In short
This Week in Startups - Episode E1811 Summary
Episode Title
OpenAI and Google race to launch multimodal LLM plus AI demos with Sunny Madra
Hosts
- Jason Calacanis
- Sunny Madra
Episode Overview
In this episode, Jason Calacanis is joined by Sunny Madra to discuss the rapidly evolving landscape of artificial intelligence, particularly focusing on the race between OpenAI and Google to launch multimodal large language models (LLMs). The conversation touches on various aspects of AI, including user interface (UI) developments, generative AI, and practical demos of AI applications.
Key Discussion Points
- All-In Summit Recap (1:34)
- Sunny shares his experiences from the All-In Summit 2023, highlighting the quality of speakers and discussions.
- The need for better organization in terms of ticketing and attendee management is noted.
- The Race for Multimodal LLMs (9:07)
- OpenAI is striving to beat Google in launching the first multimodal LLM.
- Multimodal Definition: Ability of AI to process and generate multiple forms of data (text, images, audio, etc.).
- Current AI models are often specialized (e.g., GPT-4 for text, DALL-E for images).
- User Interface (UI) Considerations (33:00)
- Discussion on whether generative AI requires a UI shift.
- Importance of intuitive interfaces for users to interact with multimodal capabilities.
- The future possibility of integrating AI seamlessly into existing tools (like Google Sheets).
- Google vs. OpenAI Capabilities (24:03)
- Google’s advantages, including access to vast data through platforms like YouTube and Google Images.
- Concerns regarding privacy and data usage in training these models.
- The potential of Google’s generative search capabilities compared to traditional search.
- Demos of AI Applications (45:37)
- Sunny showcases various AI tools, including:
- Canva's ChatGPT Plugin: Demonstrated the generation of design templates.
- Headshots AI: An app that enhances personal photos to create professional-looking headshots.
- Discussion of the effectiveness and limitations of these applications.
- New AI Models and Tools (55:00)
- Introduction of Code LLaMa and Falcon 180B as significant new entrants in the LLM space.
- Code LLaMa: Meta's coding tool that is open-source, offering flexibility and rapid development.
- Falcon 180B: A new model from the UAE with impressive capabilities.
Key Takeaways
- The competition between OpenAI and Google in the LLM space is heating up, with both companies pushing towards multimodal applications.
- The integration of AI into everyday tools and interfaces is crucial for wider adoption and usability.
- AI applications are maturing, but challenges remain in user experience and the practical application of technology.
- Open-source models like Code LLaMa have the potential to accelerate progress in AI due to community contributions and rapid iteration.
Conclusion This episode of "This Week in Startups" provides an in-depth look at the current state of AI development, particularly the race for multimodal capabilities. Jason and Sunny’s discussions reflect the excitement and challenges in the rapidly evolving tech landscape, making it a must-listen for those interested in startups and technology innovation.
Follow Sunny
- [Twitter](https://twitter.com/sundeep)
Listen to the episode
- [This Week in Startups - Episode E1811](#)
---
Sponsors
- Squarespace: Build beautiful websites easily. Use offer code TWIST for 10% off.
- OpenPhone: Business phone numbers for teams. Get 20% off at openphone.com/twist.
- Fitbod: Personalized workout plans. Get 25% off at fitbod.me/TWIST.
---
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Transcript
Automatic transcript. May contain errors.0:00You know what's really interesting is Zillow is gone. oh zillow took theirs off yeah because it was so bad well i don't know i thought it was like again heading in the right direction but i think it's sort of this you know fear of chat gpt becoming like the apex aggregator and so if everyone is just going there and searching for things and you're just a data source then your business has less value so well you lose the interface you lose the customer relationship i mean that really is why people should not do these plugins because you don't have the email of the person. You don't know why they're searching this.
0:37They can be scraping your data and you just gave up any monetization. This Week in Startups is brought to you by Squarespace. Turn your idea into a new website. Go to squarespace.com slash twist for a free trial. When you're ready to launch, use offer code twist to save 10 % off your first purchase of a website or domain. Open Phone brings your team's business calls, texts, and contacts into one delightful app that works anywhere. Get 20 % off your first six months at openphone.com slash twist. And FitBod. Tired of doing the same workouts at the gym? FitBod will build you personalized workouts that help you progress with every set.
1:24Get 25 % off your subscription or try out the app for free when you sign up now at fitbod.me slash twist. All right, everybody. Welcome back to This Week in Startups. We're back. What an incredible week at the All In Summit last week. And we're back to our regularly scheduled programming. First up, Sunny Madra, Sandeep Madra is back to do a bunch of AI demos. Welcome back, Sunny. Good to be back. uh congrats on the summit it was incredible i would say look every single speaker could have been a keynote speaker so it was hit after hit after hit got it um and it was a lot of incredible content good conversations i liked how you guys were out there with every single speaker and um you know there's probably four to six viral videos that will come from it um because the content was really incredible from dalio to girly to grandma i mean yeah and um the only thing i would say is um you need to have more people there i think and the people there are like can we have less people it's like whoever got in watch less people whoever didn't get in wants more people funny how that happens well 1800 was you know i don't want to say it's unruly but it's you're moving a lot of people around i mean it's not easy and so we are going to have a all-in board meeting in a week and just sort of debrief on what was great about the event what could be improved uh but the speakers and the content were world class i think there was only one speaker when it wasn't all three of us which was uh wolfram alpha yeah um and uh he hung out the whole conference it was really interesting to see the speakers hang out the whole conference because they wanted to hear each other so congratulations to friedberg who programmed i think 85 of the content i think the other besties like brought one or two speakers each maybe so it was uh yeah i would really like more people to be able to come i think there's two different ways to go with it one is i think we're just going to do one type of ticket next year yeah and not have like general admission versus vip and then scholarships because it people just start comparing who's got a better seat it's like you know like on an airplane oh first class seat how much different is it than sitting in coach and it's or extra leg room it's just like it's confusing for people and like then some people are like i would have rather been in coach and be like i don't want to be in coach i want to be in the first class like everybody's different you know so i think we'll we'll tighten that up a little bit parties were great people had a good time and i share one image with you both that i think encapsulates all in summit more than anything okay uh we can take a little risk here but okay it's great that's steven wolfram from wolfram alpha i gotta go let's get that right steven wolfram and he is at side the rave the blade runner rave with like 30 people 40 people around him while like grimes is on stage tearing it up it was a pretty crazy final night party and uh the pier party people had a lot of fun we took over the santa monica pier i mean it is hard to throw a party for 1800 people let me tell you something the number of locations gets really small when you get close to a thousand people and that you can have an event and your choices are find a raw canvas like we literally used a movie studio yep uh we used a giant hall majestic downtown um which was okay but and then the entire pier and you can't like you're even a hotel like if you want to use the beverly hilton or something like that and have like bad food uh you know i'm not dissing them but like you know chicken you know like conference chicken like three course bad salad bad chicken bad dessert you want to do something like that like it's going to cost a fortune and they can only fit a thousand people and then everybody's going you know like have to sit at cheesy round wedding tables i hate that so we went a different direction but um la adds a layer of difficulty as well because of traffic so we didn't want people to be in their car for an hour it was i think 20 minutes drive to each party maybe uh depending on where you stayed 20 to 30 minutes to get to each party so that was pretty good um miami was kudos yeah kudos to you guys i mean like it's last year was the first one which is incredible uplifted it this year i think next year is gonna be really hard i you know i think it was you know chmask is trying to take charge of next year's one so we'll see how that comes together because you know it gets harder and harder but it's a good challenge because it's better for all the listeners better for all he doesn't want to do the big one he wants to do like a 50 person or a 100 person yeah you know small retreat invite only kind of situation so i told him like maybe that that's the retreat all in retreat and then there's all in summit summit yeah one's like a think tank where people are kind of um you know like like the people who are on the stage are getting together to talk and do illuminati shit and then there's like all in summit which i think we could do it in vegas new york in a theater there's a category of spaces like theaters that are generally four to six thousand people uh you know you think about like radio city music hall the beacon in in new york uh the theater at madison square garden all of those they tend to be 5 000 seats some of them go as high as like seven or eight but then you could invite a large number of people to experience the content and then you'd maybe give you'd maybe separate out the dinners and the parties where that would be like an upcharge or you could go to it if you wanted to but there were a certain number of tickets you know um because again you're basically talking about taking over a mega club or something or having to take over central parks you know great lawn or something it just gets nuts everything gets nuts at that scale so we'll see i'm excited i'm excited i mean i spent two and a half million dollars on three parties that was pretty good yeah so a lot of money to spend on parties but i think it was worth it i think it was worth it it was worth it for me i thought it was a lot of fun we had a great time i think yeah If your landing page is terrible, I'm out, right?
7:46Most consumers are. It's 2023. You can't have an ugly website. Stop selling for okay or good and have great. And great means you're using Squarespace. It's out of the box. Beautiful. These websites have templates made by the world's greatest designers that are going to engage your audience, let you sell anything. And Squarespace over the past decade has just added feature after feature on top of the gorgeous templates that are designed for mobile and the drag and drop web design with their fluid engine is just perfect, easy to use. And you get built in analytics, marketing channel analysis, sales data, all that stuff.
8:20Now, you know, it goes beyond page views and site visits and time and all that. And with Squarespace, you can create an online store or you can start a blog, click of a button, right? Easy peasy, lemon squeezy. You can create a subscription business for members only content. You're seeing a lot of that out there. It's simple. It's cost effective. It's gorgeous. And they keep adding feature after feature after feature. That's when technology is at its best, isn't it? When you pay one price, but the product gets better and better and better. You get that with your Tesla, you get that with your iPhone, you get that with Squarespace.
8:49These are the legendary brands of the internet of this era. Go to squarespace.com for a free trial. And when you're ready to launch, I want you to go to squarespace.com slash twist. And they're going to give you 10 % off your first purchase of a website or domain. Go to squarespace.com slash twist because they know we sent you so there's a lot in the news where we took a little bit of time off because we had that and um i have i literally had a call today with uh some of my older people you know i got some older people in my organization i said listen okay the three of us are over 150 years old combined i need all the old dogs learning new tricks because they were asking me like what publications uh should we promote this cloud kitchens incubator we're doing too and i said okay did you ask chat gpt yet before we ever get on the phone ask chat gpt six ways to sunday ask bard and you'll saw the promise which is actually how i got a lot of the um party planners and party ideas is i sit and chat gpt and pinterest and i have both windows open and youtube and i just start brainstorming ideas and people were asking me my process and i'm like what would be good themes at a cyberpunk party what would be interesting food to serve on a pier you know what movies in the 80s and 90s represented surf culture you just instead of doing google searches it all returns everything to me in a table yeah and i move like probably three four or five times faster um which is bonkers but there was some news out here um uh there are some news out here about open ai uh going multi modal multi m-o-d-a-l multi modal llm now so i think this is the first time a lot of people are going to be hearing about this so let's talk about open ai um racing to beat google to launch the first multimodal llm what does this mean explain it to everybody as if they're five-year-olds yeah so um it's the next evolution.
10:47So if you think about our primary interaction mode with all the different LLMs today, it's text. And so you give it text like you were talking about, and it gives answers back to you. What multimodal means is your ability to give it any type of input. It could be speech, it could be video, it could be pictures, it could be a combination of those things, right? So So that's the next iteration in these LLMs where, because today it's been separated out. You sort of have Dolly or MidJourney for images, right? You have OpenAI, GPT-4 for text. And so when it all kind of comes together into a single platform and a single model, that's what people are racing towards now.
11:36So, cause today, if you think about what's happening is people are fracturing. There's some folks doing stuff with text, some folks doing stuff with audio, some folks doing stuff and not, no, there's not one clear leader across all of them. And what folk, what I guess the leaders are all trying to raise towards is they want to become the single place where people go to, to do all their LLM related, um, activities. Which makes sense. You know, uh, multimodal from the word modality modality in this case refers to, you know, like how something is expressed right and it could be input but it could also be output so when you say multi-modality you could draw on a piece of paper take a picture of what you want an app to look like or draw it on the screen you know like there are there are little drawing tools you can use on your desktop and you draw something and then you say so you could have two modalities at the same time i could draw it and then i could type in or say make this into a modern looking ios app yep and then now it's got a sketch of what looks like an e-commerce site plus it's and then the output on the modality could be code it could be images it could be wireframes So when we start thinking about modality, I think, I don't know what they're referring to, if it's just input, or if it's also output, because that's where it's both.
13:05And so this requires the LLM to really have an understanding of what the user wants. um and how to interpret it i guess normalize it so an image turning into words sounds turning into words code turning into words they already are words i guess yep and then the output you guess you would have to specify because you could say i'm going to draw something make this into a techno track i mean now we're starting to get really weird i'd make an i draw an image of a bulldog with a bunch of stars and flowers around it and i say make this into a an electronic dance track make this into some like you know house party music it's like okay i'm going to interpret a bulldog with flowers around it into a dance track is that what we're talking about here i think that's like the stretch case but yeah i think the idea is to to allow people to interact with these systems in any way they want to.
14:04So maybe it's more for the image case that you're talking about that, but like, you know, it comes down to its interpretation, right? And if it's able to interpret what you're saying and kind of glean from that, you know, what that music, what music may relate to the image that you've drawn and based on its training data, how can it come up with it? And so it's a much harder use case, I think that's why we haven't, we didn't see it up front. It's easier to predict text from text and music from music. But, you know, image to music is really fascinating. I hadn't thought about that one, but why not?
14:41I think it's going to be crazy. And, you know, I'll be honest, I haven't really used any of the image because I don't work in images all that often. and i really haven't worked in those um generative space because they're all very um uh distinct and they work in like doesn't one of them work with the interface being uh discord yeah and i'm like what are you guys doing this reminds me of crypto all over again like mid journey shouldn't the interface for mid journey should not be discord it should be i mean i guess it's clever in some ways but it should just be like a chat interface i think It allowed them to go really fast and basically have a command line interface that wasn't your terminal, right?
15:25And so, you know, the interface to MidJourney is really powerful because, like, in your command line, you can give it, you know, types of different flags. And it creates lots of – that gives you lots of flexibility. And they didn't have to go and recreate that interface from scratch. And I, you know, I think I've heard, or may already even exist, but Midjourney, I think, is pushing to have like a consumer facing interface, sort of along the lines of what you're talking about. Yeah, that's gotta be, yeah, that's gotta be available soon, right? like so the the edge was that you know by putting it in discord there's like a natural community effects there because you know discord is you know think about it just like a giant slack channel or you know go back to irc or something like that so when you're in there you have communities of you can say cool stuff people shared you can comment on things you don't have to build out all the the social and chat features right so day one mid journey if you're in the discord for mid journey um i can even pull it up if we want to yeah i mean it's yeah sure pull it up i mean the brainstorming is kind of interesting to me because what what happens is just i'll just take party planning as about one concept in year zero this first year of ai it helped me find venues it helped put them in tables it helped me find i helped me come up with ideas um and it helped me um find venues and come find party planners and then brainstorm ideas but i found myself going back to the 1.0 mentality which was not talking to an ai but you know doing google searches clicking on image the image tab in google going to pinterest and just typing in like ideas for seafood at a party was one of the things i was doing i wanted to have seafood right and i stumbled upon people who were shucking oysters you they walk around a party with buckets on either side of their belt and the buckets have oysters in them they shuck an oyster they put some you know mignette on it or whatever those things are and they hand it to you and you have an oyster and you throw it in the the other bucket it's kind of cool right people walking around shucking fresh oysters for people of course it's kind of like a fun interesting you never saw a thing i had people making crepes whatever but i would really like to be able to take a picture of a venue and say you know uh give me some ideas to make this like james bond or blade runner and um then make a list of rental firms that have these items right that would have been really cool um and uh you know that that feels like what this multimodal mode is going to do i'll be able to input an image say something i still don't understand why chat gpt4 doesn't have a talking interface yet i just seem so de minimis to build it actually does on the app if you use the app it does within the app itself like the ios app or the android app it has like an audio button so you can just chat with it yeah does that work with your airpods in in a hey siri kind of way uh i don't think there's like a trigger like hey siri yeah you know if you put it on your home screen or something like that you can do it see that's what i i'm trying to you know this is what happens when they lock down the iphone so much yeah on an android i'm sure somebody can jailbreak this in 10 seconds but i want to have i want to say hey chat gpt uh with my airpods in with my phone in my pocket and just say what are the three best sushi restaurants around me that have a reservation for two yeah and read me the top read me the top three dishes from each and not take my airpods out yeah but i guess i guess maybe chat gpt mobile does that but you have to click on the mic button and that really is i know this sounds stupid but when you think about user design sunny yeah the fact that you have to click on a button and take your phone out is friction i want to leave these airpods in all day and just be talking to it so somebody take on a memo uh clicking the microphone is death i think we'll see it soon i think is apple gonna catch up no no they're just too far behind they're gonna miss the entire AI wave after they had Siri out first.
19:34I don't think it's far behind, but I think in order to be a leading player in this, you have to take risk. And if you think about everything Apple does, just look at the last four iPhones, they've really become super risk averse. And so if you're going to really play in the LLM space, you're going to be subject. And it's even affecting everyone outside of open AI. You have to take a lot of risks because it's going to produce responses that are have hallucinations in them all the stuff that we've talked about so i think given their corporate culture um given yes you saw the mother the mother nature kind of thing what was your take on that i'm curious it's like a big misread of society and where we're at and where they're at and just it shows what's happened to the culture of apple it's kind of crazy so i saw this excellent tweet that said imagine you took that to steve jobs and said hey we're going to put this in like a keynote release of a big product of ours right he would be like what product is this yes that you're talking about recycling yes which is important and valid but this is what we're spending our time on is not the what the device can do what's the magical component or upgrade to your device what's it going to enable you to do but hey you can feel less sad about consuming this correct yeah see i like it i like some aspects of it as something to release not during a keynote so i think the context here of hey we're doing a keynote look at all these new features if they had 50 new features and that was like the last thing they showed and by the way all these new features we want to just let you know a lot of companies screwed up the environment yep we're trying to do our best we know that you're going to buy these devices every three or four years some of you buy it every year we just want to let you know you don't have to feel guilty we're going to be zero waste and we hear you on the packaging less packaging nobody wants to have plastic you know everywhere and we're selling a lot of these things like there's a way to do it that isn't so preachy apple.com recycling go check it out that would be a great way to do it or or you could just say hey here's the tonnage you know if we just so you know we sell a billion devices and if we have this much so just so you know when you get your phone it's going to be in a really small package we know you love unboxing but look if we make the package this big when you unbox it this is the number of trees if we make it smaller and we hand i mean imagine if apple handed you the phone from a stack of them in a store in a sleeve it's just like in a perfect sleeve a paper sleeve and there was just a stack of them and they just slide one off the shelves i mean that's what we all want we don't want any packaging give me the damn device are you still using your personal phone number for your startup it's 2023 it's time to stop it is a huge mistake that founders make why you're just getting started with your company and you don't think about phone numbers as being an important part of the ip collection of your startup.
22:47With open phone, you can totally solve this problem. They've rethought everything about a modern business phone and how it should work. It's super easy. You just download the app on your phone or your desktop, and you pick a number, and you're done. And you do it for just such a low price. It's so affordable. And think about it. If you have your sales team using their personal phone numbers, a salesperson leaves and goes to a competitor, you don't have any insight into what phone calls occurred, what people's phone numbers are. That's your company's database. And if you allow the sales team to run them up, or the customer support team, it's just unprofessional, be professional, use open phone.
23:24And we use it for things like event communication. So we get one phone number, but it can go to multiple people like a round robin, then we have a shared phone number. Do that for customer support. And open phone is rated number one on G2 for customer satisfaction. And I trust G2's ratings, open phone, it's ready, it's affordable starts at just 13 bucks a month but twist listeners can get 20 off any plan for the first six months at openphone.com twist and if you have existing numbers with another service no problem easy peasy lemon squeezy open phone will port them over at no cost head to openphone.com twist to start your free trial and get 20 off i believe google has the best data they have the greatest data scientists and ai so let's start reflecting back here um chat gpt wants to get this out before google they're racing yep what do you think of google's chances here because they have that giant image database known as google images they have this giant video repository known as youtube yeah i would think if they block chat gpt from using images gmail and youtube that they would have a massive advantage in multimodal well i think spot on and i think the one real just i want to add one thing to it then i'll answer your question when it comes to youtube that's all generated content that through their terms of use they have access to uh inherently versus you know search and things like that it could get a bit wonky because if they're pointing out to someone's site and using that for training data, but, um, they can easily make one change to terms of use.
25:09And unless you want to pull your videos off, uh, YouTube, you're going to, you're going to say yes. Um, similar to the way we've done in this zoom thing, right? We want to try as we just kind of said, yes, like they can train off this stuff. Um, and so given that I'm actually already experiencing an interesting, um, change in my behavior. So I use the Google search app as a, It's like a native app. You can download iOS or Android. And inside that, you can enable generative search. And I have to say, and you can do it within Chrome and other places, but I have to say 90 % of my searches now are, you know, it's a standard search.
25:50And you can say, give me the gen result. And the gen result is incredible. And it's up to date. It's not limited in time. So better than chat GPT for an average search. i that's been my experience yeah yeah see i think we're in this little trow of indifference yeah there was this peak enthusiasm this is going to change everything and then i think a lot of people stopped using it i know people in my organization have stopped using it because like i said i was on this call today and i hammered it over and over again before you ask a question on a call especially a conference call please ask chad chippity and bard and then see if they give you a decent enough answer and then you can at least bring that answer to the party um but i think people are forgetting how powerful this thing is and it's like the habit has to be built but you're telling me in the google search app it's on by default so instead of going to your chrome or safari browser and doing a search on your iphone you use the google search app yeah um and you just have to turn it on you're saying yeah it's like in their labs feature and then you can turn it on and it's it's it's really you know it's now to use the labs feature you have to use an at gmail account or yes you do see this is the thing people don't understand is you either have to in order to use these advanced features just type in google labs and you'll find it uh these experiments but if you're on a google domain like you have your own custom domain like at launch.co or at inside at definitive you then need to go into your admin if you have admin access and turn on experiments yes but it doesn't seem like yeah so search labs you know here i'll just show you it says search labs isn't available for your account so that's like my corporate account it's not available um so i have to then go switch to my gmail account which i never use um but i'm being forced to by google this is the weird thing google treats google domain users google office users whatever they call it as second class citizens it's really annoying so but if i i just did this what are the new features in the iphone 15 and see here it gives me this little tab because i have it enabled on my personal gmail account yeah and click generate and um you know basically it's going to give you the quick summary right it's got usbc the dynamic island is a cutout instead of a notch 48 megapixel camera yada yada yada a16 chip right and so um and you know i actually used this the other day because i i have the apple watch ultra and i said what i said you know what is the difference between the apple watch ultra and ultra two right and basically here you can generate it and um and like you know and even just just below that right here uh jacal this is a sponsored apple ad right which is compare models yeah and here i just it gave me the little uh thing it did its own comparison in a table generated by ai so yeah now you could do hey what's the difference between the apple watch ultra and the google watch pixel watch whatever they call that yeah yeah and it's very important you have to use if you want to use these experiments you cannot be on um your domains account you got to be on your gmail account yeah but they are enabling it for your workspace so if you know i think uh at next a couple weeks ago which was maybe even the week just before all in summit they did say they're launching all the generative features for the workspace account so you can ask your admin to go in there and probably enable it uh for for the admin for the um work account yeah but where do you turn it on here it goes it says to turn on you go to google labs scroll down to that section uh settings hit enable click on save and then there you go look for the sge card and it's on i don't know why google is making this so difficult to turn on it should just be on the top of every google search page yeah is the reason they're not turning it on because it's going to take too much cpu and they're they would yeah that's that's what i heard that's you know i Basically, it is like a kind of a compute issue.
30:20And so that's sort of what I heard in the background from a couple of different folks. But usually, I think within workspaces, if you try it on that one, I think there they just enabled it. You don't have to wait for the wait list. Yeah, that would be the way to do it. so because then for your whole org like to your point that anytime they're just searching for anything they'll see it there right and then it's not like you don't have to change the behavior but they still make you click the button it doesn't do it automatically because i guess they don't it was doing it and i think they turned it off for the same compute reasons right and maybe even monetization because remember right below that was a sponsored apple link right so now like what i'm finding myself is i'm never clicking a sponsor link because i'm getting the answer what i need right in the gen search that's super dangerous yeah and that is what we've all talked about being a very dangerous um yeah that's a very dangerous thing for google to then lose all that revenue so they are being judicious in rolling this out all right you know i've been on a health kick over the past year and you know i care about data driven solutions.
31:36And if you listen to this podcast, I bet you do too. So let me tell you about FitBot. This is a data driven workout app that blends machine learning with exercise science. FitBot creates custom dynamic workouts programs based on your fitness goals, your experience. And most interestingly to me, the available equipment, let's say you got a bunch of kettle bells, or let's say you're at some, you know, sparse gym at a hotel, or you're on vacation, you got nothing. Well, FitBot will maximize your fitness gains by varying the intensity and the volume between your sessions and leverage the equipment you have or don't have, as the case may be.
32:11You can customize the length of your workout, what muscles you want to target, and so much more. So let's say you want to get a 30 minute workout in, and I want to do chest, triceps and abs, but I'm staying at an Airbnb, there's no equipment. FitBot can create a perfectly optimized workout for me based on these parameters, and it will do it for you to Check it out. It's amazing. The design of this app is extraordinary. I was able to invest in it. That's how impressed I was with it. Fitbod takes the guesswork out of fitness. Just open the app and start making progress. You deserve it. Get 25 % off your Fitbod subscription or try out the app for free when you sign up now at fitbod.me slash twist.
32:50That's F-I-T-B-O-D dot M-E slash twist for 25 off so how is multimodal different than say um writing code right the code interpreter yeah um because code interpreter i guess when it's writing code is really just an llm that is doing code instead of words but it's still words at the end of the day yeah i mean look i i think that's like the first step into multimodal like i would view that as the kind of like very basic multimodal even though both are in language you know it's it is sort of a different um i look at it as like a different use case interestingly though they renamed code interpreter to advanced data analytics oh yeah that's because people thought it was for developers not for for data scientists yeah exactly so they've kind of renamed it recently interesting and yeah that itself feels like a different product do you think the ui for these products is going to start to break up a little bit or do you think we're still going to have this like one ui to build them all so it's a really good question and you know we've been spending a lot of time thinking about this, even in my day job, I think the way it plays out over, let's say the next 18 months is there's going to be a set of folks that try to integrate it natively.
34:24And then this is a really good segue into the first demo for this week. And then there's going to be a set of folks that say, I'm going to rely on the sort of, you know, the bards or the Google search or, you know, OpenAI to send me, I'll integrate it via a plugin and we have a good example there. And then I think where I'm the most excited is like sidekicks, where you have an existing application, it has a bunch of context and knowledge, you know, could be data, could be features and functionality. And a sidekick kind of overlays on top of it. I kind of feel like we're going to see a lot more sidekick type applications than the first two in the next 18 months.
35:06because i think what we'll have to have is people rebuild applications from scratch to make them interact with like sort of multimodal prompts we'll even get away from text i think you'll have to reimagine that completely yeah it's people are so used to chatting with each other yeah that chat felt like the natural interface um but there there is a better interface i think when you're working on a let's say a google sheet or you're working on a table or a document which is i don't want to ask a question and have it rewrite it and move everything up yeah i want to stay on the same document and just have it change that document yeah you know i'm saying like i don't want to just keep scrolling and scrolling and scrolling or if i'm working on a excel document i would like to just talk to it and you know none of these are letting me do that and I had the the CEO of Grammarly on uh recently oh great and I love Grammarly yeah it's a great product and I would like to just sit on my document and say let's talk about the second paragraph and then the AI would say okay what would you like to do with the with the second paragraph here are some ideas make it shorter make it tighter make it funnier yeah you know what I I feel like um I'd like to say this in a little bit less words and more pithy and maybe make a quip and it's like okay here are some quips here's a shorter version and i just look at and say you know i like the second version let's work on that one and you know i'm i'm talking to it and it's changing it as we go not moving it up then i gotta cut and paste it move it over to my application so that's what you mean by having it be a co-pilot yeah and grammarly does a really good job like i don't know if you're tracking their latest features you guys talked about it but like they sort of have um they've kind of built in some gen ai features and it can you know and it's always been able to take your paragraphs and help you rewrite them but now it has just some more superpowers which i find are really interesting the the latest version of grammarly is is pretty awesome i always make sure my grammarly is like up to date because they're just launching like the best features all the time um it's pretty great i have to say i am I'm a big fan of Grammarly and they've added a lot to it.
37:26Ah, you know what's interesting? I was trying to figure out why I'm not seeing it on my desktop. Google is only allowing you, because I've been testing Firefox and Brave instead of Chrome, because I found Chrome is getting a little slow. And I switched. There's some optimizations you can do I'll send you that make it really fast. They've just launched them in the last couple of months that make it zippy fast. yeah because something went wrong with chrome over the last year or two where it just gets slower and slower and slower and i'm just like this is getting very fresh they did a big push i'll share them with you some performance updates that really make things like 10 times better but if you want to use google search experiments you must be using chrome which is like that bundling issue um that the justice department doesn't like which is you know to use these things you have to and here i am now i have oh i had it all on i have it for sheets yeah yeah it's i mean go i just can somebody at google stop doing i mean listen you're google make things easy right like why is google making people jump through hoops to find this stuff just let people turn it on whenever they want uh what are great 5 000 seat theaters in new york now i got it so let me see if i try it what happens hold on i'm going to share with you because this is what i'll do to plan the next all in summit um and you start thinking about party planners right yeah i mean what do i need a party planner for anymore you know like i need them on the ground maybe yeah well you need like you know you need physical labor that's becoming more and more about oh there you go you have it yeah so you just now i have it but look um here's a bunch of like random you know venues but here i say generate and let's see how it does it's generating to my query what are great 5000 seat theaters in new york i think that's what i can rent many theaters coney island amphitheater central park oh casino central park room i don't know about that some other ones include um what are the largest movie theater screens now what are great 5000 seat theater 5000 plus see if it understands plus there's in new york city i can rent
39:44and put that in a table see this is where we'll see if it can actually put it in a table for me you're really pushing it let's see but you know and if it can't um it's gonna that's a little bit it didn't do it but if i do the same query see this is where i get confused because if i go to barred and i asked for the same query in bard yep yep i bet you it does a table right so like what is going on over here google like you're you got bard helping us do this and the part put it in a nice table with exactly what i want so why not give me that in google search it's very confusing what you're doing over there but i guess they're trying to move fast right oh i like that export to sheets thing too that's new the export to sheet thing is like an incredible feature because you click it it creates a spreadsheet and remember i said before i don't want to be cutting and pasting and i want it to cut and spit cut and paste so now it's made that spreadsheet i can go open it and you know what a great start this is because this is what i the first step i ask my people to do is this is this uh and but what i want to do is i want the ai here yeah i want to say to the ai while i'm in google sheets add to column e they have that you don't have that turned on they have that now too they have that too you you uh there's a there's a sheets ai companion yeah there is yeah yeah oh my lord yeah yeah i didn't know they had that i'm gonna have to take a look into that one as well but you know i think this is where people who listen to this show um you got to use this stuff every day you got to use multiple ones because you will find over time like radically different answers from each um like chat gpt4 i bet you does a much better job at the same query even though it doesn't have so here madison square garden king's theory beacon theater and i get links to the websites right so i can say um oh nice add 10 more but limit them to theaters 10 ,000 seats and theaters between 4 ,000 and 10 ,000 seats.
42:06Also add a column with ask her with the price to rent the theater. And this is where, like, you know it doesn't have that information because it probably doesn't exist on the web somewhere. Exactly. And look, it says finding precise rentals challenging the variable nature of this. But at least it put in something. I wonder if it's guessing here. I mean, maybe that's the correct price. I don't know. But look, it did such a better job here. Apollo Theater, Hammerstein Ballroom. and you just think about how much work this would be to do google searches to ask people i mean this was like a hard task um and now all of a sudden you know you just watch it do the work so if you're a party planner and you're not using this um and if i said hey you know what are some themes for a party in new york city that people could dress up for you know now you're like i just want to brainstorm right i just want to brainstorm some ideas but look it always puts it in a table you know and oh look and it says why it works in new york city i didn't ask for that great gatsby that's a good one studio 54 that's a great one great one new york icons andy warhol that's a great one black and white gal is good sex in the city is good gangs of new york is killer wall street oh my god so many interesting ones i mean it's pretty great yeah i mean i have to say like i was thinking of doing a studio 54 one and a great gatsby one those are on my short list of parties for the next time yeah uh but i didn't think new york icons and i I didn't think Gangs of New York, and both of those are kind of fun.
44:03Gangs of New York would be so fun. That would be hilarious. That would be so fun. You know what? You've got a bunch of demos, but we're convincing you, the audience, of how this could help you in your job. So please send this clip to everybody who works for me. Please, people, stop asking me questions and just ask ChadGPT and then come to me with what answers he gave you. Oh, did you sign up for ChadGPT Enterprise yet? Did you apply? We've had an enterprise relationship from the earliest days, just kind of the nature of how we've been working with them. But there's an enterprise product specifically now that you can sign up for, and then it shares and it records everybody's searches in one sort of grand repository.
44:48So that's what I want. Yeah. Well, the enterprise features and like, I don't know them off the top of my head, but like the main functionality that's enabled there is like sort of like a single account, like what you're saying, you can kind of look at what everyone's doing in your organization. And it's, it's all, it's not like, it's not inside your enterprise yet, right? So you still, it's not like you get it, like kind of on prem or something, but it's the right thing to do as an organization in terms of how, you know, you've been thinking about using it. A single billing account, all that stuff, everyone can go to it and all the information, you know, because if you think about IP of your business, now when people are searching with it and they're coming up with these results and you're paying them to do it, that should be your IP.
45:32You should retain that. Right. And so I think that's kind of powerful. Let's go through some demos. I mean, it's been a couple of weeks. You had a bunch of demos back up. Wow me with what's going on here. Wow me and the audience. Okay. So we're going to go back to plugins for a minute because like, you know, we kind of started here and we started in plugins. I have our definitive one enabled. We're not going to use that, but slow plug. But I have Canva enabled. So Canva is excellent. And so. Canva's in ChatGPT. Okay. Yeah. Shout out Melanie. Give me a template for a, um, a graphic that I can, uh, post on Twitter.
46:14I'm just going to say Twitter for the sake of it. Okay. And so I'm going to do that. And basically, um, things are a bit slow today. So I'm gonna have to run this one twice since there it goes. And so, and we'll, we'll iterate this really quickly and what this will do. And so here it's going to give me like a, so if I want to do some kind of post to Twitter for an event that's coming up or whatever it may happen to be, uh, it's going to give me a template I can work off of, but what's, and it, you know, it's going to give me some, some different ideas here. Right. Um, and, uh, and so what's awesome is I can go right from this and I can click on this link right here and it's going to open up right inside Canva.
46:54Oh, sweet. Yeah. Yeah. And so, and so it's given me a few different templates that I can say, now I can say, oh, you know, these weren't great. So I'll say, I'll say, can you give me some templates that include headshots? and um and so uh i find this this is like that interesting inner intersection right because canva will ultimately have some kind of co-pilot but i find this like interaction within this um you know kind of chat gpt world um really really powerful and so um going back to what you you know you worked on a bunch of things recently your team worked on a bunch of things this could really speed stuff up because you can kind of start from here and you can you know get off to the races and yeah and get going so well it's interesting you mentioned that nick and i are working on a new podcast and we were just um we were going to do it we're doing a podcast for like capital allocators um and we have a couple of idea for ideas for names and we have to make art and i said hey who's our best artist can you just work with one of our people and show me these two names as a conference and a podcast and and maybe put it on a mug or a uh you know a jacket or a t-shirt and show me the what the artwork would look like and usually when i do that i just pay an artist 500 bucks or a thousand bucks to do it they spend a day on it they show me and then i'm like okay now i creatively like this one versus that one right um but yeah this is definitely a great start um so if i did make me a motivational poster give me a motivational poster that is sarcastic and actually depressing i'm asking you to do something pretty sophisticated which is like comedy so it's using chat gpt4 sending this um and we're gonna see it didn't do it it's giving me motivational posters we got that part of it but it doesn't feel like it understand so what the goal here so why is it not using chat gpt because if i had chat gpt4 without the plugin in yeah it would come up with sarcastic yeah because he's good at coming up with sarcastic stuff, I think.
49:24Yep. But it can't put the two things together. Why is it breaking? Because it's still a light integration. A light integration. It's still like what it's doing is it's probably taking what you're saying, sending a search over to Canva. Canva has a bunch of pre-built templates and it's grabbing them. The next iteration of this is what you're saying should go over to Canva. Canva should have their own LLM that can generate these things on the fly and make them available. And so to me, this is just, it's, it's in the crawl, walk, run. This is the crawl phase of this now. And what you're asking for is more like in the walk and run, but they're going to get there super fast because you shouldn't like, I completely agree with you.
50:06You should be able to give it exactly what you're looking for and have it, uh, and have it do that. Yeah. See, if I just asked chat GPT for make me some sarcastic motivational poster slogans that are actually depressing for workers and this is where like you're asking it to do comedy here or be witty um okay here's a table teamwork makes the dream work but nightmares are also dreams that's a good one i mean you can workshop that on the corporate ladder it's just like snakes and ladders because more snakes live for the weekdays because your week weekends because days are owned by us that's darn think outside the box but stay inside your cubicle there's your winner dress for the job you want which is still probably not this one fell harder so we have someone to blame oh these are good your ideas matter just not to us nick turned that one before for me reach for the stars but don't expect a raise oh that's pretty good so if i were to take these see what i want to do is take these and look at it and they actually explained it uh which is pretty funny you didn't this success is one percent inspiration 99 perspiration and 100 office politics is that's really good it's great uh please uh remove column two and give me just column one okay now if i take these and i just have the one column i should be able to cut these paste them into the canva thread and say show me these with a motivational image behind them but it's not going to put the right image behind it you know i'm saying i don't think i don't think i think the integration is still a little bit early uh but let's give it a try i haven't tried chat gpt4 plugins okay um put these slogans on top of inspirational posters let's see let's see if it comes anywhere close to doing it see this this also makes me wonder if the people who are building these plugins are even coming up with the use case or they're just yeah so it's just giving you templates to use it's not a deep integration it's not a deep integration it's not even putting the poster it's not even understanding that i want these slogans on top of it yeah but that's not a ton of work i mean hopefully someone from canva watches this and then they can implement it pretty quickly embarrassing to my friends at canva embarrassing like just schlocky work sorry no i'm joking i mean i understand it's like this is like come back on the podcast meli come back on the podcast we love you you're awesome no i mean do better come on come on people i mean it yeah it seems like is it that the plugin architecture is janky is that the problem no no no the plugin architecture is really good in fact which one's the best plugin right now that you use or i'm super or super impressed by because this did not impress me this is more work than just using canva well wolfram alpha like if you really want access to like kind of facts does an incredible job um you know what's really interesting is uh zillow is gone oh zillow took theirs off yeah because it was so bad well i don't know i thought it was like again heading in the right direction but i think it's sort of this you know fear of chat gpt becoming like the apex aggregator and so if everyone is just going there and searching for things and you're just a data source then your business has less value so well you lose the interface you lose the customer relationship i mean that really is why people should not do these plugins yeah because you don't have the email of the person yeah you don't know why they're searching this they can be scraping your data and you just gave up any monetization so yeah but in the case like if they link you out like you're never going to be able to edit that poster in chat gpt like canva still canva still has like a lot of value proposition in my mind there well if they're going multimodal i have a feeling i could find the templates i like and then just tell chat gpt if they're going multimodal to do what i just asked it to do yeah that i mean see this is where i think this relationship between the language models and the verticalized software i think they got to put chat gpt in canva they're gonna put a language model into canva before canva is in chat gpt that's it full stop and you got to build enough accoutrement and features around it that people do not go to chat gpt for first well that that's starting to happen with the growth of the models right even in the last three or four weeks where we've been off um there's been some pretty significant news releases right from llama code yeah so llama code which is uh you know meta's version of code interpreter uh was released and um you know made available for from an open source perspective uh you know falcon released a bunch of different models and different so code llama which is l-l-a-m-a that is the coding tool this is like co-pilot for um for a code interpreter from zuckerberg from meta yeah yes and they made it open source they made it open source there you so here it is introducing code llama and ai2 for coding and so this will be better than whatever chat gpt does do you think because it's open source and people can rip on it and edit it and fork it and do all kinds of interesting things?
56:06Yes and yeah. So let me give you two different answers here. Currently, I think based on industry agreed upon benchmarks, it's not as good as Code Interpreter. But because it has the ability for folks to modify it, it's going to get there quickly. And so on this one, I know you like this, jay cal i'll say by january of 2024 this will surpass the capability of code interpreter fascinating yeah so this conversation you and i had at the start of our journey doing this week in ai basically every week the distance between chat gpt4 and everybody else is starting to narrow specifically in certain verticals code being one of them interesting and then what happened with falcon now is is falcon explain what falcon language model is to everybody so that's uh that's a language model i think created in the uae in fact yes and uh they have uh different size models um and and just to kind of go back you know what different parameter sizes i think they have everything ranging from like 1 billion to 30 billion and the way for folks to think about the parameter size is like neurons the more parameters there are like the smarter the model is and it was most likely trained even on more and more data exactly and so they now have a bunch of different models uh that they've released 180 billion is is going to be pretty significant and just for for reference um open ai hasn't published how big chat gpt4 is but i think the consensus is that it's it's um it's like four or five 200 billion parameter models working together um and again they haven't published this but people have you know kind of people are saying this is what it looks like and so um it still would be you know i guess have five times more neurons than say falcon 180b yeah and this is amazing the uae has bought a ton of big iron and they're working on this and making it open source as well so you can go see it on hugging face is the falcon 180b demo yeah and you can play with it and download it yeah exactly you do whatever you want with it yeah yeah uh and so if i were to put in here what are some great themes for a gang for a party in new york city i wonder how it does um masquerade ball roaring 20s he got that one great gatsby skyline and themed broadway musical theme party that's not bad madman not bad urban rooftop movie night classic films of new york that's not bad speakeasy prohibition era pretty good superhero inspired jazz age okay not bad yeah uh pretty similar kind of makes sense right gotham city gotham city metropolis always kind of have fun with that bunch of people running around manhattan outfits um all right it's amazing we have time for one more demo give me your next best demo here and we covered a lot of ground today we did i gotta get back on my lp calls raising fund four halfway there oh yeah man raising a fund in uh the worst economic market in 13 years 14 years what a great idea thanks jay pal i mean it literally is like you know i've never received so many i love you i want to be in business with you we are closed for business this year in my life but i'm halfway there so i feel pretty good about it in this market if you're getting halfway there that's i'll get the rest i'll get the rest it's gonna take me another six months um this one uh i'm not gonna do a demo because i was having an issue with my api key with it but i really want to push people this because this problem comes up all the time so and in fact even uh you and nick were going back and forth on something with respect to headshots but basically it's you know headshots enhanced by AI what you do is you give it you know some photos of yourself and basically give it you know some time to work its magic and then turn those into professional headshots and I know people need this all the time it's a very very simple use case but it's a common use case and we don't need it and I thought it was like super exciting and I have seen people using this there was one thing that people got a little upset about which I guess everybody reads into these but uh a woman had done them i saw this uh on the on the the twitter slash x x formerly known as twitter and it made them a little too what's the word here uh it was a little too sexy oh okay and you know like you know it was almost like making her an anime character in terms of proportions and stuff and she's like what the it's a business headshot not a anime superhero um and so and then there's also i think some people who were like wait a second this is turning me from this race into this race so uh there were there i think that's a bit of an interpretation problem like i don't think you can read into it that the ai wants you to be another ethnicity it's just probably got a bad training set of data right well it's not just training set But, you know, you got to remember these things are all tuned with a process called, you know, reinforcement learning with human feedback.
1:01:32That involves humans and it involves humans and their preferences. And, you know, we what no one really tells us is like, well, what is the makeup of those, you know, human feedback reinforcement learners? Right. And, you know, and I know like a lot of this open AI had talked about this in the past. They had it done in Africa. I think they had teams, you know, kind of spread across Africa doing a lot of this work. The reinforcement learning work. The reinforcement learning work, right? And so a lot of this can, the models themselves, or even the training data may not have as much bias in it, but definitely the humans doing the reinforcement learning, like we're humans, we're going to have bias in it.
1:02:13So I think it's something you got to be mindful of with all of these technologies. It's becoming a viral trend. And if you just type in make AI professional headshots. Yeah. You can always tell when something's like making money because there were four ads for services to do this. Oh, really? Oh, wow. Yeah. $29 per person done in two hours. Only$25 for a collection of six custom headshots. Now you think about what it would cost to get a professional headshot made. You got to find a photographer. They're going to charge you at least 100, 200 bucks to come take your photo somewhere. it's gonna take four or five hours of your day here you take a couple pictures of yourself yeah and you're done this is pretty go buy a suit if you don't have one no that's the crazy part is you can put yourself in you'd be like put me in armani suits put me in some great suits and um we're good and then you show up for work you don't own any suits they're like hey what happened to your suits sunny hilarious all right everybody metaverse yeah that's another uh this week in ai with sunny madra and jay cow we will see you next monday with another amazing episode we're going to try to be consistent producer nick in having every monday be all that ai goodness for you so wake up monday get that um this week in ai with sunday madra and your boy jay cow and we'll see you all next time on this week in startups bye bye
1:03:44Thank you.
From the publisher
This Week in Startups is brought to you by…
Squarespace. Turn your idea into a new website! Go to Squarespace.com/TWIST for a free trial. When you’re ready to launch, use offer code TWIST to save 10% off your first purchase of a website or domain.
OpenPhone. Create business phone numbers for you and your team that work through an app on your smartphone or desktop. TWiST listeners can get an extra 20% off any plan for your first 6 months at openphone.com/twist
Fitbod. Tired of doing the same workouts at the gym? Fitbod will build you personalized workouts that help you progress with every set. Get 25% off your subscription or try out the app for FREE when you sign up now at fitbod.me/TWIST.
*
Today’s show:
Sunny Madra joins Jason to discuss Sunny’s All-In Summit experience (1:34), OpenAI’s race to beat Google in launching the first multimodal LLM (9:07), whether generative AI needs a UI shift (33:00), and much more!
*
Time stamps:
(0:00) Sunny Madra joins Jason
(1:34) All-In Summit 2023 recap and Sunny’s experience
(7:44) Squarespace - Use offer code TWIST to save 10% off your first purchase of a website or domain at https://Squarespace.com/TWIST
(9:07) OpenAI’s race to beat Google in launching the first multimodal LLM
(14:42) Experiences with image-generating AI and Midjourney's interface
(19:23) The culture at Apple and where risk aversion sets them back
(22:32) OpenPhone - Get 20% off your first six months at https://openphone.com/twist
(24:03) Google’s advantages and chances against OpenAI in multimodal LLMs
(31:30) Fitbod - Get 25% off at https://fitbod.me/twist
(33:00) UI developments and what sets multimodal LLMs apart
(44:22) ChatGPT Enterprise
(45:37) Sunny demos Canva’s ChatGPT plugin
(55:00) Code LLaMa's potential and Falcon 180B’s unique features
(59:41) Sunny demos Headshots AI
*
Follow Sunny: https://twitter.com/sundeep
*
Check out Headshots AI: https://headshots-starter.vercel.app/overview
Check out Falcon 180B: https://falconllm.tii.ae/
*
Read LAUNCH Fund 4 Deal Memo: https://www.launch.co/four
Apply for Funding: https://www.launch.co/apply
Buy ANGEL: https://www.angelthebook.com
Great recent interviews: Steve Huffman, Brian Chesky, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland, PrayingForExits, Jenny Lefcourt
Check out Jason’s suite of newsletters: https://substack.com/@calacanis
*
Follow Jason:
Twitter: https://twitter.com/jason
Instagram: https://www.instagram.com/jason
LinkedIn: https://www.linkedin.com/in/jasoncalacanis
*
Follow TWiST:
Substack: https://twistartups.substack.com
Twitter: https://twitter.com/TWiStartups
YouTube: https://www.youtube.com/thisweekin
*
Subscribe to the Founder University Podcast: https://www.founder.university/podcast




