Subscribe
Share
Share
Embed
Fraser & Nabeel explore what it means to build great products in this new world of AI.
Two former founders, now VCs, have an off-the-cuff conversation with friends about the new AI products that are worth trying, emerging patterns, and how founders are navigating a world that’s changing every week.
Fraser is the former Head of Product at OpenAI, where he managed the teams that shipped ChatGPT and DALL-E, and is now an investor at Spark Capital. Nabeel is a former founder and CEO, now an investor at Spark, and has served on the boards of Discord, Postmates, Cruise, Descript, and Adept.
It's like your weekly dinner party on what's happening in artificial intelligence.
Three Layers of AI Coding, Reviewing Voice Note Apps & A Wishlist of Products
===
Fraser Kelton: you mentioned in the past that you've, built a , trading card game with your, your son. You use all AI to generate it. That's too sophisticated for us. We could use a, a little self-contained project where we are having to explore DallE and we are having to use Claude and we're using all these other tools, uh, to help reach a fun outcome together.
It's part game, like part education is something that we could do together. I'd love that.
Nabeel Hyatt: I've complained for a long time about how programming is taught even in these kind simpler mediums like Scratch and how taught. I
just wanna know like, why isn't there a startup that teaches programming to kids that approaches it more like liberal arts than math?
Like,
Fraser Kelton: Hmm.
Nabeel Hyatt: loves programming because they did paint by numbers exercises to build the game just the way the told me to build the game .
Nabeel: ---
let's get started. Uh, you
Fraser: let's do it.
Nabeel: Uh, why don't you open this time? I always open. You should
Fraser: yeah. But the fans like that.
Nabeel: the fan
Fraser: the fan. Okay. Hey everybody. Welcome back. It's Fraser.
Nabeel: and Nabeel, welcome to hallway Chat.
Fraser: Welcome to Hallway Chat. Why don't we start with a product that you've played around with this week? How's that sound?
Nabeel: It sounds good.
Fraser: you explored anything?
Nabeel: Uh, it's, it's slowed down a little bit actually. I, I was thinking we. We were talking in December and felt like there were gonna be like a dozen new product launches and app launches to finish the year, right? And, and it didn't turn out that way. It kind of petered off right after we had that conversation in early December, which only makes me think that a bunch of guys and, and women kind of, kind of had a couple bugs, pushed off a couple things. And we're gonna have a really good February. Uh, uh,
Fraser: going to be roaring.
[00:00:57] Wave AI
---
Nabeel: Be rowing. the one I've probably used the most recently, is a company called Wave ai . This is in the, um, listening, uh, an audio recorder for the iPhone where you press record, you can ramble into it for a while, and it turns those into digestible notes, which I know I've talked about before. um.
Fraser: You love these things?
Nabeel: I know it's my look, it's my to-do software of 2023 and 2024. Uh, back when the app store came out, I, I, you know, you rifle through new note taking software and new to-do list software every quarter. And I feel like I'm in that phase. I've tried in this category you can obviously use chat CPT and just talk into it and then say, turn these voice notes into a a,
Fraser: Mm-Hmm.
Nabeel: you can use.
There's, you know, audio pen, voice Pal, Oasis, there's a bunch of these guys. I, I find that there's a variety of problems with each of them. Um, but specifically my issues are I want to be able to insert custom prompts so that I can get specific responses back. I wanna be able to save those prompts. it needs to be long form.
Sometimes it's me and another person talking for 45 minutes to an hour. And many of these products cap out at 20 minutes. 'cause that's, uh, for a variety of, uh, technical reasons. Is the, is the way they, is where they cap out. Um,
and the last thing that Wave does that's really cool is you can actually have it, drop it in a phone call and listen to a phone call as well, and then take notes directly on the phone call.
Fraser: How the heck does that work?
Nabeel: it dials you, you have it dial into another number or you have it drop in.
Fraser: I see,
Nabeel: Um,
Fraser: I see.
Nabeel: and, and so you initiate the call from Wave It calls, and then you has,
Fraser: I get it
Nabeel: and it gives back. I think it's a one man band. and so it maybe is part of this conversation about the future of software and what things are gonna go on.
And there I I, I admit like there are 20 of these things. I don't know if I'll still be using Wave in six months. But there's literally no other product that does a good job of summarization, especially long form summarization. and it allows me to ramble back and forth, which, you know, I'm not a structured thinker, but I'm talking for a long time.
For me, this is usually a commute tool. It's like I have just gotten out of a wonderful meeting with you all. I'm driving home and I've got a bunch of thoughts that I wanna try to talk through. And so I just talk for 20 or 30 minutes and just ramble for a little while on a subject, and then I wanna make sense out of it later.
And I'm still trying to find the
Fraser: does it do? Does it, so you, you, you talk, you talk for 20 minutes about
your thoughts. Um, you probably meander, some are high level, some are in the weeds. What do you get back?
Nabeel: So what I really don't want back is a transcript, uh,
Fraser: Yeah, of course
Nabeel: uh, you know, that's completely useless to me.
Fraser: it would be a nightmare.
Nabeel: will only remind me of how much wasted words there were as I was trying to figure out what I really meant. You and something does a good job of summarizing the concepts in a way that's not too simplified, that doesn't treat me like an idiot.
And that usually takes some amount of prompt engineering that quite frankly, I have to do, right? the service has to be pretty good at chunking a long set of data so that it doesn't run into the missing middle problem. But it also to allow me to kind of like iterate, iterate, iterate to get what I want out of it.
It's still missing some basic things, like two basic things that's really missing are I. Um, one, I can't have a set of saved prompts, so things I go back to over and over and over again. So for instance, like,
Fraser: Hmm
Nabeel: hey, this was a
Fraser: mm-Hmm.
Nabeel: I was talking about and I have a specific prompt that I've tuned for that.
Or, Hey, this is a, some thoughts about a company or a founder that I'm working with, and I'm trying to just kind of think out loud through what to do with them. So I want a different prompt for how you summarize that, that those subjects, it doesn't let me save those prompts. That's a problem. And then the second one, which kind of gets into how I've been thinking about software generally in, in this new age, is it's not great at import export.
Obviously the point of most of these things is, I'm gonna use this 'cause it's a, best of breed piece of software at a very narrow and specific thing. And then I want that software to be able to take this information and put it somewhere else. I want it to very quickly get it into Zapier or I want to very quickly get it into reflect, which is the current note taking app that I use, or drop it in a notion or throw it into Slack or wherever it's supposed to go to go do the other things.
And it still feels very much on an island where, you know, I manually can copy and paste and share and that kind of thing. It does feel like it should be, it should feel like it's a good citizen with the other software that I'm using at this time period. And it hasn't quite integrated and that's certainly been a nudge that I think every piece of software needs right now.
If we're gonna have a bunch of agents in AI doing a bunch of automated things, then they all need to be able to talk to each other, which means. You know, you don't wanna be the last piece of software stuck on the island that doesn't have any bridges to the rest of the things that a, that a user is using.
[00:05:47] Adapting Behavior to Leverage AI Products
---
Fraser: Yeah. Yeah. Not, not to go down the hole on this conversation, but one of the interesting areas that we're starting to see some startups tinker around with is creating the infrastructure on the web for agents rather than humans. Um, and I thought that that's pretty compelling. Do, do you, so one of my observations is that I have had to change the way that I interact with AI enabled products to get the most out of them. like I, I can shift my workflow a little bit if I want to get some automation out of it. Do you find yourself having to. Change the way that you, um, are thinking aloud so that it fits the, the, product experience that wave and others give you in terms of synthesizing and making sense of your thoughts, or do you just ramble
Nabeel: I, I need to be able to just, I. Ramble, and I know what you're talking about, Frazier, it's like most of the, this new wave of early software. You can't just behave the normally way, the normal way you do. It's not smart enough yet. And so you have to kind of be okay, just like if you hire a new intern, and you're trying to teach them, you know, the first few months you're, you're probably losing 10% here and there in efficiency in order to gain another 40% in a different direction.
And you gotta go check their work. You gotta check on on 'em, how are they doing? and most of this AI software that we're trying to integrate in our workflows and kind of basically live in the future means we're living in the future in a way that is like, you know, a little bit, uh, bumpy along the way.
And you, it's still net positive on time, but it is not, uh, it is not without some behavior change. I'm not willing to do that for.
Fraser: Hmm.
Nabeel: For thinking out loud, because the whole point is to be stream of conscious. If, if the whole point is to allow me to just talk for a little while and, and then, and if I'm suddenly having to change how I speak or change how I'm thinking, uh, I'm, I, it will, you know, it'll gum you up.
You can't be in flow. And the whole point is to allow yourself to be in flow and then let the software do the work. So it doesn't mean I don't lose time, it's just the time is on engineering. The prompts to be able to pull out from my rambling things that
Fraser: Right.
Nabeel: not over summarized, but not over detailed, like at the right level of fidelity.
Um, that's where I lose time.
I'm not willing to change how I
Fraser: I, I bet you, you either have subtly started to change how you speak, how bold of me to to dictate that. I think that you're changing how you speak despite you saying otherwise. Uh, I bet you, you are in, in ways that you don't necessarily perceive or that you will. Um, and, and it'll be things like you will reiterate the, uh, importance of a point. Um, or you will, come back and say, this is relevant, and you'll give it hints. You'll say like, oh, yeah, that's a great point for what I said, uh, earlier, in this. And you would not do that if it was just you in the car because you have an AI companion helping you work through this. You're dropping these little hints or you will start
Nabeel: The one thing I think all of these transcription models, and I think about this in audio and video and all of the things that we're doing right now, is that we're very much, most of these models are still in a previous, um, world of. All, all of these models are still not taking in all of the signal that they have to figure out
Fraser: Mm-Hmm.
Nabeel: working.
Like the whisper model is just doing transcription. It's not picking up intonation, it's not picking up the gaps,
it's not doing any of that work. And the truth is that like I think we, and, and also by the way out, the output models are very similar, right? I was using 11 labs and hey Jen, and a bunch of the kind of voice tools over the weekend to test to see where they all are and kind of play with them all.
And there's very little control of the intonation of the output. I, it's, there's no easy language for me to say, please say this sentence and be a little more surprised. We still haven't
Fraser: right.
Nabeel: UX and UX for that. And on the input side with the same problem, if it's listening to my voice, it should be able to figure out when I'm like a little more excited and a little bit more into it, or like I pause because I'm really thinking through something and then I come out with something that's like slowly thought through, which might say that I've thought through that a little bit more.
All the things that humans do, like we just, it feels to me like, and all the multimodal model development that I've seen, uh, the models that'll be coming out in 2024 pretty soon all still feel like that too. Like, they're taking some images. They're taking some audio. But it doesn't feel like we're at this point where we're ev anywhere near taking in all of the fidelity of data in video and audio
Fraser: Mm-Hmm.
Nabeel: that we could
Fraser: Mm-Hmm.
Nabeel: in, in order
Fraser: Mm-Hmm.
Nabeel: signal.
Fraser: uh, um, yes. I, I, I am, I am now all of a sudden thinking about how we have adapted how we search, right? Like the, we go to Google and the search syntax that we use is like a very Google Ease.
I don't, uh, ask it in natural language. I ask it in, uh, punctuated words that I try to nudge the search engine into the right direction. Which feels, similar to what I do with AI enabled products, where I have to change my structure a little bit to be able to get them to go where I want to go. Do you think this is just a short term moment?
Uh, or is it gonna be like a search syntax experience like 20 years later? We're still now Googling with a very jilted set of terms because we tried to jolt the algorithm where we want it to go.
Nabeel: You know, that's a good piece of pushback. If I'm a founder making software, the question is how much am I meeting? I. My consumer exactly where they are with their very specific workflow that they work in right now. And how much am I asking them to behavior change or adapt to the way my software wants to work?
Fraser: Hmm.
Nabeel: I think you can ask for users to change as long as the payoff is great.
Fraser: Mm-Hmm. for sure. For sure.
Nabeel: Yeah, I, I think, I think you can ask for users to change as long as the payoff is great. I think the one exception is I call tools that rely on flow. So like
if you are painting or drawing or, or speaking thoughts out loud or playing a video game, that there are tools that rely on flow. And in that situation you are trying to construct the reality around keeping a user in flow.
And that's very different from the kind of like, stop and start. I wanna do a thing, I need to send a quick email. I need to click on this webpage. You need to buy a. Purse, all of those sets of like the 90% of the things that we're doing in life, um, I think it's, it's fine to ask for behavior change, but when you're in flow, you're in flow and you don't want anything to mess with that flow
[00:12:12] AI Enabled Software Development
---
Fraser: So this is a, a very natural point to branch into, discussing the future of AI and software engineering because, you, you have very particular interests and desires out of your, um, voice recording note app that I think, as you alluded to earlier, like we, we've seen this with task-based, uh, to-do lists, on mobile apps five, six years ago, everybody had a, their own personal preference. How do we think about the world, uh, of software development given ai? Um,
Nabeel: Well,
Fraser: we going to have,
Nabeel: of the stack and then move our way down? Right. Let, so let's
Fraser: Let's do it.
Nabeel: let's start from what is the to-do list, app of the future feel like, or the voice recorder app of the future feel like in the world of ai. And then you can move all the way down the stack to literally how do we code, which I know you've been thinking
Fraser: Let's do it.
[00:13:04] Infinitely Adaptable UIs via AI
---
Nabeel: I'll start at the top. I think there's a prevailing opinion from like the VC crew, especially the cynical, jaded old VC crew that the world of software is going away. A venture-backed software is going away because I. Like, everyone's just gonna make their own little app and it's gonna be great for cottage businesses, you know, that are selling things on the app store, but they're all gonna be small.
It's possible that that's true for categories, but let's just take a very, just for the sake of like intellectually thinking it through in really specifics versus abstracts. In the voice app software market, you could look at that market one way as a validation of that idea that, look, there are 20 or 30 competitors, none of them are that different from each other.
So it's great to build a small business, but maybe not a massive public company. And,
Fraser: Right.
Nabeel: and if that happens, that happens, great. Not everything. It should be a venture-backed business. I think founders should do what they need to do for their company, and that's right. But the other way of viewing it is if you have the best set of tools inside of a single voice taking app. Can you use AI to make that a much more modular piece of software, a much more liquid piece of software where the thing that would've made me change from one piece of say, note taking software or to-do list software to another one. Is that something that I could just now ask for or code in, right?
Fraser: Mm-Hmm.
Nabeel: Is, is the software modular enough that I can just type in a natural language, like, no, actually I want, when I open up the front page, I want it to give me this first. And then by the way, can you also add this new field that I can fill out for every single time that I added to-Do list? And that world of real modularity at the application level.
I don't think people
Fraser: Mm-Hmm.
Nabeel: experimenting with, you know,
Fraser: Mm-Hmm.
Nabeel: they want their own note taking software or their own to-do list software. And maybe the answer is they all . Do get their own. Like, not that there's five winners or 10 winners, but that there's like literally a billion winners.
And because there's a billion winners and everybody's unique piece of software its own thing, that will aggregate to literally the one company that's building it. And so you end up with one piece of note-taking software that everybody uses. It's just that the surface area looks different for everybody.
I think that's very
Fraser: Yeah. I, I, I can get my head around that. Like, the fluidity of those pieces, uh uh, is pretty simple. my guess is models today could. Handle the, the, the code behind the scenes to make that viable. I'm just thinking now about places where we have seen glimpses of this. We heard a lot of promise some number of years ago that, uh, everybody's going to have an automatic website builder. And lo and behold, we have seen certain verticals, at least in certain use cases where people have automatically built websites. But that doesn't mean that there's not, you know, beautiful, uh, websites built outside of those tools. And then there's things like Shopify and others where you can have your online store.
What's the insight there? It probably was commercially interesting enough to Shopify to take on all of the complexity Pre AI to allow anybody to have their own shop front.
It didn't necessarily make sense for people to like hard code, uh, all of the rules around different, uh, to-do list functionality and features, but maybe, um, AI allows you to, to get that for free so to speak.
Nabeel: Yeah, I, I, I just don't pres. That's right. I think that's a very good metaphor. I don't presuppose that we are going to have. There's like, you could take every category of software that people use and you can imagine like three different outcomes. Uh, that would mean different pieces of AI software solve that problem in that
Fraser: Mm-Hmm.
Nabeel: And so you could
Fraser: Mm-Hmm.
Nabeel: is just for intellectual conversation. You could take the to-do list software and you could say, well, it's gonna get solved in the like WordPress format where it's like there's a million to-do list softwares. 'cause now it's super easy to build them and it's open source to build them.
And everyone just at the code level tweak what they want. And
so there'll be a hundred winners, none of them that, that big. You could see it at the Shopify level where you get the to-do list app, no-code, app builder, AI framework where everyone can build their own or you could get the . It's almost like the amazon.com version where there's a single user interface and everything aggregates in.
I
Fraser: Hmm mm-Hmm.
Nabeel: version because everybody has bespoke needs. And if anything, AI is pushing into the idea that you get mass personalization. I could absolutely see the Shopify answer though, that there's a singular company underneath lots of lots and lots of different players.
Um,
[00:17:49] Zapier ++
---
Fraser: right. Interesting. So that, that's the top layer. What's the second layer in your hierarchy?
Nabeel: Probably How does, like how does data move, you know, below the ui ux layer? how does
Fraser: Mm-Hmm.
Nabeel: data move around the la the world, uh, in the future, and I think we would call this, you know, the API layer, the data layer, the Zapier plus Plus is what we've been talking about it internally as, uh, you know, just above the code level.
How are you thinking and working in, in, in software? Um, I. Zapier is trying to do this, but I think we've come across a lot of different AI companies that are also trying to think this way. And in a way, the GPT store, as we talked about, before is, is perfect as a natural language interface to APIs, um, or it's certainly trying to be, what do you think about that layer?
Fraser: I think it's really interesting. I think there's a lot of automation in jobs, that can be handled with a little bit of sophistication layered on top of a Zapier like user interface. I do think that we're seeing a lot of interesting efforts push into this market. I, I, I think this is gonna be great.
Like, I, I think it's going to be a very promising uh, area for transformative businesses to be built. As well as like gr obviously great end user
value. Like, think about how
Nabeel: Zapier, but just to push back a little bit like, but Zapier's also already there already at scale and already plugged into a thousand more APIs than the next guy. And so why doesn't Zapier just win?
Fraser: uh, I think, uh, I think that this new capability is so dramatically. Novel, uh, that
you would build things around it and make decisions for the end user experience that would be dramatically different from decisions Zapier made a decade ago when they started. And you, you end up with a combination of thousands of small product decisions because of that, that it is going to just have a very dramatically different end user experience, uh, and people will find one resonates and solves their problem in a way that is graspable.
And, and, and it could be Zapier. I just think that turning that around, uh, while navigating your existing user base is gonna be exceptionally hard.
Nabeel: Yeah. Yeah. And I think Zapier just announced like a big pricing change this week probably with something like that in mind, although, I dunno the backstory. Um, yeah.
Fraser: Yeah. Yeah. Yeah. Um,
[00:20:09] AI Enabled Software Engineers
---
Nabeel: And then what do you think happens
Fraser: I think that
Nabeel: levels? Like what, at the actual coding levels, what do you think happens?
Fraser: oh, well, I, you know, I think to.
Nabeel: writing Python.
Fraser: Yeah, yeah, yeah. Um, I, I, I am most excited about this layer, and I think that many people, uh, have drank the No-Code, Kool-Aid, um, in the sense that they think that AI is going to get away from software engineering. And I, I just can't imagine that's the case. Like, I think that you might be able to bring some, uh, software writing automation to, uh, to do list app so that you have flexibility in an infinite number of directions for something like that.
But I think most people don't fully appreciate the complexity of modern code bases. Uh, and to be able to do what you need to do, uh, today, you need to have exceptionally capable engineers, working their way through the code. I think that the best metaphor that I have come across, or that I've come up with here is Uh, treat AI as an entirely new capability. I've been telling people this for more than a year now that it is directionally the same as electricity in terms of a capability. And if we think about what electricity did for science and scientists, I think it's the, the right metaphor to lean into. and let me just talk a little bit about that.
Nabeel: Why? Why don't you, why do you use the science metaphor and not? I always like to try to use the. You know, as we get automation in software development, it's almost like, it's like building homes. You start out with, you know, five guys in a GC and, and they're, they're trying to figure out the plumbing and put up the walls and so on and so forth.
And you fast forward 50 years and, nobody does that. 'cause it turns out most people want homes that are remarkably similar. And so you get the modular home building movement. They all mostly feel the same. Everybody still needs a toilet. Everybody still needs, a fence. We can stamp out homes at an industrial level and larger level. And then every once in a while, there's an amazing, beautiful, somebody building the new Empire State Building. The crazy architectural feat of the age. But that's not every piece of software. Everybody builds , and it's not every piece of architecture everybody builds. What do you not like about that metaphor?
Fraser: The building one is a question of simple automation applied over and over and over again, combined with, um, leverage from mechanization. You don't have eight people with shovels in your backyard digging. You have a, a, a, a skeet steer, or you have, engines measured in horsepower, because you used to have to have eight horses pulling the
wagon, so to speak.
Uh, I
Nabeel: metaphor, and this is not, the whole point is like, this
Fraser: No,
Nabeel: this isn't doing the same. This isn't the same thing as doing the same thing faster for industrial
Fraser: that's right. Well, I, interestingly, I think there's elements of it that start with that, but I think the real promises is in opening up the cognitive and the creative Uh, Mindshare so that these exceptionally capable people are able to spend much more of their time doing those things. And so, uh, I liken it to the electrification of scientific tools is going to be the equivalent of the AI ification of software engineering tools. the first thing that happened was they took existing tools like the centrifuge. There used to be like wooden hand cranked centrifuges that scientists would like, sit around, and turn manually. And it makes sense that you should electrify that, right, and you can get, you can, uh, free people up from the monotonous routine, manual work, within a lab. And you also get better results. My Sense is that, copilot and other, code completion tools that have taken away the writing of basic boilerplate and syntax, can be thought of or should be thought of as the, uh, electrification of the centrifuge.
We have an existing tool, the IDE or taking away the, the , the equivalent of turning this wooden cog. and you, you get freed up.
it's October, 2021. We're prepping, or copilot gets
released. We've been working on it for some number of months. I ask, uh, probably one of the most capable engineers that I've ever worked with, if he found it valuable and he cough, she goes, I, I find it so valuable. I would personally pay a thousand dollars a month for it to have it at my job. And I said, what,, what, what does that even mean? And he said, it frees up so much of my time to focus on higher value work that I'm far more productive, and I find it far more rewarding that I would pay for that. That's, that's the stage that we're at. We're at the electrification of the centrifuge.
And then if you you think
Nabeel: you, you, you, it's, it's a bicycle of the mind metaphor versus an
Fraser: yeah, yeah.
Nabeel: metaphor. It, it, it, it will help
Fraser: That's right.
Nabeel: and faster. Not just do, uh, do not just produce faster. You can think faster.
Fraser: Yeah. It's not just think more faster and it's not just, you know, spending more of your time on higher cognitive tasks. the electrification of these tools allowed for entirely new abilities for these capable minds, right? So after the centrifuge gets electrified, the electron microscope comes along and it's not just making the microscope slightly better, it's allowing them to reveal molecular structure.
It is stuff that you couldn't do with the past tools and some number of years later you get the cyclotron, right? Is that, um, the electrification of a tool that didn't exist before, that actually unearths entirely new scientific fields. And, I think that we're walking down a path where we are going to have AI enabled tools for engineers, software engineers, and we are going to have a, a revolution like software is eating the world. Some people are saying that AI is eating software. I think that's the, the two layers above that we talked about where you can have a to-do list be malleable in any way that you want, or you can have Zapier plus plus that connects all your data across much of different apps. And actually, here's an interesting question, like how far do those push down, uh, the spectrum, so to speak? I don't think they're going to go down the spectrum all the way to the most sophisticated, most valuable software that gets written. I think what we're going to see is that AI enabled tools are going to allow software engineers to do things that were exceptionally scarce and rare over the past couple of decades, and we're just gonna see more beautiful, wonderful things get built. if you're not having to do like bug squashing or like migrations and these people are freed up, they're not doing boilerplates and they have entirely new capabilities that you and I on this call can't even uh, envision right now. They're going to be able to orchestrate this technology and this capability into, into ways that we can't imagine.
It's gonna be beautiful.
Nabeel: an interesting nudge there though, if you want to use the science metaphor, and I'm trying to try and draw some analogies here, it's not just about having more time to think it's about a new cognitive capability. The electron microscope isn't just a faster way of doing the thing before it's actually revealing new information.
What is that? What is that analogy in software development? What, what do we expect happen if those kinds of things happens? That sounds, that sounds more like the AI is writing software that you couldn't write yourself that is higher than your ability to write yourself.
Fraser: I, that, that may be the case. Right? I, I, I think the, the challenge is when you imagine being in the lab and you, somebody showed you the elect, uh, the electrified centrifuge, your mind would be blown, right? You
and I would be like, holy cow, this is amazing. We're not going to sit there for five minutes and say, Hey, listen, if you pass electrons, um, through an organism, we're going to be able to have an electron microscope, so let's get to work. I think we're at that phase, it's just really hard to understand all that's going to happen. we might have AI generating code that we hadn't even considered. Uh, the, I'm thinking of the AlphaGo move that no master over how many thousands of years had considered, and then it was creativity and it was beautiful. You know, I watched as the team trained GPT-4 and, uh, somebody's gonna write a book about it. It's gonna be beautiful. And it felt, it felt Nabil, it felt like I was sitting at, in the mid 18 hundreds when they were building the Brooklyn Bridge in the sense that people were inventing new software solutions to help them then go on and continue to train the model.
And, and the, these people are heroes. Greg, Michael, others. Um. that was a pretty special team. Uh, you are going to see people empowered to be able to do not train that type of model, but to do other miraculous software type projects and manage the complexity and the scope and the sophistication of them because they're no longer using.
Listen, I, I, I think we're, we're at the moment where we've just moved away from, uh, a wooden centrifuge, and I think that the next couple of decades at that moment in time would've been very hard to have, have seen how they'd all go. and I think that that's a, a, an aptt metaphor for where we are. And if you free up these ,, uh, these engineers. It's gonna be amazing.
Nabeel: I, the thing I really like about that lens, 'cause it's not, it's not how I was thinking before the. way I had thought about software automation and AI was probably kind of closer to the rept view of the world, the no-code view of the world, which was like,
Fraser: Mm.
Nabeel: know, I, I've been sitting my younger son down who doesn't do a lot of coding and just instead of starting him with the proverbial how to, you know, Python for Dummies book, or how to make games in Python book that I did for my older son years ago, I'm just sitting him down and rep it and being like, just type, let's make a thing and don't worry about it.
Fraser: Right,
Nabeel: Uh, and didn't, you don't even know what this code means. Just copy and paste the code over and then when it produces a bug, just ask Rept or Chachi PT to help you fix the bug and just like work on building a very small piece of software. And so I have been thinking about it as kind of like the starting point for anybody to be able to make software.
And what you are talking
Fraser: hmm.
Nabeel: the complete other end of the spectrum, which is. The best software engine engineers in the world and inventing, uh, new
Fraser: yeah.
Nabeel: sitting next to their AI partner. The same way that in AlphaGo you see new test moves. You see new moves that you've, uh, in go that you've never seen before.
and that's interesting. I think those are probably, I, I think the, the way you architect, huh? I think the really interesting thing that that makes me think about is that, you know, it's not just one AI coding tool. I think every layer of
Fraser: No way.
[00:30:49] Different Products Across Each Layer
---
Nabeel: just talked about is probably a very, maybe even entirely different models, but certainly very different interfaces to say, what am I doing for a beginning programmer or a programmer that has it absolutely has never experienced programming before?
And how am I getting them to full quickly and letting them iterate quickly to build a piece of software? that's wildly different from, I'm one of the top 5% programmers in the world and I want a friend who is working on me with this software and is gonna find novel new ways to do a thing.
Um, and that's both of those two things at the coding level are still wildly different from the other layers of the stack that we talked about and how you use coding to solve the Zapier plus, plus or above that, you know, modular software development and the no-code
Fraser: Yeah, yeah. A hundred percent. A hundred percent. Like you can even imagine that the no code to-do list that has malleability for any end user's need or desire is actually gonna be built by somebody who's using the tools that give them tremendous leverage to go and build that software. You know, I, I think it's just, it is impossible for me to ignore the most brilliant people that I know in, in engineering. Our echoing or or parodying, the same thing is that the gains that they have even today are so profound. My co-founder at my past startup, I don't think I've ever heard him exaggerate, and he has impeccable taste in technology.
He mentioned a product that he uses, I won't share it, that makes him two to four times more productive.
Um, and if you actually think about
Nabeel: with our, with our, with our one or two listeners? What, what, what are you doing man? We're supposed to be hot handing out hot tips.
Fraser: alright? It, uh, it is cursor. It is cursor, right? It is, uh,
it existing tool, the IDE that has leveraged, uh, a new capability in a fairly, um, obvious way. I don't mean that as a
knock, right? They've, they've electrified the centrifuge and it's making this guy two to four times more productive. And if you consider what that means, it is like rather
profound.
And then you talk to a research scientist at, uh, at Midjourney, and he's using the tool. To do in an hour what used to take him a day
Nabeel: yeah.
Fraser: and,
and it's not, it's not like that guy.
Nabeel: is not just a scrub
Fraser: Yeah, for sure. Right. And it's, it's not, it's not, it's not like your son who's learning how to code because of the AI enabled coding tool
is going to go and do what this person does is that, uh, may maybe in time right as he learns and develops and, and hones his craft and his skill, but this individual now has an entire day that they didn't have before to, to focus on, on the more rewarding parts where their, their intellect is scarce, right? Uh, rather than having to turn that
centrifuge, if you will.
Nabeel: right. That's right. So that's gonna certainly keep happening over the course of this year
Fraser: i, I think it's gonna happen over the next decade. The cycle tron was some number of years after, the centrifuge gets electrified. We have to discover these tools themselves and capabilities need to evolve.
Nabeel: Yeah. This goes back to something that you keep reminding me of, which is that the short term, some of these things will feel slightly broken and we'll feel movement and we'll almost over-index on how much of an innovation's gonna happen in the next. Quarter and over the next decade, we'll probably still under indexing on how much is gonna change.
Fraser: Yeah,
I think
[00:34:06] AI Products We'd Like to See in 2024
---
Nabeel: let's not talk about the next decade. It's the beginning of 2024 and and I, I want to get to, we started, we started chatting over the last week or two on what do we think's gonna happen over this year. And I don't, I'm not a hot take person on, you know, let's make prognostications about what's gonna happen in the world.
So I, I would love instead to just pick one or two pieces of software in AI loosely that you kind of hope somebody takes a stab at and you hope exists over the course of this year.
Fraser: Okay, I, I'll, I'll start and then you go with one. I'm reticent to frame it like this 'cause I've tried it before and people then can't help but be anchored in the mental model is I want to, I want the concept of a, to-do list rethought for a world where AI's capabilities are not where they are just today, but where they're going to be in six months, 12 months, 18 months.
And what I mean by that is, I, I think that that's the perfect place for, uh, these agentic AI capabilities to come in because what is a to-do list, it is a reminder that you need to go and take an action of some sort. And more often than not, I'm thinking of personal to-do list. And more often than not, those, those actions are a series of steps that actually have to be done in order to accomplish the outcome.
Nabeel: Mm-Hmm.
Fraser: I can imagine, um, it's going to be tax season soon and I have a whole bunch of questions. there's no reason why I can't have a a to-do item. That's the equivalent of like, Hey, prep for taxes.
and it knows my unusual situation. I have to file in two countries, all of this other stuff where it can't go and pull in a bunch of information while I'm not present.
So that when I return to the, the product, it's three steps down that path so that I'm not starting from
a blank slate. Um, and then maybe I give it the next nudge in the next direction, and I come back a couple of days later and it goes and gets it. So that's one for me. How about yourself?
Nabeel: So one for me is every time there's a new way of communicating, there tends to be a new social network for communicating that medium. And I'm thinking about, I. Long form video leading to YouTube, and then very short form video leading to TikTok text, leading to Twitter. Um, and the modality of creativity that we see right now is AI art, and it's what I play with on a regular basis.
I think that doesn't mean that I make a picture of a dog and I, and I put it
Fraser: Mm mm-Hmm.
Nabeel: been trying to think about what would be the native storytelling medium that would lean into what Midjourney and Dolly and, and other companies are, are helping us make. And the best thing I could think of is comics, uh, which are a native storytelling meeting that involves a lot of images.
And so I don't, I don't know what it would be, but it would be very cool to see experiments in the direction of . The thing I'm flipping through the kind of modality of the way I'm flipping through the product looks like comic panels and, and all of the inherent tools that go into making that. Not just that I spent five hours in Midjourney to make comics, which not enough people are gonna do.
But you could imagine, imagine it as easy to do as opening up Instagram, having a camera and taking a photo and applying a filter. Something that's as simple as that,
Fraser: Mm-Hmm.
Nabeel: communicate. And I imagine that will appeal to people in a very different way than, I don't know. Um, doing dances on TikTok would appeal.
I, I hope we see some experiments with what native AI storytelling is gonna feel like. How
Fraser: Yeah. Awesome, awesome.
Nabeel: about you got another one?
Fraser: yeah, I have one more. Uh, and it's, it's fairly nebulous, so I don't even know how to come at it. Let me come at it from two different perspectives. First of all, I have kids, um, and it is lovely to watch them discover technology. And so I, my, my eldest was doing some research yesterday and I saw her use, uh, Google Translate for the first time and Google search. She also, um, has these monthly subscription boxes that come in that help her do kind of mechanical engineering, but for kids STEM-like little
projects. We are living through this glorious moment and I would love for them to have the equivalent of a. A Crunch Labs box, but rather than a mechanical engineering project that I get to help her with a little bit, uh, on a monthly basis that there's these little projects that bring us into the, the software and the online realm that we get to build together. Um, you know, you mentioned in the past that you've, you've built a, uh, trading card game or a, a, I don't know what the
category is with your, your son. You use all AI to generate it. That's too sophisticated for us. We could use a, a little self-contained project where we are having to explore Dolly and we are having to use Claude and we're using all these other tools, uh, to help reach a fun outcome together.
It's part game, like part education is something that we could do together. I'd love that.
Nabeel: Yeah, I, I have thought, I've complained for a long time about how programming is taught even in these kind of simpler mediums like Scratch and how STEM is taught. because I, I, I just wanna know like, why isn't there a startup that teaches programming to kids that approaches it more like liberal arts than math?
That's the metaphor I've used. Like,
nobody loves programming because they did paint by numbers exercises to build the game just the way the book told me to build the game. You want
Fraser: Right, right.
Nabeel: that encourages your right brain as much as your left brain and allows the, the thing that your child makes should be remarkably different from the thing that your neighbor makes using the exact same thing.
And if you were both doing a creative writing task, the output would be wildly different 'cause it's from you. And I think it takes way too long in stem generally, uh, and, and certainly in programming to get to the place where people feel in flow. use it as expressive medium, and AI is a perfect example of being able to accelerate that curve and getting people to the fun of making faster.
Fraser: right.
Nabeel: Yeah. Good one.
Fraser: Yeah. Love it. Love it. Okay. You have a last one for us?
Nabeel: My last one is gonna loop back to a conversation we had earlier. I really expect this year, and I hope we'll see stuff that takes the idea of an API framework, like a Zapier plus plus, but punches through to the user input layer. So it, it, it's still, I can coalesce the sets of APIs that I want so I can take my outputs from my notes, from my Wave AI and plug it into my Notion docs and have it all flow into something, which is fine.
But what I, I really think the next step is, is there's no user input in Zapier, right? These are automations, these are things that run in the background. And I'd love to see something that pierced up into the app layer where I can then say, let me
Fraser: Mm
Nabeel: a couple of things, let me do a couple of things,
Fraser: mm-Hmm.
Nabeel: feedback into the product.
And I think that'll just enable a whole bunch of new behaviors that I'd love to play around with. And I hope and expect somebody will come up with.
Fraser: I love it. I, I mean that, that, you asked me the question earlier as to how a group who is building for a world of AI and connecting different tools would compete and differentiate. There's, there's part of the answer, right, is that you actually can puncture up into the UI layer and collect, uh, direct feedback
from the end
Nabeel: I mean, let's, you know, the truth is that as we all know, like founders are actually the ones that are gonna invent these things. And as, and as they're, as they're, as they're building, they're building these things, they will find the five reasons we're wrong pontificating on our podcast. Uh, so I mean, we both been founders before.
We know what it's like to walk the idea maze. So I'm, I'm sure these ideas will shift dramatically in the market. But, but some experiments in this direction just feel like they need to happen. And then mostly speaking of that, as a consumer of this software, like I, I want these things in my life.
Fraser: Yeah. Yeah, yeah, yeah. Listen, I, I, I get both of those is that, we are not going to prescribe how things get solved, but it's, I think it's also okay for us as end users to have wants and, and desires, and oftentimes, as we know, and users don't know what they
want,
Nabeel: right. Hey, uh,
Fraser: not always the case
Nabeel: let's be done for today. It was great chatting with you, Frazier. Uh, and thanks everybody. Um, we, if you have a AI app that you want us to take a look at we'd love to, that's all we do, uh, around here, uh, while we do other things. But, you know, we try to spend our time and energy. I mean, one of the benefits, by the way, for me, of doing, this podcast with you, Frazier, is not only getting your your wonderful thoughts every week, but it really puts pressure to keep using these products and thinking about how these products are changing every week and the patterns of, of software development and UI development and building, and what we're learning from customers as, as this new field emerges.
So thanks for hanging out with me, and I'll see you next week.
Fraser: Oh, see you. Take care.
Nabeel: care.
d