Scott Belsky & Steve Jang: From AI Chatbots to Copilots

0:04Hey, everyone, and welcome to Generative Now. I am Michael Magnano. I'm a partner at Lightspeed And today on the show, I'm talking to Steve Jang and Scott Belsky. Steve Jang is partner and co-founder of Kindred Ventures, and Scott Belsky is a legendary angel investor. And on the show, we talked about all things related to consumer AI, from browsers to input devices to hardware to the business model of the future internet and many other things. So enjoy this conversation with Steve Jang and Scott Belsky. Hey, Steve. Hey, Scott. Hey, Mike. Hey, Mike. Hey, Scott. What a crew. We got a good crew here that all are interested in a lot of the same things.

0:45What I want to talk to both of you guys about specifically is AI products. You know, in this podcast, we talk a lot about foundation models. We talk a lot about what's coming in terms of AI products. But the thing the thing I wanted to start off with is like the AI products haven't really showed up yet. There are a couple of them. But like, for the most part, we're all just still talking about ChachiVT and Claude and perplexity. So where are the products? What's going on? Any theories here? Where do we begin, right? You know, I think that there's, you know, all three of us have had a history in consumer products.

1:20And so we know that novelty does precede utility. We also know that it's these like crazy, weird things at the edges that may someday become the center. and yet here we are where every consumer AI experience is basically a chat. And I like to liken this to like the skeuomorphic phase of the early days of the iPhone when everything still had like stitching in the contact book and all these sorts of design tools to get us to feel comfortable with this new and modern thing. And in some ways, like I think like the chat is kind of like that. So I'm sure you both are seeing and thinking about other just weird and creative applications for AI for consumers.

2:08I always go back to what are these psychological needs that we have as consumers? And whether it's we want to feel better about ourselves, right? We want to feel loved, right? We want to feel like we are connected to our friends. Like, you know, these are sorts of things that have always been the playbook. And does AI like need to take those into account and come up with these wild new products and experiences? Or maybe not. Like maybe it's all going to be a chat. What do you guys think? I think it makes a lot of sense. I think the one part is I think everyone, all the builders were surprised by the advancement of LLMs.

2:552023 was sort of the year where everyone was trying to figure out, you know, what's an LLAMP? How do I use it? Why is it hallucinating? Why is it, you know, outdated by a year in terms of, you know, recency on information? Why is inference cost so high, right? So I think everyone was trying to figure out how to orchestrate them, how to use them. And so there's a little bit of that. I think also the skeuomorphism that you're talking about, I think that's natural and that's good, right? I think the idea of natural language models to be able to process that, it requires this text input. And I think the speech model area gained a lot from that, right?

3:40And with text-to-speech and speech-to-text capabilities. So they really took a hold of it. And there was a lot of cool stuff in consumer and enterprise there, right? Call centers, and then some early consumer apps there. So I do think that there was some consumer innovation right out of the gate, but on the more generalized LLM space, there wasn't sort of a ton of like social consumer stuff, right? We saw Meta ship some early stuff that didn't land right. We saw a bunch of other stuff try to get sort of inserted into, like bolted onto existing products. And that was also, didn't land right, right?

4:19And so a chat was very much naturally the right interface. And so I think it actually followed the right path. And that's a good path. I think now that the models are more reliable and consistent and there's higher quality outputs, we know how to work with them now. I don't think anyone, I was talking to some people at OpenAI that they said that internally it was quite a surprise to even ship chat GPT. It was sort of a command from on high to ship it immediately. And people weren't really preparing or expecting to do that. And so when it was very popular, it was a great surprise even internally at OpenAI.

5:00So I think a lot of this is due to just being surprised, right? And things happening very quickly and then people figuring it out while it goes. So if we could, one year from now, when our LLMs know us almost as well as we know ourselves, based on all of the information and the context that it's gathering, as my friend Michael McNaughton likes to say, context is king these days. Will we want to take a version of ourselves represented by AI based on all it knows about us and then put the three of us in a simulation environment and have us discuss this question? And then we can just like observe it and, you know, have plausible deniability.

5:42Like if I say something that's like super egocentric or conceited, like you can I can be like, well, sorry, like that wasn't the real Scott. And, you know, you guys can we can watch the debate ensue and like live in some ways vicariously through some like other version of ourselves. Like, is that the next consumer social experience? I definitely think there are people trying to drive things in this direction. I don't know about you guys, but I've seen a couple of products, a couple of flavors of this. One flavor I've seen is where, you know, these products, these companies basically ingest like influencers and authors, famous people.

6:17They ingest all their content, all their writing, their videos, whatever. And then you almost have like a proxy of these people where you can ask the experts questions. Another version of it, which I've seen, which I think is kind of thought provoking, is almost like the immortality angle. They talk to you more of like a consumer, not celebrity use case. Maybe they talk to an elderly person as like a biography storytelling platform. And then the ultimate output is a replica of the human being that you can then talk to forever. Right. All the stories that this person had throughout their life. So I definitely think people are working on this.

6:51I'm personally not convinced the models are good enough yet. But per your point, Scott, maybe in a year from now, they will be at the current rate of scaling. I think this is a problem space people are going after. You know, just to like even drill down a little even tighter underneath consumer AI is personal AI. I think that's what you're talking about in this concept of personal AI. I would add to Mike's saying of context is king. I would add context and memory are king. And this is really important, right? the personalization that you get from that, you know, having a long-term memory as part of your experience, not just in like one application or one agent, but across the internet in everything that you do, right?

7:34Whatever device you're using, whatever application you're using, if you have personal AI, you know, let's just define that for the sake of this conversation as what is that AI that is constantly reinforcement learning based upon not only the models and And that's not model centric only, but really learning about who you are. Right. And then you having control and input and being able to take that wherever you want it to go. So some of those product ideas that are being worked on, one that is really interesting to me is the idea. And I think this is related to the first example you gave a Scott, where a proxy for you, right, to be able to be sleeping or to be busy in a meeting and people can ask you questions.

8:20and I was listening to a podcast by Claire Vo. I was talking, it's a podcast about how you use AI personally. And she was talking with some, I forgot the guest name, but they were talking about how they created custom GPTs to reflect their own expertise and how they would answer to a certain report of theirs in their company. And I was thinking about that. I was like, that's pretty interesting. I wonder, you know, they were saying how it worked pretty well as a replacement alternative to them when they needed to ask a question in real time. I wonder how often you want to ask your boss something in a job and you don't want to ask them.

8:58But if there was a proxy, to your point, Steve, of them, that was very approachable. And maybe at the end of the conversation, you could even get the option of being like, can your boss see what just happened? Or do you want to hide it? It'd be really interesting. It's like you're practicing or you're asking without privacy disclosure. I knew that with my wife. I don't know. I feel like half the things I bring up, I get in trouble. So I got to, I got to show a portfolio company here. Hope you guys don't mind. So granola, I was having a conversation with the founders a few weeks ago and they said there was a very, very famous CEO.

9:32I won't name, I won't out the CEO said to the granola guys, Hey, like I would love to be able to have my team prompt all of my granola notes. I'm like all of my conversations, which by the way, you, you can actually do this with granola and then just have my team be able to ask me stuff without without without you know without me having to attend you know 50 100 different meetings now obviously like the the danger of this that that still needs to get ironed out is like how do you have certain aspects of conversations like opt out of the model right because like you may not want everything exposed to the chat right what about like the chatter that happens before like the real conversation like scott you and i were talking about you know friends we just saw recently like we may not want that to be in the context.

10:14So I think that's a challenge. But I think whichever platform captures all this offline content, I do think will serve this use case you're talking about. I think that's like very realistic in the near term. You know, if you were to take this into the consumer world and say, hey, dating, right? Think about like the Tinder and Match.com and all of that hinge, like all of that was based upon going to meet in real life and kind of going into this sort of like cold and hoping that it works out. Imagine if there's a way to be able to get to know someone in advance of that and have that experience in a preliminary basis.

10:52I got married before I ever got to use a dating app. So I'm speaking a little out of turn on that evolution, but I would feel that that would be a super interesting, broad, mainstream, fun use case to be able to sort of chat with a young a young scott belski or young mike vignano and uh figure out if uh if they're if they're cool if we're a fit if i agree with their politics or or uh you know things like that so i think there's something there with this i guess the question is why do why do people swipe do they swipe to actually meet their soulmate or do they swipe for that near-term dopamine hit oh this person matched with me they like me you know if the former i think it would work great first mile of relationships is like a really interesting because if you think about it the first mile is actually probably the most artificial superficial one totally that is written with thin slice judgments and like bad you know uh or very shallow assumptions and so if you could kind of deploy your you know your simulated beings and have them have first mile conversations with tons of different people and then the ones that are interesting and get dynamic like you can get brought into.

12:04And also, when Steve was talking, I was thinking, I thought you were going to say wingman as a service. Having the AI sidekick that totally pumps you up. And you're being shy and everything as you're dating. And then the wingman, wing person, wing woman comes in and is like, you should know. Steve is amazing. He's this and he's that. And you're like, please, please, please. That's actually a great idea. It would be so easy for them to bolt on. Totally. Right. And then, and it's like that, but that's like, we're getting into the social consumer mindset, right? You know, these are the things you want people to know about you, but would never want to tell them.

12:43And so if AI can sort of augment your engagement with someone you're talking to by providing all the stuff that you would never say about yourself, but you want to be said, like, that's an interesting angle. Yeah. I think you can even reason by metaphor analogy now and say, you have friendships that have different sort of functions for you, right? So there are friendships that are about being able to talk and be heard, right? And to be honest, and there's certain friendships where you're seeking advice, essentially, right? Running, you know, running ideas by them. And then there are other friendships where it's like about culture and fun and exploration, right?

13:26Like I have friends that I talk to about music and almost exclusively after we, you know, finish the small talk, we're just talking about music. Like, what are you listening to lately? Like, you know, what artists have you been, you know, putting into rotation, you know, they're DJs, producers, or just music lovers. And then there's others that are, you know, we work with in tech and we talk about, you know, ideas and AI, or we talk about software, hardware, and then there's others there, you know, just like, hey, I'm having trouble with this. How do you handle this? What do you think? And in a very personal AI perspective, we just haven't seen a lot of exploration yet of people doing the very skeuomorphic, very sort of like, you know, anthropomorphic version of that, which can accelerate you in the same way that co-pilots accelerate you on code 247, the way agents accelerate you when you're async and not paying attention and they do the multi-step work for you while you're away.

14:20We don't have those purposeful friendships because of personal AI yet. And that to me is low-hanging fruit. It's a really good use case to actually have a product that has a real human proper name. I kind of get bothered by SaaS products that use a human name. I'm like, what? Well, like your product. Yeah. I'm like, that's just really corny, but, or like it's cornier than, you know, Western companies using Japanese names for their product. But I think that personal AI aspect, like, let's just look at like all the different types of relationships that we wish we had more of, that we had more knowledgeable, more powerful, more amplifying.

14:57That's a huge opportunity right now. Let's talk about memory. Let's go back. We touched on this a little while ago, and I feel like it's something that is getting discussed. I feel like every day on Twitter, I see a tweet that's like, memory is the ultimate moat. I have a bit of a reaction to this. Number one, I think in the near term, I feel like it is not yet the ultimate moat. I don't know about you guys, but I'm constantly switching back and forth between different models. I feel like certain models are better for other things. So the moat has not really established for me yet. I'm curious if it has for you guys.

15:29But secondly, despite like how in the near term, I don't think it's a great moat. I do think it's going to be a very, very strong long-term moat when you could take it sort of portably outside of the model and onto different websites and services. and, you know, imagine log in with ChatGPT or log in with Claude. And all of a sudden, these different services and stores and platforms will have all of my preferences built in through the memory. I think it's something you've been talking about for a long time, Scott. How do you guys think about this? Like, do you think memory actually is a near-term mode?

16:03Is there an opportunity for memory to more entrench, I guess, the leading models right now once they get sort of opened up through a login capability? First of all, I know that Steve also has been thinking about this. And I think there's a company that both of us are supporting that's exploring some of the memory stuff. But I think that one of the things I would say is I like to back out of where I believe the future needs to be. And so if you think about the future a few years from now, every website should welcome you by name, know your shoe size, know everything that's in your closet, know what shoes you purchased before, know that you're a vegetarian, know that you like to run this many miles a day.

16:39everything that you want every entity in the world to know about you, it should be able to know about you as you have these sort of agentic experiences with brands. And so if you believe that that's true, and that currently the world is sort of generalized for all of us, and in the future, the world will be personalized for each of us, then the question is, how do you get there? The only way to get there is for the context window of our experiences, but also just the memory across all the various LLMs that we are encountering and the agents in different places. like the memory has to be stored to some degree in a centralized profile.

17:13And that's for the benefit of the consumer, but also for the benefit of the brand. I mean, if Delta knows that I'm a silver medallion flyer, that's good. But if they don't know that I am global services and United or that I fly JetBlue Mint a lot, or like those are, if you think about it, status has never been portable. Status has always been owned by the brand, even though we earned it, not the brand like we did. So again, if that becomes portable, that's great for brands. And so I actually believe that brands are going to want portability of memory for their own benefits. Consumers are going to want portability of memory for their benefit.

17:49The one player that won't want portability of memory are the LLMs. They are going to want to own that, right? When you see OpenAI come with their login with OpenAI, they're trying to get in that as well. I just hope that the solution is better for both ends of the users and not just the LLMs. I hope that that's the winning path. Yeah, the opportunity to have that portability or that agnostic relationship with data and all your services, it's got to happen either on a third-party, independent, sign-in, like an OAuth for AI type of service, or it can be operating system level, right? But what's incredible is that the operating system makers today, and there's really like two that matter, haven't really like worked on this area.

18:39I feel like a lot of the operating system companies, Google and then Apple and then Windows, right? So you have, I think on the Windows level, I'm not expecting there to be something really creative there on that side. I think they've always been a fast follow on whatever the new technology approach would be. And that's fine because they're really appealing to businesses and enterprise. But when it comes to consumer, I feel like with iOS, macOS, and Android, there's a huge opportunity to do that. But, you know, Google's trying to get Gemini, and they're doing a great job of it, by the way, Gemini and DeepMind to catch up to OAI as a frontier model lab.

19:22On the Apple side, you know, I think they're chasing how to catch up across the board, right, on every level. But it's either going to be a third-party independent company that offers a LLM agnostic memory service and layer that allows for that personalization and consumer that we all need, or it's going to be operating system level and we just haven't seen it yet. So the field is wide open on that. I wanted to go back to something that Mike said too, which is when you talk about a memory, you know, long-term being very, very important. I think today, you know, we haven't experienced what memory feels like, true personalization feels like with LMs.

20:06Like we're prompting, I think one of the reasons why we put so much context window and so much work into prompt engineering is because it doesn't know about us. It doesn't know our intent. It doesn't know our context. It doesn't know about our preferences. So I feel like we'll go through a step function in experience, user experience. and the value that all of these orchestrated AI agents and applications will provide us once we have that memory. So I do feel like if we can solve that in the short term, it really makes a huge difference. Not a minor difference, but a huge difference. I actually wonder, and I realize this is a cynical take, if this actually ends up going in the opposite direction of what you guys just laid out and ends pretty badly for the consumer.

20:53So I think if an independent party, like you're talking about, it can establish this portable layer that that is very aligned with the consumer's incentives, right? Perhaps we pay for this third party service to host this memory and we get to bring it around to these different services and have these services selectively pull from it to the best of our liking. That's really good for us. Right. But I think if the model, you know, you know, just going to pick on open AI as an example, like if the model layer ends up controlling the portable memory, which I think they naturally have a good chance at doing, given that they're like the leading, you know, they're the leading platform for this type of behavior right now.

21:31They're becoming, you know, there's a world in which they become the dominant search engine, let's call it. We probably just end up with another repeat of the ads model in which our data is sold to it, like to such an extreme degree, probably more so than we saw in traditional search, because the memory is that much better. And what I mean by that is if let's say ChatGPT effectively replaces the search engine, I have to imagine they're going to get into an ads model where they're going to start taking bids for driving, you know, for returning certain results on search terms. And if they also own the memory layer and see and have visibility into all the different places we're browsing and the stores we're visiting, I think the quality of those results are basically going to be driven by whoever pays the most in terms of the ad dollars.

22:19And we're actually, again, the consumer is probably going to end up with like the worst outcome there. Do you guys see that as a potential risk here? And how does that not just automatically happen, given, again, like the open AIs of the world are sort of running away with the new sort of AI search market? Is this sort of about the fate and the future business models of the LLMs? I think it has a lot to do with it, right? Right. Yeah. I think I think I think it has to do with the fate of their future business models, which if they become the dominant place that people search, it's going to be an ad.

22:51It's going to be an ads model. Like, I don't see how it's not an ads model. Yeah, it has to be. I agree with that. And the best thing to feed that ads model is all the data about the user. Right. Like every every piece of information about the user these models could get, they will want. And so they're naturally going to want to own this memory layer. And they will fight like hell to get it through things like a login, you know, login with OpenAI, etc. It's a great conversation because we're kind of articulating the chess as it is right now, right? We are debating the interests of the LMs versus the interests of the end users on both sides, the brands and the SaaS tools and the end users that use them.

23:34Those both sides will want portability of memory and they won't want to be all reliant on only one LLM and one agent from one LLM forever. Structurally, though, I'm optimistic. And the reason why I'm optimistic is that most of these LLMs, most of these model labs are providing inference through API. It is not necessarily a full two-way street where all that data is enriched in both directions. So, agents and applications and these products, whether they're hardware, software, downloadable client, web app, doesn't matter. If you assume that there will be products that will be independent of those LLMs, they will actually collect much richer data about people.

24:19Um, if you assume that there will be monolithic, like one or two, uh, products only in AI and they were owned by the owned and operated by the LLMs, then yes, that dystopian future that you described where everyone's like hyper investigated and overly understood Michael, uh, and everyone's being served ads that are hyper personalized, but we feel like it's like a little creepy and too much and we don't have choice. Uh, that's a dystopia to me, right on the internet. But a utopia to me on the internet is that there's a diverse set of product companies that are actually collecting much more rich data about everyone and competing on fairness and terms of service and privacy, as well as quality of product.

25:09And they'll actually have a much more rich memory contribution and holding than the LLMs. Because the LLM APIs that they're using are abstracted. They're using inference. They're storing information on their abstraction layer as a product company and not necessarily sending all of that back. And this is the opportunity for open source models. So for consumer apps, we have a lot of consumer AI companies in the portfolio, but we also meet with 100 times more that pitch us. And one of the things that's so great about these conversations is talking to them about how they feel about open source and closed source models.

25:50I actually think open source models doing well is of great, great advantage and benefit to consumer AI companies and consumers, right, for people. We want open source models, not just because we want those LLMs to have choice on LLMs, but we also want open source because it allows us to store locally and not unintentionally create these monolithic entities that feel dystopian, right? So I don't know, I want to plug for open source models again. Plug for open source. So we've talked in this conversation, I feel like we've talked about sort of things kind of being rebuilt. We talked about sort of the interface and how we're sort of back in the skewmorphic phase to compare it to the mobile era.

26:36We're talking about how the business model may or may not be rewritten of the Internet as a result of AI. We got to talk about hardware. I feel like right now there's another opportunity where hardware, especially for consumers, may be rebuilt. Obviously, Steve, you've been investing in AI hardware. You're one of the earliest investors in AI hardware, obviously with Humane. OpenAI just completed its acquisition of Joni Ives I.O., which, you know, we believe they're working on a new hardware product. I guess my main question for you guys to kick things off is, is it possible for a startup to compete in this space?

27:09Can a startup emerge that establishes the next ubiquitous consumer hardware product in the age of AI? Do we even need a new hardware product in the age of AI? What do you guys think? I go into AI hardware investing with my eyes wide open. I had worked at hardware companies before and it's, it is, as they say, you know, hardware is hard, but I think, you know, hardware and software are both hard. Getting it right when, you know, the number of winners in hardware that become sort of global products that are household names, it's very, very thin, right? It's, it's a very small number, but as you can see, the prize is gargantuan and it's amazing.