Dan Shipper: The First MultiModal Media Company

24:51If it can respond, it will draft a reply for you so you don't have to do the repetitive stuff. Everything else gets automatically archived and twice a day you get like a really beautiful summary of all of the emails that you normally need to look at need to like look at and archive and whatever but now you just scroll through it and you're done um so your inbox is just way quieter and and more beautiful that's fascinating to me that ai can be used to gain leverage in these non-technical industries and and i think quora is like a great example of of of something people can use thank you i appreciate that so quora um has a gm his name is kieran okay um and he built quora end to end he's also written articles on every okay uh he is like weirdly like a uh he spent a lot of time as a professional composer of music and a baker.

25:34Okay. So he does everything sort of full stack and we provide him a lot of help and when there's a lot of collaboration and all that kind of stuff, but like he owns it end to end. And he owns Quora. Yeah. Got it. Like he owns, you know, he owns the product. Okay. Yeah. Does he know how to code? Yes. Or is this pure vibe? No, he knows how to code. Okay. He's founded several companies before, was a technical co-founder of several companies. Okay. A lot of the people that work for Every have that like sort of multi-dimensional skill set and particularly for the eirs they're like owning the full stack but like we also have like um for example we have this with this writer um katie when we have eirs right sometimes they're good writers sometimes they're not and she does a lot of like co-writing with them to help them like get out their ideas um but she's also like using now using like lovable all the time to like build little tools for herself oh wow um or we have another writer her name is ria who's doing the same thing as like building custom gpts and like so it's all kind of they're all these people are like now curious and interested and like they sort of see how powerful it is.

26:32And once you get them going, they're like, oh my God, this is amazing. You know? So I think we have an environment that stimulates that and values. I think in general, startups really value generalists. But this is like incredibly empowering for generalists if you kind of like allow them the space and freedom to play. So like one of the things we do every quarter is we do this thing called Think Week where we don't publish anything new. we don't do any meetings um every day there's a theme but the idea of think week is to sort of recognize that most of the time in a startup you're spending time being like sort of very reactionary yep um and you're just like under like constantly under fire like trying to make sure things are not breaking and like whatever the best creative work comes from a different sort of place where you're not reacting to your circumstances you are like proactively kind of like playing around and like following like that thing that you just are psyched about and Think Week is really about getting back in touch with that.

27:29So we don't do any meetings, we don't publish anything new to the extent we can. Like we're like, you know, pausing a little bit on some of the product stuff and every day there's a theme and the idea is pay attention to stuff that inspires you, pay attention to whether you're not, whether or not you're in that sort of reactive mode and sort of get into a play mode. And then like one of the days this time we did like a day where it was like experiment with a new tool that you've been meaning to experiment with, but you haven't. And that's how Katie started using lovable. And I think that that's actually so fucking important for businesses right now.

27:59And we do this with, we have a consulting arm where we do consulting with big companies and help them use AI. And this is like one of the big recommendations is, um, find space to play. Yeah. Like every employee has this, like, um, it doesn't matter if you're employee executive, whatever founder, you have this decision you have to make every day, which is, do I do things the way that I know how and get them done? And then I have so much work to do that. Like if I worked the way I know how for 20 hours a day, it still wouldn't be enough, but like I can just get it done and I can go home and whatever.

28:28Or do I spend like two hours like playing around with this new tool that may not work and probably won't. And then I will have to go and do it the old way I knew how anyway. Um, and I'll be further behind and like everyone, unless you're someone with like a really, you know, curious early adopter mindset, everyone is going to do the first one and they're just going to get their work done. But the skills are changing really rapidly. And if you have invested the time and you are a little bit more familiar, your rate of progress and your rate of productivity is so much higher. You just need to be given the space to realize that.

29:01We do that as a concerted regular practice at Every, and it also, I think, works for other types of companies too. I'm curious, with something like Quora, again, that product is super cool. If that takes off, what happens? Or any of them, you know, Quora or any of them, what happens? Does Every like reorient itself around that product, to spin it out? Each product is its own separate business. Oh, okay. Technically. Technically. That's majority owned by every. How much we own is dependent on the circumstances. So they are already spun out. Yeah. Some of them, actually, not everyone. Some of them are still internal.

29:36It depends on like how big a potential there is. Like there's a lot of circumstances where it makes sense to keep it in. How many are there? There's four. Okay. But Quora is already a separate business. What we have found is if those products start to take off and some of them are, Like Quora is really, I mean, I worked on a lot of stuff and Quora is like the most promising thing I've ever seen. Wow. Yeah. That's really good. When people start using it, they just use it. Like it just, it's so sticky. They really love it. And it solves a big problem. It's very cool. The reason I say code is content is like they push so many more people into the every ecosystem.

30:08Like we're growing just in terms of audience size so much faster. So there's this really nice symbiotic relationship where people want to try Quora because they know every did it. and then Quora starts to grow and it becomes this like thing that's like has an audience in its own right but they're pushing people back into the every ecosystem because it's like made by every and then we have a you know we have an onboarding sequence it's like you should read every because a lot of people who are signing up for Quora are AI early adopter people so they kind of it's good for them to be in the ecosystem and Quora for example we can't let in everyone at once because we're scaling it and it's it's a hard product so like we have a 10 ,000 person wait list so So rather than let those people just like dry up on the waitlist, like they're now in the every ecosystem.

30:47They get emails from us. They kind of like they're in the whole thing. And then they start to get, you know, introduced to other products. And so it becomes really this whole ecosystem that I think is really valuable. And I think all of the like ARs and stuff recognize that. And everyone's sort of on board to make every this like sort of special place that I've been talking about. So if it gets big, I think that's amazing. And I think the bigger the bigger these things get, the probably the less reliant on every itself that they'll be. But they will prop up every at just as every props them up and make something that's bigger than either one individually.

31:21You said these things are their own entities. I mean, you could end up fundraising just for Quora or you could end up. I mean, yeah, like you can end up putting in a different office. And I'm guessing, you know, if you're telling me like it's super that one is super promising, it probably will grow to some extent. And I hope so kind of get spun out or I guess it's already spun out. What are some of the the other products you guys have built? So there's Sparkle. So we have Quora. We have Sparkle. Remind us again what Sparkle is. Sparkle is an AI file cleaner. OK. So it just looks at your files and on your desktop downloads and documents folders and then creates use AI to create an ontology of like here's.

31:57I just use that. My desktop is pretty cluttered. So my desktop always looks clean and I don't have to do anything, which is great. I love it. I love it. So that's really cool. And that's actually growing pretty fast. We have someone who's really talented doing that. We have Spiral, which is basically does marketing automations for you. So there's a lot of repetitive work in marketing and writing and anything you might do. So like for me, like I do a podcast every week, you do a podcast every week. Like there's a format to the tweet for that podcast. So like you can build a spiral that like you train it on examples and then it like forever will like get you to 80 % on the like.

32:34That's really cool. Yeah, it's really cool. I think it's really valuable. We've got a really talented EIR Danny working on that. And then the last one is Lex, which I think that's how we met. That's right. Yeah. When Lex was still inside of every. Oh, right. So Lex is a great example of a spin out. Yes, exactly. So Lex was built by my co-founder, Nathan. That Lex was launched like pre-ChatGPT. So it was like really early. It went super viral. And then we ended up spinning that out. That's his own business. Nathan runs that. raise money for it. Different businesses take different paths, basically.

33:04And we'll do a bunch more. We have a bunch more in the pipeline that I'm really excited about. So the people that build these sounds like it can happen any different way. But are you bringing in EIRs dedicated to pursuing specific products or is just like, hey, one of your writers might end up making a thing and then they're an EIR now? The answer is yes, both. So basically, like sometimes we'll make a product like I built the first version of spiral or sparkle and then we will go and like recruit any IR to run it and the recruiting is more like we like put a little thing in the newsletter and then we get a bunch of applicants and like some people stick you know yeah then sometimes we've what we've done is like we have a little bit more of like a program where we're like you don't have an idea but you kind of want an idea and you love every so like come hang out with us and build experiments, write about them.

33:54And then like we'll launch them to the audience. And if something works like great, it becomes a business. And we have like three or four EIRs that are like that. That's really cool. It almost reminds me of back in the like the early iterations of Betaworks back in the day, if you remember what that was like. I've definitely gotten a lot of Betaworks sort of comparisons, which I love. And I was not really around, like I was sort of in New York at that time, but like it Betaworks like had like it still has, but like really back then really had this like aura, which is really cool. Yeah. You know, they were the first check I raised for, I was so drawn to Betaworks.

34:30I, you know, we were building Anchor. I was like, I got to work with Betaworks. Yeah. Yeah. Yeah. Maybe you already do like sort of have that same sort of, oh, I got to go build something at every, I don't know what it is, but I want to go build something at every, right? And the podcast now, that's another media format or property you have tell us about that I mean I guess maybe it's kind of obvious but that is just it's all the stuff you're doing yeah now in a new format more or less yeah I mean it's called AI and I started as a podcast called how do you use chat GPT oh I remember that yeah and that was like sort of at the very beginning of chat GPT where everyone was like I know I should be using this but I don't know how and there were some people who had really figured it out so that was like the first I don't know six to nine months of the show, which was awesome.

35:10It was very timely. It was also at a time when video podcasts were really starting to take off and go viral. So that was very cool. What I found after the first six or nine months is like most people use ChachiPT in the same kinds of ways. And so there are still people who are doing crazy stuff, but like they're harder to find. What's the craziest thing you're seeing people do? I think the biggest, like most crazy stuff right now, and I've been doing this, but you know, it's sort of it's starting to happen is like doing notes and stuff with cursor it's like a really a really cool one i keep saying this what explain this i haven't tried it yet yeah this is where i started with every with super organizers is like i want to organize what i know that's like the core question for me and the and ai just like totally changes the game where you're not like you don't have to do these like backlinks and like whatever tags or like it just it solves every problem that we thought we wanted solved um in a much better way than than previous technology you could solve.

36:04At least where I started with this, there's a guy who wrote a blog post about this that originally inspired me that I had dinner with who I whose name I'm completely blanking on right now. So please excuse me, whoever you are. I think you're amazing. Poor guy. I talked to him about it. But basically, the way that it works for me is I take a bunch of notes. Some of them are like from my Apple notes or some of them are from like my export from Rome or Readwise or Evernote or whatever. And they're all in text in a folder that I then open Cursor. And Cursor has this like agent experience, right? Where you can be like, find me a note that whatever.

36:38And it's supposed to work for code, but it actually is quite good at just like searching through text files for particular things, which is kind of cool. But the really cool thing is if you want to do something more advanced and you're like, okay, I want you to gather all of the notes that I have that seem to be about, I don't know, the Civil War. and I want you to organize them by time and then put them all in a file so I can read them through. So it can't just do that in the agent, but the agent can build any kind of software you want. So you can be like, I want you to make a tool that is really good at searching through and finding bits of information for my notes that are like this and then outputting them to a text file.

37:16And it's like, cool, I just built it. And then you update your cursor rules and you're like, cool, here are your powers, here are your tools. And so the cursor rules and tell it like what it can do. And it becomes this like kind of That's really organizing thing. Do you have to recreate it for each use case or build on this one? You kind of build it on this one thing. That's really cool. Obviously, Chris and I are not built for it. I think someone's going to do it really well. There's a big opportunity for that. Another thing that I've been doing, which I think is really cool, and I will probably release this as a product.

37:45It's called Deep Background. And it's sort of like a play on deep research. Yeah, of course. So one of the things about deep research, I think deep research is just like the coolest AI tool that's been released in the last year. Yeah, it's insane. And one of the things that's really interesting about it is the responses get better if you give it better sources. Yeah. Which we've known for a long time. You get to direct it quite a bit. Yeah. If it's too open-ended, it doesn't really perform that well. I spent a lot of time, like, I don't know, a couple years ago talking about, I was calling it at the time, knowledge orchestration.

38:11But basically, like, the context you provide to a model governs a lot of whether it's good or not. It's not just about the raw intelligence. And deep research, like, it has a search tool. But a lot of the stuff it finds is, like, blog spam because that's just, like, how the internet works. Yeah, I noticed that too. Deep background allows you to provide your own context to deep research. So basically what it is, it's a web app. You can drop any file into it, and then it organizes it on a deep research-friendly web page. It's just plain text built for AI agents. And then you can kind of like say deep research, like I'm interested in this question here.

38:45Here's all the sources. And then you can put books. You can put like all your notes. You can put whatever you want. Why couldn't you just put that into the context window on deep research? I've found that it's not very good at, if you give it a bunch of files, like exploring all the files well, they may fix that at some point. It's much better at browsing through a bunch of webpages. And also it makes it more portable. And like, you can just have this like one repository of everything that you might want to give to an AI agent that you're constantly adding to. Like notebook LM meets deep research.

39:14Yeah, sort of like that. That's pretty cool. I think it's really cool. We might merge it with Sparkle. We're not sure yet. So you have it on your computer. But I think just generally, like the way that I think about it right now is it's sort of the about.me, but for AI agents where everyone's going to have their own like little knowledge repository that is like publicly accessible to AI agents that they like take around with them. One of the things I keep hearing a lot about is the opportunity for these MCP servers with Claude. That's one of the things I'm now thinking about for this tool. Like I'm going to do that this week.

39:47Basically, like MCP, it's called Model Context Protocol. It stands for Model Contact Protocol. And the idea is it is a way to build little APIs or website type things for agents. So if you're using an agent and you wanted to have access to a certain type of information, like let's say it's like GitHub repositories or maybe it's your deep background website, it provides a format for the agent to understand what information is available and what tools it can use to search through that information. it's sort of a layer between like whatever context you want to give the model and the model itself so that the model can like optimally access and find what it needs to find inside of the inside of the information and it's like a standard protocol that I think Anthropic created so there's a standard way for you know cursor to go access like all the files on a particular website or whatever got it yeah and so I guess the idea is people are going to start to build all these hooks into all these different things your email your notes your whatever your github repository.

40:47That's pretty cool. And they already are. Okay. Oh, they are. Where do people find these things? There's a bunch of like GitHub repositories that are like little directories right now. And then there's on Twitter and whatever. It's just starting, but it seems like it'll be a big deal. All right. So back to the podcast. So this was a way to explore the way people were using ChatGPT. And that's how we got on this tangent around using cursing for notes. Right. And then I got bored with that. And then I was like, I just want to talk to interesting people. So like every podcast starts with a very specific idea and that just becomes an everything.

41:20Yeah. I mean, you know, so now it's AI and I, we rebranded to AI and I, and, um, I'm really interested in how AI changes, um, how we work, like the I in AI and I is also about like how it changes like ourselves and what it means to be human and how we see reality and how we see ourselves and all that kind of stuff. Um, and so I tend to talk to a mix of like startup B type people and then i guess i'm like sort of starting to branch out a little bit more into like scientist philosophy creative type people who are working at that intersection because that's like the stuff that fascinates me yeah with the podcast do you imagine that you will maybe eventually take a similar approach as you've taken with the writing and have different hosts or different shows like does this part this become part of the collective yeah of every i mean it is it is already i think like we've we've had this is not our first podcast we've had a ton of different podcast.

42:10We used to have a podcast with Legion called means of creation. Oh yeah, that's right. We've had a couple other different ones. Um, me and Nathan used to have a podcast. There's just, there's a bunch of them. Well, we'll definitely do more. I think like the core is the writing. And then I think the big exciting thing is the software. And then I think we're also starting to branch out into video. Um, so I I've done some videos. We have a really good in-house beyond the podcast. You're saying like one-off type. Yeah. A lot of like, I got, I got this access to this new model. Like, let me tell you how it works.

42:37Yeah. That's really, and those go really viral it's really fun yeah i noticed that it seems like you do seem to get access to these things early you're almost you're almost becoming one of these ai influencers that's part of you know the influencer program yeah some new video model is that was that intentional definitely yeah okay for a lot of different reasons the core is just it's super fun i yeah like if you told me when i was 12 that i would be able to just like get access to new stuff before it came out i'd be like yes sign me up did you have 4.5 gbt 4.5 what are your thoughts on it i know there's like people are very divided over gpt 4.5 so we got access to it for a couple days before it came out um and the the take that we published initially was like this thing isn't isn't mind blowing yet but like it's very hard to evaluate models really quickly and so like everyone should just like take a deep breath and like play around with it and um right now it seems like it's sort of like claude sonnet 3 5e and that's what that's what um uh openai is saying but like we just We just don't know yet.

43:38And it's not one of those things where it just obviously knocks your socks off the first time you use it. Now it's about a week later and I'm like, oh, this model is great. I really like it. An easy one is it's funny and it's really hard to be funny. Yeah. These models, they really struggle with comedy. But this is good. Yeah, it's really good. It's the first model that made me laugh. I actually think it's the first model that like I could have it sort of swipe on hinge for me and it would do a good job. it being funny as a sign of intelligence because we're so used to the default picture of intelligence being something like, oh, it can do math well and or it can code really well.

44:15And it turns out that that was like sort of it's still hard, but it was like relatively like the lower hanging fruit and actually just making a good joke requires a lot of intelligence because you really have to know a lot about the world because a lot of humor is saying the exact opposite of what's true. Right. And so you have to have a good sense of like what's true and a good sense of being able to say it in a way that like comes off as a joke and is the opposite of what you actually mean. And that's that's a very subtle thing to be able to do. So it's funny in the sense that it could write a stand up routine.

44:47It could tell a joke. One of its best tricks is, you know, the like 4chan green text where it's like be me like Google exec or whatever. And then and then it will just like write out just like a funny narrative about like a person's day. Like I'd had to do that for me and I thought it was so funny. And I've seen a lot of other people do stuff like this. The thing that actually just totally changed my mind is I started getting it to do chain of thought reasoning. And once it started to do that, it was like incredibly good, which is interesting because chain of thought is like one of those things where it was like a big deal like a year ago.

45:19And then you started to feel like you didn't really have to use it, especially with reasoning models, which like just internalize the chain of thought in the in the reinforcement learning. some of the things i've been using it for are i had it write some poetry with with a chain of thought which was like incredibly good wow um and it made it wrote some lines that i was like oh man like i'm i feel touched um but i've also um had it help me like um edit writing where george saunders is a short story writer like one of the top short story writers in the world and um he has this editing technique you read one sentence one sentence at a time and you just feel whatever that sentence makes you feel and it's either going to be like positive like it's starting to make me like into this and there's something interesting happening or it's negative like i got pulled out and i'm bored or like i'm confused or whatever and all of writing um aside from the first draft is basically just like doing that process of one sentence at a time just like saying like how do i feel do i like this do i not like this and why and embellishing the stuff that you like um and and editing the stuff that you don't like and sort of like radically infusing your taste over and over and over again into every little thing that you do.

46:26Sounds like a model in a way. Like, yeah, it's like token. Yeah, there's something about that. There's some there's some overlap there. But like I had GPT for 4.5 do that with a couple of pieces of writing that I have where it's like reading each sentence and then outputting what it's thinking and feeling. Oh, wow. And it's really good. Wow. It's very helpful. That's really cool. Yeah. It feels like there's a product to be built there. There's a lot. Yeah. An editing product. Yeah. there's there's editing there i'm one of the other things i've been excited about recently because i fucked up my my wrist as as we've uh as we've discussed um we should discuss it off the show yeah yeah um i i was uh living in panama for the last like three and a half weeks again most interesting i had a little it's called a pocket bike so it's like a motorcycle but it's like small okay i was um going downhill on it and i just like braked a little bit too much with the front brake and i just like went over the handlebars and braced my fall and i don't think it's broken but it's like uh you know i sprained something or there's some tendon that's like if all you did was sprain your wrist getting thrown over the handlebars in the motorcycle i'm just glad you're here yeah me too i was wearing a helmet um my mom doesn't know about this so sorry mom um hi i'm fine um what's funny is like that was the day where i got access to gbg 4.5 and i had to write the article for the next day or something like that so i was i was like writing one-handed this this this review and and the thing that saved me is our speech to text is so good it is really good so I was using whisper flow a lot yeah and I would not be able to work right now because I can't type so I would have to be like so you just dictated it just yeah GBT 4.5 like edited and I dictated it what's profile did it I had a lot of I was there's a lot of back and forth with what you be 4.5 to like make it better I have a co-writer that I do a lot of the model releases with his name is Alex and he's he's fantastic and is such a big model nerd and also like really he is dramatically improved as a writer over the last like two months like dramatically more than I would expect having worked with a lot of young writers and it's entirely because of AI it's it's wild like he's made progress that I would expect to take at least a year in a month or two anyway so he he did some of it and then Kate who is our editor-in-chief who was previously the editor-in-chief or the publisher of Stripe Press.

48:46I just got on the phone. Kate Lee. Kate Lee, yeah. I just got on the phone with her and I was like, look, like, we need to get this out in like two hours. Like, I just need to tell you what it should say and like you write it and then I will like edit it a little bit. And she was like, I got you. Because just glad you're alive. Yeah. So it was a true team effort of humans and AI collaborating to get that article out. but it's been an interesting forcing function to like use all the text-to-speech stuff and now I have so many ideas for for this like it just feels and I've actually heard this I'm curious what you think I've heard this from a lot of parents that voice is the native interface for computers for their children yeah for a lot of different reasons and that I think is like really intriguing and makes me feel like it's important to interact with a lot of things using voice yeah Yeah, I feel the same way.

49:37I, you know, as somebody who built an audio company, I've always wanted voice to be like the primary input and, you know, output modalities or whatever. I'm hopeful, but like still somewhat skeptical because it's, I think it is the most natural, but it's just, at least for consuming, it's just not as fast. It's not listening. It's only dictating. Yeah, dictating. It's voice as the interface just to get your thoughts out. I use, you know, voice transcription or dictation on my phone all the time. And it's like, it's never perfect, which is frustrating. I think if it was perfect, yeah, I would use it constantly.

50:11And I do always find myself wanting it on my desktop. I'm always like, I want to just talk. Have you tried Whisperflow? A little bit. Yeah, I need to really commit to it. It does that pretty well. It's so funny. I usually write every Friday and I haven't been doing it this quarter because I'm writing this big, long piece that I think is turning into a book. Oh, wow. That's pretty cool. We'll see what happens. I got up this morning. We're going to record this morning. And I try to do a thousand words a day. and I just did my thousand words in like 30 minutes and it felt like cheating because I was just like talking and it was done and I was like, is that it?

50:41And just cleaned it up. Yeah. I mean, it's not final at all, but it's like, it's enough where I'm like, this is, this is good enough. You know, like I'll come back and, uh, and like really, really edit all this. Like once my hand is healed, but yeah, it's like when people said, uh, audio books or not read it, like it feels like it's going to be the same thing for writing. Yeah. Um, what about some of the other models? Like, how do you feel about let's say grok in my limited testing it seems good yeah i haven't really needed to use it that much i think that the interesting thing to me about grok is one of my very early days of ai i wrote or this generation of ai i wrote this piece called artificial unintelligence um and the basic idea is that the models are already smarter than we allow them to be like they're being made dumber on purpose by the model companies because it's risky for them to be too smart yeah the easy example is they're not going to tell you to make a nuclear bomb which is good they should not let you do that but there's all of these ways in which like we we're artificially constraining their capabilities already even in the gpt-3 days a place where there's going to be a competitive advantage is um companies that allow um their models to um return risky results which which seems like rock rock is like it's the whole idea exactly yeah ground truth and there's like you know there's you can interpret risky in like lots of different ways so like grok is like not safe for work things or whatever another form of risk is like um it will give you legal advice yeah um but so whatever whatever form of risk we're talking about but like grok definitely fits into that category and so i think that that's really interesting i think in a lot of ways open source models are are more fit for that yeah except they're far less accessible yeah for most for now yeah for now um i think at some point like more on device stuff is going to be a big deal i totally agree so i think grok's cool i mean claude yeah what about claude everyone inside every is like holy shit this is like yeah so useful um they're just going a lot faster especially people love claude code oh you think you think claude's going faster than the other yeah for code specifically for code yeah because i feel like some people feel like it's the opposite claude's moving slower but i guess for code like i feel like everyone loves claude for code one of the takes that Alex who's my co-writer some of these pieces one of the takes he's been playing around with that I think he may maybe post this week I don't know if it published yet I think it is is a lot of the you can see some of the model providers starting to pick their lanes a little bit and it's not just like totally general intelligence like actually Claude is like well we should really be good at coding because that's what everyone uses us for you know and I think that that makes a lot of sense and all the revenues and it's not all the most fits in the enterprise yeah right like so that makes sense Just to piggyback on that a little bit, I just think that that's so it's so clearly where things are going.

53:23And we had this idea of, you know, general intelligence is this like God thing that could just do anything. And one thing is becoming clear is that different models have distinct like sort of views on reality or different vantage points that they come from. And that is not something that is like easily you can't easily get rid of that. which is good at everything all the time um because the way models work is knowledge is distributed implicitly across like all of the neurons in the network and all the connections in the network and so each connection relates to thousands and thousands and thousands of things that um a model knows and so if you change one of them you're changing lots of different things in the network and so good things or things the models are good at tend to be like very highly interrelated um and so if you make claude really good at coding you're you end up unintentionally damaging it's like ability to tell jokes um and you can there are things you can do to like kind of make it better at jokes but then there's something else that changes right and so and i love that i think that's a really beautiful and interesting result and it's also you know you can see that in people um everyone that you meet is a general intelligence and people look very different and that's actually a good thing pretty clearly like everyone has their own arsenal of tools or models that have different viewpoints on reality that they use for different things and i think that's definitely the future what is gpt's lane is like just yeah consumer and enterprise basically most of their they're the inverse right all their revenues for most of its consumer a lot of it but i still think like if you go talk to businesses like yeah no one's heard of clod everyone knows open i think so yeah everyone's like what is that um open ai is the standard um And I think that they're just hitting the gas on that and they're becoming an application company.

55:08They still care a lot about research, but like the application layer is like where they're going to compete and win. And they have the distribution and the kind of like name brand advantage. And I think they're going to be that. To be honest, I don't know like Anthropics revenue figures off the top of my head. And I think they're amazing people and their model is amazing, but they're very, very far, far behind even in the enterprise. the lesson that opening I learned from chat gbt is that the power of the model doesn't have that much to do with how well it's adopted because it was the same model but when they wrapped it in chat gbt it took off and I think you can see that in all the rest of their product releases so like their computer use um operator it's an app right anthropic released an api they released an app right um I think you can see it with Sora it's not an api it's a it's a product it's a product dedicated product yep so that's the lane that they're picking is like mass adoption consumer and enterprise products yeah and i think that's i think that's valuable i think there's lots of other places in the stack to play it seems like claude is picking the code path i think that's super valuable too um or anthropic is yeah they have to figure out something with the model model names though because i'm i'm getting really confused personally it's very bad yeah but i saw somebody maybe they're in some testing group or whatever uh somebody on twitter yesterday is getting like the automatic switcher action yeah that makes sense i don't know if you have that i haven't seen i haven't got that yet um that makes a ton of sense i think you know i think especially as more and more people adopt this stuff like they're not going to want to pick them they're just going to want to do the thing right and and the model will get to some degree the models are already like that like yeah they're a mixture of experts models so like they have different models inside of them more or less that that they route to automatically but then there's there's layers of this um so yeah i think the automatic routing makes a lot of sense maybe GPT-4-5 can help them figure out how to name these models.

56:55You know, you're playing with all these models. Obviously, you're also playing with products. What products are exciting you? Obviously, Cursor. You're spending so much of your time in Cursor. What else is interesting to you? Cursor, WindSurf. It's really cool. Obviously, Granola. Granola. Yep. Thank you for that. All the credit is due to the team, Chris and Sam. I'm just very, very lucky to be a tiny part of the journey. Yeah. It's a great product. I use it as well. I'm biased, I use it every day. Yeah Obviously all the internal products we have those are really the main ones. I think people ask me that question a lot and I Really think a lot of the you know, the clods and the Chachi bichis of the world are so still so untapped Mm-hmm.

57:39I think there will be a lot of really great AI products that are standalone and not part of that ecosystem But yeah, I think of them as being this generations um excel where excel is very easy to start with um anyone can type into it yeah into a cell but you can build mario kart in excel you know it's it's endlessly powerful and i think chat gpt and claude are like that easy to start with but there's there's so much we're just scratching the surface of what you can do with just chat models and um it will take some time before Or people have built up these really specialized power workflows that they are like, I know how to use this and I can do it in cloud, but it's like hard.

58:24I would like a special tool. Same thing for Excel. Like Excel taught a generation of people how to use computers. And eventually they made powerful workflows that they were like, it'd be great if I had another app for this. And then B2B SaaS became a huge thing. I think the same thing will happen to AI. You do. Yeah. So you don't you don't think because a lot of people think that chat, GBT, Anthropic, like the models ultimately going to take over these individual sort of verticalized workflows as well. You still think there's a huge opportunity for companies, startups, people to build dedicated products.

58:53If I had to pick what OpenAI's product strategy is going to be, I'd like probably pick Salesforce. Salesforce. Salesforce has like Salesforce Marketing Cloud and like CRM and like all the sales cloud, like whatever. that every vertical they have every like i don't even know if every business sector or area they have a product you know so i think they're definitely going to it's not just going to be like chai gpt like they i reports are they have a sales product sales automation product you know um i imagine they will do a customer service product eventually but they won't do everything i don't they definitely won't do everything and and there will be a lot of room for specialized players to to to do amazing stuff what about this whole theory that um you know when the model gets good enough, you know, we have AGI or super intelligence or whatever products will just get spun up dynamically for all these things.

59:40So business needs, you know, I don't know, business needs note transcription, like you just spit out scrawl, you know? Yes. And I think that's not going to happen. There's two things about this, that vision of the world that I think people are missing. One is the thing I brought up earlier, which is this sort of like bundling of capabilities. So as you get good at one particular thing, you get less good at other things. Reality is very complicated. It's very high dimensional. You get one get good at one part of reality. You start to get less good at others. It's just sort of somehow the way things work because of the way that these networks are built, which parallels a lot of the things that happen in our brains.

1:00:13That's one reason. Another reason is something we talked about earlier, which is the power of the model is in large part governed by what context you can give to it. Being in a particular area of the world where you know a lot about the world allows you to prompt the model in ways that like a generalized player they could do it but it would it like it's not just about raw intelligence you have to actually go into that part of the world and get the data yeah so there's like this whole empirical aspect of intelligence which is like yeah you need a lot of experience to like actually in a particular part of the world lots of worlds have the world has a lot of these like little niches with like um things that you can't just like learn by like going and having a conversation you need to like spend years like yeah there's lots of long feedback loops that it's it's incredibly complicated and so i think that there's a lot of room for people to sort of specialize even in where we have agi and and the a way to think about it is okay you're starting a like a waste management company i don't know i just picked it i just finished re-watching the sopranos to me that's why i'm thinking about this amazing amazing show it's about that it's about that time again time for a re-watch and who do you want to hire for that um people who are in waste management i'm just you know using my own experience for other for other things like they have a certain like flavor for them to them or like way of seeing the world um people who work at every have a certain flavor like way of seeing the world and you want to find people that are like that or you want to find young people who can train to see the world in that way yeah and i think models are going to be quite similar i think people forget about that when they think about AGI because they like we actually have very little intuition for what an AGI is going to be like.

1:01:53So we imagine it to be like this all powerful thing. And if you imagine all powerful thing, there's a lot of conclusions you come to that are wrong. If you imagine it to be like just like a person, which is sort of what you're trying to do anyway, which is like, yeah, it can just learn anything that it can that a person can know. There's still lots of differences between different people. Yeah. And even someone who's a genius at one thing like einstein like i still wouldn't want him to like run an app run a be a gm for a product at every like he would suck um so um so i think maybe yeah maybe i you know no you know no shame to einstein uh but i really think that's that's the case and that's that's where we're heading and and i think that's really cool i think there's a a huge emphasis on um intelligence as uh sort of abstract reasoning, which is definitely part of intelligence.

1:02:44But there's another part of it, which is, um, uh, built up through experiences in the world, um, that I think we're underweighting in our thinking about this stuff that will become much more, much more apparent and much more, uh, critical to our thinking over the next year or two. So the future is bright for startups and app builders in in ai the models aren't just going to take everyone's jobs and take away all the opportunities i mean i do think i think there are important social conversations to have i think jobs will change and some jobs will be eliminated but uh definitely the doomsday scenario is definitely not happening that sounds like a perfectly optimistic place to leave it dan thank you so much this has been awesome i can't wait to do it again thanks for having me yeah thanks so much for listening to generative now if you liked what you heard please rate and review the show on Spotify, Apple Podcasts, and YouTube.

1:03:36And of course, subscribe. All that stuff really, really does help. And if you want to learn more, follow Lightspeed at Lightspeed VP on X, YouTube, or LinkedIn. Generative Now is produced by Lightspeed in partnership with Pod People. I am Michael McNano, and we will be back next week. See you then.