#59 Gabor Cselle — He Sold Two Startups. He's Launching Another Two.

All right, let's jump into it.

Sweet.

So you've sold a startup to Google and Twitter. How did you get into startups in the first place?

Okay, so my story is actually that I grew up in Germany. I was born in Hungary and I grew up in Germany after we left sort of communist Hungary in 1989. I moved to West Berlin and then I wanted to go to a great school for CS because I read like the Bill Gates biography and I was like, I want to do startups just like Bill Gates. This was like 1999 or so. And so I went to ETH Zurich, which is in Switzerland. And then Switzerland is like the loveliest place on earth. It's like literally paradise, but I really wanted to do startups. So in 2007, I moved to SF and I joined a startup called Xobni, which was inbox backwards. And that was kind of like an instant hit when we launched that, a million downloads in 2 weeks kind of thing. And I was like, okay, now everything that I touch is going to be just like this. And it wasn't because I started my own startup, which was a YC startup called Re:mail. And it didn't go to a million downloads in 2 weeks. It was like a almost 2-year slog to get something working. And in the end, we sold it as a talent acquisition to Google. And then I was at Google working on Inbox. And I was there for 2 and a half years, worked on Inbox and Android, and then left again because I was like, I have one more startup in me. Started Well, took some time to come up with an idea, and then started Nemo Media, which is like an, like right time, right team, right execution, right everything, like just like amazing. And it was an ad tech solution for apps with streams to monetize better. And we sold it to Twitter like a year and a little bit in for an alleged $50 million. I'm not supposed to comment on the number, but here we are. Then I was at, I Twitter, went back to Google to do internal incubation at Google, which was really fun. And then started an ill-fated social network called T2 after Elon took over Twitter. It didn't really succeed. But, but here I am kind of launching things on Product Hunt many, many years later.

Still launching on Product Hunt. I love it. We have this, you know, philosophy, I talk a lot with Rajiv, Product Hunt CEO, for those listening, about the idea, uh, you know, startups don't just launch once, they launch for life. You see some of the best startups, Notion, Figma, Linear launch again and again. And it's the same with people, you know, whatever they're doing, they end up launching on Product Hunt again and again. Um, I'm curious, uh, yeah, selling to Twitter. What was that like? What was Twitter like then? What, what year was that exactly?

Yeah, so it's 2014. When— if you want to sell a company to a larger company, I think the best time to do it is just after the IPO, because Twitter had IPO'd in 2013. And then just after you IPO, you have this like new currency of shares that you can just issue and get essentially free money that you can spend on new acquisitions. And that's exactly what had what was going on. And Twitter, if you remember, I think, bought like 12, maybe even more companies in 2014, because there was a fresh currency to acquire new things with. And honestly, I think, you know, much credit to our investors, we kind of were— I was kind of lucked into this meeting, where Andreessen Horowitz was one of the investors in, in Naimo Media. And they had a policy that you could get invited as one of their like larger Series A investments, you can get invited to these industry roundtables. But if you are a seed funded company, you can only be sort of like the slot-in replacement. So one day, like I literally get an email from Pernille Andreesen, and she says, we have this thing tomorrow at 5 PM. And it's going to be all the ad tech experts from all these like large companies. And so, of course, I said yes, because someone else had dropped out. And so I'm sitting there with like all of the potential acquirers for our company that in this room at the, I think it's called the Rosewood in on Sand Hill Road. And it was literally like myself, another startup, and all of our potential acquirers. And so I spent the next 24 hours before this event actually started. Coming up with like just a couple of smart things to say. And by the end of this, like, you know, dinner kind of buffet thing for that lasted about 2 hours. I ran out of all my smart things, but I didn't get any more questions. So that was great. So suddenly we were like very much on the radar of these larger companies as one of the players in this emerging field of native ads. And so we just kind of lucked into this and that started actually all of our acquisition conversations.

Wow, that's amazing. I love the preparation beforehand. I think it's underrated. Just come up with some smart things to say. You know, it's so good before big meetings and stuff. I'm curious, Twitter ads, they get a lot of shit on the internet, on Twitter. You know, some people are bullish on it. Maybe they're underpriced. Other people think that they're a waste of money. Being inside of Twitter, like, Do you know kind of why Twitter ads never got to the Meta level of optimization or popularity?

I think Twitter spent a tremendous amount of time and money rebuilding or building the ad stack that was kind of modeled after Facebook's ad stack into 2013 era. And it was actually not bad. Like it had all the bells and whistles. It had the features, it had the targeting options. Facebook has billions of users and Twitter at its peak had something like $350 million MAU. And I don't know, obviously, it's now a private company. So we don't know what the numbers are, but the audience is much smaller. And so you can target and optimize as much as you want. But if your audience is much smaller, you're just going to be not as big of a potential avenue that advertisers will consider. Twitter, I think, was profitable for one quarter of its existence as a public company. So it does actually work. But I think one of the big problems that Twitter had, and they got a lot of shit, And, you know, Jack and Dick get a lot of shit for like the, the, the always struggling with profitability, but it was kind of a hard thing to do with a much smaller network that was not growing as fast and was overtaken by Instagram, I think roughly in 2015 or 2016. And two, I think Twitter overhired and, you know, Elon solved that quote unquote, I think pretty inelegantly, but Twitter, part of the reason why Twitter wasn't profitable is because there were just too many people. People doing too few things. And it's easy in these large corporations to want to hire larger teams because you get promoted if you hire more people that you then manage, because then you can be a senior director instead of director. So all the incentives are kind of badly set up. And that's why Twitter was never profitable. I think there was a path to profitability, but it would have involved a lot more discipline around spending.

Yeah, that makes sense. It seemed definitely bloated and seems like X, you know, formerly known as Twitter, is still running. It's working.

That's right.

You know, some hiccups, but it's working. So let's jump into Choosey Chat. Tell me and tell the listeners about Choosey Chat and why should we be able to choose our LLMs? And I guess the question is, why would somebody want to choose different LLMs? Like, aren't they all the same?

They're actually not. They have pretty different performance and they keep getting better all of the time, but they still don't necessarily output what you want. So the idea is this is Truzy Chat. I launched it on Product Hunt on Friday after I spammed all of my friends. It got to number 5 and that's when it petered out. We couldn't get any better.

It was a competitive day. That was— Oh my God. I was, no, I don't think anybody was expecting Wordware went insanely viral. Like that was 4,300 upvotes. That does not happen very often. That— they broke the internet. Yeah, I don't know how that happened on a Friday, like a random Friday. That was crazy in the middle of—

yeah, I mean, I don't have any expectation that the first launch of anything will make it to number 1 or even number 2. So my goal was just to get above the ad.

Nice.

That was my ambition, and we like kind of scraped by. Most of the people that I spammed are on West Coast time, and Product Hunt starts the leaderboard at midnight Pacific time. So you kind of got to like, we lost a lot of momentum early because everyone else got higher ranked. And so we kind of like scraped by at number 5.

Yeah.

Okay. Choosey Chat. Choosey Chat is a very, very simple idea. It is literally, we take the best 3 LLMs, which are currently Gemini, GPT-4o, and Claude 3.5 Sonnet by Anthropic. And we run them in against each other in parallel. So you ask your question and you see the 3 LLMs responding and streaming in the responses. And then after that, our mascot, Choosey Cat, chooses the best answer. And Choosey Cat is a critic model. It's based actually on a UC Berkeley paper that I took a little bit further than they had gone in comparing LLM responses. And I made it, I transfigured it into a cat That's like with a French scarf. So it's Choosy Cat and cat in French is chat. So that's, we got chat. So that was the whole, whole thing. I like sort of like playing with these like branding element type things. And yeah, I mean, we got like 590 users, we got 300-ish upvotes. Um, and it was really centered on this, like choosing the best answer, thus Choosy Chat. And I think that worked relatively well. Why that space? I think with T2, which was more or less a straightforward clone of Twitter, I learned the hard lesson that you need to not only build something that's like a little bit familiar, but also build a new angle on it. So I'm, I think that new angles can be most easily found with new technology and LLMs are kind of very much still developing actually. And so that's why I went for the space. I've also tried this year to build and launch stuff in GenAI video, uh, which is a whole nother story, but I, I want to really get to a place where I'm working on stuff where new models are dropping every single day and I can get to the forefront of it.

I love it. And Choosey Chat allows you to do that. I'm sure you'll be integrating more models or maybe replacing. I'm curious, do you have stats on which models win most of the time or what kind of the percentages are?

Yeah, so I think the LM-Sys benchmarks, lmsys.org, are probably the best ones. They essentially do pairwise comparison and what they call the chatbot arena. And currently the best one, literally this changed on Thursday, is the Gemini model that just dropped. It's sort of like the Gemini 1.5 model, but a newer checkpoint of it. And that one is currently the best one. And then number 2 is GPT-4o. Number 3 is GPT-4o Mini, though I'm surprised by that. We can get into that if you want. And then Claude. So that's why I've chosen those 3 models for Choosey Chat. And the other thing is like I don't think that people care about like comparing the number 10 ranked model for their actual questions. If it's a human spending their human time on wanting the answer to a question, you only want the best models. You don't care about the cheapest models because your time is the scarce resource, not whether you're saving a couple cents on inference.

Is that the reason why you're not using Llama out of curiosity, or where does that fall into it?

Yeah, Llama just doesn't beat these 3 on the benchmarks.

That makes sense. I mean, there's so much like hype around it. I wasn't sure about that. I've been really loving Claude. I love using Claude and Sonnet and Artifact is so cool. As somebody who's, you know, I've coded a bit, but not super technical, but I've been able to build these little like prototypes and stuff. And I love how shareable it is. I find that really cool. If you're listening to this and you haven't tried Claude, you're living under a rock, go try it. It's so cool.

Yeah. Yeah.

I, I was thinking—

and well productized and kind of this lovable, like hairball character. I think that's, I think people really respond to that.

Yeah. I think the branding was done very well. Very delightful, classy. I wanna say the word like almost, you know, Claude makes it feel like just classy. The entire thing, the colors, the muted colors, the tan, the orange. Even the billboards aren't obnoxious. They're kind of nice. Like they're kind of artistic. I think it was just done very thoughtfully, uh, overall. Um, very impressed. Um, something I was thinking about is, and I wanted to talk to you about, was on Claude, on ChatGPT, on some other stuff, Wordware, I was trying out that the other day. When you type in something, you could see almost it, it thinking as it goes through. So like, you know, Claude, it, it starts typing all that stuff. Wordware is a series of workflows, but it shows it creating. And I personally, I'm really loving that because it's a little bit different than just saying waiting for 30 seconds, right? It's so much nicer to see it going than seeing, you know, the ball of doom or, you know, 10%. I don't know. Have you, have you thought about that? Is, do you think that's like a conscious decision or on like the, the builder's parts to show that process? Because they could have easily just said loading, like they could have done that. And that's what we've done for years, but.

Okay, and this is like, how nerdy do you want to get on this one?

As nerdy as possible, you know.

All right, so these are— okay, so all of the LLMs today are what's called autoregressive models. So what they do is they are fed the text that you want to have an answer to up to the point where they've completed, and they're fed that and they're basically asked, what is the next character? And then they produce logits for the next character and choose the sort of like best one. And then they are re-fed the entire text that was up to there, plus the new character or the new token, actually, which tends to be multiple characters. And thus they are autoregressive because they are— they get fed their own output. And this has a number of both consequences, like the fact that it kind of feels like slow and it reveals itself over time, which is actually kind of interesting to watch. And every single token that gets added, you kind of get a little dopamine hit of, let's look at that and let's see what it's just typed. So I think that's good. The negative pieces is that these models have something called drift. So if early on they choose kind of the wrong direction of tokens, they can end up with an answer that doesn't make sense or it goes, it goes a different way than you would have expected, or they hallucinate or they produce incorrect results. When in fact the model has all the information to produce the correct results. So there's, there's some— this is a Jan LeCun reference that he calls this drift, and drift is a real problem. There's actually like really hot off the press new models that Stanford is working on that are called text diffusion models. The specific one is called SEDD, and they use diffusion to generate text. So it's like Stable Diffusion, which You give it a text prompt and it generates an image. And the way that, that these diffusion models work very simply is they're given sort of a fixed amount of noise. So, you know, I don't know, like 1600 by 900 pixels. And then there, the models are asked, okay, is this a picture of a cat? And it's clearly not a picture of the cat. So which direction do we need to wiggle the pixels in order to make this more like a a picture of a cat. That's called gradient descent. And then there's a bunch of math, uh, circus circuitry around it. And essentially, if you apply this diffusion a lot of different times, eventually a picture of a cat will emerge. So text diffusion, this new model, SEDD, is the same thing applied to text. So you feed it an arbitrary amount of noise, essentially, and then you say, I want the code for the backend of Choosey Chat. 'Generate me that,' and it will, you know, essentially diffuse that random noise into code. And so this is far faster inference, but you don't get that blinking cursor thing that keeps moving. So this is like a very exciting thing. This paper just dropped, just won the best paper award at a very prestigious AI conference. So I think that's a space to watch for sure.

That's very cool. I'm excited that For Choosey Chat, do you think you'll also integrate images like Midjourney versus DALL-E or stuff like that?

So it does support image, um, feeding in images and image understanding. It does support that. Uh, just added that last week before the launch. It doesn't support image generation yet. Um, I think actually the best productization of image generation currently is the new thing that GPT added. That ChatGPT console added, which kind of takes you into an image edit and generation mode where you can click on the image and you, you pause the main chat and you go into a chat around that image that was just generated. I think that would be something pretty awesome. I'm not totally sure how to, how to generalize that with multiple models, multiple image generation models that you would be chatting with, but maybe I just need to think more about it.

Yeah, because when I'm creating images for my newsletter covers, I bounce back and forth for a while between ChatGPT and DALL·E and Midjourney, and now I'm back to full Midjourney. But it's nice kind of seeing the comparison every now and then just to see what it's like.

Do you have like cinematic images for your newsletter?

Yeah, yeah. So I'll, I'll flash them on the screen for listeners, but I do this whole, you know, cyberpunk style. And at first, you know, it was just like very basic early DALL·E. Like when that came out, I was blown my mind and now it's gotten like deeper and deeper and I'll input photos of myself, add filters, combine them with other things. I'll take past images and sometimes, you know, it's, it's getting to the point where I could even create videos, right? With stuff like Runway as, as these covers or create GIFs. And I've experimented with that a bit. It's a little time consuming to do just for a newsletter cover. But I think it's really fun. I personally, I treat my newsletter covers like MrBeast treats his YouTube thumbnails. Like I, I put a lot of, lot of love and love and time in them. And it's, it's a lot of fun. I, I really yeah, using Midjourney is quite fun. Something I was actually thinking about on this topic yesterday was that the time waiting for, you know, things to load like GenAI tools, Midjourney or ChatGPT. It's a lot of actually just waiting, you know, and maybe, and maybe it's, it's because things are slow right now, but I think like it's very exciting when it gets, the hit comes, but there's a lot of just waiting and like chilling. Like I just watch YouTube videos or listen to podcasts to not go fricking insane, you know, like it's, it's really fun. Is that gonna like speed up? Is it gonna be the speed of light kind of thing or are we always gonna, you know, need some entertainment when we're using our gen AI tools.

Yeah. So I think there's two problems. One is you don't know if— one, you have to wait for a very long time. Essentially inference is slow. And the second one is you don't know if what it outputs is gonna be good, especially the video models. It's like, I don't know if you posted that meme, but that meme of all the gamblers in the casino at the, uh, you know, one-eyed bandits just keeping pulling. That's kind of what video generation feeling Yeah, it does. These things have like enormous drift as well, where they kind of like take the video in a certain direction that you don't really want and it's too late to stop them because you just used your runway credits on generating that video and you're like, oh, that wasn't the way I wanted it. And then you like fiddle with the prompts and, and oh man. Yeah. So this is like whole, so very frustrating situation right now, I think, waiting for these models. All of the stuff that you see on Twitter is like, look at my cool new model. And then it's, you then see that the timestamp is accelerated by 8x in order to get that Twitter video to be really cool looking. So when you go to the actual product, it just takes forever. Okay. Inference is going to get better. Either it's like a Groq type solution where they just make custom chips just for inference and maybe just for LLMs.

Can you explain a little more about Groq? What do you mean by Groq type?

Do you mean— Oh yeah, so Groq, G-R-O-Q, is a startup. They literally just raised $680 million. And yeah, exactly. But they— what their chips do very specifically is they accelerate inference for LLMs. So it is like all of that, like, you know, encoding the input, using the transformer stuff, applying that to it, outputting the logits for the next character. They're just super, super fast at that. That's called inference. Generating the next token is inference for the next token. And they just do that extremely fast by having the hardware built specifically for that. So I think that is a legitimate approach to solving this issue. And I think we're going to see a lot more of Groq in the next couple of months to years. I think one of the issues that we're facing right now is like no one has enough chips or like even the big player, it doesn't have that, like, OpenAI is sometimes slow because they're just like overloaded. So, uh, Uncle Jensen Huang is going to keep churning out more H100s and Blackwell hopefully is eventually coming soon. And we're going to see availability of compute being higher. And so our experience will be faster, but this is going to persist for a long time. I think like for all practical purposes right now, the demand for compute for GPU compute is basically infinite. So, uh, it's gonna be a long time to where we get saturation in those markets. And then there's a whole nother like planetary scale problem of how do we get enough power while also not burning the earth down? Uh, which is a whole different issue. So there's like a bunch of correlated issues here.

Yeah. Talk to me about that. I remember during the, during the crypto hype waves, there was a lot of climate change, uh, you know, arguments, uh, or, or, you know, beef being thrown or shade being thrown. Uh, I haven't seen that around AI too much. I feel like maybe it's just because it's less financialized in the sense that random people aren't losing their life savings on JPEGs. But I, I don't see people really attacking AI for, you know, environmental damages right now.

Yeah. If you go to Hugging Face and you scroll through a couple of the models, especially the one, the ones that are corporate models, typically they'll have an environmental disclosure of this was trained with this kind of power and this much CO2 was emitted in training this model. But if you read those disclosures, like the CO2 emitted is non-zero and like far non-zero, but you don't typically think about it because you're just sitting in front of your computer waiting for it to spit out the next token. So Yeah, I think the compute needs and the power needs of AI are pretty dramatic. Can't quantify myself because I'm not very deep, deep into the space. But we are going to be at a point where either we need to dramatically make these models more efficient, and maybe it's Groq-type hardware, or we need to just figure out how to produce fusion networks or just put solar on every rooftop because the the compute power needs of these things are really, really high.

Makes sense. Um, all right, let's jump into automations. We were on the phone last week talking about automations. Um, I, I kind of want to bring up a few that I'm thinking about and maybe we could talk about a couple as well. So, you know, automation number one that I, I just set up, excited about. So I just brought on a podcast editor to help out with this. And so when people schedule a call, you know, schedule a podcast with me, right? I just set up a Zap from Calendly over to Slack, right? Very simple thing. But now, you know, my podcast editor is alerted, right? Super nice. He's able to plan ahead, know what's on the calendar, able to see all that. Another cool one that I'm thinking about, and this is actually related to what you're working on with video. So when a product hits 100 upvotes, Or maybe when a product wins, say it's like 12 AM the next day and a product wins, I want a video to be generated.

Ooh.

Right. Yeah. Like a faceless video or, or maybe a HeyGen avatar of me, like an AI avatar, or maybe like of the product kitty talking about that product and saying, you know, what is this product? And like, why was this product of the day? Right. And I feel like, you know, I know the team, our team at Product Hunt, we're working on the Zap and stuff. But I feel like that's going to be doable. Like there's just enough, you know, APIs and Zaps where I could just have that as soon as the day ends, product of the day video, it's sent everywhere on Instagram, TikTok, YouTube, every single day there's a new product video. Wouldn't that be cool?

That would 100% be cool. And I think it's going to be doable in the next couple of months. Um, we're already seeing, so I think what you need for all this to work is I think you have the Zap infrastructure once it hits 100 votes. So the trigger is there, but the production of the video Right now, you can go to HeyGen and type in what you want your AI Jason Levin to say, but that still takes manual labor to do. I'm seeing APIs come up from a bunch of different players that would automate this. There's a class of startups, some of them are launching. There was just a YC startup, I can't remember their name, but maybe you can add it to the show notes once I send you that, that just launched that basically is like, Auto grow your channel with AI content. And so they have 2 minutes of training video and they'll just like figure out new stuff for your AI Jason Levin to say on your TikTok channel, and that will auto grow your channel. So I think video generation automation is coming. There's a ton of like issues around, you know, what is real? What do you disclose? How much do you disclose? Will people actually like watch this? Should they know that you are an AI? So there's a ton of stuff around, around that. Um, my personal experience and having experimented with this a little bit is that TikTok especially immediately downranks AI-generated content. Like they can tell it's, it's, yeah, they can tell and it's basically instant. I think I did disclose on all of my AI-generated content. There's now like a little switch, uh, where you can say this is AI-generated. But the moment you flip that, you're just not going to get that many views on this stuff. So they do prefer authentic content.

I have a friend who's running a business. It's called Faceless.video and you just kind of sign up. It takes a subreddit and tells stories about it and puts a video out every single day. They're doing great, doing pretty high MRR and killing it. So I think—

He just launched on Product Hunt, right?

Um, I know they launched recently a few months ago and now they might be launching again. Yeah, it might be them, faceless.video. They're killing it. And I think, yeah, the next step is automated versions of ourselves. So I was playing around with HeyGen trying to create AI versions of Jason Levin, right? And my voice just sounds so flat. It is, it's so bad. Like everything looks right human-wise. Um, but the voice is just really bad. I do wonder also, like, yeah, it's, it's, uh, that's, that's the problem is the pitch and then making a script that sounds like a human and not boring. Right. And I don't know if AI can do that or, or what, cuz I don't use AI for writing at all. Any of my Twitter, any of my, you know, ghostwriting, any of my content writing because it's just so obviously AI. And so I feel like even with speaking like a human, that's gonna be worse cuz you could see the person.

Okay. So copy generation, you gotta automate that. And the way you automate it, it's, you don't publish what it writes, but it gives you like a starting point. So for, yeah. Uh, so like Choosy Chat, I set up my own GPT that's called like Choosy Chat Marketer is what I called it. Uh, only I have access to it and I fed it like literally the about page, the pricing page. The stuff that I wrote on Product Hunt about the launch. I also put some stuff on Reddit and stuff and other channels that are not Product Hunt that maybe shouldn't be named, but Hacker News, for example, they all tanked. Like Product Hunt is by far the number one best channel, hands down. So I wrote Choosy Chatmarketer, I fed it a bunch of stuff and now all I need to do is like, I just went in before this call and I said, hey, tomorrow I wanna launch the new Google model that just dropped. It has these characteristics, it's better in these ways, write me an email to all the users. And it just wrote me a pretty good email that I had to like slightly tweak because ChatGPT for some reason was like, it's always so joyful and exuberant.

Maybe that's why I can't use it. Cause I'm just a sarcastic asshole. Yeah.

Yeah. Yeah. Exactly.

Like I'm not, you know, like I, it's just too, it's too nice for me. Yeah.

Yeah. So I always type in like, you need to tone it down. And then it turned out—

Chill ChatGPT.

Exactly.

Have you, have you checked out Glyph? Glyph.app?

No. Is it kind of like a Jasper?

No, it's, it's totally different. So it, uh, it's the first AI meme app that's actually really funny. And it is just, you know, like unhinged funny. Um, it's, it's crazy. Like Andreessen was posting some of them and a bunch of big tech people because it was the first time where it was like, oh my God, like AI isn't being PC. And I was, I was really excited about that. Like I've been playing around with it a lot. I've been using it a lot and it's cool because you could go into Glyph and actually, you know, type in the prompt and, and say like, you know, be as unhinged as possible and it, it'll be pretty unhinged. I'm loving it. I think that's been pretty cool. Yeah.

Memefying things is an underappreciated aspect of LLMs. I feel like what percentage of memes do you think are already LLM-generated or AI-generated on Twitter?

I don't think too many. I think, I think a small percent, honestly, because there's just so many idiots on the internet doing memeable stuff already that like, I don't think it's too much, but, but a non-zero percent because a lot of people make funny AI art that gets turned into a meme. But, but I'd say a non-zero percent. Um, yeah, yeah.

I, I, um, I use Tweet Hunter for automating quote unquote my, my tweets. Uh, yeah, it's not good. And it has its own AI compose, but it's too nice.

Exactly. Exactly. It's, it's too nice. It's too like upbeat. But I think what you were saying earlier, which, um, is like AI can give you a starting point, right? It can give you ideas. Um, and then you go in and edit. I think the problem is, oh man, nothing pisses me off more than the AI reply guys. My God, it is just such a great idea, Jason Levin.

I'm so conflicted about it.

Oh man, I hate it so much. It's just like reply for yourself, man. I have a, I have a friend in New York and I think, I know he's using one, but it's awkward because I can't ask him if he's using one. Cause if he says no, he's going to be insulted by about his writing. But I like, I think it is, it's like, it's like asking a girl if they're pregnant. Like asking somebody if their writing is AI is like, like, that's a little, I don't know. I'm a little worried about asking, but I gotta tell 'em to stop cuz it's so bad.

We had this problem with T2 a little bit where, so we didn't have GPTs didn't exist back then, but this was our Twitter clone alternative. The main, what we tried to do was have trust and safety built in from the start. Turns out that isn't the feature that people will come for in droves. That's more like a feature that is really relevant once things start going south for users and not until then. And we just never had enough usage on the thing until for that kind of stuff to happen. So with T2, actually, we came up with a bunch of interesting things to do with LLMs. Like we would take people's tweets. Once you're the platform, you can do way more fun stuff. We would take people's posts as we called them. And we would write a poem out of them and kind of trying to match their voice. And we would put that in our weekly reactivation email. And people like love that. People love things where you take what they've already created and you, you take it in a new creative direction that they wouldn't have otherwise explored. And a lot of that you can do in, um, in a— if you're the platform yourself. But we, we all— we eventually got backlash of like, okay, this is cute, but it's not actually that useful for me, and please stop. So you can only take, uh, AI-generated content, especially in marketing, so far before people start complaining.

Yeah, I think the best AI reply guys are when you reply something goofy though. Like if there was an AI thing that replied, like it turned my tweets into very, very funny unhinged memes and replied to me, I'd be okay with that. Like maybe that's, maybe that's a business idea in itself.

It's the Glyph.app integration into Tweet Hunter.

Exactly, exactly, exactly. And so like I think that would be nice. If I, if I've done that to, you know, big, you know, Twitterati people, right? I'll, I'll make a goofy AI meme on Glyph or something from one of their tweets and then respond. And it's like a good way to get their attention because you take something they said and you take it in a new direction, right? And it's, it's goofy. So yeah, maybe that's, maybe that's something there.

But 100%. Um, yeah, launching on Product Hunt next week now that it's been in your podcast.

Oh, there we go. There we go. I'm, I'm, you know, thank you for subscribing to Meme Alerts, by the way. Really appreciate it. So I was just on my, the phone with my engineers this morning on this who are helping me, you know, SaaSify this, this newsletter. Super, super exciting. Um, and, uh, you know, I think, I think a lot of people have tried to use AI to make memes and I think there's a lot of use cases around AI and memes. that aren't AI writing the memes. It's AI finding the memes. It's like a better search database for memes using AI. It's like a lot of other things that aren't necessarily the actual AI making the memes. Um, but we'll see. Yeah.

But even then you couldn't— what if you could— there's all these embedding models now for, for images. So you can do vector search, you can do a bunch of stuff. Like generally I think image search is underdone. And meme search is maybe a subcategory thereof. It's, uh, there's not enough like, just find me the image that does X and find it for me very quickly. So the meme searching, uh, reply to auto reply guy, I think that is a product still to be done.

I was on the phone with somebody earlier talking about this idea, which I think is kind of crazy. I was at the movies last week and I'm watching, you know, the previews as you do at the movies. You're forced to. And for the first time I was watching, I was like, oh, that's going to be a meme when that movie comes out. Yeah, it's gonna be a meme. And I am kind of tempted to build something where it's like a trailer gets posted on, you know, the YouTube channel of Universal Pictures, and then AI scans through it, finds the memeable images, and then reminds me on the date that the movie comes out. Like, it might be stupid, it might be stupid, but I kind of think it's a hilarious idea. And I might just build that for fun.

I think there's a lot of stuff. This is like marketing automation, but it's not the corny cheesy, like just random copy text written by AI stuff. That is the auto video generation, the auto meme generation. I think that stuff is really interesting. I think we're, we're just getting to, to that. It feels like.

Yeah. Talk to me about auto video generation and what you're trying to build there.

Okay. So I am building this app. It basically gives you your personalized financial news for your portfolio every single day. And I'm like, you know, in test flight right now with just a couple friends and it's narrated well in this beta version, which I'm not gonna release because I don't want Hollywood to sue me. It just has like celebrities narrating it, which is very, very easy to do with today's tooling. And I built my, like. You know, with a friend at South Park Comments, I was building this like, you know, 6,000-line Python script to automate it all. And I think in— it actually gets— it's really funny to see, you know, a celeb, let's say Kim Kardashian, narrating your NVIDIA position. That has been really fun. The problem that I encountered though is like, I think for financial news specifically, just narrating that, I don't know, NVIDIA is down 5% today because the entire market crashed before we recorded this. The, maybe even more, it's just not, not as fast as just see like opening up Yahoo Finance or your Apple Stocks app and then just seeing what your current stock portfolio is like. So I'm, I'm curious about if I can tell a little bit more stories or give you investment ideas and, and stuff like that. But it's like this, I think this auto video generation will be huge. I think there's a whole nother issue of where do these videos go? 'Cause the natural place to put these is TikTok and YouTube and Instagram. But as far as I can tell, all of them auto-downrank the AI-generated content and the faceless accounts and all that stuff was like a growth hack that worked for 3 months at the beginning of this year and not ever since because now they tweaked the parameters of their algorithm to downrank that stuff. So is there gonna be like a new AI video only TikTok? Is there gonna be meme talk, something like that?

Yeah.

It feels like there needs to be an outlet for all this creativity.

I agree somewhat. No, I think I disagree. I disagree. I think the best way is mixing it and I know it demotes on the profiles or on, on the platforms. I know it demotes there, but I think the solution is just to kind of handcraft stuff after the AI. you know, makes it.

Yeah.

And then, yeah, yeah. Post-produce, like have it made, sent to you, and then you add a little bit on it. So it's not just AI or something like, I just, I don't, I can't, I, there was this entrepreneur trying to do stuff like, I don't know, a year or two ago. I remember a few people have tried to do this where it's like a social network just for AIs to like talk and you kind of just watch them. I think, and, and the problem with AI music and a lot of AI art and AI friends in general— we could talk about that too— is like, people like human stuff. We like seeing what's created by humans. Um, we don't even like text read by AI. Like, I haven't read an AI-written book. Like, I want to read Ernest Hemingway, I want to read Vonnegut, I want to read people who felt things, you know? And, uh, I think it's the same with all art and content. And marketing in general is like, we wanna read things and consume things created by people with real life emotions that are similar to our own. And so I think AI is, is gonna be a tool, but not the creator is my belief around that.

Yeah. I think where we kind of are is probably a lot of the stuff that especially you see in marketing is actually AI generated and then a human kind of went over it and fine-tuned it. And that is already happening. I think video is a special case because you, the whole reason, like you want to do this Product Hunt bot, right? Video bot. And I think the content would be interesting. You would probably put the content on TikTok to get like followers and to get a bit of distribution. And TikTok famously has an algorithm that's very tuned towards discovery and very tuned towards like discovering new things. And you could maybe even ride the Product Hunt brand with, with it. So get a little bit of built-in distribution around that. But would it be interesting and would the algorithm unfairly downrank it, especially if you auto-generate this stuff without you fine-tuning anything? You would be downranked like immediately.

Yeah.

So I agree that, that people wanna see stuff generated by humans for whom the human experience really matters.

Yeah. So I, I'm, you know, talking with another friend who launched on Product Hunt and he's building a different app, right? And it clips long-form podcasts like this and auto clips it into shorts and then posts them, right? Like Google Posts? Yes, but it auto posts them. And so it auto posts, but right now, you know, as we're speaking, I have the settings on for it posts unlisted, right? It posts on private. And then I approve the ones that are good because most of them are pretty bad, right?

Yeah.

And I think that is where we're at for now, and probably maybe even forever, is like, if you are a human person running a channel, you know, your own channel, and you're trying to have AI generated, you always have to be like a curator, almost a taste curator. And like, maybe it's different if you have a channel that's just AI videos, and that's like what the channel is, is AI food and it's only AI, but if it's Jason Levin, I, I can't have AI auto-generating stuff because it's just, there'll be too much slop. I don't have a huge following on YouTube, but on Twitter, if I start posting too much slop, like, you know, that's bad. That's really bad.

You're gonna get downranked.

You're, you're gonna get downranked. And, and more importantly is your reputation, right? Is like, you know, uh, that's everything in our, in our business. Um, and so that's, yeah, I would never let AI just auto-post on Twitter for me.

Uh, yeah.

Yeah.

That is, that's real humans talking to each other, which I think is per your earlier thesis, interesting content. And I think I agree with that thesis. It's just, you're making the editing process more efficient. You're making the editing process of getting to these short-form videos that just perform better because you get right to the point. It just makes that more efficient. And there's a ton of that stuff that's still coming. I think it's gonna probably right now it auto, it sort of does the cutting of the clip, it auto-captions it. Maybe it gives it like a, an interesting title. Maybe it will give you the description to post. But I think the next generation will like, you know, pop in little clips and animate like you're putting a little diagram of what we're explaining and and auto-generate all of that. I think a lot of that's still coming and I do think that's gonna make it visually more interesting, which performs better on these types of platforms.

Yeah, I, I definitely agree with that. I guess we'll see what happens. I have been seeing a lot of, I guess some people experiment with HeyGen. I know Ryan Hoover, founder of Product Hunt, has been experimenting a lot. I know HeyGen launched a weird thing. I'm not sure how I feel about this. They launched a thing where you could almost pay to talk to a person based on yourself.

Oh, it's like Intro, but for—

Yeah, but for AI versions of yourself. Like, I'm not ever going to pay to talk to an AI version of— even if, even if it's, you know, Elon, I'm not paying to talk to AI Elon. I'm not, I'm not too bullish on that.

You might pay with your time. Like it could be an entertainment thing that it's monetized with ads. It's kind of like, okay, yeah, but that's fair.

That's fair. I, but I, yeah, you're not going to pay like a, an intro fee. Nikita Beer's $1,400, a 15-minute intro for Nikita AI, right?

Apparently well worth it.

That's what I hear. I mean, I had, I had the dupe founder on my podcast and it is well worth it. He's up to 2 million plus users. One last thing before we, uh, we kind of sign off. I'm curious, speaking of AI, AI friends, what's your thoughts on the AI friend that was recently launched?

What AI friend recently launched?

The, you know, the like super viral AI friend necklace.

Oh, friend.com.

Yeah. Yeah.

What are your thoughts on that? I know about the memes about the $1.8 million domain or whatever, but yeah, I know nothing about it. Um, yeah, the whole necklace thing is getting kind of crowded. Like there's Limitless. There's, which I'm a tiny, tiny investor in, so maybe I'm biased. There's Friend, there's Based Hardware. That's right. Yeah.

Yeah. I had Based Hardware on the other, the other day on my podcast. I'm bullish on the idea of AI assistants, not bullish on AI friends. If I, if I had to, you know, say my thesis is like, I like my dog. I like my wife. I like my friends. But I could always use an AI assistant. That's—

you like human content created by humans, but sort of enriched by AI. Is that the Jason thesis?

I think that's my thesis. I think, I think you've, you've nailed it. Yeah. Yeah. I, I think we're going with that.

Okay. What about this? It is a, it's an app. Uh, like you're a large corporate worker. You're like a developer at one of the mega big tech corps. And you don't like meetings because you are an introvert like myself. Instead of real Gabor attending the meeting, like, I'm there and I'm maybe listening in, but there's like AI Gabor, and he sort of like nods and listens intently. And then when actual Gabor is being called on, like, I get a little alert, like the little siren goes off. That's And then it gives you like what the current question is being discussed and you're, you're being called on and what the question is. Is that too far for you or no?

I don't think that's far enough. I love it.

I want, you want it to be answering the question for you.

I want everybody on the meeting just to be AI talking to each other.

I think that's, it's coming. I think it's in the future. This is gonna, that meeting is gonna happen, but it's not gonna be the majority of the meeting.

No, it's not. But there'll be one meeting where it's like 3 AI dudes. I remember, you know, a similar way it was like, I went on some Google Meet and it's 3 NFT profile pictures and me. And I'm just like, what's up guys? Like, how's it going? I don't know. But they were real people. Yeah.

Yeah.

I love the idea of sending my AI to a meeting that I don't want to go. And like, if you're, I remember, so I used to be, I used to work at Amex. I was a data engineer and there are these like town halls, right? 1,000 people. It's like, oh God, this is such a waste of time. Like, I'm not learning anything from this. I'm not, I'm not getting anything out of this. Like, I wish I could just say, but like, you just go cuz you're afraid your boss will see that you're not there. Right. And so like, that is that killer use case is sending your AI version so you don't get fired. That's like the entire business model.

That's it. Well, but if you get discovered, you might get, you might be more likely to, it's like the equivalent of like the mouse jiggler, but as, as AI. I'm not like 100% in love with this future, but I think it is going to happen because like the email, the meeting could have been an email. That's the whole— maybe that's the name of that startup, but the meeting could have been an email. And why couldn't we culturally allow in these like large tech corporations for the meeting to be an email? And I think the incentives are kind of need to be revisited around meeting culture. Shopify does the thing where you get a counter of how much the meeting costs. I think it's a really good, a really good thing to do.

I think so too. My friend Shane, he's a designer. He just posted a, one of those goofy designs of, um, it shows when somebody quiet quit on the meeting. It's like this guy quiet quit 3 months ago.

Like, so like isn't listening or isn't watching?

Yeah. Yeah. Like hasn't actually shown up to work in 6 months.

Like, oh, holy crap. Okay.

Wow. Yeah.

Full on.

Yeah, I think that'd be funny. Cool. I think that's, those are all kind of the main things. This was, this was fun conversation. I'm, I'm excited to share it with the world. I think people will learn a lot about AI and can't wait for the highlights where it's at. I know, I know. This is awesome. Um, all right. Yeah. I will, uh, yeah. Talk to you soon.

Cool.