SPEAKER_00: so this is from our friends over at deepgram i don't know if you remember this i'm just

SPEAKER_03: play a little snippet uh steve there's a thousand bloggers at microsoft and uh bloggers at people at

SPEAKER_05: apple i understand are not allowed to blog you remember this right yeah deepgram has released its new tool where if you give it an audio file they can do an intelligence analysis of it so you can basically have it break out the intents of that conversation so 25 they felt was about requesting clarification about blogging yep nine percent was about the rss feed 11 was about

SPEAKER_07: inquiring about podcasts and then the last bit was the zinger about the email oh ask for zinger email

SPEAKER_09: i love it that's hilarious you remember the end of that one i made the joke and the audience left yeah

SPEAKER_16: wow this week in startups is brought to you by linkedin ads to redeem a 100 linkedin ad credit and launch your first campaign go to linkedin.com this week in startups coda a new doc that brings words tables and teams together all your valuable data plans objectives and strategies in one place go to coda.io twist to get a 1000 credit and wistia is the all-in-one video platform for business with tools that help you create manage and measure the impact of your videos try wistia for free at wistia.com startups all right everybody welcome back to this week in startups i'm here with my guy

SPEAKER_21: sandeep madra he works at a company called grok yeah definitive intelligence was acquired by grok g-r-o-q

SPEAKER_25: q is your grok g-r-o-q is the grok that elon spun out from twitter so just so we're clear how's uh your

SPEAKER_29: two weeks into the acquisition how's it going so far i see uh the number of developers using the grok

SPEAKER_23: platform has gone parabolic or it's doubling every couple weeks it's going really really fast you know

SPEAKER_33: we are over 60 000 developers over 15 000 apps that's awesome yeah we are rocking and rolling so it has been really fun uh we're adding capacity really fast it is a bit of cat and mouse for us because every time we add capacity people consume it ah as fast as we can and so you know we love all the support we're getting i was actually at a hackathon yesterday uh that we we kind of co-sponsored with

SPEAKER_39: the mistral team oh yeah you know i saw that was that in the city yeah it was it's a great space

SPEAKER_05: there i don't know if you've been there shack 15 wow yeah it's it's the second level of the fairy building and i didn't even know this thing existed it's a little bit like a we work slash soul house so you buy a membership i did hear about this yeah and you know the here's the

SPEAKER_28: thing like there's so much space available when you go through the boom bus cycle and i've seen this before during the dot-com and the great recession people after the dot-com bust or after the great recession were like you know there's tons of office space for like 1200 bucks a month the five of us

SPEAKER_29: could be in a cool office maybe we'll spend 15 grand a year having an office it'll be fun right

SPEAKER_28: then it became oh my god you know my we work for like five desks like nine thousand dollars or something it was like yeah okay we're spending a hundred thousand dollars a year and this just makes no sense what are we doing yeah yeah and that boom bus cycle is now we're in that trial where people are like oh there's space available and it's free basically or close to free or yeah minimums and cost oh yeah i might as well get a space and so that's really interesting when you think about space people get inspired by space too they see a space they're like wouldn't it be cool if a

SPEAKER_53: bunch of people came here and did something like a restaurant this space was a 10 out of 10. it was a 10

SPEAKER_33: out of 10. it's the best space i've ever seen in the city for like uh startup slash kind of uh

SPEAKER_57: you know community vibes yeah so i think we're going to start to see a lot of that you know because

SPEAKER_58: why not we should do this we can start up's ai event there maybe oh yeah we should totally do it

SPEAKER_59: yeah absolutely easy to do we just pick a day and then we just put a bunch of speakers and we're done

SPEAKER_60: yeah yeah great idea some demos too so we'll get that going okay yeah actually that's what we do is

SPEAKER_62: just demo or die kind of situation a couple of news items jaco yeah there's news yeah okay so

SPEAKER_63: the news and then demos okay so quick news in the ai world yeah let's go so i would say there's like

SPEAKER_05: one interesting thing which is the ceo and founder of stability diffusion uh mad he basically um stepped down from the role and you know that that's that's one thing in of itself let's not speculate too much on that i think he said something interesting in his comment in stepping down which is he feels as though there is yeah and exactly this is it because you're not going to beat centralized ai with more centralized ai and so i find that to be a really interesting comment in terms of um you know stepping down and you know they're open source focused so i'm not fully sure well i can speculate on what he's trying to mean here but i want to see what what what you have to think when you read that

SPEAKER_74: comment because they're open source first okay i met him uh the gentleman in question briefly when i was in the middle east he was in dubai and we were at a oh okay awesome a salon together uh seemed like

SPEAKER_28: an interesting cat but yeah he's doing an open sourced image creator right and what is the market for these things i guess is going to become the question so i think we're past the honeymoon phase and we're now getting into brass tax and so what investors are starting to ask i believe is okay how many subscriptions do consumers and businesses need to ai okay if you are using notion i'm paying extra for ai we're using zoom right now we've got the ai summary ai companion i pay for everybody on our team to have open ai i paid also for poe and you know now i'm like i'm pretty promiscuous and trying all these things but i do think businesses are gonna and investors are gonna start to say okay what what's what's essential what do we need what are we going to use and maybe the idea is not many people need image generation maybe that's like a niche thing you know how often are you creating images maybe for fun you and i could sit there and yeah you know go crazy all day but i think maybe we only need two of these in the world three of these in the world maybe we don't need 20 people making image creators maybe there's just not that much work and so the one built into adobe's suite right they have one that's legal and built into there then there's dolly you know unstable diffusion maybe there just doesn't need to be that many of them and that's my read into his comments which is we can't compete with open ai that would be my read open ai plus microsoft plus google gemini or maybe the google apple alliance which seems to be emerging we talked about last week so if you have these two competing factions open ai plus microsoft and then over here you have google and apple and then the open source community and the google app and then you have the meta open source community like maybe

SPEAKER_25: there's enough going on here that we the world just doesn't need number four or five and maybe they

SPEAKER_28: yeah he believed they would be number five and you know that's like this i watched this happen with uber lyft doordash just eats postmates sidecar there there was a long tale of people going after that space and at a certain point consolidation happens and this has been a rapidly evolving space maybe

SPEAKER_07: we're witnessing consolidation happening or capitulation you know what that that was a very similar read to mind which is the big folks are already here you know it wasn't hard for them to get there from a

SPEAKER_33: technology standpoint and it's very difficult to be you know like you said third you know maybe not

SPEAKER_05: even third but fourth fifth sixth player because you've got the consumer folks covering it off you've got the original pro folks covering it off with adobe and then you have the up and coming new folks covering it off with mid-journey and so really like you know to your point that it's it's difficult to

SPEAKER_43: win in that space and maybe you know he'll come at this from a different point of view um as he

SPEAKER_29: rethinks uh what he's going to do next yeah or there could be something else to him leaving the company maybe people didn't want to work for him we don't know right you sort of speculated who knows

SPEAKER_28: why he's leaving the company when ceos leave a company there could be burnout personal issues that

SPEAKER_59: they've got could be you know bad news inside the company maybe people don't want to work for them

SPEAKER_28: maybe they're not a great manager it could be any of those things and so sometimes when people leave the playground they take their ball with them so he could be leaving the playground or being pushed out of the playground it's like i'm going to take my ball with me this space sucks you know but

SPEAKER_29: it might just be he couldn't compete maybe he just didn't have the energy to fight this fight for the next five or ten years which would be a mature thing to do if you don't think you can fight a

SPEAKER_25: dog fight for five or ten years to just be honest about it and just be like it's hard what's the point you know and sometimes you have to face that as an entrepreneur i've had to face it in my career many times where like you're playing a certain game on the field the game changes i did a print magazine and then the world didn't want print magazines and i did blogs and the world didn't

SPEAKER_99: want blogs and i did podcasts maybe someday the world doesn't want podcasts i want shorts actually

SPEAKER_93: the world wanted blogs you sold too early there yeah you know i had a little bit of scar tissue from not selling too late the first time so that was like my 18 months in i'm selling this thing and

SPEAKER_07: i'm getting my first check and then so i don't have to sweat the world wanted mahalo just google

SPEAKER_04: didn't want you to be it uh and now we have a version of it and with ai so and perplexity yeah

SPEAKER_93: and the process is using the same concept we're using just ai wasn't ready 12 years ago when i thought you know these 12 blue links would be better if it was like content plus services plus q a i mean

SPEAKER_25: i had the right idea i just was really too early and i think that's like you could forgive yourself

SPEAKER_28: as an entrepreneur i think is important message here sometimes you're too early sometimes you sell too early sometimes you're too early to the party sometimes you're too late to the party and who knows what's happening here yeah but it does seem like um maybe there just wasn't an opportunity maybe they didn't perceive there was a huge opportunity and also with the sale of um to microsoft of um hi yeah what was the parent name um last week inflection election yeah that feels like okay and we said this like how many large models do there need to be how many assistants do there need to be yeah you know it does seem to me like verticalized services in ai will win the day and then

SPEAKER_120: there'll be three or four of these foundational models and that's exactly what's happening right

SPEAKER_123: navigating the b2b maze can feel really tough huh you're trying to hit the mark with all those top

SPEAKER_125: tier executives you want them to pay attention to your enterprise product but where can you find all those big fish the whales the ones who call the shots and make the buying decisions for corporations for startups and everybody in between well here's where linkedin ads is going to solve that problem for you and i've used this it is one of my secret weapons linkedin means business business equals linkedin in people's minds when you're on linkedin you're in the business mindset so you're going to really be thinking about business products and services you're open to those opportunities and linkedin recently passed a billion users 180 million of those billion are senior executives 18 but hey we all know about the one percent 10 million c-suite executives that's your cfo cto cio these are the people who are always looking for a new product or service to make their organization run better but they are on linkedin that's why linkedin's ad platform delivers two to five times greater return on investment compared to other social media platforms so easy to understand why this is because this is where all the business people are and they're in that business mindset super easy call to action make your b2b marketing everything it can be and get a hundred dollar credit on your next campaign go to linkedin.com this week in startups to claim your credit that's linkedin.com slash this week in startups no spaces no dashes terms and conditions apply because they're giving

SPEAKER_05: you a hundy jaco you always have this magic to do it a great segue into just the last news piece because i do want to touch on it which is a16z at least a really good report and i just want to

SPEAKER_33: touch on a couple things for all the builders out there because we we do this for a lot of builders and and yesterday a lot of folks came up at the hackathon and they're like oh it was so awesome you guys do the demos so they're really cool to see it but you should read the whole report but i'm going to touch on three interesting bullets from this report and it's generated by the enterprise first you can see here the amount of average spend for llms is going from 7 million last year to 18 million this is for

SPEAKER_128: enterprise companies so large companies are saying their average spend in 2023 the average

SPEAKER_29: spend across foundation model apis self-hosting and fine-tuning models was 7 million of course dozens of companies we spoke to so they spoke to a couple of dozen companies yeah it does look like the average spend per enterprise yes from 7 million to 18 million is the estimate so the actual was seven so they're two and a half times in it what would this be the equivalent of the cloud computing revolution

SPEAKER_33: even faster right like you know this this is a cager which is like you know growing really really fast right and i think it's faster than anything we've seen before and look like you know they call it's like anywhere from two to five x right and more workloads going into production so this is good

SPEAKER_68: for folks that are building okay i thought this was interesting in terms of like where the allocation

SPEAKER_23: yeah where the money is for general where is the money for general ai coming from so they asked 25

SPEAKER_136: enterprises i'm guessing here it says dozens they didn't say two dozens they said dozens dozens of

SPEAKER_23: companies okay oh dozens of companies yeah so when you say dozens of companies that would mean like

SPEAKER_105: 24 36 companies probably so in the second one where do they allocate this from or reallocate so they're

SPEAKER_23: saying their the budget doesn't exist well one company cited saving six million for each call served six dollars which will serve by their llm power customer service for a total of 90 cost savings that's a reason to increase their investment in jenny i eight fold okay so they're using one anecdotal

SPEAKER_139: story there got it but what i would say here jay cal is when you're building in the earliest phases and

SPEAKER_33: i've seen this in you know two previous lives i've seen it when i had a mobile business and i had a cloud business a lot of the early budget comes from the innovation teams because they're the ones within the enterprise that you know are tasked with looking at this but to see that we're already at a place where 75 or percent of the budget is not coming from innovation this is really good got it so this

SPEAKER_141: isn't the innovation budget where like experimental stuff goes okay you know vision pro comes out

SPEAKER_28: ar vr okay we'll put a little bit of money into that okay yeah here 24 innovation 76 other business unit budgets 21 is the next highest that means somebody in a business unit like customer support said hey i want to allocate my budget towards this and away from something else reallocated from the

SPEAKER_148: it group 15 percent new net new it budget for llm so that's 19 so somebody said hey i need to increase

SPEAKER_28: our budget yes give me some money for this reallocated from other ai 13 r d product development 8 so yeah it's definitely immediately having an impact on enterprise is your point i would can't argue with

SPEAKER_07: that here yeah and i have one more that i want to pull up and then we'll we'll get into our demos here

SPEAKER_33: which is interestingly showing you know which models are enterprises using and so let me bring

SPEAKER_154: the share back here and by the way i found in the first paragraph here over the past couple of months

SPEAKER_23: we've spoken with dozens of fortune 500 and top enterprise leaders and surveyed 70 more to understand how they're using buying and budgeting generative ai we were shocked by how significant the resourcing and attitude storage gen i i had changed over the last six months though these leaders still have some reservations about deploying generative ai they're also nearly tripling their budgets expanding the number of use cases that are deployed on smaller open source models and transitioning more workloads

SPEAKER_28: from early experimentation into production that says to me that they know it's not ready for prime time but they're trying to make it ready for prime time so there's some like recognition

SPEAKER_05: how would you interpret that i would i would say that they are they're ready to push into production okay and these next two charts which will be the last ones and we won't go to the whole report but yeah this is really fascinating because you can basically see which model providers or enterprises using

SPEAKER_07: and you can see here that um you know overall um you know open ai still the clear leader which is you know which 100 of the enterprises

SPEAKER_21: uh open ai are open ai is that that's that's impressive so 70 of the 70 yes in fact 70 companies or 70 leaders said they were using open i yeah and 63 said they were using microsoft google no i'm sorry google google yes yeah um and 50 of it was testing 13 of it was production yes llama 30 testing 11 production

SPEAKER_23: anthropic 34 we're using but not in production uh mistral 17 oh here 17 so basically yeah people are uh

SPEAKER_141: testing everything and slowly putting stuff into market makes sense and then the last one and we'll

SPEAKER_05: go we'll get into our demos which is um right here open source is booming and you can see here where the shift is really happening um across you know this little uh pie chart okay so we have a pie

SPEAKER_20: chart here with three sections and the three choices are one for those people who are listening yeah

SPEAKER_33: yeah the the three sections are no plans to increase open source that's this uh that's 18

SPEAKER_05: then 41 say they will switch to open source when the performance matches and 40 another 41 saying will increase or already use majority open source this is really our so this

SPEAKER_28: means 82 percent of enterprises want open source they don't want a closed model yes wow you know this is so interesting because you remember in the early days of open source 20 years ago everybody was like yeah nobody wants to use it now enterprises are basically defaulting to it yes it is it's incredible what open source has done and this means you would be long llama anthropics um

SPEAKER_07: no anthropics not open oh no you'd be long llama llama you'd be long one of the mistral models mixed row yeah right um you would be short and then the open source google ones yeah and then there's an open

SPEAKER_28: source google one yeah it makes sense i think this technology is so powerful that putting your entire company in the hands of microsoft and open ai or google and google cloud some folks might feel uncomfortable with them having so much say in what they're doing right that's how i would read it and they want to be able to not be to not have that dependency right so that's what i think i would read in here is open source means not dependent not taking a huge risk which is so funny when you think about it because we used to have this expression nobody ever got fired for um selecting microsoft and then microsoft yeah so you know now it's kind of the opposite nobody would ever got fired for picking an open source project because with an open source project you know

SPEAKER_25: that you can't get rug pulled you can't get price gouging you know you're just fascinating which

SPEAKER_28: which makes it even more perplexing why open ai went closed because the statistics here are showing

SPEAKER_55: don't go closed right that's what the statistics are saying the statistics are saying the market

SPEAKER_59: enterprise at least the enterprise market wants this solution uh in a certain way and people are just ignoring them so there you have it folks the great ignoring of the consumer base okay listen i got a

David Friedberg: lot on my plate gotta do a couple podcasts i got found a university launch accelerator and all my personal life and you know things i want to do so there's just a lot but i'm able to manage it all with an amazing piece of software called coda coda is the all-in-one platform that combines the best of documents spreadsheets and apps here's an example we use coda to run the founder university every week we ask all of our founders to submit a progress update and then all of that goes into coda as a database and then we can sort through all of those weekly updates and look for trends and coda also allows us to send automated reminders to all those founders to send in their updates and then we track week over week growth by generating charts so if we see strong growth we'll reach out and we'll invest

SPEAKER_205: in that company and it's all done through this beautiful product called coda coda here's your call to action you can use this software to solve any business problem it is extraordinary and if you

David Friedberg: want that platform to empower your startup to strategize plan and track goals effectively you can get started for free coda wants to support founders so they can give you a thousand dollars in credit one thousand dollars in credits at coda.io twist what a generous offer it's a limited time offer so i want you to get it right now coda.io twist that means you can begin planning right now to make your startup just really tight coda.io twist to get started for free and get that thousand dollar

SPEAKER_05: credit can't beat the price i love coda and you will too you know there's there's this notion that

SPEAKER_33: is you know kind of cross-pollinated between enterprise and consumer i believe definitely in silicon valley more than other places which is moat and you know obviously in consumer that makes you know it's important but i think in enterprise the notion of moat being tied to something proprietary or closed is a bad idea but you know folks maybe tend to go in that direction versus something that is like you know performance or price you know so this lock-in thing i think silicon valley and sas and everything that's happened in the last 10 plus years are still

SPEAKER_21: working through this yeah makes sense all right let's do something else let's do it yeah of course let's go well i mean we're starting to see where the industry is going to wind up and it's it's you

SPEAKER_93: know which if you dovetail the two stories like stable diffusion then is an open source project

SPEAKER_28: with a hosting project that would argue that they were more valuable because enterprises would want the open source solution so there's something with the founder leaving that's a little bit wasn't

SPEAKER_29: that the same founder who um allegedly screwed his co-founder out of their shares and bought a bunch of the shares back from them there was there was some story related to yeah getting rid of a co-founder

SPEAKER_217: or something and then getting all their shares yeah so okay let's get started okay first one today i

SPEAKER_05: really like this one um they are under a lot of pressure today because it's it's been going viral what it is it's called gatekeep.ai okay and what they do is uh you you generate something i i did this already because it's it's it's been uh slammed today but you ask it for a concept they have some pre-built ones here uh i did one myself which i'll share here which is tell me about put options

SPEAKER_07: because i know you know i know i know how much uh jay trading loves you know it's trying to i don't do

SPEAKER_219: options trading i gotta get to that nancy pelosi level here of taking exactly just follow her trades

SPEAKER_05: there you don't even follow her trade so what i did was i just said tell me about put options and what this did was and i really like this a lot it created a little short video i'll play it in a second it gave me a chat about it and then it gave me a little uh summary down here in a transcript

SPEAKER_33: so i'm just going to play this for let me 30 seconds put options are financial contracts that

SPEAKER_222: give the owner the right but not the obligation to sell an underlying asset at a specified price known as the strike price within a specific time frame let's visualize how a put option works imagine we have a stock whose price is decreasing over time okay i'm not going to play the whole

SPEAKER_05: video yeah i think this is awesome it made this explainer video the graphics the voiceover and the

SPEAKER_64: graphics the voiceover and gave me a little chat so imagine you know you're trying to work through

SPEAKER_05: trying to understand something or you're doing something with your kids when they're trying to

SPEAKER_162: learn you know this is a real combination of the video the chat experience it's khan academy in a box

SPEAKER_233: it's like a khan academy oh yeah good way to put it yeah it's a icon academy let's try i want to try one

SPEAKER_93: um explain to me okay how a keystone in an arch arch works to me uh in uh an arch works it does take them

SPEAKER_236: a few minutes and yeah it's all good you know well because it's generating a video this is generating

SPEAKER_93: a how-to video that you could then publish on youtube and now youtube is going to be flooded

SPEAKER_28: with these and hopefully they're correct because if they suck then it's wrong information that's not helping so this is where synthetic data or the moment where data can be degenerative data revolution is going to pollute the hell out of the internet so it's almost like you want to have the human internet ends at 2023 and going forward youtube is going to be filled with this stuff now is this stuff been

SPEAKER_93: checked is it correct or not we're going to have to create a whole cleanup crew because as we know it gets the things wrong all the time and so what if this is wrong what if it explains the pythagorean

SPEAKER_28: theorem wrong what if it explains how you know the keystone works incorrectly and now we start flooding the internet i my replies are frequently on x twitter generated by ai i can tell somebody's got an ai oh really reply bot and it's like oh you enjoy beef ribs in texas beef ribs are a nice thing to enjoy i hope you enjoyed your beef ribs continue enjoying your beef ribs from terry blacks and i'm

SPEAKER_241: just like uh what's going on here yeah okay and it's obvious like we're going to pollute the hell out of

SPEAKER_25: the internet uh with this garbage and so i i'm just worried about youtube the incentives are

SPEAKER_28: really messed up right now so we had content farms which hollow was definitely not part of but there was like ehow and other people who were doing content farming which was just throw anything on a page get seo link it together grab the ads and if you don't answer the question or if it's incorrect that's even better because then people click on the ads so it's a frustrating page what do you do when you land on a frustrating page that's supposedly going to answer your question yeah you either hit the back key and go back to google or you click one of the google or the ads on the page yeah and so that's what led to the whole panda update and google not wanting to link to content sites that were scaling up content and it killed mahalo and a bunch of other sites along with it if it was accurate what was

SPEAKER_21: their issue with that their issue was that there were a group of people who had figured out seo and

SPEAKER_28: had to build a large number of content pages and that people would start going direct to those and they just didn't want these content networks to take over the top ranks and so they just said this is the maximum amount of traffic that can go to each site and they wouldn't tell you what was going on so we went from having like the nut as but one silly example how to cook salmon we had the number one result on and for some period of time mahalo just had you know like literally 25 000 people a day day in and out going to that page so we just started investing more in that page more in that page than we did how to cook cod of course yeah we do like little how-to content and then links to the best stuff on the internet but yeah they started to consider this like a little bit competitive and then they started taking from other sites content and then putting it in the one box so their vision of the world was don't go to these pages we'll just put the information on the google search results so they started getting competitive and you see that today where google is competing heads up with a lot of the sites that they used to link to so you used to get stock quotes sports scores recipes shopping at other websites those got built into the google search results right if you ask for

SPEAKER_148: yeah yeah yeah you know if you ask for a stock price google just gives it to you if s for sports score it just gives it to you if you ask for a recipe it just gives it to you or if you're asking like what temperature should medium rare salmon be it just gives it to you yeah yeah and so yeah you

SPEAKER_28: know this is the pollution of the internet is going to be crazy in the coming years and that that's where figuring out if this is in fact ai generated or human generated and then as long as it's correct then it doesn't matter but that's the question i have for you is when when do we think we can start trusting the results from ai and in what sectors it's just the the amount of wrong

SPEAKER_25: information is the challenge i have right now with fully embracing a lot of this let's do it let's do a

SPEAKER_05: real-time test okay this it just got generated this is the one that you asked how a keystone in

SPEAKER_181: an arch works yes here's here's the summary okay the keystone is a wedge-shaped stone placed at the

SPEAKER_265: top of an arch it's rolls you should wait evenly a little bit faster here let's play faster just

SPEAKER_267: here we go welcome to this video explaining the importance of the keystone in an arch an arch is a curved structure that supports weight above an opening and it has been a crucial architectural element for centuries let's first look at an arch without a keystone the stones in the arch push against each other and rely on friction to stay in place it's not an arch now let's introduce the keystone the keystone is the final wedge-shaped stone placed at the top of the arch when the keystone is inserted it locks the other stones in place by distributing the weight evenly the keystone that is not even close

SPEAKER_99: outward providing stability to the arch bunch of bricks on top of each other totally well i think you

SPEAKER_270: didn't get the drawing right but it's it's concept is correct but it's concept is correct so proving my

SPEAKER_266: point is crucial for maintaining the integrity of the arch it got this is the final piece utterly

SPEAKER_28: in terms of drawing it it got the other parts correct yeah yeah i mean for anybody doesn't know if you think of an arch there's a keystone that's the yes middle stone all the stones lean into that and all the pressure goes into that stone and keeps the arch up and for those of you who are listening not watching this video just showed a stack of squares yeah with a triangle laid in the middle of them it's exactly the opposite of how it works so gatekeep.ai great but i guess garbage in garbage out or good data in garbage out like it's almost like we have a problem here where you get good information into these llms and then they screw it up so still work to be done here i think it got the information

SPEAKER_23: from wikipedia relatively well so the copy wasn't bad but the drawing was terrible and wrong so i don't

David Friedberg: i don't trust any of this stuff right now okay so your boss tells you you got to make a product tutorial and a customer support video and i know what you're thinking i need a video producer i need an editor and oh man i need a new boss well think again with wistia anyone can make professional quality videos right in their web browser do you need a script just use wistia's ai script generator and it will work magic need to be on camera use wistia's built-in recorded to capture your face your screen or both at the same time and when it comes to editing wistia's text-based editor makes things a breeze just highlight the mistakes and hit delete take out those mistakes the ums the ahs the hiccups whatever you got you can edit it in real time and export your video you can even add music with wistia's smart tracks and these are composed specifically for business videos so they're not going to be corny most importantly they're royalty free and they automatically adjust to the length of your video seriously if you need to make polished videos fast wistia is your go-to because they make amazing products for startups so i want you to go to wistia.com startups even we got a big company they're great for big companies too wistia.com startups yeah i mean look uh conceptually the way

SPEAKER_05: you put it which is like kind of an on-demand con academy awesome the idea is very powerful i mean this just yeah it's got to be right so i think this is an interesting spot where there could be a

SPEAKER_33: convergence of human you know what what do they call them on uh on twitter or x now the people that do

SPEAKER_07: those um oh the community notes yeah yeah community notes yeah like wikipedia editors so yeah you need

SPEAKER_21: to have teachers authority you need to have humans who can say this is correct getting into this model i

SPEAKER_287: give this an f for the output and an a for the concept okay okay so what is it average youtube

SPEAKER_289: i give it a d as in it's dumb it's a dumb answer i give it a d okay as in dumb well the answer was

SPEAKER_05: right no the answer was right the image was wrong it didn't okay and in learning an arc there's plenty

SPEAKER_130: of videos on youtube right now where you can learn it so this is d for dumb i give it a d

SPEAKER_123: okay what are you okay you'll make you um no no i think the concept's in a okay i love the concept

SPEAKER_94: that it would you know make me a unique video of anything i'm trying to understand it just it's worthless if it's not correct it's actually harmful yeah i might go to an f now i'm talking myself back

SPEAKER_07: oh my god i think um you're so kind it's yeah it's no it's it's hard to build these things shake how you're yes i know um maybe i'm frustrated on a monday yeah exactly um you know for me

SPEAKER_05: it's a c because you know if there's not a way you know for the system to tell you this might not be right and so they should maybe provide some some type of score on that one to say we're not really sure because um like i said i was playing around with the pre-built ones and the stock option one was pretty good um but you know that one wasn't and so uh but that's that's an easy thing for the team to iterate on is provide a confident score or something like that and look they did do one thing and uh when i refresh the site they said they did change their model because they were getting overloaded maybe their older model is not running as well they might have gone to a cheaper one since

SPEAKER_141: everybody came there to make their anyway i'm giving you guys and gals whoever's over there say them

SPEAKER_28: the whole group that's a unit it could be one person it could be a hundred but i'm going to give

SPEAKER_305: i just want to share it it says right up here on this you see that right yeah yeah yeah our servers

SPEAKER_123: are overloaded so anyway if it was correct i would have given this like a b plus an a minus like it's

SPEAKER_28: really a great compelling concept so i i but i give it a d on the execution but i think the concept is wonderful so i grind it out and get the results to equal the vision like anything else listen

SPEAKER_99: first tesla was a little rough too man yeah the roadster was a was a bit of a rough experience uh

SPEAKER_33: and uh you know no charging network none of that stuff right you were just no supercharger you know

SPEAKER_28: 200 and something range it was a little rough you know when you hit a bump pothole in the roadster you'd feel it in your teeth you know it was a boom it was a hard experience you know i drive that car

SPEAKER_313: sometimes and then i drive my model y and it's like oh my god you want to talk about going from like

SPEAKER_314: a d to an a plus like yeah there you go you know the roadster was a d like it was it was i mean but

SPEAKER_105: it was also delightful in that it was like so peppy or whatever but you do have to make it comfortable

SPEAKER_28: for everybody and it has to be accurate and safe and yeah here we go so you're on your journey i think

SPEAKER_25: how long before these things in your mind start to give enough consistent good answers that you

SPEAKER_321: would trust it to teach your kids so when would you trust it to teach your kids for the sats so here's

SPEAKER_05: my issue i think if you asked me the question three months ago i would have given you my really conservative oh sorry my aggressive answer like oh within six months this will be solved there's a weird thing that's happening where and i'm probably not like uniquely positioned to say i'm being fair about it because of you know what i've been working on for the last couple years yeah where so many people just blindly rely on whether it's chat gpt or which llm of your choice or something like this that it's there's like this virtuous cycle of people saying oh you know what let me just count on the llm to produce whatever it does and i think now there's so many of these things out there that do that um it's been stretched out so i would have said by the end of this year we'll be in a spot where you can trust these things and it'll be grounded um now i think that's going to be harder because there's so much being created that we have to go through the first arc of many people um kind of dealing with that and it's a unique thing for computers because we never really count on computers to be fundamentally probabilistic right you know you don't break out a calculator and go like that might not be

SPEAKER_28: all the time you know so yeah no i mean the whole this is the conundrum of this technology is we're so used to that if you put numbers in a spreadsheet there is no way for the spreadsheet to say you know

SPEAKER_240: cell a1 plus cell b1 equals cells c1 like there's the formula can't be wrong it just it's not possible

SPEAKER_93: for the computer to do a wrong formula in excel or yeah for spell check not to work like we

SPEAKER_25: it's fundamentally when you hit the you know when you type the word the we expect it to be correct and you know if it's not correct like your keyboard's broken like something is something really wrong that's easy to identify here it's like you you get these hallucinations or whatever it's it's like

SPEAKER_28: it's almost like a a drunk human you know like if you were to had a great let's say you had a great

SPEAKER_54: history professor and they got hammered yeah so like you were listening to them like they gave a great talk like and then that night they got bombed and you're at you know uh the kitchen table and

SPEAKER_45: they're drunk out of their mind telling you history and it's not accurate yeah but during the day they

SPEAKER_94: told you really accurate stuff like that's what we're trying to deal with this is jekyll and hyde

SPEAKER_333: i think we we got a lot of work to do here i don't know if it's two more jumps or three more jumps but

SPEAKER_28: yeah we're not being i think harsh enough on accuracy and so this needs to be the year of

SPEAKER_25: accuracy i'm gonna dub this the year of accuracy i'm looking at every single demo going forward

SPEAKER_28: accuracy is my number one thing i'm not wowed by it anymore i'm not wowed by what i just saw i like the i'm wowed by the vision but i'm judging things on accuracy going forward i'm just letting people know

SPEAKER_339: starting today march 25th 2024 i think when it comes to building quickly and like you know putting things out there when it comes to accuracy that's something you really should like you want to go

SPEAKER_162: fast but i think on accuracy it's important time for accuracy to play a role here yeah more no more

SPEAKER_145: gifts of uh you know a pluses but you you got okay not accurate i want to see a little accuracy okay

SPEAKER_33: let's go to another okay okay so the next one you'll hopefully get a kick out of this one jake how this is a good one so this is from our friends over at deep gram and i don't know if you remember this

SPEAKER_01: i'm just play a little snippet uh steve there's a thousand bloggers at microsoft and uh bloggers at

SPEAKER_04: people at apple i understand are not allowed to blog and okay you remember this right you're uh basically

SPEAKER_05: asking about blogging at apple right yeah and what was going to happen so deep gram has released its new tool we're going to give it an audio file what they can do is they can do basically an intelligence analysis of it so you can summarize it topic detection intent detection and sentiment so i took that video and basically you know obviously you have the summary which is the speaker discusses the

SPEAKER_33: issue of blogging at apple the need for employees not to spread false information ironically right yep

SPEAKER_348: um they also asked the other speaker if they can add a feeder into their podcast like i think you were

SPEAKER_179: talking about something about open i was talking about itunes uh supporting podcasts back yes 10 plus

SPEAKER_05: years ago exactly so now what they have this tool is so you can basically have it break out the intents of conversation so 25 they felt was about request requesting clarification about blogging yep uh nine percent

SPEAKER_07: was about the rss feed 11 was about inquiring about podcasts and then the last bit was the zinger about

SPEAKER_09: the email oh ask for zinger email i love it that's hilarious you remember at the end of that one i made

SPEAKER_54: the joke and the audience left yeah wow yeah this is really scary how interesting this is and then you

SPEAKER_351: have word level sentiment wow yes so negative positive that's interesting yes yeah so this is i think really

SPEAKER_123: important um this is really cool yeah because you know you could take an entire podcast yes and run

SPEAKER_28: it through this and you know start to understand what i and people are doing this for sales calls

SPEAKER_25: obviously with okay things like gong or whatever yeah so you know i feel like this is another one of

SPEAKER_28: these really cool tools for you to understand how you're being perceived and um i could see this working really well we do 60 or 70 calls with founders per week and we invest in okay on average two companies a week right um okay so if you're doing 70 calls um you know and we're doing two investments a week you know it's like one in 35 eventually wind up getting an investment now they had to beat out a bunch of people to even get on that phone call yada yada we have the accelerator but we now are recording these calls um and transcribing etc so be really good to have the associates do this and we could see yes you know both the associates at our firm we call them raas uh researchers analysts associates those are the three levels you start as a researcher out of school you become an analyst after 500 falls and

SPEAKER_25: then you become an associate after a thousand calls or something like that all right is the program uh created at launch so i would like that launch.com here you go i would like to actually have uh run some of my raas through this and and see how they are perceived if they're perceived as smart or not

SPEAKER_28: and then i the way i hire people now is i have them i have somebody on our team pitch i do a mock pitch and i pay people a hundred bucks who are considering hiring to do a call and we record it where they're give them the script of how you take a pitch call right an introductory call so we pay them for that but i would run it through this now so i'm going to use it i give this an a um this is a for awesome

SPEAKER_180: yeah and awesome yeah a for awesome okay great and you can see and it's and it is accurate yeah and

SPEAKER_07: you know that one's particularly interesting because it shows how the sentiment changed over time throughout the you know very short conversation it was like a minute long yeah q a and you know

SPEAKER_05: that aggressive and then obviously you know the uh steve says um hey you'll be able to add whatever

SPEAKER_07: you want we're working with open standards and yes uh ends with you know the email comment so yeah i

SPEAKER_05: thought this was really really great really cool there's a lot of other things you can define custom intents that you want it to detect as well so if there's in your case if you're there's like a certain type of intent like question or uh you know listen you know things that you're focusing on

SPEAKER_07: with your um employees very very cool i mean i want to i want to keep the calls positive

SPEAKER_28: um and i want to keep them moving quickly right and keep them positive and be respectful of people's time and also make sure that we under my most important thing is making sure that the uh associates and researchers at our firm understand the product so the founders feel that we understand their vision and so i have them say at the end of the recording okay can i repeat back to you um your vision for definitive intelligence i just want to make sure i understand it let me know if i got it right and by adding that to the script i got rated higher by founders on our calls because sometimes founders were like yeah i don't know if jcal understood our vision now if you don't invest in somebody they're like they didn't invest because they don't understand our vision and how genius we are i get it like i'm an entrepreneur too um but if i say at the end of the call hey just to be clear sandeep your vision for definitive intelligence is to use ai to sort through enterprise companies and their big data sets and help them leverage that to achieve their business goals and you charge them a fee and the fee's not cheap it's a couple hundred thousand dollars a year you would feel like i understood what definitive intelligence did yes and so they're my custom thing here would could we tell sentiment of feeling understood feeling heard oh so you would put custom intent in there my custom intent would i want the founders to feel that we understand them or they feel heard so i

SPEAKER_374: don't know how you're high on the empathy you're high on the empathy these days well you know your

SPEAKER_28: reputation in venture is really important and you know you have moments in this job where you have to be a bit of a hard ass you know like people want to sell their company and they put like 90 of the deal value in an employment contract and then screw the investors like we deal with all kinds of weird stuff that occurs you know people rent themselves an apartment for 20k a month and you know you're like what are you doing like this is not ethical or it's you know would be perceived as like impropriety and

SPEAKER_25: so when we don't have to when we don't have to deal with those situations i'm trying to have like a very high level of empathy and connection with the founders yeah i know that sounds groovy and hippy

SPEAKER_86: dippy but you know there are unique individuals in the world founders right they are and they're

SPEAKER_07: unique you know we but feeling heard like having been on all sides of it right having been a multiple

SPEAKER_05: time entrepreneur having an investor having been you know selling companies i've acquired companies i think that little trick that you're talking about is really important i've felt that frustration many times i felt it selling a product to someone not feeling like the person i'm selling to is not

SPEAKER_33: understanding felt during a pitch to people um so i like that um okay i'll give these devices i call

SPEAKER_28: these devices and when you have a device like this i try to make devices like we're doing here when

SPEAKER_21: we give it a rating a device you know can in a in a process or a system these kind of devices in a system

SPEAKER_28: or a process can really manifest like um a certain brand right so if you even have the concept of an introductory call we frame it as an introductory call because it's not designed to be the the be all end all it's just to introduce you to our team and what your vision is and then for you to understand who we are and how we invest oh we're not series a leads we're seed okay we have programs right so sometimes we need to you know help with them understanding what we do right so that's why i created the concept of the introductory call and the device of let me repeat back to you the vision

SPEAKER_25: so i make sure we understand it i do it myself now it's like slowing down to speed up in a way you slow down and you reflect back to the person what they're saying right it's also a good way to

SPEAKER_05: be a good partner uh to a spouse or to be a good friend hey it is the other thing is reflection is a big thing in ai as well um and i'll explain and i explain why when an llm starts to produce output right it's you know predicting next tokens but one of the things it doesn't have is a backspace key so once it starts down its journey and if if in its prediction tree it makes a prediction that's incorrect it just has to continue but when you ask it are you sure are you sure it gets a chance to run it again and take a different path and so the same thing works with uh with llms yeah so that's

SPEAKER_28: interesting so confirming are you sure like if we did that when we did the keystone thing yes in the

SPEAKER_21: training if we said are you sure that's how a keystone works and we're like let me go check

SPEAKER_05: but yeah in the inference again let me check because i i kind of made a mistake and i just have to continue with it because i've absorbed and there's no way to stop because it's just streaming the tokens at you so it's an important concept it's reflection there's some good papers written about it we'll

SPEAKER_393: drop a link to everyone yeah awesome all right let's do one more demo here again another demo yeah

SPEAKER_394: i do okay i i'm gonna give that one an a as well though so that's a a for both of us yeah i mean

SPEAKER_21: it's so awesome i'm a sucker for these kind of tools that help you understand humans more and it

SPEAKER_28: seemed i give it the a for accuracy because i know that was me talking i know steve jobs replying i know yeah you know it was walt mossberg you know teeing it off to me like i i've seen that clip a couple

SPEAKER_05: of times i i know it's accurate so it felt very accurate so credit to them okay we are going to close out with our friends over at anthropic okay and so um you know i have this queued up and here we go i basically have a famous picture of uh you know michael jordan uh oh the dunk yeah it's a famous

SPEAKER_401: dunk of his where he takes off at the free throw line yeah sam down competition you know we've all

SPEAKER_05: seen it and basically the the team over at anthropic have launched a new model uh or a variant of a model in this particular one it's called claude 3 haiku so it's smaller and faster and it's multimodal so in this case what you know uh for folks that are listening uh and i can run this again just real

SPEAKER_43: time here and you can see it's going to come back pretty fast you said describe this image and it was

SPEAKER_148: instant the image depicts a professional basketball player in midair dunking a basketball through the hoop during a game the player is wearing a red jersey appears to be in the midst of an impressive dunk and with the crowd in the background visible and cheering the court and basketball arena suggests this is a high level basketball competition likely an nba game the dynamic action and energy

SPEAKER_152: captured in the image convey the excitement athleticism of the sport and so um what what we

SPEAKER_05: can see here is you know what we said at the end of last year that multimodal um ai is going to become really important and uh you know these folks have a model it's fast um it's accurate um and it's incredibly useful for where things are going now and what i wanted to really talk about with respect to this because we've seen a few of these now this is one of the ones that's available like hosted along with like gptv i want entrepreneurs who are building to think about this in the following way

SPEAKER_07: if you were to oh let me actually find another image uh really quickly and pull it up well let

SPEAKER_20: me also ask you a question while you're doing this if you were to increase temperature and the max

SPEAKER_25: tokens to sample um there's two bars there explain to the audience what that would do here yeah so

SPEAKER_05: temperature is going to increase the variance in its response right it's going to basically it's going to make it much more i would say uh i'm just going to just try to describe it when it's going down its path of probabilities it's going to make it a little bit more random and sometimes it won't always pick the highest probability that's what i would say yeah okay and so if we run that again and then i will pull up another image here as well so we'll run this uh and so 0.6 it's kind of very

SPEAKER_406: similar but i'm going to pull up an image yeah 0.6 and this model haiku is the this is their cheapest

SPEAKER_25: fastest smallest right like this is there you want to grab an image and describe it real fast this is the way to do it this is the way you want to do it this is exactly the way you want to do it okay so

SPEAKER_05: what we're going to do is and i want really entrepreneurs to think about this particular example i've got this image of a nice healthy food bowl that is a healthy food okay and what i've done is i'm saying describe this image and return the results in json format include the type of food the size and calorie count and we're going to do this and why this is really important um uh this one knows

SPEAKER_303: i okay i would say um i want more details okay um so it simplified oh let me turn the temperature down this is probably why as well let's take this here let's run this again um why this is very powerful um i'm going to do this one more time hold on okay okay each ingredient and for people who don't

SPEAKER_418: know json is a open standard file format yeah human readable text to store and transmit data in a radio

SPEAKER_303: right okay so here's why this is important if you were to try to do this a year ago you would have to

SPEAKER_05: have had something take the image then process the image and then go through an individual element on on each thing that was there and then cross-reference that against another database with food information right now all of that's been collapsed into a single model which is basically saying telling me there's about a cup of roasted potatoes in there sweet potatoes calorie count calories it's telling me

SPEAKER_07: there's about a half a cup of sauteed mushrooms um there's some of you know friedberg's favorite it's a half a cup of quinoa in there there's half an avocado some roasted tomatoes and some turmeric roasted cauliflower about half a cup how does it know it's turmeric because of the color yes

SPEAKER_05: yeah so wow what why this is amazing is if you're now developing something you don't have to use a whole bunch of services inside aws or gcp or azure to to decompose the problem you can basically have the llm decompose the problem for you yes and so you've if you wanted to make uh like the next generation calorie counter app it's literally maybe five lines of code to pull that off amazing and you

SPEAKER_74: know if you also wanted to take uh i mentioned this like every charlie rose episode ever you know and

SPEAKER_25: understand like who's a film director who's a an author who's a journalist and then you know start to understand what topics they're talking about then looking for those topics you know in history in journalism like stitching this stuff together is getting faster and smoother you don't need to do multiple choices like a quentin tarantino interview on charlie rose in 2002 would include these films but it wouldn't include the following films a 2010 interview would include those and so just understanding you know the 17 seminal charlie rose interviews with quentin tarantino or just quentin tarantino generally putting those all together and then understanding the arc of what he's talking about what topics over time and being able to pin those to uh his filmography etc all that's starting to be

SPEAKER_05: included in these models yeah yeah and you know what i think for developers and i saw this at the hackathon i was at yesterday to close out is what you can pull off in a short amount of time has been fundamentally rewritten i think i'm excited to do we should do you know like we said a demo hackathon thing together jcal and i really want you know folks that are out there that are building and i know many already are but if the folks that haven't made that jump yet it's easier than ever to

SPEAKER_432: do these things you're going to be able to really solve a lot of problems very quickly uh as these

SPEAKER_28: platforms you know just keep extending the data and the reasoning right i mean the reasoning part of this is super interesting because you could ask it once you have this information you could say what should be removed from this and what should be added to make it healthier yeah right so if you could say like hey um optimize this salad for me to reduce the amount of calories in it but increase the amount of fiber you didn't ask it for fiber but we could get fiber and you could say i need to you know get more

SPEAKER_25: fiber in my diet and i need to reduce you know the number of calories so what what's a way to do that oh you know get rid of the quinoa and go with china doll you know like here's a different this would

SPEAKER_160: be the easiest way to solve that problem right this is just absolutely astounding and how fast it is and

SPEAKER_21: how cheap it is which means i think on a venture basis we were talking about this um 40 billion dollar

SPEAKER_25: fund that uh the pif was considering doing in ai and how we might allocate it it does seem like the fundamental language model it's like they're getting cheaper and cheaper i'm wondering like how much money is to be made there i mean i like your business the hardware business the hosting business running them but the models themselves if they're open source and they're going to trend

SPEAKER_05: towards free right like how much is it going to cost to do this i think you know the way the allocation is being put out was accurate i think it's like just give developers access to these models and you know let them build and i think the market will stabilize itself when it comes to who's running them

SPEAKER_25: and who can run them the cheapest and fastest it's going to be amazing i mean they yeah the the cost plummeting and the quality increasing means this accuracy issue i think it's solved i don't know when but i think today's theme is accuracy and um yeah so i'll give this one that seemed highly accurate to

SPEAKER_418: me i'm going to give this one uh a b plus yeah just having not tested it across many things but i

SPEAKER_444: just feel like uh yeah significant i think they've done a good job they've made it fast they've made

SPEAKER_05: it accessible um and now we just want to see a lot more people building with it i think that'll be really exciting keep building with it keep building and you know the number of developers is going to

SPEAKER_74: greatly increase i think with these tools being able to teach you to be a developer or be a maestro

SPEAKER_25: amongst a bunch of you know agents that can write code so i'm really excited for that i know that's going to take a little while but i really want to see the number of people who can write code or push code let's just say push code because they're not writing it but who can oversee code production i want to see that increase you know and the velocity of new products increase all right man it's been another

SPEAKER_21: great episode we'll see you all next time follow x.com sundip x.com jason you get us uh this week in startups.com bets to see the bets we've placed on our ai future in this week in startups.com ai to get the ai plus playlist go to this weekend startups on youtube so you can watch the video and

SPEAKER_28: see the keystone not get placed in the right spot if you want to uh great job studying and we'll see you

SPEAKER_257: all next time bye bye awesome bye