This Week in Startups - Threads, ChatGPT usage drops, and AI demos with Sunny Madra | E1774

Episode Date: July 11, 2023

Fin can’t burn its mouth on hot pizza. Or wave at someone who wasn’t waving at them. Fin can resolve half of your customer support tickets instantly before they reach your team. Meet Fin. A breakthrough AI bot by Intercom – ready to join your support team today. Visit https://intercom.com/fin Eight Sleep. Good sleep is the ultimate game changer. Now you can add the Pod Pro Cover to any mattress! Go to eightsleep.com/twist to check out the Pod Pro Cover and get $150 off at checkout! Carta now lets you launch and administer SPVs for your syndicate. Share your knowledge, capital, and network to launch your syndicate SPVs through Carta. Get 10% off your first SPV with promo code TWIST at http://Carta.com * Today’s show: Sunny Madra joins Jason to demo VenturusAI (11:01) and other tools, before discussing Sunny’s new AutoGPT project (34:38). They wrap up talking about Meta’s launch of Threads (49:08), Google’s attempts at building a social network, and Inflection AI’s new supercomputer (1:00:13). * Time stamps: (0:00) Sunny joins Jason (1:49) ChatGPT sees a decline in growth (10:22) Fin - Try Fin, Intercom's new AI customer support chatbot, at https://intercom.com/fin (11:01) Sunny demos VenturusAI (24:29) Eight Sleep - Go to https://eightsleep.com/twist to check out the Pod Cover and get $150 off at checkout! (26:01) Sunny demos Vercel (34:38) Sunny's new AutoGPT (41:58) Carta - Go to http://Carta.com and use code TWIST to get 10% off your first SPV (43:29) The decision to be open-sourced or closed (49:08) Meta’s new platform Threads (1:00:13) Google’s attempts at a social network (1:03:28) Inflection AI's supercomputer, roundtripping and training LLMs * 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

The best way I heard it explained to me was when you're behind, you open source. When you're ahead, you're closed. And if you look at, say, Windows, they had a monopoly on the desktop, closed. They don't need to be open source. But then you look at Google very far behind. They have a monopoly on search. So when they talk about search, their algorithm for search is closed. You can't understand how pages are ranked.

But then you look at Android, they went open because they were so far behind. iOS and stuff like that. So they went open source. And they wanted to be on everybody's phone. So I think Facebook going is closed when it comes to their graphs and everything, right? Facebook, Instagram. They used to be open when all the companies like were created, you know, like Zinga and all.

And then they close the graph up at some point. When they had a lead. And nobody can compete with them, which is exactly what open AI did. They became closed AI. So fascinating. This week in startups is brought to you by Finn can't burn its mouth on hot pizza or wave at someone who wasn't waving at them. Finn can resolve half of your customer support tickets instantly before they reach your team. Meet Finn, a breakthrough AI bot by Intercom.

Ready to join your support team today. Visit intercom.com slash fin. Eight Sleep. Good sleep is the ultimate game changer. Now you can add the pod cover to any mattress. Go to 8Sleep.com slash twist to check out the pod cover and get $150 off at checkout. And Carta now lets you launch and administer SPVs for your syndicate. Share your knowledge, capital, and network to launch your syndicate SPVs through Carta. Get 10% off your first SPV with promo code twist. Hey, everybody, welcome to another episode of This Week in Startups and Sunny Madras here. My guy, Sandeep Madra, from Definitive Intelligence.

if you don't know what that is. I placed a little bet on that company. I was able to get a little, little tiny sliver of Sunny's cap table, a serial entrepreneur, and one of the smartest, most fun guys I know. Welcome back to the program.

Good to be back. Good to be back. Very excited. All right. Missed you guys last week. I know, I know. And I wasn't on All In, and these guys did a rogue podcast without me.

Go on vacation. Crazy. These guys were out of their minds. But it was pretty much. Did Friedberg really do it an I-movie? No, I think he was... No, I think it was joking to do an I-movie. But he did edit it himself.

I mean, that's why it's like a little bit janky. But, I mean, if you record a Zoom podcast and people have good microphones, eh, you know, it takes a bit of... The visuals weren't edited. So you get that four-by-four frame, which I hate. Yeah. It's like a bad experience to watch people drinking coffee or, like, checking their email or, you know, pulling up the next topic.

It's much better when it's a single or... you sometimes go to it. But yeah, I mean, I was impressed. It, you know, having hundreds and thousands of edits from sacks and everything. Like, these guys really edit the pod. So to make themselves, like, they like focus on every little sentence and edited in post like lunatics. I don't know. I just, I'm just like, whatever I said, I said.

But these guys are like super precious and have their marketing and comms people like review everything for compliance. Which I understand. Like if you have funds, you have to be careful. But yeah, I mean, we all agree to be off. and then they're like at the last minute and I gave everybody producer Nick and I tried to shut my company down that week and

I guess like I gave everybody the week off you know like they deserve to get a vacation day but yeah not a bad episode but we should talk about there was one thing on there about AI which was the drop in AI usage which I would have liked to comment on

because I do think you know that whatever 10 or 20% drop is notable in some ways and not notable what's your take on that I mean obviously kids out of school is I was going to go there. That's super interesting. Yeah, so I think there's the kids out of school

problem and then I do think, you know, if we look back and it's kind of really timely, in the next, you know, in the last 48 hours, OpenAI just released code interpreter, which, you know, we've demoed here early on

because we had early access to it, but they just released that to everyone. And so I think it's there isn't, there's a definitely, part of this being impacted by where we are in the school cycle and obviously that's going to impact things. But I think more so

having code interpreter available to all paid users is going to create a massive uplift for them. Because that's, you know, we've seen the functionality before. We don't have to demo it again, but it's really powerful.

Well, explain to people who maybe are hearing about Chat ChAPT's code interpreter. What is that used for? Just one more time for the audience. Yeah. So the code interpreter allows you to give some data, usually in the form of a CSV file, into chat GPT, and then have chat GPT help you with analyzing that data. And, you know, the example that we did was like some output of electric car registration data, and then we input it and we had, we asked some questions,

we had some charts created. So it's a way of like having like your own personal data assistant around a smaller data set available to you. And that is now available, code interpreter, to the people who are paying the 20 bucks a month. 20 bucks a month, exactly. How many people you think are paying the 20 bucks a month now? I think it's a million, two million? Yeah, I would say it's, yeah, my guess would be somewhere between two and five million. If you think about that, it's going to blow past the New York Times, which is, I think,

at 9.7 million last I checked. They might be, they're about to pass 10 million. Yeah. And that's not insignificant. If 10 million people, well, 5 million, that's 100 million. month. Yeah. 100 billion a

month, 1.2 billion a year in subscriptions. And subscriptions are generally 100% profitable, right? Yep. They do have some

infrastructure costs. But I think they could become, if people are willing to keep paying the $20 a month, that'll be the real test. Yeah.

I also, people don't understand that all web traffic, YouTube, Twitter, Facebook, everything goes down a certain percentage during the summer

because people go on vacation, they go outside. So that's the obvious thing. I also think there are some people who, and I guess this is the more interesting topic,

the press immediately went to, oh, it's waning, people are less interested in it, which is such a stupid take. Because obviously, when a new technology comes out, everybody tries it, because there's no cost to trying it, but you find your natural audience for your podcast, for your software, whatever. So what are your

thoughts on that angle, that some of the press were like, oh, it's a fad, it's a fad, it's crypto. Well, you know, like, it's definitely not, right? I mean, the use cases are there. You know, we've talked about this.

The origins of this podcast was in crypto. And so, look, the other thing that, that it's hard to take into account until OpenAI starts publishing some type of data is many use cases have made their way into other products. So where you may not have to go to openaI.com and, or, you know, chat, you know, chat, GPT. OpenAI, you may be using OpenAI indirectly, either through Notion or through any number of products that now offer integrated experience with their APIs, right?

And their underlying LLMs. And so I think it's like almost like let's think about AWS, right? When AWS initial customer was primarily Amazon, but then as they made it available to others, you have to look at AWS as, you know, from a revenue perspective, not just as

a, you know, sort of a end consumer site. So I think, I do think, you know, these numbers and the way to look at it until we get some data or like maybe someone publishes something around their API, I don't think we have the full picture. So I think it's, I think people are just jumping to a story, which is easier to do than saying, because my guess is their API usage through all the startups and enterprises that are out there is, you know, increasing week over week at a pretty significant clip.

That is, I think, where the rubber meets the road. When developers use an API, that means, in some cases, they are playing with it, but in majority of cases, I think there's some application that's going to hit consumers, and they just

don't see it. And so you do not judge Amazon Web Services or Azure or Google Cloud, but the number of consumers talking about it, it's the number of developers talking about it, and that is Yeah, and then ultimately their revenues, you know, those things report now separately, and we can see, you know, what kind of huge impact that they've had on the, you know, top and bottom line of those companies.

Well, I mean, the growth of cloud computing was spectacular up until 2022, when it still was spectacular, over 20% growth month or year over year. But it did slow a little bit, I think because of belt tightening people during a recession, or recessionary-ish kind of thing we're in a down market intact. a tech depression. Yeah, and you know what? It's like a good kind of tie-in to this topic. Like a lot of folks have been pushing to the cloud for years, right?

And we've seen those phenomenal growth numbers. The one thing that I think companies struggle with as they were moving to the cloud was the benefits that the cloud provides. Because if you were a legacy business, right, and you were running either something on-prem or maybe in, you know, your own data centers, those things were probably really efficient. and moving to cloud doesn't immediately get you that efficiency. Where the efficiency starts to amplify now is when you want to start using additional services, whether the services came from the cloud providers themselves or like third party services,

which requires your data to be in the cloud. So I do think what we're going to see very shortly is a huge uplift in workloads in the cloud being driven by AI applications because that's the place you have to drive it. And so I think that's something we'll see that. will play out, I think, in the next 18 months. Finn can't go through a golf phase or still be haunted by a bad haircut they had in middle school. Finn can resolve half your customer support tickets instantly before they reach your team. What is Finn?

Finn is a breakthrough AI bot from Intercom. Designed for customer support teams. It learns your entire knowledge database and has the ability to carry conversations, remember context and nuance, while slashing your resolution times and support volume. Meet Finn, a breakthrough AI bot by Intercom, ready to join your support team at A. Visit intercom.com slash fin.

I mean, let's just get to demos. That's why everybody's here. We did it a little couple of minute preamble. We got on the same page here. But I love the fact that you're obsessed with this like I am. I have been doing a couple of projects myself over at inside.com. I won't talk about them yet. but one of the things I was doing

I want to get some advice on is how to tag things by category you may have seen, I'm trying to have instead of human editors tag the stories. I was trying to see if I could get chat GPT4 to tag the stories correctly and I need to get a prompt

that really does a good job on that so we'll talk to about that offline or if anybody listening. I'm trying to find like a database of like the most important topics in the world that somebody has come to the conclusion that these are like the actual topics. So Google has one, it seems, for advertising. Yeah, the 2,700.

Yeah, 2,700 verticals that kind of represent almost all kind of topics. Yeah, I'm trying to figure out where to get that list from exactly. And if Google lets you use it, I can send you a link. Yeah, no, they publish it. They publish it because they want people to be able to download it, incorporate it into their websites, etc. Exactly, exactly. I'll send you to link. Okay.

First demo is coming up. Of course, we'll sports guests us, if you're listening and not watching, well, go to YouTube and type in this week in startups and go find the channel, subscribe to the channel, put the alert on because I'm going to be doing some breaking news alerts over the summer from time to time, but go ahead and check that out. Okay. All right. So this is a fun one. And actually quite useful. I'll speak to this. Like, you know, when I was earlier in my career and I needed help and so I'm kind of creating a basic framework.

I don't have an MBA when you're trying to, you know, basically understand other. aspects of the business other than technical, you want to have some framework. So this Venturous AI, you come to their website and basically you can give it a topic and it will come up with either an advanced or simple. The free ones are the simple thing and it'll do a business analysis. And so, you know, there's a great movie from the 80s called Brewster's Millions. And I don't know if you remember in Brewster's Millions. one of the ideas that Richard Pryor's character is pitched on is a guy wanting to sell ice that's breaking off of icebergs.

Artisanal ice. Yes. Yes. Your artisanal glacier ice. I mean, I'm crazy, but I think I literally heard a pitch on people who wanted to get artisional glacier ice to put in fancy cocktails. I don't know if that was real or I imagined it or it was from Bruce Stour's millions, but okay. Yeah, and so basically here is it, you know, generated by me. So what's this website called?

Venturous, V-E-N-T-U-R-U-S-A-I. Venturist A-I. Terrible name, okay. Okay. I kind of liked it, but... Venturous? Oh, I like Adventurous, but Venturous. Okay, I get it.

Or maybe like a venture, you know, as well. Okay, so you put your startup idea in. I want to start a company that sells ICE, that breaks. off icebergs. Very simple prompt. And it basically turned that into a business analysis and feedback. So it gave me a brief description. It basically did that on its own. Then it did a SWAT analysis for me. It subsequently did a pestle analysis, right, which is political, economic, sociological, technological, technological, environmental, legal. The target audience and user stories, business strategies, business frameworks.

And, you know, I'll just, I won't read off all things here, but it basically gives you a solid framework for a business. And in many ways, J-Cal, you know, you guys do this at Inside, right, when folks, or launch, I'd say, right? At launch, when folks show up with something. I thought they did an incredible job. And this was the free version. So, Porter's Five Forces Analysis, this is,

what's really interesting about this is they went into sub- categories or other people's frameworks for analyzing a business. One of those is Portis Five Forces analysis. I've heard of that before. I've never actually used it. But can you maybe read some of those? Yeah, sure. So that's like number 13 here. It's like so threats of new entrance. Moderate as barriers to entry include sourcing ice from iceberg, established partnerships and brand reputation. You know, bargaining power of suppliers, right? Like do you have some kind of edge there with the suppliers, right? Bargaining power of the buyers, like who's going to buy this. You know, threat of substitutes. How easy will this be for someone to choose another party that's

doing it? Intensity of competitive rivalry. Like how quickly will someone- What does it say about this threat of substitute products, i.e. the ice machine in your refrigerator. That's already there and you've already paid for. No, the five forces so people know, I'm just reading from Investopedia here, competition in the industry, potential of new entrance into the industry. So the first to our competition. The power of suppliers, the power of customers. In other words,

how much power do they have over you as the provider of these? And then threat of substitute products. So a substitute product would be slightly different. It would be something to cool drinks, right? As opposed to ice itself. That would be a direct competitor, right? Yes, correct.

That's super fascinating. So essentially what this is doing is they have a series of prompts they run your idea through, is what I'm guessing, right? Is that what the same thing? is my guess is like sort of the rough structure behind this is yeah they take a prompt and what they do behind the scenes is they have a set of you know prompt templates that walk it walk your idea through each of these and they have 14 sections here and so they take the idea and then they work with an l-lm to create a brief description and then each of those 13 sections which i think is really powerful like i i think it's this is great i mean this is like this would be a whole semester at a business school, you would work on something like this. Yeah. And now you can basically do an

approximation of it. Who knows if it's actually of the quality of what you would get in a course, right, to, you know, do your first mock-up of a business. Yeah. But it would literally whip you through multiple of these. So I just did one. Yeah. And I said, pull it up. Let's see yours. Yeah, yeah. Let's see what mine does. Because this is actually something I'm thinking about doing. So I'll make a little bit of an announcement here. Venture Capitalist Training School. Comprehensive analysis and feedback. So you know, I have Angel University where I teach people to be angels and we've donated 200,000 to charity. I've taught it 35 times, I think. And we've basically got maybe, I think, four or five thousand people have taken the course now. And so we've created a lot of angel

investors in the world. But my idea is now that I'm going to have my venture studio, my accelerator in San Mateo, I'm trying to find a nice garage or something, put it in a big open space. I was thinking of starting a competitor to Kaufman Fellows. You know that program? It's $80,000 for two years. Yeah. So I want to create a Kaufman Fellows killer. That would be half the price or maybe one year intense or six months intensive and maybe be 20K or something. Have 10 people come to each one and create a program where they basically get to draft off my deal flow.

And the core of it would be they would sit in on all the investment team meetings and do all the front line meetings. It's like a launch EIR program. Like an EIR program, but like an associates in training. Yeah. Instead of EIR, AIR, AIR, associate in residence or AIT. So anyway, here's the business idea. The business idea is to establish a school that offers comprehensive training programs to individuals aspiring to be venture capital. So it took my, my prompt, by the way, was venture, uh, what was my name?

Prample's training school. Oh, yeah. Um, yeah, Venturellia training school. Um, the school aims to provide participants with the necessary knowledge skills and practical experience are required to excel in a highly competitive field of venture capital. So we'd added all that. It adlib that. which is quite accurate.

The venture capital industry has been witnessing significant growth in recent years. That's true. Fueled by increasing startup activity, true, and continued interest of investors in high potential early stage companies. However, there is a shortage of skilled venture capitalists. That's true, who can effectively identify value in investment. That's very true. By establishing a dedicated school for venture capital training,

this business can tap into the demand for professional education as field and potentially bridge the skills gap. This is true. SWAT analysis, strengths. Unique business idea with limited competition in the market. Taylor trading programs can address specific gaps in the industry, potential to establish strong industry partnerships for internships and job placements. And Jacob, I can pause you for a second here. You know, some of these things also, just a framework is good.

Like if you're trying to do this, like the SWAT may not be fully right, but it can get you thinking and get you going, right, as well. I think that's the key point is that. You, as somebody who didn't go to business school, and just to be with those SWAT analysis, strengths, weaknesses, opportunities, and threats. I had to go take a look at that because I didn't remember. It's been so long. I don't go through that.

Limited market size, venture capital training is in each field. That's true. This is weaknesses. Demanding and resource intensive curriculum acquiring experience instructors. I don't think it has to be resource intensive, but okay. But maybe. Well, you do need to have somebody like myself who's been doing it for a while. You need to continuously adapt programs to incorporate changing industry trends.

That's actually very true. Huh. Threats. High competition for top talent from established venture capital firms. That's not true. Yeah. They're going to try. I don't think it's a threat to our business.

Okay. Right. Like if there's actually high competition for top talent from venture capital firms, that's what you want. That's actually a benefit. Fair. That's true. Yeah, because then these people graduating would be rapidly evolving industry dynamics and regulatory changes. Nope.

The last one is most. Economic downturns affecting investor confidence and startup funding availability. That's, it nailed it. So it must have just said, what are the threats to a business that did this? Yes. Wow, this is incredible. And I've never even heard of a pastel analysis. Yeah.

Yeah, you didn't spend enough time in corporate America. I did not. economic, sociocultural, technological, environmental, and legal. When you were, what was the name of your consulting firm called? Extreme Labs. Extreme Labs. So when you did Extreme Labs, when you had customers,

they would pay you massive amounts of money to write this kind of stuff up and included your analysis, or they did it themselves? They were doing this themselves. We were more of like on the development side, right? Got it. So they would come to you with this stuff. Yeah, but as our business expanded, we were doing more product. management and product incubation, then we would do these type of things.

Got it. Yeah. Oh, my Lord. Suitable business strategy. Yeah, this is incredible. What a great. This is the free version. They haven't advanced.

I didn't get to try the advanced version, which they don't allow for free. You can do 10 of these for free. The advanced version is maybe something you should try for your business idea. And it lets you make the report visibility, public or private. I just put it on public so people could go see it with books to guide you along the way. Venture deals, the lean startup. Angel, how to invest in startup? Timeless advice.

Wow, I can pick that one up. Who's the author? This book offers a first-hand perspective on how do I identify, evaluate, investment problems, and technology starters making it highly relevant to your business idea. Who's the author? Oh, Jason Gallaghanas. That's pretty funny.

You can download it and export to a Google Doc. Wow, what a great service. So shout out to whoever made this. Yeah. Venturous AI. Congratulations. And let me see the pricing here. It's got on the pricing tab.

Start or free. Yeah, 10 standard reports a month. Got it. Pro, 20 bucks a month, 40 standard reports. This is great. I could see doing this with every startup. Yeah. No, I was going to say you could use it as, like,

you could do this as part of, you know, the university and all the things you're doing, yeah. Yeah, I mean, what's interesting about this, oh, it says you own the commercial report rights. That's interesting. What I like about this is I could do this. If they had an API, every time we meet with a company,

we put it in our database, is we have a summary of that business. They have API access right there. What's that? What's that? Oh, it says, contact us for API access. Perfect, yeah. What I've been doing now is, in preparation for the great AI overhaul of our industry is,

we have a programs team call every day where, you know, the people who run Founder University Launch Accelerator get together for a 30-minute stand-up, and then we have two investment team meetings for two hours each, twice week, Tuesday and Thursday. because we process 60 new companies, well, we do 60 new meetings per week, intro meetings. And so I'm gathering all of those, and I am recording the Zooms now, storing the Zooms, transcribing because Zoom does transcripts automatically, putting the transcripts into Notion. And then I'm summarizing, I think we're using the Notion API or we're using ChatGPT4. I'm not sure which one.

Just summarize the call transcripts for our internal meetings. Summer is reaching its apex. And there is nothing worse than tossing and turning and sweating in the night because of all that summer heat, don't I know it? But the pod cover by eight sleep will keep you cool all night all the way down to 55 degrees Fahrenheit. This is going to help you wake up fully refreshed. And it's so easy. The pod cover by eight sleep fits on any bed, just like a fitted sheet. And it improves your sleep by automatically adjusting the temperature on each side of the bed based on your and your partner's individual

needs. You just set your eight sleep side of the bed to the cool temperature you like and maybe, hey, maybe your partner wants it hotter. Okay, Vivalenteferance. It can cool down and warm up and adjust based on the phases of your sleep as well as your environment. I love my eight sleep. Got one up in Tahoe, got one in the Bay Area. You know, sometimes I'm in Tahoe, it's freezing. I need to get nice and toasty warm so I get a nice sleep and I get cozy-wozy. Hey, but then sometimes in the summer In the Bay Area, we get this heat waves in August and September. Oh, my Lord. But then I bring my 8Sleep way down.

55-60 degrees. I love it. Freezing cold. Those crisp cold sheets are waiting for you. Eight Sleep's cooling technology is a lifesaver in the summer. Take it from me. Go to 8Sleep.com slash twist for exclusive summer savings on the pod cover until July 10th. That's today.

Today is the last day to get that July 4th deal. So go right now. 8Sleep.com slash twist. Hey, they're now shipping in the USA, Canada, the UK, and select countries in the EU. as well as Australia. Way to go, A-Sleep. What should I do

with that data eventually? Well, it's a good segue, J-Cal. I think, let me show you that... That's why I'm the world's greatest moderator. Yeah, exactly. Let me get my...

Or you can have C3PO moderate your podcast. Yeah. So similar to, you know, kind of what we are talking about here, what a team launched

is a chat app specific to the hacker news, hacker news. And what they've done is they've taken all what's going on in hacker news. One size are bigger? One size is bigger. Oh yeah, sure. Yeah, we can up that a bit. Welcome to chat, hN.

So this is chat hn. Vresel. So Vrescel. Explain Brcele. Yeah, so Vresel is actually crushing the game. We should spend a little bit time giving them a shout out. They are a modern hosting. like an app hosting service.

And they've been doing this for a while, but in the AI game, they're the go-to. So if you want to host any type of AI app, most of the frameworks that people are using are designed for like sort of one-click deployment into Vurcel. And they do a great job. They have very flexible plans.

And so they're like sort of the modern up-and-comer in the cloud war. I mean, you know, they're more than up-and-comer now, but they've been really crushing it. And they have a bunch of their own additional frameworks. Like in this case, they have an AI SDK to help you with, you know, sort of the chat and all those other things that they're doing a really good job of as well.

So big shout out to the Vrcel team here. And so you get this prompt. Give me the top five stories on Hacker News in markdown table format. Seems like doing tables and formats is like a great use for AI. You click that. And it gives you the title, link, score, and comment score is something they use on Hacker News to kind of give it, you know, how popular something is. And it gives you the first, it gives you the top five.

And then it says, hey, send me a message. So I guess we could ask it, what's the, give us the top, give us the five funniest comments on the first story. Let's see if that does anything. So not enough amusing comments to provide the complete five list. Yeah. Oh, interesting. How about give us the most...

Oh, what is the sentiment on threads? Interesting. So this is a proprietary data set, and it just gives you a new interface for how to process all that information. And this is the work of an analyst. So when I see this work, I don't know what you see. I see a $40 an hour person.

And if you ever want to, when I talk about hourly wages, I always extrapolate as a business owner and somebody who invest in businesses as the hour of what an hour of work costs. And then you just times that by 2000, right? Because it's 50 hours a week, 50 weeks a year, 40 hours a week, 2,000. You know, listen, if you work 50 hours a week, it's 2,500. 60 hours a week, it's 3,000. But 2,000 is a pretty good multiplier.

$40 an hour, 2,000 hours, $80,000. a year. It's a really good paying job, especially from home. And that's what an analyst would get. A researcher would get $20 an hour and a data processing person, like an offshore kind of person, would get $5 to $10,000 in Manila. So five times $2,000 is $10,000, $20,000 is $40, and $40 times $2,000 is $80,000. Just so you get an idea. $80,000 is a U.S. smart person who reads books, you know, English as a native language. A researcher is something. somebody just out of school or maybe went to a two-year and then offshore, they don't understand the context in America probably, and it's English as a second language in many

cases. So whose job do you think this replaces most? Because we keep seeing this. And I say the researcher analyst and the data formatting person keeps coming up. Yeah. So let me answer your question a little bit indirectly, which is, so first, I think given large proprietary data set, you can see. see the value of putting a chat interface on top of it, right? And so I think for you guys,

what you need to do as the next step is you have all this proprietary data now, which some of it's even being created by AI or enhanced by AI. We need to stick a chat interface in front of it. So that should be sort of one of our projects so that you can go through that and ask a general question and say, hey, have there been any other companies that have come through that have pitched us on selling ice from icebergs? And then it can go through that data and then you can get that answer quickly and you can see, you know, what was a call about and things like, you can kind of dive into it. So I think that's the, that's the first thing we're trying to show there is that the world of deploying your own chat app on your own data is really simplistic now.

And I think, you know, it's something that we should, you know, explore for, yeah, for, you know, for launch. And so we should kind of kick that off. Yeah. I think the value is, you know, in one sense, in a replacement, but I think it's the enhancement. I think if you put this in front of everyone and ask yourself how many times, Jay Call, are you, are you, how much are you relying on your memory to go back?

Oh, there was a company that did that or, you know, where did these guys end up? Got it. And now to basically have that enhancement is more valuable from a, rather than a replacement of a person or an analyst, but to basically kind of give yourself that superpower.

I think that's where it's a much better framework for value. Yeah. So let me explain this to folks. because I think you just hit a key insight. When you're running a business and at-scale business like ours, 15,000 people emailing us and filling out our forum

and sending us a pitch for a company, over 1,000 people coming to Founder University a year. We have so much data and we have 19 people in our little investment team, our little company. Now imagine, with those 19 people, we then do an analysis. We do that analysis.

And if we do that analysis of what we're doing, I would normally go to somebody in my team and say, Hey, we met with that company. They were doing a marketplace of diamonds. But we had heard two other pitches about diamonds, the one who was doing the fake diamonds and another one who was doing, you know, setting your diamonds and it was using AR. I can't remember any of them.

Now, normally you just search for diamonds. And then you get all this croft. Diamond in the rough. Somebody would be saying like, oh, this person's a diamond. Here you could say, tell me all the startups that are working in the diamond industry that we've met in this over the last five years and pull clips from their video, because we have the video now, and make me. a little dossier of that.

Show me all the marketplace companies we met Mith last year. Yes. Then go on the web and tell me how many employees they have on LinkedIn. And this is where LinkedIn has to start in paid API. I don't know, this is, you know,

Twitter and Reddit having their APIs. I think the LinkedIn API for LinkedIn team, please let us just give you money because everybody's scraping your data anyway and we can buy your data from like Israeli or, you know, companies in the Philippines have scraped all of LinkedIn already. They have it all.

And you can't stop the scrapers, and it's legal in other countries to scrape this stuff. So your terms of service means nothing, and those are the jurisdictions. So now you're left with going with gray hat people for data sources, and I'd like to put this on you for next week. If you could find some gray database sources of like Instagram, Facebook, profiles, whatever, I'm curious about the gray market underground. So anybody has information, send it to producers at This Week in Startups.com or Sannie, what's your Twitter? At Sundeep. At Sundeep. That's what I think is going to be super interesting is some of this great market data. But imagine if I could ping the actual API and it would come back and tell me, hey, this company has 60 employees when you met with them. They had 20. And I would pay for that. I would pay some amount for database calls. It could be an incredible revenue stream for LinkedIn. And I get pinged by people. Or like, would you like a billion LinkedIn reference database records of?

CTOs of founders. Like, this stuff's all been scraped already, so don't be precious. But all of that back and forth in meetings and research, somebody's like, oh, I'll get back to you in an hour.

It's going to be like, I didn't even need to waste somebody's time. So, Jake, I know you've been on this for a bit. And one new segment I wanted to start. And so we're not fully up and running with this yet,

but we have the framework. And I think we're going to start building this out. We're going to start building this out over the next couple of episodes. And so you've been going on about auto GPDs. And so what I have here is a basic framework

of an auto GPD that I've created. Define auto GPD for the audience who's catching up. So yeah, an auto GPD is an agent

that uses LLMs to work through a problem through what's called like a chain of thought. So you'll give it like a high level problem statement and then it will come up

with a set of tasks to solve that on its own. And then it will use it's access to different sub-agents it has to solve that task. And so this is, you know, from a demo sense, and J-Cal, I'm going to add you actually to this replet so you can basically participate in the, you know, kind of in the live, you know, kind of the live experience that we're going to have.

And so, like, I'm going to just start with this. This isn't with LinkedIn, and you'll just see the power of what we can do here. So you can say, can you come up? with a schedule of summer league. Let's say, let's call it NBA summer league games for me to watch. And so what this is going to do is, so I just gave it something generic. And it's going to say here really quickly,

well, I need to start by searching for summer league games. And then it's going to figure out, okay, I found a place to get it. So I need to analyze those results. And then it found that it can get. it from this MBA page, right? And then it'll start working to put it into. Now, this is just, it's a basic framework.

We're not, like, it's not fully up running yet. Now, what are we using here again? One more time, what is this called? So, so this is basically something we've built from scratch, but it's using two main, it's, it's, it's, we're running it in replet. So we're going to give them a shout out. It's using language. Yep.

And we're using Langchane, which is sort of like a, um, a language model by Facebook. No, no, no. It's not a like, we're using opening I. That's the other one, yeah. Yeah, Langchain is a scaffolding framework for working with LLMs. Yes. Right. And we're using another, another server called SERP API,

which is basically a search engine results API. And so it basically interacts with Google. And so that's what the two keys you see here. Is Google allow that? Or it's just, how does it actually do that? It has your machine do the search and then rips the HTML page apart? No, no, no. there are whole companies that exist for this now, right?

And so that, so this is a company called SERP API. Wow, I've never heard of this. I'm learning it. Yeah. And so they exist. You obviously have paid use cases with them. And they basically, you give it a search and they'll return you back like the search engine results in a in a consumable format, which is like sort of on the right hand side

what I'm describing here, like as JSON. SERP API.com. which is search engine result page. Wow. You're very familiar with this, JCal. Yeah, sure, of course. Yeah. And there's a few of these,

but this is the one that I've decided to use in this particular demo here. And then that's how this auto-GPT, which I gave it, like, can you come up with lists of Summer League games? Obviously, this is not, it doesn't have the 2021 limitation. It uses a SERP API to figure out,

well, that result's going to come from this MBA.com page summer league schedule. and then it will start to kind of parse through that document to come up with your list. And so this is sort of the beginnings of our auto-GPT, and we're going to start working through this. So I wanted to kick this off today for us. Right. Yeah. So how do you propose?

What's the next step in this? Well, the next step is like you've had a bunch of different use cases, right? You rattled one off today. Let's just build towards now that we have the Bracic scaffolding, and we're going to get you back to writing code again, J.Kal. We'll basically, yeah, so, and you can see this thing is not very long. It's only like 60 lines of code. That's the beauty here.

Because all of that is being abstracted by API calls. Exactly. These days, writing code is really like hitting 20 different APIs and blending whatever you get back, right? I mean, it's really fascinating how code has changed. Yeah, and honestly, like in this particular case, like the, The code to use that SERP service is right here, right? This SERP API wrapper. And then this is my API key that I have with them.

In the example you were talking about with LinkedIn, if LinkedIn would want to work with us, they would offer a key that we would pay for. We would import their agent. And then we have another agent here that wasn't search, but that was like LinkedIn that we would go and get that information from. But there's plenty of other services out there. LinkedIn isn't there yet.

But that's how we're going to get back into it. All right. So here's what? I would like to do. I want to create one of these. Okay. I'm going to run this up to flybook. That finds new startups that aren't in our database already.

Okay. And then finds the founders, puts them into a category, does some sort of analysis of their business, right? Like, finds out what the startup does, sends it to that other API from the people who do venture, what's it called? Ventura. So we find a startup somewhere that didn't exist before. So an announcement of a new startup. That could happen on Hacker News, Reddit, Twitter, LinkedIn. People could announce a new startup.

So we find announcing my new startup, date being today. So it was published in the last 10 days, let's say. So in the last 10 days, somebody publishes, I'm announcing my new startup. And we could do that with the search engine result page by passing a query to it. So we could say Google, and you can do this in a chat window, or I could do it in one. It would say to Google, new startup announcement. And then you would go to all filters, advanced search. Yeah, no, no, where is it?

You go to tools, and then I would pick not any time, but I would say in the past week, it would come up with a search result, which is crunch base, EU startups, Alleywatch. Yeah. And we would try to find on those sites, new startups, and find the URL. of the startup. Once we have the URL of the startup, then we could find their LinkedIn profile page. We can find their Twitter profile and then try to get some information on that startup and then propose them to a researcher analyst inside our team and then click book meeting. That would be amazing. Because by the way, we do that. We call it qualified hunting. We try to have our researchers

and analysts hunt for companies that never applied to our programs. And so that's hunting. And you product hunt, speaking of hunting, product hunt has like every day new products coming in there. So we just look at the URLs of every new product on product hunt, which is why I think Angelus bought it. It's because they wanted to have first dibs on all the new startups and technology. So very cool. All right, listen, you're in the technology industry. You know what CARDA is. Carta is the leading venture capital and equity management platform.

What does that mean? That means they manage your cap table. They manage your venture fund. They are the experts at that. and they do such an amazing job. All of my startups use it. They all love it. But I've got huge news to share with you.

Carta now lets you syndicate an SPV. So you can create your own syndicate through Carta. That's right. That's what I've done. I have a syndicate. When I have a great investment, like say,com. I put a little bit of money in 50K in Com. And then $328,000 came in with the syndicate.

You can then go build your own SPVs and use Carta as your partner. They are one of the most trusted names in business. You know that. And they're used by more than 4,500 funds. that represent over $120 billion in assets under administration. And they will support you on every stage of the fundraising journey from your first syndicate to building your global venture capital firm. You can raise and deploy from anywhere because Carter offers U.S. and international SBVs.

And they fund in different domiciles. That means different locations, right? Fancy word for locations, GEOs. Carter provides this automated back office solution so you can focus on what's important. building and finding great startups, building these amazing relationships, investing in great startups. So it's a very simple call to action.

I want you to go to carda.com, C-A-R-T-A-com, you know that, but use the code twist to get 10% off your first SPV. That's Carta.com and use the code twist. Yeah, so that's our, I think that's our little project. So we'll take that and we'll kind of expand it out and we'll kind of work through it and we'll show the results every episode.

You know, I'm very interested in Facebook's approach to AI. There was this, they have their own language model that was leaked, quote, unquote. Lama, Lama. And it's very popular on hugging face and other places. So Lama, they claim, was accidentally leaked by Facebook itself so that it would kind of undercut Google Bard and proprietary stuff like closed AI, chat chp T4. Who knows if that's true, but they keep putting out public stuff. So they're still on the public releasing of information open source tip, correct? Or are they now circling the wagons and being closed?

They've been very open and public. If there's things that they haven't released yet, they've just said they're going to release them. There's one that they did recently. We can pull it up as well. But I think, you know, this is probably a better segue into threats. and you know, I actually was having this thought over the weekend, which is, you know, is threads perhaps a way to get the data set for a hive mine? Because if you look at any of their existing products, they've all evolved, right? You know, Instagram is, yeah, oh, yeah, there we go.

And, you know, Instagram is not going to give you sort of hive mind because it's going to give you video and pictures, right? Yeah. WhatsApp, if you mind that data, it's going to give you proprietary chats. And Facebook has just evolved into something I don't really understand anymore. Maybe you can, you can chime in there. If you want birth announcements and bar mitzvahs and retirement parties, you know, kid photos. It's basically for moms and dads and grandma and grandma. I use it like our Yahoo groups.

There's some great groups I'm in there. Yes, the group's product is very subtly, I think, put a lot of new life into that. Because the general feed is kind of like boring and repetitive. Oh, it's your birthday, happy birthday. Literally, it's a birthday announcement website or a birth announcement website even. So why, let me ask you one question here. I know we'll get back to threads. Facebook, I have my own theories, but I'm curious of yours.

Open AI went closed. Google published all this stuff, including TensorFlow and kind of regrets it, I think, now. They're kind of closed. Transformers and everything. So they're closed. or somewhat closed and they're not doing they're not announcing the papers anymore

is what I heard the scientists are not announcing their work as often no I would say that's not fully true you know they just launched a paper that we're leveraging like around SQL Palm

just a few weeks ago so I think they're continuing to do that their approach from a cloud perspective Google cloud perspective has been hey we're going to have

our own proprietary models and we'll also host open models and models from other proprietary companies as well. So if you are a Google Cloud user, you can use their Palm models, which are their proprietary ones. They also support all the open models inside of their, you know, inside of their frameworks, right? Got it. And also they have models from companies like Anthropic, right?

And so I feel like they have a really kind of a great open approach. Why do some people choose open and some people choose closed? It's sort of the natural arc of the tech industry, right? Like when operating systems first started, they were all closed, right? And then quickly we had an evolution to more open source Linux or Unix-based operating systems, right? You know, BSD and then Linux. And so I think people try to build. Now, there's always these conflicting forces.

I'll speak from a developer standpoint, right? When something is closed and owned by a company, it can move very, very fast. Because when it's open source, you have to work your way through the community. Now, the way companies have got around this, let's talk about say Red Hat and Linux, is that they represented almost like 90% of all the folks that had the control on the project, that were the developers. And so big companies can embrace open source and not get stuck in sort of some of the politics that emerge inside as well.

And then there's the reason that, you know, you think you have a lead and you want to basically create a moat for yourself. And so those are the main reasons that generally pop up. The best way I heard it explained to me was when you're behind, you open source. When you're ahead, you're closed. And if you look at, say, Windows, they had a monopoly on the desktop, closed. They don't need to be open source. But then you look at Google very far behind.

They have a monopoly on search. So when they talk about search, their algorithm for search is closed. You can't understand how page. are ranked. But then you look at Android, they went open because they were so far behind iOS and stuff like that. So they went open source and they wanted to be on everybody's phone. So I think Facebook is closed when it comes to their graphs and everything, right? Facebook, Instagram. They used to be open when all the companies like were created, you know, like Zinga and all. And then they closed the graph up at some point. When they had a lead and nobody can compete

with them, which is exactly what Open AI did. They became closed AI. So fascinating. Yeah. So on threads, what are your, I logged in. I immediately got dunked on because they're like, oh, friend of Elon is on the meds. So I like literally did like two, I did like two posts to it, just like, hello world. And then I was like, oh my God, Zuck, sucks is a little copycat. And then Zuck winds up replying to me.

Oh, okay. Yeah. And he put concerning. What was his reply? I think he did concerning with a wink. You know how Elon will just give a one word, reply? Yeah. Like, interesting.

So not only is Zuckerberg copying Twitter now, which in fairness is copying Jack more than copying Elon, right? Yeah. But he's obviously still obsessed with Elon as well and Twitter. I mean, Zuck's been obsessed with Twitter from the beginning. He was going to buy it and he's got a lot of comments. But he actually took on Elon's replying with one word replies, which I thought was hilarious.

Super Elon thing to do. Huge engagement with like people, right? Well, no, now he's like, yeah, I'm going to engage to get my social. It's like, okay, yeah, we know that. like, and then he's like, I'm going to be outrageous. I'm going to fight Elon in the Octagon. So Zuckerberg is clearly like obsessed with Elon and, you know, copying him. Yeah.

Just like he was with Snapchat for a while and before that. I guess Instagram, which he bought, he was obsessed with that for a little while. But I think it's a perfectly serviceable copy of Twitter. They obviously rushed it because it's feature light. It doesn't have a lot of the features. But it's a different graph. Yeah. I found it was like a lot of people who don't have interest.

things to say, trying to say things with words instead of pictures. Does that make sense? You nailed it. Like, that's the issue, right? Is that, look, I think Instagram has a huge community. I think they have a ton of engagement, probably, you know, maybe orders of magnitude more than Twitter does. But it's for a different purpose.

Yes. And it's kind of focused on video and photos. Yeah. Many of times those videos and photos are not a real, representation of what's happening. It's sort of like, hey, look at me. They're staged. Yeah, they're staged, right?

Highly produced. Yeah, highly produced. And so, and look, people want that. And I think it's great. And I think subsequently, the graph you create when you have an Instagram account is around that as well, right? You've decided to curate something there. And I think when you go to Twitter, whether you have your list of folks that you follow

or like a 4U feed, it's a completely different input into the system, right? Which is, you know, no one on Instagram is sharing archive papers, right? You know, from these research papers, right? No. Or like debating the subtle points of the Ukraine war. Yeah.

Russia's invasion. China, Taiwan is not a big topic. Yeah. And, you know, it's so funny. Like what I kind of posted this thing, which was like, oh, like, for like 24 hours, it felt like all the threads were disappearing on Twitter, like how to, you know, do the, you know, make money from AI or, you know, you see all. They were gone for about 24 hours and then they all came back because that, again, vice versa, that community there doesn't want to read that stuff, right? They don't want to read the, what's the latest in AI and let me see the last five cool AI tools that were launched and how can I make money using them?

This is the key thing that for founders who are listening, this is the key thing you have to understand. about making a clone. If you make a clone of another product and that product is doing a good job and it's like it's servicing its purpose and it was the first, it's the at-scale one, I wouldn't say the first. You have to be so much better.

And it might be that the team that built that product has already gotten it to so much better that the users can't tell the difference. And so if you look at threads versus Twitter, there's nothing there that's different or better. And in fact, obviously, still playing catch up. Until they have something that's dramatically better for

some reason, it's going to have a moderate success, right? It's the same thing held true for search engines for a long time. People just kept making search engines until page rang came out and the results were noticeably better, like 10 times better. There was no reason to go

over to Google. Yahoo got you or Lycos or excite. Yeah, they all got you, Alta Vista, Magellan. These things all got you to a similar result. You were typing Pepsi or Coke and it would get to the Pepe or Croke website. Yeah. But when you typed in Pepsi versus Coke and it got you to some scholarly article about the Pepsi challenge, you're like, oh wow, this is interesting. It's better. Which is why ChatGPT feels so uniquely different. So there has to be something uniquely different for people to change. I've learned this the hard way many times building products. Yeah. Now, look, the one edge where, you know, Instagram threads,

meta will have is brands. Because a lot of, you know, when I look at accounts, like it's a mix of things you follow. And if you follow, you know, like kind of larger brands for a reason, then, you know, I think Instagram is starting to surface that a little bit better. And at least from my understanding from a few different folks I spoke to is they went and targeted folks. And I don't mean brands just by company, but brands that are people as well. Because, you know, they had Mr. Beast on there and other folks.

And so did they pay for him to give away at Tesla? that was like a very specific troll. Yeah, I'm not sure what the economic arrangement was. I guess Zuckerberg gave him a million dollars or something to start posting over there or paid for the giveaway. Well, I definitely heard that from someone, you know, is reliable to say like they definitely had a pretty, it started actually when the whole thing happened with the subscriber ticks. The verified tics, right? The verified was the start. Yeah, they copied the verified checkbox. No, no, no, no.

I'm saying when, when Twitter started, no, when Twitter started to change the policy around, yes, that's when they started going after the high profile folks. Got it. Smart. Yeah. Yeah. They saw that as like sort of a, like a misstep.

Yeah, it's an attack vector. Yeah. And so that's where they kind of went after these brands, right? And so now the challenge is, is like, so Mr. Beast is the most interesting on YouTube of all places. And his Twitter and his, you know, threads are sort of, okay, I don't know where to go. I also think there's, yeah, yeah, they're secondary. The struggle also is, and, you know, people start saying this thing, well, you'll just post

in both places. But if you're trying to maintain a conversation and you're trying to kind of be, it's very hard to do that across both platforms now. And so I think that's, that's going to be the challenge and what will end up emerging. And I think, I think we had it in our show. show notes. I think Adam had a really good post saying, look, I just think we're, they themselves are saying, I think we're just going to become two different things. I don't think they're trying to, here we go, right? And so why don't you read this out, Jake House? The goal isn't to replace Twitter. The goal is to create a public square for communities on Instagram that never really embraced Twitter and for communities on Twitter and other platforms that are interested in a less angry place for conversations, but not all of Twitter. Politics and hard news are inevitably going to show up on threads they have on Instagram as well, certain extent, but we're not going to do anything to encourage those

verticals. So he's basically telling journalists, we really don't want you on here, talking about Ukraine, talking about AI threats. Yeah. Because that Debbie Downer news is not where advertisers want to be. So this is another, I guess, if you were going to make a, if you're going to list all the attack vectors for Twitter, even pre-Elon owning it, the fact is this highly intelligent people debating very controversial subjects, gender, woke politics, politics, wars, geopolitics, technological, edge cases and dark stuff. Advertisers don't always want to be next to that. Some don't mind.

They just want audience, but other ones might be very brand conscious, right? So that's another attack vector. But I'm surprised they didn't embrace journalists because journalists tend to be on the woke side of things. There may not be fans of Elon, and they probably feel very bad about losing their blue, I know they felt very bad about the blue checkmark thing. I mean, Kara Swisher talks about, and Professor Coltakes, like they talk about Elon incessantly on Twitter. And they've invested in post news or something, like the competitor. And I'm like, why are you guys on Twitter if you're investors in post? Go to post news post. Well, and also Meta's had this like battle with news organizations.

Yes. Like the big thing happened in Canada recently in Australia as well. Very good pull. Yeah. Right. where, you know, they have this challenging relationship with news organizations now, which I think why that was called up. Where news organizations want to get paid and Zuckerberg does not like to share revenue. So I have a prediction here. You know what I think? Let's hear it. Well, I think this is a fight that Zuckerberg does not want to lose.

So I don't think, like, remember he came up with like poke to attack Snapchat and he did it like, I think he did four or five competitors. to Snapchat before just saying, you know what, screw it, put it in Instagram, I give up. I'm not going to make a standalone thing.

I have a feeling that threads will become a feature inside of Instagram and they'll just be like threads next to photos. Just like they've like a tab kind of a situation. Because I think as a second app, it doesn't work. I would rather have a single app like Instagram.

I don't want to give them like the roadmap here, but I think having a tab with threads without images. So images are not allowed in it. It's, you know, you can't attach an image. You can only do text. That would be a better place for it to live. And then you get 100.

I don't have to rebuild the graph. Like, I literally did not check, follow everybody. Because I don't, I didn't want to do that and send a bunch of alerts out. I accidentally did it. And it was like, it's been a mess. Well, yeah. So I was like, I'm not just going to follow everybody. I'll restart my graph.

And then I was like, I do not feel like playing the rebuild my social graph for the 50th time in my life. I don't know how many times. I don't know how many times. I don't know the last. last time I had, oh, Clubhouse was the last time I started rebuilding. I was like, this is just not worth it. I think the audience is exhausted with that, but I do think he's, I do like your angle of this is how to get a data set. So if they did this and it didn't make any money for them, but they did get like more text and discussions that they could use. And topics of today's,

like what are people talking about that we can use for an LLM? Because remember, right, that's what this is all going to go to is having these LLMs to help people in decision making and things like that. And I think this is a great way to get one of those datasets. This is why Google should have not given up on doing social. They should have kept doing it. They did Google Buzz, which was an extraordinary social network before Google Plus, which was actually very well designed. But when they did Google Buzz, they had another one with a weird name too. Orkin was one they did in South America.

during 20% time when you could just release products with like you're on your Fridays at Google there's something called 20% time where Larry and Sergey

let you work on whatever you want on Fridays pretty cool idea but what Google Buzz did was and maybe producer Nick you can go find Google Buzz screenshots

and I wrote a blog post about this holy cow Google Buzz is going to like kill Facebook now of course it's easy to dunk on me but they gave up on it for privacy reasons

which they shouldn't have what you see here when you look at Google Buzz and this is your inbox, right Gmail, and then right under it was Buzz and it told you how your updates. So you get in there and you get like a Twitter

or a Facebook box, hey, what are you doing? You type into it and then you see everybody else's. It was brilliant. It lived in your Gmail box. Google should go back to this because you can write an email or you can just give an update

inside of your email and it's right there. It lived in a perfect spot. Google gave up too early. Sometimes it takes five or six swings to get something right. Google did three swings. swings, Orkitt, Buzz, Google Plus. If they had done the fourth and the fifth swing, I believe they would have built a co-exister.

Maybe not one that beat it. And then do you have my blog post where I wrote about this? Google Bug is brilliant, like groundbreaking, game-changing brilliant. This is when Business Insider used to ask to reprimis. This is 2010. Google Buzz 1.0.0 was better than Facebook after six or seven years. True statement. Facebook's history is one filled with stealing other people's innovations.

I wrote this in 2010, 13 years ago, and doing them better. Zuckerberg has stolen every idea Evan Williams and the Twitterium have released. How ironic that now Google has out Facebook. Facebook 3. Google has an excellent privacy record, and Facebook is a disaster. Most folks do not trust Zuckerberg and Facebook because of their privacy record. It's pretty crazy. Google Buzz auto-generates your network.

This is much better a process than Facebook's. Google Buzz is way faster than the sluggish Facebook. This is a huge advantage. Buzz puts relies and updates into your Gmail as threads. This is brilliant and a huge advantage. Anyway, my assessment was perfect. They just gave up. They turned off Google Buzz because they got too many privacy complaints.

And this is what happens to a big company. Oh, look at this. You can sign up for Jason's excellent email here. That's hilarious. Anyway, there have folks. And so I think if threads just keeps going and Zuckerberg is really good. And so social, like no one. new, at least, you know, obviously back then, the end game now is for the world's best

LLM, which will be sort of the underlying API for everything. And it, you know, circles back around to what Twitter's real value means in this ecosystem today versus everyone else, especially with all the work that's happened around, you know, scraping and turning off scraping and monetization from scraping. I think it's really, really fascinating. All right. Any lightning round stuff you want to go through here? I had a couple of other arms of docket or we can leave them for next week. The other story that I thought was really good was the inflection AI.

Oh yeah. Explain this. If you wanted to get into that. A huge number of GPUs, right? Yeah. I saw this go by. Yeah. Yeah. So Inflection AI, it started by a co-founder of Deep Mine, Mustafa, and Reed Hoffman.

And, you know, they raised a massive amount of money, I think, $1.3 billion. dollars on a significant valuation, maybe $4 billion. A couple of interesting things here, right? The list of investors like Microsoft, Nvidia, Reid, Bill Gates, Eric Schmidt. Wait a second. So wait, Bill Gates and Microsoft are mortal enemies with Eric Schmidt and Google. And Microsoft and Bill Gates are massive investors or Microsoft's massive investors in Open AI. So they are obviously hedging their bets here.

why not invest in two language models? Two better than one. Is that what I'm seeing here? Yeah, and this one's even more interesting because alongside of the language model bet, this one is a significant hardware spend, right? And I believe, you know, they sort of are planning to have something like

20,000 H-100s available in a cluster. These are Nvidia's AI computers, cards, whatever I call them. Exactly. Exactly. Yeah, their AI system on a chip or something. Yeah. More than that.

And so, yeah, it's really... So they're going to spend hundreds of millions of dollars building that cluster. Correct. And now is that... Maybe even close to a billion. I think if you do the numbers, like the majority of that 1.3 billion would be consumed by the creation of that cluster. Wow. Because those cost how much now?

150,000 or something? 40,000? I think retail is like, no, 20,000. Oh, 20,000. Okay. So a thousand would be 20 million, 10,000 would be 200 million, 20,000 would be 400 million. So 400 million as just a floor number. Yeah, yeah.

And you know, you got to add on other things. That's just the immediate cost. And you have to rack them somewhere. So wait a second. Why isn't this part of Azure? Why are they just, I wonder if they're buying them and putting them in Azure's cloud or if they're buying them and building their own cloud and own co-location facility. Well, I time Microsoft's part of a deal that involves, you know, some kind of cloud infrastructure.

It's usually part of their trade is that, hey, you're, you know, you're going to use our infrastructure in some way, shape, or form, right? And they've been really good at that. So that could have been part of the deal. But we don't know that. I wonder if that would fall into round-tripping, you know, where these deals could greatly enhance Microsofts and NVIDias. These kind of deals could optically make NVIDIA's stock and Microsoft stock look more. valuable and the value of the stock would go up because you just got a $400 million order. Remember they said they were going to... Actually, in the notes here, it's 40K, so it's actually $740 million. So $700 million comes in, right? Or something like that. In new orders, that's going to make

Nvidia stock a way up. So let's say, Nvidia gave them $500 million, let's say of the $1.3 billion, Nvidia gave them $400 million, a third of it. Yeah. They gave them $400 million. And Then they bought $800 million worth of hardware. They're basically just shipping the $400 million back to them. Yeah. And they must have a 50% margin on those machines. So essentially, Nvidia stock goes massively up by billions of dollars because of that order.

Yeah. And they got their money back. And the startup is essentially, in a way, painting the tape or wash trading in some way, Nvidia stock. Now, this could be completely inadvertent, but that is the result that will happen here.

this order will make more people buy Nvidia stock and the money comes right back to Nvidia. This is something the SEC is going to be all over. Yeah. Well, Nvidia is the real winner here and even Microsoft, right? Because maybe all of those end up in a Microsoft data center, right? So if they're a big part of this. Microsoft put in $400 million. Yeah.

And then they host them and then they send the credits back and they give $200, $500 million back to Microsoft with the coming years. That's a round trip too. Yeah. I don't know how people who are so sophisticated are doing something that sends up so many red flags they must have some plan to

make this clean. It's fascinating and we're just talking about the hardware side of it, J-Cal but like, you know, flipping back around. I don't know if you've had a chance to talk about it, but you know, Twitter did shut off access to like basically almost every service, right? And it was even breaking eye message and signal and everything else. And so in today's world, like, how do you stand something up?

And I know you had a tweet. You asked, hey, what's the go-to service? And I think I linked to something there, right? There's a couple of things out there. But it's also fascinating, where are these folks going to get data from? And then we had to follow on saying, hey, look, if you've ever trained a model, if you've ever trained a model and you have some type of restricted data in there, It is in the model forever until you retrain.

And so this whole world is super interesting because if you have that much investment on hardware, where are you getting the associated data from that is not restricted at this point, such that you can leverage such a huge amount of hardware that you need. Do you think the next set of models will be weaker than the previous set because of this? I think if you're creating a model from scratch, the answer is yes. Because when the models were previously created, so I think, And, you know, I say this with about like 90% certainty. Twitter changed their policies post Elon's takeover.

And so if you were trained off any data from before then and your, you know, your models are using that for, you know, their training data and they can use it for their own reasoning, I think it's fine. But I think if you have any data beyond that point, you can't. And I think it's going to create a real problem for folks that are starting from scratch today. That's what I want to be. Let me ask you a technical question. If you did train on Reddit, Twitter, or Core's dataset. In the past, which these things did. In the past. And we know they have because people have proven it, right? Because you can ask the LLM and it will pull information

from it, right? So there's no doubt that they did that. Correct. I'm not going to pick any particular company. So if you did train on that, does that mean you have that information stored in some giant database. In other words, you took every single core question and you're storing them somewhere, or you have the resulting hashes of those. So it's masked. Therefore, if you were to say, give me, I want you to go into your LLM, like I have a cause of action, like a legal action, against somebody who made an LLM. Let's say an open source one even, you know. And let's say the Facebook one, which is called Lama. Let's say Lama had crawled Reddit. and we know that.

Could they rip out what it learned from Lama? What data is still stored in the language model? So, you know, it's probably worth a longer discussion with some demo, and I will queue it up for the next one, and I'll put a demo together. When these things are used, when data is used to train a model, the data is basically turned into an embedding. And then embedding looks like a number between minus one and one. And so if you remember from the summit, I kind of gave an example. So it's never stored as its kind of holistic nature of whatever you took. It's basically broken down by the model tokenized and then turned into a probability,

which is then tied into, you know, what probabilities does this mean to the previous word, the next word? And so it's not kind of stored as a holistic thing where they... Can that be untangled? could it be reverse engineered to prove that these words it's giving in an example came from this place? You will be able to do that, right? Because in those cases, you can ask it a question, and you'll just use an example. Like, you can go to Open AI and you can ask it a question about, you know, what does it know about Elon's tweets? And I'll say, well, I don't know anything after September 2021, but before that I know the following.

And I think the tweet thread that I had that you retweeted that Elon commented on, I had a little share, you know, like the. Open AI share in there that showed it's kind of data. I don't know if Nicky want to pull that one up, but that one showed the history of what it understood of the training data from tweets that it had prior to September 2021. Yeah. So this is kind of interesting because I think they know this. And now when you ask it to give you tweets,

it says as an AI language model, I don't have real-time access to specific individual social media accounts or tweet history. Or their tweet history. Therefore, I don't have information on Jason Kalakannis' top tweet topics from 2019 to 21. So they are preparing at OpenEI, I think, for eventually having to rip out all the tweets. So the balkanization has happened. It's so funny, the guys on All In were like, this will never happen.

And I was like, I think this is guaranteed to happen. Well, they don't have to rip out all the tweets, right? Because it'll depend on when the terms changed. You know, they could argue, even if the terms of service didn't say that, that they've created a derivative product and they want them to remove it. Yeah, but prior to, prior to Elon's takeover, what if they were paying for? Oh, that's different. Yeah.

Who knows what that contract said, yeah. Exactly. We don't know that. Even if there was a contract. The contract might not have taken into account AI. So they can then, but it is interesting. Like, you can't get tweet data any, or I don't know if you ever could, but I'm using GPT3.5. So here's my, you know, from the tweet that we had together, which is, you know, this one.

And so this is what I said. I said, hey. Summary of the last five tweets, email, Elon speech I've accessed. Okay. And it says, I don't. And then I said, what is the general theme of Elon tweets in your training data? And then it says, you know, as of the cutoff, SpaceX, Tesla, AI, cryptocurrency, humor, and pop culture, personal benefits and opinions. I wonder if we got that from like business insider, Wall Street Journal.

topics about his Twitter or Twitter data itself. So being able to rip it out, possible, not possible? Do you think they built in a kill switch to be able to remove stuff knowing that this would happen?

I mean, Sam Holman and Greg are smart. They had to anticipate people would not be happy about this and they did it anyway. They broke the rules to make the model. As far as, you know, my understanding and experience,

you have to retrain from scratch. you cannot take things out. Now, are they going to retrain from scratch anyway? Is that the best practice when they make GPT-5? Are they starting from scratch? Or are they taking GPT-4's learnings and then building on top of it? What's the better thing to do? So this is an interesting topic.

Maybe we're spending a minute or two on. One of the things that Sam has been saying and others have been saying something similar, including Brad and he talks to a lot of people, Sam has been pretty open about they're not training another model right now. and where the majority of their energy is focused on is taking the models they already have and enhancing them so that they can have more memory and become personalized. And so today, whenever you go to Open AI, you start from scratch. You have a history of everything you've done, but there's no collective learning from all of what you've said to say,

okay, I kind of know the theme of what Jason wants, and maybe he wants me to always answer things like a pirate, because every time he comes in and says, reply like a pirate, right? He always wants things in table format with short sentences. Exactly. And so one of the things that he's been very explicit about, they're not training a GPT5 yet, but they're spending their energy around these kind of incillary things to make the model much more personal and have memory related. If they have customers now, see, this is the burden of customers. Yeah.

Because they have customers, the customers are pointing out all the weaknesses. So now they start getting into the edge cases or how do I, make this more polished. Once you have a car on the road, you know, when the Tesla Model S comes out, now all you've got is feedback about the Model S. This should change, this should change, and you start going down the punch list of to-do items as opposed to making the Model 3 or the model Y and having a fresh start with a new platform. So this is going to be their challenge. I think it's partly that, which is they've got customers and they're stuck on it. I also think it's partly maybe it's good enough. And it's become such a gray reasoning engine. Like,

our little experiment we're going to work on J-KEL, where it already knows sort of how to take on tasks and figure things out. It just needs a set of sub-agents to do what it needs. And we don't need to make the larger model any better because we don't want it to be an information retrieval system. We'll have agents

use SERP API and other things we've talked about. And it's good enough. Like, they may have just realized that at this point. Yeah. Okay. This has been another episode of this week in startups. Our AI edition. Sandeep Madra, Sunny.

You're so great. Please don't sell your company. Make it into a unicorn. Let's get to a billion dollars on this one. Please. Okay. I don't know what I invested. I'd probably invested at a $30, $40 million valuation.

I need to get a $20,000. Anything for you. I mean, I just want to turn that $250 I put in into $25 million. Is that too much to ask? You're an LP in the funds. I mean, just get me $100. On this. God, 100x is just so great.

I know you're working with some companies. You can't say the majority of the names, but if there is a corporate enterprise company out there or a category of company that you can do, definitive intelligence can do great work for right now. And I know you have a short, you don't have an unlimited list of open slots on the dance card, as it were. But if there were one or two dream customers for you,

which would be the dream customers? Yeah, I think folks that have made giant investments into data infrastructure, so, you know, folks that have put a lot of money in to creating data lakes or warehouses, and they want to extract more value from that. They want to do it in a way that leverages AI, not just from humans, but AI automatically. So imagine there's agents that can look at your data,

whether, you know, all day, all night, and kind of find the insights are looking for. I think those are the ideal set of customers for us. So I could see an e-commerce company with a lot of data, a finance or a fintech company with a lot of data, your Robin Hood, you got a huge amount of trading volume. Yep. Can you just sit there and ask questions to the Robin Hood data set?

Would be incredible, right? Tell me about trades. Tell me about, you know, what was popular, what were the most popular stocks last year compared to this year? You know, which ones have fallen the most? Which ones get the most amount of chat? Or people with a lot of data like Reddit. So if someone like Reddit needed AI help, they could hire your firm to make interface or even Twitter. Or even Twitter needed people help with data.

They could hire your firm. Yeah, yeah. Got it. That's kind of the ideal type of customer. Yeah, and look like, you know, that's what we're woken on. We're very excited. So you're sunny at definitive.AI or IO? Definitive.io.

Definitive.com. All right, everybody. We'll see you next time on this week's startups. Bye-bye. Thanks, partners and sponsors.