SPEAKER_00: the fact that they would even consider doing round tripping where you know defined as we invest in this company that company buys compute from us so the money goes into their bank account and then comes back to our bank account even if it's not part of the deal yeah that's that's fugazi it's uh

SPEAKER_05: you know like something's wrong there this week in startups is brought to you by and broker startup insurance program helps startups secure the most important types of insurance at a lower cost and with less hassle save up to 20 off of traditional insurance today at imbroker.com twist while you're there get an extra 10 off using offer code twist masterclass learn from the world's best minds anytime anywhere and at your own pace get 15 off an annual membership to masterclass at masterclass.com startups and roots invest in the only real estate investment trust that creates wealth for you and its residents at invest with roots.com twist hey everybody welcome

SPEAKER_00: to this week in startups it's monday so sandeep madra is with us from definitive intelligence and we are going to break down everything that happened in ai over the past week and what we're going to see this week going forward you know this ai thing never stops in videos reporting uh their earnings that's going to be a blowout it seems they just keep selling more and more of this hardware but i was thinking about nvidia specifically and i was wondering and i wanted to ask you if they are going to have a durable franchise similar to and a monopoly is really what i'm saying by durable franchise i don't want lena khan to get in here so lena khan uh translation monopoly are they going to have a monopoly in this space on building ai super computers to just use a an incorrect term here or is their margin and their revenue and customers going to inspire all the other people making chips and platforms to go get it what do you think because google became an enduring franchise

SPEAKER_14: with 90 market share and search and search advertising uh and then you look at apple didn't think that's kind of a duopoly people say but if you look at the actual revenue and profits androids make no profit yep 90 i think of the profits from smartphones go to apple so it's it's

SPEAKER_00: essentially a de facto monopoly what do you think nvidia is it a monopoly or is it a uh at-risk franchise

SPEAKER_17: um it's it's the latter and and look i don't want to call it at risk i think nvidia is going to have

SPEAKER_20: an incredible business but you know the reason we're seeing this explosion today is a function of uh you know there are way more demand than supply of nvidia chips um secondly what really leads to that is the tools and frameworks that exist for you know at scale training and at scale inference are primarily built around the nvidia infrastructure but we're already starting to see that break um you know carpathy had a really really awesome tweet last week where he did a bunch of inference on his macbook and what we're going to see is the frameworks adapt in a couple of different ways one for at scale training we'll see frameworks start to pop up that can be used on um you know other silicon that's already going to start to happen the nvidia one is leading cuda i believe right but others are going to get there really really fast i also think what we're going to see is inference move to different workloads in fact you know there's a uh here you go right so he was running 16 tokens a second on his macbook right and so and and you can you know he kind of maps out the difference and where things are going to go and and everything else and so take a look at this chart can you explain

SPEAKER_24: this chart to the listeners it's uh looks like there's a bunch of red dots a bunch of blue dots and

SPEAKER_26: a and an average line going up but what are what is this chart explaining to us yeah basically it's

SPEAKER_20: showing like you know flops right which is like operations per second right and correlated against like memory bandwidth in this particular case and what it's showing is that you know obviously nvidia is the clear leader and top into the right in this chart at you know a 100 uh which is kind of you know the world class but what you can see on the right hand side in terms of like memory bandwidth and cost right and you know how much these things cost so if you just scroll nick if you can just

SPEAKER_31: scroll the thing on the right hand side right there yeah exactly i'm just down a bit right there it's

SPEAKER_32: on the bottom axis and then the up and down access is is is scaling yeah it's like just normalized scaling

SPEAKER_20: exactly right and so what you can see here is you know the memory bandwidth and the uh flops out of an a100 but then a macbook m2 which is just a commercial cpu you know on your your device which

SPEAKER_00: is the greatest laptop ever made yeah i literally bought it the second i started using it it was so

SPEAKER_14: good it was better than my m1 uh macbook pro and i had the giant one and it was so heavy and uh i just

SPEAKER_36: gave that to my assistant and then i literally went and bought my wife one because they're so great yeah

SPEAKER_20: and what's the interesting thing here is when when you're taking on this category of problems there there's there's two things you have to think about which is compute and memory bandwidth and what he specifically calls out here is although the compute is 200x more which you know you think is a problem i'll explain why it's not the memory bandwidth is only 20x more so it's you know sort of within an order of magnitude and you can scale compute by just going broad you know by pushing the problem out to more and more devices like sort of like a seti at home laptops and then this would go down exactly memory bandwidth you can't scale that way because memory bandwidth needs to be you know the memory needs to be very close to the cpu right and so you probably remember this from your days of building machines yeah sure but the fact that you can see that one is within an order of magnitude and the other is like two orders of magnitude but it's solvable i think we're going to see solutions

SPEAKER_43: to this problem very very quickly right so and this is what people have been doing since uh the beginning of server farms which is throw hardware at it yes you know one chip can't do it well what

SPEAKER_14: if i had five chips and that's what i think we're going to see and and also what jobs are you doing it might turn out that 90 percent of work um it doesn't matter if it takes one second or 10 seconds or one second or even 200 seconds if i'm building a model and i want to refresh it every hour and it takes 200 minutes i could have a rolling refresh going on you know and taking the new data and what's that going to cost me so it's amazing that we're even talking about silicon right now uh in 2023 but we do see this happen every decade or two something bandwidth gets clogged because movies come to the

SPEAKER_00: internet or here you know language models cripple uh compute power or outstrip compute power it's great people have been looking for ways to leverage the amount of bandwidth we have and to leverage the amount of cpu we have it used to be like only people who could really use up the cpu were people making using video games or like pixar making movies a person doing you know email serving the web can't even use a fraction of what's happening on your laptop so much so that those chips were

SPEAKER_49: designed in large part to increase battery life that's how little compute matter to the average user

SPEAKER_14: that battery life was a more important metric you literally come into the store and they tell you 20 hours of battery life you're more interested in that than you are in the memory and the cpu 99 out of 100 cases exactly so i wonder if video editing will work nick will video editing work on the apple uh macbook air m2 with a lot of memory in it yes absolutely ssd drive yes so this idea that people would buy that mac pro studio for five or six grand or the mac pro tower for 10 grand what video editing

SPEAKER_57: difference would you see between a max faster rendering of like large massive files imax 4k stuff so

SPEAKER_00: if you're editing this pod or you're editing all in when you hit the export button the difference would

SPEAKER_60: be one minute versus ten no probably versus 20 probably 10 versus 20 25 i think it'll double

SPEAKER_57: the time got it yeah and we've already seen that from the you know i'd say the 2018 2019 macbook uh the show is exporting two times as fast from there maybe faster how long does it take to do export an hour or two hour show i'm curious our show is about 15 to 20 minutes depending on how much like

SPEAKER_63: graphical stuff more than i would think yeah wow it's actually more than i thought it would take

SPEAKER_64: from video for audio it takes about you know a minute 10 seconds okay yeah so what yeah i wonder when

SPEAKER_00: video is going to catch up here um and i wonder what the fastest you could do if you spend 10 000 or 20 000 on a computer because you get it done in one minute or no yeah well it depends on the

SPEAKER_67: compression so we compress the video pretty heavily so people can download it on spotify right so work on the more compression you're using the longer it typically takes if you're using a pretty light compression it'll be a larger file but it'll be quicker it's very interesting the two things that

SPEAKER_43: are c three things that are cpu heavy right now in the world video video games and uh language models

SPEAKER_74: those are gpu heavy specific they're gpu heavy now but and when i say cpu and gpu i just think we

SPEAKER_00: just put it all into the bucket of compute now right it's trending in that direction and i think you know

SPEAKER_20: calling them gpus is like confusing at this point because they're used like we're not using them for

SPEAKER_81: graphics right using it for matrix math listen i work with super early stage companies at launch

SPEAKER_71: like literally year zero they haven't even incorporated yet and then we hit the series a people have thousands of dollars in mrr and they maybe they've only raised a couple of hundred thousand before that series a and they don't have their insurance set up and in fact we recently had a great startup that didn't have dno and we had to really stop everything because they were having board meetings they were making massive decisions there were legal issues and they didn't have the basic dno insurance that protects directors and officers so we send them right to a broker and broker is business insurance built specifically for startups a single application will help your startup get four quotes for four lines of coverage in 15 minutes think about that four quotes four lines 15 minutes and they're going to connect you with one of their expert brokers for unmatched service that goes beyond your policy we use it at launch it's easy peasy lemon squeezy it's easy breezy what more do i need to tell you i use it i love it a lot of our startups use it they love it try and broker today with the code twist and you'll get 10 off their startup package in broker.com twist that's e-m-b-r-o-k-e-r

SPEAKER_63: dot com slash twist and use the code twist for 10 off okay let's get back to this amazing episode

SPEAKER_26: all right well listen nvidia's price targets just keeps going up people saying 800 bucks 780 bucks

SPEAKER_20: um it's nuts the warning i will say here on this one is if you remember the late 90s there was a boom in optical or networking chips let's put it that way and i started my career in that space and what happened was um it was a category of companies like one specific one maybe a bunch of folks that are you know from that generation will remember called jds uniface i don't know if you remember it but it was like one that made like optical transceivers for you know fiber optics and what ended up happening is when they couldn't meet supply you'd have to basically if you wanted like 500 transceivers you have to order like 5 000 from them and you'd get 500 so that you know their stock went through the roof because it's a huge you know deluge of orders and then ultimately the orders get fulfilled and then you know cisco in early or 2001 or two they basically all that supply showed up then had to write down like five billion dollars of inventory um because they bought way too much and so one of the things i fear in this era right now is there's this mad grab and everyone is going to everyone asking for compute and so there's you know there's a weird dynamic and you've talked about it before where there's companies doing these like weird trades of like oh you know take an investment from equity exactly and you know sometimes i fear that there's a lot more kind of artificial demand than actual demand created through a lot of these deals that are out there so i just say you know

SPEAKER_31: look i nvidia's incredible company they're doing great but the demand may not be as big as we think

SPEAKER_43: it is but yeah i'm not j trading it i'm not j trading it yeah i'd rather j trade the folks that

SPEAKER_00: people think in video is going to run over but that have great management team so if tsmc or intel or amd think they're going to make progress in this area and or you know intel i don't know who are leading candidates or new companies i would rather make the bet on those which are undervalued right now but that have great management teams and great revenue and reasonable pees because when you're

SPEAKER_43: investing in companies your entry price and the valuation matters the p of nvidia right now is like

SPEAKER_96: 230 240 depending on the day you know which is crazy like high growth companies are typically 30

SPEAKER_00: pe ratio so yeah be careful folks be careful out there i really do like your insight though about the bulk ordering and the shenanigans going on with the round tripping uh potentially i'm gonna put the word potentially in there i don't want to insinuate that people are breaking any laws here but the fact that they would even consider doing round tripping where you know defined as uh we invest in this company that company buys compute from us so the money goes into their bank account and then comes back to our bank account even if it's not part of the deal yeah that's that's fugazi it's uh

SPEAKER_57: you know like something's wrong there jason can i ask a theoretical question for you about the yes you may producer nick um nvidia uh thesis so amd is up 68 year to date intel is up 22 year to date sort of on the entire broad spectrum float of the ai mark right and the market returning and the market returning as well yeah yeah um do you think this that should eliminate them from a j trade just due

SPEAKER_103: to them being sort of kind of already a little bit more over value than they would have been otherwise

SPEAKER_24: i feel like the whole market's gotten a little bit ahead of itself those uh names could also have gotten

SPEAKER_00: ahead of themselves and i think when you'd have to sharpen your pencil in that analysis look at the price earnings ratio look at the growth and then uh look at the product roadmap and do you trust that management team how long have they been working together you know and the management team matters uh

SPEAKER_43: say what you will about dara he's not he's no tk i would rather travis be running the company obviously i think the company would have a greater stock price if he was but um you know he did get it to profitability he did create stability and it would have been just two different looking companies one would have been amazon under bezos and you know and one would have been you know amazon under um andy jassy and you know a professional operator versus a visionary so you get two different valuations there so yeah i would sharpen your pencils and look at it um how long the management

SPEAKER_24: team's been working together product roadmap you know customers buying stuff and then uh what's

SPEAKER_115: what's the financial fundamentals yeah but on the flip side jensen and team have done a great job

SPEAKER_00: you can't take anything away from jensen and the products they've made you know i watched them go from being like this you know dorky video game yeah to dorky crypto to like what's the point of all

SPEAKER_98: this i thought for sure with the crypto collapse like that company was toast you know and then you look

SPEAKER_00: at uh all of a sudden this customer race that has more money than they know what to do with apple google amazon facebook you know and then the investment community and that listen even sovereign

SPEAKER_14: wealth funds the uae uh and abu dhabi are making their own falcon language models the saudis supposedly stopped buying ferraris and uh yachts and now they're stocking up on h100s like literally and

SPEAKER_106: martin screlli's uh wears an h100 around his neck with a gold chain i mean this has become a status

SPEAKER_57: symbol h100 and just this morning the uk prime minister said they are investing 100 million pounds

David Friedberg: into in via amd and intel compute oh interesting fantastic okay so there's the ground there there's

SPEAKER_43: your fundamentals those you know have the industries growing this is not a drill all this compute i believe is going to get used you know maybe people are overpaying for it maybe they'll have extra capacity who knows but um you know there are a lot of jobs going to these servers so let's talk about

SPEAKER_00: how this uh infrastructure is getting leverage what do you got for demo it's demo time everybody type in this weekend startups on youtube and go subscribe to the channel hit the alert bell you get alerts

SPEAKER_20: when we post a new uh episode awesome so one of the things one of the themes i'm just trying to build on here is the evolution and usability of these models and you know even just in the short while that we've been doing you know ai demos we're starting to see like an evolution and advancement in that place and so the latest one here jason is um one where it's called rask.ai and what they make it it's

SPEAKER_31: basically a you know a gooey interface to convert models from one language to another language and so

SPEAKER_57: i'll play uh the original sony i i created like a a strung together clip of these if you want me to hear it and play it oh yeah let's do that please let's really do that right so you got a whizzy wig

SPEAKER_76: interface anybody can do it the url is once more rask dot a r-a-s-k yeah exactly and it's terrible

SPEAKER_131: super straightforward you basically you know um here we go pay 100 bucks a month here you go all

SPEAKER_00: right so here's a picture of me in a tick tock vertical format uh making a comment about something

SPEAKER_68: so let's play uh me talking i love most about this is like we have this oppenheimer film comes out all

SPEAKER_132: of this energy at the same time of people who are just really stoked do material science to do basic

SPEAKER_136: science to figure out big problems this is jason in spanish and espanol okay here we go on espanol

SPEAKER_139: lo que mas me gusta de esto es como tenemos esta pelicula de oppenheimer sale toda esta energía al mismo tiempo de personas que están realmente entusiasmados para hacer ciencia de los materiales para hacer ciencia básica para resolver grandes problemas and this is jason in hindi oh i was

SPEAKER_142: waiting for this hindi is que baré me joe baat mujhe sabse adik pasand hai bah ya hai ki oppenheimer ki ye film ekhi samayame sari urja bahar lati hai aise log joe vastav mein badi samasyao ka pata lagane ke liye bhautik vijyan se lekar bunyadi vijyan tak ke liye utsahit hain

SPEAKER_00: so the question here is do i start doing this as a business opening up a youtube channel on youtube espanol i think there is a separate like login and yes or something yeah yes whatever

SPEAKER_150: do i take the time to do that translate all this do the expense post it and hope that a business

SPEAKER_00: emerges and that an audience emerges or do i just wait another two years before every video on youtube just like it has captions is going to have the ability to dub it because that's what's going to

SPEAKER_122: happen isn't it like these tools exist now yeah one is about capturing the audience today you know if if

SPEAKER_20: you are trying to grow your audience as a creator these are barriers to that and i think even mr beast right jimmy talks about one of the things that was frustrating for them is people were ripping off his videos and just dubbing them and re-uploading them and you know profiting to it um and so now you can you know jimmy now has large infrastructure but you know you could think about it if you're a solopreneur or a small company and you want to get your word out there or you just you're a creator um think about it you know those i did four of those in four different languages i didn't even send the greek i did greek as well and uh just paying you know homage to your history jcow and uh um and you know i did them in like a minute each like and that goes to you know combining the conversation before the speed the ui like how fast everything's evolving it's incredible do we have a clip of mr

SPEAKER_154: beast in another language like is he actually doing that yeah he's doing that now officially

SPEAKER_00: i wonder if he's doing with ai or he's hiring people to dub because at his level he could just hire

SPEAKER_16: people to dub it and just give him a thousand bucks to be his voice or whatever or a production

SPEAKER_20: company probably cost i would say in the samir and colin pod uh there's some places where he's got hired actors that he actually went and met with so he talked about the whole experience and it's weird because he's used to listening to them and then so that's like their primary in the bigger markets and then the other ones they're starting to leverage ai great yeah i mean uh this is something

SPEAKER_68: i've been thinking about you know i you have to look and be cutthroat about your time and your

SPEAKER_00: effort would this be a good use of our time well we're not investing in companies you know in other non-english speaking countries right now because we don't know how to invest in those countries as well as the domestic folks right so china india japan you know we're gonna be at we're gonna get our butts kicked by the investors there who understand culture understand regulation understand founders understand society and products and the history of the products going to market there will never

SPEAKER_43: compete so you know i would be afraid almost to put this weekend startups in japanese and have too many japanese founders contacting me because i i can't work uh in japan right now i don't have like

SPEAKER_00: the infrastructure set up there so while this is great that it's happening i wonder uh about you know the scale you need to have now if you're uber or airbnb and you want to do

SPEAKER_43: uh an update to all your hosts and you've got some new regulations or something where you just want to

SPEAKER_14: have the ceo you know give a an inspirational message or do a q a man this would be incredible to just do a quick get it out your host in every language in every country you operate in so i see

SPEAKER_162: that i see a very big corporate value here but on the first one i disagree with you a little bit

SPEAKER_31: jcal because think about you know twitter now does the creator payments and if all you're doing is

SPEAKER_20: english all the only thing you can get paid in is english right because you're gonna lose out on everybody else from all those languages and so you know would your and i saw you posted something about a creator payment you got and yeah it's weird i didn't even know i was in the program i guess

SPEAKER_80: they selected some folks yeah you have some friends there i don't know i met somebody who works there

SPEAKER_00: yeah yeah exactly employee so yeah i mean it's yeah so i guess your point is free money why wouldn't

SPEAKER_43: you pick up the free money laying all over the ground and i do think that there's a possibility of that and so the question is how much would this cost to do an hour because when i got quoted this you know doing an hour of my podcast i think it was 500 to a thousand dollars to use actors to dub it and then i think the ai people were like it was 100 or 200 an hour to do this i wonder what they're at

SPEAKER_122: now which means it's going to be a dollar next year so what they've gone to here is um what it looks like is uh about a hundred dollars for 100 minutes a month is what they're asking for and then a dollar

SPEAKER_00: a minute so a dollar a minute means a 75 minute podcast is 75 bucks not bad but that's pro language

SPEAKER_67: that's about what it costs to transcribe a podcast before all the new tools by the way like five years ago the best transcriptions were about a dollar a minute if you wanted to do a transcription adam

SPEAKER_00: manila with somebody with english as a second language who doesn't understand you know the names or the organizations you're talking about will have some mistakes in it a raw transcript in other words it was a dollar a minute then it became free on youtube they got there first now if you use temi otter or descript it's it's free and included nick it's part of the subscription the standard

SPEAKER_179: subscription package basically for descript yeah and they're incredibly fast number of minutes

SPEAKER_57: yeah it's they do have a cap on it yeah yeah i think we pay like 120 a month for all in features

SPEAKER_69: and there's like no cap yeah um yeah incredible do you want to see by the way the mr beast dubbing feature it's pretty impressive i just looked it up yeah it looks like it looks like youtube is testing with a couple of thousand creators i believe you have to upload your own dubs i'm not sure if they mr beast uses ai for it but let me know if you could just hear this in english the latest videos can you guys hear that yep okay they have 30 seconds to get a question it's pretty impressive they youtube in the settings tab has a audio track dub and you could switch from english to spanish to vietnamese to french just like netflix yeah this is the french

SPEAKER_191: that's a human yeah totally wait a second they have humans doing this

SPEAKER_20: well like i said that doesn't sound like a computer no no so i listened to uh like i said a uh colin and samir podcast where they they he talked about that for these main languages because he went and met

SPEAKER_31: these folks that that do it and not only for him but for his uh his you know co-hosts right so they've

SPEAKER_36: got characters that act as all of them yeah but google is abstracting this for the top level creators which makes sense because google wants to grow in those markets so why not take the top

SPEAKER_43: 10 creators and then offer this for free for them uh probably cost them you know so that means google

SPEAKER_199: is hiring voice actors to do this or outsourcing it to some third-party company to do it it's pretty

SPEAKER_201: crazy when you think about it or mr b's upload his dub mr beast uploaded his dub yeah it looks like they're oh i see so they put it on mr b's that makes more sense yeah yeah got it so that would

SPEAKER_00: be the equivalent of like netflix you know buying a movie from you know an independent studio and then the studio has dubbed it and they put it up there and then when it gets bought by you know whatever some other regions you know the chinese uh version of netflix they can upload those same dubs

SPEAKER_209: there brilliant we all have self-doubt i can promise you if there's a hundred people in a room a hundred people have some level of insecurity you can't have any self-doubt about your commitment and conviction to do this you might have self-doubt about your ability to succeed to raise the money

SPEAKER_210: to attract the right people those are two different issues what an incredible clip self-doubt we all have

SPEAKER_213: it as founders as human beings and the voice you just heard one of my favorite business leaders i've been trying to get him on this week and starts forever somebody help me out here starbucks ceo howard schultz now i may not have gotten howard schultz to come on this week in startups but you know who did get him master class and all of the lessons he learned are sitting there waiting for you at master class in addition you can get chris voss on negotiations kim scott friend of the pod on radical candor bob eiger on leadership and so many more think about if you get one insight just one and it changes the course of your startup your life your career i bet you there's 10 incredible insights that will change your life in each course all of these legends are sitting there gordon ramsey on cooking serena williams on tennis steph curry on shooting and so much more just waiting there for you get unlimited access to every class and right now as a twist listener you can get 15 off when you go to masterclass.com startups that's masterclass.com startups for 15 off an annual membership masterclass.com

SPEAKER_35: startups all right let's go on to uh what's this pink floyd one okay yeah this this is interesting

SPEAKER_215: so let me let me just cue this one i love pink floyd yeah i know it's listening to the dark side of the

SPEAKER_98: mule dark side of the mule if you know you know okay yeah dark side of the mule it's a cover band

SPEAKER_00: that does a live version on new year's of dark side of the mule it's a uh live album by government

SPEAKER_208: mule and they just cover like the great songs by pink flood pretty dope yeah i want to give

SPEAKER_124: rowan credit again for this one because you know he found it and i and i went and read the article but

SPEAKER_20: basically yeah uh high level what they did was uh researchers at berkeley and you know we'll include links to the this stuff for folks but shout out to berkeley what they're able to do was uh implant some electrodes into a person's brain that had you know some issues regarding uh speech and they were able to recreate a song through a person's neural activity associated with the song and so wow i'm

SPEAKER_220: just going to play this here hopefully the audio comes through so the original song

SPEAKER_224: is this short 30 seconds sweet okay sweet okay and now this is the recreation wait a second what do

SPEAKER_122: you mean recreation who's recreating that an ai so here here's the article let me sorry i got a lot

SPEAKER_00: of windows open here so an ai created the brain wave version of that song so i could experience it

SPEAKER_124: no this is not so you could experience it so you know these are the electrodes that were hooked into

SPEAKER_31: the person's brain okay so it's a skull cap yeah yeah well no but these are actually on the brain it's not like so just they call out that it's not you can't read people's minds with this but if you

SPEAKER_20: actually can hook up into the brain an ai basically can recreate audio from brain waves of you recalling how you're hearing the song and so think about this oh i see that was the output of

SPEAKER_238: my brain that was the output of your brain yes oh so it's like humming yes and it yeah oh wow well

SPEAKER_24: it's really weird um what's the application of that i can replay my experiences what if you can't speak

SPEAKER_43: what if you have a speech a speech impediment so that i could have my computer i could read a poem

SPEAKER_00: and the computer yeah i would figure out what my brain waves are saying and then you could actually

SPEAKER_246: hear me read a poem or sing a song yeah or it sounds like it's underwater yeah it sounds like

SPEAKER_14: it's underwater or over like a phone which is exactly how like a lot of these things start out marconi sending telegraphs and it's oh you know yeah it's going to evolve from here you know last week

SPEAKER_20: we didn't get to talk about it but like you know there was like a negative uh p article out there of like an ai being used to listen to keystrokes did you see this one where yes explain that to everybody yeah yeah so basically someone had used you know one some form of of ai i don't know particularly which one it was to use uh an audio of you typing in your keystroke to basically determine what you were typing and you know the idea is that it again pattern recognition that's all these things are about right is you know understanding patterns so you know i guess when we type we follow certain patterns and if you give enough of that sample data to an ai it can just recreate whatever you're typing between telling the difference between how fast you type er versus et versus ep and what yeah

SPEAKER_22: exactly because you know the speed from those you know the distance those characters and all that matter

SPEAKER_43: and so um that was like i wonder if it needs training data from an individual or if it has training data from just you know i don't think it was a hundred hours of people typing the most common words i think

SPEAKER_00: it's a ladder okay yeah wow so then it what could happen is if and i think this probably already exists in the cia fbi musad everything kgb that they have ways to do this already because they also have ways to look at your phone and when you use your fingerprint on your keypad you could look on an angle or with certain you know chemicals figure out where you press the button most often from the oil on your finger to figure out figure out your code and so here you could test this so the

SPEAKER_259: researchers used a 2021 macbook pro to test their concept a laptop that features a keyboard identical in switch design to their models for the last from the last two years and potentially those in the future typing on 36 keys 25 times each to train their model on the wave storms associated with each

SPEAKER_00: key so that to me indicates it's not just the typing of um it's not just the typing of like common words like the um it might be that certain keyboards give off a certain sound for qwerty and each one of those

SPEAKER_36: might be slightly different that's wild or it could be the acoustics because if you had the speaker in the right spot it was stationary the time from for the q versus say the p opposite ends left and right

SPEAKER_00: of your keyboard could come in at a different angle because there was a company called shot spotter it was a really cool idea i didn't invest in it but it's active in a lot of locations i think they had oakland they put microphones all over a community when a gun goes off they figure out yeah and they triangulate where the within like a hundred feet where the gunshot was from really powerful and then

SPEAKER_36: they just send the cars right there cops go right to that location very powerful okay what's this uh

SPEAKER_233: god mode what's god mode yeah so this one i think you'd really like uh let me just pull this up as well so i'll pull up the github project so you know folks can um uh folks can see where this is from

SPEAKER_162: so let me share this as well it says github github is a repository of uh code and people share their

SPEAKER_249: code and uh so it's by the developer that created small ai which is uh you know a bunch of ai

SPEAKER_20: enhancements to help people be more productive uh and this god mode project is super interesting

SPEAKER_233: and i'm going to just pull it up here in a different share now which okay give me a second i'll pull it

SPEAKER_36: up and so this means by default it's open source there's some sort of license here anybody can take it and fork it create their own program and build on it this is the power of open

SPEAKER_249: source and and this is like uh repositories exactly and so what this is it's like a browser window that

SPEAKER_20: has all of the major ai's running in like a phone mode and so you know and i i didn't log into you know po here but like basically you have gpt on the left here you have uh claude here you have uh the bing here perplexity and po and it lets you basically just you know if you're working on a problem just uh you know try it across the different ai how do i get this on my desktop yeah so you can

SPEAKER_31: just go to that github url and you can download it's a there's a dmg like a mac dmg file that

SPEAKER_20: you can download and install so yeah it's uh it's just there i'll pull it up and show you there and so and i think yeah i thought you'd really like it because you can just go there and uh i do this

SPEAKER_273: already just cut and paste from one window to the other and i have a 50 inch monitor i have one of those

SPEAKER_20: giant ones yeah it's like a little form of uh it's like yeah it's like a little form of like i you know it reminded me a tweet deck a little bit yeah so you just go here yeah there's a url in the uh in the in the window here and yeah you download this dmg and you're off to the races

SPEAKER_31: you just got to sign in to all the different use cases and that's it love it yeah so download it try it have all your ais you know instead of being enhanced by one ai be enhanced by five of them

SPEAKER_16: what i would like to have happen is i'd like an ai on top of the ais so i'd like a meta ai

SPEAKER_00: that i could say it to once it would look at the five different versions and then it would look for differences between them and i'd say which one do you think is the best answer if i asked it to look for family offices that i could meet with to be lps and launch run four it would say yeah um you know here are the you know we came up with 117 unique ones here i deduped the list and given you the you know whatever yeah and here's the links you know it's a pretty great idea so and this is where the reinforcement learning um as these data lakes and these lms start to emerge and become refined

SPEAKER_14: right and we're still in year zero here i was like you're one of this you know by year three of this these things are going to be very sophisticated there will be ais talking to ais ais managing other ais baby gpts you know managing a process that you say hey i just want to know uh all of the important uh new restaurants in my neighborhood uh within a you know and these are my and these

SPEAKER_00: are my three favorite foods just read the reviews and bring me the best reviews and tell me where i should eat next and it's going to go out your personal ai maybe it's claude maybe it's another one and it's going to give those searches to multiple ais consolidate the results clean them up and then maybe put citations hey bard found this new sushi place you love sushi and these are the three best dishes here are the reviews of it boom and then when i talk to mine i just say give me uh five reservations for when i'm when i'm at lake tahoe next week i want to eat dinner i want all of them to be at the on

SPEAKER_199: the early side between five and 6 30 and make them all for five people boom yep and it just does

SPEAKER_20: leave right we've done this in definitive where you know we have an automated data science agent that has like 50 or 60 personas of different things that data scientists try and basically that's what it

SPEAKER_122: you know it actually does for that particular use case so um this is already here we're going to see

SPEAKER_287: it sooner than later oh wow hey everybody today i'm joined by root ceo dan dorfman dan welcome to the show thanks for having me jason tell everybody here in the audience what is roots

SPEAKER_289: and what makes it different than the other real estate investing platforms i'm a complete neophyte

SPEAKER_290: roots is a reit with a little twist sorry i had to do it we are the first real estate portfolio that we know of that builds wealth for both our investors and our residents and we've created a unique win-win model that creates partners and not tenants how does the resident get their equity

SPEAKER_291: do they put in 100 bucks just like i might as an investor or is it blended into their rent what we

SPEAKER_290: do is we do not take security deposits at roots instead our residents fund a roots wealth building account and invest in the reit and so from day one when they move into one of our homes they're actually an investor just like me or you would be an investor the second step of our program is what we call living it like you own it and so if they pay rent on time for three months in a row and then they'd be a good neighbor so like no noise complaints and then the last thing we want them to do is a quick walkthrough video of the inside and outside of their unit we're actually going to give them a rebate of rent of usually around 150 dollars into their investment account so just for being a great partner with us they're going to make 600 extra on top of the growth of the fund for

SPEAKER_289: helping us take care of the asset head to invest with roots.com twist to sign up and start investing today that's invest with roots no spaces no dashes dot com twist to sign up today

SPEAKER_43: the user interfaces haven't been built out yet users haven't been trained on how to do it just

SPEAKER_00: reminds me of like the early days of pcs dos windows where you just you know some of the stuff can be done it's just a little bit of work to figure out how to do it hey you uh built a

SPEAKER_71: uh little search project for everybody's favorite social network can you show that

SPEAKER_293: oh yeah um yeah let me finish my demos and pull that up next yeah but you know kind of leveraging a

SPEAKER_126: little secret but yeah a little secret yeah we'll pull that up as well let me do this last demo and then pull that up as well for you awesome yeah um this one i think you'd like it's kind of related to

SPEAKER_31: you know day trading so this is called pluto.fi and so it's a finance chat it's a you know a chat bot that's been you know built and customized around you know we talked about this everyone's saying oh these are just chat gpt wrappers well no like this is when you take a generative i endpoint and build something around it this is quite powerful and so in this particular case you know

SPEAKER_20: i've logged in and i started saying you know tell me something about give me a chart about ford and it did and then it gave me a little history and then it said hey create an automation for me that that buys ford when it drops by five percent and it creates that automation this thing has its own

SPEAKER_124: platform for uh you know basically putting money in so you can buy and sell stocks and so uh it's

SPEAKER_31: really really powerful right and um you know they've yeah i was really really impressed by it they have an ai co-pilot they have ai charts they have sort of you know all the different use cases of automations

SPEAKER_124: you can create and um i was very impressed by it you can chat about a company

SPEAKER_297: yeah well you think about you know robinhood did to e-trade uh what e-trade did to charles schwab

SPEAKER_00: used to have to call charles schwab on the phone e-trade made a webpage robinhood made an app and now pluto is making it ai based every time a new platform comes out and your technology comes out you can ask the question why now and you can create a new service based on that new platform and right

SPEAKER_68: now your choices are ar vr or i guess cloud computing is kind of over everything that could be in the cloud is in the cloud so it's it's ai or ar vr the two new platforms that you could build for yeah and so

SPEAKER_124: obvious why nows the why now and this is really well and i think you know jaykel you should give it a little try for j trading and and see i would love to yeah you know and it gives you a little analysis here it tells you how ford is you know doing over the last year compared to everything else and just i i thought it was really well done you know i was thinking about you know this is

SPEAKER_281: getting towards the end of the year and so i was thinking about uh doing because i did so well on

SPEAKER_259: my j trades i didn't sell anything because i was like this is all going up i got in at the right price there's no reason to sell here but i got my ass kicked by disney and uh one or two others

SPEAKER_199: like stitch fix i think didn't work out great and uh uh what was the other one warner brothers maybe didn't work out great but disney kind of killed me and so i was thinking you know what

SPEAKER_00: maybe i take some losses here take the losses for the end of the year taxes get some losses on

SPEAKER_68: my taxes and then i would redeploy that into one of the winners and start consolidating down

SPEAKER_305: into my winning position so i wonder if we if we if we gave it your portfolio how it would perform

SPEAKER_43: to those use cases yeah no that's what i what i would like to do is give it my portfolio and just

SPEAKER_00: start asking you questions like you know tell me about the debt of each of these companies and then tell me the debt ratio to their earnings give me the debt ratio to like so if you did that right now if you asked it tell me about the debt to earnings or debt to revenue ratios of disney and warner brothers and just see what it comes up with because you know you can do that on something like why charts which is pretty nice or yahoo is coming out with uh which is different than white charts as an independent company um yahoo is coming out with a new has a beta testing a new charting tool that adds like a lot of the features that white charts got ahead of mon that are really available on bloomberg terminals and other like more sophisticated paid stuff and just knowing what uh you know being able to quickly figure that stuff out and then say hey what other companies look like this would be interesting so what are the companies that are similar you said tell me i'm just trying

SPEAKER_310: it here what is the debt to yearly revenue ratio of disney okay what is i mean this is a question yeah

SPEAKER_313: yearly revenue ratio of disney that's it i'm just yeah sure i'll start with disney saving i mean this

SPEAKER_36: is going to confuse the heck out of it i don't think it's going to get a rant i think it's going to

SPEAKER_199: i mean and then nick if you could do this on white white charts maybe that'd be pretty cool if we have a white charts account still we don't so yeah this is interesting because it's going through

SPEAKER_20: do it on any charting software it's definitely going through like an agi style process here which is gathering its thoughts trying to figure out how it's going to get there and it's kind of um you

SPEAKER_319: know let's see let's speak on the current shall we no i'm not looking for you to be cute uh disney

SPEAKER_321: debt to yearly revenue ratio is a bit like a roller coaster ride at disney what is this what are they doing yeah it's elucidating yeah currently disney debt to equity ratio is 0.48 and is the debt to asset

SPEAKER_325: ratios between two three simpler terms every dollar of equity disney has about 48 cents in debt for every dollar assets it has 23 cents in debt that's interesting so if you thought about it like

SPEAKER_259: a mortgage uh or something or you know your own revenue uh to your mortgage ratio that's how it is so their market cap would be dollar of equity disney has 48 cents in debt so if i own

SPEAKER_00: a million dollars worth of disney shares disney has mortgage 48 cents of that according to this if this is correct but they must have put in something here be cheeky uh and make jokes yeah so i don't like that um just a note i'm not here to for comedy if i want comedy i go to netflix uh and watch chapelle or

SPEAKER_24: whatever so no bueno on the comedy let's just get dollars and cents here and remember it's a small

SPEAKER_332: world after all what are they doing over there a famously uh joking product famously jokey bunch the uh

SPEAKER_14: the wall street traders famously jokey yeah i mean if you want to post or you want to troll me on my trade i'm here for it but i don't want you to me yeah you should that's what i want i want and i want to be able to tell that ai to question every trade i make and give me the other side of it to steal man my trade and then to um straw man it criticize it whatever you know every possible

SPEAKER_00: analysis like um ventress ai is doing yeah just give me every possible analysis of my trade did i make a good trade or not who's with me who's against me what are their arguments right because like that's what i do with uber you know i keep holding my shares because the press says stupid stuff like uber will never be profitable on cnbc on everything you know these guys on the show or

SPEAKER_14: in the incubator for sure so all right they you know keep doing that and then you watch the analysts and they're like yeah you know here's how this gets to profitability eventually in the

SPEAKER_00: same way amazon did i'm like i would feel pretty good if i was a private investor in amazon and i held my public shares forever and here we are you know uber you know having his first profitable

SPEAKER_31: quarter so the analysts tend to go ahead jake i'll finish sorry i got one surprise for you then i'll show you no no i got a surprise for you all right nice um sorry let me cut you off because i want

SPEAKER_20: to get through these i think so jake you'll remember you this should look familiar for you and i can make

SPEAKER_342: make the font bigger here right and has a singular gold oh this is my book yes well this is a big problem

SPEAKER_00: right now is that people have started training open ai um has been hit with this yep there are authors and there are on the web stolen books right you know you go to pirate bay or any of these bit torrents whatever uh people around the world steal books you know it's like in the west it's like well

SPEAKER_199: why would i steal a book i can get it for 10 bucks or whatever audiobook kindles but you know if you're

SPEAKER_346: in china or you know you can't afford it and you're in an emerging market of course you're going to

SPEAKER_259: steal it so all of these have been ripped and are on the web in pdf form if you want to search for this you can just search that's where i got this i actually just i needed it yeah somebody ripped

SPEAKER_94: it but i'm going to play this for you honestly i mean it's so cheap to buy a book compared to the

SPEAKER_00: value that i i think that's they've they've priced it so piracy in the united states or the west

SPEAKER_259: doesn't even make sense like who wants to go find it and then use a pdf instead of use it in your kindle player yeah kind of like what happened with music like why would i even bother but now what happened was all of those since they're on the web the folks at the claim is folks that like

SPEAKER_00: open ai and we'll see if this is knowingly or just part of the generalized crawl looked for pdfs to train their data so you could say like oh just find every pdf on the web that you find in google drives and so one of the great ways hacking is going on right now or piracy is if you want to look

SPEAKER_14: for something you do site like docs.google.com or drive.google.com so people if you're looking for a movie and you were looking for blade runner you just do blade runner site colon docs.google.com and you will find like i just did it and i found the blade runner script it's a second result on

SPEAKER_00: google i'm sorry to expose like hacking techniques this is a pretty obvious one that most people don't know i think yeah so here's how to be a hacker and find pirate stuff sorry um type in blade runner and

SPEAKER_259: then do site colon docs.google.com and you can do this for dropbox or other public things uh and you find and you can also do file type i think in the operators in your google search but here you see blade runner the final cut that's my version uh blade runner esper retirement edition this is esper retirement edition is uh someone says no preview available i'm not sure why but here blade runner final cut i can just download it i could literally download blade runner right now for free um and here's the blade runner movie script so if they pointed if if openai or anybody else doing language models just did script script docs.google.com they could just take every script from hollywood and do this yeah or they could take an imdb database scrape the imdb get all that data and then say take every movie name then do the search movie name site colon docs whatever and then put that into my natural language model and we don't know for sure but like something like this has happened

SPEAKER_20: jacob because if you go into openai and ask it to write you like a modern seinfeld episode it can

SPEAKER_366: do a really good job yes and you know and very we know how it happened i just showed you yeah yeah

SPEAKER_20: yeah i just do keep talking yeah yeah and so that's where we have to get to that's why open source models are so important because you know everyone will understand the the training data that was used in them because everything is made available yeah so i mean you just go here and you can start

SPEAKER_36: finding all different scripts you know uh friends is a tough one to do because or seinfeld yeah

SPEAKER_99: tv show scripts season one see oh simpsons i mean it's just it's gonna happen everything so

SPEAKER_305: you have to do a little bit of refinement but you'll find it pretty easily yeah i think actually

SPEAKER_43: what you have to put in here is that it's a pdf so if you put pdf in yeah then it will probably find

SPEAKER_373: you do file type file type colon pdf yeah exactly all right here we go file type colon pdf or now

SPEAKER_374: they're filtering it out now they're like google knows you're trying something bad it's like telling

David Friedberg: you no uh well then you just use any other search engine yeah there isn't there isn't there is a correct me if i'm wrong there's an open crawl so there is an open source project to crawl the entire

SPEAKER_376: web and anybody can use open crawl what is it called common crawl yeah yeah who has access to common crawl

SPEAKER_246: you can just go download it anyone yeah but what is common crawl who who did you know like it's a

SPEAKER_122: non-profit i guess you can donate to it yeah it's a non-profit and it's uh i believe it's related to like the internet archives oh look yeah it's rich granta yeah i remember this yeah yeah i know these

SPEAKER_00: guys rich granta was doing uh he was doing his own like search engine thing yeah uh oh and gill is um elbaz is on the um he did applied semantics which was that was adsense bought by google

SPEAKER_199: so yeah yeah we should have gill back on the pod he's been on the pod so this is fascinating so

SPEAKER_259: what they do at common crawl is i'm sure read what they do everyone should have the opportunity to indulge your curious analyze the world and pursue brilliant ideas small startups and even individuals can now access high quality crawl data that was previously only available to large search

SPEAKER_36: engine corporations for more information on the corpus let's go to get starting our google group is

SPEAKER_109: active hub so what's really interesting about this is you know they started this before people were

SPEAKER_00: doing ai and um you know other people who were doing search engines or wanted like you know data about the web could use this but it was a very small use case now the use case is caught up common crawl.org that's it uh amazing look at this wow and i guess they do it look they and they have all the images of it like from 2014 on oh yeah yeah you could do really yeah so if they do this monthly what's really great about this is you could compare may june 2023 to you know pre-gpt you know go to 2019 uh and man wow that would be super powerful to look at the

SPEAKER_40: differences all right go ahead uh what do you anything anything else well i just i just wanted to play this for you get your take because i've been refining this voice of jacal a little bit well what

SPEAKER_20: about your little project i'm going to show that let me just about your little search engine let's

SPEAKER_109: do that god you won't take any credit for your work uh c minus student from brooklyn before brooklyn was cool clawed his way into the tech industry got lucky seven times and counting and made tens of millions

David Friedberg: of dollars yum yum that's it i mean it's 90 of the way there um if you put this on my website and

SPEAKER_00: did my blog post or my sub stack or sub stack built something like this in after and it just trained your data like this is a great sub stack feature it's a great twitter feature like why not just have twitter read my tweets out loud in my voice it's killer it's killer you know and what's interesting about how great that quality is i use something to called speechify speechify lets me send it a pdf so i'll send it like a long read you know an ft article a new yorker article so i can then listen to it when you know i'm going on a hike out here or skiing or something like that so before i go out skiing i may take three or four long reads put them into speechify and then play them in a playlist on speechify and i can pick barack obama's voice or gwyneth paltrow did a thing with them and she

SPEAKER_24: approved them so i listened to gpo big fan of the all in pod coming to the all in summit by the way talk about goop and uh celebrity and business and everything life should be a great talk uh so they haven't um pretty good i would love to see that reading that feature should be built into your browser so that's like essentially what speechify does for me is it builds it into my browser big

SPEAKER_397: fan of speechify i know shout out to the speechify team um they wanted to do my voice as one of the default voices i will totally do that nick um let's see if you can get in touch with the speechify

SPEAKER_14: guy and see if we can get that going again i want to be one of the voices be like barack obama they call it mr president then they have a picture of a cartoon character of a black president you know it's like i think this is barack obama and it sounds just like barack obama um yeah so great awesome uh because you know they don't want authors to read their audiobook did you know that

SPEAKER_122: do you know why no see i i didn't know that because i really enjoyed bill clinton's own

SPEAKER_366: uh read if it's somebody that iconic yes who's a great speaker okay okay most authors can't get

SPEAKER_00: through it it's too hard to read 50 or 100 000 words in three or four days i did it in two and a half days they begged me not to do it i told them i'm a podcaster people want to hear my voice not like some professional voice and they're like you're wrong you're wrong you're wrong i said i tell you what what's the real reason they said the real reason is most people suck and then they quit after two or three days and then we have to pay for that i said i tell you what i'll pay for it if i suck you can bill me yep uh and if i suck just say jacal you suck and then i'll take your word for it so at any point you can say this sucks and i'll literally pay for the studio time that got burnt and they're like not necessarily we'll give it a shot give a shot they were like you're great awesome but yeah this is amazing all right now you were playing with something i know that you have an affinity for the website twitter slash x you like to post x i understand yep yeah i do and uh the search on twitter has sucked since day one they had to buy a company called surmise back in the day to add search and then the search never got better like literally how long has twitter been

SPEAKER_405: around now is it 18 years 17 years the search has always been six yeah right yeah 17 years 17 years

SPEAKER_406: it's never been good it's never been a priority and they've got the greatest data set in the world and search sucks it's like what are you doing well now there's a new proprietor and i know he's working on

SPEAKER_31: it but you decided you would work on it yeah you know we spend a lot of time organizing data and

SPEAKER_20: understanding data and uh you know this url is not openly available so this is something that we built like for a demo which is basically yeah yeah for definitive and um and the idea is like look you know in the work that we're doing in helping enterprises organize data we took some initiative we we bought the api access and we basically built a natural language search and so uh just some examples here of uh you know obviously you know last week there was some stuff going on with uh you know barstool sports and day portnoy he bought it back for a dollar and you know if you want to find out what's going on you want to just go into twitter and type what's going on between barstool sports and day portnoy then you want to get sort of the summaries there uh you know it's a language model

SPEAKER_11: trained against the search results or you get the search result then you have the language model

SPEAKER_175: analyze it this is what's called a semantic search that's built for twitter data yeah exactly and so

SPEAKER_20: and it you know leverages some of the things that you know we've built and understand around like llms and just also vector databases and so you know how bad is crime in san francisco right now is it's just natural language you know what's hot in llms and so you know i actually use this when i'm doing research for for the pod i'm trying to figure out like hey you know what's out there and i can

SPEAKER_235: kind of find interesting things this is amazing type in um who are the knicks going to trade for

SPEAKER_199: okay i wonder if it will give me because that's what i'm always looking for next twitter i don't know you know some highlights workout videos whatever i just want to care about the trade market

SPEAKER_68: so i'd love to have like that trade running all the time um let's see if there's any if it

SPEAKER_80: could semantically figure that out back here oh look here we go yeah they did the inflexible after

SPEAKER_126: signing josh hart right and uh there's not there's not a lot so this isn't the full twitter data set

SPEAKER_374: this is not the full twitter data is only exactly so but you know that's the idea is that you want to baseball search this way without kind of doing the the broken things that we do somebody clip this

SPEAKER_424: and uh send it to elon we'll put it on twitter and put it on x and send it to elon all right

SPEAKER_426: everybody this has been an amazing episode of this week's service we'll see you all next time thanks sunny