This Week in Startups - AI DEMOS: Google’s NotebookLM, Bard’s Gemini upgrade, Magnific’s image upscaler, & more! | E1862

Episode Date: December 12, 2023

This Week in Startups is brought to you by… Embroker. The Embroker Startup Insurance Program helps startups secure the most important types of insurance at a lower cost and with less hassle. Save up to 20% off of traditional insurance today at Embroker.com/twist. While you’re there, get an extra 10% off using offer code TWIST. 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. Northwest Registered Agent. When starting your business, it's important to use a service that will actually help you. Northwest Registered Agent is that service. They'll form your company fast, give you the documents you need to open a business bank account, and even provide you with mail scanning and a business address to keep your personal privacy intact. Visit http://northwestregisteredagent.com/twist to get a 60% discount on your next LLC. * Today’s show: Sunny Madra joins Jason to demo Google’s NotebookLM (5:00), Bard’s new capabilities after Gemini upgrade (17:13), Mixtral’s 8x7B model (53:59), and much more! * TIMESTAMPS (0:00) Sunny Madra joins Jason (5:00) Sunny demos Google NotebookLM (11:22) Embroker - Use code TWIST to get an extra 10% off insurance at https://Embroker.com/twist (17:13) Bard's Gemini upgrade and Google's branding challenges (24:45) Ars Technica's experimental test bed (28:17) Squarespace - Use offer code TWIST to save 10% off your first purchase of a website or domain at https://Squarespace.com/twist (29:14) Unpacking Bard's enhanced capabilities (38:44) Northwest Registered Agent - Get a 60% discount on your next LLC at http://northwestregisteredagent.com/twist (39:50) Evaluating Gemini, Its market position, and the GPT Landscape (53:59) Sunny demos Mixtral's 8x7B model (1:07:38) Sunny creates an AI influencer using Juggernaut XL model in Google Colab (1:18:24) Sunny demos Magnific’s powerful image upscaler * LINKS: https://arstechnica.com/ai/2023/12/chatgpt-vs-google-bard-round-2-how-does-the-new-gemini-model-fare/https://twitter.com/sundarpichai/status/1732433036929589301https://www.youtube.com/watch?v=K4pX1VAxaAIhttps://www.youtube.com/watch?v=kna9E_3kFF0&list=PL24nOpPUQlbYd1U349UDH2rrPaWWreM79&index=3https://notebooklm.google.com/?pli=1poe.comhttps://magnific.ai/upgrade/ * Follow Sunny: https://twitter.com/sundeep * Follow Jason: X: https://twitter.com/jason Instagram: https://www.instagram.com/jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis * Great 2023 interviews: Steve Huffman, Brian Chesky, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland * Check out Jason’s suite of newsletters: https://substack.com/@calacanis * 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

We now have one, two, three, four, five, six brands. This is the message to Sundar. We've talked about this, you and I. Okay. Branding at Google is challenging. It's challenging. You call it Google Bard. A much better idea would have been to take the YouTube approach and just call it Bard.

And get the domain named Bard.com and let people go to bard.com and have a completely unique, verticalized experience without the Google branding. Yeah. And then nobody needs to know about Deep Mind. Nobody needs to know about Gemini. It's just Bard. And Bard is a new product from Google. Yeah.

The end. This weekend startups is brought to you by Embroker's startup insurance program help startup secure the most important types of insurance at a lower cost and with less hassle. Save up to 20% off of traditional insurance today at imbroker.com slash twist. While you're there, get an extra 10% off using offer

code twist. 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. And Northwest Registered Agent. When starting your business, it's important to use a service that will actually help you. Northwest Registered Agent is that service. They'll form your company fast, give you the documents you need to open a business bank account, and even provide you with mail scanning and a business address to keep your personal privacy intact. Visit northwest registered agent.com slash twist to get a 60% discount on your next LLC. All right, everybody, welcome back to this week in startups. Let's get to work. It's Monday.

That means it's a Madra Monday. My God, Sunny Sandeep Madra. People don't know this. He's one of my besties. When I go skiing, every season, Sunny and I do a bit skiing. We love to hang. Lifts. We talk on the lifts. We talk on the lifts. We have talks about life, parenting, business, everything. You blast down the mountain. I think your high speed is far faster. I think my highest ever was 40.2 miles per hour and you blew past that. I am slowing down. I don't want to tempt fate here, but the 54 mile per hour record is not a record I want to repeat. We got old man moans. We got to That's it. Under 50.

Do not do it again. So, yeah, but this year I'm going for it. This year, I've set a target of 50 days on the mountain. 5-0. 5-0. But, you know, I like to go for two hours a day. I don't, you and I like to go out. We do a little bit.

We have a little, maybe we have a cocktail. You know, it's that pre-discan today. We get that little, you know, spiked beverage. Not bad for the old bones. I like it. And I'm not a drinker, but there's something about that, you know, a little something at the end of the day, a little collo in the coffee. But let's get to work here.

Everybody loves Mondays. Mondays is our AI Mondays, and I want to just announce that for 2024, we're locking this in for the entire freaking year. 52 episodes next year of the AI Mondays with Sundee Maudra, and we're actually selling the ads specifically for Monday. So if you're an AI company, you can talk to our partnership team or email me, Jason at Kalakanis.com. There's only three ads every Monday. So if you get one of those in your hugging phase or your AWS or something, you wanted to reach people who are obsessed about AI. We're going to make this the greatest AI habit that anybody who's an AI developer, startup, or investor has to come every Monday to listen to this. And you've been getting

great feedback. Thank you. It's been really fun. And honestly, it helps me in my day job. And it basically is, you know, really insightful. I even went and did a little session with our other bestie Chimoth. Oh, yes. Yeah. You're pushing code with Chimoth. We did. He was standing next to you. You guys did pair programming. I love it. We did pair programming. Yeah. He was on the keyboard. board. I got to give them credit, so it was really good. It's a nice thing to do. I don't have the head for programming. It's like a little too much focus and I get distracted or whatever, my ADHD, but I do like the idea that you help me sort of pop it up and learn about it. We'll do one. We'll do a pairing session in 2024. For sure.

Actually, we should do it on the air. That would be a fun for people. All right. But you know what? Everybody loves to come here to watch us and listen to us do these AI demos. So if you're listening, no problem. We're going to explain what's happening on the screen why it's important. If you're watching, great. Go to YouTube.com slash this weekend. And then you can just hit the videos and you'll find a playlist there. My team made a playlist for the AI Mondays. We just look at the playlists.

You find that playlist. You can watch all the demos and you can see, you know, even if you missed a couple, going backwards is a great idea. So we're entering year two. Consider this year two of the AI revolution in terms of chat GPT 3.5 coming on. November 30th was the one year birthday. Yeah, exactly. Happy birthday, everybody.

Congratulations. And, you know, the big news, obviously, was Google has a bunch of new updates. So we'll talk about that today and see in the notes. But what do you got first? What's on first? What's our first demo? Well, I want to start with, it kind of got lost in the shuffle. And I think it's a really, really cool product.

And it's called Notebook L.M. I know about this. This is a product by Stephen Johnson, who is a writer, who I knew from New York. He's written a ton of books. Stephen Berlin Johnson is a writer. and there is a tool, I think it's called Scribner, that is a writing tool that like Tim Ferriss might have gotten into and he was into.

So when you're writing a book, you probably have 20 different sources. Let's say you're writing a book like Malcolm Gladwell does, like the tipping point. You might have 20 different sources of information, right? And you, they write these kind of meta books where they read 20 studies and then try to make an overarching thing. So what's very interesting about this is I was on this product, this weekend because I saw a wired story about it trending in the news and I played with this

weekend. So this is great. The Google product. The Google product. Notebook L.M is a book for, it's basically there's a, like I said, I think it's Scribner is the product. Yeah. And Stephen Berlin Johnson.

I remember him from New York, really cool cat. He's at a Netflix show. He's written a dozen books. He's like a real old school wired guy, you know, like a Gen X wired guy. Yeah. Yeah. And so this is a tool based on my understanding of it to help you collect. collect lots of thoughts and then make sense of them, perhaps in a book format or whatever it is.

Yeah. So look, I think you nailed it. And what Google's done and I've kind of, you know, started this demo here is, and I wasn't thinking about it in the context you talked about, which is interesting because you and I just bring different perspectives. But I think about this as a collaborative space. If you've got a document or something and what, you know, Google's done an excellent job of what we want to keep giving them credit. Here, I can add documents by uploading them, but I can also grab stuff from my Google Drive. Got it. Okay, so we're looking at an interface.

It's almost like a whiteboard in front of you or like, would you say a conboard where you're putting like little objects, boxes? Yeah, I think so. And on the left, you upload your source material, the source material is typically copied text, a PDF or something on your Google Drive. So that could be a Google sheet or a document. Amazing. Yes.

And you add those in and then once those are in, you can start a chat and you can say, tell me about this NDA. And I just uploaded the definitive. mutual NDA. And basically it will start to do its analysis, you know, using the LLM. And it'll explain, you know, I just asked it here, like, tell me about this NDA and explains, you know, what it's for. And I can ask, like, is this a mutual?

And yeah, yeah. How can I be sure, right? And then, you know, you can get me that question. You can also do the sharing. So you can add people to this note. book. Ah, so you could add me right now. Yeah.

Yeah. So I can add you to the notebook and then we can kind of collectively jam on this. And then I can add notes to this. And I think sort of the use case you're talking about is a little bit different, but I like it as well. And, you know, if you go back into the kind of the, and this is a good job with they done. They have some example notebooks here where you can kind of see, you know, how to use it. And so this is a little bit more like what you're talking about, like a bunch of notes and other things that are leveraging. Right.

So on the left here, let me describe it. In this one, they have an announcement, a letter from the CEO, pros and cons of, I guess, a product, some product specs. So if you were thinking about a project and a project had a collection of documents, everybody says, oh, I want to create a language model, but that's, you don't want to get a developer and hugging phase, replet, whatever. I may just want to take 20 documents that my team's worked on. So here are the 20 documents about project conduct.

door, whatever we're doing at the company. Could be an acquisition. Could be an investment. And I just very simply upload those 20 documents. And it could be the pitch deck from the founder. It could be the deal memo internal, deal memo, external, last deal memo for this. Let's say that we were doing an investment deal or an acquisition. Project Condor was, you know, Google acquiring YouTube.

You might put all this different information in here. And then people can ask questions. But when they ask the question and they get the answer, all of those become little notebooks. in this shared space. So I can see what you're asking the LM. I can ask the LM. And then we both get the benefit of those queries. Am I correct?

100%. What I really love is you can basically also then use kind of LLMs to ask about the documents like I was doing in my example, right? And I can kind of build on that. So I think that's really powerful. I think this is a really nice interface, done super clean, collaborative. I like integration into the Google ecosystem.

So I'm going to start out, and I'm going to give this one like an A minus. I think improvement is basically give it more sources here beyond sort of these. And it's really up there. There's not a lot more they can do. I want to start using this kind of in a work process of things. And I like it. I think it's a great tool. I like it.

I give it a B. I think it's pretty straightforward right now. It's a little confusing until you look at the examples of what you would use this for. But I think we gave you too. If you're writing a book, if you're doing a, an M&A acquisition. If you were doing a term paper, I could see you using this.

You know, if you were working on a project, you know, in graduate school or, you know, a term paper with a, you know, a study group. You might put all your notes of each week's notes in here. You might put every chapter of the book your reading. So let's say you were doing, I don't know, a cinema studies class. I might put in the script and we were doing Japanese film studies.

I might put all the Kurosawa scripts in here. A perfect example. I might put Kurosawa's biography in here. And that was the first thing. I was thinking was, I want to put some books in here. But there's a limit, I think, to the size of it. And then I was like, wait a second, how do I get a book in here? Because they're all locked down with digital rights management, DRM.

So I was like, oh, so how do I get something like a biography? What I wanted to do with this was take all my favorite business biographies, put it in there, and be able to query my top 10 business biographies or 50 business bars, but I can't do it. I have to get a podcast. I read a copy like a PDF of it. Now, I don't know ethically, you know, how you would do that. All right, listen, we work with startups and they are all over the map. Most of them, very early stage. Pre-seed, seed, you know, going on to their Series A. But some of our investments have gone on to raise those late-stage funding rounds. They've gotten acquired. Hey, and a number of them have

gone public. And there is one thing that unites them all. They need to have their business insurance tight if they want to succeed. This is obvious. A lot of founders ignore it. And they ignore it at their peril. If it's not tight, it's not right, and we need tight and right, and we send them to in-broker. And broker is business insurance built specifically for startups. Their single application helps startups get four quotes for four lines of coverage in just 15 minutes. In broker, they'll connect you with one of their expert brokers for unmatched service, and it goes beyond your policy. They'll make the process painless and transparent, especially when you compare them to the incumbents, which are slow. So try in-broker-ta.

Today with the code twisting at 10% off, they're already amazing prices, their startup package inbroker.com slash twist, E-M-B-R-O-K-E-R-com slash twist, and use the code twist for 10% off. We love in broker. Thank you for all the amazing support over the years, both on this program and the love and care you give to our startups. Let me ask you a morality question here. Let's say we had a friend, not me, not you, but we had a friend. The friend said, I have an ethical question for you.

I've bought these 50 business biographies. I also see, you know, if I do a search of Google drives on the web, that, you know, something like an autobiography or born standing up, Steve Martin's great one, or, you know, shoe dog, somebody has the shoe dog PDF in a Google drive, right? You can just search site colon drive.gulgoogle.com for any PDF. You'll find every copyrighted PDF if you want them. I don't suggest you do that. And don't steal my book. Go buy it. Support authors by everybody's book.

But let's say I bought all 50 books. and then I put my top 20 in Google Notebook from the web where I found them without DRM on them. Is that morally and ethically acceptable to you? Not saying you'd do it, but what would you tell your friend? Grayzone, okay? So here's how I think about it. It's like the time of music sharing before we had iTunes.

So, and look, I participated in, like, downloading music because it was the only way to get digital music. Right. But once iTunes came around, I've been using it probably since 2002. It was around very early. And I've been using it because, you know, I think artists should be compensated for their work. Of course. I'm happy to pay them because that's how the ecosystem works. And so I would say until that type of functionality is available in Apple, iBooks, or, uh,

Kindle or other places. Yeah, because, you know, I bought it. Yeah, I bought it and I want that functionality and the company's not making it available. So I feel like it's okay. I think the minute those platforms offer that functionality, then I should do it correctly. Do it correctly. That's kind of how I would think about it. You know, you should be able to.

So this is just for Stephen Berlin Johnson, I believe is his name. It's been 20 years since I seen him. Cool cat. Hey, Stephen, come on the show. would be really cool if you could press a button there and instead of upload from PDF, it said, sync with Google, Apple, Apple Books, sync with Amazon, Kindle, sync with Audible.

Like, why can't I just sync and put those things in there? That'd be an incredible feature. And so it should be allowed. And so until it's allowed, I would tell my friend, I'm okay with it. I'll allow it. the spirit of the law as you bought it, you should be able to do what you want with it.

This is why I may start going back to buying Blu-rays, just to not go on a total tangent here. But I feel like, you know, like when sometimes the internet's down, or I'd like to have physical copies of like the top 100 films that, you know, are rewatchable or whatever to have. Yeah.

And, you know, then trade them with friends and stuff like, because you can't trade your digital library. So you, if you buy a movie, I can't like just give it to you whatever. And then they have all the, these cool things like the directors, cuts and the audio tracks.

And they, I can't find those online. Like, if I want to listen to the director's commentary, I can't find the director's commentary, uh, like on Apple. It's not there.

So anyway, this is a really cool thing. So it's called Google Notebook LM. Google Notebook LM. Correct. Go search rate. It's pretty cool, I have to say.

Um, and they should actually know what they should do is Google has Google books for all the books that are and magazines and stuff like that. They have all that stuff there. So it'd be very cool if they connected it to their magazine and newspaper archive, and you could put the New York Times in there and pull New York Times archival stories or whatever. So if I have a New York Times subscription, why can't I add that to here? You know, and just add a couple of New York Times stories.

I'm constantly, that's another thing I'm constantly doing is trying to take a New York Times story or a financial time story and share it with a friend or discuss it with a friend. And it's just like so many roadblocks. It's so hard. I think the New York Times gives you like a little gift thing now where you can give a couple of articles. but I think it's like maybe five a month or something. I don't even know what it is. So I think that's a nice pressure cooker to take it off.

I do like also putting stuff in Speechify, which now has Speechify Studio. I saw that for the first time this week, where you can create your own voice. And then you can read to yourself news stories. So you can train your own language model with Speechify Studio. We'll show it next week. Okay.

But yeah, you can make like a J-Cal voice from this podcast and then have me read your news to you, I guess. Yeah. And train your own. I took a note for it. Okay, so Bard has been updated with Gemini Pro functionality, I believe. Whatever that means. So now they have Google Deep Mind, Google Bard, and Google Gemini.

Yes. Well, what's going on with branding Google? Tell me what I'm supposed to explain to the audience the difference between those three things, just so we can level it. Okay, well, deep mind and brain were two independent teams. site of Google that have been merged into a single team. That's now called DeepMind. DeepMind is the remaining brand. That's their AI team.

That's their AI team now. That's like saying, that's equivalent of saying open AI. It's equivalent of saying open AI. Yes, correct. And historically, those two teams had different approaches and they came from different worlds. DeepMind was an acquisition that Google did years ago that actually led to some of the fallout between Elon Musk and. and some of the Google founders.

Yep. And, you know, some of the key folks that were involved in the drama with Open AI this week, Ilya in particular was a deep mind person that was recruited out to come to Open AI. So it's very kind of some important details. Those two teams now have created a new model, a multimodal model. We've started talking about these. We've used them. And that model is called Gemini.

And Gemini comes in three different versions. And those three different versions basically are, different size parameters, which have different capabilities. And Pro, which is the middle-sized model, is now available for use inside of BART. And on an API level, it's going to become available on Wednesday. So if you're working with Google Cloud, you can use it as an API, like the same way you can build things on top of OpenAI and Chat GPT and GPTs. And then Ultra, which is the kind of the most advanced model, they said will be available.

in early 2024. Okay. Does that help? Yes. We have the DeepMind team. That's the group inside of Google. The consumer product is called Bard. BARD.

Now, BARD is a way for them to say, hey, this isn't Google Search. This is a different experimental product. It's going to hallucinate. Click here to understand you shouldn't like base your medical decisions or financial decisions based on BARD. Bards an experiment. Then BARD uses an experiment. it uses an underlying technology called Gemini.

Gemini. Gemini comes in three of flavors. Nano pro. Ultra. Ultra. So we now have one, two, three, four, five, six brands. This is a message to Sundar. We've talked about this, you and I.

Okay. Branding at Google is challenging. It's challenge. You call it Google Bard. A much better idea would have been to take the YouTube approach and just call it Bard. And get the domain name barred.com

and let people go to barred.com and have a completely unique verticalized experience without the Google branding. Yeah. And that nobody needs to know about Deep Mind, nobody needs to know about Gemini. It's just Bard.

And Bard is a new product from Google. Yeah. The end. Well, Microsoft nailed this too with the naming co-pilot, which, you know, we had a little private chat where we were sharing some names and ideas and we said, co-pilot is the way to go.

It's the simplest. And nobody can own that. So Google co-pilot would have been a better name for this. They just call it co-pilot. But you know what? Google's strength is in the underlying technology. Their weakness is in branding and often in U.X. And the problem I have with Bard right now is there's no app, right?

There's no Google Bard app. And the interface is just, I don't want to say it's terrible, but it's like they, I'd say the UI is like a five of 10. And I feel like chat GPTs is like an 8.5. Yeah. Well, no, can I say one thing to, not to Google's defense, but like very similar in branding and naming. We have open AI. We have chat GPT and we have GPT 3.5, GPT4, GPT4. Sure.

So they've kind of taken the same flavor. Consumers, but they didn't have a legacy brand of Google, right? So there is no legacy. And so they can start from a little bit of a cleaner position. People know chat GPT. They don't really know OpenAI. They really don't care about the version numbers. You just search for chat GPT.

This is why Google you should just search for Bard. And then all they should say is Bard let you do this. You can use Bard for this. You can use Bard for that. And then, yeah, when you go to the Dev Day, they can talk about all the underpinnings. but the consumer should just experience Bard and Bard should have a logo and it should be barred.com.

This is why Google Plus failed. I don't know why Google keeps doing the same mistake, which is putting things as like a sub-topic under the word Google. YouTube is your biggest success. It is an independent, strong brand with its own domain name. Google Plus got lost because it was

plus.gougal.com. It should just been called plus.com. And you went to plus.com and you had that experience. Yeah, fascinating. So here end is the branding lesson. You have to get this right. Google. Maybe a better name than Bard too, because no one knows what a bard is. I mean, yes, I agree.

I mean, they should have just bought Chat.com. I would have just bought Chat.com. And you know who bought it? Darmesh. I think bought it for a million dollars over at HubSpot. Yes. And the huge mistake. Wow.

What a, oh my God. I mean, for Google, that's like one millisecond of search. Can you imagine if they just had chat.com? And it'd be like, hey, yeah. If you want to chat, go there. Or AIChat.com. or AI.com. You know, it's just, there's so many different ways to do this, folks.

But the way Google does it is always a mistake. There's no branding person there. And Nest had a great branding team. And then you screwed it by having the Google team shoehorn it into and break it into Google Home. There should be a five alarm fire. Actually, you know what? This is a Sergei thing. Sergei, find somebody who you trust, who's your Johnny I.

I heard Sergey's in there right in the code, right? Yeah, Sergei needs a Johnny eye. He was the one that pushed for the release, apparently. Oh, really? Let's go. Yeah. Yeah. I like him. I like this out. I like Sergey after Doc. He's, I like, I like, billionaires writing code.

We have Jemoth and we have Sergey. I love it. Samurai Sergey. That means worth 100 billion who's writing code. I so much respect for Samurai Sergey. Oh, I like, Samurai, he's in the game. Oh, yeah. But this is a message. Somebody sent this message to Samurai Sergey.

I know him. Jake out. Samurai Sergei, please hire your Johnny I, find a branding person, and stop taking Google's innovations and wrapping them in bad U.S. This is a place to invest. Like literally just find 10 great designers, put them in a room, don't let anybody in the room. Well, I think Johnny Hive has a studio now.

He could just acquire it. Yes, go get. I mean, go give Johnny I have a billion dollars and have him come in for 10 hours a week. He'll do a better job than the Google team. No offense, more respect to the Google team. Yeah, they're doing a bunch of them. of work for like Silicon Valley companies. Like a little bit of a design studio.

Yeah, design studio. Just buy them. Okay. Easy breezy. Okay. Let's keep going here. Okay. Because I'm in Barr right now and I asked Bard.

Is Barr using Deep Mines Gemini and I said yes. It is using it. So what I want to do here, J.Kalz, I want to give the R's Technica guys a little bit of a shout out because they put together a really world class test bed. And I'm going to show some of these, but I just wanted to touch on this. But they did a great comparison between original. original barred and updated bard and chat GPT. And so they have a lot of examples in here.

And I've kind of picked on a couple of them for our demo here. So I wanted to give these guys a shut up. They did a great job. And so one of the ones that I found that was incredibly interesting from that example was, and I think you'll like this, which is I gave it some text to summarize. And it was a bunch of texts that I just copied from an article. What it was able to do. was it was able to, and this was sort of the article here, it was like an AI generated video

of Will Smith eating spaghetti. And basically what Bard was able to do was tell me, you know, what this, like, what this was about and was also able to link me to the original article. And I think that's really powerful because, you know, one of these things about attribution is going to be, I think, really important with respect to understanding the ground truth of things and also perhaps monetization in the future. So this is one of the enhanced powers of Bard in that when you give it text from the internet, it will summarize it and then use Google's vast knowledge base to kind of pull in the original source of that content. So two things you're pointing out here. One, attribution. This is something we've talked about. Who should get credit for this? You took a cut and

paste of the text from an article, did not give Bard the URL. It then found the URL. It then found the URL and gave proper credit to it. So all of this like, oh, the LLM, we could never undo it. We can never give attribution. That is absolute BS. AI has a very easy time doing attribution. And so this entire concept that we don't know how it came up with the answer, I call BS. You can obviously figure that out.

And then, so that's great for two reasons. One, it's fair. And it'll probably get you around a lot of issues of a lot of lawsuits. But number two, monetization. If this sends traffic to the original article, the original article gets to get some money, which is the very delicate balance Google made

with content creators. Hey, if you index it, we're sending you traffic. If we send you traffic, you can figure out how to monetize it on your site. You could even use our ad network, if you like, or somebody else's.

So it's worth giving us the snippet, just a little bit of the content to put in, to give people a preview. And so much so that they even did those one box where they kind of gave you a more specific section and that might give you the answer. And content providers were like,

you know what? I'll let you do at Google because net net, I get a third of my traffic from Google or half of my traffic from Google, whatever the website's doing. So I think those are two very good reasons and it places their strength.

They now know because they have that corpus of the web crawl. So very good points. Yeah. And so I thought like really good example and, you know, touches on a bunch of things that will,

could really put Google in a in a good position with content creators all over the internet and and you know attribution which is driven much of the internet business for for the last few years love it um okay so that was like sort of the first example i wanted to touch on with bard which i thought was really good if your landing page looks terrible i'm out we all know that you see an ugly website you skedaddle you leave you're done so you need to stop selling for okay or good and and start using Squarespace so you can be excellent and extraordinary. It's an out-of-the-box business solution to build beautiful websites, engage your audience, and sell anything you want.

You know Squarespace's amazing features, gorgeous templates that are always optimized for mobile, drag-and-drop web design with their fluid engine, advanced analytics, marketing analysis, sales data, and more. And with Squarespace, you can create an online store or start a blog at the click of a button, create a subscription business for members-only content, and so much more. And you can do this all simultaneously.

It's the simplest, most effective, and best-looking way to start a business online. So here's your call to action, squarespace.com slash twist for a free trial. And when you're ready to launch, go to squarespace.com slash twist for 10% off your first purchase of a website or domain. Another example, which I thought was really powerful, was its ability to go deeper in its responses. And so, you know, this is something, you know, talking about, like an era that, you know, you were probably closer to this stuff back in the 2000, but like, you know, PowerPC versus Intel processors. This was a big debate when Mac was switching over to, you know, the Intel these processors. And so it's able to just kind of dive into much more detail than before. So if you ask this

question in a previous bard, it just wasn't as powerful. And so the, you can basically, what you can kind of glean from this is they've done a good job of just increasing the amount of material that's used in its training. So it's just as a deeper knowledge set, it's not as surface level. A lot of times before Bard was a little bit surface level and that's called out there. Yeah. Fantastic. Yeah.

Yeah. So like, look, there's kind of a few of these improvements that they've done. These are kind of the more interesting ones. Other ones are a little bit subtle. Like it's better at jokes and things like that. I thought the thing that was most important was for me, multimodal looked very, very interesting. And there was a couple of video demos. Maybe we could do a reaction to.

a video here. So the video is going to play here and they're going to show some capabilities. What they were showing here in this video for the folks listening is someone on a sticky note pad drawing pictures and then increasingly adding detail to the pictures and Bard responding to it. And so this is really interesting because what it made it seem like was Bard was able to watch video and then in real time explain what it was seeing. What was subsequently shared by... casting. It was sports casting what was happening. So when you drew it, it was like, oh, it's a scribbly line. Then you drew a little more. And it was like, well, that looks like a bird. Oh, because it's got a duck, because it's got a bell, it's a duck. It's in the water because

you just drew some water. And oh, it's a blue duck, which really doesn't exist. So it was like really giving a commentary in real time of what's happening. Yes. And what, and this is very powerful. What, there's some criticism around in controversy. I'd say more criticism was that nothing is faked in this video from a sense of that, like, you know, the answers are coming from Bard. But the interface to Bard was not a real-time video. So what they were doing was they were taking pictures, they were giving them to Bard, and then Bard was replying with the answers on the right-hand side. But the interaction wasn't real-time video. Got it.

So they faked a bit of the demo for dramatic effect. What it really was was static images. So they were just showing, like, essentially, one post-it note at a time. And they would come back and say, this looks like this. And what is this now? And what is this now? But the truth is, they probably could build this in a weekend. So Sergei's probably coding it right now just to dunk on everybody and they'll release it in the next version. You know what? I was going to say that is like the best way to kind of, you know, demystify everything is actually just have Bard do it with video. And at the end of the day, you know, take screenshots of the video along the way and submit them to Bard and have it answer the question. You know, I had something similar to this happen when I was watching Han Solo on a long distance flight. I clicked the wrong version.

And the version I listened to was for people who were blind. Okay. And it described what was happening on the screen. And it's like, Hans Solo jumps into a speeder. This person jumps in with him. They speed off into an alleyway. They turn. Now they're being chased by three people.

And it was explaining, it was a narrator describing what was happening on the screen. And there was still the sound effects, the dialogue. And I was like, I didn't know that existed in the world. There is, and there must be a technical name for it, but there is a technical name for, like a commentary track that's describing what's happening on scene. Now you think about that, we could take any video. And if somebody was blind, you could play a music video for them. Obviously, they can listen to the song. But you could have a narrator describe what's happening on the scene.

And the person could say, give me more detail or less detail. or pause the video, tell me what's happening, and then restart it every 30 seconds or whatever at every milestone. So they could say, okay, wow, I'm going to pause here. Now Britney Spears comes onto the scene. She is a ninja, then she turns into this, and there's enough time to explain, you know, something like everything, everywhere, all at once,

like some really dense movie with a lot of activity. And just think about what that would do for somebody who was blind. You can do this in real time on any screen for any video, including live sports. So, you know, you have a sportscaster and we're sports casting here. Sports casting could become, could be done better by a computer, by an AI. So if Bard does this for a Warriors game, you can say, you know what I really like when George Michaels did this or somebody who's passed away who was a sportscaster. And you could have like some sportscaster describe what's happening in the game and clone their voice.

And man, that might be really interesting, huh? Yeah, it's probably, I think it starts off in like maybe different languages. versus that because like, you know, if you watch NFL and you get to watch a game that Tony Romo is sports casting, you get his knowledge of football and being a recent football player, which I don't think the AI can replicate us yet. I don't know about that. If you could have Jimmy the Greek, right, who is very famous, you know, Monday night commentator that people loved.

And nobody remembers Jimmy the Greek, but I do because I'm Greek and it was like a big deal to have somebody called Jimmy the Greek. But don't talk about me, Jason the Greek. I mean, you can. Anyway, J. J. the Greek, you could bring him back and have him doing Monday Night Football. You could actually probably get all of his insights from previously and get his real-time

reaction. Yeah, I mean, if you built, yeah, if you took a model and fine-tuned it with all his previous commentary combined with video with a multimodal. Yeah, that's actually an interesting idea. Mohamed Ali, Mohamed Ali doing, you know, MMA fighting or boxing. And, you know, what about having Steve Jobs? I mean, listen, we talk about the recently dead. I want to just be clear here to family members,

etc. Who might be watching or friends of his. I'm not making light here. But sincerely, like, are we eventually going to have people who've moved on and passed on, you know, like do a keynote or send a message?

And they did that with Tupac, right? He did a very hard decision. So I think this is a fan, it's a very hard decision. I think they did like a, like a what do you call a hologram thing, a holographic.

I mean, so if Tupac can come back, could, you know, Steve Jobs come back. It might be too soon. It might be too painful for people. But eventually could Steve Jobs. come back and release the next Apple product.

You know, it does so many ethical, moral, emotional things to think about there, but we seem to be on the cusp of that. You know, the other thing, I was so impressed with what they showed with Gemini. The one I found really interesting is multimodal where they took a math test. So maybe we could pull that video up,

and my team can do it too. But they were explaining. of in math and physics. All right. So this is super interesting. Here you see, they are taking a picture of the answers.

And explain the concepts that... So it scored the test, Sunday. Scored the test. Yeah. Right. With problems one and three here.

Let's take a look at three. That's incredible. Teachers don't have to do. Teachers no longer have to score tests. That's a big part of the job. That the formula was correct. But there was a mistake in calculating height. We can not.

Ask Gemini to explain and more. That's powerful. The height is 50 meters instead of six. Yeah. Not even teachers so much. Not even teachers so much parents. Like I don't know if you ever have to kind of help with homework. Like sometimes you really got to dig deep back.

And especially as they're getting into algebra and calculus. But like, let's say here, you know, they're, they've identified which question got wrong. Then they're looking at your solution as a student. And finding. And they're finding your mistake. And then explaining to you how to fix it. Yeah.

This is an example of the teacher doing their work when they're not in the classroom, which is scoring test. That's done. Then sometimes the teacher will circle the place where you made the mistake. That's done, the detail. And then you have to get tutoring and you have to do corrective work to not get it right the wrong time. Those are three components of work humans do that has now been replaced by an AI.

So let's just think about teachers in general. Now teachers don't need to score tests. The test would be scored in real time. So you take your test up to the front of the class, and they literally just put the task in front of a webcam. Yep. You have a webcam facing down. They just put it there, move, put it there, move, put it there, move. And it's sent to you what you score in real time, what you got wrong.

They could send you to another room to go over the answers, then have you come back and take three new versions of the three questions you got wrong and rescore your test. So think what that does for education. You know, there were those really cool teachers who would let you retake a test. Yes. Because they wanted you to do the best possible. You could be doing that in real time. You would learn more by fixing it and coming back than you're doing the real learning. But it's not scalable.

If they have to do that for 50, you know, 20, 30 students, they can't. This is a great aid for teachers. I think it's even a greater aid for students because they can sit there while they're doing their practice, their homework. And, you know, when they get something wrong, they can highlight, you know, why they got it wrong and explain to them how they should have done it differently. Starting a business used to be such a painful process. You needed to get a lawyer. There were tons of fees. It was a mess, but not anymore.

Just check out Northwest registered agent. They're going to help you form your company fast. Remember, speed matters. And then they're going to get you the docs you need to open a business bank account instantly. Then they're going to provide you with mail scanning and a business address. And they're going to do all that keeping your personal privacy intact. Northwest can form LLCs, corporations, and nonprofits, and here's why founders love Northwest. There's no hidden fees. There's no upselling. You can call them or cancel at any time.

And Northwest has the best of both world solution. It's simple and self-serve, but they can be hands-on if you need help with their amazing registered agent service. Northwest provides everything you need to start and maintain your business. And they're giving Twist listeners a 60% discount. For just $39 plus state fees, they'll form your LLC corporation. or nonprofit. So visit Northwest registeredagent.com slash twist today. Northwest registered agent.com slash twist. All right.

So we should give a score for Gemini here. Yeah. I have to say, I was blown away. I think that they have leapfrogged chat GPT 4.0 in some modest ways. And obviously on the tests, like there's a battery of tests. have you explain that in a second and pull up the list. So people understand how that works. They seem to have smashed

chat GPT4. Smash. Now, who knows if chat GPT4 is going to take this lying down or not? I suspect they won't. And maybe they'll drop like a 4.1 or a 4.5 to smash barred back and smash feedline back. But this to me means Google's in the game. You're taking it seriously. I mean, people are not going to believe this, but I'm giving this an A plus. Because... Okay, I'm going to pull up the benchmark.

Okay. Oh, A plus. I'm giving it at A-plus because I do think to surpass OpenAI on this many items, even though I think the interface is still janky and needs work and they don't even have an app and they don't even have a brand and it's just confusing. You know, if I were, I'm rating this release as an A-plus release. Let me say that. Okay. Wow.

Product-wise, I think Chad GPT is still a better product because their interface is an 8.5 of 10 and I give the Google interface like a four or five. I mean, the Google interface looks like it was not made by anybody. It looks like they literally went with some default template, you know, that was built into like WordPress a decade ago, right? What are the differences you see between the Bard interface and the chat GPT interface? Okay, so first and foremost, majority of people and majority of work, consumption and creation is now being done on smartphones.

So I always judge it on the app and work backwards. There is no app. It's huge. There's no app. There's no app. So that means you lose of 10 points,

you immediately lose five. So now you're out of five. Yeah. Now we're going to just grade you on your web and your web base app. When you pull up barred in a web app, when you pull it up on your desktop, it looks cluttered,

confusing. They just put it under the Google task bar, you know, with all the accrued amount. It doesn't have its own look and feel. So I'm minusing a point for that. Okay. And then I don't think the pagination and the formatting or as clean and crisp looking and the fonts and the topography, I'll take a minus one on that too.

So I give them a three of ten actually now that I'm doing my interface. Three of ten. You know what? I was kind of looking at it different, but the, sorry, go ahead. And then I just think chat, JBT, if you use that app, it is so polished that it's delightful to use. I've been using Dolly now to write my blog, to make my blog post headers. and my illustrations on my last two blog posts look like a million bucks.

If I had hired an illustrated to do that, it would have been $5,000 per illustration, I think. At a minimum $1,000. You couldn't do what I just did in Dolly. You can pull them up my team in a minute. But if you go to calicanus.substack.com, I'm doing a weekly startup post and I changed my list to CaliC Jason on startups or something.

And just the two header images are gorgeous. So let's talk about scores here. Let's talk about scores. And by the way, just hire somebody, To do the mobile interface, Samurai Sergei, I need you, Samurai Sergei to take out the sword. Whoever's on the design team, you just, you just, too, cut the team in half, fire them.

And then you find an external team that's great at making apps, and you ask them to make three versions. You give them $5 million. And then you come back to your team and say, why aren't we as good as this independent firm? And then you pick the people from that firm, and you give them an offer they can refuse. You go Samurai Sergaon.

Samurai Serga on. You coach him. I'm all right, Sergei. Let's do it. So, you know, a couple of things there, Jacob. We don't have access to ultra yet, right? So what we're playing with in Bard is pro. So pro, you can see here across the board, scored lower than GPT4. And so even though in your experience, you're feeling like it's equivalent. Explain to people what they're seeing on the left. Their mass, big, reach hard. What is this? I know what it is, but explain to everybody. Yeah. Yeah. So there's, I think there's eight critical benchmarks that people use to score different models. And what you can see here in this chart that I have pulled up is Gemini

Ultra and Gemini Pro versus GPT 435, Palm 2, Clod 2, Infliction, GROC 1 and Lama. And so these are kind of all the major LLMs that are out there right now. And it's, you know, changing rapidly. And what we see here is Gemini Ultra on almost all benchmarks scored higher than GPT4. Can you give us an example of but one of these benchmarks? and what it is actually testing? I mean, the simplest one, which is the one we were just looking at, is GSM8K, which is grade school math. Got it. And so that's the second one in this chart that I'm looking at.

You can see that GPT4 gets 92% of the questions, right? And Gemini Ultra gets 94.4. So this is, to be clear, they took grade school math, and they have a battery of questions, like a standardized test. So this is kind of like the SATs, or I forgot what we called them in New York, I think they were called your regions. But these are your testing. These are your tests. Everybody takes the same test and it's basically not, you can't cheat on it, essential. Well, so that's a very good question you asked.

You are not allowed to include the test in your training material. Because if you include the test in training material, you'll probably get 100%. So the researchers and people that are doing this have to, hold to that standard and say, you cannot include any of the questions from any of these different benchmarks inside the training material. And I like this, MMLU, multiple choice questions in 57 subjects, professional and academic. Correct. And this was done by some university or some researchers, correct? And then they put this out in the world. And if we look at that one, Gemini Ultra, 90% on this multiple choice test. Gemini Pro, 79%. Okay, great. Chat, CPT,

87%. So Gemini was, you know, whatever it is, a couple of... Gemini was a student and GPT4 was a B plus student. Got it. But it was interesting about this is if you go down, Grock 1 is at 73 already, and that's under six-month-old. Claude's been at it for a while. They're at 78%. So they're not applying themselves well enough.

And then Lama 2, which is Zuckerberg's one. That's an open source. Am I correct? Yes, correct. Is that 68%. And so we're seeing open source far behind the class. It'd be interesting to see if they catch out. But the thing, can I just comment on that?

The idea with Lama 2 is it's not meant to be a model you use out of the box. It's meant to be a model you take and you fine too. And we don't have it in this example. But I'd be willing to guess there's a Lama 2 fine tune that can score higher than both GP4 and Gemini Ultra on MMLU, which is the... Isn't the Falcon project on here? I guess it's not doing as well, or they didn't think of putting it on here? I wonder.

I don't know for sure, but my guess is it's not scoring as high as any of these other projects right now. Yeah. Okay. So, yeah, great. So this is amazing. What you find here is

chat, cheap ET4 has not run away, has not, I think what we've learned now is as an organization, Open AI has an Achilles seal, right? The Franken corporate structure, the Frankenstructure. And then also, there are two other criteria. There's the core power of the language model, which we're seeing here in these tests, and chat GPT and OpenAI can be beat.

That's what we've determined with this, correct? They're being able to. I think that was the biggest thing last week, which is folks got comfortable that, you know, what Chad GPT had wasn't, you know, magic or it wasn't so far of a technological lead that it couldn't be caught up. And the fact that someone else got there means other people can get there as well, which means we're heading towards commodity. Got. Which you have always said. That's always been your position. And what's very interesting about this is if you look at this chart, GPT4 was trouncing the other competitors. So when you compare Claude to chat GPT4, I'm sorry, GPT4, yeah, Claude or to Lama or to Palm,

it really did feel insurmountable. You know, there are scores here, 92 to 80, you know, 92 to 88, 52 to 34. I mean, these were big leads. Yes. But then now, so there's that piece. Then there's the piece of user interface. And then there's a piece of proprietary.

data. And so if we were to look at those, chat GPT for an open AI team, they have some, I don't know who the web designers are on there. Somebody tip us off of who's designing the app for open AI, but that's the person Samurai Sergei's got a poach. They got to get in there and steal the design team from Open AI

and just double their salaries and double their options and get them on the team or find somebody who can. So they've got a massive lead there, but that's an easy one to close. And then they've also, the question is, who's got the lead on data? So I need you to answer that question for me because the language models are at parity. The interfaces can be at parity, but the data that doesn't seem like it can be parity.

So explain. So I think this answer comes in two parts for me. And it's actually one of the parts come from your recent interview with David Luan from Adept. And you did a great job. I want to give you credit as one of my favorite twist episodes. And David shared just a little bit of kind of nuggets along the way as you were interviewing him. And one of the nuggets that he shared was in terms of the training material, it wasn't all public information. They went in licensed information.

And one of the things that might have happened, and this is like speculation, but I think we can see it sounds reasonable, which is prior to the explosion of, you know, last November 30th of Chad GBT and then early next year with GPD 4 or GPT4, if you were some organization and this open nonprofit shows up and said, hey, we want to. license some material for you. You'd probably like, yeah, sure, you want to pay some money. You're doing some kind of nonprofit research great. They may, and you know, they have a contract for it, and it's all good. I think, you know, that's an interesting thing in terms of what they were probably able to do in a past life. And what we're seeing now is, in over the last, even the weekend and last week, some people were complaining about GPT-4's capabilities are coming down. And that may be, some of that material may be having to come out of the training material that may be updating their models.

We'll talk about that in one of these demos that are coming up. And so I think between Open AI being at this for a long time and their approach and then Google with just their vast access to information, you know, being kind of the company that organized the world's information, they both had this interesting approach. What I will say, and I tweeted this yesterday, I'm saying right now, GPT-35, has been commoditized. So, yeah, and we're going to do a demo for coming up next. And it's using a specific model we'll talk about,

but I'm willing, I just want to say here, GPT35 level model has been commoditized. There's nothing special about them. Wow. Okay, this is big news, folks. And this is great news for the overall ecosystem, correct? Now startups are not beholden to open AI. When the APIs become more rich and the costs go down,

this is all going to be commoditized in the same way, storage has. This was a big question. If you paid a lot of money to invest in OpenAI, buy the secondary shares, and listen, I'm not trying to screw up anybody's deals here, but if you invested in a large language model at a very large dollar amount, I'm concerned about your investment. Now, you have downside protection because, you know, even if you invested in something at $10 billion, $10 billion, you have the preference stack. So don't cry for any of these folks. They knew that going in. But the fact is, a large language model on the internet is worth a billion dollars if it's one of the top five.

It's not worth $100 billion. Am I correct for that statement of you think you're great? I think you're right, right? But literally last year, you would value a large language model at $50 billion, $25 billion, $100 billion, because there were only two or three really good ones. Yes. Now there's going to be 50 good ones, probably. What do you think?

And there'll be 50 worth a billion and none of them will. You wouldn't even be able to tell the difference between them for, Nine out of ten searches, am I correct? I don't think we're going to be able to tell the difference. And that's maybe it's a good jump off point, JCal, to the next thing. Okay, here we go, everybody. More demos coming. Do we need to give a score to anything?

You didn't give your score for this release by Google. So in totality, I gave it an A-plus because Samurai, Sergey, just proved. He just commoditized the market. Yeah. Deep-minded commoditize the market. I'm only going to give it a B-plus until we get API access to you know, pro, which is coming on Wednesday. And then when we, and I want to also reserve until we get to actually play with Ultra

because we, we haven't had a chance. And so, but I do think they have an, it's, this is like the example you gave of, you get to come back and get your score updated. So either they release like an updated video of Gemini Ultra kind of doing it all, like as they showed in that video or they just release it to us developers and a limited sense. And you can move your way to an A plus. All right. And that David Luan episode is 1855 if you're looking for it, folks.

It'll be linked to the show notes. Okay. Okay. So next, we're going to, actually, I'm going to just jump around in our notes to this because it's a natural flow and then I'll come back to the next thing I want to show you. So next, this weekend, the team at Mistral released another new model, and it's called an 8x7B. And you can play with this model at Poe. So one of the things that, you know, if you're a subscriber to,

Po, you're able to pick which model you want to chat with, and they have all kinds of models there. They have API access to GPT4, they have a clod, and so you don't need to have an API key. They abstract that, they use their API keys, you pay them, and then

they figure it out, yeah. I'm going to go get a subscription to Poe because I really, and Poe is by Cora, which is Adam DeAngelo, who would be great to have it. We should have a guess. We should do like a Tuesday with a guest,

you and Adam. So let's do Adam on a Tuesday and we'll have a back-to-back. But it would be great to just go through Adam and his thinking, but he's also the one on the board of Open AI and at the center of all that nexus of

yeah, chaos the last couple weeks. Not that he caused the chaos, but he was there in the room where it happened. It was happening and he was there. He was there. Yes, he was at chaos. So why this model is really interesting,

and this is when I saw this model and I got to play with it, it basically led me to the conclusion, which I said 3.5 has been totally commoditized now. So this is what's called a mixture of experts model. So this is eight, seven billion models working together. And why this is really interesting is in a few ways. One, when the models are smaller, the resources that you need to run them become much easier to obtain. And the resources required to train them become much easier to.

obtain. And so generally a rule of thumb that you can use is if you look at a model size and if it says $7 billion, you need 70 billion tokens to roughly train it, 10x the amount of parameters to train it. And so when you get these really large models that are $70 billion, $270 billion, you need trillions and trillions of tokens, which then leads to, you need very, very specific compute clusters that only a few people in the world have. have. When you have the smaller models, you can use much more kind of readily available compute to train these things. And what Mistral released with this model is a mixture of experts, and this is generally considered how open AI is structured. It's, I think, you know, they may

have changed it recently, but before people were saying it's like 620 billion models working together. So they did 8, 7 billion models. And what they were able to show is 8, 7 billion models can outperform Lama 270B. They can outperform 170 billion parameter model. And that, you know, as we saw on the previous one, the Lama

270B was lower in score, but it's designed to be fine-tuned. And if you fine-tune it, you could probably get it up to, you know, matching all the other better models. And now that you can take a smaller model, which is much more realistic for smaller

companies to run and fine-tune, you know, the 3.5 era is commoditized. And so you can play with it here and I gave it, they're going to say, tell me about the founding of Atari, and it was able to tell me who founded it and what went on. And, you know, you can run it through the benchmark. But the key thing here talking about, you know, Po and this Mistral model is we've, you know,

the open source community combined with Mistral has found a way to take on these large models. And just one year ago, GPT-35, chat GPT-3-5, chat GPT was announced, now we have an open source model that anyone can take and fine-tune and run without having ridiculous infrastructure. And you have the equivalent of what chat GPT was a year ago. So I'll pause there and let you react to that and ask me questions. Yeah, I mean, here we are. I just subscribed to Po. I paid $199 for the year. Yes, and while we were talking as I was like, well, that's the... That's like, you said, Harpooh.

Well, I mean, as the cheapest I'm going to spend and I want to be up to date on all the models, So the fact that they have abstracted it, so let me just open up the conversation to Poe, generally speaking. Yeah. I've been looking for this because like this show kind of helps you level up. I feel like letting people see everything is great. And then they also have bots and they let you create bots and stuff like that. Yeah.

So that's kind of what you're... That was a little bit of the tension that was there with GPTs got launched, right? Because they had just launched these bots and then GPs got launched like literally a few weeks later. So that was part of the speculated tension. Maybe we can ask Adam about it directly. Well, I mean, also, No conflict, no interest, right? We always say that here in Silicon Valley.

So listen, you're on the board of this thing. Who knows who's an investor? Maybe, you know, Sam has shares in Cora. Maybe Cora's got a licensing deal with Open AI and they're sending a billion dollars back and forth. Who knows? You know, I think Cora's got this incredible data set. I think what Cora should do is come up with a revenue model

for people like me who are super answers in the system. Like I did a thing with Cora twice where they had me do a live Q&A, like an AMA, and I just took 25 questions. And it really builds their corpus up with an expert. Yes. So, you know, I did, and it grew my following on core. Now, I never figured out how Cora would work for me, how it would help me. I'm sure people see my answer there.

Maybe I get a little bit of play. But I know that people who are lawyers, accountants, real estate people, they love answering questions there and pulling bullet points on it because it makes them famous, right? So if you're not famous, you can become a famous real estate broker in New York city by sharing your knowledge, right? And being number one there and then people call you. They know your name. They see it. You can send your profile to people. But what if, you know, every time my answers were used in an LM, I got a licensing fee. So as the number one angel

investor, early stage investor, and you as the number one, you know, AI developer and Susie as the number own, you know, tech lawyer and John as the number one real estate agent, every time they got their licensing Fiat Poe, they distributed to us as being, you know, X number of answers in the corpus, X percent of real estate or whatever, $10,000 a year. The Spotify for everyone else. The Spotify for everyone else. Spotify for everybody else. Spotify for knowledge.

Where is that model? And I think that Adam could do that where he just said folks, hey, we know, we've, here's the verified list. So if you move to verified on Cora, and we know your credentials, we have your driver's license or passport, we know it's really a human, you're not stealing the answers, and you've done over 100 answers, and you've been on the system more than two years, you get into the licensing pool. And if you want to be in the licensing pool, you've got to be an expert and get this much. Like, just take that and then for all time. And then they could be telling me, hey, if you do 100 questions, we'll pay you in advance. These are 100 questions we want to have in the corpus.

If you answer these hundred, we'll give you $100 each. We'll give you $10,000. And then that'll be against your earnings in the future. There's a really interesting model there that could work. Yeah, I mean, I love it. I think, you know, what you talked about is sort of what Sam hinted at with GPTs. Yes. And, you know, we demoed them as soon as they came out.

And I think, you know, the enhanced ability of Quora slash Poe is that they're not limited to offer it with just the, Open AI model because you can mix and match and you can have models that are lower cost to run that maybe the monetization is better on because they don't have when you know it's mistral they don't have to pay as much money to Open AI for using chat GPT especially GPT 4 which is quite expensive so okay so to recap what we just saw yeah the high order point here is there is a new model that has commoditized the 3.5 level correct and it's essentially now free to use these yes you can run it on pretty standard compute. Define pretty standard compute.

A MacBook Air, M2? I'm saying for a service, right? So for a develop, like, you know, if you want to run like a service, you're going to run it in the cloud, you don't need to go and procure a farm of H-100s to run it. God. So you can run this with just standard CPUs.

You should use, you can, but you should use GPUs, but you don't have to buy the latest and greatest from Nvidia that has... So when the models get bigger, J-Cal, you definitely need to use the bigger chips because they have a huge memory footprint that's required to run them. And we're going to show that in the next demo

when you're trying to run these things. So when they're really, really big, you have to have a memory interconnect amongst these chips, and they have to run in like racks and even more even beyond racks, like multiple racks. Because you're giving it big chunks of information at a time to process,

and it just can't be done without... Yeah. It's not going to read that from a hard drive. That'd be far too slow. It needs to have really high-speed memory. What do they call the memory now? There's a specific name for the memory chips inside of these GPUs. Just like V-Ram, like, you know, kind of, yeah, for that they use for inside GPUs.

But the idea is that when you get these large frontier models, like, you know, 100 billion, 200 billion, 300 billion parameters in size, they are hundreds of gigabytes, even approaching, you know, like a terabyte. Well, a terabyte of memory, as you remember, even from your days working in the computer lab, a single computer can only support so much memory. So what you have to do is have multiple machines, stitch them together, have them working amongst each other,

and very close so you can support it. When the models become smaller, they can all fit on a single machine. That really lowers the type of, you know, hardware that you have to go after to running these things. Got it. Awesome. Okay, let's continue. Okay. And so, yeah, last thing, just on Poe, these are the different models that you can access if you're using Po today.

You can use GPT4, Playground, Cloud Instant, Dolly, you know, Mistrol. And I was just showing you there, Cloud 2 with 100K. So all the different models. So it's really, really great to leverage if you want to try the different models and get on the subscription tier, 200 bucks a year. Great. Awesome. So I guess we got to grade that. Yes. It feels like a B plus.

I don't know. What do you feel? Well, I'm going to grade two different things. I'm going to grade Mistral 7B. I'm going to grade Mistral 7B as an A plus. Okay. Because it will, like I said, commoditize the 3.5 level of... Okay, I'm going to give it an A for that.

That's how we're basing it. And then Po, I'm going to give Po. You know, it kind of suffers a little bit of UI wonkiness in the same way. It's like very engineered kind of like tech heavy. So I think it's a great tool. It's quite powerful. It has lots of different ways for you to try all the different models, all in the same place.

And so I think from a UI perspective, I kind of knock its grade down to probably a B, but I think the fact that they just make it all available, all the different models, they have all the partnerships done. Some they run locally. Some they run with other, you know,

startups in the ecosystem. So I think it's like a B right now, but if they continue to improve it, yeah. I give, I give, I give, I give Poh a B.

But I will give them credit that they do have like a desktop app now, and they have mobile apps. Yeah. So I think that they, I think, Adam understands consumers pretty well. Now, you know, when you do these desktop apps. And they have a good monetization model, right,

that, you know, people can participate in. Yeah. Explain the model of bots and what they're doing. Yeah. So the different bots can exist like either as core language models, or you can take one and you can find two and eight. So you can create the J-Cal bot that is trained on your books, maybe all the episodes,

and it's been fine-tuned, and you can put it up there, and then people can use it, and basically you can monetize it that way. You'll like this one. Let me share this. I'm going to share my screen. I may have to go B-minus now.

I may have to go to a C because of this. Oh, okay. That's a joke. But I just asked their photo underscore create-e, which is one of their bots operated by, Angel A. Chen. It's got 12,000 followers. Angela Chen, maybe Angel, Angela Chen. Yeah. 1947. I said, hey, she asked, hey, can I create images with your prompt?

I said, make me a handsome version of Jason Galickana's. Yeah. And it came out with this chubby version of me as like, I don't know, like some dude in San Diego who's, you know, going to get sushi. Okay, this is good because we have a kind of an updated version of this. And let me just share the monetization model here. Oh, I can't. You can have to kill the shares. I'm going to ask him to make him thinner. Let's see.

Okay. Let's see what it does. Generating first. I mean, it is, I will say, it's pretty fast. Oh, that is not good. That's not good. I said make him dinner and it made a pig horse and a suit. Oh, because I may be because I said make he dinner.

I don't know. Yeah. All right. So this one is clearly hallucinating. But I'm still going with Poe as a B. And I'm going to do it. Let's do a deep dive into Poe when we have Adam here. So let's do it.

Awesome. On the call. All right. And just on monetization, I want to close it up. It's up to $20 per user that subscribes. And so if you put your model in here, you can make $20 a user. You have a thousand subscribers. You're going to get $20,000 a month.

Okay. Here we go. Yeah. All right. Like it. Okay. This one is really interesting. It's been going around the internet.

I'm going to see your reaction. We're going to get like a JCal reacts moment here. Here we go. Live and direct. Live and direct. Live and direct. Live and direct. live and direct. Okay. So one of the things that I've been seeing

kind of in the last couple of weeks, like really kind of uptick in is folks creating these influencers. Yes. This is a big deal. Okay. So this, you know, it's a good lead up here because what I did was I created a co-lab workspace. This is a Google product. It's sort of like a different version of RepLid, and it's all within the Google. Exactly. It's an IDE that has access to compute. And what I did, I was able to basically get a model.

In this case, the model that I am using is called Juggernaut Excel. So this is what would be called like a fine tune of stable diffusion, right? SDXL, stable diffusion Excel. So it's a fine tune. And the fine tune's been really kind of focused on making people. And so I, oops, let me go back here. And what I did was I used. this framework called Focus, and basically Focus lets you run these models locally.

You can see here in this little interface, when you kick this off, a model gets pulled in. That model is about six gigs in size. You see that right here, 6.62 gigabytes. And you can see its utilization, how much GPU RAM it's using, system RAM it's using, and whatnot on the right hand side here. And so now what I have is I have my own model running that I downloaded from Hubbing Face. On my own, well, yeah, hardware that's in the cloud, right? In this case, it's in Google's cloud. Nothing too spectacular.

Like, it's, I don't have to pick like H-100 or anything like that. And so let's say a blonde woman standing in front of the Eiffel Tower. Right, this can get dangerous. I can see he's getting dangerous quick. Yeah, we might have to edit this out if it goes wrong on us, but it should be fine. Let's be careful, folks. This could go. I mean, no, this is true, though.

When you do Dolly, it has a lot of rules. You can't do a regular person. You can't do nudity. You can't get too risque. But I do see people releasing AI, you know, pinups, let's just say, or even more graphic. And people are doing this. I see a lot of women doing it online. Like, as a joke, like professional women are like trying to make.

LinkedIn photos and then it comes back with something that looks like they're a runway model and they're like, not exactly appropriate. So here is a stunningly beautiful, almost impossibly perfect. Well, it's still working. It's still working. But even at halfway mark, it's like Scorley Johansson in a very revealing dress, nightgown. I'm trying to figure out where it's going with this. The left hand side's a little bit more PG, but like I think they're both okay. That's just a really nice dress. I mean, listen, these are not, you wouldn't use these. These would be fine if you were going out for a night on the town. These, you know, going to the club, but you wouldn't, this is not an outfit you would wear

in a professional setting. Yeah. But if you're an influencer, you might wear something like this soon. But an influencer would certainly wear something like this. This would be the average Instagram influencer's photo. And so here you go. So like, you know, we just did this. I mean, that's super impressive.

And you, I guess my question is to you on County Valley, if you were scrolling. Yes. Not stopping and pinching and zooming and like doing like a detailed thing. But if you were swiping through Instagram, would you swipe past us and think it was a person with a filter. Yeah. And I think you would. I think you would see this and think it's a real person.

Now, we have one more demo after this I want to do, but let me show what you can do here. So I downloaded that picture and let's just change it. Let's say in front of Eiffel Tower. Let's pick something else. Let's say, how about, say, I don't know, beach or New York City. Yeah, doing a yoga pose on the beach, right? Okay, great. Well, this is, yeah, now here we go.

This could get a little risky, folks. So try to keep it. I hope it keeps it PG. And so basically what I've done now is I've said, I want to say, you know, the same thing. I'm going to give it an image input. So the model is going to use this person. And I've obviously changed it doing a yoga pose, you know, on the beach wearing yoga clothing. So like, hopefully all that comes true for us.

Amazing. So the first one you did of an influencer, it picked a Caucasian, blonde, whatever. And that was its own ideas. We didn't tell it that. Correct. And now you're using that as the foundational. Because that's the influencer that I want to have, right? And so...

Here it goes. Here's somebody in a yoga pose, which is on a beach and, yeah, it looks reasonable and not too risk. It should be a version of her, the first one that we created. And I want you to think about this, Jason, in terms of like, you know, the world of Instagram and everything that we've seen in our world and like the simplicity by which like these things can be created now. And what is your reaction to this?

And what does this mean to content? So a person's been on content for very long. So here's what I would say. If you are a yoga brand, let's say, in your first year. And so you're, you know, Lulu Lemon and it's a startup to, you know, 20 years ago to do these images for your catalog would be a $20,000 photo shoot. Plus the flights. Plus.

Well, yeah. No, I'm saying 20,000. all in. You would book a model. You would need a hairstyle, hair and makeup. You need a photographer, a photographer's assistant. I'm just giving you like a baseline. And let's say you probably need three models because you were doing a catalog or whatever. And you need a wardrobe, hair and makeup, pull permits, security, photographer, photographer's assistant. And, you know, photographer's assistant would handle lighting and then you have to develop the film,

etc. So it'd be a $20,000 process and oh, you'd also need a location scout to find the place and you'd have to pull a permit to take pictures there if you were going to have lighting or typically if you're putting a tripod or lighting down, that's when you need to get a permit. If you're just filming and, you know, it's for your personally, you're probably not going to have an issue. So this is why people use like a Canon 5D or like, you know, like those kind of handheld ones because if anybody comes and there's those things can do low light and everything, you're just like, Yeah, no, no, it's just for personal use. We're just, yeah, I'm influencer.

It's not a whole, yeah. Now when influencers came out, this dropped down to, let's say, $500 to $1,000 for a suit. People would have their own, they would get their own wardrobe. So let's just call it $1,000 for the day. Instead of paying $1,000 a day for the model, the model would do it for free for credit to put it on their Instagram or maybe give them $100 per diem. So it went down 95%. Now you've gone down 100%.

You now are at freight. And so the question is, if you're a Lulu Lemon, in your, or the competitor to that, you can make an unlimited brand on Shopify, up and becoming brand on Shopify. You literally don't need models, and you could be doing your pictures based on an editorial calendar,

and if F1 was happening in Vegas, you could make these people doing yoga in front of one of those F1 cars, or you could have the same models, you know, at Fashion Week in Milan or New York. You could have them at Art Basel. So now you're traveling around the world with a cohort of models.

you own 100% of their identity and their rights, and it takes one person an hour a day to do this. It is insane. This is saving millions of dollars in marketing collateral. So this is called marketing collateral in the business. And this is the end of needing to hire models. And so if you did hire a model, you'd be hiring them because they were unique to the brand.

So Suki Waterhouse, I met her one time at a party in L.A., quite charming. and she's like an influencer, musician, et cetera. You're getting a brand endorsement from that person because she's hip and she's got this incredible fan base, etc. Yeah. Kim Kardashian fan base, traffic. But for the next tier, you can't afford to pay them 500K or 50K, whatever they get. You know, I think it's probably 50K to 500K per engagement. You're not, people aren't going to pay that.

And so. Or what if Kim made a model that was of her and made it available to download for use? Like cameo. Yeah. I think that if it was for non-commercial use, it would be fine. And then the second you go into commercial, you got to pay a little bit more. So I think Cameo came up with this. I saw somebody was using Russ, the guy who plays Russ Hammerman or whatever that guy's name is the Tracecoma guy with the

Yeah, yeah, yeah, yeah. That guy was doing a bunch of stuff for Acquired.com. Okay. And he was doing it for Andrew Gazdecki. And he would make it like really funny, Andrew Gazdeke, and he would do it in the Russ Hammerman voice. And it was hilarious. And he would, I think like, he could.

use that to promote his company and instead of paying $50 a cameo, you pay $500 or something. So I was talking to Steve from Cameo and like they just basically came up with a cheer for sort of commercial work and I think you can use it for one year. So just, you know, a set of rights. Like Kim Kardashian could let you do it on your site and it's non-commercial and it's $100 an image. Or if you want to use a commercial, it's $1,000 in image and somebody's to approve it. So yeah, yeah, there could be something like that. Kim would do it, but other people would.

Other would, yeah. Like Grimes is doing. Grimes is doing that for, hey, yeah. can make. Yeah, and that actually what you would do is a royalty share. Now, royalties don't exist for e-commerce, but you could say, if you want to use this on Shopify, Shopify will make an API, anything that sells. Someone to do a Shopify app for this.

Yeah, so do it as a Shopify apps. Shout out Toby, I'm a fan of the pod. We have Toby set that up, and then let's say you put yourself, let's say they put Kim Kardashian or Shuki Waterhouse, and they say, I approve me being in these 50 outfits. And that's 50 of 250 skews. Anybody who sees her in the outfit gets tracked and she gets 10% of this out. Yeah, like a super affiliate. Okay. Okay.

So letter grade, one more demo after this. Man, I got to give it an A minus. I think it needs a little more refinement, but it's almost perfect. It's very technical because I had to load a model in CoLab and all that. But this needs to be an app. If this was an app format, I give it an A plus. But I just actually, if it was an app format, I'd give it a plus. A.

In this format, I give it either a B plus or an A minus. I'll give it a B plus. I don't want to go too crazy with my later. I give it a B plus because I don't, it doesn't cross the uncanny valley enough for me. And it doesn't do video yet. These need to do short videos. So if I'm going to reserve a grade point for app and a grade, well, some grade escalation for those things.

I'm going to give it an A minus because the next product, we'll wrap up quickly here because a bit over. Bit over time. This is, so this is a new tool called. I like that picture of me. I look okay. And that's like medium fat, j-cal. Magnific.

And what it can do is you drop an imager and it's an upscaler and it will and see what it's doing here. It's kind of, Oh, it gave me 5 o'clock shadow. It gave you 5 o'clock shadow and like it improved your eyebrows and hair. Yeah. Wait, wait, which version is me? This is the one with that.

That's me. Okay. That's you. And this is like a little bit older. Oh, it meant older. Older, because that's what I was trying to go for because these are younger pictures, right? Yeah, it gave me a little turkey neck. So this is this photo of yourself.

Got it. And then you made the older. Made you older. It gave you a bit of a thing. And like it cleaned up the photo. Like it made the hair look a little bit more realistic. It fixed your suit. Yeah, that's me at 65.

Yeah, maybe. Yeah. Fix your search. Yeah. Okay. So then I took another one of you and I said, this is a quite young photo of you. I said, yeah, this is me at 30. I'm about 30 years old there.

Yeah. Make an older J-Cal. Wow. I look so distinguished when I'm thin. Right? Look at that. I look like Pierce Brosnan, a little Pierce Brosnan going on. Actually, it's from 2009.

It's a 15-year-old photo. Yeah. And did you tell it how many years you want to add? I just said older. I think the prompt was older. Yeah. And, you know, and so. Yeah, I mean, it gives me my, it actually doesn't, it looks like the waypoint between now and 65.

So I'm 53 now. That was 14 years ago. So that was 40 or something. Yeah. Maybe I was 39 in that photo. So 39 to now. Yeah, it's probably right in between. Not bad.

Okay. And then for the influence. influencers we were making, and I'll just do this quickly here. You know, we were saying there's a little bit of uncanny value, but if you put it over them, you see here, it just gives them a little bit more realism, wrinkles and... Got it. So this is the opposite of a touch-up. This is a touchdown.

Touch-down to make them upscaled and more realistic. Wow. Oh, so if you put these two things together, it's A plus. Yes. This looks like a real yoga mom. Yes. This looks like a 40-year-old yoga mom from, you know, Pacific-Bowel. The United, Malibu, whatever.

Watch the bicep. Watch the bicep there. Yeah, it added, instead of it all being perfectly smooth. Yes. You know, I like about this, I have three daughters. I like not putting a standard out there that's absurd. And that's what this AI is doing is putting out an absurd standard. None of us can sort of ever achieve.

And I like this, you know, instead, I like the touchdown. I think touchdown for the win. Stop touching everybody up. Let's go touch down. Yes. Let's own it. I like being, I like moving into my Harrison Ford years. I'm hoping my Clint Eastwood years. I hope I look grizzled and, you know.

Chiseled, grizzled and chiseled. Griseled and chiseled is what I'm going for. And I've been doing my two, I'm starting my two mile run program, trying to see if I can lower my two mile time. Okay. And then just, you know, I got my, I got my 20 pound waist over there on the floor and I just. That's it. Yeah, I just hit him once in a while.

All right, this has been amazing. Awesome. So for the touchdown product, I'm going to give that a B plus. and giving the other thing a B plus you put them together, I give them an A. Yes. But I'm going B plus on both. I think they're both fascinating.

I think it's good. I think they've done a nice job like making the photos look more realistic, which is not the touch-up and what we see in the Instagram filters. It's like it's an upscale, but making them more realistic. And I think it really changes a game for e-commerce entrepreneurs or folks that want to, you know, sell products and want to have, you know, marketing material around it.

What was the term you used? You called it marketing. Marketing collateral. Marketing collateral. So, you know, marketing material. collateral is, let's say you had that picture from the beach yoga session and you had the three models. You know, okay, we can use this when we go to a trade show and we put that as the backdrop. Oh, on our website, it's the marquee image.

Just really quickly, if we could pull up my substacks, I'll just show you what I did here. So go to the top there. So this is Jason Kalakhanes on startups. And I said this is the greatest moment in time to start a company, the greatest moment in time to start a company question mark right now. And I explain why. And I gave it a prompt that I wanted something that would be the roaring 20s. So combine 1920s with a futuristic world and we'll call it the roaring AI 20s. And I said, add robots, had flying cars, yada, yada, and then put the roaring 20s at top. And it did s the roaring 20s.

Like, it still has text. And then I said, put somewhere on this picture, start a company. So you see at that store, it says start a company. Now, what's interesting is, this took me five minutes because at the bottom of my story, I just uploaded a gallery of all the other versions because I thought it would be funny. So here's the Warring 20s. This is a cyberpunk version, start a company now.

It's spelling things wrong. This one is weird. Here's another one that is a bit more like Art Deco, next one. And I did all this with Dolly. This one was kind of ugly. This one was Blade Runner-ish with a bunch of drones and neon. This one was also cyberpunk neon. on.

So anyway, these illustrations, and then if you go back to my previous substack, I did the productivity one. And so now, in this one, I did a prompt, you know, I wanted a bulldog in it,

a cup of coffee and, and, you know, a table. And so now I have like a concept here. I make these marquee images for my blog post, and I think it makes them very appealing

for people. And I don't need to rely on somebody to make, I used to go find a gift that I like and throw a gift in there and that or make a gift from a movie. So I'd go find. a movie I like, I download the video clip. The greatest time ever to be an entrepreneur.

All the tools that just right at your fingertips. Sit there and just do it all. Which is a good at times as any for me to just give a final plug here. Definitive.io, if you need work done by a genius, that genius would be my friend, Cindy Mondra. And our team. And our whole team. The whole team.

Okay, whatever. The team works for you. If you go to If Only I Started.com, you will see a video of me talking about, you know, starting a company. And my team just got the domain name if only I started.com. And so I just made a 15-minute video. And this is about Founder University.

Go watch the video. Go to Founder. University. We're starting this. We give people 25K checks, two or three-person teams. We accept 200 teams. We give the $25K check to about 30 of them,

40 of them per cohort. So I put about a half million dollars to work, a million dollars to work every quarter. So about, my budget for this is about $2 million a year. $2 million a year. Yeah. $2 million a year is what I'd like to do at scale. That is obviously 80 companies, $25K each.

So I'd like to, or 100 companies, $25K each would be $2.5. I would like to actually get this to 200 companies, $2.5 million a year. No, yeah, 200 companies, $5 million a year. So that's where I'm working towards. I want to put $5 million a year to work in the fund. You're an LP. I'm an LP. Yes.

In 25K checks, we take $2. 5% of the company. It's the first check-in to the company. Helps you get organized. Help you get your lawyer. And so just let us pick up the first 25K of your expenses. And then you got what your excuse.

And 2.5% is like nothing, you know. Nothing. And if the company shuts down and it doesn't work out, no harm, no foul. We lose 25K. You lose six months of your time. If it does work out, all we ask is that we get to invest in the next two rounds,

another 100, 250K. And so we can get that 2.5% of 10 or 7 or 6, whatever works out. Whatever the number is in. Whatever the number. as the number is. I think it's great. As an entrepreneur,

love it. I call it a pre-accelerator. And so I'd like you to do a talk at it. So, hey, producers can, I want Sonny to do an AI talk and an AI segment with the Founder University companies. I've now done,

I just want to let people know 60. This year I did 60-25K checks, I think. So I'm deadly serious about this. I am investing in 100 new companies a year and doing follow-ons with 50 companies. And I think we're going to catch up to Y Combinator. I don't, you know, listen, launch.

1 4, launch. launch. launch.co slash memo. There's a little bit of time for QPs to get in, but I think the accredited investor slots are filled. All right. Everybody, we'll see you next time. Follow at Sundee. Follow at Sundeep. Follow at TWA startups. We love you guys. Give us feedback. Sunny, this has been a great joy. Now we play poker every week. We do our podcast episode every week. And then we're going to start chasing that pow-pow. Let's go. Let's do it. We need that weather to comment. Everybody, see you next time.