then\nthe second one is a is is also in a some\nsemi- related space it's just\nprogramming agents and and that's not\nlike that's a three-month thing it's\nbeen around for a while but in the last\nthree months the the evolution has been\npretty interesting to see well and\nthere's a cursor competitor coming\naround now too I mean I love competition\nso I'm glad to see it yeah it's uh well\nchis is doing really well who's the\ncompetitor you uh windsurf yeah correct\num so it's it's awesome to see it'll be\ninteresting to see how this all out okay\nso two categories so first one sres Sr\nyou know tons of companies doing various\nversions of this um and I say Sr but\nit's going also like support is kind of\nincluded in that Technical and\nnon-technical support the idea\nbeing look I have a problem in the\ncompany code goes breaks in production\nthings go wrong that is money on the\nline especially from I'm a big company\nlike every you can do the math every\nhour can be like tens of millions of\ndollars and it's kind of crazy that\nright now the best thing we have is we\nsend a guy like that's the the the\nresolution is a guy goes and looks at it\nand tries to like figure this issue out\nand it I've done that at at glean and at\nGoogle it is one of the most\nnerve-wracking things you can do as a\nsoftware engineer which is like\nsomething's down in production there is\ntime on the line and there are literally\npedabytes of data you can look into like\nwhich dashboard do I look into what do I\ncheck what do I check next what do I\nquery for how do I fix it when does the\nrelease go out like your brain's going\nlonger and this could be 3:00 a.m. in\nthe in the night ping and it's like oh\nboy it's crazy uhoh it's crazy so it's a\nreally bad hard thing to do and there's\nways to make um make this easier and and\nthere's like a couple of companies that\nare that are doing this pretty well\nthere's there's resolve um there's\ncompany traversal there a bunch bunch of\nthese other companies trying to do this\num super interesting sort of sort of uh\ncategory of problems and just to me uh\nand and then the last one like one I was\ntalking about is just coding agents nice\nyeah um there just so much more stuff we\ncan build there just so much more and\nit's so cool it's a it's a it's a place\nthat's ripe too because everyone talks\nabout you know code is um easily\nverifiable does it work or does it not\nright the hardest part is still just\nwith the planning piece and do you get\nthe intention right and can the human\nadequately Express their intent to the\ncoding agent in the first place exactly\num it's amazing one of the side thoughts\nI've had is I bet the outside of just AI\nin general um like epistemology as a\nstudy like the study of thought and the\nstudy of how knowledge Works must get\nsuch Tailwind from AI the AI move that's\nhappening so people are thinking so much\nmore about how do I express language how\ndo I formulate these thoughts how do I\nthink about planning how do I think\nabout communication um all that wouldn't\nhave happened unless we had these Tech\nspas llms that we now have a financial\nincentive to really dig into how we're\ndealing with these things absolutely I\nmean I think it's incredible where we\nare in the world um so let's move on to\nthe Builder side yeah so what\ntechnologies are you personally building\nwith right now that you're you're\nexcited about right now I think the\nbiggest unlock for me personally is it's\nreally simple it I I use cloud computer\nuse for like half of the stuff no joke\nand I think there's with the caveat okay\nthere's two caveats actually there's a\nversion of cloud computer use that came\nout they wanted to do it safely\nanthropic is all about safety and as\nthey should be like I don't think this\nis something that you should just let\neveryone willy-nilly go and and run a\nmock in their computer um but I like to\ntake some risk and so I Clon the version\nthat they have for Linux and you have to\ndo a couple of edits to that I go into\nall as much detail as you want and and\nI'll show it to you later if you want to\nH but you can make it work on your\nactual Mac and makes things a lot better\nthat's caveat number one second thing\nthat you can do is I realized over time\nthat as cool as this is It's Claud\ncomputer did this really cool thing\nwhere it can control your mouse and view\nyour screen and then take an action but\nas cool as that is it's kind of the most\nlame part in the sense this is something\nthat takes a human a second and now it\ntakes this 20 seconds and it's really\nreally frustrating when it does all this\ncool stuff and then it gets stuck trying\nto find a button and click it sure on\nyour computer so when I prompt it and I\ndo a couple of alterations I just say\ncan you not do that like just try to\nstick to the command line you can pretty\nmuch run anything on the command line U\npretty much just stick to that and I\nfelt like the progress has been\nincredible that's cool that's very cool\nso is cloud computer use is that going\nto be the product or is that going to be\nthe platform that other people are going\nto build really cool products on top of\nit's unclear like now they have mCP\nwhich is I like I don't I don't want to\nspeak for anthropic but I feel like mCP\nis the version of safe Cloud computer\nuse well so can you define what what is\nmCP could you define it and in layman's\nterms\ntoo the way I see it and this is\nprobably not doing Justice to everything\nit can do is mCP is a way for you to\nsafely say hey Claude you can use these\ntools um I can give you some memory you\ncan go use these tools and uh customize\nthe way you can you can do things with\nClaud but it's very Dev Dev Centric it's\nnot really you know consumer facing\nready there's a whole spec that comes\nwith it right like yeah yeah yeah so\nwell I don't know I don't know where\nanthropic goes with this but I do know\ncoming back to the startup thing it's a\nbig thing that startups are thinking\nabout um computer use how to use things\nlike computer use to automate more kinds\nof tasks like the first era of\nautomating tasks was all browser based\nand the general way it would worked is\nscrape the Dom everything on the HTML on\non a web page and then try to figure out\nvarious hacks on what I can click on\nyeah and you know like you could think\nabout that for two seconds and go like\nthat's a pretty shitty strategy because\nthat's not how humans work it would be\ninsane to read thousands and thousands\nof lines of like HTML code and then\nfigure out what to do on a page the\nwhole point of the page is you see\nsomething and you do something and\ncomputer use makes that a lot better\nbecause you have the vision angle but\nit's slow so but there is going to be a\ntime where the vision latency comes down\nand you can start doing actual actions\non pages that get to automating a whole\nother boatload of tasks both on web and\non your desktop so an exciting sort of\nhigh level technical thing that startups\ncan do now um one last question before\nwe dig into computers we'll pull\ncomputers out here so there's an open\nquestion on if the future of automation\nwhat's the interface for an AI type of\nthing with call it third party software\nwill it be more like Cloud computer use\nwhere you're actually interacting\nthrough the UI or are the AP is the API\nlayer going to get built out and that's\nwhere the automation is going to happen\nI think the jury is still out what\nwhat's your opinion on which way it's\ngoing to go there have been a couple of\nstartups I've seen\ntrying to sort\nof have various approaches to how that\nlayer should look so there's one called\ncompos and compos is a way of saying\nthere are a bunch of tools that are out\nthere for you to use I'm going to\ntoolify it for you so that you can go uh\nand and the llm can see a prompt and\nunderstand an API spec of say GitHub and\nthen figure out what to do so you can\nthink about it almost like documentation\ntechnical documentation for llms and\nthen I will teach an llm how to use the\ntools that it has I will keep it up to\ndate and you just feed it to the LM it\ncan use my tool that's one angle there's\nanother company called Anan and what\nAnan does is it's more they almost think\nof it as the Plaid for apis okay which\nis like the same way you do Bank\nauthorizations by a protocol with plaid\nthis is like trying to do to extract\nauthorization out to the user but let an\nagent act on the behalf on behalf of you\nand the funny part about plat is it used\nto be just via in and now it's like done\nproper and that's almost like the\nmetaphor for how this is probably going\nto go too maybe uh I I I still think\ntool use has not got mainstream adoption\nyet um so it's it's TVD where that\ndirection goes obviously a lot of\nstartups looking after it looking out\nfor it a lot of Enterprise companies\nthat do a good job there but end users\nstill kind of haven't done it most end\nusers are still just kind of talking to\nclaw talking to Chach BT and going like\nhey what the heck's going on right now\nright they're just waiting for more\nBuilders to go build stuff to go use it\nfrom there to me it's crazy maybe as\ngood aside there but you step out of the\nSilicon Valley bubble and nobody knows\nwhat's going on it's it's it's just it's\nalmost baffling sometimes because when\nyou spend so much time on AI Twitter and\nyou're literally in San Francisco in the\nheart of the city right now it's hard to\nget away from it you know you go to\nthese meetups and it's just like oh my\ngoodness I'm so far behind because\nthere's so many other people that are so\nfar ahead of me um but like you said you\nstep outside of it and it's almost like\nyou're speaking different languages when\nit comes to all it yeah yeah and and\nthere's like a soou cultural uh element\nto this too I don't know if you've been\nseeing this but like a lot of the Youth\nof of the world today like gen Z I don't\nknow what Jen and young Alpha yeah\nthere's like this hatred of AI inter in\na in a huge way where it's like it's not\ncool to talk about it to use it\ninteresting it's it's the most weird\nlike like I feel like an uncle now even\nsaying this but I'm like dude when I was\na kid all new tech stuff was cool it was\nreally um nerdy but cool uh what is what\nis that reaction and and maybe it came\nfrom the world of art where I know AI\nhad a really really negative reaction\nfrom all artist but now it's it's it's\nit's seeped into just AI at large people\nare just like uh yeah all right well\nwhatever they're they're lost they're L\num cool well I tell you what let's do\nsome building so DD one of the cool\nthings is that a lot of the projects you\nhave surround or revolve around diagrams\nand that's cool because it's not only\nvery pretty for humans to look at but it\nalso actually helps AI\nunderstand the flow of something a\nlittle bit better that language may not\nExpress in the first place so why do you\nlove diagrams so much I I think diagrams\nare just way more expressive it's it's\nbringing structure to unstructured data\nand I think the caveat is look it's not\ngoing to be perfect you're going to get\nthings wrong but just the fact that I'm\na huge book nerd and I love reading\nbooks and it was always this like thing\nin the nagging back of my mind where I'm\nlike I wish I had a character diagram\ncuz my memory is terrible like I I watch\nGame of Thrones and every episode I'm\nlike who's this who's that why do they\nknow them and I just I don't want to\nread a Wikipedia again I want to see the\nrelationships and that kind of inspired\nthat you know viral Harry Potter tweat\nback in the day where I was like let me\njust throw all of Harry Potter at um in\nthat case it was Gemini for the context\nwindow and just get a graph out of it a\nget a diagram out of it so maybe I can\nshow you that that's one and I'll\nanother one I'll talk about and we'll do\nit is um getting diagram from code he's\nthe literally biggest problem in the\nworld when you look a new code base is\nit's not the time you spend reading it\nit's the time you spend not doing\nanything cuz you're overwhelmed sure\nyou're looking at it you're like nah I\ndon't know that's cuz it's just a bunch\nof text files it's like it's where's the\ninsertion point where do I go start\nwhere do I start right yeah it's crazy\nif you think about it where do I start\nand so I we'll do a demo but what I did\nwas I like like dumped the entire\nrepository into a plain text file put it\ninto an llm and go like hey tell me tell\nme what the code looks like love it um\nso let's start with start with one uh we\nwill um what do you want to summarize\nwhat would be cool what's something\nrelevant um okay I have a good one I\nhave a good one let's let's start with\nthis can you you can see my screen right\nnice yes I can it's funny that you\nmentioned the the the G thing because I\nthink I might have that already\ndownloaded oh that would be nice cool\nlet's see what we can do here this is\nall live I have not done anything before\nhave not prepped for this and I think\nthat is going to be the beauty of it so\nwhere are we let's so you like ter like\nworking in the terminal almost yep I am\na purist I guess they would call me at\nsome point\num so I'm going to use I like to use Zed\nso sorry cursor guys but this is awesome\nso this is a VT file and what this means\nis I use this tool called YT DLP YT DLP\nnot only helps you download YouTube vide\nbut it helps you download transcripts so\nI'll show you a that's just the\ntranscript that comes straight from\nYouTube itself right yeah okay so we got\nit so what that did was I generated this\npreviously with Claude if you look at\nthis code it is basically trying to take\nout the timestamp artifacts which take a\nbunch of tokens uh and it gives me some\nplain text and this plain text is not\ngreat and there's some repetition but\nit's it'll serve our purpose it'll get\nyou there um it'll serve our purpose so\nlet let's see if this one\nfits well we're going to go we're going\nto put it all in oh great it fit and I'm\ngoing to\nsay draw a graph I know mermaid is the\ndefault but I don't like using Mermaid\nso draw a visually aesthetic graph this\nis what I just like to do there you go\nin as a DOT file why aren't you fan of\nmermaid out of curiosity I think mermaid\nis too simplistic and I'm just like more\nold school and mermaid comes from\nsomething called Dot and Dot or graphis\nis what I used to use in college and so\nI'm just more used to that there you go\num as a DOT\nfile for those that don't know DD was a\nhardcore engineer before investor I know\nwe actually had our first conversation\num when you were still at glean a long\ntime back yeah I I was an engineer I was\nan engineer um with with\nsubgraphs demonstrating their\nrelationship between various things\nthat are talked are spoken about in this\npodcast what about prompting techniques\ndo you do you like to think out loud do\nyou like tips or what what's your take\non those I wouldn't say like I'm a pro\nprompter like my my advice is don't\nthink do um just go try something out\njust go try something you'll figure out\nin the end like I a couple of these\nthings that I've done I've just figured\nout over time sure so here it generated\na bunch of stuff now what I do is I will\ngo and take this to online graph VIs\npaste it in and just one shot too that's\nnice yeah this is so cool so um we have\nsome personal background hearing\nimpairment social isolation reading\nFocus talks about rabbit holes pattern\nsynthesis um D and it gets it gets it\ngets you some somewhere and it looks\npretty nice this is just such a nice\nlike better way to look at a long body\nof text yeah it just makes more so much\nmore sense makes it so much easier to do\nand and we'll do a little bit more on\nthis because I think this is a question\nI get asked a lot which is how do I get\nout of this context window limitation a\nlittle bit so I like to do this is one\nof my techniques my prompting hacks\nassume all of these nodes and edges\nexist keep going add new ones and make\nit\nexhaustive and so you wanted to go you\nwant to double click here I just wanted\nto go deeper I think this graph was a\ngood start and let's see let's see where\nit can get me and\nso it does regenerate some of\nthe boiler plate here but it it's going\non and on and\non oh look we have some glp\ndrugs some China some great stuff great\nstuff so now what I'm going to do is I'm\nactually not going to copy the whole\nthing I'm going to copy just the\nsubgraph thing and let's see if it works\nso go back to this\nand oh look at that and it got even more\ncomplicated that's nice so that's cool\nso now it's like okay you're like if if\nyou're a hyper productivity guy you\ndon't have that much time you want to\nlisten to this podcast but you don't\nhave three hours yeah download it and\nnow you just know it'd be so nice to\nhave time stamps for each one of these\nand just click on it and then listen to\nit it's funny that you ask it's like\nlet's say I'm interested in Ai and\nintellig\ncan you time oh well it can't do the it\ndoesn't have the time stamps right it\nwould be really cool it would be really\ncool yeah but look in a world where we\ncould Fe fit the time stamps and I've\ndone that before too which is like just\ngive me quotes with time stamps for for\nthese areas that I like that's cool so\nthen other than just looking at the\ndiagram will you use that diagram schema\nanywhere else or like any other the\nproblems like does that make it its way\nto an input to a prompt for you I\nhaven't tried that yet yeah um one\nexample I talked to I talked to another\nuh person on this uh interview series\ncalled arjin and he wanted to judge how\nwell his support people were following\nthe call script that they were\ninstructed to do so we first started\nwith the with the diagram flow of the\ncall script and what it should go here\ngo here if this then go here blah blah\nblah you give that pair it up with a\ntranscript and then the AI will llm as a\njudge will judge how well he followed\nthat script I've seen some\nvoice pretty cool yeah okay so that's on\nthe diagram side um let's go into\ncomputer use let's do it yeah so we're\ngoing to uh and again this is my Fork of\nof of anthropic computer use totally\nunsafe do not use it for for it okay\nyeah do not use it it can do all sorts\nof weird stuff one thing that went wrong\nwhen I when I was trying to make this do\nsomething was they started a rogue\nprocess and just created like a million\nfiles in a directory which then stopped\nloading because the computer just went\nBonkers um so let's try to do something\nhere what I'm going to do first\nis we're going to go to Claude and we're\ngoing to\nsay instruct a computer to and you know\nthere are some people who like to save\nProms and they're like very meticulous\nI'm not that guy I kind\nof instruct a computer in as much a\ndetail do not write the code now if tell\nCLA to do that because it left to write\ncode um to do the\nfollowing create a directory a called\nDev slash we'll call it Greg\ndemo and in that\ndirectory create and no you tell me Greg\nlet's we can interact through we're\ngoing to create an app with a front end\nand a back end give me some ideas like\nwhat kind of stuff do you think and I\ncan push back I think it's but for a\nfront end and a back end um let's I mean\nlet's just do it on brand topical no\nit's kind of lame I was going to do like\nVC inbound blah BL\num how\nabout like simple restaurants in San\nFrancisco uh think about that one apis\napis will be tough because you have to\ndo oh you want to pull some a real data\ntoo we could\num I'm trying to think you know what\nI'll do one now get I don't know how\ncomplicated you want to go I want to\nscrape Open Table to know when\nreservations are opening oh interesting\ncuz there's this one spot in San\nFrancisco that every time I look it's\nalways booked and I'm never there early\nenough you know what um it's not going\nto be as visual but let's let's see how\nfar we can go all right let's see how we\ncan do that um in that\ndirectory create an app to scrape Open\nTable open\ntable and I have the\nrestaurant reservations in San\nFrancisco to find\num do you have a restaurant in mind I do\nI\nuh I just put it in the Riverside Studio\nchat it's called Four Kings San\nFrancisco how do you spell that I just\njust four like the the number four to\nfind Four Kings\navailability and\num let\nme you know I I'll I'll make it\ninteresting and\nlet's keep it simple for now tell\nme when it opens up nice\nand so let's see how this goes and it\nmight it might say something about okay\nso are you're making the prompt right\nnow that you're going to go then give to\nthe computer use yeah okay yeah man this\nis if it could I mean oh my goodness\nthis is a lot I don't even know this is\na lot this is a lot so yeah this is\ngoing to be fun cuz I have no idea if\nthis going to work we we'll see how far\nit\ngoes would you like me to proceed with\nthebody know I don't want you to proceed\nuh try to use this is another one try to\nuse command line only as much as\npossible and build an MVP yes boom so\nnow okay let me open up what's going on\nhere this is how I typically look at it\nokay it is trying\nto it's creating all these files it's\ntrying to it's it's trying to create a\nconfig it's trying to download all the\nrequirements for python this is where\nit's running at and then real quick let\nme just go into that demo but I think\nfor a lot of people if you do prompt it\nthe right way absolutely like I have a\nfriend one of the one of the people who\nwork at one of the companies that I I I\nnamed before let's let's pause there and\njust bring it up for a second you had a\ntweet that said everyone thinks this is\nan exaggeration but there are so many\nsoftware Engineers not just Fang who I\nknow that literally make two code\nchanges a month and they're getting paid\n2002 to $300,000 a year it went viral it\npopped off you listed the name of\ncompanies so who who's this friend at\nthe company there's many friends at many\ncompanies um Andre karpati uh tweeted\nback at that and I love like using other\npeople to validate that this is a real\nthing and he's like people brag about\nthis all the time there's entire\nsubreddits dedicated yeah overemployed\nis an entire sub dedicated to it so I\nthink like for me as somebody in the\nsoftware engineering community I don't\nthink this was like a denovo type\nthought everyone knew this happened I\nthink outside that Community maybe\npeople didn't know as much maybe the\nexecutives didn't know as much and now\nmaybe with this tweet some people do\nwithout question there are people who\ndon't do anything okay and is it their\nfault is it their fault no I mean\nthey're not sometimes well just it's\nwhere incentives are right incentives\nand the incentives go up the chain right\nthe managers aren't giving them work\nthey don't have anything to do because\nthe managers don't expect them to do\nanything managers at any big company\nnever have incentive to get rid of you\nyeah they always have I have had a\nfollow-up tweet about that they always\nwant to increase headcount because\nthat's the only thing that gets them\npromoted sure so no matter how bad you\ndo they have no they have no reason to\nget rid of you sure and look the nature\nof work is you're not going to have work\nfor five people I say you're a team of\nfive you're not going to have consistent\nwork for five people consistently uh the\nproblem happens is when you have\nbasically no work for no people and then\nyou're like uh I can't like I'm a\nmanager I can't tell my manager that I\ndon't have any work for these Five Guys\nso I have to then pretend that I have\nwork so that he like I keep my job and\nthen it's just this thing goes up the\nchain and then nobody does anything but\nto get back to it there was a followup I\nI'll mention in that thread where\nsomebody said something something what\ndoes this have to do with like does AI\nhave anything to do with it and the\nanswer was no like the people who are\ndoing these jobs have no idea what\ncomputer use or Claude orbt they're\npractically not even they don't care\nthey're in Mexico they're in Hawaii they\nare vacationing all the time so so crazy\nbut when you think about and I can give\none example of a couple of friends that\nwork in a job like this they were\ntelling me I'm like dude what when you\ndo work what do you do and he was like\nwe maintain the service it's a big Java\nservice and like I just had to do like\none dependency upgrade once a month\nthat's so crazy a that's ridiculous and\nB talk let's talk about that dependency\nupgrade yeah what how much time do you\nspend on it and he's like oh I spent\ndude it was hard on it spent like 10\nhours and I'm looking at him like okay\ncan you can you show it to me show me\nshow me what you're doing there a buddy\nof mine and we we met and he pulls up\nhis computer showing me a little bit I'm\nlike dude why are you doing this\nmanually that's insane you're like try\nto resolve Java depths and I know that\nthis is insane\nsomething yeah he's basically Googling a\nbunch of things to try to figure out how\nto use this dependency system that he\ndoes not know how to use I'm like dude\nyou can run this through like Claude and\nit will fix it for you yeah and and his\nresponse was well they don't let us use\nAI at work H and I'm like okay well\nthat's a pretty shitty response because\nlike you're the person who's going to\nhave to do it now yeah exactly not not\nme and so that's kind of the that's kind\nof where it is all right and so where we\nat now with this thing let's see um ah\nthere are this is a problem that happens\nwhere here's what happened it ran the\nscraper the scraper ran and it blocked\nthe thread okay it waited four minutes\nand then it timed out okay well there\nyou go that's the work in progress with\nit so we'll see whether and and there's\na log so there's a little bit of a log\nand it\nsaid nice at least we're at the right\nURL which is good loading reservation\nBri yeah oh look I mean they're probably\npretty good about scraping yeah yeah if\nany open table Engineers are listening\nto\nthis so it looked like the error was it\nwas a timeout and then it schedules to\ncheck every five minutes and then\ntherefore computer use timed out\nschedules to check every five minutes\nthere you go nice\nman but look it wrote a lot of\ncode it somehow L wrote this like\nnotification sending code which that's\ncrazy just send from an email to you\nit's like sometimes when I look at AI\ncode that's like this I just get\noverwhelmed it's like man I don't have\nany mental bearing or anchor points to\nto lock myself into this and it makes me\na little nervous sometimes but I tell\nyou what when you click run and it works\nit's like well all right it's true I I\nfeel a little bit better because I like\nsometimes understand what's going on\nI'll give you one example now that we're\nthat while we're at it um and I know I\nknow we're time but we I can go to to\n130 Flats don't worry about beautiful um\nhere I will do one of these and actually\nyou know what let me do a simpler one\nokay\num and while you're doing that I would\nlove to hear about what's the stack\nyou're using right now where do you like\nto deploy what's your DB what's all that\nstuff I keep it really simple my this\nwas true of when I did this at companies\nwhich is do what you know don't do what\nyou don't know sure um sometimes I I\nwant to learn something new I'll learn\nsomething new I know Google Cloud really\nwell so I do Google Cloud a that's one\none command after one of my command line\nand then I say go deploy it with Google\nCloud I ask computer use tell me what to\ndo yeah it just does it nice so gcp um\ngcp uh and then\nfor what else what what other let's see\nI mean gcp you you're using Zed you're\nusing a lot of cloud oh I use I use Zed\nI use cursor to um I I uh I use a lot of\nClaude and computer use sometimes when\nClaude doesn't work I will go to 01 and\nI'll try to like go back and forth and\nget different answers I use Gemini when\nI need long contact Gemini that um yeah\nthat's that's kind of my stack when it\ncomes to languages um there's another\nreally cool project that that I was\nworking on were in in Rust and so\nsometimes I'll use rust when I need to\nbe super fast um but I'm mostly doing\nstuff in Python and then if I do front\nends it'll be in react uh that's mostly\nit beautiful love it man really quickly\nI want to tell you about like the\ncoolest example of something I leared\nwhen usually it's one of the hard mental\nconstructs with AI is you know what it\ncan do so there's like this unknown\nunknown problem where I know what I can\ndo therefore I ask it to do things that\nI know that I can do I don't know what I\ncan't do therefore I can't ask it to do\nthat but there was this one example\nwhere I asked Claude again meta\nprompting to our point I'm like give me\nsome random ideas and I just get out\ntheir ideas and then I took some of\nthose ideas and one of those ideas was\nget it to make a song and in my head I'm\nlike oh that's stupid cuz obviously\ncomputers can't do that yet and but I'm\nlike okay well what would happen if I\nasked it to do that and this one blew my\nmind because I didn't know it was\npossible and I'll show you this\nexample that you could create waveforms\nfrom scratch just like with code like\nwith code\nand yeah look at that and he just built\nthat from absolute scratch using a thing\ncalled Socks which is a library I had\nnever heard of and I did not know and it\ndownloaded it and it used it and this is\na really simple one but I've gotten to\nget some really really great that's wild\nokay so say we were going to take the um\nOpen Table example we just did and it's\nlike okay it works it's cool it's there\nyou're about to package it up and you're\nabout to put on Twitter yeah you're very\ngood at this what is the mental process\nyou're going to think okay here's how I\npackage it up and everything up until\nyou click send or post okay so step\nnumber one is get Hub repo uh usually\nget computer us to take care of most of\nthat for you um so put it on GitHub\ncouple of General sanitary things which\nis get get some documentation going so\npeople like know what it's so you can\njust spit out all the code and say now\ncreate a read me for it um get a nice\nreadme whatever that's step one step two\nis I think about like videos and there's\ntwo ways to do this there's like the\nvideo of the process which has a certain\naudience and there's a video of the\noutput which has a certain other\naudience sometimes uh just a picture\nworks better than video it's pretty\nlaborious cuz what I have to do is Quick\nTime screen record I'm not super fancy\nwith it do a screen recording run the\nentire process back literally like load\nit into iMovie but sometimes I use some\nother video software and like clip it up\nsee like make it faster cut the ums cut\nthe pauses yeah and I I usually do no\naudio for this CU that's a whole other\nthing but um I just want to like show\nthe the demo because it's usually for\nTwitter not YouTube cut it up make it 4X\nfaster if it's a computer used thing and\nExport boom boom I use an app called\ntype fully to schedule tweets so I'll\nput the video on uh you never put a link\non your first tweet you put the link in\nthe thread so I link to my repo in a\nfollow-up tweet and then I just compose\nlike I'm like okay what do I and how do\nI want to talk about this so um one one\nexample of that was uh say I I built a\nthing on on on my friend's computer\nwhich is I always thought this would go\nbe a really nice concept for people to\nunderstand which is everyone talks in SF\nespecially is like\nmake me a bot that swipes on hinge and\nuh iPhone screen mirroring came out\ncomputer use like put two and two\ntogether like let's make this happen and\nso that was a cool one and so take the\nwhole video get computer used to do it\nand then package up the demo get rid of\npii Boom tweet and then go so you have\nan idea you go and build it you share it\nand then it seems like it goes viral\nevery other post in fact when you were\ndoing the preall for this one I said I\nwant to talk about your viral post and\nyou said which one and so it's like yeah\nyou have so much good content out there\nso how how do you build so much like\nwhat's the secret behind it um firstly I\njust want to say well I I'll tell you\nthe story of how I got into Twitter cuz\nI think I think people misconstrue a lot\nof the the online stuff and I hate like\nthe word influencer because that's not\nwhat I Tred to do at all um the reason\nwhy I'm able to Output so much is\nbecause I I think I just genuinely love\nbuilding stuff and in the past one of\nthe biggest constraints I think mentally\nnow that I introspect about it was\nwhat's the point of me building this no\none's going to see it and no one's going\nto care it's just going to be a little\nproject that sits on GitHub with no\nviews and I can tell one buddy of mine\nwho cares and nobody else gives a\nand now that I have an audience I'm so I\nI feel really like genuinely grateful\nthat it lets me go like I'm going to go\nbuild it because at least some people\nwill see it and um and and it kind of\nbrings me and that's kind of all I\nwanted to do and that's so cool uh the\nreason I got into writing on in the\nfirst place was because I mean the story\nis me and a friend were out at drinks um\nwe were having a great time and I gave\nhim a like like an idea and he looks at\nme and he goes like dud that's a\nterrible idea and I I'm arguing and I'm\nlike what are you talking about like no\nit's a great idea and he kind of does\nthis flip on me and he goes like look\nand this is before I really tweeted but\nI did write online a little bit and he's\nlike I trust your judgment if you sober\ncan go online and write about this\npublicly I'll believe it's a good idea\nand I'm like oh well that's a that's an\ninteresting thought right there it's\nliterally just framing the idea exactly\nand then I\nrealized that gap between you having an\nidea and thinking you're awesome versus\npressing the send button so big for so\nmany people so true and I realized that\nthat was stopping me from doing a lot um\nI was I I thought I had great ideas\ndidn't put them into produ or put out\nthere and the second thing is everyone\ndoes this on the internet we go into\nrabbit holes we learn things but you\ndon't remember much yeah and this was a\ngood way I found I promised myself I\nwould go online I would write one thing\nevery day ni One thing that's somewhat\nuseful to other people that I learned\nand that's how it all started and I just\npromised myself to do that and it gave\nme a reason to go and learn that's very\ncool instead of aimlessly exploring I\nwould say I learned a thing here's a\nthing and that's kind of how it evolved\nbut then obviously you know it goes from\nthere I was just listening to the daresh\ngorin interview and I think gorin said\nthat one of his Northstar metric for\nhimself is to maximize the number of\nrabbit holes he goes down so it's\nliterally just getting in Flow State\ninto an esoteric topic and just going\nand having fun with it um so when you\nsaid Rabbit Hole it made me think of\nthat um one thing I would love to do to\ninspire people to share more here so um\ngoing from the Journey of having no\naudience all the way up to hundreds of\nthousands of followers on Twitter I'm\nsure there's a lot of like cool\nserendipitous moments that happened\nalong the way maybe somebody famous\nreaches out or an opportunity prevents\nitself or maybe a new career thing what\nare some of the cool things that\nhappened along the way of your journey\nthat are due to having a bit more of\nexposure oh this is my favorite one\nwhich is I was on um the first date with\nthe the girl I've been seeing now for a\nwhile and I remember we were we were\ndriving in my car together and uh across\nacross the bay bridge uh going to\nOakland for msf and it's the first time\nwe ever met yeah and my phone starts\nbuzzing and it's like I see a\nnotification like I'm driving and the\nphone's up there the on the dash and it\nbuzzes and it says something something\nElon tweeted something and and but it\nwasn't the normal Elon tweeted something\nit was I couldn't tell what it was I\nwasn't really looking at my phone but\nthen my phone just starts blowing up\nit's like notification notification and\nI'm like what is going on and probably\nshould publicly admit this but then I\nput my car on self-driving and I'm like\nwhat's going on you're like babe got to\nlook at this no it's my phone's blowing\nup and I saw an Elon thing and then it\nblew up and I looked at it and it's like\nElon had retweeted and written something\nto me and in my mind I'm like oh my God\nwhat is going on and to her it's just\nlike meaningless she's like I don't know\nwhat you're so obsessed about but I have\nfriends reaching out to me going like\ncongratulations oh my goodness and um\nand and obviously like I'm a huge Fanboy\nof a lot of things Elon does and so that\nwas that was one moment where I'm like\nDam\nthe richest guy in the world knows I\nexist that's kind of like nice so those\nare the positive ones like yeah the Elon\nis I think we've had 10 replies and\nstuff now and Paul Graham Mark andri a\nbunch of these like pretty esteemed um\ntech people but I think it's worth like\nyou take the positives with the\nnegatives and I think the hardest part\nabout writing online is that there's\ngoing to be times people tear you apart\nand that's happened a ton of times too\nand what percentage of the energy when\nyou when when it ends up blowing up like\nthat what percentage of the energy is\nnegative that comes up it depends on the\npost and I think you know there's some\nthere's some threads where you're like\nokay 90% positive and there's a couple\nof people who are kind of dumb and\nthey're saying like nonsense and that's\nfine you you you build thick skin to\nthose things um the hardest part is when\nyou get a full-blown cancel um thing and\nI don't think there's anybody I've met\nwho has some scale online who hasn't had\nat least one of those um but when it\nhappens for the first time in a big way\nit messes with you man like as much as\nyou want to say I have thick skin you're\njust like you're like looking at people\non the street and going like do they\nknow do they know do they know that I'm\nlike I'm a villain online to millions of\npeople and uh and you get over it but\nthat's happened a couple of times too so\num and I will share that there's so many\nyounger people especially that I've met\nwho literally have had one of those\nhappen and have stopped posting online\nfor years wow it kills people\num energy and motivation what's the\nadvice to them for me it's like it's the\nsame advice in you know in real life\nwhen people say you care too much about\nwhat other people think of you they\ndon't really care it's kind of that it's\nlike people are going to say Yeah\nkeep doing you just keep going do it and\nthat's the other thing too is like the\nalgorithm wants to promote good stuff\nand if you post something and it's not\ngood don't worry nobody's going to see\nit they're all going to see the good\nstuff there's only one caveat which is\nlike take feedback so if there is\nsomething you're actually doing wrong go\nlike okay well what am I actually doing\nhere but in many cases that feedback is\nyou know just keep doing what you're\ndoing DD this is fabulous awesome thank\nyou very much for thank you so much",
  "transcript_chars": 55577,
  "ingested_at": "2026-05-15T04:40:12.189422+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 7122,
    "like_count": 249,
    "channel_id": "UCyR2Ct3pDOeZSRyZH5hPO-Q",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}