Open-source vs. "ClosedAI," demoing new AI tools & more with Sunny Madra & Vinny Lingham | E1742

0:00it should have a button that would be like the jay cal button which would be like troll this person and it would be the troll options and like my troll option would be like why do you why does hamas why does everybody hate hamas and then link to the clip of sasha bar and cohen talking about hummus and hamas and he's he's like uh trolling them trolling options should be there to do like jokes and stuff like that that would be a great startup is uh an agent that follows you around the internet and then optimizes your trolling and comedy and just makes a, you know, a mockery of everything. Mockery AI.

0:37Somebody go grab that domain name. This Week in Startups is brought to you by Miro helps take ideas from in your head to out there in the world with its ability to democratize collaboration and input. Sign up for free at miro.com slash startups. Lemon.io. Need to speed up your product development without draining your budget? Hire vetted engineers from Europe at Lemon.io. Go to Lemon.io slash twist to get 15 % off for the first four weeks. And iConnections is a platform to connect and meet with elite capital allocators through their online platform and bespoke events. The first 25 VC funds to sign up for iConnections Miami 2024 event in January of next year will receive a 20 % discount.

1:25head to iconnections.io slash twist to sign up today all right everybody welcome back to this week in startups is once again it's our ai round table call it this week in ai if you want uh every week we're going to do this until this ai thing slows down which basically means until the singularity agi happens and humans are retired with us again sandeep madra we call him sunny he's a co-founder of definitive intelligence that lets users view on and off chain data but let's be honest he's pivoting this whole that pivoting this whole crypto on chain analysis data thing to ai right you're all in you're full board we are all in on ai yes yeah so enough with the chains and the tokens and the nfts it's enough they're important what'd you say they're important they're going to be what's going to the source of truth is going to be powered by that so yes so it'll be a feature like of the wider new intranet uh but this is really ai is the new web 3.0 i think that's going to be how i look back on it vinnie lingam's here any credit for that yeah yeah i'm giving vinnie credit here you go i'm finding i'm founder of weight room one-on-one video conferencing for your corporation your enterprise including a bunch of ai stuff how you doing vinnie hey jakel good to be on all right let's get started i'm in new york city i just got i got my coffee i had a little sticker shock eight dollars for a cold brew new york's wilding out here i walked over i'm at the salt conference just got back from uae and i'm speaking at the salt conference so my world wind tour continues did you get an hiss a bagel you know i went to sullivan street bakery i'm over here on 48th and 9th by on the uh hell's kitchen where the irish you know my irish brothers and sisters we we ran this part of town for a little while uh but i just went to sullivan street bakery and like a big argument broke out there the tension is very high in new york right now yeah so i just step in it's a little everybody's on a little bit of edge usually new york is a little chippy but it does you know usually dissipates pretty quick but there was a on the line there was a little bit of a brouhaha woman asked for a piece of bread the woman who was taking orders her first day she didn't speak english perfectly they got into it a little bit and then the manager said listen i gotta protect her when we came back i didn't really appreciate the way he talked to me well well you know i gotta protect her she was like protect me from what protect her from what it's like a white woman black woman yeah manager comes out and then i was like well i gotta get in the middle of this and i was like listen you got to apologize in a more sincere way and i can understand why you would take that the wrong way that didn't come out well but listen let's all get a coffee maybe we can get a couple pastry going on here complimentary whatever and i had to dial the whole thing i brought the whole tension down i had to explain this woman was the manager of solidry bakery that when the words come out wrong i don't think she was being racist exactly but it kind of came out a little bit racist so i had to step in and be like hey listen you know and her apology was very weak very weak apology okay and sincere so there's your story everybody a little story time and then i come around the corner and there's a girl in a bikini and another girl dancing in the middle of 9th avenue and i'm like this place has gone crazy since i left and then i look down they've got their camera mounted on the lamppost they're doing a tiktok they're doing a tiktok dance in the middle of 9th avenue so that's it those are my two stories in one block of getting my coffee uh yeah it's new york and springtime it's just people are losing their mind the snow has started to melt and it's you know like in tahoe the snow melts and we get on our mountain bikes in here people just go crazy um but you you guys been to the uae have you have you been to the region at all i've been to abu dhabi and uh i've been to dubai yeah that's where i went yeah pretty wild huh yeah you know it's um it's it's very hot even this time of year it's pretty hot it's like you know it's 100 degrees yeah it's like going to vegas yeah yeah it's uh yeah speaking you know i don't know i prefer vegas well i mean but if you lived in that region it is kind of like the it's a combination of vegas new york and washington dc or new york abu dhabi is very impressive give us a 90 second summary the startup vibe the investor vibe the lp vibe give it four seasons was like rose uh you know like the rosewood you know i came downstairs and 90 seconds four people from slogan valley stopped me to say hi you know like who were in the region doing meetings or had moved their companies there and tons of startups i came out of a dinner and there was like a table of 10 startups and so startups are going there for funding startups are opening offices there and during covid they remained open so it's kind of like what was it sweden that remained open in europe or i think of sweden the state opened yeah sweden yep did it yeah and so dubai and abu dhabi became like sweden they they had rules but they were like we're gonna stay open um and everybody goes to their office and it's incredibly international there's only 500 000 nationals but there's 10 million people and they built everything in the last 20 years so it reminded me you guys ever go to like shanghai or china or shenzhen in the like but in like the 2005 2010 period when they were building there were sunny cranes everywhere and they have like a cleveland clinic nyu the louvre i mean and not like little tiny outposts like rivaling the rival but like um rivaling the actual ones like giant outposts so they are they're they're for real and they have a 20-year window as explained to me multiple times we've got 20 30 years to convert oil wealth you know the geographic lottery that they hit into something else and something else is going to be finance and tech and they don't want to just be lps they see themselves as becoming vcs so they're looking to build bridges with the venture community not just to lp them that's great they want to be lps and funds but they also want to be uh doing directs and so they they mubadala i met with and a bunch of other firms but it wasn't a fundraiser but i just went with brad gerster our friend of the pod to just uh hang out and learn about the region and then i wound up getting 10 meetings you know 48 hours before i landed people wanted to meet with me so i did a bunch of meetings and you know it's uh it's a very progressive place it's kind of like new york in the 90s i would say equivalent you know socially and how are the builders so interestingly people from europe from asia singapore india which is a hop skip to uae they're putting their companies in dubai uh because of tax because of 10-year golden visas so like if you want to get a visa you know how it's hard it is around the world to get it across borders they're just like yeah come here and then they get all kinds of tax breaks for your company so if you were going to put your company somewhere if you go to abu dhabi they're going to support the heck out of you and they have like their own we work and they have this place hub 71 so there were a lot of startups there these were legit startups that had gone to yc you know international startups but instead of basing themselves in india singapore china australia london they're just in dubai and they can get anywhere and anybody can get there right and they get good um you know canada gives um uh what do they call it when they give you subsidies for your developers where you do it tax creditsrd credits so i think they're they're they're getting in on that a little bit and so good place if you are not in the united states to to land your company i think um and the tax treatment's sick like you don't pay any tax base taxes yeah probably darby's investing heavily in in crypto and blockchain they've actually um they've they've funded this blockchain called venom i think they put a billion dollars into the foundation i guess just years of building bucks going seed the ecosystem and going it's crazy it's like there's so much money that's going into blockchain and it's exactly what you're saying jay jay carl they're trying they're trying to buy the tech future and get you know knowing that oil is limited i mean california we're going to be off oil in 10 12 years yeah we're the tip of the spear yeah yeah so like and not just us sweden i think scandinavia a couple other countries like you know get normal gas vehicles in about a decade from now and so the writing's on the wall long term for i think for for oil uh if we move to electric and if that you know that that's a 30-year story 40-year story in california and then there are countries that haven't even started yet so i think that's how they look at it um and so it's pretty impressive that they have this kind of 30-year vision to to sort of convert it uh and some of these sovereign wealth funds you know they have a couple hundred billion dollars And so for a fund manager like myself, if I wind up going to Harvard or, let me, let me take that again, Nick, I don't want to mention Harvard.

10:37For a fund manager like myself, who is a new fund manager, you know, I've only had funds for whatever, eight, nine years before that was angel investing. A lot of the endowments here in the United States are full up. They've built relationships with the top firms. The top firms come back with new product, new funds. They don't add new names. and so even with my track record and you know my notoriety or brand it is hard for me to break into one of those endowments that are now over these endowments are over committed to venture so if you have five percent in venture and you were some endowment and then all the public market equities came down the denominator the size of your endowment went down so that and if it went down a third you know now all of a sudden your venture position went up because the venture position doesn't get marked to market yeah but the public equities do so now they're you know 10 or 8 venture and like oh we got to get down to five again that's our mandate how do you do that you've got to either wait for the marks to come down on the venture portfolios where the fund managers are lowering their marks or you're gonna have to invest more money in public markets i mean just basically a lot of them are taking a pause on venture so i think that's a very unique moment in time it's like a poker game where oh this week four seats are open those four seats like it could go to you know the singapore sovereign wealth fund uaes saudi whoever they're going to have a unique moment in time while the while the large endowments here are taking a pause well you have a unique chance too with the story you know around building around what ai represents and how that's more aligned to what you've been doing for a long time right and so early stage is very attractive right now people are like the late stage is kind of messy because the cap tables are so screwed up and they're overpriced in many cases still so that whole late stage game was too many players too high price is not enough traction and so early stage you're starting fresh and look at all the activity we're seeing founders always ask me and my team to punch up their pitch decks right you got to have a great pitch deck if you're a startup That's obvious.

12:47It's table stakes. Great news for you. My team just worked with a team at Miro, that awesome whiteboarding software I've been telling you about. And we created an amazing pitch deck template for founders. And you can see it if you're watching video right now, or just head to Miro.com slash Miroverse and search for pitch deck, M-I-R-O.com slash M-I-R-O-V-E-R-S-E. And then just search for pitch deck. You'll see it there. This is basically going to help you take your pitch from zero to hero. Be VC ready. Go check it out right now. If you're a hybrid or you're a fully remote team, Miro is incredibly useful for you.

13:24It's like an old school in-person whiteboarding session, but it's distributed. It's asynchronous and it lets your brainstorm ideas, collaborate on projects and more from anywhere in the world. So when you think of Miro, think zero to one, but even faster. And it's so much more than just that simple digital whiteboard. your team can collaborate on all kinds of important projects planning research design feedback cycles for your product and remember faster inputs faster outcomes and velocity is how startups win we all know that to access our new Miroverse template and thousands of others sign up today for free and get your Miro account going miro.com slash startups that's miro.com startups to sign up for free what are you seeing sonny you're you're you like to tinker yeah so give us a you know this is the best part of the show everybody's talking about this you know uh regular series we're doing every monday here on this week in startups is our ai day we start the week with ai because this is a tread this is a tread that's moving at a daily pace and so we're covering it weekly which means every week there's gonna be three or four important things that we'll just demo for you every monday morning get your coffee get your cold brew whatever you're doing uh and we'll just get into it so let's get into it sonny yeah well you know here's a good one that you know talks about pace of innovation and startups versus kind of established companies so i think you had aaron levy on a couple weeks ago now right to talk about last week yeah last week yeah so okay you know about sort of the new features that they're launching with Box AI, right?

14:59Where you can take documents and you can upload them into Box and you can have it turn them into Q &A. Well, just to highlight someone who's doing something quite similar, I'm going to pull up here this service called Chat PDF. And what this allows you to do, and it's openly available, you don't even have to log in. You can, for the folks that are listening, I'm just logged into chatpdf.com and I'm uploading PDFs. I uploaded kind of three different ones i uploaded the berkshire 2022 letter by warren buffett i also uploaded the actual law of the inflation reduction act and then a summary of the plus plus on that just of course here we go perfect yeah yeah one more and one more all right okay boom now it fills the screen good for everybody yeah all right jcal mode and i don't have to put my grandpa glasses on Yeah.

15:53And so, you know, it's, it's amazing. And look like this is happening in a lot of different places, but what's amazing is like this service has been stood up. It's available. You can use it. Obviously it's using LLMs and it really highlights like how fast things are moving because with box that's still in, in like a private beta. and so i think this is kind of i'm trying to tie it back into what we were talking about a few minutes ago like the pace at which this stuff is happening is incredible because you have to wait you have to sign up and obviously box will give you all the enterprise features and lots of people will use it but this is something that everyone really needs to think about because you can come here take your pdfs i think is really useful uh i've been using it for the last couple of days to just summarize uploaded berkshire's 2022 letter this is warren buffett's letter and it said welcome to the exciting world of berkshire hathaway's performance this pdf file provides an overview of berkshire's annual percentage change in per share market value compared to the s &p 500 with dividends included from 1965 to 2019 here are three examples here are three example questions you may have about this file so what is this called when it anticipates your questions has the industry come up with a prompt for this where you anticipate what the person's questions suggested future queries it's it's something new right it it is um you know and this is like a really big area where you know things are going from the uh going towards i would say where it's it's starting to anticipate what a user would like to learn about what has been given and it's doing that based on sort of all the learning that's going on in these systems and you know sort of the prompts that are coming in and i think this is a really powerful aspect of uh the lead in towards ultimately agi is i wanted to you know predict what we're doing vinny maybe you know you can jump in here because you guys are obviously adding these features into weight room like how do you think about this um you know so here's how we're thinking about it right now um it's really good for i mean you take a pdf there's some context to the pdf but it doesn't really understand the bigger picture of what's going on it's based upon the content in it so for a large scale pdf three four five hundred pages it'll look it can probably do a pretty good job the small pdfs there's just probably a lot of context missing so the way we think about it is when we're working with companies and they're using wait room and they're um we're extracting the data from a single conversation it needs broader context for the whole company it doesn't you know it won't understand what is going on example of that yeah okay so here's an example let's say we have we have a a company that manufactures a certain uh you know you know some device okay some box and they have a shipping delay because there's a component that's missing uh that they haven't received for the manufacturing process and now they've got a logistical problem Now, imagine there's a meeting about this and people are having this conversation.

18:55And the AI can transcribe the conversation. It can infer what's going on based upon what it knows so far. But it doesn't know what the far-reaching effects of that could be. So, for example, let's say it's a really large order. Now, all of a sudden, that means that you can't bill the customer, which means that the financial projections are now out by 30 days, let's say. and that means that the companies are not going to have enough cash to pay payroll at the end of the month and now it needs a lot of credit now these are the things that takes days to surface up in a company uh and the ai can do it immediately can flag it saying this order is critical we're behind now we have a cash flow problem alert finance let them go to the the bank and you know give some sort of facility or speak to the ceo or whatever it is um but the without the context of how this functions in the big organization it's just one conversation that could seem very arbitrary yeah that's a great example so we're here on the phone talking about our new product it's shipping from shenzhen it's delayed and then when a product is delayed the llm might be able to say okay when products are delayed what are the downstream effects of that customers are unhappy returns or refunds may happen extra expense we're going to burn another month's runway but if it doesn't know your uh payroll system or your orders it can't really anticipate that yeah so so we're going from a world where you know like let's not confuse what ai is supposed to be or what it is with like what transcription is for example or just basic interpretation basic transcription interpretation is we've had that for years you know we've had it's table stakes now it's table stakes yeah you have to just be able to understand what what the what the audio means you know um yeah now it's about how do you take that and provide context and so i i think the future and i'll make a prediction i think i mean i put up a tweet about this actually um but i think the future is is really about um every company one day is going to have its own llm like of some sort so every major company within i think within five years i found my tweet but um will have its own its own llm on some level just because like how many conversations are happening in a company every single day within yeah they're happening in slack they're happening in email they're happening over iMessage signal there's they're happening in notion or coda documents slack's a good example so now yeah so now imagine the ai the llm for the company has access to every conversation in slack obviously private ones will be excluded whatever every every sort of group conversation is now being fed into the llm live every video conversation uh yes all your ones or transcripts it's all going in and so you know basically i think within five years the if you're not running an lm a custom lm for your company that's been trained on your business five months well i think things no no but i say i say five years before like every company will have to do it including the laggards yeah because it does seem to me it'll put them out of business if they don't do it it'll put them out of business and yeah and like i was at this conference locology and my quote that they quoted me on this tweet was you will have virtual agents sitting in every virtual meeting that happens within a company so like well yeah you know sunny this was something we went through we've been on this concept of doing transcripts for this week in startups since you know maybe the third fourth fifth year of the show because people want to get the knowledge out of it or get summaries and i have done multiple tries at this but human beings like to do a transcript is a couple of hundred bucks if you want to do it properly with some level of fidelity you know people's names etc now you can't not get a transcript and the transcript is being done in multiple places youtube was the first with the really rudimentary closed caption one but spotify people are creating search engines now of twist and all in uh new podcast players it's built into zoom descript which people are using to edit it's built in there so now we have five transcripts of every episode for free or close to free i'll take i'll take you a step further jacal knowing how much content you've got on twist for the past decade plus whatever it's been running um you could actually build some sort of an llm or model or you can just plug it into an existing model like open ai and you could actually create a business coach yeah actually understand so from every conversation you had with brian chesky and whoever else yeah all that knowledge you could capture it all you could you could transcribe it and incorporate in the model and now people could basically go and chat to uh this you know chat twist chat whatever you want to call it and say hey i'm struggling with like my business what is the um you know how do i how do i like make the numbers work how do i you know get more runway give you a reference or it could just give you an answer yeah um now i guess the then the question becomes sunny who who gets to use the data from the show because people are showing me experiments of like all in and this speaking startups and doing that and i'm like that's fine for an experiment for a year i just tell them all the same thing feel free to experiment for a year but i'm starting to think like i uploaded this to youtube they have the rights to the transcript so i guess they're going to try to claim if i put this on youtube or i put it on spotify or if spotify grabs the rss feed that they now have the rights to the conversations that occurred here i'm going to need to like have a conversation with each of these platforms i don't want my stuff in their llms necessarily unless I'm getting some quid pro quo.

24:24Imagine this, you've got the greatest idea ever for a tech startup, and it's going to change the world. But you got a problem. You don't have the engineers you need to make this a reality. Why? It's hard to find engineers, right? Everybody's in competition for those great engineers. And you've got to manage your burn rate. You don't have unlimited resources like those big slow incumbents. You've got to be efficient. So now imagine you had a partner who could provide you with more than 1000 on demand engineers. And these devs were vetted, experienced, result oriented, and passionate about helping you grow.

24:58And they charge competitive rates. Sound too good to be true? Well, you need to head to lemon.io right now. It is not too good to be true. Startups choose lemon.io because they only offer handpicked developers with three or more years of experience and really strong portfolios. Only 1 % of candidates who apply get into Lemon.io. A couple of great launch founders have worked with Lemon.io and they had great experiences. So here is your call to action. To learn more, go to Lemon.io slash twist and find your perfect developer or tech team in 48 hours or less. And twist listeners get 15 % off the first four weeks.

25:34Stop burning money. Hire developers smarter, faster. Visit Lemon.io slash twist. let's unpack a couple of things so you know one regarding the transcripts and translation um we'll kind of maybe lead into google io in a second but they they announced something called universal translator all right here's the 44 second clip let's uh listen and then we'll see on the other side of 44 seconds universal translator is an experimental ai video dubbing service that helps experts translate the speaker voice while also matching their lip movements. Let me show you how it works. What many college students don't realize is that knowing when to ask for help and then following through and using helpful resources is actually a hallmark of becoming a productive adult.

26:25Muchos universitarios no comprenden que saber cuándo pedir ayuda y usar recursos útiles es en realidad una clave para convertirse en un adulto productivo. We use next generation translation models to translate what the speaker's saying models to replicate the style and the tone and then match the speaker's lip movements then we bring it all together that's wild so that sounds like the saturday afternoon kung fu movie problem it's a it's a film from china putting american dubbing and you can see the lips are not matching at all i i love this because i hate what even on netflix there's so many good shows people have told me about and either i wouldn't watch it was like i would rather watch subtitles than watch the english dubbing and they can change that i'm all in like that's great and that creates it opens up a universal content pretty amazing we got pitched on this week in startups being uh re acted reenacted i guess where they were going to hire voiceover actors to do my part you know and aaron's part or you know your part sunny and your part vinnie and then you know basically republished the podcast in german and it was going to be i think i don't know if nick remembers but i think it was 500 an episode and they were like yeah you know for whatever that is 150 000 a year you could have an entire german version then you get just one german sponsor to do all 250 episodes for thousand dollars an episode and you'll be a prophet and i was like ah it's a lot of work and then you'd have to have voice actors full-time you've heard the scale of the show you'd have full-time voice actors and this just eliminates all of that work it's wild yeah and so coming back to what you said here so first of all the platforms that you're publishing to are going to start making these features available they started to your point with transcripts they'll just make this available so i think that's something mindful there's something you have to be mindful about and then two in terms of like how are people leveraging all that that i think that's the grimes problem so first you know this functionality is coming so if you're kind of building in around it be very careful because the platforms are moving very quickly and adding you know this type of functionality there's a lot of other great things that were launched at io which probably hit on what you know 20 or 30 different startups are working on right now yep um and so you know you have to really think about what what is your moat in that particular situation or what advantage do you have to not have to have a moat but like something proprietary maybe someone can show up and cut a deal with you jcal right and say hey we'll do this and then you know we'll be responsible for you know sharing the rev share that comes from it so i think that those are the type of things that people really need to think about now yeah by the way the um google calls these uh you know questions suggested questions proactive prompts and so this could be a very interesting um experience i think for a lot of people to when they open an app or they open up some you know their slack for the day or their box and they're working work will be the ai is going to be telling you what to do next at your company it's going to be like you know what you have enough customers but you're losing customers so you really should be working on churn today here are three customers that aren't using your product anymore do you would you like to schedule calls with them you know and like it'll just the ai will just know that these customers are going to churn and do something preemptive with them it is really could be uh transformative for people at work who have this sort of in in writing blind sheet of paper problem you know where you're like it's very intimidating to have a blind sheet of paper for startups or any company you come to work sometimes you're like what do i do now what's the most important use of my time what's the most effective thing for me to do today and this could really help drive that you know everybody's been hearing about hugging face sunny maybe you could explain what hugging face is um and why uh everybody in ai is talking about it right now yeah so think of hugging face as more of a collection of models and it has a wide range of models so you know we actually you know vinnie talked about this a few minutes ago where um we think there's a long-term belief that there'll be a lot of different AI models for a lot of different use cases.

30:40And Hugging Face really was a leader in that. And it continues to be where they don't have a single model. They have lots of different models that have been uploaded by different users. And so it's a combination of what you think of as a, like a GitHub and a repository for AI, where there's lots of different models that don't necessarily have to be for generative AI. They've been there for traditional machine learning for for years and actually amazon partnered with them so that's who aws's partner has been up front in terms of giving users the selection they need for their you know particular use cases and i think it's really fascinating pull it up and then walk us around the side give us a little tour of hugging face if you will in real time and uh you know make sure you sports cast it uh as you do so well now on the show just telling us what we're seeing um yeah because they have the data sets there they've got models yeah uh etc yeah let me just pull this up here so and people know github is a community it's really like a community of developers so when somebody figures something out they post it and they show their work yeah and then if you show your work you get kind of social credit for that people follow you and if people follow you it kind of builds your career correct that's like part of the dynamic here is that people who are developers are looking for recognition they want to move up leaderboards and or get exposure at a site like github or here correct that's a great great explanation and so you know here we've kind of gone i've just clicked into the models area i'm on the website and what you can see here is different models and on the left you can organize them you know being multimodal computer vision natural language processing audio tabular and you can see there's lots of different models here there's like 6 000 pages jcal right so if you went to audio there's a collection of them like audio classification voice uh activity uh text to speech so if you were to click text to speech uh as but one example it would give you a bunch of the models but it's sorting the models by most downloaded so if you didn't know where to set where to start can beashi i guess is uh been used 28 000 times in the last month this is the most downloaded text-to-speech model and you can take this for free and start using it correct and so this is you know like a collection of and you can see here like just the sheer number right there's like thousands and thousands of models here and this goes back to again these are not all you know focused on generative i this is like sort of the broader reach of um you know ai and machine learning and so this is what you know one of the things i get confused when we see about regulation because hugging phase been around a very long time and these are all uh ai machine learning models i mean where are you going to cut this off there's so many here that are serving different purposes and what's interesting is if you were to click on like the first one um when you look at that one uh it actually has a hosted version of it on the on the right underneath downloads in the last month there's a hosted version so if you typed in welcome to this week in startups and you hit compute you could actually kind of see a basic version of it without a lot of operators or interference um yeah and so uh we can put that in a post but yeah so this basically is why things are moving so fast because people are sharing their work yeah and this is the argument you know there was a uh sort of a leak of a internal google document uh last week and you know tough to keep up with all this now too and in that google document what the you know sort of the summary of it was the highest level summary was open source is moving so fast that open source is going to win here and it talked about just sort of the pace of innovation has come down to two weeks um you know paul graham had a really good tweet this morning days yeah days yeah yeah so and that you know paul graham had a really good tweet this morning i can try to find it and pull it up but we talk about it actually was trending yeah yeah and uh i i think it's it's really incredible he basically said what we've been saying on the show which is hey this is moving um much faster than anybody anticipated something major is happening when the time scale is days i was there for the last one of these the web and even then the time scale was closer to weeks than days and he's absolutely right back then every week or two something would be added to the html spec so people would be like hey you know i want to put a background on this it's a gray background can i put a background color can i change the background yeah okay background colors is up and running a week later then somebody would say hey you know what background colors are nice i would like to put an image in the background with the texture can i you know point to a jpeg like okay sure put a jpeg in the background yeah and and that's what we're seeing right now in terms of you know that that pace and i he really nailed it and i think vinnie first called it out in terms of like anxiety for for developers i think that that anxiety probably even continuing move up the stack all the way into companies and projects and venture capitalists who venture capitalists i've had two startups now that we back to do something in ai i won't say which ones and they came back to me like this is kind of built in everybody can do this now all the work we did for the last three months is now built in native everybody has this superpower so it's like it's like buildings like some weapon and you're like oh i put a scope on this rifle i can now you know have 50 more accuracy and it's like oh yeah scopes are open source everybody's got a scope it's like okay what's the next weapon and you're like arms dealers in startups if when you're building tools and this arms race is like a free arms race everybody gets the latest weapon everybody gets the latest armor everybody gets the latest counter defense and so a lot of founders have been telling me what should i do and i'm like just keep racing and building and build an experience and a brand that speaks to people if you have a nice brand and a community people will stick around i i really think that's great advice i think the community aspect of what's going to happen here is and it's a good kind of tie into what paul talked about it's important because it's so early the communities that establish around your projects are as critical now as more than ever before and i also think we're gonna go in a new world where you know we've seen the internet in the last 15 years become about sort of giants titans um you know whether There's like the Facebooks of the world where everyone had to be there for kind of the network effect.

37:11I think we'll now see a world where you can have smaller communities. It could be 10 ,000, it could be 50 ,000, it could be 100 ,000. That's not internet scale, but I think it's valid to build businesses around as well. So I think we'll see kind of a move towards that. Listen, when you're a fund manager like me, you're trying to raise money. The hardest thing to do is get in front of those elite LPs, limited partners, right? How do you get those connections? Well, iConnections is how you get them. So check out iConnections. This is a platform to help connect and meet elite capital allocators. What a great idea.

37:44And these are the biggest LPs in the world from every category, endowments, foundations, sovereign wealth funds, and iConnections speciality is building and connecting their global network of capital allocators with a diverse range of fund managers. They do this two different ways. They've got their iConnections platform and they do bespoke events, including one I'll be speaking at. The events are amazing. as a guy who knows how to throw great events. And the iConnections platform offers a bunch of great tools like an LP meeting schedule, document sharing, a database of all these industry contacts that they keep up to date, all the world-class events, and some exclusive content.

38:17So if you're a fund manager of any kind, you need to sign up for iConnections at iConnections.io slash twist. The first 25 funds to sign up for iConnections Miami 2024 in January of next year will receive a 20 % discount. This is the largest capital allocator event in the world with trillions in lp money represented so sign up at i connections.io twist i connections.io twist yeah um i thought one of the most interesting things i saw in hugging face was they're just speaking to this uh exact you know impact of like who's going to win the open source or the closed proprietary and if you look at the closed proprietary so open ai open was the name of it elon funded it because he wanted it to be open then they closed it so it's really we should start i'm going to start referring to open ai as closed ai so closed ai their approach is to not share their work not share their data so if you use closed ai everything you do as a founder just keep this in mind they get the benefit of and they don't share back so if you is that accurate you think sunny vinnie that using closed ai means you're giving but not get it i think it i think it's unless someone cracks ui in a way the same way microsoft and cracked ui and distribution um this is this is linux versus um you know and and the the era is very different right so in the 90s like communications and was costly and the internet was expensive it wasn't as widespread as it is today so there was remotes yeah yeah so those are most that that that microsoft actually yeah there was a lot of uh information asymmetry between people not knowing Linux was out there or getting information on how to use it and the developers were, you know, UI, UX wasn't something which they were really focused on at that point.

40:10It was very developer-centric. I think with BARD, in particular, Google's resources around this, you could find that actually that open source is the way to go just because of pure scale that it gets to and then the UI, UX issue gets resolved to just, you know, experience that we've had and And now we have a better global communications network. So where open source failed in operating systems, I think it could succeed in AI. There was, in operating systems, this idea that Linux or Lindos or any number. Ubuntu. Ubuntu is another one. That like, hey, Windows could be displaced. And it never was, except Google then created Chrome.

40:51And really, to have a desktop, you really have to be making user-friendly. And these things are not user-friendly. so the fit and finish never competed with mac and the os but i think this time i think this time is different i think this time could be different yeah this page is really interesting this is the open llm leaderboard and it has a little description here and what it says is uh with the plethora of large language models and chatbots being released week upon week often with grandiose claims of their performance it can be hard to filter out the genuine progress that is being made by the open source community and which model is the current state of the art the open llm leaderboard aims to track rank and evaluate llms and chatbots as they are released we evaluate models on four key benchmarks from the luther ai language model evaluation harness i think that's how you pronounce it a unified framework to test generative language models on large numbers of different evaluation tasks a key advantage of this leaderboard is that anyone from the community can submit a model for automated evaluation on the gpu cluster as long as it is a transformer model with weights on the hub we also support evaluation models with delta weights yada yada yada evaluation is performed against four popular benchmarks ai2 reasoning challenge 25 shot a set of grade school science questions heliswag 10 shot a test of common sense inference which is easy for humans 95 but challenging for sota models mmlu five shot a test to measure text models multitask accuracy test covers 47 tasks including elementary mathematics u.s history computer science law and more truthful qamc zero shot a benchmark to measure whether a language model is truthful in generating answers to questions we choose these benchmarks as they test variety of reasoning and general knowledge and then it lists them and you see here 11am-65b is the model with the highest average and you can see their arc shot and their hello swag and you know like where they're done so the the race is on boys um i don't know have you seen this before sunny or vinnie this leaderboard yeah you know i did and i was looking for a tweet and i just couldn't find it there was a leak of a google document and or maybe you could look for nick in the background yeah yeah and in the google document they basically said in just a few weeks and what you know we've spent hundreds of millions doing people spending tens of thousands that are you know getting there just as quickly now because of you know and that was all driven by the leak of the llama which is you know the facebook model um and then you know there's just sort of some hot off the presses news i also dropped it in here it's like you know a mirror from the information basically saying and i like i'm just reading it as we're doing this here but i guess open ai is now moving to open source some of their models maybe even in reaction it's going to be open again yeah i know it's just funny because you just giving him that nickname and so i just dropped the link in the chat i'll believe when i see it if they're going to do it but i think there's going to be pressure on them to do it because they sam is so smart i would say sam is clever he's even which is even better than smart in this instance he's clever and i think sam altman is going to be like you know being closed helped us because we could build in stealth but now that everybody's adopting this and all the best developers are spending time on this close is going to hurt them uh so let's read from this uh i mean story my fear my thesis on this is that that the real um you know the real revenue and and and the real revenue revenue and strategic differentiation you have right now is the hardware not the software the software can be easily replicated but the the gpus that we're running out of that this is where i think like the big iron yeah exactly so the ability to produce these chips aggregate them and and basically you know gpu processing becomes the utility of the future that's why i'm into stuff like render i think render is interesting for this sort of thing as well like this is where the world changes we have you know if we can utilize efficiently the gpu chips the software just runs on top of it you still need the hardware to do the processing you can't have all this open source you know all these different models without like hardware becomes a limiting factor not software and of course you're referring to this a100 this is the that's just one of the a100s are i mean they're powerful but you could line up you know you can line up a bunch of you know 4090s and get similar output you know it's it's not too far i mean they use tesla cores in the a100s and i think they use cuda cores in the in the 4090s but you know it's it's um who is going to put up the biggest fight against nvidia here and is there an open source hardware project as well like i don't know yeah there's no open source project so like the the direct competitor that like the most meaningful one is the tpu right by google that's the most meaningful one um so google's making their own hardware well amazon as well amazon announced it as well they're doing the same thing and google has been for years they're they're on i don't even know which you know third fourth generation of the tpu yeah um so they've been doing it for a long time um so does their tp is the google tpu available to the public or is it only for their cloud it's only for their cloud got it so they're making proprietary and this is something he talked about at google io a long time ago they've been talking about this for a while yep yeah yeah they They had a moment, you know, years ago where they were doing more in TPU compute than they were in like kind of x86 compute.

46:28Yeah. And that wasn't a recent thing. It was several years ago that they had made that, you know, point. So really it's going to be GPUs versus TPUs is the battle we should be monitoring. I'll give a counterpoint to Vinny's point there, which is, you know, from this, we have the Google, we have no mode. And so, you know, I thought this particular paragraph, this is a really good article. It's worth reading. Read it, yeah. So people who are listening. Yeah, for sure. I was going to read this one paragraph. While our models hold a slight edge in terms of quality, the gap is closing astonishingly quickly.

47:02Astonishingly. Yeah. Open source models are faster, more customizable, private, and pound for pound more capable. They're doing things with$100 and 13 billion parameters that we struggle with at$10 million and 540 billion. Right? and they're doing so in weeks not months that has profound implications for us yeah so the interpretation here um is that and he says this in bullet points after that paragraph we have no secret sauce there is nothing proprietary there's nothing owned by google um and then his second bullet point people will not pay for a restricted model when free unrestricted alternatives are comparable in quality and this is something i said on the all in pockets and i think here you know this lead that open ai has it feels tremendous until it doesn't right because if in a vacuum it's the only one you could use and the only one you can see other people using it feels like it's an insurmountable lead but then bard came out and we started playing with bard and it feels like okay bard is at you know chat gpt3 or 3.5 right so the gaps will close and then ultimately the race will be won by the platform that provides the most for the least cost and the platform that's going to provide the most for the least cost is going to be an open source project so let me let me just i want to just jump to bard versus um jackie we tea i've been playing about a lot this week it's actually very inconsistent yes it's delusional it says we're no no you can run the same query twice and it'll give you different answers that's how chat gpt used to be yeah yeah so so they're very much behind i mean uh i i asked it to rank i went through a process i said rank the top 10 vehicle manufacturers in the world based upon gross profit generated right so uh it ranked the 10 tesla wasn't in there so i said to it where is tesla and it responded to me barred responded to me saying oh my apologies i didn't see that now i didn't realize that tesla was a vehicle manufacturer whatever excuse it gave me and then it had and then it had to update the list for me again and then updated it for me i mean imagine not having tesla in that list how is that possible i literally just did it here are the top water motor models of the u.s in 2021 and 2022 according to motor intelligence ford f-series chevy silverado ram pickup toyota rav toyota camry toyota honder yeah there's no tesla yeah so i did tell it to include 2022 it put tesla model y as number six yeah interesting so so this is this is the uh this is the problem i don't think it's i don't think it's close i mean i don't think it's good enough yet um i think what it does really well by the way is it takes feedback and then incorporates that feedback and like for example i did i said tell me about vinnie lingham and it tells it says that gifts was acquired by first data in 2015 for 390 million dollars the year is wrong the amount is very wrong uh you know it's just like all the stuff it said to me it said i sold another company for 100 million which i didn't i don't know where it got that from um but you know it's very inaccurate and here's the problem and here's the point i'm trying to make and highlight here is i think we're going to have um a lot of bad ai be based upon false information on the internet so the more fake information that people populate with the more these ai tools are going to search read and then you know i know it's trying to give authority to certain sites but you know people can like it's putting data from um you know wikimilly.com so can i give an example here um and it's a bit of a plug and sort of what we do at Definitive, but it actually gives a really good example.

50:53Would you guys mind? Yeah, please. No, plug live. I have to leave. You guys can keep going. Okay, yeah, get to the airport. Cool, thanks guys. See ya. Everybody visit weightroom.com. Vinny Lingham, blah, blah, blah. Follow him on social media and see him at the club, at the music festival. Live the life of Vinny Lingham on Instagram. And the All-In Summit. And the All-In Summit. See you at the All-In Summit. I hope you bought a ticket. I hope you bought a ticket, Vinny. I'm selling out fast. I don't want to apply for one, but i'll i'll buy one this is a big internal debate do i have to apply i mean seriously no you don't have to apply the issue is each bestie this year only gets 25 tickets for their friends and family so now i am in regular we have 25 vips each there's 100 vips for this so now there's a little bit of brinksmanship going if you're in the poker group you know and you want to get the free ticket which of the four besties you know you might want to go to all four and try to lock up one of those tickets if you're you know xander or you or okay sunny who's going to get you your free ticket right like you two guys how do i say no to you if you're coming on the pod every week i gotta kind of hard for me to say no right then we get 25 tickets that are 50 off i think is what we've settled on okay but now it's like this thing's going to sell out we already sold out the general admission it's this is such a hot ticket that now it's like they're counting how many tickets each bestie gives away so so are you saying that sunny i getting two tickets for you jay well i mean how do i say no is the issue but i would very much hope that you lobby your i think sunny's made a lot of money a lot more money for chamath so sunny should vinnie you guys have business relations i've made money for chamath as well yeah so i think you guys got a lobby chamath so i can save those two tickets for friends and family i'm actually taking my tickets and i'm selling them on the black market and this thing sells out i'm selling them for 20k each so i'm gripping this big time that's like a half milli right there I'm going to sell those 25 for 20k each okay Jack I'll reach out to one bestie and if I say no you're the next guy just work on four please please please I'll do one I'll do one I guess you know like whichever you got to work your besties I think you've made the most money to David Sachs though oh really Vinny then you got yeah if you got that Solana Solana bag you got to go tell him he's got to give you a ticket yeah yeah I'll bring I'll bring Sachs okay cool yeah I would do that now because now the word's going to be out when this podcast comes out okay cool that's the biggest one see you guys he dropped that bag on the table he broke the table in half yeah he's like hey i got your envelope drop the envelope on the table sax is like marble table broken half he's like i can get a new table see you guys that's a lot of bad love you both love you besties look at this um all right let's keep going you got any more examples for us i mean god i just wanted to you know like it just kind of builds on there and and uh we can see if it's useful or not but i'll I'll share with you guys here.

53:44And this, hopefully, I was going to send this to all you guys because this came up in the all-in pod as well. So, and this really shows like where, and Ben Thompson had a tweet this morning as well. And if you can pull it up in the background, let's pull it up right after this. But really, you know, where these LLMs are good and where they kind of have some challenges. And so in this particular case, you know, we asked Bing, which is powered by ChatGPT4, what is the average delay time for each airport in 2022 it comes up with an answer that we would have normally not typically seen i'll read it quickly here it says i found an answer on straddle's jet that says the average uh time for airport delays is 15.3 minutes but this is for all airports not for each airport um right you know chat gbt is getting better where it doesn't you know give you one of those type of standard responses but it starts with its hey i don't have data beyond you know september 2021 but it then offers a method which says hey you know to find the average delay time for each airport in 2022 i would recommend going to the you know authority's website the faa and you could pull it from there and that's how you could get the answer so it's helpful but it doesn't give you the answer but it's at least um giving you the right directions right um google is finding a result and just kind of flipping through some sides there for people listening google finds a result that comes from a page that you know a website euro control.int which has published delay times and it you know provides the like this is where google's crawl is so valuable to them they're crawling the web and they've indexed the web so granularly over so many decades that they've got a big advantage and they're doing that in real time right and this is like sort of in this came up in the all-in pot but this is a little bit of like the you know where you can see some of the hallucination happen.

55:33So in this particular case, you know, we asked the same question, what is the average delay time for each airport in 2022? And it comes up with a list, not all the airports. And it says, Hey, LaGuardia is 19 minutes. Now you or I know, J. Cal, if you've flown in LaGuardia, it's not 19 minutes. Right. And so, but this answer looks very, very, you know, it's very confident on this answer. So we asked the same question inside definitive. This is a screenshot of our product. we actually go to the airline database and write a SQL query, which is shown here. And we come up with a response that has all 393 us airports and the average delay time.

56:10And in this particular case for LaGuardia says 74.54 minutes. Okay. So now we want to go and verify this. And so the only way we can go and do that is we can go off to the, you know, Bureau of transportation statistics site and go to, uh, you know, june 2022 and we can look it up uh here and say uh here is the average delay time um in in that time frame for 2022 74.55 minutes at la guardia and so what really this shows here is the llms and we use an llm to come up with a sequel for that answer but they can be really good reasoning engines but when you're using them as information retrieval systems and maybe if you pull up the um the uh the tweet from ben this morning nick um they try to give an answer which it has seen someone come up with before which is not always practical in a scenario when you're looking for a specific answer that's based on data that may not have been published in a web article that it could have crawled got it so what it's doing is it's taking the corpus of journalism blogging and using that as a proxy of the data what you're doing is saying okay if a journalist had gone into that database and then created this document i'm going to just go directly to the database database yeah and that's really more definitive intelligence if you will correct which is the name of your company definitive intelligence happens to be the name of your company and so this is like beyond a crawl is the actual sniper shot so like a crawl is kind of like spray and pray you know what can we find and leverage the collective work of humans and you're saying well screw that there's data there's got to be a database somewhere we can go ingest but that is not what google has done google cannot it's not allowed to go ingest somebody's database that's against the rules um in some cases it can you know if they're public data sources it can sure and right but in many cases it cannot because the it doesn't allow for it right um and so uh and sorry i was i was misquoting it wasn't ben it was benedict evans and uh i think yeah if we pull this one yes benedict evans formerly of a16z yes yeah and and i think you know it sort of highlights what we just talked what we just locked about what we talked about here right llms do not tell you the answer to your question they tell you when people ask that question like this this is what the answer that other people tend to give tend to look like i think this is a really really interesting and you know thoughtful explanation and you know when you're trying to apply the technology i think is something really to think about you know whether you're using it for information retrieval or you're using it for you know logic or reasoning yeah so if you were to ask you know um what's the best peking duck in new york city people have asked that question many times yes it's been answered by time out new york zagats yelp you know some forum a newspaper and it gave red farm and peking duck house as the first two those are the first two i would have given you yeah uh red farm is the peking duck that i order on gold belly as just about one example and then peking duck house is a place where you can bring your own bottle of wine no quirkage fee and it's 50 bucks a person for peking duck so you can have a table of 10 500 bucks you bring 500 bucks in like really great wine bring five bottles of great wine and you're in and out for a thousand bucks for 10 people 100 bucks a person okay uh yeah i'll take you there next time with a review if you like picking duck highly recommend now this transcript is going to get crawled by someone as we were talking about earlier and then that's going to get added into the corpus of training material for these llms so basically it's going to like it's getting all the chit chat out there and you know and what google will do is and this is where it gets really interesting i've had a conversation about peking duck a hundred times in my gmail yeah with various people now is that something that it can use publicly of course not but maybe they'll figure out a way to finagle if you're using gmail for free and there's some knowledge in there maybe would point it in the right direction who knows what you know insane justification they'll use for taking conversations that were not meant and using them you know well yeah i mean especially buying free products like you said right where you know there you have to understand there's got to be some trade-off of your data or your knowledge that's now being put into other systems yeah that's kind of scary when you think about it but that's where i think you know i really think google's uh and i i bought google shares um i don't know when i've checked j trading but when people got down on google i was like i think they're going to incorporate this into everything and i think at google io they incorporated it into most things most things yeah most things now did they i didn't see i didn't watch io but what was their gmail strategy because to me it's really incredible is the most important one there's a yeah yeah so there's a short there's a 15 minute summary video which i can send you jay cal it's awesome that right yeah yeah no no someone else did it was like uh yeah it was it's that was something we created at uh in gadget that verge which is like a bunch of being for gadget editors uh it was uh it was it was one of the steve jobs keynote in three minutes we do a steve jobs keynote in one minute three minutes and 10 minutes so we'd make like whatever version you want to watch we'll do it but now yeah i could do that yeah it was a it was a really really incredible summary i'll try to find it it's uh i tweeted about it here i hear it is i'll just drop it in here you can pull it up but we're not going to watch it here it's too long but But really, in terms of the video, Nick, if you want to pull it up after, is that what they're doing in Gmail now, so basically they're adding a button, which is a prompt.

1:02:12And when you hit that button, you can actually write a prompt to reply to a message. And then when the reply is generated, they have like sort of some options. And I don't remember exactly what they were, which is like, be more aggressive, be more neutral, be more passive, you know, something along those lines. and so if you're fighting with someone about like uh you know canceling something along those lines you can say hey draft me an email that's a response to this and basically and then you know it generates a response and you can tell it to be more passive saves but same more aggressive it was really really cool i don't want to like tout my j trades but on uh march 21st i bought like a hundred grand uh in what a thousand shares at about a hundred bucks a piece uh and uh it went up 17 in a month that's another great j trade my j trades are right now and google who would i got the meta one and the google one my j trades are now 7.45 versus the s &p equivalent of 2.5 so i'm i'm beating the market by 3x apparently on my j trading this is not trading advice do not follow my j trades but i think it's just like timing i bought when everybody was scared i'm getting killed on my disney trade yeah just just absolutely slaughtered and my amazon trades getting killed and my warner brothers but i'm gonna buy more of them those three companies are gonna figure it out disney's gonna figure out ai yeah amazon's gotta figure where's amazon and all this where's amazon well so they partnered with hugging face okay so they What does that mean?

1:03:48So they're making it such that, you know, like what we were just showing earlier, if you want to pick models and you want to run them in an infrastructure, they'll make it as simple as possible within their infrastructure. Got it. So if you really think about sort of that open source argument, they are embracing that the most kind of, I guess, at the forefront. So if I want to grab the model and start playing with it, it'll just dump it on my AWS account and let me start playing with it. That's smart. exactly and why don't they just buy hugging face man they should just bought it now yeah i mean that would be the move like you know a microsoft and github type thing yeah yeah and you know similarly google did a big investment in replet so yeah that would seem like a good move for them yeah that's i think that gmail thing where it understands tone is going to be really really powerful so aws then switcher landed it they're not like azure's like use open ai aws is like listen use whatever you want here's a bunch of options yep that's what google will do what google cloud so google cloud and then this was also at io they are basically going down the approach of uh they're going to be multiple models of different sizes and we will allow you to run google models of different sizes uh within your infrastructure with your own data so um like allow you to have a model that's private running within your vpc like your virtual private cloud connected to your data yeah and i think it's a it's a really smart approach for enterprises that already and because you know google has one of the leading data product or a couple of the leading data products but like the largest one probably being big query which is a large analytical database and so a lot of companies rely on that and so it's a really smart place to you know plug into uh and they gave some also really great demos there as well i think nick has the gmail video if you have the gmail video nick we should pull that up real quick for for json because didn't see that last week all right here's a 44 second clip we'll see on the other side say you're writing to your neighbors about an upcoming potluck now as you can see sidekick has summarized what this conversation is about last year everyone brought hummus who doesn't love hummus but this year you want a little more variety let's see what people signed up to bring somewhere in this thread is a google sheet where you've collected that information You can get some help by typing, write a note about the main dishes people are bringing, and let's see what we get back.

1:06:18Awesome. It found the right sheet and cited the source in the found in section, giving you confidence that this is not made up. Looks good. You can insert it directly into your email. Wow. That's so powerful. It should have a button that would be like the JCal button, which would be like troll this person. and it would be the troll options and like my troll option would be like why do you why does hamas why does everybody hate hamas and then link to the clip of sasha bar and cohen talking about hummus and hamas and he's he's like uh trolling them trolling options should be there to do like jokes and stuff like that that would be a great startup is uh an agent that follows you around the internet and then optimizes your trolling and comedy and just makes uh you know a mockery of everything mockery ai somebody go grab that domain and mockery ai are anything else happening this week that we should talk about as we wrap here i think we did a lot i think there's a lot to process everybody yeah all right and then tomorrow on the show we'll uh talk about karen ai i don't know if you saw that one go by but some influencer uploaded herself and she's now made a they used to have these in the back of the village voice like you could pay a dollar a minute dollar a minute yeah talk to somebody in a um adult fashion i'm gonna be careful not to get a censored here on the youtube um but there's a dollar a minute virtual girlfriend for lonely people i think she made like 70k in the first week i mean i think people are gonna try that yeah that seems dystopian and crazy um and sensational but i don't know is that gonna be a long maybe they'll just build it into only fans and then whatever you upload to only fans it then makes a model of you that's interactive i mean that makes sense to me that you could have an interactive version of every person there and then you can it's like a real upsell premium if you want the person live all right everybody we'll see you next time on this week in service bye bye