Join Free

​Chapter One Of My Book – Investing In AI: Six Mental Models To Tell Moats From Traps

Happy Sunday and welcome to Investing in AI! Be sure to check out our AI in NYC podcast, (spotify version is here). Also, if you are an openclaw user, we recently open sourced a tool we use at Neurometric called Clawbake that allows you to spin up individual openclaw agent instances at work inside Kubernetes. Check it out.

I’ve mentioned this before but, about 18 months ago I started writing a book. No one has written one on investing in AI so, I figured I might as well start with my best posts here and turn it into something. The book is done and will come out this summer. I’ll have a PDF version available to all Investing In AI subscribers so you can get it for free. Today’s post is the first chapter of that book. I’d love any feedback you have.

Chapter 1: Why Investing in AI Requires New Mental Models

In the late 1990s, I was part of a group doing freelance work for a company that wanted to build its first website. Not knowing any better, leadership viewed the web as a digital version of paper and proceeded to follow their existing analog graphic design process, which they had traditionally used for physical materials such as brochures and flyers.

Here’s how the process went. Their graphic designers designed the webpage on paper. Then, they handed the design to the web developer and asked them to make the website look like that piece of paper.

Surprise: it didn’t work.

The designers weren’t bad. They just lacked context. They were used to operating under the constraints of paper and didn’t yet understand that those constraints were meaningless on the web. Websites are dynamic, enabling visitors to scroll around and click links however they see fit. A website built out of an analog design process wasn’t doomed to fail. They needed a new approach—a new way of thinking that incorporated the attributes of the web.

Once they understood this, the team was able to shift their thinking and design a functional website.

Fast forward to 2008.

I founded Backupify, my first startup, which provided backup for cloud computing applications, including Google Apps, Salesforce, and Office 365. We followed a software-as-a-service (SaaS) model, offering a monthly subscription for access to our applications.

Many investors didn’t understand what we were doing at the time, because SaaS was still a relatively new concept. They were more concerned with bookings and maintenance contracts than annual recurring revenue (ARR) or churn. Several investors told me point-blank that they didn’t believe cloud businesses would work because the economics weren’t as good as traditional software companies.

Time proved them very wrong.

Just like the graphic designers trying to design a website on paper, these investors were applying old mental models—software licensing, maintenance contracts—to a new paradigm. It’s what they knew, but it also didn’t make sense.

Sometimes, this lack of understanding is relatively harmless. The company I worked with in the late nineties was delayed in getting its website up, but they eventually figured it out. But sometimes, this lack of understanding can prove incredibly costly.

Take Yahoo, for example. Initially, they didn’t understand the importance of search. They were a directory business. They saw search as just an add-on. So, they contracted with Google to power their search tool. That choice helped make Google—and stripped Yahoo of a crucial growth engine. By the time they realized their error, it was too late.

The point is that the history of new technologies is littered with examples of people misunderstanding how a new technology works and where its value accrues within the new technology stack. Just as investors in the mid-to-late aughts sought to apply software licensing rules to early SaaS companies, investors today are trying to apply SaaS and software rules to AI companies.

The problem? It doesn’t work.

That’s why, to invest in AI and AI-related industries successfully, we must first understand this unique moment, how it differs from previous eras, and how to maximize value in the age of AI.

Invest with Caution: The Ups and Downs of AI Development

AI has a long track record of impressive demos that fail in practice. When that happens, capital behaves predictably. A bunch of money flows in, the returns aren’t there, and then funding dries up. This pattern of “AI winters,” as they’re referred to, is essential to understanding how investors have traditionally behaved toward AI and why that pattern no longer holds.

Broadly, the pattern goes like this:

You can see the pattern: Tech entrepreneurs overpromise and underdeliver. Investors get cautious. And everyone treats AI more like a science project than an industry.

This cycle has perpetuated itself since the 1950s—the early days of artificial intelligence. Computers were young, AI was new, and many thought they would be able to do everything with these new technologies. With what we’d now consider naive optimism, Marvin Minsky, a professor at MIT, gave one of his graduate students what he thought was a simple summer project: solve machine vision. The idea was to build a thing so that a computer could identify images. The student made almost no progress, and Minsky and others began to realize the complexity of the problems they were trying to solve.

So went the early days of AI in the 1950s and 60s. Ambition was enormous, but reality was stubborn, and progress came slowly—too slow to justify the expectations people had set. Frustrated by the slow progress and questioning whether AI was worth pursuing, investors turned their attention elsewhere.

By the 1980s and 1990s, that same gap showed up in a more commercial form through what we called expert systems. The premise was simple: If you were an expert in something, a team of engineers would interview you and try to create an algorithm that would standardize and scale your expertise. These expert systems functioned as long as the process lacked nuance, at which point they broke down. And since most processes involve the infinite nuance of human operators, these systems did break down—a lot. Most processes, it turned out, simply had too many use cases for these systems to work predictably.

Once again, the tech world overpromised, investors grew impatient, and we entered another AI winter.

Still, while funding would often slow to a trickle, research into AI didn’t stop. Computers did get better at machine vision. Expert systems did develop methods to account for nuance. Many other new tools, use cases, and applications emerged. But the progress was slow and incremental, and most of it didn’t feel like a revolution.

Finally, by the 2010s, all this slow progress culminated in a series of remarkable breakthroughs. In 2012, AlexNet beat the accuracy benchmark for machine vision by eight to ten points. By 2016, neural networks could identify basic images as well as the average human. With these new advancements—including the rise of GPUs and neural networks—we were no longer eking out single-digit percentage improvements every year in AI. We were getting 50 percent improvements. Hundred-percent improvements. Emerging tools and techniques were beating out the old techniques by significant orders of magnitude.

It was an exciting time from an investor’s perspective, but also a dangerous one. On the one hand, there was a real sense that this was it; AI had finally arrived. On the other hand, many remained conscious, convinced that another AI winter was just around the corner.

As I write this a decade later, the next winter doesn’t appear to be coming anytime soon. At long last, the pattern is broken.

Here’s why. AI hasn’t advanced merely because of increased compute power and data, even though those have helped. It has advanced through the diversity of techniques and models that have been used to build it. As a broad category, AI advances through branch shifts, new architectures, new training methods, and new assumptions. That’s why the best AI investors don’t treat “AI” like one thing—and it’s also why the winters keep recurring: it’s easy to overfund a story about “AI” and miss what’s actually happening in a specific pocket.

We’ve reached a new plateau of adoption because many of these pockets have become robust and useful. Even if progress slows in some of these pockets—which it inevitably will—nobody gets to un-invent what’s already been adopted. AI tools are now deployed across too many workflows, too many companies, and too many tool stacks to put the genie back in the bottle. The deployment rate is simply too high—and in terms of investment, deployment should be the main metric of success.

So, why does this matter? Why have I written so many words about AI winters and investment cycles when you came to this book to learn about how to invest in AI today?

Because the history of AI isn’t just a chronology. It’s a behavioral pattern, it’s a capital pattern, and it’s the reason investors keep reaching for the wrong mental model.

If you assume investing in AI progresses the same way investing in SaaS did twenty years ago, you’re in for a rude awakening. AI has already told us what it does when you mix hype, demos, and reality. If you want to invest well, you have to start by recognizing the pattern.

Four Fundamental Reasons AI Is Different

Now that you understand why we’re not doomed to another AI winter, let’s examine the specific reasons why AI is fundamentally different than what’s come before.

1. AI Replicates Cognitive Capabilities

In the nineteenth century, machines began to replace physical labor. Horses became tractors, and people had a simple mental model: “Machines can lift more than humans.” Today, AI replaces cognitive labor. Software can write, analyze, and decide. Analysts become algorithms.

This is a first in human history—and it’s jarring. We’re not used to thinking about software performing tasks at the level that humans can.

That’s why investors keep reaching for the wrong comparison. When investors look at AI companies, the default mental model is: “How does it compare to a human doing the same task, and how would we price the human?” From this perspective, the obvious investor move is to price AI as if it were human. A writer charges $100 an hour, so AI writing should cost something similar, right?

Wrong. That misses everything. AI scales differently. It improves differently. It fails differently.

Instead of thinking about “software market size,” think about “tasks humans currently do that AI could do,” and then ask what happens when those tasks become cheap, fast, and widely available. Think about the second order effects of having access to that kind of intelligence. That shift changes how we think about pricing, ROI, and competitive dynamics, because the real competition isn’t always AI versus AI. It’s AI versus the human workflows the software is replacing.

This is what columnists Steve Cohen and Matthew Granade were getting in their 2018 op-ed “Models Will Run the World” in The Wall Street Journal. To Cohen and Granade, much like the expert systems of the 1980s were supposed to do, in the world of AI, all jobs will eventually become models. As a person performs a job or task, they generate data about how they do it, and at some scale, AI tools could use that data to create a model that performs the job.

It’s more complicated than that, of course, and different tasks will turn into models at different times, but the point is the same: Work generates data, and data can become leverage.

That mental model is hard to hold for two reasons. First, in the short term, it can make a good business look worse. Imagine two call centers with similar revenue and cost structures. One spends an extra 2 percent of revenue each year gathering, acquiring, and labeling data. In the near term, its EBITDA margin is lower, and the other business appears more profitable and better runcleaner. But if the data-rich call center hits the threshold where it can build models that do some of the work, margins start improving faster, and the other company is suddenly a couple of years behind.

The other reason is that AI applications present a wider distribution of potential outcomes. For investors, most of our project intuition comes from domains in which the distribution of outcomes is tight. Most AI projects aren’t like that. For instance, the call center might expend that effort and collect that data only to discover that it lacks sufficient predictive value.

In other words, there’s a hit-and-miss component to investing in AI. You cannot assume that spending money to collect data guarantees that the model will be useful later. Still, that’s the goal: to capture and replicate a person’s cognitive capabilities not on a one-to-one basis, but at scale. This concept that it is hard to predict in advance what AI can do well and what it does poorly is called the “jagged frontier” and it is a concept we will address several times throughout this book. Investors must understand this—and understand the uncertainty inherent in investing in this world.

2. Software That Learns and Adapts

Traditional software is written to behave predictably. You program it, it does exactly what you told it to do, and it stays stable until an engineer changes it.

AI software behaves differently. It learns from data, adapts to patterns, and can improve without new code. In other words, AI can learn and adapt to the world—and you don’t always know where it is going to end up. While this is a key differentiator between AI and traditional software, it also introduces a degree of outcome uncertainty that most investors have not had to underwrite.

The AlexNet moment is a good illustration of why this gets slippery. Before AlexNet, researchers trained neural networks on CPUs. GPUs were not designed for AI, so they were not part of the standard toolkit.

Then someone realized that the chips driving graphics cards were well suited to the mathematical operations neural networks perform, and asked a simple question: What if they trained these networks on a GPU rather than a CPU? The result was dramatically faster training times, and it changed what suddenly felt possible.

When a capability jump like that happens, product roadmaps get harder to predict. Use cases emerge that were not planned. Failure modes appear that were not obvious on the whiteboard. Companies pivot in unexpected directions because the models keep opening doors.

This is why traditional software due diligence does not fully capture an AI company’s potential or risk. Reviewing a codebase and understanding a feature list is not enough when the core system can change as it learns.

Instead, to be successful, we investors must:

With all this, of course, the question of safety becomes central. Nick Bostrom’s paperclip maximizer exercise from his book SuperIntelligence offers a famous, almost cartoonish example: Task an AI with making as many paperclips as possible, and it might learn that humans are in the way, or that humans are made of atoms that could become paperclips. These conclusions are science fiction on the surface, but they also point to a real issue: systems designed to optimize can optimize in directions you did not intend.

And you do not have to go to science fiction to see how badly this could play out. In 2016, Microsoft launched a chatbot called Tay. Its purpose was to entertain Twitter users aged eighteen to twenty-four, and it learned and adapted through interaction with these users. In less than a day, it started producing racist and misogynistic content and had to be shut down.

When a program changes as it interacts with the world, you cannot assume it will stay inside the behavioral envelope you tested. So, we need a different mental model for preparation—one that includes rails and circuit breakers, so to speak. Instead of imagining we can design a perfectly secure system, we start by assuming the system will eventually be compromised, misused, or pushed into weird situations. Then, we design the system so that, even when it fails, it fails in a direction you can live with—in other words, it doesn’t become a racist and misogynistic trollbot.

For instance, I’m an investor in Aescape, say you’re designing a massage robot to help that features two large robotic arms sitting over a special table and an ipad you can use to control the pressure and movement of the arms. rushed travelers take a moment to relax. Concerned for safety, As a programmer, the very first thing the Aescape team you would builtd (outside the model itself) wais a hard-coded rule: if the person on the table pops up, the robot’s arms pull away immediately. This rule is not routed through the model. It is not negotiated. It just happens. Overrides like this are essential when the model is uncertain.

And the uncertainty is not only technical. Whenever a new tool or capacity becomes available—AI or otherwise—that brings social uncertainty as well. We don’t yet know how people and institutions will respond to this new development.

Chatbots offer a perfect example of this. On their own, these tools can both expand into new markets and create new problems for questions concerning companionship and behavioral health. For instance, a chatbot might help an older adult who has no one to talk to, or it might help someone who cannot afford a human therapist receive more frequent support. At the same time, it might reinforce a trend where people socialize less and form unhealthy attachments to synthetic relationships.

It is rarely one thing. Email was a great invention, and it also created a great opportunity for scammers and spammers. AI will be both too—and we won’t fully understand why or how until it happens.

3. Reflexivity

Many investors—chief among them George Soros—believe in a concept called reflexivity theory. Essentially, this concept states that when you work inside a system, you influence that system, and the act of participating changes the properties of the system you seek to understand.

Most technologies do not exhibit strong reflexivity. They are closer to one-for-one replacements. Replacing the phone system makes calls clearer, but the nature of calling does not fundamentally change. Replacing the database improves performance, but the relationships in the code do not rewrite themselves.

AI is different because it does not just improve an existing workflow. It can reshape what customers expect the workflow to be. Once a capability becomes visible, it changes the baseline.

Before Siri and Alexa, people were not walking around demanding that they be able to talk to everything. Then those products showed up, and suddenly it felt normal to ask why every device could not respond conversationally. The existence of the feature created the demand for the feature, and that demand then forced product roadmaps to shift.

In reflexive markets, early adoption creates feedback loops. Success can accelerate success because the environment becomes more favorable to the thing that is winning, and failure can accelerate failure because expectations move on without you.

AI has proven to be an incredibly reflexive market. Customer expectations change in response to AI capabilities, which means product developers are chasing a moving target. Meanwhile, investors are chasing a fixed target. Traditional competitive analysis struggles with AI because it assumes a stable market definition, whereas the AI market is anything but. The market definition can change as the tools change, and your own deployment decisions can be part of the reason it changes.

4. Generational Thinking

The final fundamental reason investing in AI is different has to do with us humans and the slow march of time. There is also a generational shift underway, and the way we think about and interact with AI is not the same as the way younger generations will.

Older people tend to think about AI as something you can bolt onto an existing workflow: “This can solve a problem I already have.” That produces incremental improvements and retrofit thinking, and it can be valuable.

But that’s not how younger generations will think of AI. When you grow up with a capability, you do not experience it as a feature. You experience it as the environment. Native AI users will build behaviors and products with AI as a baseline, not as an add-on.

We have already seen this pattern with the internet. Digital natives did not ask, “How can the internet help me shop?” They just shopped online. They did not ask, “How can I use social media?” They just lived in social media. Those behaviors looked strange to older generations until they became normal, and now they’re not so strange anymore.

Similarly, AI natives will not ask, “How can AI help me with this task?” They will assume AI is present in every tool and will use it in ways that are not obvious if your native frame of reference is the old digital (or even analog) workflow.

So, what does this mean for investors? New markets will be created by people who do not feel the friction points older users feel, and new product categories will look weird until they do not.

If you are constantly judging AI through the lens of how you personally would use it today, you might make incremental progress, but you’re also missing the full scope of possibilities.

If you want to invest well, you have to resist the instinct to dismiss something because it does not match your own habits. “That’s not how I would use it” is often just another way of saying, “I am not the future user.”

Four Broken Assumptions

Now that we’ve established the four fundamental reasons why investing in AI is different, let’s turn inward to our own assumptions about investing—whether in AI or otherwise.

The biggest mistake investors make when investing in AI is failing to adjust their mental models. When you apply old frameworks to a new paradigm, you don’t just get a little off. You can miss the whole point.

We’ve seen this pattern before. Early SaaS investors applied enterprise software multiples and missed massive growth. Early mobile investors treated apps like websites and missed the platform shift. Early internet investors tried to map it to TV, radio, or newspapers and missed how value would accrue.

AI represents a similar shift, except the pace is faster and the uncertainty is higher. Companies in this category can achieve enormous scale on a timeline that was once reserved for once-in-a-generation outliers. If you bring the wrong mental model, you won’t just misprice a company. You’ll back the wrong bets and pass on the right ones.

The four fundamental reasons AI is different don’t just change a few line items. They break some of the default assumptions investors carry over from software.

Assumption #1: “It’s Just Software with AI Sprinkled On”

Many people view AI companies as ordinary software companies with a new feature. That framing makes the space feel familiar.

But AI companies routinely experiment with business models that don’t map cleanly to SaaS. Some give the product away to collect data, then sell access to the model trained on that data. Some charge per task or per outcome instead of per seat. Some start as services, capture the workflow data, and then use that data to become software.

That last pattern is easy to miss if you only look at the P&L. In 2017 I invested in a company called Botkeeper that was trying to automate bookkeeping. BotKeeper did this. On the surface, it looked like a services business. What many investors didn’t see was the embedded optionality: if you capture the data from services, you can use that data to turn the services into an algorithm. Botkeeper’s use of humans in the early days made their gross margins low – 20-25%, and many people who considered investing told me the margins weren’t software margins and that was a red flag. What they didn’t understand is that the humans Botkeeper was using were labeling data as they worked. That data was used to build models, and those models replaced much of the human work. Four years later botkeeper had margins over 70 percent and looked economically more like a software company.

If you use SaaS valuation multiples and SaaS metrics as your default, you will misjudge these businesses because you will be evaluating them as something they are not.

Assumption #2: “AI Is One Thing”

“AI” is a label we use for a collection of technologies, not a single capability. That’s why you can hear one person say AI is terrible at something and hear another person say AI is great at that same thing.

Machine learning alone has different approaches, like supervised, unsupervised, and reinforcement learning. Neural networks have different architectures—as do natural language processing, computer vision, robotics, and others. Each can be called “AI,” but each functions in a radically different way.

In other words, using the word “AI” is like using the word “vehicle.” Do you mean a car, a boat, a train, an airplane, or a motorcycle? They all move you from one place to another, but they do so in radically different ways. They have different cost structures. They have different failure modes. They become commoditized at different timescales.

If you treat AI as a single bucket, you’ll keep making blanket statements that don’t hold up, and you’ll keep being surprised by results that look contradictory.

Assumption #3: “Predictable Adoption Curves”

Investors like smooth curves. If you get early growth data, you want to extrapolate it.

AI companies typically don’t behave that way. Their innovation path is discontinuous. Their adoption path is discontinuous. Earlier I mentioned the “jagged frontier,” a concept popularized in a 2023 researcher paper by Karim Lakhani from Harvard and Ethan Mollick from Wharton. It proved that the capabilities of AI models were not easily predictable. Consultants using ChatGPT at work often predicted it would be good at a task and it wasn’t, and often predicted it would be bad at a task that it actually performed well. This jagged frontier makes it hard to predict where AI goes next. The space of possibility in AI is still massive and largely unexplored. The possibility space is not even largely explored.

That’s why you can get shocks like DeepSeek. When DeepSeek—a Chinese lab—first emerged in 2023, the predominant narrative on AI was that it was expensive, expensive, expensive. Then, DeepSeek introduced a new approach, and the narrative changed overnight. The market was shocked at DeepSeek’s ability to match the output of OpenAI’s GPT‑4 for only $6 million in training costs—a fraction of the hundreds of millions of dollars OpenAI spent to train their models. This revelation sent the market reeling; Nvidia alone lost an estimated $600 billion in market cap in one day.

You don’t get that kind of discontinuity in most other industries. Nobody appears out of nowhere in real estate and says, “I can build the same building for 10 percent of the cost.” And you also don’t get moats that crumble this quickly. Coca‑Cola won’t vanish just because someone made a slightly better soda, for example.

In AI, you can watch companies spike revenue and then crash when a better product launches and users switch. Jasper raised $125 million at a $1.5 billion valuation, but its valuation declined sharply after ChatGPT launched its consumer product. One popular software coding agent is rumored to have gone fromCursor went from $0 to $400 million ARR to $200 million ARR in a matter of months. While each of these companies contained great promise, their performance in the market didn’t just depend on what they did, but on what other companies were doing as well.

If you build financial models that assume steady growth when investing in AI, you will get blindsided. You have to underwrite disruption as a normal part of the landscape.

Assumption #4: “Compute Doesn’t Matter”

In software, we haven’t typically treated compute—or computing power—as a moat. It was assumed to be replicable.

In AI, that assumption breaks. AI mModels are very large by computer program standards and often have to run across multiple machines – a single GPU won’t run a large model.. Bigger models require more machines and longer time periods. Training can take months. This represents a kind of digital capital intensity we haven’t had to model in most software businesses until now.

When we don’t have an equivalent, we’re not sure how to think about it. But the implication is straightforward: some incumbents will be hard to dislodge because they have the hardware, the compute capacity, and the ability to catch up even if they fall behind on a specific model. They have the raw materials. So, while many startups offer a lot of promise as disruptors, you can’t count companies like Google, Amazon, or Apple out of the fight just yet.

The Choice Ahead

If you’ve read this far, you’ve probably felt a particular kind of tension as we’ve worked through this chapter. Right now, you can see that AI is real and accelerating and that the old ways of valuing and evaluating software companies don’t seem to fit. Still, you’re not sure how to proceed: the demos are impressive, but the business models are unfamiliar and the path from “cool capability” to “durable returns” is harder to see.

My point with this chapter is simple: The confusion and uncertainty you’re feeling aren’t a sign that you’re behind. They’re a sign that the underlying rules are changing.

AI has repeatedly produced exciting promises that failed in practice. It behaves differently from traditional software and forces you to rethink where value accrues, how moats form, and what “progress” even entails. If you want to invest well in this environment, you need new mental models—and you need a decision about whether you’re going to keep using the old ones anyway.

Knowing this, right now you have two options.

Option 1: Keep using your existing mental models. Evaluate AI companies like software companies. Wonder why your picks underperform.

Option 2: Spend the next few hours learning new mental models that took us ten years and 135 AI investments to develop. Show up to your next AI investment conversation asking questions your competitors haven’t thought of.

Option 2 doesn’t guarantee you’ll be right every time. But it does guarantee you’ll be asking the right questions.

We learned what questions to ask the hard way. As we learned those lessons, we lost some money along the way. Fortunately, with a little bit of luck, we had some early wins—as well as some would-be losses that ultimately turned into wins. Those victories gave us the insight and awareness to stay in the game, refine our thesis, and keep learning.

And the most important lesson we learned?

Think about AI companies from first principles and not from comps.

Investors love to compare one company to another and place their bets accordingly. But that doesn’t work well in the AI space. A first principles approach allows us to break the comps trap and ask:

These are the kinds of questions that have helped us hit more than we’ve missed over nine years and 135 investments. They came at a cost of time, money, and experience. But that’s what makes them work: they’re battle-tested. We know they work because we know everything that doesn’t work.

We’ll start exploring the six key mental models for investing in AI in Chapter 3. But before we do, in Chapter 2, we’ll take a closer look at the work of AI right now—where we are, where we’re going, and why the prospect of artificial general intelligence (AGI) isn’t an end, but a beginning.

Thanks for reading.

Leave a Reply

Sign Up for TheVCDaily

The best news in VC, delivered every day!