How Agentic AI is Transforming The Startup Landscape with Andrew Ng

Hi, listeners. Welcome back to No Priors. Today, Elad and I are here with Andrew Ng. Andrew is one of the godfathers of the AI revolution. He was the co-founder of Google Brain, Coursera, and the venture studio AI Fund. More recently, he coined the term “agentic AI” and joined the board of Amazon. He was also one of the very first people, a decade ago, to convince me that deep learning was the future. Welcome, Andrew. Andrew, thank you so much for being with us.

No, always great to see you.

I’m not sure where we should begin because you have such a broad view of these topics, but I feel like we should start with the biggest question: If you look forward at capability growth from here, where does it come from? Does it come from more scale? Does it come from data work?

1. AI Progress Has Multiple Vectors

Multiple vectors of progress. I think there is probably a little bit more juice to squeeze out of the scalability lemon, so hopefully we’ll continue to make progress there, but it’s getting really, really difficult.

Society’s perception of AI has been very skewed by the PR machinery of a handful of companies with amazing PR capabilities. Because that number of companies drove scale as a narrative, people think of scale first as a vector of progress. But I think agentic workflows, the way we build multimodal models, and the work we have to do to build concrete applications are all multiple vectors of progress, as well as wild cards like brand-new technologies. For example, can diffusion models, which are used to generate images for the most part, also work for generating text? I think that’s exciting, so I think there will be multiple ways for AI to make progress.

You actually came up with the term “agentic AI.” What did you mean then?

When I decided to start talking about agentic AI, it wasn’t a thing when I started to use the term, and my team was slightly annoyed at me. One of my team members, whom I won’t name, said, “Andrew, the world does not need you to make up another term.” But I decided to do it anyway, and for whatever reason, it stuck.

The reason I started to talk about agentic AI was because, a couple of years ago, I saw people spending a lot of time debating, “Is this an agent? Is this not an agent? What is an agent?” I felt there was a lot of good work happening, and there was a spectrum of degrees of agency. There were highly autonomous agents that could plan and take multiple steps of reasoning to a final step by themselves, and then there were things with lower degrees of agency, where we’re prompting an LLM and affecting its output.

I felt that rather than debating whether something is an agent or not, we should just describe the degrees of agency and say it’s all agentic, so you can spend the time actually building this. So I started to push the term “agentic AI.”

What I did not expect was that, several months later, a bunch of marketers would get ahold of this term and use it as a sticker to stick on everything in sight. I think the term “agentic AI” really took off. I feel like the marketing hype has grown insanely fast, but the real business progress has also been rapidly growing—maybe not as fast as the marketing hype.

What do you think are the biggest obstacles right now to true agents actually being implemented as AI applications? To your point, I think we’ve been talking about it for a little while now. There are certain things that were missing initially that are now in place, in terms of everything from certain forms of inference-time compute through to forms of memory and other things that allow you to maintain some sort of state as you’re doing things. What do you view as the things that are still missing or need to get built, or what will foment progress on that end?

2. Talent Limits Agentic Workflows

I think at the technology-component level, there’s stuff that I hope will improve. For example, computer use kind of works and often doesn’t work. I think guardrails and evals are a huge problem: How do we quickly evaluate these things and drive evals? So the components have room for improvement.

But what I see as the single biggest barrier to getting more agentic AI workflows implemented is actually talent. When I look at the way many teams build agents, the single biggest differentiator that I see in the market is whether the team knows how to drive a systematic error-analysis process with evals. You’re building the agents by analyzing, at any moment in time, what’s working, what’s not working, and what you should improve, as opposed to less-experienced teams that try things in a more random way. Then it just takes a long time.

When looking across a huge range of businesses, small and large, it feels like there’s so much work that could be automated through agentic workflows. But the talent, the skills, and maybe the software tooling just aren’t there to drive that disciplined engineering process and get this stuff built.

How much of that engineering process could you imagine being automated with AI?

It turns out that a lot of the process of building agentic workflows requires ingesting external knowledge, which is often locked up in people’s heads. Until and unless we build AI avatars that can interview employees doing the work, and better visual AI that can look at the computer monitor, I think maybe eventually that will happen. But at least right now, for the next year or two, I think there’s a lot of work for human engineers to do to build more agentic workflows.

And so that’s more the kind of collection of data, feedback, et cetera, for certain loops that people are doing? Are there other things? I’m sort of curious what that translates into tangibly versus—

I’ll give you one example. I see a lot of workflows where maybe a customer emails you a document, you’re going to convert the document to text, then maybe do a web search for some compliance reason to see if you’re working with a vendor you’re not supposed to be working with, and then look up a database record, see if the pricing is right, save it somewhere else, and so on.

So it’s a multistep agentic workflow, kind of next-generation robotic process automation. We implement this and it doesn’t work. Is it a problem if you got the invoice date wrong? Is that a problem if you routed a message to the wrong person for verification?

When you implement these things, almost always they don’t work the first time. But then you have to know what’s important for your business process. Is it okay that I bothered the CEO of the company too many times, or does the CEO of the company mind verifying some invoices? All that external contextual knowledge often requires thoughtful human product managers or human engineers, at least right now, to think through and make these decisions. So can an AI agent do that someday? I don’t know. It seems pretty difficult right now. Maybe someday.

But it’s not in the internet pretraining dataset, and it’s not in a manual that we can automatically extract.

I feel like for a lot of the work involved in building agentic workflows, that dataset is proprietary. It’s just not general knowledge on the internet. Figuring that out is still exciting work to do.

What is, if you just look at it, the spectrum of agentic AI? What’s the strongest example of agency you’ve seen?

3. Coding Agents Lead The Pack

At the bleeding edge of agentic AI, I’ve been really impressed by some of the AI coding agents.

In terms of economic value, I feel like there are 2 very clear and apparent buckets. One is answering people’s questions. Probably OpenAI’s ChatGPT seems to be the market leader in that, with real takeoff velocity.

The second massive bucket of economic value is coding agents. My personal favorite coding agent right now is Claude Code. Maybe it’ll change at some point, but I just use it. Love it. It’s highly autonomous in terms of planning out what to do to build the software, building a checklist, and going through it one at a time.

This ability to plan a multistep task and execute the multiple steps of a plan is one of the most highly autonomous agent capabilities out there that is actually being used and actually works. There’s other stuff that I think doesn’t work, like some computer-use stuff. For example, “Go shop for something for me and browse online.” Some of those things are really nice demos, but not yet production-ready.

Do you think that’s because of looser criteria in terms of what needs to be done and more variability around actions, or do you think there’s a better training set or sort of set of outputs for coding? I’m curious why one works so well, or almost feels magical at times, and others are really struggling as use cases so far.

I think engineers are really good at getting all sorts of stuff to work.

Mm-hmm.

But the economic value of coding is just clear, apparent, and massive. So I think the sheer amount of resources dedicated to this has led to a lot of smart people for whom they themselves are the user, so they also have good product instincts, building really amazing coding agents.

You don’t think it’s a fundamental research challenge. You just think it’s capitalism at work and domain knowledge in a lab.

Oh, I think capitalism is great at solving fundamental research problems.

At what point do you think models will effectively be bootstrapping themselves, in terms of 99% of the code of a model being written by agentic coding agents?

So I suspect we're slowly getting there. Some of the leading foundation model companies are clearly—or they've said publicly—that they're using AI to write a lot of the code. One thing I find exciting is AI models using agentic workflows to generate data for the next generation—

Mm-hmm.

—of models. I think the Llama research paper was talking about this, where an older version of Llama would be used to think for a long time to generate puzzles that you then train the next generation of the model to try to solve really quickly, without needing to think as long. So I find that exciting, too.

Multiple vectors of progress. It feels like AI is not just one way to make progress. There are so many smart people pushing forward in so many different ways.

I think you have rejected the term “vibe coding” in favor of “AI-assisted coding.” What's the difference?

Vibe coding leads people to think, “I'm just gonna go with the vibes and accept all the changes that Chris has suggested,” or whatever, and it's fine. Sometimes you can do that, and it works, but I wish it was that easy.

So when I'm coding for a day or for an afternoon, I'm not going with the vibes. It's just a deeply intellectual exercise, and I think the term “vibe coding” makes people think it's easier than it is. Frankly, after a day of using AI-assisted coding, I'm mentally exhausted, right? So I think of it as rapid engineering, where AI is letting us build serious systems and products much faster than ever before, but it is engineering, just done really rapidly.

Do you think that's changing the nature of startups—how many people you need, how you build things, how you approach things—or do you think it's still the same old kind of approach, but you just have people who get more leverage because they have these tools now?

4. Product Management Becomes The Bottleneck

So, finally, we build startups, and it's really exciting to see how rapid engineering and AI-assisted coding are changing the way we build startups. There are so many things that would've taken a team of 6 engineers 3 months to build that today, 1 of my friends or I—we just build it in a weekend.

Mm-hmm.

The fascinating thing I'm seeing is, if we think about building a startup, the core loop of what we do, right? I want to build a product that users love. So the core iteration loop is writing software—it's software engineering work—and then the product managers may go do user testing, look at it, go by gut, whatever, to decide how to improve the product.

When we look at this loop, the speed of coding is accelerating and the cost is falling.

Mm-hmm.

So increasingly, the bottleneck is actually product management. The product management bottleneck is now that we can build what we want much faster, while the bottleneck is deciding what we actually want to build. Previously, if it took you, say, 3 weeks to build a prototype and you needed 1 week to get user feedback, it's fine. But if you now build a prototype in a day, then boy, if you have to wait a week for user feedback, that's really painful.

I find my teams increasingly relying on gut because we go and collect a lot of data that informs our very human mental model—our brain's mental model—of what the user wants. Then we often have to have deep customer empathy, so we make product decisions like that, right? Really, really fast in order to drive progress.

Have you seen anything that actually automates some aspects of that? I know there have been some versions of things where people, for example, are trying to generate market research by having a series of bots react in real time, and that almost forms your market or your user base as a simulated environment of users. Have you seen any tooling like that work or take off, or do you think that's coming, or do you think that's too hard to do?

Yeah. There are a bunch of tools trying to speed up product management. The recent Figma IPO is a great example of design being highly enabled by AI. Dylan did a great job.

Then there are these tools that are trying to use AI to help interview prospective users. As you say, we looked at some of the scientific papers on using a flock of AI agents—

Uh-huh.

—to simulate a group of users and how to calibrate that. It all feels promising and early, and hopefully wildly exciting in the future, but I don't think those tools are accelerating product managers nearly as much as coding tools are accelerating software engineers. So this does shift more of the bottleneck onto the product management side.

It does make sense to me that my partner, Mike, has this idea that I think is broadly applicable in a couple of different ways: computers can now interrogate humans at scale. There are companies like Listen Labs working on this for consumer research-type tasks, right? But you could also use it to—

Mm-hmm.

—understand tasks for training, or for the data collection piece that you described. When you think about your teams that are in this iteration loop, has the founder profile that makes sense changed over time?

5. Technical Founders Gain The Edge

To me, there are so many things that the world used to do in 2022 that just do not work in 2025. In fact, I often ask myself, is there anything we're doing today that we're also doing in 2022? If so, let's take a look and see if it still even makes sense today, because a lot of stuff, a lot of workflows in 2020, don't make sense today.

I think today, the technology's moving so fast that founders who are on top of gen AI technology—tech-oriented product leaders—are much more likely to succeed than someone who may be more business-oriented and more business-savvy, but doesn't have a good feel for where AI is going. Unless you have a good feel for what the technology can and cannot do, it's really difficult to think about strategy and where to lead the company.

We believe this, too.

Yeah.

Cool.

Yeah. Yeah. Yeah.

I think that's old-school Silicon Valley, even. If you look at Gates, Steve Jobs and Steve Wozniak, or a lot of the really early pioneers of the semiconductor, computer, and early internet eras, they were all highly technical.

Yeah.

So I almost feel like we lost that for a little while, and now it's very clear that you need technical leaders for technology companies.

I think we used to think, “Oh, they've had 1 exit before, or 2 exits even, so let's just back that founder again.” But I think if that founder has stayed on top of AI, then that's fantastic.

I think part of it is that, in moments of technological disruption, when AI is rapidly changing, that's rare knowledge. So, actually, take mobile technology. Everyone knows what a mobile phone can and cannot do, right? They know what a mobile app is—there's GPS, all that. Everyone knows that. So you don't need to be very technical to have a gut feeling about whether you can build a mobile app for something.

Mm.

But AI's changing so rapidly. What can you do with a voice app? What can engineering workflows do? How rapidly are foundation models changing? What is a reasoning model? Having that knowledge is a much bigger differentiator.

Yeah. It's an interesting point, because when I look at the biggest mobile apps, they were all started by engineers. WhatsApp was started by an engineer, Instagram was started by an engineer. I think Travis at Uber was technical-ish.

Technically adjacent. Yeah.

Technically adjacent. Instacart—Purva was an engineer at Amazon.

Yeah. And Travis had the insight that GPS enabled a new thing.

Yeah.

But you had to be 1 of the people who saw GPS or mobile coming early to go and do that.

Yeah.

Yeah.

You have to be really aware of the capabilities.

Yeah, you have to know the technology.

That's what I'm hearing. Yeah.

Yeah, it's super interesting. What other characteristics do you think are common? I know people have been talking about, for example, it almost felt like there was an era where being hardworking was kind of poo-pooed. Do you think founders have to work hard? I'm just curious: aggression, hours worked—what else may correlate or not correlate in your mind?

I work very hard. There are periods in my life where I encourage others who want to have a great career and have an impact to work hard. But even now I feel a little bit nervous just saying that because in some parts of society, it's considered not politically correct to say, “Well, working hard probably correlates to your personal success.”

I think it's just a reality.

Yeah.

I know that not everyone at every point in their life is in a time when they work hard.

Mm-hmm.

When my kids were first born—

Yeah.

That week, I did not work very hard. It was fine, right? So, acknowledging that not everyone is in a circumstance where they can work hard, the factual reality is that people who work hard accomplish a lot more. But of course, you need to respect people who aren't in a phase where they—

Mm.

I'd say something maybe a little less politically correct, which is, I think there was an era where people thought startups were for everyone, and I do not believe that's true, right? You're trying to do a very unreasonable thing—

Yeah.

Create a lot of value, impacting people very quickly. When you're trying to do an unreasonable thing, you probably have to work pretty hard, right? I think the work ethic required to move the needle in the world very quickly disappeared.

Yeah. There was a quote—I wish I remembered who said this. Was it, “The only people who would change the world are the ones crazy enough to think they can”? I think it does take someone with the boldness and decisiveness to say, “You know what? To save the world, I'm going to take a shot at changing it.” There are only people with that conviction who I think can do this.

Strikes me as being true in any endeavor. I used to work as a biologist, and I think it's true in biology, technology, and almost every field that I've seen. The people who work really hard do very well.

Then, in startups at least, the thing I tended to forget for a while was just how important competitiveness—or people who really wanted to compete and win—was. Sometimes people come across as really low-key, but they still have that drive and that urge, and they want to be the ones who are the winners. I think that matters. Similarly, that was put aside for a little bit, at least from a societal perspective relative to companies.

Actually, I've seen two types. One is that they really want their business to win. That's fine; some do great. Others really want their customers to win.

Mm.

They're so obsessed with serving the customer that it works out. I usually say that in the early days of Coursera, yes, I knew about competition, but I was really obsessed with learners, with the customers, and that drove a lot of my behaviors.

No, that's a really good framework. When I say competition, I don't necessarily mean with other companies. It's almost with whatever metric you set for yourself, or whatever thing you want to win at or be the best at.

One thing I found is that in the startup environment, you just have to make so many decisions every day. You have to go by gut a lot of the time, right? I feel like building a startup feels more like playing tennis than solving calculus problems. You just don't have time to think; you just have to make a decision.

This is why people who obsess day and night over the customer and the company think really deeply and have that conceptual knowledge that, when someone says, “Do I ship product feature A or feature B?”—

You feel it now, yeah.

You just have to know all the time. Not always. It turns out there are so many, to use Jeff Bezos' term, two-way doors in startups because, frankly, you have very little to lose.

Mm-hmm.

So just make a decision. If it's wrong, change it a week later; that's fine. To be really decisive and move really fast, you need to have usually been obsessed with the customer, maybe the technology, to have that set of knowledge to make really rapid decisions—

Okay.

—and still be right most of the time.

How do you think about that bottleneck in terms of product management that you mentioned, or people who have good product instincts? I was talking to one of the best-known tech public company CEOs, and his view was that in all of Silicon Valley, or in all of tech globally, there's probably a few hundred, at most, great product people.

Do you think that's true, or do you think there's a broader swath of people who are very capable at it? And then how do you find those people? I think that's actually a very rare skill set in terms of the people who are—just like there's a 10x engineer, there's 10x product insight, it feels.

Boy, that's a great question. I feel it's got to be more than a few hundred great product people. Maybe more—just as I think there are way more than a few hundred great AI people.

I think one thing I find very difficult is that user empathy, or that customer empathy. To form a model of the user or the customer, there are so many sources of data. You run surveys, you talk to a handful of people, you read market reports, and you look at people's behavior on other parallel or competing apps or whatever. There are so many sources of data.

But to take all this data and then get out of your own head to form a mental model for what your ideal customer profile, or some user you want to serve, would think and act like so you can very quickly make decisions and serve them better—that human empathy is difficult.

One of my failures, one of the things I did not do well in the early phase of my career, was that, for some dumb reason, I tried to make a bunch of engineers into product managers. I gave them product manager training, and I found that I had foolishly made a bunch of really good engineers feel bad for not being good product managers, right?

Mm-hmm.

But I found that one correlate of whether someone would have good product instincts is a very high level of human empathy, where you can synthesize lots of signals to really put yourself in the other person's shoes, and then rapidly make product decisions on how to serve them.

Going back to coding assistance, I think it's reasonably well known that the Cursor team makes their decisions very instinctively, rather than spending a lot of time talking to users. I think that makes sense if you are the user, and then your mental model of yourself and what you want is applicable to a lot of people.

Similarly, these things change all the time, but I don't think Cloud Code incorporates, despite the scale of its usage, feedback data today from a training-loop perspective. That surprises people because it is really just, “What do we think the product should be at this stage?”

One advantage that startups have is that, while you're early, you can serve one user profile. Today, if you're Google, Google serves such a diverse set of user personas that you really have to think about a lot of different user personas, and that adds complexity to the product changes.

But when you're a startup trying to get your initial wedge in the market, if you pick even one human who is representative enough of a broad set of users and just build a product for one user—one ideal customer profile, one hypothetical person—then you should actually go quite far.

For some of these businesses, be it Cursor or Cloud Code or something, if they have internally a mental picture of a user that's close enough to a very large set of prospective users, you can actually go really far that way.

6. AI Forces Leaders To Adapt

The other thing that I've observed, and I'm curious if you guys see this in some of our companies, is just that the floor is lava, right? The ground is changing in terms of capability all the time, and the competition is also very fierce in the categories that are already obviously important and have multiple players.

Leaders who were really effective in companies a generation ago are not necessarily as effective when recruited into these companies as they're scaling, because the pace of operation, or the pace of change, is so high. It's interesting to see you say, “I'm looking at what I was doing today and in 2022 and asking, ‘Is that still right?’” Versus if you're an engineering leader or a go-to-market leader and you've built your career being really great at how that's done, that may not be applicable anymore.

I think it's a challenge for a lot of people. I know many great leaders in lots of different functions who are still doing things the way they were in 2022, and I think it's just got to change.

When new technology comes, once upon a time there was no such thing as web search. Today, who would you hire for any role that doesn't know how to search the web, right? I think we're well past the point where, for a lot of job roles, if you can't use LLMs effectively, you're just much less effective than someone who can.

And as it turns out, everyone on my team at AI Fund knows how to code. Everyone has a GitHub account. I see that for a lot of my team members—my assistant general counsel, my CFO, or my front-desk operator—when they learn how to code, they're not software engineers, but they do their job function better.

By learning the language of computers, they can now tell a computer more precisely what they want it to do for them, and the computer will do it for them. This makes them more effective at their job function.

Mm-hmm.

I think the rapid pace of change is disconcerting to a lot of people. But I feel like when the world is moving at this pace, we just have to change at the pace that the world demands of us.

Yeah, I've seen that, to your point, show up in hires, particularly around product—or product and design. So, one...

...sort of later-stage AI company I'm involved with, they were doing a search for somebody to run product and somebody to run design, and in both cases they selected for people who really understood how to use some of the vibe-coding and AI-assisted coding tools.

Mm-hmm.

Because they said your point. You can prototype something so rapidly, and if you can't even just mock it up really quickly to show what it could look like or feel like or do in a very simple way, you're wasting an enormous amount of time talking and writing up the product requirements document and everything else.

I do think there's a shift in terms of how you even think about what processes you use to develop a product or even pitch it, right? What should you show up with to a meeting when you're talking about a product for the first time?

The whole thing, apparently.

It's completely changed. Yeah, no, you should have a prototype in some cases.

Actually, I'll just give you an example. Recently, I was interviewing engineers for a role. I interviewed someone with about 10 years of experience, full-stack, with a very good résumé. I also interviewed a fresh college grad.

The difference was that the person with 10 years of experience had not used AI tools much at all. The fresh college grad had. My assessment was that the fresh college grad who knew AI would be much more productive, and I decided to hire them instead. It turned out to be a great decision.

Now, the flip side of this is that the best engineers I work with today are not fresh college grads. They're people with 10, 15, or more years of experience, but they're also really on top of AI tools. Those engineers are just completely in a class of their own.

I actually think software engineering is a harbinger of what will happen in other disciplines because the tools are most advanced in software engineering.

It's interesting. One company that I guess both of us are involved with is called Harvey, and I led their Series B. When I did that, I called a bunch of their customers, and the thing that was most interesting to me about some of those customer calls was that legal is notorious for being a tough profession when it comes to adopting new technology, right? There aren't a dozen great legal software companies.

Those customers that I called—big law firms or people who were quite far along in terms of adopting Harvey—all thought this was the future. They all thought that AI was really going to matter for their vertical. The main thing they would raise were questions like, “In a world where this is ubiquitous, suddenly, instead of hiring 100 associates, I only hire 10. How do I think about future partners and who to promote if I don't have a big pool?”

I thought that mindset shift was really interesting. To your point, I feel like it's percolating into all these markets or industries. It's slowly happening, but industry by industry, people are starting to rethink aspects of their business in really interesting ways. It'll take a decade, 2 decades, for this transformation to happen.

But it's compelling to see how people in the earliest-adopting verticals are thinking most deeply about it.

That should be really interesting. I actually have a legal startup called Callauses AI that AI Fund helped build. It's doing very well as well.

I think the nature of work in the future will be very interesting. I feel like a lot of teams wound up outsourcing a lot of the work, partly because of costs. But with AI and AI assistance, part of me wonders: Is a really small, really skilled team with lots of AI tools going to outperform a much larger, maybe lower-cost team that may or may not be—

And they have less coordination cost.

Yeah. So, actually, some of the most productive teams I'm on now, teams that I'm a part of, are some of the smallest teams. Very small teams of really good engineers with lots of AI enablement and very low coordination cost because everyone's together in person.

Mm.

We'll see how the world evolves. It's too early to make a call, but you can see where I'm thinking the world may or may not be headed.

I work with several teams now, one of which is called OpenEvidence and has pretty good penetration—50% of doctors in the US now—where it's an explicit objective in the company to try to be as small as possible as they grow their impact. We'll see where these companies land because there are lots of functions that need to grow in a company over time. But that certainly wasn't an objective for, like—

I've heard that objective a lot.

...5 years ago.

I've actually heard that objective a lot in the 2010s, and there are a bunch of companies that I actually think underhired pretty dramatically or stayed profitable and would brag about being profitable, but growth wasn't as strong as it could be. So I actually feel like that's a trap. Partly from—

How would you calibrate, then?

...helping think about this.

Yeah.

It's basically—it's almost: Are you being lackadaisical or too accepting of the progress that your company's making because it's going just fine? It could be going much better, but it's still going great on a relative basis. And so you're like, “Oh, I'll keep the team small. I'll be super lean. I won't spend any money. Look at me, how profitable I am.”

Sometimes it's amazing, right? Capital efficiency is great. But sometimes you're actually missing the opportunity or not going as fast as you can.

Usually, I think what happens is that in the early stage of a startup's life, you're competing with other startups. If you're way ahead, it feels great. But eventually, if there are incumbents in your market, they come in, and the faster you capture the market and move upmarket, the less time you give them to realize what's going on and catch on.

Often, 5, 6, or 7 years into the life of a startup, you're suddenly competing with incumbents, and they just kill you with distribution or other things. I think people really miss the mark, and you could argue that was kind of Slack versus Teams.

There's a few companies I won't name, but I feel like they were so proud of their profitability and they kind of blew up. I guess on the design side, that was Sketch, right? Remember? They were—

Yeah, Bohemian Coding, yeah.

They were based in the Netherlands. They were super happy. They were profitable. They were doing great, and then the Figma wave kind of came. And then—

Do you think your companies stay this small?

What? Do I think what?

Do you think your teams stay this small?

Do I think my teams stay this small?

Yeah.

What do you mean?

In terms of just efficiency, can you actually affect millions and billions of people with 10-, 50-, or 100-person teams?

I think teams can definitely be smaller now than they used to be, but are we overinvesting or underinvesting? And then also, to your point, to the analysis of market dynamics, right? If it's a winner-take-all market, then the incentives just—

Gotta go.

Yeah, it's gotta go.

You gotta move.

Yeah.

Minecraft, I think, when it sold to Microsoft, was how many people? Like 5 people or something. It sold for a few billion dollars, and it was massively used. I think people forget all these examples, right?

Yeah.

It's just this: “Oh, suddenly you can do things really lean.” You could always do things lean before. The real question is: How much leverage did you have in headcount? How did you distribute? What did you actually need to invest money behind?

And then I would almost argue that one of the reasons small teams are so efficient with AI is because small teams are efficient in general. They didn't hire 30 extra crappy people who get in the way, and I think often people do that.

If you look at the big tech companies right now, many—not all of them, but many of them—could probably shrink by 70% and be more effective, right? And so I do think people also forget the fact that, A, there's AI efficiency; B, there's high-value capital being arbitraged into markets that normally wouldn't have it.

Legal is a good example. Great engineers didn't want to work in legal. Now they do because of things like Harvey.

Or healthcare.

Or healthcare, which, again—

Yeah.

Suddenly, you have these great people showing up. But I think the other part of it is that small teams tend to be more effective, and AI helps you argue for other reasons to keep teams highly small and performant, which I think is under-discussed.

Yeah.

I feel like one of the reasons why that AI instinct is so important is that I remember one week when I had two conversations with two different team members. One person came to me and said, “Hey, Andrew, I’m going to do this. Can you give me some more headcount to do this?”

Mm.

I said, “No.” Later that week, independently, someone else said something very similar: “Hey, Andrew, can you give me some budget to hire AI to do this?”

Yeah.

I said, “Yes.” That realization—that you hire AI, not a lot more humans, for this—you just have to have those instincts.

Yeah.

You hire AI—

Yeah.

—not a lot more humans for this. You just have to have those instincts.

Yeah, that’s very interesting.

If you think of what’s happening in software engineering as the harbinger for the next industry transformations, you spend a lot of time investing at the application level, or building things there. What do you think is next? Or what do you want to be next?

I feel like there’s a lot of—

I’d actually prefer a ranked list for all of us investing in this stuff.

Oh.

You know, for—

Yeah. Oh, sure.

7. AI Targets Rich Environments

There’s actually one thing I find really interesting: We have economists doing all the studies on whether the jobs at highest risk of AI disruption. I think you’re skeptical. I actually look at them sometimes for inspiration for where—

Yeah.

Where we should find ideas to build projects.

That’s true.

One of my friends, Eric Brynjolfsson, is often very insightful on the nature of our work.

Yeah, I like him. Yeah.

Yeah, good. I find talking to him sometimes useful. Although, actually, one of the lessons I’ve learned is that, in view of top-down market analysis, I think AI will target rich environments. There are so many ideas that no one’s working on yet because the tech is so new.

One thing I’ve learned is that at AI Fund, we have an obsession with speed. All my life, I’ve had an obsession with speed, but now we have tools to go even faster than we could. One of the lessons I’ve learned is that we really like concrete ideas.

So if someone says, “I did the market analysis. AI will transform healthcare,” that’s true, but I don’t know what to do with that.

Mm.

But if a subject-matter expert or an engineer comes and says, “I have an idea. Look at this part of healthcare operations and all this,” then I go, “Great. That’s a concrete idea.”

I don’t know if it’s a good idea or a bad idea, but it’s concrete. At least we could very efficiently figure out: Do customers want this? Is it technically feasible? And get going.

So at AI Fund, when we’re trying to decide what to build, we screen a long list of ideas to select a small number that we want to move forward with. We don’t like looking at ideas that aren’t concrete.

What do you think investing firms or incubation studios like yours will not do 2 years from now? Not do manually, sorry.

I think a lot could be automated, but the question is: What are the tasks we should be automating? For example, we don’t make follow-on decisions that often, right, because of a portfolio of some dozens of companies. So do we need to fully automate that? Probably not, because we’ve already looked at it. It’s pretty hard to automate.

I feel like doing deep research on individual companies and competitive research seems ripe for automation. I personally use OpenAI’s Deep Research and other deep-research types of tools a lot to do at least cursory market research. LP reporting is a massive amount of paperwork that maybe we could simplify.

Yeah. I’m taking the strategy of general avoidance, besides basic compliance. One of my partners, Bella, worked at Bridgewater before, where they had an internal effort to take a chunk of capital and then try to disrupt what Bridgewater was doing with AI.

It’s macro investing, so it’s a very different style, but I think it probably gives us some indications about where the human-judgment piece for our business is not obvious. Does an entrepreneur have the qualities we’re looking for? When you’re a new grad, your résumé on paper, your GitHub, or whatever work history you have isn’t very indicative.

People have other ideas for doing this. I know investors who are looking at recordings of meetings with entrepreneurs and seeing if they can get some signal from communication style, for example. But I think that part is very hard.

I do think you can be programmatic about looking at materials, for example, and ranking the overall quality of teams.

There’s actually one thing. I feel like our AI models are getting really intelligent, but there’s a set of places where humans still have a huge advantage over AI, especially if the human has additional context that, for whatever reason, the AI model can’t get at.

It could be things like meeting the founder and sussing out how they are as a person, their leadership qualities, their communication, or whatever. Maybe reviewing video will eventually allow us to get that context into an AI model.

But I find that, in all these things, as humans, we do a background reference check, and someone makes an offhand comment that we catch, which affects a decision. How does an AI model get this information? Especially when a friend will talk to me, but they don’t really talk to my AI model.

I find that there are a lot of these tasks where humans still have a huge information advantage because they haven’t figured out the plumbing or whatever’s needed to get information to the AI model.

The other thing I think is very durable is things that rely on a relationship advantage. If I’m convincing somebody to work at one of my companies, and they worked at a previous company and trust me because of it, or for whatever reason, all the information in the world about why this is a good opportunity isn’t the same thing as me saying, “Sally, you’ve got to do this. It’s going to work.”

It remains to be seen whether company building is actually that correlated with investment returns, but I do think that side of it feels harder to fully automate.

Yeah. Yeah. No, I think trust—because people know you, and people do trust you. I trust you, right? You can only say so many things; it’s very easy to lose trust. So that makes sense.

But actually, one thing I’m curious to get your take on is that we increasingly see highly technical people try to be first-time founders. How do you set up the processes to help first-time founders learn all the hard lessons and all the craziness needed to be a successful founder?

I spend a lot of time thinking through how to set up founders for success when they have 80% of the skills needed to be really great, but there’s another little bit that we can help them with.

That’s a very manual process.

I don’t sweat it.

You don’t sweat it?

I just view it as a mix of peer groups. Can you surround people with other people who are either similar to them or 1 or 2 steps ahead of them on the founding journey? The second thing is complementary hires.

I think, in general, one of my big learnings is that early in their careers, people try to build out the skill set they don’t have. Late in their careers, they lean into what they’re really good at, and then they hire people to do the rest.

If the company’s working, I think you just hire people. Bill Gates would notoriously talk about how his COO was always the person he learned the most from. Then, once he’d reached a certain level of scale, he’d hire his next COO.

I almost view it through that lens for founders.

I see. Yeah.

Yeah, complementary hires make sense. But I think the best way to learn something is to do it, so just go. You’ll screw it up; it’s fine. As long as it’s not existential to the business, who cares?

I tend to be very lackadaisical.

I probably—

So I tend to be very lackadaisical.

I probably think too many things are existential for companies.

Yeah, it’s something like: Do you have customers, and are you building product?

To most of it, yeah.

Are you building a product that users love, right? And then, of course, go-to-market is important and all that is important, but you solve for the product first. Usually, sometimes you can figure out the rest, too.

I agree with that most of the time, but not always. I think there are some counterexamples, but I generally agree with you.

No, yeah. Sometimes you can build a sucky product—

Yeah.

—and have a sales channel you can force it through. But I’d rather not. That’s not my default model of operation.

Yeah, I don’t want that either. It does work. There’s a lot of really bad technology in big companies right now.

Yeah.

Yeah.

Okay, if you have these first-time, very technical founders with gaps in their knowledge or skill set as the core profile of the folks you’re backing, do you augment them somehow? What helps them when they begin?

I think a lot of things. That’s actually one thing I realized: at venture firms and venture studios, we do so many reps that we just see a lot that even repeat founders have only done once or twice in their lives.

I find that when my firm sits alongside the founders and shares our instincts on when to get customer feedback faster, whether they’re really on top of the latest technology trends, how to speed things up, or how to fundraise, that helps. Most people don’t fundraise that much in their lives, right? Most founders just do it a handful of times. That helps even very good founders with things that, because of what we do, we’ve had more reps at.

And then I think hiring others around them and building a peer group—I know these are things that you guys do. I think there’s a lot we could do. It turns out even the best founders need help. Hopefully, VCs and venture studios can provide that to great founders.

You’re a lot wiser about this than I am. I mean, I can’t help myself but want to specifically try to upskill founders on a few things they have to be able to do, like recruiting, right? But I would agree that the higher-leverage path is absolutely that you can put people around yourself to do this and learn it on the job.

Last question for you: What do you believe about the broad impact of AI over the next 5 years that you think most people don’t?

I think many people will be much more empowered and much more capable in a few years than they are today. The capability of individuals who embrace AI will probably be far greater than most people realize.

Two years ago, who would have realized that software engineers would be as productive as they are today when they embrace AI? I think in the future, people in all sorts of job functions, and also people doing personal tasks, will just be so much more powerful and so much more capable than they’re probably even imagining.

Awesome.

Exciting. Yeah.

Thanks, Andrew.

Thanks for joining.

Thanks. Thanks a lot. Thanks, Sarah.