SPEAKER_00: Microsoft had a big announcement that they've gutted a company and

SPEAKER_02: stolen all their employees inflection.

SPEAKER_00: Congratulations, I guess, to the team, but maybe not the investors.

SPEAKER_05: Uh, congratulations, uh, Satya Nadella to running the table on everybody.

SPEAKER_06: I mean, my Lord, this guy is like, this is getting super cutthroat. Yeah. Be careful out there.

SPEAKER_07: Yeah. Satya will take all your employees. Satya might steal your girl. He might steal your girl. Exactly.

SPEAKER_10: This Week in Startups is brought to you by 8sleep. Good sleep is the ultimate game changer.

SPEAKER_15: Now you can add the pod cover to any mattress. Go to 8sleep.com slash twist to check out the pod cover and get $200 off the pod plus free shipping. 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 Gelt. It's time to take control over your taxes. Discover how Gelt can help you to manage and optimize both your personal and business taxes.

SPEAKER_20: Visit joingelt.com slash twist now.

SPEAKER_19: All right, everybody, welcome back to this week in startups.

SPEAKER_21: You know, we had so much news breaking in the world of AI that, uh, Sonny and I had a conversation about it. We said, you know what? We didn't do any demos this week because we had to talk about, God, so much going on.

SPEAKER_00: And then more news drops in the meantime. Yes. Microsoft had a big announcement that they've gutted a company and stolen all their employees

SPEAKER_02: inflection.

SPEAKER_00: So explain just briefly, we're going to do demos today, I promise, but we do need to address what happened here.

SPEAKER_21: And there's a lot of feelings, I guess, or debates about it. And so maybe you can level set and tell us.

SPEAKER_27: So inflection was this creation between Greylock and a bunch of folks from the deep mind team, maybe mainly one of the co-founders, Mustafa, co-founder of deep mind, incredibly talented person who's built several iterations of AI. And when they came together, the promise of what they were doing, which was really awesome. They raised a lot of money for this was they had built a huge compute cluster. They were one of the largest purchasers of Nvidia hardware last year, and they were claiming to create both a new chat bot and a brand new foundational model. So it was a direct competitor to open AI and being built from folks that come out of the same lineage, have the same experience. The team had raised, I think, upwards of $1.2 or $1.3 billion to compete. And everyone was excited to see what they were building, what they were launching, what we mean more competition in the ecosystem. And then we wake up yesterday and we find out that Microsoft has recruited the team away. There's a new CEO in inflection, and we're all kind of scratching our heads.

SPEAKER_29: And we also learned that there was some kind of relationship between Microsoft and inflection as well. That was the other thing that kind of came together.

SPEAKER_19: So I did see that they also had like some number of investors in common, including Bill Gates

SPEAKER_21: and some other folks. So this, and then Reed Hoffman and then Reed. So LinkedIn to Microsoft, and I believe Reed is still on the board of Microsoft. And so they've gutted the company. And this was Microsoft's approach as well with open AI. When there was a little bit of friction there, they said, oh, you know what? We'll just take the open AI team. So this is like a very cutthroat Microsoft approach now of, Hey, we'll just steal your team. If you don't sell your company to us. And here they are, they've now essentially stolen a team. And then I guess there's going to be some backend payments, but they're giving up their chat. GPT. Competitor. I pie. Yeah. So that's over the dream of that front end consumer facing agent has been taken out of the market. So Microsoft now doesn't have that competitor for their windows operating.

SPEAKER_36: Co-pilot. Co-pilot. So they eliminated a competitor. They took all their employees or their main employees.

SPEAKER_39: And this, and they stood up a wholly owned sub called Microsoft AI, which, uh, Mustafa is a CEO of. Hmm.

SPEAKER_38: That's interesting. Yes. So what does that say to you?

SPEAKER_44: Okay. Well, let, let me, let me frame it in a way that's different, at least, uh, from the way

SPEAKER_27: I see it and what's going on in the market. Why is this happening? I think it comes down to the following reasons. I think one for large enterprises and Microsoft being, you know, the biggest one these days, AI is so important and their need for resources that can build defensible AI for their business. They're willing to do anything now that stretches from higher, the entire open AI team, when they had that big falling out a couple, maybe months ago to taking a company that their investors and that's had been funded through a well-known Silicon Valley investor and guiding it. This is how important building AI technology is to enterprises. And I think what Microsoft is showing is that if they really want to play to win, they must own that technology. We give them credit for pseudo owning it through the open AI relationship, but clearly they don't own enough of it. And they sort of pseudo owned this one through their investment relationship, but that wasn't enough. Yeah. And what it really means is the basics of business. If you really want to own something, you actually have to own it. And this is how important it is for them. They're willing to kind of, you know, not quite blow up, but do these funky things with these businesses in order to be in the market. I think that's a, I think B is these things require so much money and this kind of got uncovered in the Elon messages, right? Between open AI and Elon, where it's almost likely impossible to build one of these things as a standalone company, because if you don't have a giant machine that's generating cash, like in the case of Google, Google and search and the case of Microsoft and their core business plus Azure, it's probably next to impossible to build out the infrastructure needed to pull this off. And so I think that's the second reason, you know, what if the team was like, Hey, we're not happy with the 10,000 GPUs we have.

SPEAKER_39: We need a hundred thousand.

SPEAKER_49: Yes. So this is super interesting.

SPEAKER_21: Microsoft obviously is, you know, become the world's most valuable tech company, huge market cap. Now they added to the market cap after this, and they are under obviously regulatory scrutiny, like all of the major big tech companies, Apple and Amazon included. And what people are saying is this is a way to not have regulatory scrutiny, just hire the team and Lena Khan, the UK, we know, which screwed up the, um, we'll put the kibosh on the, uh, Adobe Figma deal. You know, you could just, Hey, I just hired a bunch of people. Yeah. It is what it is.

SPEAKER_51: And so now this is what happens in a free market. If you try to over-regulate a market and you get rid of M and a, well, now people have found a backdoor.

SPEAKER_21: I'll just hire the whole team. And. And that's.

SPEAKER_53: The outcome is probably as good as an M and a.

SPEAKER_03: Well, we'll see, because now what you have to do is, you know, how do those investors get paid their money back? Is the company going to buy the investors?

SPEAKER_54: Not for the investors.

SPEAKER_57: It's for the investors.

SPEAKER_55: For employees is probably even better.

SPEAKER_21: Yeah, who knows, but I mean, you would think in order to pull this off, they must have figured out a way to make those investors whole. And I think there's been some back channel about. But they're finding some way to do that, but this is putting somebody, you know, who's super qualified. And, you know, he was on the podcast. Mustafa was on the podcast this summer, August 17th, episode 1794, when his book came out. And, um, this is specifically, he's going to report to Satya Satya, and he's going to have Bing and their co-pilot is going to be his area. So this is super interesting. You have to wonder, like Apple and Google are not being super aggressive.

SPEAKER_28: It's kind of weird. Like, well, I don't know. Why isn't Apple getting more active here?

SPEAKER_27: Did you see this one thing? I didn't see it, but I saw this article. It said Apple acquired 23 AI companies last year. What? Yeah. Very quietly, I guess, huh? Yeah. I saw it and I was totally surprised by it.

SPEAKER_19: That's a new piece of information for me.

SPEAKER_21: I know they've bought some small ones. Six days ago, Apple buys Canadian AI startup, Darwin AI. So that one I know about.

SPEAKER_37: And there it is, nine to five back. Apple bought 30 plus AI startups last year.

SPEAKER_03: Wow. This is incredible. Let me take a look at this here. Okay. New information. Apple is said to have bought companies at an earlier stage in their development.

SPEAKER_66: Apple's pursuit of AI innovation has been evident in recent years. The tech giant has made a series of strategic acquisitions, including staff hires from AI

SPEAKER_03: startups to bolster its AI capabilities across various product lines. Apple purchased up to 32 AI startups by 2023, the highest number among tech giants in the overall AI startup acquisition.

SPEAKER_68: Hmm. When did you get a list of these companies?

SPEAKER_44: It says by 2023, not in 2023. Yeah, exactly. Ah, okay. That makes much more sense.

SPEAKER_70: Elite performers have a secret and it helps them stand out in their field. You want to know what that secret is? I can tell you it's sleep. Whether you're an athlete, an entrepreneur, or you're just a career driven professional, like the people who listen to this podcast, quality sleep is the fuel that powers your

SPEAKER_73: success. Well, how do you get deep, restful sleep? Very simple. You do what I do. Eight sleep. I got two eight sleep mattresses. I never miss a night on my eight sleep and they're selling more than just mattresses right now. You can sleep like a champion with the eight sleep pod cover. That's right. You don't have to get rid of your existing mattress. You put the cover on, slips right on top of your bed and instantly transforms your sleep game. We're talking about dual zone heating and cooling. So if you have a partner and you're both in the same bed, you know, one might sleep hot, one might sleep cold. I like it a little chilly willy. And this smart temperature adjustment keeps you in your ideal zone for a deeper uninterrupted rest. In other words, you set how you like it, but it adjusts the temperature to make you sleep more. This is really interesting science, folks. And this advanced sleep tracking will help you monitor your metrics, like your heart rate. And if you can measure it, you can manage it. That's the secret of eight sleep. And get that eight sleep pod cover. If it's any mattress, it's going to work well with whatever you got. And when you maximize your sleep, you maximize your potential. You know that. So whether you're having sleep problems or you just want to optimize, here's your call to action. I want you to invest in the rest that you deserve. Go to eight sleep.com slash twist and get 200 bucks off the pod cover.

SPEAKER_70: What a great deal. Eight sleep.com slash twist for $200 off the pod plus free shipping.

SPEAKER_66: We have to find this list of companies. This may have been over a very long period of time. According to Stocklytic, Apple purchased up to 32 companies. Back in 2020, Apple purchased Voices, B-O-Y-S-I-S, an AI startup involved in making digital voice assistants that naturally aid in comprehending natural language.

SPEAKER_03: They did that to improve Siri, ostensibly. They also acquired Wave One in March of 2023, whose technology is helpful in ample video compression. That doesn't seem like AI. Hmm. They bought drive.ai and AI.music.

Chamath Palihapitiya: They are known for making what they call tuck-in acquisitions at Apple.

SPEAKER_51: So some team of 20 people figures out like a technology, gets some patents. They're just like a really good SWAT team, but they don't have a public-facing product.

David Friedberg: Or if they do, they shut that down immediately, get rid of the website, scrub as much information as possible.

SPEAKER_57: They did this with an acquisition of a company called Swell, which was a podcast player. And that team became...

SPEAKER_84: They were a customer of ours at Xtreme Labs. Yeah, Swell was great.

SPEAKER_21: Yeah. Yeah. Awesome podcasting. We doubled our money. I was like, please don't sell. Keep going. It's going to be great. You know? And I think the reason why Apple's podcast player got a lot better and they took it more service, it was because of that application. But they immediately killed Swell. And Swell had this really great interface. It was tile cards. So you'd swipe the tile cards. I guess you were involved in building it. Maybe you did the Android app for them. But what was interesting is they would drop you off, you know, into a moment of the podcast. So if they knew you were into Apple, they might skip the first part when we're talking about Microsoft and just drop you into that Apple. And that was their kind of idea was like TikTok-ish understanding what you like about podcasts and

SPEAKER_86: then drop you off at the moment in time that you would like. That's a pretty interesting idea.

SPEAKER_88: Yeah. Way early on. Yeah. I was actually... I was an investor in a company called Locationary that they bought to clean up Apple Maps.

SPEAKER_66: Okay. Here we go. Apple is looking to add more acquisitions, even as it plans to bring on board Brighter.ai. Jeremy may start to improve its new headset, the Apple Vision Pro. Hmm.

SPEAKER_03: So let's get to demos. A lot of demos. Hey, congratulations, I guess, to the team, but maybe not the investors.

SPEAKER_05: Uh, congratulations, uh, Satya and Nadella to running the table on everybody.

SPEAKER_96: I mean, my Lord, this guy is like, he, I mean, this is, people were like, Microsoft is cutthroat

SPEAKER_97: back in the day. It's, this is getting super cutthroat.

SPEAKER_13: Yeah. Be careful out there.

SPEAKER_97: Careful out there, folks.

SPEAKER_13: Yeah. Yeah. Satya will take all your employees. Satya might steal your girl. He might steal your girl. Exactly.

SPEAKER_102: Be careful.

SPEAKER_101: Don't leave your girl on the dance floor.

SPEAKER_102: Satya comes right up. We're going to insert an AI video of Satya and Nadella stealing somebody's girlfriend on the dance floor. Oh my God. That's so funny. All right, let's do our demos.

SPEAKER_103: Come on. Let's get first one.

SPEAKER_27: We're going to, we're going to rock and roll here today. So Suno is back. And what I would say is, uh, they have a brand new model V3 and what they've done is, so I don't know if you remember this one, this is where you give it a description and it creates. Yes, I remember this. Yeah. And so they have a new model. That's why we're bringing them back here. Okay. And what I was going to say is I did some ones beforehand, but these were kind of the defaults. Give me a little description, Jake. How we'll do one live here.

SPEAKER_38: Um, a rock song in the style of Dire Straits. Oh my God. Okay.

SPEAKER_113: So I know what's going to happen here, but again, the style of, uh, because I'm saying

SPEAKER_112: Dire Straits. Yeah. Yeah.

SPEAKER_111: I want you to see what they're doing. That's it. Rock song in the style of Dire Straits.

SPEAKER_44: So if I do that, what you're going to see is right down here, uh, almost immediately, it's going to kick an air back, basically saying could not generate that song description contains artist Dire Straits.

SPEAKER_117: Ah, okay.

SPEAKER_44: So let's take out Dire Straits.

SPEAKER_118: Yeah. Yeah.

SPEAKER_21: They're being respectful of this, um, but it kind of kills exactly what we all want to hear here.

SPEAKER_36: So what we would say here, a rock song in the progressive rock style with finger picking electric guitar about a rock band playing in a dive bar and kind of alluding to Sultans of Swing here.

SPEAKER_113: Yeah. Yeah.

SPEAKER_44: Yeah. Yeah. So it's so funny. J. Cal, this is what this song up here was. And while that's coming up, I did it. And I basically had like chat GPT explained to me what the style of Dire Straits was.

SPEAKER_03: Ah, I mean, you can see how much fun this is going to be. I mean, literally this is going to be so much fun when they figure it out.

SPEAKER_126: This is, this is your one. Okay.

SPEAKER_130: That is literally, it doesn't sound like Mark, exactly like Mark Knopfler, but it is definitely

SPEAKER_57: like a bluesy, it's more John Mayer who is influenced by Dire Straits, but it's similar to Dire Straits.

SPEAKER_118: I mean, they understand electric guitar, progressive rock, they understand soulful, they understand finger picking.

SPEAKER_36: Wow. Wow. I mean, it's actually starting to sound like Dire Straits there. That is really impressive. That's super impressive. I mean, so let me hear the one you did.

SPEAKER_119: Oh, mine was not as good. So, uh, cause I, you're much better at describing songs than I am. Oh, got it. Let me do another one. Rock intro set.

SPEAKER_136: Yeah, that's, that's melodic. You got melodic. Let's do another one.

SPEAKER_36: Let's do one that is progressive punk rock about breaking the law and getting caught by the cops. I thought you were going to go for West end girls for a second by, uh, the pet shop boys. I mean, that would be more of skydane's playlist, but not exactly much.

SPEAKER_21: He's into those emo, pet shop boy. Okay. Here we go.

SPEAKER_142: Pretty fast paced punk rock. Okay.

SPEAKER_144: Okay. That's kind of pretty. No, didn't do a good job. No, no, no. Hold on. Play it again.

SPEAKER_145: Let me see something here. It's got the drums. Okay. Whoa. Did you say F the police? Yeah.

SPEAKER_147: Yeah.

SPEAKER_149: Right here.

SPEAKER_144: That's the police. Giving the finger to law enforcement F the police. Oh my God.

SPEAKER_155: I mean, that's a little hardcore.

SPEAKER_156: Yeah.

SPEAKER_163: Um, you know, maybe because I put progressive punk rock in there, but it went for fast paced punk rock. I just want people to understand here who are not watching. They should go to YouTube and type in this week and start up some subscribe and hit the

SPEAKER_00: bell and watch us do these because it's giving you in under a minute, a musical track with lyrics that are somewhat sync to it.

SPEAKER_03: Yes. And if you were to play this to me, I would say that is the, what did they call the tapes when

SPEAKER_130: somebody would give their audition to to sounds like a demo tape, like it's, it's bad, but it's a demo tape. Right. It feels like it could be a lot of the tape. You can just play the tape. Yeah, that's right. It's a demo tape, right? It's a demo tape. I mean, I think when I think this is about a instant. this to me, I would say that is the what do they call the tapes when somebody would give their audition to this sounds like a demo tape like it's, it's bad. But it's a demo tape, right?

SPEAKER_36: It feels like it could be good. If they were more produced, like if you gave this to a professional producer, they

SPEAKER_173: might be able to make something out of it. Yeah. I give this an egg. This is awesome. Check this out. We just built a new site

SPEAKER_57: instead of using a template like we normally would, I was prompted with a series of questions. Our preferred website structure and

SPEAKER_73: objective could be a new business model, whatever you're doing, the color palette, design elements and font pairings. And within minutes, it generated a beautiful website for us that was ready to go live and custom for what we're doing. This sounds like science fiction, doesn't it? Well, this is the new blueprint AI product from our friends at Squarespace. It's their guided design system for building new websites. And it's going to blow your mind because Squarespace is more than just a website builder. It's an all-in-one platform loaded with powerful tools designed to launch and grow your business online. They've put over a billion design combinations right at your fingertips. So you can make something beautiful and bespoke to you. So you know this AI will get your brand online. And this means you're going to have a unique online presence that's as unique as the brand and the vision you have. This is the use of AI that we've been waiting for. With Squarespace's blueprint AI, your website gets up and running quickly and it's completely custom. It's tailored to your business. So here's your call to action. I want you to try this. Squarespace.com slash twist for a free trial. And when you're ready to launch, you go to squarespace.com slash twist to get 10% off your first website or domain purchase. Squarespace is super affordable. It's super valuable. And when you partner with Squarespace, like I've done for a decade, you constantly get new product releases. This is one of the great product teams of all time over at Squarespace. I know them and they launch incredible things like

SPEAKER_57: this blueprint AI that you got to see to believe. Squarespace.com slash twist. Squarespace.com slash twist. What's really interesting here is this is it's not a proof of concept anymore. These are actually demo tape quality. These are demo tape quality.

SPEAKER_125: I'm going to play this one for you. This was your second dive bar one.

SPEAKER_178: So okay. Yeah. Oh, it's got lyrics. More country. Yeah. Yeah.

SPEAKER_183: No, that one wasn't good. That one wasn't good. So what's

SPEAKER_181: interesting there is they put auto tune on the person, they made it

SPEAKER_19: into more of country. And they might finger picking it didn't understand that I didn't want a song about finger picking electric guitar. I wanted the finger picking electric guitar

SPEAKER_21: science. So that sounds like my our description wasn't crisp. Yeah, yeah, it should give you tell me about the lyrics in one

SPEAKER_36: box. Yes. Tell me about the guitar. Tell me about the drum. So I wonder if we did that a drum style. That's, you know, hard charging an electric guitar style. That's finger picking a vocal style. That's a female vocalist. I wonder if you could even

SPEAKER_189: give it all those. Yeah, you can. It's it's, you know, kind of lets you describe it and how much detail you want to. I mean, here we

SPEAKER_21: are folks. It's pretty clear what's going to happen here. How this was trained is interesting because although they are not letting you put in dire straights or bad names, I believe this was trained

SPEAKER_111: on copyrighted music. I don't know that. Yeah. But I'm wondering, it does feel like, so this is going to be an interesting trick,

SPEAKER_21: right? Yeah. We don't let you say you want to create Darth Vader or Jedi's because we know that's protected IP. You can't make Marvel characters. You can't make dire straight songs, but we kind of trained it on Marvel and dire straights. So if you do enough, you know, prompt engineering, you can still get

SPEAKER_03: that. I think that, and is this tell everybody the, the URL of

SPEAKER_193: this so they can try it themselves. I'm just pulling up this. It's

SPEAKER_194: yeah. Suno dot AI, right? S U N O dot AI. Okay. What I want you to

SPEAKER_27: see here is from a computer commercial use, uh, aspect, you are allowed to take these songs and put them in monetization services. And so as long as you're a subscriber, and basically if you are a paying subscriber, you own the songs you generate while subscribed. So as long as you subscribe, you own it. So what's interesting though, is

SPEAKER_105: while you are subscribed. Oh, while you are subscribed. So if you

SPEAKER_168: unsubscribe, they get the rights back. Yes. Oh, that's it. Oh,

SPEAKER_111: maybe that's just a weird way of saying if you are paying

SPEAKER_03: subscriber to sooner, then you own the songs you generate while subscribed to pro or premier subject. Okay. So if you stop your subscription, that's not worded very well for the team over

SPEAKER_38: there. You know, interesting business model, make a great song, subscribe to them. In one of those plans, use it in your

SPEAKER_44: YouTube for audio or, you know, background music. That's pretty incredible.

SPEAKER_21: Yeah. I like that. They're dipping their toe into the rights issue and trying to explain it. Yeah. I would like to know the

SPEAKER_204: training data. Yeah. So we'll, we'll look into that. Hey, if you work at Suno, uh, let us know how did you train this show? Exactly. I'd like to have them on the show. That'd be a great time. Come on. Great job.

SPEAKER_27: And I want to give it a name. I give it a name. Okay. I'm going to give them, I'm in the same spot as you. I also want to say there's one more thing I just wanted to show that they had here is they kind of have this trending. So you can see by the number of plays. And so I'll just play one of these for us right now.

SPEAKER_113: Unless something jumps at you here.

SPEAKER_214: Yeah. One of the top, any one on the top one sounds good. Definitely not speed metal, but yeah, this one sounds cool.

SPEAKER_216: So electric punk.

SPEAKER_219: I mean, this is like, I was listening to some of these, honestly,

SPEAKER_27: for me, I couldn't tell if one of these songs made it into the, you

SPEAKER_44: know, billboard top 100 and whether it was real artists or not.

SPEAKER_222: Yeah. I don't know if we have a bet on this, but if we do, the under is going to win.

SPEAKER_21: I mean, it's pretty clear that we saw that they could do Gwyneth Paltrow's

SPEAKER_03: voice, for example, in speechify. So, you know, or getting me to speak Chinese and do you and I can do this in French or Japanese or Chinese, like it's okay.

SPEAKER_36: We know we can get speaking to the point at which you really can't tell the

SPEAKER_21: difference, whether it's a robo call of Biden or lip syncing, you know, us

SPEAKER_03: speaking in another language. So, okay. Check box there. And now it's kind of clear, like we knew lyrics would get done. That's not difficult. We know. Artwork for albums can get done. So the real question was, could you

SPEAKER_21: actually make a melodic, interesting song? So can music and the singing

SPEAKER_03: voice get there? And it's pretty clear that it can.

SPEAKER_227: It's impressive. A's, A's all around. A, A, A, A, A for awesome. It's going to be great. Yeah.

SPEAKER_69: We got a little breaking news here.

SPEAKER_66: Oh, okay. Wow. Open AI faces multiple lawsuits over chat, GPT's use of books, news, articles, and other copyright material. In its vast corpus of training data, Suno's founders declined to reveal details of just what data they're shoveling into their own model. Other than the fact that its ability to generate convincing

SPEAKER_03: human vocals comes in part because it's learning from recordings of speech. In addition to music, naked speech will help you learn the characteristics of human voice that are difficult. Showman says from Rolling Stone.

SPEAKER_21: And I think our bet was, um, that there's a top 100 song by AI.

SPEAKER_232: Okay. And so, you know, it's, I'm going to make a thousand of these in one. I don't win that bet. I'm going to be like the streaker.

SPEAKER_21: People don't know somebody bet, whatever, 10 to one odds. There'd be a streaker at the Superbowl. So then they decided to get a place to bet, get butt naked and run across the field.

SPEAKER_235: Sometimes you got to do what you got to do to win.

SPEAKER_227: And we'll, we'll put a link to the, um, the Suno story from Rolling Stone, uh, by Brian Hyatt in the show notes. So you can read that directly. Thank you to Rolling Stone for asking that hard question. Let's keep going here. Let's keep going. This is a good area.

SPEAKER_27: I think you're going to like this one a lot. Go here. So we haven't really talked about them a lot. They are one of the foundational model providers and they released a new model

SPEAKER_38: called command R command R, which is, I believe to refresh the page.

SPEAKER_238: Yeah, it is. It is. Yeah. You got it right.

SPEAKER_44: But what command R is, you know, claim to be focused on and

SPEAKER_27: really good at is reg stands for retrieval, augmented generation. Effectively. What that means is using a model to go get, uh, not the source of truth from its training data, but to go use an external source to go get the training data, to get the information you're looking for. So using it as like a, almost like an agent. Okay. And so in this case, I've pulled up their playground and what you can do is, uh, you can say, Hey, get me a quick overview of, um, global market for solar panels. It's just one of their built in defaults and it's going to run here, but while it does that, I'll just kind of, I did this a few minutes ago. And so what you'll see here is it went and did this search. It did this grounding search and it put all the references right here.

SPEAKER_66: Okay. So you say, can you give me a global market overview of solar panels? And then it went to the web and it found a bunch of different articles.

SPEAKER_03: Um, that's a solar panel market size, share, et cetera. Okay.

SPEAKER_38: So it's doing a web search. Yup.

SPEAKER_44: And then what it did was it put this like, you know, short summary together with the links to where it got all the data from.

SPEAKER_244: Okay. Citations, which is what we've always been asked.

SPEAKER_21: I've always been asking for, and some people don't want to do it because I think the

SPEAKER_03: reason people don't want to do it Sunday, they don't want to get caught with their hand in the cookie jar, so they'd rather pretend that they didn't take this information.

SPEAKER_05: This is different.

SPEAKER_27: Cause it's searching the web. Exactly. So this is using the model as a reasoning engine to search over the web and then it's putting it here. So this is not coming from the models training data without it being connected to the internet saying, oh, well I went and I went over this site and I realized that in 2023, the market was 165 billion. This is a distinctly different thing. And I think let's spend another minute making that clarification. What you talk about a lot is when someone takes a model and basically in the training data, they go over things and when they regurgitate this information without a reference, that's not, um, you know, falling under, I guess what you would say is fair use.

SPEAKER_232: I think that's how you've categorized it. Yeah.

SPEAKER_51: And let's just even use the word fair, like put aside all legal concepts. It doesn't feel fair that somebody did all that work and then somebody else, you know, got the value from it.

SPEAKER_21: And here when you do a citation and you take a small portion of it and you link to the source, I wouldn't mind that if you, I was literally talking to somebody who's got like a, a Twitter and Instagram handle called startup archive or something. And he like every week does one of my videos. And I said, Hey, listen, I'm okay with you doing it, but would you please put at Jason and at TWI startups in your first tweet, because he wasn't actually linking back to this week in startups. So people had to guess who I was in the video. And I'm like, Hey, here's the conditions under which I'll let you do this. Put at Jason at TWI startups and in the followup, put a link to the source and then I'll let you do this, you know, once a week or whatever, which I try to be reasonable about it. So if they're a fan of the show, I thought that was fair, are you grinding hard to grow

SPEAKER_253: your business? I bet you are.

SPEAKER_73: You're listening to this week in startups. Of course you are. But don't let your hard on profits slip away because of overpaying on taxes. You need to check out GELT is the secret tax weapon trusted by savvy founders and CEOs. Their elite solutions will transform how you handle your taxes. You can integrate your personal and business tax planning with one provider. They have tech-driven efficiency that simplifies tax management and in-house advisors with startup and business expertise. This means no more overpaying your taxes. You're going to save time and these expert opinions will boost your financial health and

SPEAKER_57: help you grow your business. This helps you stay compliant so you can focus on your startup's mission with year round tax advice, personalized strategy sessions, and a comprehensive tax library for continuous education. When you optimize your tax strategy, you optimize your competitive edge. So here is your call to action. It's time to bring GELT's elite tax team into your business. Visit joingelt.com slash twist and get 15% off your first year. What a generous offer. Thanks team GELT. Join gelt.com slash twist to get 15% off. From your taxes from a liability to an asset. That's 15% off your first year at joingelt.com slash twist. So anyway, this feels fair. It's great.

SPEAKER_05: The other thing is data and facts are not protectable. You cannot copyright data and facts.

SPEAKER_21: So if you were to make a website and you were to put, um, these are the heights of all of the NBA players in the NBA. Um, now I could take that data and put it into my database.

SPEAKER_36: You cannot claim that those facts are, you know, copyrightable now.

SPEAKER_21: So does that you can say scores?

SPEAKER_262: Okay. Interesting.

SPEAKER_21: Yeah, sports scores, whatever it is. There are some exceptions that were people have tried to put copyright onto it.

SPEAKER_51: One of them is my terms of service are that you can't scrape data. And if you do, you're breaking into my server. This was the LinkedIn case versus an Israeli company. I believe that LinkedIn wound up losing. Okay.

Chamath Palihapitiya: So, you know, scraping data off of crunch base or pitch book or other places, like if it's

SPEAKER_51: behind a paywall, you know, and you take that data and you've scraped it, you know, they can make some legal claims against you.

SPEAKER_21: It depends on jurisdiction as well. Obviously in Israel and some other countries, they, they, they don't stop me from doing that, uh, but you would have to get past the paywall. And really, it just depends on if you're being fair with people, but I could go on to at, for example, crunch base or LinkedIn, take the CEO's name of every company. In San Francisco and make a new website with that information. But if I scraped it, they might be able to make the claim, or if it was your website, you know, San Francisco startups, you could make the claim that I invaded and I broke into your servers.

SPEAKER_44: And it's, that's why they probably put them all behind logins now too, right? Cause then that forces you to agree to a terms of use.

SPEAKER_204: That's exactly why they do it. So the name of the company was high Q, um, and, uh, LinkedIn.

SPEAKER_191: Yeah, so if, um, I'll just read it to you really quick, cause I think it's, I'm super

SPEAKER_03: fascinated by this high Q ink versus LinkedIn core case about web scraping. The ninth circuit affirmed the district court's preliminary injunction, preventing LinkedIn from denying the plaintiff high Q labs from accessing LinkedIn's publicly available LinkedIn member profiles. High Q is a small data analytics company that uses automated boxes, scrape information from public LinkedIn profiles. The court ruled that high Q had the right to do web scraping. However, the Supreme court based on van burn versus United States decision, vacated decision and remanded the case to further review in a second ruling. They affirmed the decision.

SPEAKER_66: And then in November 22, the usage court, Northern district of California, where the high Q had breached LinkedIn's user agreement and a settlement agreement was reached

SPEAKER_36: between the two parties. So even though they had the right to scrape in, they didn't have the right to break the terms of service. If that makes sense. Right. Um, and so that that's how that sort, and then I've said this before on this podcast and on all in most of these things get settled.

SPEAKER_21: Right. And so it's very hard to actually know what the law is in a lot of these cases. And there's not like a bright line. It's always like interpretation. And the interpretation here, I think was, Hey, you can scrape and read data off the web and you know, the, the data is not copyrightable and an article might be, but not data in the article.

SPEAKER_167: However, you also have to, you can't break somebody's terms of service to use their service, right? Cause that's an agreement you're making when you log into the website.

SPEAKER_39: Super complicated. How does, um, well, I guess in this case, let me just put my share back up.

SPEAKER_44: Um, in this case. We're okay because they're citing it. So we don't run into that issue.

SPEAKER_19: You know, even if you cite it, if you did scrape the person's website and break the terms of service, they could still say you're broke into my servers.

SPEAKER_21: Basically you broke our agreement. I gave you an agreement to use my website. You agreed to it. Therefore you've broken it. So if these folks have a terms of service, it's possible that this is, they broke the

SPEAKER_167: terms of service by scraping it like that.

SPEAKER_123: So when I clicked on this right here and it took me to this, this came from grand view research, solar panel, and they've ingested the entire thing and they've, they're

SPEAKER_36: presenting it on their website, that's illegal, they didn't ingest it.

SPEAKER_27: I asked this question. Let's go through this workflow. Okay. It then went and searched the internet and then it basically used, um, its web search connector to go through these, you know, different references. Yeah.

SPEAKER_21: But see, it's got the whole entire article there. So they have republished the entire article on their website. That's the part that would be illegal. Oh, but do you consider this a republishing rate? Sure, the whole thing's there. You scrape the whole thing and publish it on their website.

SPEAKER_51: So you don't have to visit their website. Now they did not have the right to republish it on their website.

SPEAKER_143: If they put like, if they put a snippet and linked back to it, that would be fine. Yeah. Okay.

SPEAKER_51: So on the left where it's citing like one sentence, yes, that would be the equivalent of Google doing that one box where they just take a clip. Yeah. But on the right here, they've, they've republished it. So if I was advising them on fair use and legal issues, I would say, Hey, you know, here may be truncated and say, visit the website because then grand view research

SPEAKER_03: from the web solar VP house, whoever this is grand view research on solar panels, market size, they're getting no value here. You've just stolen all their value. This page that got it.

SPEAKER_21: So this page that you linked on and they're trying to sell reports, you've basically sucked in the report, you've intercepted and given all the value, essentially in my

SPEAKER_03: mind, stolen value that would have happened if they went there. And so that's the challenge, I think with these kinds of services now, what somebody like Freeberg will say, or, you know, a lot of like Google techno, you know, first people will say, well, if it's on the open web, then you can do whatever you want. And that's just like a level of naivete or like techno might is right.

SPEAKER_51: That's just ridiculous. Like just because something's available on the web, like listen, Netflix is available on the web.

SPEAKER_291: That doesn't mean I can take, you know, queen's gambit and republish it.

SPEAKER_113: And, you know, it's a little copyright immaterial. So your recommendation would have been, they can look at this entire page, which

SPEAKER_27: is, this is the page they looked at. Yeah. This, this is the whole page. Yeah. They can go and extract this 165 billion to 355 billion, which somehow came from this page somewhere. Right.

SPEAKER_165: I'm sure if we look for, unless that page has the terms of service are, you cannot republish this, you cannot pull this data and republish it.

SPEAKER_21: Yeah. And they agreed to it. Now, if that page is on the open web and you didn't agree to the terms of service, but the bot, now the question is, does the bot have to agree to the terms of service? Oh man.

SPEAKER_51: And the bot does, if you send the bot to go do this stuff, you know, it does. And then this is why people create bot for watering software. Yes.

SPEAKER_36: And so I bet you the way this is working is they're firing up a browser to do it. And this is where like being fair is the easiest path forward. You know? Okay. And I think the new standard will be robots.txt. And, um, uh, language model.txt or AI.txt, where they can say, Hey, here's our rules. You can, um, use this one time in a session.

SPEAKER_51: If you put at the start of the sentence, the following information is from this website for the complete information, go here.

SPEAKER_21: Uh, and then you can present up to 10% of the page or something like that.

SPEAKER_193: How do we make a better internet? Because that's still kind of is a disjointed experience.

SPEAKER_27: Like I like what they did here and honestly, I wasn't thinking about it. Like, I was like, Oh, this is great. They got this information. They linked to where they got it from. And, you know, but I see your point.

SPEAKER_38: How do we make a better internet?

Chamath Palihapitiya: I think the robots.txt and a lot of the early Google work was pretty good. They always took a small portion of the original work and they linked back. And so they could always say, you know, and then they said, if you don't want to be

SPEAKER_51: in the index, you don't have to.

SPEAKER_21: Yeah, so you could opt out of it. Now, most people would say opting in would be the higher standard, but okay, let's just say opting out at least gives the person who owns it some power. And all you have to do is just put website.com slash robots.txt and you go, but they also said, we're only going to take a small portion of the original content, the abstract. Yeah. And they even got rid of recently the web cache. Remember they had web cache where you could see the whole page.

SPEAKER_03: They got rid of web cache. Cause I think that was causing issues around copyright as well. I suspect that's why it went away, but I can't confirm that.

SPEAKER_21: And so a better way to do it would be to say, if you want to use me in your language model or your AI product, um, we have a license and here's how you can pay for that license. Go here. And it might be, you pay automatically through an API.

SPEAKER_51: It might be a clearing house, like the music industry has, or it might just be talk to our sales team. You know, if it's ready, they might just say, Hey, you know, if you want to license it, talk to this person in sales and set up a meeting. Um, and so, yeah, some sort of automated rights would be good. If you look at YouTube, they also came up with a solution, which was if somebody is using my content, I have the right to claim the advertising.

SPEAKER_36: Right, so if you were to do a Mark Knopfler song on your channel, if we played Mark Knopfler right now, his representatives would claim this video, put ads against it and they would get the money.

SPEAKER_305: They get all the monetization, all the monetization.

SPEAKER_36: So I, as a creator can be like, you know what? I just wanted to play a dope song. I don't care about the $6 I'm making from this video a year. Yeah.

SPEAKER_51: Go ahead and take the money. I'm not in for the money. Now, other people, there's an entire genre of reacting to songs. Yes, and they claim fair use and it's a bit of a hack because people type in, I want to listen to Sultans of Swing.

SPEAKER_56: Oh, yeah, there's a ton of these videos. It's like the first time someone's ever heard.

SPEAKER_03: Yeah, so just type in Sultans of Sing reaction video and pull it up on the screen and you'll see it on YouTube. I like watching these sometimes because it's kind of funny for me to see people, but it's now become offshore.

SPEAKER_51: So there are some people in like, I think the Philippines or it might be somewhere in Africa are, they have armies of people doing this

SPEAKER_21: format, which is react to videos. And there's so many look reaction comes up automatically. Wait, do you see how many there are?

SPEAKER_167: Okay. So this one's from two years ago. Next one, four months ago, next one, three months ago, a real number of views too.

SPEAKER_310: Yes.

SPEAKER_36: Next one, 10 months ago, another one, another one, keep going. And those guys are like the originals by the way there. And then I know that guy is a famous person, but like you can see how many of these videos there are. And then what you'll notice is a lot of the folks are from foreign countries and you're like, wait a second. What's going on? I think that some group of people in other countries have figured out this is a hack.

SPEAKER_21: Now, if you pull one of these up and you look at the that guy, classical composer, reacts to the Daily Doug, I love the Daily Doug. This guy's great. He's a professional composer and he's had a problem because he responds and he goes really into detail. It's super educational and he doesn't play the song all the way through. He pauses it and stuff like that, so it's not a replacement for Spotify. Okay. And some of them, they have disclaimers and then some of them, these get, you know, he loses his monetization. And then he said in his videos, I, you know, I, I don't fight every claim because he makes his money from Patreon.

SPEAKER_51: But then there are some people who insist he can't publish it.

SPEAKER_57: So then he gets multiple strikes against his account. And if you get three strikes on YouTube, your account gets turned off forever.

SPEAKER_19: And so three copyright claims are done forever. And so you have to be careful if you do too many of these and you have to fight them.

SPEAKER_21: So he'll spend all this time on a reaction video of a certain song. Okay. The artist will not say, I want to claim it. They'll say, you can't use it. Oh, okay.

SPEAKER_39: Stop. Wow. Cause I was going to ask, can't they just claim it? But they, they know they're so pissed off. They just say, you're done.

SPEAKER_113: You're done.

SPEAKER_21: And then he gets caught in that copyright claim. So anyway, there's abuse in every system. And so YouTube and Google have navigated this over decades and I think there's

SPEAKER_36: something to this give and take where people get to opt out or they have a way

SPEAKER_21: to collect the money from it that, you know, feels at least somewhat fair. You can actually publish an entire version of star Wars. If you edit. Yes. Yeah.

SPEAKER_36: So people have re-edited like the three prequels into one movie and then put a clone wars video into it and they can just have like their own movie, like

SPEAKER_21: their own two hour movie out of like seven hours of videos and they publish it on YouTube or other places.

SPEAKER_03: And they, as long as they don't claim ownership of it or monetize it, it seems

SPEAKER_316: like Lucas art is okay with that. We started in one place, which is totally different place.

SPEAKER_27: No, but it's all related though. We started in a place where these guys have created this, I think a really good model. I thought, you know, they did a really good job and I think you agree. They've done a good job of citing where this has come from to me of all the models that are out there right now, this is awesome because you can now at least press

SPEAKER_29: that this is not a hallucination and you can go find out where these numbers come from.

SPEAKER_318: Yeah. I've always felt sourcing is super easy to do the way they're doing it is they

SPEAKER_36: put the sources on the side, as opposed to putting them in line, like you would do in a college paper. I like when you put the number one at the end, um, like you do in a college

SPEAKER_21: paper on Wikipedia.

SPEAKER_51: So I would like to see, I would just make that one little edit for the citations, but

SPEAKER_21: otherwise, I give it a B plus. I give it a B plus. I would use it. I like it. I think it should exist in the world and people should have the ability to opt out of it. And what they should aspire to do is send as much traffic and give as much

SPEAKER_51: recognition to the people who they get the content from so that they don't piss them off.

SPEAKER_27: Yeah. Maybe it's just like you said, a little bit, a little short snippet here on the right or down below and then basically link people out so that the traffic, I give it a B plus.

SPEAKER_321: What did you get?

SPEAKER_27: Well, so for me, like, you know, we've been spending a lot of time in and around reg for the last year and a half. I was blown away because the number one issue I have, and it's funny because I ended up having this argument with friends where they pull something up from Chad GPT and I said, well, you know, that doesn't necessarily mean that it's real. And then, you know, we'll do like a search and find out that it wasn't real.

SPEAKER_44: So for me, this was an A because I would send everybody to command R and say, if you really want a Q&A style approach to a language model, this is where you can do it and you can get the references so we know that what it's telling us is real.

SPEAKER_324: So what is your grade again? I'm giving it an A.

SPEAKER_44: Okay, perfect. I'm giving it an A. A and a B plus.

SPEAKER_324: Yeah, it's great. Yeah.

SPEAKER_27: Awesome.

SPEAKER_44: Yeah.

SPEAKER_27: I had a friend that recently got in trouble on a flight. He was having a couple of drinks before and then when he boarded, he was carrying a drink in and then he got rid, yeah, exactly. And he got rid of it just as he was sitting down and he wasn't trying to do anything nefarious. Then they asked him to get off the plane and he pulled out Chad GPT to ask what the rule was that he was trying to show them. And then basically the, you know, the, the agents and everyone else like, yeah, that's

SPEAKER_113: not, that's not the rule. Like you can't walk on a plane with an open drink. So. And did he wind up getting bounced? Yeah, he did.

SPEAKER_36: Yeah. I mean, my advice is, yeah. If you, if you really like alcohol that much, there's something about people who love alcohol. I'm not a drinker, as you know, I've seen you have a cocktail once in a while. You know what you can do if you're really that into it is you could get a bottle of Snapple and put your vodka with Snapple and not have this issue. But I always see these dopey people, not your friends, Toby, but other dopey people who pull out, like they pull out a bottle of like vodka on a plane.

SPEAKER_00: And I'm like, are you dumb? Like, just pour it into a Snapple bottle and you're drinking a Snapple. No, or if you're going to a concert or whatever, you were walking down the street. And then second, you can't have a open bottle of Snapple's way to go.

SPEAKER_44: Like you got the, I think he was just, you know, there's like, he was, his plane was delayed and he was drinking at the bar and he just took the drink from there. That was it. Like, sort of. Yeah.

SPEAKER_336: He gave it wheels, as we said, he gave it wheels. Oh, I like that. Gave it wheels.

SPEAKER_19: All right. Well, you know, this was a drunk driving, um, very poor taste joke.

SPEAKER_00: When I grew up in Brooklyn, people would ask you, Hey, you want to give it wheels? And people were like, yeah, give it wheels. And they would just put your drink in a to-go cup for your ride home and people were tanked.

SPEAKER_336: So don't drink and drive folks. Don't drink and drive. Don't do it. Don't do it. Okay. One last one.

SPEAKER_27: This one's really cool. So, you know, uh, end of last year, we talked about multimodal models becoming really, really good. This is a Chinese company actually called deep seek. And, um, they have this new model, uh, very small, only 7 billion parameters.

SPEAKER_44: And, uh, you know, they have some different examples here, uh, then I, I just pulled one of them up, which is how many people in this image and why. And so basically there's one person is wearing a pink shirt, standing in front of allow, you know, large mirror.

SPEAKER_27: So it's able to figure out that it's a single person standing in front of a mirror. Why I really liked what this team has done. All these models are open source. They're available to use. And to get a 7 billion parameter model. That's able to, you know, reason, this type. reason exactly over images, this is going to give us superpowers.

SPEAKER_03: And so this is going to end up everywhere for us. This is really interesting. You can see this on hugging face.co. Yeah. Yeah. And the name of it is deep seek. Just do a search for D E P S E E K. And the image given to it is how many people are in the image period space. Why question mark. And it's a person holding a mirror as if they're putting the mirror on the wall, like they're installing it. So it's a round mirror and you can see it's got the reflection of the person.

SPEAKER_57: So you would think the computer would get it wrong. That there were two people in this image, but it says, this is one person in the image. Period. The individual appears to be a man wearing a pink shirt. She doesn't want to assume gender. It's a little bit woke standing in front of a large round mirror. That reflects his image. He is seems to be adjusting or touching something on the wall beside him, possibly checking for any infractions or making sure it's aligned properly.

SPEAKER_00: So it knows, cause I said, it looks like he's trying to adjust, uh, this action suggests he might be involved in some sort of home improvement project or simply ensuring everything looks neat and tidy. I mean, that is, that's just next level. Come on. It's next level. They're, they're describing the person's actions through the image. Image that's bonkers. Like how, how many images must it be trained on to understand? Somebody's involved in a home improvement project that they're in a mirror, that they're a man, you know, that they're wearing a pink shirt.

SPEAKER_36: I mean, this is nuts and it is kind of a pink jacket.

SPEAKER_51: So it's not a human would know that's a pink members only jacket as opposed to a shirt, but Hey, at least there's something that a human can figure out in this image. This, this is an a, yeah. Yeah.

SPEAKER_96: I mean, based on this one, but is there another one? Click on one more and let's take a look. Yeah. Yeah. So there, there's, yeah, we'll just do, uh, which one of these do you like help me write Python code based on it? Oh, what is this app about? I like that one too. Both of those. I like it pick either one. Let's do this one. Okay. Let's do it. All right.

SPEAKER_03: So it's showing, okay, help me write a Python code based on the image and the image is a thing. And if this, then that, right. Kind of situation image depicts a flow chart.

SPEAKER_66: Okay, for a Python code that simulates a game where the user is asked to guess a number between one and 10, the flow chart includes several condition statements, which are represented by different color boxes with text in them.

SPEAKER_03: Here's how you can write the corresponding Python code. Okay. There we go. I mean, yeah, it's just getting so good. And I've come to the conclusion that even though many of these things are not ready for prime time, I'm convinced that they will figure it out at a rapid pace. Unlike say self-driving, which is incredibly hard because of the edge cases and because of like, you know, what's at stake, like what's at stake describing an image or writing some software code that you're going to edit anyway.

SPEAKER_51: Cause you could take this and put it into Devin, right? Yep. Yeah.

SPEAKER_44: I mean, it's an, this is a good, another one, which is like, what are your, it's like, Hey, what is this app about? So it's like minutes, uh, listen, the meditations points.

SPEAKER_27: Courses saved. The app appears appears to be a meditation or mindfulness tracking application.

SPEAKER_53: This is indicated by your statistics section, which shows metrics such as meditation minutes, listen, and points. Right. Additionally, there's a, your mood.

SPEAKER_357: This stuff is incredible.

SPEAKER_51: Yeah. It's really, and if you can describe these things, what's interesting is then you could write the next series of prompts. So by describing what's in the image. Now you could say, what questions would humans have next?

SPEAKER_03: Yes. Right. Well, what would a human being like to do next, you know, um, oh, okay. Yeah. Let's go ahead and ask it.

SPEAKER_359: Given this information.

SPEAKER_03: And let's see what it does. Yeah. Like who makes the app maybe, or. Yeah. Yeah. Basically provided in the image. Humans might have the following questions. How many minutes did I spend meditating today? What is the total number of meditation sessions completed? How many courses have I done?

SPEAKER_363: Oh, yeah.

SPEAKER_03: It's incredible. Can I set reminders for when to meditate or listen to content? Yeah. I mean, it's incredible. I mean, it's totally incredible.

SPEAKER_21: Give us 10 questions that somebody who is using, they took the perspective of the person using the app, what they might be.

SPEAKER_51: Yeah. Wow. Are there any specific activities or challenges within the app that contribute to earning points? Like how do I hack this app to be, to level up?

SPEAKER_136: I mean, wow. Isn't this just wild? And with 7 billion, uh, parameters and the size of this, I could run it perhaps on my M3 laptop.

SPEAKER_113: You can for sure run it on your laptop. And this is getting to the place where you can probably run it on your phone.

SPEAKER_36: I mean, here we go, folks. It's a brave new world. And Sonny and I are here to help you with it. This week in startups.com slash AI. This one, an A. Let's give them, this was an A, right? It's an A. A. Yeah. A. A. I was talking to somebody the other day and they were just telling me, you know, ah, you know, these big models are expensive.

SPEAKER_51: Yep.

SPEAKER_36: I was talking to the CEO of Glean, I think, uh, who's just in the program and, uh, I think Mamoon from Kleiner backed him. And so the, the, the head of Glean was on and he's like, you know, sometimes like we're seeing customers who can use a, I think it was him, but I may have juxtaposed with somebody else.

SPEAKER_51: Like, it might be better to ask the question three or four times to these other cheaper, faster models, and then ask a follow up question on that and then give an answer. Yep. Than to use the most expensive model. Yes. That's called reflection.

SPEAKER_27: It's actually a term with an X. Ah. They call that reflection, just like humans reflect. And so one of the things is all of these scores, these benchmarks you see, J. Cal, they're based on one shot. If you do it multiple times and if it goes fast enough, then who cares? Because if it's just running fast enough, do you really care that it answered in one, you know, because you can get the score way up.

SPEAKER_03: So do a bunch of smaller models, get five answers real quick, but then how do you determine which is the best answer? Or how does AI determine what's the best answer?

SPEAKER_108: Well, that's really the grounding. The grounding. Yeah. I mean, there's different ways.

SPEAKER_27: One, the humans obviously do the reinforcement learning, but two, the grounding against actual data. Right. So like what we saw earlier with command R, that's the intersection of these things. That's why I don't think we'll ever have a moment where they'll exist without the internet. They'll always have to have the internet and the internet will become, it's sort of the way for these models to go ground themselves in whatever resolution they've come to.

SPEAKER_113: Got it.

SPEAKER_36: This will be like their fact-checking department, their reference desk, their card catalog, their library, their Wikipedia, just a way for them to level set that they got it right.

SPEAKER_03: All right, listen, if you guys want to get it right, this week in startups.com slash AI and x.com slash Sundeep, x.com slash Jason. You can follow the show TWI startups on Twitter. And I'm doing some TikTok experiments, so go ahead and search for Jason Calacanis on TikTok.

SPEAKER_204: I did my first TikTok today where I just mentioned Tiananmen Square, the Uyghurs, and banning TikTok.

SPEAKER_51: And I just want to see if, like, it gets past the sensors and anybody sees it.

SPEAKER_341: Don't, don't, don't do it, don't do it.

SPEAKER_388: I'm tempting the, uh, the CZB. It was nice to know you. It was nice to know you.

SPEAKER_391: Hey, listen, it was good to know me. I'll be in, uh, I'll be getting re-education with Jack Ma. We'll see you all next time. Bye-bye.

SPEAKER_393: Bye.