Worldbuilders: Why Most AI Startups Won't Survive | The Model Economy by Sumeet Singh

0:07I'm Sumeet Singh, and I'm guest hosting a new series for the Village Global podcast called World Builders. I run World Build, a thesis-driven investment firm that backs creative technologists. We spend our time exploring rabbit holes because some of these rabbit holes are actually tunnels that open into the future. In this show, I'll be taking you down those tunnels. We'll have conversations with the people shaping what comes next, break down the frameworks for where value accrues when the rules are changing and figure out what the builders actually see that the rest of us don't. So come take the leap with world builders.

0:38For this episode, I want to lay out a framework I've been developing called the model economy, which I briefly touched on in our last episode with Evan Conrad, the CEO of the San Francisco Compute Company. It's a way of thinking about where durable value gets created in the age of AI and why most of what's being built right now might not survive. I'll walk through the history with the bitter lesson, what the mobile era teaches us about what's coming, and the two types of companies I think actually win in this era. Think of this as my investing thesis out loud.

1:07Sumeet Singh:The playbook of the last era of software tech investing is over. For a while, we were locked in this era of homogeneity. It was dictated by a well-trodden path of fundraising, I would argue, rather than true innovation. It was like, how do we make the best-looking spreadsheet company? Right. Every company was almost fungible. There wasn't change. And how do I not optimize for new experiences and new products, but how do I optimize for a fundraise? Honestly, that led to companies raising too much capital, too high valuations. And that obviously, you know, corrected itself. I think that era is over.

1:41Sumeet Singh:We, of course, have the advent of transformer models, generative AI. Also combine that with the end of easy monetary policy. And as an investor, I'm actually excited. AI has finally opened up the potential for real innovation that's been missing since the mobile revolution. Post-mobile, until just a few years ago, there wasn't much change. But I'm a bit worried about how founders are reacting to this. I often see founders building these specialist AI products. Imagine an AI product for marketing or an AI product for finance. As if they were still building for the same sort of subscription software era of the last decade.

2:18Sumeet Singh:They're treating AI like it's just another feature in a SaaS app. And those who are playing by that old framework, though, are about to make a big mistake, in my opinion. I think the old framework fails because of what the AI pioneer, Richard Sutton, one of the early reinforcement learning pioneers, calls the bitter lesson. In traditional SaaS, right, again, the last era of investing and building, you won by building a moat of hard-coded rules, vertical expertise, systems of record, et cetera. But the bitter lesson teaches us that brute force compute and scale almost always beats domain-specific cleverness.

2:53Sumeet Singh:And that's been the case over the 70-year history of AI, right, which has been its own category outside of software. In that world of AI, scale always wins. That's the bitter lesson. So, look, if you build a specialist, you know, AI app for finance or marketing today, well, for one, there's probably a dozen competitors because what it takes to build software has gone to zero. But I think more interestingly is that the models themselves will likely outperform you in just a few months. You know, there's this beautiful chart from the research nonprofit meter, METR, which has shown that since the advent of GPT-2, which was released in 2022, the task length that a frontier model can carry out has been doubling every six months.

3:35Sumeet Singh:We've gone from two seconds to 6.6 hours actually as of yesterday. If your value prop then is, hey, we taught the AI accounting rules, I think you're fighting a losing battle against the scaling laws. The model itself, the models themselves, will swallow the application layer. So if models are eating software, then the question is, where does value go? Where does it accrue in this era of investing? I have this concept, what I call the model economy. And the thesis is pretty simple. It's don't just build lightweight applications with models. Instead, build for those models, help improve the model capabilities, help the models proliferate.

4:15Sumeet Singh:That's an entirely new market. And I think durable value will accrue to the infrastructure that keeps those models alive, growing, and improving. And look, when people think infrastructure, when I say this infrastructure for models, they of course just point to the chips, right? To NVIDIA. But I'm actually looking for pretty specific things. One is what I call the exchange model for compute and for power. I think when we zoom out 30 years from now, we're going to see that everything around us is compute driven and that this demand for compute from now until then will be exponential. However, when you zoom in, I believe it will be incredibly volatile.

4:50Sumeet Singh:One month, we'll have a GPU shortage. The next month, Microsoft says we have a glut. And so the venture opportunity, I think, is not just, hey, let's go buy chips. I think it's building marketplaces that smooth out that volatility, perhaps even benefit from that volatility. You know, we're looking at startups that trade flops, like commodities. They also trade the watts. They trade the power, helping increase utilization, helping get power where it's needed and when it's needed. Another key piece of the bitter lesson is data. Data is required for scale. And, you know, what I'm looking for in data is experiential data.

5:25Sumeet Singh:And what I mean by that is, you know, when you had hardware companies, let's say tackling the built world, the physical environment. They had to deal with raising a lot of capital to get to market, to serve that end customer base. And unfortunately, for many of the companies, the CapEx need was too great. What's different today is that there's almost this new business model where you can capture data that, let's say, you place on a site or on a bulldozer. You can capture that data and now sell it to a world model builder that needs to understand physics, need to understand the real world. That almost gives you an entirely new lifeline of revenue that did not exist before, right?

6:07Sumeet Singh:The models are these new customers. That's a totally different environment than the last era. Another piece of the model economy in which we're looking at is security. And, you know, when people think security, they think firewalls, defensive, protect me from bad actors, protect my infrastructure, protect my product, protect my company, my customers from bad actors. But I think off security in the model economy is offensive. These models themselves can be tricked. They can be jailbroken. And if models proliferate, if they touch all aspects of our lives from the cars that we sit in to the, you know, humanoid robotics in our homes, we need to be able to jailbreak and red team these models so that the vulnerabilities can't be exposed by any other bad actors.

6:54Sumeet Singh:So security looks much more offensive. It looks like, say, a Navy SEAL Team 6 of people who are actually trying to break the models, expose their vulnerabilities before a bad actor can. Offensive rather than defensive in the model economy. When people hear my model economy thesis, they think that I believe all applications are doomed. But that's not exactly true. The applications that I think are doomed are skeuomorphic in nature. They are existing workflows that are essentially AI-ified, Right? The marketing apps that I talked about, the finance apps that I talk about. What I'm actually looking for is post-Skeumorphic apps.

7:29Sumeet Singh:And to take a step back, like what is skeumorphism? Skeumorphism was a design term that I think got popular around the time of the iPhone, around the time of the original mobile apps on the App Store. If you recall the beer drinking app on the iPhone, that was a skeumorphic application. But no value accrued to those skeumorphic apps. Where did value actually accrue? They accrued to the post-Skeumorphic apps. Right? They accrued to Uber, DoorDash, Instacart. When your phone became a remote control for the world, as Matt Kohler from Benchmark has said, that was post-Skeumorphic in nature. And that's what I'm looking for in the application world today.

8:02Sumeet Singh:It's not just taking an existing workflow, like a human writing a report, and having an AI write it for you. It's not just that same process, just faster. That's fragile. It's instead building workflows that were impossible without AI. Now, one example of that is multi-agent coordination. which I think is becoming very real. Instead of, you know, one AI co-pilot, let's say, just helping you, imagine a hive mind. One agent writes the code, another reviews it for security, a third generation tests, and a fourth manages it. You know, if you've played around with Claudebot, I think you could see where this is going.

8:38Sumeet Singh:They argue, they collaborate, they fix each other's mistakes. The value isn't in the single model. It's in the emergent behavior of the swarm. And that's something a human organization can't easily replicate at speed. I would also argue it's not something that one lab will be able to do themselves. You know, I find that the most interesting multi-agent collaboration happens when you set up distinct models in its own team, in its own council, if you will. You should check out Andre Karpathy's LLM Council, which alludes to this, where you could set up your own team of different AI models working towards each other with different personas, different characters, et cetera.

9:14Sumeet Singh:Another way that I'm thinking about post-geomorphic applications is you have to start with the nuances of the models. Like, what can we uniquely do today that was not possible before? And one area that I'm excited about pursuing that substrate in is in the world of science. Take in biology, for example. You know, for decades, we had to hand code biological pathways or do physical experiments in a lab. Now companies are using these models to simulate entire cellular systems or protein structures. like we're collapsing decades of trial and error into minutes of computation. If you can build a platform that simulates millions of materials or drugs to find the one that works, like that's a massive durable business.

9:54Sumeet Singh:Very different than saying, hey, let's give a scientist a co-pilot. Instead, what if the model itself, right, is the experiment? Another nuance that I've picked up upon and that I've been excited about is what I call 24-7 feedback loops. You know, take, for example, in the observability world. Right now, if software breaks, if your infrastructure or application breaks, we get an alert, a human wakes up at 3 a.m., and they have to fix it. A post-skeomorphic observability platform wouldn't just give you an alert, it would generate a hypothesis, perhaps write a fix, spin up a test environment, verify it, and push it to prod, close the loop.

10:29Sumeet Singh:Instead of the human ever realizing that something broke, the software healed itself. That's what a post-skeomorphic observability platform would look like, for example. Look, I think something that I'm cognizant of as an investor is that timing is everything, right? Being too early is basically the same thing as being wrong. You know, part of the model economy thesis is that we're betting on infrastructure for scale that might still be years away. You know, if the scaling laws don't hold or, you know, the new chip architecture doesn't lead to breakthroughs from a model perspective, perhaps we experience a lot of volatility.

11:04Sumeet Singh:And so for me, something that I think about as I invest behind this model economy and in post-geomorphic applications is the timing right. Secondly, when you think about the application layer, I'm excited about these post-Skeomorphic applications, doing things that can only be possible today. But is there a world in which take some of the skeomorphic applications, do they just get to distribution faster than the post-Skeomorphic apps can catch up to? That's something that keeps me up as well. This is where the incumbents come in as well, right? Network effects, distribution is still powerful. Maybe Reddit selling data to Google proves that the old school aggregators will still win.

11:41Sumeet Singh:And it's possible that these embedded workflows, you know, just being the tool that people already use, beats the better AI native post-geomorphic application. But here at WorldBuild, I am betting on the better lesson. You know, history rarely gives us a clear playbook, but I think the better lesson is the closest thing we have. You know, brute force and scale keep closing the gap on cleverness, on clever workflows, on specialized workflows. So, you know, we're betting on the infrastructure, the pipes that feed the beast, and also the applications that ride the wave rather than getting crushed by it.

12:14Sumeet Singh:So if you're building in the model economy or if you're building a post-queomorphic application, I'd love to talk.

12:35love to see you for the next one.