April 11, 2024

Devin & the Autonomous Engineer Wave. Plus, .Com lessons as a model for today

Devin & the Autonomous Engineer Wave. Plus, .Com lessons as a model for today
The player is loading ...
Hallway Chat

What is the extent of autonomous coding engineer Devin’s ability to generate real, functional applications with little to no help? Nabeel and Fraser dive into the buzz about Cognition’s ‘Devin’, what makes it different, and the transformative potential of AI in software engineering, particularly focusing on autonomous coding software. Later, they get into the innovator's dilemma and the lessons the .com era can lend to this new time in AI.


(00:00) Intro

(01:23) The buzz about Cognition’s Devin

(07:44) What makes Devin different?

(12:19) Tolerance for time

(16:07) The interface of the future

(22:19) Innovating around the incumbent’s advantage

(25:30) Cutting edge products mean new user bases

(29:52) Netscape was the Open AI of the Mobile Revolution

(33:42) Optimism as the engine of capitalism

(37:56) The model is not the product

Transcript

Nabeel Hyatt:
Is what you just said to me literally what you told the program, "Make me a Pokemon lookup database thing."
Fraser Kelton:
Pretty much exactly that. I gave it a task. I said, "Create a web app." The first thing it does is it creates a plan.
Nabeel Hyatt:
Much like your coding intern, "Thanks for the advice, man."
Netscape is basically OpenAI. I mean it basically is. It was an enabler to all of this. It was the way you get onto the internet, the way you get access to everything is going to be Netscape and so it just felt like the big company. And 16 months to IPO for Netscape and a $3 billion market cap on day one under 18 months old, it was pretty crazy.
The consumer companies that made it really made it. The Amazons of the world became unbelievably dominant. Cisco is a 200 billion company, Akamai, who haven't thought about it in forever, we looked up in the middle of the meeting, it's a $16 billion company, who knew that? And so for all the talk of application software, AOL is not really around, and Excite@Home is not really around and so on and so forth, and a lot of these infrastructure companies really didn't last.
Fraser Kelton:
Hey everybody, welcome back. I'm Fraser.
Nabeel Hyatt:
I'm Nabeel.
Fraser Kelton:
We are back for Hallway Chat, and we are back to talk about a couple of things that have happened over the past week and things that we've tried. Nabeel, I have to start with what is one of the more remarkable products that I've explored, at least at a V1 stage, in a long time. And I'll preface to say that it's remarkable, not necessarily for where it is today, but because of the glimpse that it gave me of the future. And I'm talking about Devin, the autonomous code software engineer created by Cognition, and I'm sure a lot of people have seen the demos online. We've been playing around with it for a little bit.
Nabeel Hyatt:
It reminded me a lot of the boom of Agentic software last summer, which turned onto largely too early to ship, which we should cover, but nonetheless, it hit all the right synapses.
Fraser Kelton:
Yep, you tell it what you want it to do, you can share links with it, you can upload files and then it goes in and it does it, as simple as that. For writing code and then building software, it is interesting because it is a break from current products and it has its own UI and the product has access to an IDE, a terminal and a browser. And then you have a chat window, which we're all familiar with now, which is the primary way that you interact with the product. You can follow along progress Live or you can go in a TiVo-type scrub back into the past and see what's happened or pushed forward to real time. I gave it a task. I said, "Create a web app where my daughter can enter in a Pokemon card and it will fetch from an API details about that card and then display it on a website and that it should persist so that other people can see the cards in her collection," et cetera, et cetera, et cetera. And you just hit go and the first thing it does is it creates a plan.
Nabeel Hyatt:
So you had this idea, what was your product document or PRD like? What was your actual prompts when you were trying to tell it what is really an application description. Is what you just said to me literally what you told the program, "Make me a Pokemon lookup database thing."
Fraser Kelton:
Yep. Pretty much exactly that. And I will come back to talk about some of the earlier prompts that I gave it because I think that that lets away some of the surprise if you will. And so that's what I said, and it creates a plan that says, "First I'm going do this and then I'm going to do this, and then I'm going to do this and then I'm going to do this." And then it starts trying to do those things. And then if it hits a blocker on step one, let's say, it creates derivative plans to try to overcome the obstacle to then solve the first step. And to make that tangible, the first step was to set up react and you just see it go into the terminal and start doing this and opening up shell sessions where it's standing these things up and then it switches back into the IDE and it starts writing the code. But far more interesting, I didn't specify an API and it went out into the internet and it found an API.
Nabeel Hyatt:
I've used that Pokemon API actually, I know the API you're talking about.
Fraser Kelton:
Yeah. Well, [inaudible 00:04:33] Devin. Go watch the demos. You don't want to hear me talk about it, but the things that I want to touch on before we then talk about what maybe some of the more interesting observations and takeaways from it, my framing is that it's exceptionally stubborn, Nabeel. It hits an issue and it doesn't come back to me and say, "Meh, can't do it." It hits an issue... In one case it had an issue with rendering something with react and so it tried to debug it, it tried to debug it. It tried to look at the code and see if it could figure it out.
And then I saw it switch to the browser and it navigates to Google and it types in the error that it had and it lands on the react FAQ page and then it goes back to Google and it lands on a stack overflow page and it gets some sort of insight and it comes back to the IDE with that and then it writes the code to overcome the issue. And I'm thinking, "This is crazy." It measures the time in minutes. So what is different is that usually we're used to inference or AI enabled features measured in hundreds of milliseconds, in this case it grinded away for hundreds of minutes and you just see it do it.
Nabeel Hyatt:
How long did it take you to make the app for you?
Fraser Kelton:
Oh, 280 minutes, so a good amount of time. I think I interacted with it maybe 10 messages where I gave it feedback or it asked for something from me and I was able to let it go and then get on its way.
Two other things and then we'll move on. This blew my mind, at one point it couldn't overcome a bug. It integrates front end logging, it deploys the product, it uses the web app to generate logs and then goes into the logs to figure out what's going on and from those logs is then able to deduce what the issue is. It goes back to the IDE and writes the code to overcome the issue that it fixed and then it moves on to the next task. And it took 40 minutes of trying different things before it said... You could almost feel it throw up its arms and be like, "Okay, I'm putting it in front end logging, let's see what we can do." And it was amazing.
Maybe the other thing that is worthwhile calling out is it clearly had a bug that I could see and reason through what the cause of the bug was where it had deployed something, it then had called it, it ran into an issue, it tried to fix the bug and then there was a caching error from the past time that had tried it and it stubbornly hit his head against the wall 10 times to try to fix it in chat. I'm like, "Hey, Devin, I actually think that those two things are remnants of a cache issue and if you just delete them, you'll see that you've actually fixed the issue and things will be fine." And it came back and it said, "Actually that's not a bad idea. Let's see what we can do." And it deleted it. And it came back to me, and it closed the loop. Much like a programmer would [inaudible 00:07:28] like, "Hey, thanks for that, it helped me figure it out."
Nabeel Hyatt:
Much like your coding intern, "Thanks for the advice, man."
Fraser Kelton:
That's right. It helped. I'm now doing the next step because of that. All of it felt like a glimpse of the future, not just for software development but for a lot of the ways that we're going to work.
Nabeel Hyatt:
How did it feel... The cynic side is like, "Great, this is all the Agentic chain of thought reasoning stuff that we saw last summer that everyone went crazy over." And so help the cynic listener get over that a little bit. Why are we choosing to decide to talk about this today when the core, very basic concepts, make a plan, go step by step through the plan, don't worry if it takes a lot of time, reflect back on that plan over time, those kinds of things, are now... Again, that's over a year old. What do you think is different about this?
Fraser Kelton:
Because this one worked and the other ones showed promise, but they never worked. Listen, I deployed Natalie's Pokemon card site to the internet without touching a line of code myself, and she was able to go in, add her cards and it saved them and it [inaudible 00:08:39]... It worked. It worked. It actually added an Easter egg where I was able to upload a CSV for us, and then it went and it scraped the Spark site for the images. All stuff that is not sophisticated from a software development standpoint, but it did it and it wrote the code, it integrated the APIs and it worked.
In another example, I had it query the Anthropic API to translate from English to Spanish or Spanish to English, and it was remarkable because it also then figured out prompting. You saw it in the browser, navigate to the Anthropic documents and figure out that prompting was a thing that it needed to tend to, and it adjusted the prompt to work for Spanish to English, which is autonomous, and it worked. It actually created a translation app for me, and I don't think anything that we saw last summer was remotely close to that.
Nabeel Hyatt:
Yep. I think that's right. For me, we should talk about what we take away from this and what we think this means. I think I learned a couple of things from using and watching this product. I think the first is the thing we've been learning overall for a little bit, which is, it turns out coding works really well in LLMs and probably maybe slightly more profoundly, starting a course as a founder to not something that is maybe three or four years ahead in capability of the models and where they're headed, but maybe just about a year ahead. A lot of founders, frankly, seem to chart four years ahead or a lot of times four months behind. They wait until they can really get it working in the model and then they're like, it's a bit tough. This is a good example of a team that seems to have charted about a year ahead of time, nine months ahead of time. It's on the edge of working and coding is one of those use cases right now. They're obviously not the only people now working on this. I think we've come across as a Princeton group, John Yang and Carlos Jimenez and others that have released a GitHub Repository for doing something relatively similar, and I think there's another directly competing project. Of course OpenDevin, which is an open source version of Devin. That also uses similar agentic frameworks and a custom terminal that is tuned for LLM interactions and so on and so forth.
Fraser Kelton:
Well, listen, that's to the side that you mentioned, but I think it's too interesting not to say, because I'm not sure if everybody notices that the Princeton group had great results with their agent that could do full software engineering. And at the end of the day the paper isn't released. They say it's coming soon, April 10th or something like that, but they attribute it to basically what amounts to product work. One of the product works that they focused on was when they changed the terminal to not just be a terminal that's appropriate for a human, but a terminal that is great for an AI interacting with the terminal.
They saw better results and I'm not entirely sure what that means in this context. We'll read the paper and figure it out, but it was really enlightening to hear that, like is that... There is a lot of opportunity building on top of these models, both today and then accounting for where they're going to go by doing just like thoughtful product. I love that.
Nabeel Hyatt:
I glossed over it and I forget always that not everybody is obviously reading these papers and going through these things, but you're right that we have had... They're not the only ones now. I think we come across several teams that have found that... And it's intuitive. Creating a framework that is more natural for an LLM to interact with is going to get better results. That's not changing the model, that's changing the air, the model's breathing, the interaction layers the model is trying to use. What else are you taking away from this?
Fraser Kelton:
There's something that you and I have talked a lot about with the partners here is, what is possible when you can wait an hour or a day for an outcome rather than looking for an immediate response? And again, I think that there are many tasks where we will be tolerant of that.
Nabeel Hyatt:
What do you get? This took you four hours to make an app and that's crazy. There's plenty of other situations in many other startups where if you think about how long Google would allow a search result to last before it give you back an answer, they're going to measure that in milliseconds and want to drive it down the, "Hey, I'll get back to you over the course of the next three to four hours [inaudible 00:13:03]," is not... If you change that time affordance, what new products can you make? And the cost affordance, obviously. It's also not going to cost a Google query. That's probably a multiple dollars of cost in making all those queries and call, but so be it. You got what you wanted.
Fraser Kelton:
Well, listen that I can't even begin to fathom how much my app for my daughter's Pokemon collection, probably cost an inference, but it worked and it's great. So that is one. I think that there's a whole paradigm and then you have to think about what is the right user experience that account for this. Halfway through my using of the product, they enabled push notifications as a feature within the desktop experience, because I kicked this off and then need my help or not.
Nabeel Hyatt:
Of course, because you can walk away and then you want it to be able to ping you 45 minutes later and be like, oh, I need your help now.
Fraser Kelton:
Actually, way too early I'm sure on their product roadmap, but my one thought was like, man, I'd actually like some of this experience within an app, because if I'm on the road and my agent gets stuck, I want to be able to give it a little nudge to keep going. Another observation that we had and that we talked about that I thought was interesting is, if you are building a product that is supporting and enabling technology that's not quite mature enough, just fail gracefully. Find ways within the product experience to accept failure and do it elegantly for the end user. And in this case, what I wanted it to do was to write a full web app. I didn't care if it took four hours to get there and I didn't care if it bumped up against a dead end five times on one task, if it eventually overcame that obstacle and move forward.
Nabeel Hyatt:
Just keep brute forcing through every error, keep eventually end up on stack exchange. Just keep trying and trying and trying. Maybe at the 50th time, ping you and send you a text message and ask for help, but just keep going.
Fraser Kelton:
Just keep going. And that harks back to the early days of the GPT-3 API with some of the early groups who had traction where like the copywriting apps, they failed gracefully. If you tried to one shot your marketing copy, it was not very good to be able to do it. But if you showed the user 10 results and let the user easily pick from those 10, it was not friction filled and they could find the good outcome pretty consistently from that, and so I think that that's my other takeaway on that. Do you have any other takeaways from playing with the product and watching some of the demos?
Nabeel Hyatt:
No, I liked that analogy. I'm sure the copywriter, the folks that were building the copywriting apps in the early GPT-3 days started with something that gave you one example or two examples and were like, "Oh, this isn't good." And so how do you solve that? In that case, give a 10 shot and not just feel gracefully in a way, but there's a delicate dance between how much weight we're putting on the customer versus the product. And we're, listen, we can't do the whole thing fully, but we're just going to keep trying and then we'll ask you for a little help. We're going to give you 10, pick the best. We're just trying to pick the best. That makes sense.
Fraser Kelton:
Right, that makes sense. And actually that then leads me into a question for you. It failed gracefully. It failed ugly at times as well and it was probably something that would've benefited from a more natural interface with me collaborating with it. I think this is clearly a glimpse of the future. I think everybody agrees that a lot of the software that gets written is going to be done by AI in time, but we're seeing two very different approaches to try to get to that future. We are seeing groups start with the workflow tools that exist today, and layer on AI to those incrementally.
So here's a ticket to PR. I'm going to integrate into your linear and to GitHub, and I'm going to be able to just stitch those together and find leverage for the end user through that. And then there are other groups like Devin as we've seen, who are just basically breaking free from today's interactive models or interaction models and saying, here's an entirely new product experience that we are going to allow you to do novel new things within. The technology is not there yet today. It will get there today and we're going to figure out the right product experience as the technology matures and we're going to meet the moment a year from now, two years from now in a very large market.
Nabeel Hyatt:
Well, if you're meeting the world where they are today with a interface that feels familiar to the way that we work today, then the positive from a founder standpoint is they know how to use the product immediately. The negative is that those interfaces generally favor incumbents and new workflows generally favor new companies. This kind of goes back to some of the frameworks we've been using around is it an adaptation evolution or is it revolution?
You have to ask the question that people ask all the time, both founders and investors, like, what if the big boy integrates this tool? And that can be easy or that can be hard. It's not just a one-off flippant question, what if Google does this or whatever? There are points where it is very easy to say, "Hey, you're still in the same IDE that you were in before. Now there's just a little window up here. You click a button and it does a little thing automatically for you. Use the same IDE you've always used. It just does a little bit extra." Guess what? That's obviously going to favor the incumbent. You are probably going to get squashed. The complete opposite is does the AI afford a new workflow or a new UX that is just wildly and crazily different that you haven't seen before? In which case, the problem with any incumbent is that they can try and compete with you.
Maybe the technology they have is slightly even more superior. Maybe they catch up with you. But the whole point is they have millions of users who are used to the old workflow, and so they're going to have to start over from scratch. And we have several example of these that we've worked with at Spark, and so for me it's just an old innovator's dilemma kind of conversation. AI will be proven over time to have areas of the market where incumbents obviously are going to succeed and where incumbents have massive disadvantages.
And the places where incumbents are going to have massive disadvantages are likely places like the interface layer entirely changes. That one I already mentioned. They could launch a new product from scratch with your interface, but then they're starting from zero as well and they can try and use bundling and all the things that large companies do, but that doesn't necessarily mean they're going to win.
The second one is use business model advantage, and the best example I have this one is ChatGPT and Google. That's just everybody talks about. You have this situation where now there's a cost per search that's really high. Google is a free ad based model. How do they compete with the idea and how do you merge your Google business model with the way that ChatGPT is a paid model is going to be, it's going to be marketed. The third one I've been thinking about lately, Fraser, I haven't brought up with you is areas where in particular sales and marketing were previously a slog and may now be automated by AI for that particular industry. There's a whole bunch of companies in prior ages that were really good at either long tail self-service marketing to everybody on the planet, go be Squarespace and build your own webpage, or really big enterprise customers. The ones that were always the hardest to solve were mid-market players.
Fraser Kelton:
Right.
Nabeel Hyatt:
Where the sales and marketing was just a slog. You had to touch everybody. This is like construction and legal and there's a whole bunch of industries where essentially, there's no major technology player incumbent because the go-to-market was so crap in that market.
And so if you could work on an AI enabled go-to-market to that system, which is by the way very different from some of these construction legal companies. I'm not even saying that it's some AI model that writes your legal briefs. That could totally be the product. I'm actually saying something different. I'm saying the reason there's no incumbent player there that's technology forward is because the sales go-to-market motion is terrible and you can probably use AI to take advantage of the sales go-to-market motion.
And then my last one is a situation where the data structure itself, the way we store the data is entirely different and therefore has an innovator's dilemma. And the best example of this is it's not just that it's a vector database that I'm now storing everything in. It's that storing things in these loose relationships where I can get summarization and bring up new products. And actually, image diffusion models are kind of this. It's not like Google wasn't already caching all of the images on the web, but storing them in a different way where I understand the relationship between them gives you an image diffusion opportunity to create something like a journey. And I suspect there are lots of other situations where it's not just that you have the data. It's that the way the data is structured is entirely new and that will enable new products that someone else will have an innovator's dilemma to build unless they rebuild their entire data structure the way that you're built.
Those are my off the cuff before. I'm sure there are another three or four or five that we really went through all of these industries and thought through like, "Where are their incumbent disadvantages?" I think it's important though, as a founder to think through what your incumbent disadvantage is going to be for the large players because they're going to come at you. This is very much like the dot-com era. It's not like AI is catching anybody by surprise at this point, right?
Fraser Kelton:
Yep. Yep. So, let's see if we can summarize that. In situations where there's either an entirely new UI or business model changes significantly such that it's against what has historically been present, or there's an opportunity to solve the mid-market slog of marketing and sales with these types of technologies, or if there's new data structures that afford themselves or present themselves to product builders, those are at least four opportunities where new entrants might be able to thrive and win.
Nabeel Hyatt:
Where if you look at the old players, you either say that they're not... This is the reason those old players are not there and entrenched or this is where as an old player tries to move into this world, they're going to have some level of a problem. The last one we're not talking about at all, which is the more frequent conversation we have internally, to be fair, is entirely new market creation, where there's no incumbent because it's a new job to be done and everybody uses the Uber example really, which is like a total disruption of a market. Yes, it's taxis, but it's not really that. It's a whole new thing, frankly. Airbnb is another example and those are the canonical versions.
And I actually thought Sarah Tavel, who's at Benchmark, wrote a nice piece this week on... Her headline was something like What Midjourney MLEVM Labs and HeyGen Have in Common. I actually thought the post had some good points to make. She used, first of all, an analogy that I use a lot, which is stop thinking about it as an AI model or stop thinking about it as code, and just almost think about it as a person doing a job, thinking of it, outsourcing. And if you think about it as outsourcing, it puts a different lens on things.
And then the thing that she was coming to in grouping all those together is in a way, in this case, it's outsourcing, not to replace a job that was currently being done, but it's really a net new job that wasn't going to get done otherwise. It's not faster horses, "Hey, let's take our SDR sales function, outsource it to a model so it's 50% cheaper or 1,000% cheaper," whatever.
Fraser Kelton:
Right.
Nabeel Hyatt:
Look, I use Midjourney all the time. I would never, for hobby projects and things that I'm doing with my kids, as you know, I would never hire a concept artist for those jobs. It's a net new job. And similar with HeyGen, you've probably maybe used it for a marketing video, but you probably have never have hired an actor in your life.
Fraser Kelton:
Right. Right.
Nabeel Hyatt:
And so it's net new frameworks or net new markets that were created simply because the way you use this product is either 10 times cheaper, 10 times easier. It's not as much friction. I'm not trying to find talent in the world to go do the thing. By the way, I actually think, to loop back to it, that's the comment I would have about Devin. I don't think of Devin and that whole affordance, that whole type of product, I don't think of Devin as a better software engineer. I think of Devin as an entirely new software engineer. It is not a replacement market. It's that how many more people would hire coders to make small programs if it wasn't a total pain in the butt? And all of us would do that.
Fraser Kelton:
I made a web app.
Nabeel Hyatt:
Right? Exactly.
Fraser Kelton:
I never would've made that web map. I wouldn't have hired an engineer to build that. I wouldn't have built it myself. Interesting. Very interesting. What are the implications then for the product? Let's take Devin out of the picture so that we're not talking about a specific company, but clearly there's going to be... You will have at your disposal if you're a software engineer, I have to assume, AI that is the equivalent of a whole bunch of L4s that complete tasks for you that you then orchestrate. Is there going to be two different products? Is there going to be the product that is for the senior software engineer who then is orchestrating a bunch of AIs at RL4s, and then is there a different product for the person who's in this previous framework that you just shared is building software that they wouldn't have otherwise hired a software engineer to do?
Nabeel Hyatt:
I think if you're a startup and you're trying to figure out your go-to-market, you're trying to figure out your growth trajectory and you're also not trying to get stopped by the competition over time, then these things are in direct conflict. You want to move fast, then all your advisors and all your board members and all your friends are trying to help you do what is ostensibly TAM analysis. Here's the market for software coders or lawyers or sales agents or whatever. And then you go after the ones that are out there, please get me traction. Three months, go, go, go, go, go. That necessarily puts you into a red ocean situation.
Pulling the thread on, great, what new behaviors does this enable for? What new types of customers? Puts you in a much more uncomfortable place, especially with a bunch of VCs that like to see TAM slides, knowing laugh, but it puts you in a situation where you have not only less people competing with you over time. You don't have a red ocean situation, you have a blue ocean situation. You don't know where the fish are, but you know it's an open sea. The better thing over the mid to long-term there is also that frankly, you are going to learn about a new customer set. You're going to discover things about that customer set. And then frankly, when incumbents and other startups try and run at you, you have had time learning about this new type of customer that they just didn't know about.
And I had this conversation with Andrew Mason over at Descript recently that was kind of similar to this, which is just they started out in the podcasting market and most of the major podcasts all use them, This American Life and all the rest of these use Descript, but they had a hard time sizing their TAM because they kept getting new users that use the product that don't fit any other category they previously defined. And so they invented internal things like Communicator and so on and so forth that explained why they're using this product, but they just signed an enterprise contract for a whole company to use Descript. That's not in the TAM slide initially. Right?
And I think there's so much more purchase there on just pulling the thread on how your users are using a new type of workflow and following it. And you don't know how far it's going to go, but it certainly leads to new pastures. I think there is some analogy that we could probably make back to companies of the 90s and those who chased Tams that existed versus Tams that were net new. Right?
Fraser Kelton:
I'm laughing because yeah, this was a topic of discussion internally recently. And in an offsite we had recently for us at Spark, Kevin presented observations from the dot-com era to help us figure out and see if there's anything that could help us figure out this current moment in time. And he had slides that showed early winners within various markets of the dot-com moment, and then he showed who ultimately arose as the long-term winner.
Nabeel Hyatt:
Yeah, I loved it because we spent so much time talking about the mobile revolution as a context window and those feel very fresh to us. We were both founders through some of this, and so it feels still very fresh to us, right? We were both founders through some of this, and so it feels really native and I definitely am old enough to have, as a very young person, gone through some of the internet stuffs, but it's also just like the late 90s internet stuff was also just far enough away and fuzzy enough that I had forgotten a lot of the information and just kind of going back through that historical cycle again was pretty great. Netscape is basically open AI. I mean it basically is, it was an enabler to all of this. It was the way you get onto the internet where you get access to everything is going to be Netscape. And so it just felt like the big company and 16 months to IPO for Netscape and a $3 billion market cap on day one under 18 months old. It was pretty crazy. And then his whole conversation about how then the way that they were going to monetize Netscape was not to sell the browser but was to sell web servers, which again felt very similar to this like, oh, we're going to sell AI models obviously.
Fraser Kelton:
Yeah.
Nabeel Hyatt:
That doesn't mean by the way that I think Netscape and Open AI are going to have the same kind of endings. I think we can't over index on all of these things that are in different market positions. He went through individual categories and all of them I think are relevant in application. And to think about incumbents versus net new companies today. Obviously the enabler was Netscape and then he went through the application companies. Companies like Yahoo, which is very much almost, it's somewhat analogous to ChatGPT. It was the directory of the internet at the time. It was how you thought about how you're going to use the tool and get access to everything out there. He talked about web builders, which folks like Razorfish, Sapient and Viant. Do you remember those companies? I was very surprised that those companies felt iconic and big to me. And some of the younger folks in the room had never heard of Viant and Sapient.
Fraser Kelton:
I had heard about half of the previous companies from that era who didn't necessarily persist. And actually it was very interesting. I've heard of more of the consumer facing ones, which you would expect than the infrastructure or other layers, but even of those, I certainly remember Razorfish, remember everybody was going to have an agency, you were going to have a web agency that was going to help you get on the information Super Highway. So it was interesting. I feel like we haven't done Kevin's presentation justice because the insight that he was trying to instill was there are clear winners in these moments who might not persist as the winner. They also still might be great businesses and investments in that moment. And then to remember that there's likely to be long-term winners that are still getting started or will get started in the next couple of years.
Nabeel Hyatt:
Yeah, for sure. I mean both of those points. One, there will be obviously durable companies that will last. There'll be in the next a hundred billion dollar companies in the last 20, 30, 40 years that we'll still know about, companies like Amazon or eBay. But also that there were some really important companies that moved our entire market forward that were wildly optimistic and they didn't last and that's okay. Founders and investors made money, but maybe more importantly than that, they actually moved the technology environment forward. And you need that kind of optimism and a little bit of frankly, overconfidence when you go into these markets, cynicism doesn't help you that much. I've been listening to a bunch of old podcasts from Danny Kahneman this week. He passed away and he has this saying that's like overconfidence can be beneficial. It's a good thing in certain scenarios because first of all, it makes the individual being overconfident more appealing, which helps them acquire resources and encourages them to take risks and that can contribute to societal progress and economic growth.
And so while the odds of success are often exaggerated, that optimism drives economic progress. I think he called it the engine of capitalism. And I think we all can poo poo the late nineties because of the companies that didn't make it. But level three laid down fiber for the very first time, which we all benefit from. It changed the world for us. And some of those companies also, especially the infrastructure companies, they really lasted. I think one of the takeaways I had from that process was that the consumer companies that made it really made it, the Amazons of the world became unbelievably dominant, but a lot of infrastructure companies kind of also made it through. I don't think about them as much, but Cisco is a $200 billion company, Akamai who haven't thought about it in forever. We looked up in the middle of the meeting, it's a $16 billion company, who knew that?
And so for all the talk of application software, AOL is not really around and Excite@Home is not really around and so on and so forth. And a lot of these infrastructure companies really did last. I'm not sure exactly what the corollary is to those infrastructure companies in this environment, whether they are the model companies or the eval companies or chip companies, or is it going to be NVIDIA? Who are those players? I think that's a different conversation to have just trying to evaluate all these things through our current framework. But I love his takeaways.
Fraser Kelton:
I thought even just this obvious reminder that we're watching this every single day week after week, and these are our iconic companies in the moment, and we are cheering for them and we are getting excited to see what they bring to market because it's profound. And I think a lot of people think that these markets are done. And Kevin's reminder was that no, there's a very long, long arc of history that we're just one or two years into and we're going to see a lot of change and a lot of opportunity still. That I thought was just so amazing and motivating to me.
Nabeel Hyatt:
I agree. It's really interesting. If there's one area in conversations about AI in this current moment in time where I've seen a massive age gap, it is in predictions about where we are in the cycle right now. You've referenced the S-curve before and where we are on the S-curve for these individual technologies. And I think that's actually probably worth a 10 or 15 minute conversation on its own, and maybe next time we have a hallway chat.
Fraser Kelton:
Next time, maybe.
Nabeel Hyatt:
But I would say on average when I talk to folks that are younger in their careers and earlier in their careers, they will put us much further into this AI cycle. And you look at these companies and then you look at the 5, 6, 7, 8, 9 year cycle of the internet getting built out and you're talking about Yahoo and Excite@Home and AltaVista and Google is still not even invented. That's years away. It's just not even there. And yes, our technology curves are accelerating, but it is going to take a long time for these models to get to both their peak level of abilities. We're still not even there yet. And then secondly, to work their way through the GDP of the country to work their way through every single industry. And sometimes there's fast moving industries and they're going to flip immediately as we've seen, but there'll be other ones that are just not going to be ready to flip for another four or five years.
Fraser Kelton:
Oh, sure. Longer for some of them.
Nabeel Hyatt:
Yeah.
Fraser Kelton:
I really like that. Maybe that's a natural place to wrap up then.
Nabeel Hyatt:
The last thing I'll say though is the other thing I took away from that whole Kevin thing that really stuck in me and I've repeated a bunch of times is I have a model is not an investment thesis or a founding thesis for a company. Nobody-
Fraser Kelton:
Oh no, we don't have time for this.
Nabeel Hyatt:
I'm just going to rant it and it will stop. It's just that the beginning of the internet, you don't say, I should exist and be an amazing company because I happen to have a website. It's not enough. Like this is an enabling technology and it's what it enables, not the fact that it's enabled, that matters. And I'll stop there. I will go to the rest of the rant. Call it a day for now.
Fraser Kelton:
No, I feel like the rest of the rant is actually the really interesting part. You have been saying frequently that even if somebody is pitching, here's a product with the model, you just want to say, let's ignore the model. Let's just not even discuss that. Let's not even think about it. Let's just move that off to the side and then evaluate it on the merits of the product and what the product enables for the end user. And that's like a big shift from a lot of the conversations a year ago. And it's reasonably counter to a lot of the discussions that are happening to the market. And so it feels like that deserves a little bit of space to breathe.
Nabeel Hyatt:
Well then let's talk about it next time.
Fraser Kelton:
Love it. All right, well thanks everybody. What a teaser. I can't wait until next time.
Nabeel Hyatt:
I just did the comic book, the end of the Batman episode, the comic book ending. Until next time, Fraser.
Fraser Kelton:
See you.
Nabeel Hyatt:
See you.