AI Abundance: a16z Partners Justine Moore and Anish Acharya Forecast a New Era for Consumers

Our philosophy is like for creativity, everybody is born creative and everyone has this sort of latent desire to be creative. But as we get older, what happens is one, our sort of taste improves. So like the kind of work that we would want to create gets more sophisticated. For most of us, that muscle atrophies. So our creative skill tends to diverge from our creative taste. So what does that mean? It's like hard to compose the kind of music you want to hear. It's hard to paint the kind of pictures you want to see. So one of the cool things about the new technology is it feels like if you have good taste, you can make good art. So that sort of lowers the floor for participation and being creative.

And then some of the more, I guess, niche subreddits are really good for specific sorts of topics. I think Local Llama and AI Video and then ComfyUI are probably my favorite 3 right now because they tend to be curated group of folks who are really like pushing the forefront of language models or video or kind of image generation.

Our hot take is that maybe the human part of human connections overstated. And actually, people just need connection. And look, I think there's an opportunity for these companion experiences to, you know, be tailored to people in a way that really helps them flourish.

Hello, and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas. And together, we'll build a picture of how AI technology will transform work, Life and society in the coming years. I'm Nathan LeBenz, joined by my co-host Eric Thornberg. Hello and welcome back to the Cognitive Revolution. Today's episode is a fun one, as we'll be exploring the frontier of AI-powered consumer products. My guests are Anisha Charya and Justine Moore, partners at A16Z Consumer, the division of venture capital mega fund Andreessen Horowitz that invests in consumer product companies. They have recently published an investment thesis called the Abundance Agenda. which outlines how they see the market for AI-native consumer applications developing and what they're looking for in new investment opportunities. If that sounds a bit academic or vague, I assure you that this is not armchair theorizing. To develop their investment thesis, the a16z consumer team has spent countless hours trying a huge number of different AI applications hands-on. Justine in particular shares many of her experiments and findings on Twitter and is a true must-follow for anyone who wants to keep up with this fast-moving space. Between her feed and the Abundance Agenda presentation itself, and I must say it was created with AI-native presentation app Gamma, itself the subject of a Cognitive Revolution episode back in August of last year, I was inspired to try at least 10 new AI-native applications just while preparing for this episode. For what it's worth, of those, music generator Suno AI impressed me most relative to what I had experienced previously. I was able to generate remarkably high-quality music in minutes, including the revolutionary reggae track, complete with lyrics about AI, that we used as intro music for today's episode. It's really a great tangible example of how empowering and how fun AI products can be, and I definitely look forward to making more music with it in the future. In this conversation, we cover a16z Consumers' investment approach, including their foundational belief that AI represents one of, if not the, biggest platform shifts of our lifetimes, their view that consumer trends can't really be predicted in advance and must instead be spotted in the wild, and their preference for products that are, quote, weird but working, even if the use cases seem unserious or potentially even Black Mirror at first. We also discuss how they are thinking about the retention challenges that many AI application developers are facing, the prospects for an AI product bundle and/or a personal AI procurement department, The reasons so many applications are still missing APIs, and the potential for entrepreneurs to disrupt regulated industries with unapologetic direct-to-consumer approaches. It was a ton of fun to compare notes with some fellow AI scouts, and at a high level, as regular listeners will certainly know, I share Anish and Justine's enthusiasm for the prospect of AI-provided abundance. From the creative superpowers we enjoy today with apps like Suno, To the productivity tools that are just starting to become really useful, to the fitness and mental health coaches that we might not otherwise be able to afford, and even to the medical and legal advice which can supplement human experts in critical moments— it's all super exciting to me. At the same time, I do have what I like to think of as a healthy fear for just how weird things might get. And I really do worry that the whole field is moving so quickly that competitive pressure, especially when combined with founder optimism or naivete, could cause real harm to consumers at scale and potentially even damage the AI application industry itself if such harm is sufficiently outrageous as to provoke counterproductive regulation before we have even established best practices to guide the industry. Balancing the benefits and risks of AI will, of course, be a society-wide challenge in which everyone has a stake and many millions have a voice. But even thinking small and reflecting on my own experience as the father of 3 young boys— and parenting is arguably the quintessential thing that doesn't scale— I unfortunately don't have any easy answers. Of course, I will want my kids to take advantage of the best AI tutors, and absolutely I will want them to pursue their own weird interests, many of which will likely involve AI in some form. And yet, I currently feel quite reluctant to expose them to applications whose developers are under intense pressure to grow and retain users. I certainly don't have all the answers, and we only had an hour together for this episode. But my hope is that in addition to helping advance our shared understanding of AI, at least some of the conversations that I have here on The Cognitive Revolution can serve as a gentle but credible reminder that with great power comes great responsibility. As always, if you're finding value in the show, we'd appreciate it if you'd take a moment to share it with your friends. I believe this episode will be of broad general interest for anyone who's investing in, building with, or just exploring their own creative and productive potential with the help of AI. With that, I hope you enjoy this conversation with Anish Acharya and Justine Moore of a16z Consumer.

Anish Acharya and Justine Moore from a16z Consumer, welcome to The Cognitive Revolution.

Sweet, Nathan. A longtime listener, first-time caller, so thank you for having us.

That's kind. I'm very excited to have you.

We're here obviously to talk about, you know, a new framework that we're proposing and excited about and investing in, really called the Abundance Agenda. And this is all the work that we've been doing and seeing and thinking about in the world of AI. And of course, the key word here is abundance. So let me kind of talk you through why we're so excited and fixated on this concept of abundance. And there's really kind of 2 reasons. Like every time there's one of these big platform shifts, everything from the internet to mobile to, you know, smaller shifts like fintech and crypto. There's been this, uh, huge sort of explosion of new companies being built and all these benefits for consumers. And if you look at consumers, every time one of these shifts happens, their lives just get better. Their lives get better because they can, you know, afford more stuff, both products and services, and things that are luxuries become a lot more accessible that were sort of out of reach become accessible and maybe even commodities. So, you know, the AI thing that's happening, it feels like consumers are going to benefit in this enormous way. It also feels like so many of the most interesting companies that we invested in were built around the last platform shift. This feels like kind of the moment for consumer investing. So hopefully you can tell from our vibes that we're very fired up. And then in terms of like the way we're actually looking at the opportunity set, there's really 3 markets and 3 things that we see that are resonating with consumers. So the first is everything in the world of creativity and productivity. So I think our philosophy is like for creativity, everybody is born creative. And everyone has this sort of latent desire to be creative. But as we get older, what happens is one, our sort of taste improves. So like the kind of work that we would want to create gets more sophisticated. And 2, look, for most of us, we sort of, that muscle atrophies. So our creative skill tends to diverge from our creative taste. So what does that mean? It's like hard to compose the kind of music you want to hear. It's hard to paint the kind of pictures you want to see. You know, sometimes it's even tough to compose an email in a way that really captures exactly what you're trying to get at. So one of the cool things about the new technology is it feels like you can, if you have good taste, you can make good art. And that just hasn't been true. So that sort of lowers the floor for participation in being creative. And we're seeing a ton of that. Like, what is the market size for that? I don't know, but, you know, Midjourney, Ideogram, all of these companies are seeing enormous consumer pull, both in usage and revenue. So there's obviously something to that idea of helping everyone be their best, most creative self. I think the second concept is around productivity and like how we do our work. And that's almost the inverse of the creativity observation, which is, you know, the thing that prevents us from doing great work is one, the kind of administrative overhead that we all pay, that tax we pay in doing our work, you know, and two, it's great that we have tools that help us do work, but sometimes you just want to delegate the work to someone or a team. And I think with the AI, you can imagine a lot of the administrative overhead goes to zero or approaches zero, and a lot of the AI can do the work instead of helping you do the work, which should sort of raise the ceiling. So creativity, productivity, sort of lowering the floor and raising the ceiling for creative work and professional work. So that's been like one of our biggest areas of focus. We're super excited, and Justine will walk you through it. Look, I think the second thing that's really working is companionship. So relationships, you know, we all know that there's like this fundamental need for human connection. I think our hot take is that maybe the human part of human connection is overstated. And actually people just need connection. And if you look at the public data, you see that people are using these companionship apps twice as often as they're using social apps, twice as often as they're using TikTok or Instagram. So something about these experiences is resonating with consumers in a really deep way. And look, I think there's an opportunity for these companion experiences to, you know, be tailored to people in a way that really helps them flourish, right? Helps them sort of feel good about themselves, help them feel accepted, help them feel like, you know, they're not the only one that is their own unique kind of weird. So there's something really powerful that's happening and really life-affirming. And, you know, consumers are behaving accordingly. And I think the last thing is, you know, wellness and personal growth. I mean, It's kind of well established, right? If you have personalized teaching, you perform 2 standard deviations better than the average student. If you have personalized wealth management or sort of financial advice, you just tend to manage your money a lot better. And of course, if you have a team of people that's looking after your physical health, mental health, you know, it just improves your wellbeing. The problem with all those things is that they're really expensive. So it's like a great vision for people who have resources and, you know, AI should collapse the cost of those things. And it should mean that things get dramatically cheaper. So yeah, those are kind of the 3 markets that we're looking at. I think the only thing before I hand it to Justine that I'd say is like the way we're prosecuting these markets is a little different from perhaps how others have done or how we've looked at it in the past. Number one, I think we're really looking at the intersection of weird and working. So like we think that weird products or products that feel like they might not be serious, tend to be the ones that end up being the winners. And this is one of the reasons that incumbents don't just win everything in consumer. This is also one of the reasons, by the way, why first-time founders tend to outperform second-time founders in consumer. Because I think their tolerance for being embarrassed and building weird things is actually much higher. So that's kind of the, the area that we've looked for, weird and working and being nonjudgmental about that. You know, related to that, I think we've got a real belief that we can't predict consumer, we can only observe consumer. It's just really hard to know what is actually going to be built. And that's why we're sort of in the business of observation instead of having our own vision for what consumers should or will want. I think the last bit is just the importance of culture as an overlay here. We've always as a firm been all about the intersection of culture and technology. And AI is such a sort of, it's enabling such a set of creative fields, it's hard to talk about the technology without talking about culture as well. So.

I love the notion of an abundance agenda for starters. And that is what has me so excited about the AI moment. When I first got access to GPT-4, it was within the 60-day testing window that I started going to it with my medical questions. I felt that even within that short time and even as rough as the product was, and obviously there was going to be a million extensions and access to databases and stuff, which it didn't have at the time. I felt like for many things, it already was a superior experience to the all-in experience of going to the doctor. And I was like, man, the AI doctor for everyone in the world is just going to be such an unbelievably empowering, democratizing, just world-changing phenomenon. So I do love the notion of the abundance agenda. And I also do share the sense that, that it's like very hard to predict where this is going. And I always say my crystal ball gets foggy, you know, more than a few months out. So Justine, do you want to kind of take part 2 and give us a little bit more into the things that you're looking for and the things that are exciting you in the market early as it is?

For sure. Yeah, maybe, maybe I'll start with like the kind of how we look at everything prosumer, which is essentially enabling everyday consumers to do things that before were considered professional-grade work or that required very specific skills, knowledge of complex software, things like that. But I think like the first killer use case probably of consumer AI beyond chat, like just chatting with ChatGPT, was content generation. So this kind of magical experience of, hey, I can type in a prompt and I can get an image, I can get a song, I can get a video. And that has a ton of use cases across both just like everyday consumers making fun things to share with friends, to creators making things to help them grow an audience, to people doing this for a business, whether they're like a small business making an ad or promotional materials, or they're maybe even a creative professional at a large company that is making their brainstorming and iteration much faster. I think we started with content generation around kind of the base model, like How do you make the magic happen? And so we needed all of these foundation models to be built around all of these different content categories. And now we're kind of hitting the point where there's enough publicly accessible models, whether it's open source or closed models that you can access via API, that we're starting to see people build the workflow and the features around these models and make it truly easy for everyday people to use. So you don't have to be a prompt engineer. You don't have to go into a bunch of tools to edit and piece together content.

Hey, we'll continue our interview in a moment after a word from our sponsors.

And actually allows you to like iterate and refine on said content. Like you have a person working alongside you, like a brainstorming partner who's also an editor who can like make changes on the fly. So content generation and editing, we've been spending a bunch of time on the workflow, have also spending— been spending a bunch of time thinking about like, how do you translate all of the incredible things that the open source community is doing into consumer accessible products? I think the amazing thing about AI right now is that, especially in image, but we're actually now starting to see it a lot in language as well, and probably eventually video too. There's a really vibrant group of people all over the world who are taking these base models and fine-tuning them and creating all of these LoRAs and checkpoints and then combining them in interesting ways to make very cool content. You've probably seen even with things like ComfyUI, like people are making whole videos out of image models. People are generating 3D textures. People are doing very incredible things, but historically that's been limited to folks who are technical, who have like a 4090 at home or more professional-grade hardware. So we're very much hoping that in this next wave of consumer AI, we start to see a lot of companies that are bringing these like truly cutting-edge open-source products to everyday consumers and more accessible like browser-based products, apps, things like that. We're also, I think, starting to see kind of on the editing front, like, how do you combine the generation and editing experience? How do you go from, hey, I made an image and all I can do is download that image or regenerate the same prompt to like, I want the same thing, but a little bit different. Like we've seen with the Midjourney variations or even with things like we've seen with Pika's ability to modify a canvas. You make the video and then you can say, I just want to change this small part of it, or I want to expand this, or I want to change the dimensions. So it's not kind of like a one-off generation artifact. So that's kind of like productivity for and workflow tooling for consumers or creative individuals. We also spend a bunch of time thinking about like maybe the less creative, slightly more professional use cases around productivity, things just helping you get your everyday job done. Like basically the idea of can we use AI to give every individual an assistant or a chief of staff or a data analyst, kind of whatever makes most sense to their role to give them a lot more leverage on their time. There's so many tasks that we can now automate. Before AI, it was largely kind of like data or rules-based tasks. And now there's so much more you can do with kind of natural language prompts to say, go out and research this for me, go out and go through all of these numbers, come up with a pros and cons list. Generate a ton of emails, whatever, whatever the task is, how can those things be automated? And thus far, I think we've mostly seen that in like exactly what I mentioned around text-based interfaces. I think moving forward, the question is like, how do you have products that are more integrated into someone's daily life? So you're not stopping to enter a prompt to the AI and wait for it to come back, but it's maybe listening to you, whether that's through software or through hardware on a regular basis. And then Interjecting both to answer specific questions or do specific things, but also just to make like smart recommendations as you go through your daily life. So as you can probably tell, there's a ton of stuff that we're excited about. Happy to go through any of those in more detail.

I have a million different questions and I'll try to ask them as crisply as I can to get through as many as possible. For folks that want to, you know, go absorb your full document, I've had the chance to do that with an advanced copy built using the Gamma app, a former guest, which is a great generative AI tool for slides that really has some nice UI patterns that I recommend. It's actually a pretty frequent recommendation of mine to go check out the Gamma UI. I would put it up there right alongside my own Waymark in terms of a very nice fusion of the generative AI, the chat back and forth, but then also the like controls that you need to actually make something that you can put out there into the world that's kind of a finished product. So let's start with maybe some of your methodology. So much stuff happening, right? I describe myself as an AI scout, and that means trying to keep up with everything. You're kind of trying to keep up with everything. People should follow you on Twitter, Justine in particular, for just kind of a firehose of product experimentations and funny findings and all that kind of stuff. But how do you spend your time? Where do you go to figure out what's cool? Is it just trying products? Is it Discords? Is it Other stuff?

I mean, Justine's going to have a great answer to this too, but like, look, it feels like a lot of the communities on Twitter and Reddit as well, different aspects of the community, it feels like you can't be a part of the community if you're not using the products, which is why we're using Gamma and all these, you know, our compliance team hates that, but we are using all the products because it just, it feels like it's hard to be authentic to the space and have real intuition if you're not using the product. So like Twitter and Reddit are the main 2 places and, you know, Justine's really led the way on this for our team, but We're just doing a lot of work to make sure that we've tried these products ourselves. So whether that means ComfyUI or, you know, Automatic1111 and just getting them set up on our own laptops and trying ControlNet and just really trying everything so that we understand what's going on at a sort of product instinct level instead of just an investor metrics level.

Yeah, totally agree with Anish. I would say Twitter, Reddit, also Discord obviously is huge. I think for specific communities and people who are really deep in In certain areas and trying a different thing, a ton of different things and troubleshooting. I think the very cool thing about a lot of these AI products is that it's not just super technical users. Like, we kind of have a joke on the team of we have various Discord channels or Reddits or all these anonymous Twitter accounts. Like, it could be the CEO of a giant company. It could be a 16-year-old in the Philippines who discovered a really cool app and has spent enough time on it to like be become an expert. And I think that's one of the things that is so exciting about these new platform shifts is like it truly enables so many more people to become experts and to share their thoughts and to be bringing something new to the community. So yeah, we probably spend too much time online, but we spend a kind of a massive amount of time. Like obviously the Twitter algorithm is great at surfacing things that kind of get attention and get a lot of engagement, but I think Reddit and Discord in particular are like various communities like Local Llama or AI Video, like where there's really folks who are not like posting something to get attention or to get engagement or likes or like drive eyeballs to a product, but who have just spent a ton of time tinkering around with something and are trying to make something cool or have like a real question. And I think engaging with those communities and seeing how people are using these products and, and what they like and what they don't like, it's like Incredibly valuable. There's kind of no, no substitute for that.

I think my biggest takeaway there is I need to spend more time on Reddit. Are there any particular subreddits that you would say are the absolute place to be?

Oh my gosh, yeah. So I did— I'll send this to you— I did a big recap at the end of the year on the biggest AI subreddits and kind of where all the activity is happening. I think it depends on what you're looking for. I think some of them are like, you know, the ChatGPT or the Stable Diffusion subreddits are really good for, hey, I just want to see like what's popping off in AI and like what people are interested in. And then some of the more, I guess, niche subreddits are really good for specific sorts of topics. I think Local Llama and AI Video and then ComfyUI are probably my favorite 3 right now because that— it's— they tend to be curated group of folks who are really like pushing the forefront of Language models or video or kind of image generation. There's like a running joke in the community that AGI is going to be discovered or released or launched or whatever your term for it is, depending on how you interpret AGI, by somebody like running a 4090 in their basement who then posts about it on Local Llama. And it doesn't seem too far from the truth.

We've had Sam Altman possibly shitpostingly teasing certain developments on Reddit as well. So yeah, all bets are off, I would say, in terms of just how this all develops. For the business side of the consumer startup, this is something that we've kind of wrestled with at Waymark. And I don't know if you've seen our product, but basically it's another kind of text-to-output type experience. You get to say what sort of video you want, what you want to communicate, et cetera. We typically serve small businesses and we make commercials, video ads, but especially focused on TV commercials. And we have found that we do great when we partner on a B2B basis with the TV platforms. So, you know, whether it's broadcast or cable or, you know, OTT, whatever. On the contrary though, when we go direct to an SMB, we've had what I would say is a pretty representative, I think, challenge across the kind of consumer prosumer space, which is The economics are not quite figured out, right? Inference is kind of expensive. So we do give a free trial, but then we need to convert a certain percentage to get our unit economics to work. And it's like, geez, well, it's got to be kind of at least $19 or maybe $29 a month. And then we have a bunch of people that will convert, and then a significant number will immediately go and cancel. And we'll say to them, well, what was wrong or whatever? And a lot of times they'll say, well, nothing was really wrong. It's just that It was worth it for me to pay this $19 or $29 or whatever right now to get this problem solved, but I'm not sure how often I'm going to come back. And there's a ton of stuff like this coming out. And so I'll just kind of come back if I need to. So I imagine that's a huge challenge. What are you seeing right now in terms of retention? Has anybody solved that? Is there a good model for it? I have an idea for you I'll pitch next, but have you seen anybody with a good way through that problem?

It's not very well understood how well this media format can perform. So I was at Credit Karma for 5 years, you know, Credit Karma is one of the biggest consumer products out there, 120 million US members, I believe. And most of their growth story, actually, it was a paid growth story, but it was mostly through television and kind of long tail television inventory, you know, as well as video to some extent. And it just, you know, I think it was a poorly understood format by digital SMBs and digital startups in particular, you know, it just felt kind of legacy and dated, but it performs really well. So one, I think there's just a little bit of a marketing to marketing challenge and like showing people how well the format can perform. Look at 2, of course, like the cost of inference is going to go down and the models are going to get more efficient. So like that's inevitable. It's happening every day. I think the last bit though, and our partner Alex talks a bunch about this, we sort of look at prosumer products and think about high NPS churn versus low NPS churn. You know, low NPS churn being like, I hated your product and I'm never coming back again. And highest NPS churn being like, look, I'm an episodic user of this. And there's probably some way of modeling that that's somewhere between, you know, recurring sort of SaaS revenue and purely transactional one-time revenue. So I think we just have to have a slightly different framework because a lot of these content generation products are going to be episodic and that's okay. We can build great businesses around them. It's just hard to look at that churn and treat it the same as you would low NPS churn. Does that make sense?

That's really interesting. To be honest, we have historically struggled with that argument and maybe we just haven't come up with quite the right way to frame it. But this is maybe the first time I've heard, you know, from the investor side of the table, somebody saying like, we'd be open to a financial model that isn't really anchored to strictly recurring revenue model. In the past, people have told me outright, we'll give you zero credit for anything that is a one-time purchase.

You got to come talk to us, man.

Well, okay, well, let's set something up. I like the sound of that much better. I guess an alternative path perhaps would be this concept that I've been kicking around for a while of the AI bundle, which Because we're in the TV business, it's sort of inspired by the cable bundle. And I've just been thinking, it seems to me like we're all headed for a world in which we pay like $100 a month or something, or we slash our employers pay $100 a month for AI tools writ large. And for that, I kind of want an experience a little bit more like I get with cable, where I have 1,000 channels and I may watch some, most never, and some rarely. But I have access and that kind of reduces the friction. And then, you know, they split up the money, obviously, depending on their deals. Would you be excited about an AI bundle? Do you think that that has any legs? It seems like every app developer I ask is into it and the consumer side seems to make sense, but like who actually would put that together and, you know, why will it fail?

I don't know.

Hey, we'll continue our interview in a moment after a word from our sponsors.

I think it's, it'll probably happen eventually, some sort of AI bundle where you have broadly maybe the most popular products in every category that you can kind of tap into or use different credits for or things like that. I think thus far we haven't seen that sort of thing take off partially because there's still so much fragmentation going on in products and in various markets. Like people use different products for different things. And people, there's so many of the products now are starting to fragment into like, hey, this is an image generator, but this one is really good at anime. This one is really good at photorealistic images. This one is really good at text. And so it's kind of hard to have a bundle that maybe combines what everybody wants in these 4 GPUs when there's like so much kind of diversity in product options today, if that makes sense.

I agree with that. I also think that, you know, bundle, it I mean, it's a great idea. It might be a sort of a bridge solution. So like, I think the long-term trajectory of this is that if you're a business owner, you just don't want to think about marketing. You just want customers to show up and you want to, you know, your business to have paid some profitable acquisition costs. So when you think of like, what is an AI CMO or what is a sort of entirely AI-powered back office? I think it's one in which an AI system can choose what products to buy. How to be a subscriber and when to churn and how to use them. And you just sort of work on whatever your craft is and customers sort of show up and get charged and pay and all of that.

So just to make sure I'm understanding that correctly, the alternative to an AI bundle that I as a human would subscribe to and sort of let ride like my cable bundle is an AI agent that is in fact managing my individual subscriptions more actively. And I'm going to, by the way, need that for content too, right? Because like now everything's fragmenting off the cable bundle. So I almost need like my content subscription AI agent to just manage, you know, what have I actually watched, cancel what I haven't. But that's basically the idea you're sketching out.

Totally. And I think it's actually bigger than just subscriptions. I think it's your entire back office. Like a great example is a company called Format, right? Justine, that sort of site generation company where they will actually, you know, they'll generate N variants of a website. And they'll do all of the multivariate testing so that like the best one just gets chosen for every consumer. And you as someone who, you know, is a D2C seller online, like you don't want to think about creating variations of your website and doing performance marketing that way. You just want as many customers to go through the funnel as possible. So that's kind of like speaking in the direction of like an AI CMO or an AI-powered back office.

The threat to the AI bundle is the AI procurement department.

That's right.

That's right. Okay. Another question I had on this, and this is, I guess it may be a 2-parter. So at Waymark, we fine-tune models specifically focused on the script writing, and then we use any model that we can get our hands on, open source, commercial, whatever. Our goal is to make the best end product that we can. And I have not yet found a good image-to-video app. So I guess first question is like, do you have any recommendations for me? I tried a number, even just off the slides last night. I have a picture of my— I kind of use the same image for all this stuff. I have a picture of my kid making a snow angel in the snow. And I prompt it like, here's the image, kid making a snow angel in the snow. And it's always like his face is melting. We're in still kind of the Will Smith eating spaghetti thing. So I'm wondering, first, have you seen anything that is great in that respect? And then second, if there is something great, I kind of expect it's not going to have an API right now because most of these things, not all, there are exceptions, but very frequently I see the app does not yet have an API. And that kind of surprises me for some of these image generation or these sort of what feel to me like point solutions, it feels to me like a lot of them are still even like an image to video, right? Like that's still an input to something most of the time. I don't just take that clip and like go publish it, right? I'm going to use it in some other context. So is there a good solution that can get my kid to make a snow angel? And why are there not more APIs for these sorts of products so far?

Yeah, it's a great question. I think what we've seen thus far in video is a mix of things. So like you mentioned, a lot of closed models, people training these foundation models, folks like Runway or Pika or Genmo, who I do think all do a good job at kind of different styles of, of image to video. Like Runway, for example, is maybe if you have a photorealistic image of your kid, especially with all of the new things around motion brush to try to prevent the face from moving or warping too much. But you might be moving the arms if they're making a snow angel or something like that. Pika might be great if you want to do kind of like a stylized, like a cartoon or an anime. Like they all kind of specialize in different things. So we've seen that kind of closed model ecosystem in video. And then we've also seen, especially I would say in the past couple of months, a ton of work happening in open source video. The first half of last year, I think the kind of only broad open source video model was the ModelScope and then the iteration of that, which was ZeroScope, which are great, but get kind of warpy, very short output, stuff like that. I have seen a lot of folks iterating on That with things like Hotshot that are making it easier to do consistent image to video. In terms of the API question, I think video is tough because it's relatively expensive to generate and it takes a long time. And so, and the quality is not quite there yet for folks to be able to use commercially. So you have this weird combination of like costs a lot of money to generate and the outputs aren't great enough to be So exceptional yet that you can put it in something that you use to generate revenue or things like that. So that's why it feels like video is maybe a little bit more at the kind of creator prosumer stage than the actual kind of like enterprise API stage. We've actually written a massive amount about AI video. We spent a ton of time in the space. We're really excited about it. I think this will be the year that there's a lot more development happening in kind of the underlying model architectures as well that enable us to be able to generate more temporally coherent, consistent, non-warping and morphing clips. So I expect we, if we talked about this in 3 months, 6 months, definitely by the end of the year, we'll hopefully be having a very different conversation.

Gotcha. So to summarize that, the lack of APIs is a reflection of the fact that the products are just not there to the point where other application developers are Jumping to integrate quite yet.

I think that, and also I think like, I think we see at the early stages of these sorts of developments and categories, the companies at the forefront are often purely focused on developing best-in-class models first before saying like, hey, how do I enable other people to put this in their application? So I also think it really hasn't been a focus of many of the teams yet. Like they're focused on Getting to the very best model. And then once they have that, and once they can see that in some of the consumer prosumer numbers, I expect we'll start seeing APIs come out.

I guess a couple of things that jumped out at me that were maybe not as much represented in the taxonomy of the different kinds of things that you were looking for. One was just things that are a little less weird. I was recently introduced to Shortwave, which is a kind of originally a Gmail competitor before there was generative AI to take it to the next level. And now they've built, for me, the best implementation of RAG that I have seen possibly anywhere. I've been super impressed with it, and that's a subject of an upcoming episode. But Gmail is a big product, right? Email is a lot of time, and we generally, I'm very impressed with Shortwave, but there's not yet a killer email assistant. And that is like not a weird thing. So I guess I'm wondering why, why weird versus just like, hey, kill it on this like super obvious, you know, everybody's sending emails all day use case. Is that your, is there a reason that that's not as much of the focus?

I mean, I think so sort of made 2 things to consider. I think one is if it's a feature you can get promoted for building at Google, then it's going to get built at Google. And it feels like, you know, email, AI email composition, is, is like a sort of straightforward sustaining innovation for an incumbent. I've heard that Shortwave is great. We're of course investors in Superhuman. So like, I like the category, but you know, one, I feel like that, that sort of thing is something that incumbents are going to build and we're just less focused on incumbents. I think the second is like some of the things that we speak to here may actually be, end up looking like the products you're describing. So voice has been something we've talked a lot about. I'm really passionate about it. You know, voice has just never worked as an interface to technology. Like it's, it's just never worked, period. And now it works. And there's a sort of question of like, well, what is a voice-first productivity experience? What's a voice-first inbox? That feels like a stretch for someone at Gmail to build, whereas someone at, you know, Superhuman or Shortwave or a company that's just, you know, starting today might build it. And it might actually feel a lot more like you're, you know, reviewing your inbox with your EA. Versus, you know, sort of linearly going through these blocks of text and context switching. So yeah, I do think some of those things might get solved, albeit in a way that feels a little weird.

I'm for, I'm, by the way, generally quite pro-weird. I do enjoy all these, um, you know, strange AI experiences. How do you think about the breakdown? You know, we've heard like Elad Gil was on the show some months ago and said he sort of expects 80/20, you know, 80% to go to the existing platforms as sustaining value and 20% to go to startups, but he's quick to add, but that's huge because this is probably the biggest platform shift of our lifetimes. Is that basically how you see it also, or would you put a different spin on that macro lens?

Yeah, I mean, I think we'd almost take the opposite view. It's like 80/20 to startups. I mean, look, I think that Incumbents will become more efficient and more profitable. And that's great for them because that means their products will get cheaper and more people will have access to them. That's kind of the abundance concept as defined by sort of existing companies and existing products. But, you know, the whole history of these platform shifts is that the new markets are way bigger than anyone predicts. And they always favor the kind of platform native startup, the AI native startup in this case. So, you know, if we have this conversation again in 2 years, And it's 80/20 in the sort of direction that Elon's predicting, I'll be surprised. I'll be disappointed.

So how do you think about that from the standpoint of like Salesforce? That's always my kind of canonical go-to. I always think Salesforce will add an AI layer before you will build an AI-first CRM that can really displace them. Is there something wrong with that analysis?

My challenge to you is like, do you even need a lot of like, You know, we don't focus on B2B, but just to weigh in for a second, you know, what is BI, for example? The entire category of BI is an interface between a CEO question and a database, right? Like, I would argue that you actually don't even need any of those products anymore. And in fact, you can just have a CEO asking a question and having an AI or LLM generate a query that goes to a database and a bunch of dashboards are returned. And that same system can decide what dashboards should be in a CEO's inbox every morning. So I don't know. I mean, I think that there, yes, I think that in terms of adding AI to the existing products, yes, incumbents and Salesforce is set up well to do that. I just sort of question how many of these categories will be either unnecessary or dramatically different thanks to the tech.

Yeah, this is kind of a, if software was eating the world for a long time, now AI is going to eat software. And it sort of strongly suggests that you believe that the models will continue to get a lot more powerful. Because even like GPT-4, we just did a show with somebody from, with Juan from data.world, and they've got a benchmark now of like kind of complicated queries and, you know, can AI generate the query effectively? And the answer is like, even with GPT-4, not really yet. And so for me, like maybe the $100 trillion question is, do the models just continue to advance to the point where, yeah, they don't even need all this scaffolding and structure and you just go straight natural language to database and boom, and it's like all that crap gets kind of washed away? Or do we sort of find that next couple rungs on the ladder to be harder to climb? And so a lot of the scaffolding kind of continues to be important, whether it's Gmail or Salesforce or anything else. But it sounds like the implicit belief underlying this projection is that we're nowhere close to the models topping out. Is that fair?

I think so. I think the models are getting, you know, better, faster. Justine, what do you think actually? Because I also think you've got to distinguish between the language models and the media models. I mean, they're all moving in different directions.

I think we're big believers that the models will continue to get better. I think we're also big believers that the workflow you build around the models is very important and that there's more and more people building that workflow now, sort of like I mentioned in content generation, but it applies to every category. The first wave of AI was like, you can have this magical product where you put in a prompt and you get text back or you get an image back or you get a video back. But I think a lot of kind of what you're mentioning, like the challenges around it is like, okay, what if it's not exactly what you want or what if the answer is wrong? That's not to say like, oh, this is never going to work or the models have to get much better or like, That this means that the incumbents are going to win. Like, I think we're big believers that there will be startups that tackle those challenges. Okay, what is the accuracy layer? What is the layer that double checks? Like all of the stuff we're seeing around RAG now, like people run into these problems and innovate on solutions, which I think we believe will just, will just continue to happen.

You know, Nathan, I think sort of related, one of the most interesting things that feels underdiscussed is how how artistically opinionated you can sort of tip models in, especially, you know, visual media models using fine-tuning, you know, both the DreamBooth approach as well as LoRAs and, and all of these things just allow you to create these really opinionated models using fine-tuning techniques instead of training new foundation models. So it also feels like some of these techniques will allow you to have a lot more leverage in sort of the specialization that the models have.

Yeah, I definitely don't rule that out either. The leap that I felt in the first days of GPT-4, even relative to having spent tons of time fine-tuning just one generation prior of models for very specific tasks throughout that summer of 2022, I was already bought into the idea that AI is going to restructure the economy. And then I got the GPT-4 access. So I think the prospect of another leap is probably the biggest thing that is still coming at us and going to shake the snow globe again. I think that's probably more likely than not. And so I think if I have to bet, I'm probably betting on the same side of the question that you guys are. Although it's hard to know exactly what shape that'll take and how soon it'll come. And it's easy to sort of say, well, let's just keep refining what we have. And layering on the assistant to this great platform that we've built over decades in the meantime. And hard to tell those people that they're definitely wrong. So yeah, it's interesting. That does suggest that there is perhaps a lot of disruption coming for all corners of the economy.

I mean, I also think some of the primitives, for me, they feel underexplored so far, like GPT-4V. It's so fucking cool. I mean, it's amazing. Computers can see. They can see. And like that, you know, it, I've been very excited about it pre, you know, release and now post-release. It's also super cool, but it also, it just feels like it's underexplored from a workflow and a product concept perspective. So I actually think there'll just be a lot to be built, even if there isn't another leap in models in the near term, which I'm sure there will be. There's a lot to be built with the stuff we already have.

Yeah, totally. I think if we don't get any better models, we have a decade in front of us of figuring out how to plug GPT-4 in everywhere. And if we do, then we might have just a few years of ripping a lot of things out and plugging new models into where the stuff that we ripped out was. One thing I've been kind of expecting is, and this goes directly to the abundance agenda, is somebody to come out and kind of be like, okay, we're doing the Uber strategy, for AI doctors, and that is to say, we don't really care what you tell us we can or can't do. We're just going to do it. We're going to go straight to the public. We're going to be so damn good that everybody's going to love us and they're going to hate you for trying to stop us. And that's our play. That hasn't really happened yet. I haven't seen anybody come out with that positioning. And you could kind of extend that, doctors, lawyers, and a number of other things. What do you think is up with that? Why is there no just unapologetic, you know, uh, Kalanick-like push on the AI doctor front?

That's a good question. I mean, I don't want to weigh in too much because it's not the area that we focus on day to day. But look, I do think there's a lot of companies doing work in this area. So, you know, we're investors in Hippocratic. They're building a foundation model that sort of enables things like doctors and nurses as a service. So like one, I do think that there's a specialized dataset and knowledge set you need to fine-tune or train the foundation model to be able to just be really good. Because I think even if you're, you're Uber and you don't give a shit, like, you still do give a shit about at least getting the right product outcomes. You know, Uber may not have paid much attention to regulators, but it's not like their cars exploded on the way to your destination. So I do think that there's, there's a little bit of that. And then look too, all of these companies are at best 14 months old. You know, Midjourney, at least their latest sort of iterations, are 14 months old. ChatGPT is 14 months old. So it's just like really hard to overstate how early it is. And, you know, if we have this conversation in 12 months and there isn't something that looks like that, and that's really getting to scale and hopefully it's democratic, then I'll be surprised.

I see the, the drgupta.ai link in the chat as well. And I've heard a little bit about that and heard that that is kind of the positioning of that product. And there's a somewhat notorious entrepreneur behind it. I don't know if you want to tell more of that story at this point, but I'm looking forward to that hitting the broader consciousness and seeing how that shapes up.

For sure.

Yeah.

No specific comments on that one. I would say there's probably a reason that Uber was in rideshare and not in like, I don't know, financial services or healthcare. Like there's a couple of these categories that are extremely highly regulated where you can get extreme, very heavily fined as an individual or arrested or get your medical license taken away or something like that, where you're directly impacting someone's kind of like core livelihood and career. And so I think in those categories, there's massive room for improvement and AI will definitely come to them. It will just take a little bit longer probably to change some of the systems, the structure, the regulation to enable these folks who have spent many, many years in school to become a doctor or something like that to feel comfortable doing this type of thing, whether they're overseeing chatbots, whether the chatbots completely replace them. We should have GPT-4 take the MCAT. I think that would be very interesting. But I think a category similar to video where in a year we'll be having a different conversation.

I'm honestly confused by the relative lack of this kind of frontal attack. I'm not necessarily sure it would be a great thing. What I always say is I want both for now at least, right? If I have a serious medical issue, I will absolutely want to talk to a human doctor and I will also absolutely want to talk to AI. And I would feel that I'm missing out if I'm not getting the best, if I did only one or the other. So I'm not recommending people move past the human doctor just yet. But I am surprised that there hasn't been somebody, you know, if only for just kind of the distribution and, you know, sort of Streisand effect potential, who would come out and say, you know, this is what we're doing. And, you know, there could be a jurisdictional arbitrage angle to it as well, of course, that seemingly would make it kind of a cool opportunity for someone.

Well, one kind of like related concept I think that we're seeing more activity in is things around therapy might be too strong, but just, you know, journals that talk back to you, sort of things that help you with your mental health. I mean, I think one of the clear trend for the last 50 years is that, you know, one, the stigma around things like therapy has really collapsed. And, you know, 2, it's sort of gone from something that was available or needed by a small number of people to something that a lot of people want to have access to. And there's a shortage of therapists and costs are enormously high. So I'm actually fairly optimistic that a product like that will really break out. And look, if you spend time with the products, I You know, I've been using one called Rosebud, which is really cool. And I know Justine has used a couple of them, our partner Olivia. They're really satisfying, actually. You can learn a lot about yourself and, you know, get a lot of feelings of wellbeing, which now feels weird to say because you're chatting to a computer in a chat box and, and yet the feelings are there. So something like that feels like it's going to happen in the near term for sure.

Yeah. You talk about weird, but working. So actually my, later this afternoon, I've got Eugenia from Replica coming back for a second appearance. on the cognitive revolution. And they've recently put out a paper, I'm sure you've seen, that showed, and this is actually basically pre-generative AI, that interaction with the app reduced suicidal ideation for like 3% of the users that they studied. And this was in late 2021. So, you know, this is, we're talking pre-ChatGPT here, right?

Pretty remarkable.

One thing that I worry about in the sort of extreme speed with which everything is happening is I look at somebody like Eugenia and I think, man, she is the perfect person to do this, right? She came at this from having felt a lot of loneliness in her own life and a real deep concern for people that are struggling with these issues. And she built something pre-AI. One of the most memorable things anybody said on this show to me was, we couldn't build a bot that could talk, but we could make one that could listen. And just her empathy for that user profile was cut above anybody else. And so she's been early to this space and she clearly really cares about her users. But now I'm like, geez, the App Store is getting crowded and the sort of blitzscaling philosophy obviously makes a lot of sense for maximizing your return on investment, but does create some weird incentives where if I'm trying to compete in the virtual friend space, I need to make that economic flywheel work, right? I need to pay per click and I need to monetize and I need to kind of hook people probably. So I wonder if you guys have a sort of companion framework to all this that sort of governs me. I mean, we're headed for regulation if we don't figure it out, I feel, as an industry. Ourselves and maybe anyway, but I'm kind of wondering, do you have some guidance for founders or do you have guidance for yourselves for what investments you would and wouldn't want to make to try to make sure that people are doing this work in a pro-social way? It's easy for me to imagine that we could end up in a sort of parasitic or predatory dynamic and that if people are paying per click for users, That might even be the dominant strategy. So how do we avoid that trap? Which, by the way, I think will be the thing that brings regulation if it doesn't come, you know, for other reasons.

I think we spent a ton of time in the companionship space, as you've mentioned, started with a few products, ended up getting tons and tons of products, both on web, things in the app stores, things getting banned from the app stores and going back to web or Telegram or Discord. I think at the end of the day, it's Consumers decide what products they want to use and what they spend time on and where they spend money. And for us, we haven't really seen these kind of like clickbaity, scammy, not quality products driving a lot of user acquisition, engagement, and retention over time. I think it's really easy to be judgmental about like, I don't think people should be talking to a bot about these topics, or I don't think people should be using a bot for this. And what we're looking for is more like kind of taking off the lens of what we would, would or wouldn't do with a bot. What is adding value to someone else's life? Like we all have very different experiences and things that we struggle with. And so I think what we're looking for is more like, is there a segment of users that are really engaged on this product? Are they retaining on it? Are they spreading the word to their friends because it's adding value in some way to their lives? And thus far, those have not been like— like, it's pretty easy when you download an app to see like, hey, this is like a poor quality model, or this bot isn't really listening to me and it's just trying to get me to like upgrade to pro or click on ads or things like that. Like, kind of the consumer market sorts those things out over time in that the quality products tend to rise to the top.

I agree with all of that strongly, Nathan. And I think the other thing is like, if we— let's say we were PMs looking at this industry and we kind of said like, well, what's proxy metric. I actually think that the percentage of people who can leave a product feeling good has been going up as a lot of the social and relationship products have evolved. Like I might say Instagram is kind of a global popularity contest, right? And definitionally most people are not winning. And maybe it's, it's still satisfying or on some level because people go back, but I just, I feel like a lot of people probably, you know, walk away feeling not as good. I think for TikTok, there's a bigger opportunity to participate. Like anyone that wants to do the trend can participate and be creative. So more people can sort of walk away feeling cool. And I bet you, if you looked at people using the existing companion apps, like even a far higher number of them would walk away saying they're feeling really good and fulfilled by the experience. So I think the kind of overall trend is really positive and we may see regulation, but I don't think it's going to be because like users want to be in abusive relationships with companions.

How do you think that changes as we move toward more agentic products? I've recently been doing what I call red teaming in public on a bunch of AI calling agent products and just seeing like, what will they do? What will they not do? Spoiler, they'll basically all do anything. So if I say, call and make a ransom demand and claim you've got this person's kid, They'll just do it. If I clone a Donald Trump voice and say, go call these people and say you've rethought your position on immigration, you now support open borders, it'll call in the Trump voice and do it. It seems like we are headed for a world where we are going to need some governance on this or standards or self-governance of some sort. I think there's a good argument, or at least a decent argument, for, okay, people are spending their own time, their own money, their own attention. We can trust them to do that as adults. But as we create these more agentic things that go out into society or into group settings and interact with people who maybe don't even know that they're talking to an AI, it seems like things are about to get even weirder. Maybe we're just not there yet to worry about that, but how do you guys see that world starting to shape up?

Number one, I think that the intentions of the builders in the AI space are extremely positive. Everyone here is trying to do good work. That benefits everyone. So I think that's like just an important, obvious but important thing to state explicitly. You know, every time there is a technology change, of course there's going to be bad actors and ways to exploit the technology, but we're already seeing a lot of companies, including in our portfolio, trying to, you know, take deliberate steps to make it easy to understand, for example, what is AI content and what is not. Look, I think the third point and what gets underestimated is just this note about media literacy. You know, for a generation that has not grown up with AI, some of these things may feel hard to grok and indistinguishable, just like our parents can't tell what's a spam email. Like you can tell what's a spam email. And in the same way, I think the next generation is going to have more literacy and intuition around what is AI. So I think that the sort of the benefits dramatically outweigh the costs.

Yeah, I believe that too, for what it's worth. I want my biggest fear as an application developer myself. is that we'll end up with GDPR for AI before we've even got the benefits of AI. And that, you know, I think a little bit of self-governance will be very helpful for us to avoid that, or at least kind of postpone it to where, you know, we've built up the portfolio of the actual good things that, you know, can hopefully carry the day. Do you want to put a request for startups out there? Is there anything that you want people desperately want people to bring you?

All of them. Anything that's weird and working, certainly anything that's sort of that we've touched on in the deck, but I think we try to be very intellectually humble about not sort of presupposing what's going to work or not. And, you know, we're just very, we're just very excited to meet more founders and see what they're working on. So yeah, come one, come all.

I think exactly what Anish said, nothing is too weird for us. Anything that is a new, exciting, insightful take, Anything that people are like, that feels like a Black Mirror episode is probably something that we're interested in.

All right. I love it. Weird and working Black Mirror episodes, come one, come all. This has been a lot of fun. I really appreciate the conversation and I definitely look forward to doing it again. We could probably do it, uh, every few months with the rate that things are going. So don't be strangers. For now, I will say Anish Acharya and Justine Moore, thank you for being part of the Cognitive Revolution.

Thank you, Nathan. Good vibes today, man. Can't wait to hear it.

It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.