The 7 Most Powerful Moats For AI Startups

8,637 words · about 35 min to read

Moats protect AI startups from infinite competition

0:00 This idea of moes is so pervasive and so important.

0:03 It is interesting how moes have just become much more discussed by aspiring startup founders now than they were pre- AI.

0:11 What is going to prevent you from being basically subject to infinite competition?

0:16 Like a mode is inherently a defensive thing and you have to have something to defend otherwise like

0:20 if you got nothing to defend, don't worry about your mode. Welcome back to another episode of the Light Cone. Today we're going to talk about Moes. So, in your head you might be thinking about barbarians storming your gate. You've got this little startup and you've got every other company out there who wants to come and eat your lunch. Uh, and you know, right outside your castle is a moat that keeps them away. Jared, when you were going to college campuses, this isn't sort of this trivial thing that people are thinking about. It's actually uh something that keeps them from starting companies right now.

1:01 Yeah, this is a question that we got from a lot of very smart college students on on our on our our recent scholarships. And basically, their question is like they don't see how these new AI agent companies like a lot of the ones that we've talked about on on this podcast could have moes. um it plays into this meme of like the chat GBT rapper that like all of these companies could be easily cloned and so they can see how you could build a business that makes some amount of revenue, but they don't really see how you can build a long enduring business.

1:28 And so I think it's actually not true. I actually think these businesses do have quite deep and interesting modes, but they're not totally obvious what they would be. So I think this is an interesting topic for us to to explore. At our recent AI startup school backstage, I had this exchange with Sam Alman that I thought was kind of funny. You know, we spend a lot of time thinking about, you know, make something people want. Very simple maxims that are sort of anti- business school. And yet this idea of moes is so pervasive and so important. We sort of remarked how funny it is that uh one of the more important books to read these days is actually business school fodder. um this book called the seven powers. So today we thought that we would actually go through those seven powers. What are they? What are some concrete examples and ways that a startup founder who's just starting out uh could or should be thinking about these things from real world examples that we've seen.

2:26 So Diana, can you tell us a bit about this book?

2:29 This book was written by Hamilton Helmer who taught at Stanford Economics School and was published in 2016. And the book title was the seven powers the foundations of business strategies. And a lot of the examples are more with the era of uh internet companies from the 2000s. So a lot of the examples are like Oracle Facebook Netflix, which is a older generation. So we want to do a bit of a reboot right now how he applies now 2025 with AI.

3:01 I think it's a little bit confusing the way he uses the terminology in the book. It's called the seven powers, but it would make a lot more sense if he just called the thing the seven moes because that's really what he's talking about. He's really talking about seven categories of moes that a business can have. And I it's true that the examples are out of date, but I think the framework is actually pretty timeless. Like it turns out there's just only so many kinds of modes that a business can have and they don't really change. And so like even though the specific like versions of these modes are different in the AI agent world, like the categories haven't changed. Thankfully, we live in a world where there's markets and there's free markets where there's lots and lots of competition. And these moes in a lot of ways are the only way if you're running a business, you can sort of fight against all of the other people who might want to do exactly what you're doing. And um you know, famously Peter Teal talks about uh competition is for losers. And so the profound view there is that given infinite competition, what is going to prevent you from being basically subject to infinite

4:02 competition and then as a result uh you know your margins, how much you can actually profit off of what you're selling goes down to zero. And what that means is like actually your business will die. And so you know having a moat is uh relatively existential eventually. You made a great point earlier, Gary,

Speed is the only early-stage moat

4:21 that like this is actually like you kind of have to worry about this at the right time of of a startup. Do you want to talk about like how like early stage founders should think about Moes?

4:31 I mean, this is sort of why we generally tell people to go find a person with a real problem and then go solve that problem first. It's um what's funny about the world uh that's a little surprising is that you can go almost anywhere and find some painoint, some problem that could be solved with software and especially with AI that frankly just isn't being solved. And if that and they're they're so numerous and so severe that if you find that thing and solve it, you literally can mint a billion dollar or 10 billion or even hundred or hundreds of billions of dollars uh market cap business and it's just lying in plain sight. That's really the first thing that people should do.

5:15 Like you should just find a problem and go solve it. And then along the way you will probably as you work with customers as as you build the product itself and engineer it and figure out what data you need for it and all of these things like you will stumble upon these seven powers.

5:32 Yeah. The moes come later like it would be like pretty dumb for somebody to decide not to work on a startup idea because they can't see what the long-term modes of that idea could be. Right. It is interesting how moes have just become um much more discussed by aspiring startup founders now than they were pre- AI. Seems like the main reason for that presumably is just that big the original chat GBT rapper meme and that the the moat that most people are worried about is moat against the big model companies and how like are you not going to get crushed by one of the big labs when they decide the product you're working on is really valuable and they want to own it too. And I think Varun uh from Windsorf who we hosted some time ago he said it himself the early stages at the beginning the only model that startups have is really just speed. Once you pass that and build something that people want then you figure out and go deeper into these type of modes that we're going to discuss.

6:31 I really like Verun's point that the only moat is speed. That is not one of the seven powers in the book, but I think it probably should be.

6:39 I think it also comes with a lot of the essays from PD because one of the tenants really at the beginning is yes, your big company, let's say OpenAI at this OpenAI is the new Google. It's like sure OpenAI or Anthropic could build all these features let's say like cloud code and then compete directly let's say with cursor or etc. And for a startup like cursor to really win even in the beginning is they had relentless execution because a larger company like a Google or Anthropic it they just have a lot of uh more craft that they need to do in order to ship a product. They just have all these product managers all the operations. It needs to go through a PRD some spec dog and it takes mo a lot more time to ship a feature as opposed to cursor. the incredible story about

1 Highlight · Cursor's one-day product development cycles

7:28 cursor. When we hosted Michael Truel to come talk to the badge, he was sharing how his product development cycle for shipping features and sprint cycles were one day

7:39 one day. So one day sprint

7:42 in the at the beginning during a 2023 2024 around era they would start the every day would restart the clock and try to ship things every day. I mean that's like insane speed. Like there's no big company that could ship something at that speed at most weeks, couple weeks and maybe like larger companies. I don't know your Google maybe like

8:04 multiple months or sometimes years. I mean they had Google Bard or Gemini a long time ago that took years to get out right

8:11 I think cursor and winds are great examples of when you should start thinking about the modes because for the first few years I don't think it really matter that much they just had to like they proved out that hey like codegen is going to be a really valuable application of AI the development environment is going to be very very important to own. they like got rapid growth and then it's only when they're at scale that you know like they have to start thinking about how are we going to defend against like clawed code or codeex or all the other things coming in and sort of like the mental model that's really stuck with me is when we spoke to Bob Mcgru a couple of weeks ago um and how I think Jared you brought this up actually was one way you could think about it is that sort of all of these startups are kind of forward deployed engineering teams like for for the labs maybe and so like early on actually because this is all green field we don't actually know what the valuable verticals and products to build are. So in a sense you don't you step one is just figure out what that is and it wasn't actually even two years ago it wasn't actually clear it was codegen or um the IDE once you figure that out and

9:12 you find and you sort of struggle then you keep digging that's when you have to probably assume at some point you're going to get more competition because people are going to realize oh this is really valuable there's lots of money to be made here and then you have to start like defending like the treasure you found. So, I mean, all the things that we're about to cover aside from speed are sort of 1 to a billion, 1 to 10 billion, 1 to a 100red billion, one to a trillion dollar sort of problems and then uh the real stupid thing that people might do is watch this and look for this as a reason to not even get to one.

9:46 Yes. So that would be

9:47 they try to use it to like pick between two different startup ideas because they're like trying to forecast five years in the future which one will have a greater moat

9:54 which just isn't how it works. I mean literally you shouldn't do that.

9:58 Like a moat is inherently a defensive thing and you have to have something to defend otherwise like

10:03 you have nothing.

10:04 Yeah. Hey nothing to defend. Don't worry about your moat.

10:06 Yeah.

10:07 Otherwise it's just like a puddle in a field.

10:09 Yeah. Exactly.

10:11 Let's assume that someone has found something that's valuable that is worth

Process power comes from mastering edge cases

10:13 defending. Should we talk through what some of the modes they they could think about are?

10:17 Yeah. So process power again like the terminology is kind of funky but like basically it means you built something that's like you built a very complicated business with a lot of stuff that's just hard for people to replicate just because you like built all this stuff. Um and so the example that he uses in his book is like the Toyota assembly line. And I think the AI version, the AI agent version of this is just a really complicated AI agent that's been like finely honed over like multiple years to work really well under real world conditions. We've we've talked about a bunch of these on this podcast like Jake Heler with um Case Text is like the original example. A couple other ones I was thinking about from more recent companies. We have like a couple companies that sell AI agents to banks.

11:00 We have Greenlight who worked with Tom. They do KYC for banks. And we have Casa which like does loan origination for banks. So it is essentially tells banks like which loans they should give. And I think these are interesting examples because for all of these AI agents you could build a version of green light or CASA or case text like a like a demo version in like a weekend hackathon. And I think when college students are thinking about these AI agents I think what they have in their mind is like the weekend hackathon version of the product and they're like like I could build that in a week. Like how could that be defensible? And like the reason is like the the version you build in a hackathon isn't useful to anyone. It's like like like like if Casca or or Greenlight fail like the the banks will lose millions of dollars. This is like missionritical infrastructure. So it's it's more like a self-driving car.

11:47 One way to look at it is way better engineering uh is actually that's like the most profound form of process power. Like one example might be Plaid which you know the surface area of the number of uh financial institutions that they have to support is so giant it's you know probably thou on the order of thousands to tens of thousands of different different websites different crawlers and then all of the different you know can you imagine like Plaid's CI/CD structure and then you know this is pure speculation but if I were uh Zach running plaid like I you know know that I would want to be using codegen tool, the latest codegen tools to be able to uh you know basically add every new financial institution on the planet quicker than anyone else. Like that's sort of a very profound form of process power uh in the modern AI age.

12:38 I think this is probably the main form of defensibility for the existing SAS companies. Like if you look one generation before the AI agent companies like why is Stripe or Rippling or Gusto

2 Highlight · The painful drudgery of achieving high reliability

12:47 defensible? I think it's mostly this right. It's just like they've just built a lot of software and it'd be really expensive and hard to replicate all of it and like you can't just copy it from their landing page. Like there's like like the backend logic is like super deep.

12:59 There's also I feel like kind of a schle blindness aspect to this going on too where like the the hackathon version of any AI tool is like quicker than ever to get to. But actually the last like 10% of getting it to work reliably across like tens of thousands of KYC requests like per day is sort of like a particular type of painstaking drudgery work in a way that I think like lots of engineers just not excited to do. And then that is also kind of like the teams at OpenAI are going to experience this too, right? Like if you're if you're working in one of the big model labs and there's teams of people trying to invent AGI, um it's going to be hard to get jazzed about nailing the like final 5% consistency on your like KYC tool.

13:42 Yeah. And so I I think this is

13:43 especially true for like verticals like KYC that are require specialized knowledge to even know what to even to know what the edge cases are. Like if if we had to pick from the seven powers like I think speed and this these are probably like the two dominant ones that come up the most often

14:00 and those are most related to execution

14:02 is where uh the hardcore builders win. having really good product taste and building the best product really matters and I think it comes to a lot of the point maybe the the misconception is I think a lot of these products you can probably build the 80% solution with 20% of the effort but for these solutions and products to work you need the 99% accuracy one which then takes like 10 times or even sometimes 100 times the amount of effort right it's sort of that parto principle type of thing what about uh the the other power for corner for resources.

14:35 I think the classic view is they're just

Cornered resources secure unique market advantages

14:37 coveted assets or things that uh you know they're not arbitrageable. Um they must be independently valuable and then I sometimes they offer preferential access with you know rates that are way lower. So uh the classic example that you know you could look at is you know pharma companies have these patents that are very hard to get. Um they have to come up with them and then prove them and get through regulatory approval. And the sheer fact that they have a patent plus you know uh getting through FDA approval is something that can be very durable and it's uh you know so powerful that patents have a uh limited lifespan because you know you don't want people to have that forever. A more modern example I think you know on the regulatory side might be you know scale AI is doing a ton of work with the DoD um you know Palunteer as well. Uh in order to even get there, it's you know a painstaking process. You've you got to hire the right people. You've got to spend a lot of time in DC and Langley or

15:37 wherever you know you're trying to sell to. And uh you've got to literally build um uh skiffs like these like sort of you know special data uh centers where you know it's at great pain and expense. um you have to get embedded with the government but then when you do like well you've got it you know the corner resource in some sense is even the brain space in people who work in the government like you right now if you're working with AI like you've got to go through a palunteer or a scale and that's like literally written into their uh public documents around like how they're thinking about the nature of warfare and the nature of uh you know everything that they want to do having to do with AI moving forward. So, you know, the corner resource doesn't have to be a diamond mind. It could be the diamond mind in your customer's heads.

16:27 Those examples are sort of uh closer to like being way up in the sky having this like insane decacorn like worth hundreds of billions of dollars sort of situation. But what's relevant for startups that I think all of us uh sort of see every day is sort of what you were mentioning with uh this forward deployed engineer you know FTE forward deployed engineer model that uh that is what a lot of startups that are extremely successful today are literally doing like they're going out and getting a cornered resource in the form of real data and real workflows um literally sitting with a customer who normally would never get access to good software and then spotting Okay. Uh this is sort of the tailored time in motion. You know, first the uh you know a request comes in by email. Then we take this and we enrich it in this way. Uh sometimes we have to have a call center call this person like you know actually understanding what um might be a very boring process um and then translating

17:30 that into your own prompts, your own evals, eventually your own uh data sets to tune your own models. Like those are all things that are incredibly valuable. And then uh clearly there are examples you know earlier uh we're saying like character AI for instance um you know took LLMs you know obviously built some of the first LLMs then took many of them and then fine-tuned them in a way so that they could bring down the cost of uh serving those models by 10x and so you know that itself is also a form of a cornered resource. the best cornered resource to have is your own model that can like do the specific work. Yeah.

18:09 Better right?

18:09 And for a while, people thought that that was the only mode that you could have in this space. If you didn't have your own model, like you were totally hosed. Turns out that's not true. Turns out there's just one of the possible modes.

18:18 Partly that is a threat people are worried about in in the big picture. The 10,000 foot scary thing is if the labs at some point decide to treat their models as a cornered resource and they restrict access. I guess the interesting thing right now is like it may well be true that the you know platonic ideal perfect manifestation of an AI system will require a lot of both you know

3 Highlight · Context engineering is enough for early startups

18:41 maybe uh pre-training post-training RHF like just so many different things that you have to throw at it to get it to like chat GPT level but we're also so early in the revolution that um you know even if just context engineering gets you 80 you're 90% of the way there. That's plenty. That's actually all people need to do for like the first 2 years of their startup almost always.

19:08 You know, Cursor didn't start out by doing, you know, full parameter fine-tunes of GPD5, which they probably have access to now. Um, they started just by making something people want.

Deep enterprise integrations create massive switching costs

19:20 You earlier we were saying like don't use these frameworks to count yourself out prematurely. And this is a very profound version of that.

19:28 So the third power we're going to discuss is switching costs. That is uh the concept where you get a mode when your customers are kind of trapped because it becomes very expensive for them to find a other solution. Even if the other solution might be like a little bit better, it's just very painful for them to switch financially or in terms of the operations times or effort because they just have so much of it in the current solution. And examples that are given in the book are um like databases like Oracle. When you have all of your system of record and all your data in Oracle, it becomes incredibly hard to migrate.

20:12 like database migration is something that people don't do. Other example given is a Salesforce and because once you have all your customer records in Salesforce you build all these workflows the UI and it's just a lot to retrain a lot of your sales team to use like a new software you need to like migrate all the data and then at that point for the company to switch to a new CRM is probably going to take I don't know lose

4 Highlight · Converting long enterprise pilots into sticky contracts

20:35 like a whole year of productivity or something even if the new solution is a little bit better. I think how AI companies are building mode with this has to do with a version of what Gary mentioned with the forward deploy engineer. We've given examples of this with Happy Robot or Salient where they start with specific workflows that are very customized per company and they work uh with large enterprises and part of it is actually with the forward deploy engineer they may have actually very long pilot pilot periods which might last like six months to a year but if they succeed these convert into seven figure contracts and the reason why these pilots are so long is because They're very much building custom software for the specific operations in these companies. And the examples for uh Happy Robot, they got customers like DHL

21:30 where they went deep into integrating into a lot of the workflows for how all their logistic operations are done which is very accustomed to the DHL operation or the example for salient who's building AI voice agent for the financial industry. They integrate with banks and a lot of the banks have very different workflows on how they do a lot of the loan consilation, how do you do the debt recovery, how they do a lot of the fraud monitoring and risk and compliance and it's all a little bit different because all these companies have built kind of internal tools and the whole part of u being an AI company that builds these workflows They build custom workflows then that work with them. But as a result, the trade-off is you do have very long pilot cycles, but the pot of gold is worth it because you end up with this big contract and once you're in, you're kind of minted and the big enterprise is not

22:31 going to do another bake off because it's gonna it's going to be a huge waste of time for them to let's try the other whatever cool AI voice agent company. At that point, it's like we just want to get the benefits. So that's how these AI companies are winning. I think it's like at once a moat and it's also uh in it's interesting in the age of AI that uh simultaneously you could how see how AI brings down the cost of switching by a lot and that's you know sort of another lever that a startup could use like if you can write um use codegen to basically extract data out of old oified systems or your competitors then you you know there are things that might have really relied on switching costs that you could potentially bring it down to zero.

23:18 Yeah, there's actually two different flavors of switching costs, right? There's the the old school ones from the SAS era, all the system of records like Salesforce, but also ATS's like like Lever and Ashb where the switching cost was the painfulness of migrating data from one system to another. And I agree with Gary. LLMs might significantly reduce the switching cost because the

5 Highlight · Custom agent logic creates new switching costs

23:41 LMS can figure out how to like morph the data from the old schema into the new schema. You use brower like use browser automation on both sides to like solve issues where like people don't let you export the data. But then there's this new form of switching costs that I think is pretty native to the AI era like you're talking about to Tayiana which is like this these these lengthy onboarding processes that lead to like deep customizations of the logic of the agent not just the data that didn't really exist in the SAS era. Like I guess you'd like customize your like your Zenesk implementation a little bit but like not that much.

24:16 Yeah. I mean and then for AI companies on the consumer side, I mean this is all very nent, but like I think memory is already becoming a bit of a switching cost for me. Like it actually blew me away that Claude was so behind on memory and then you know uh my relationship with CHP I feel like has evolved very significantly in the last year where I'm like oh I actually just generally it

Per-seat pricing is an Achilles heel for incumbents

24:40 seems to know you know what I'm into and what I care about. So you know that switching cost I think over time will only become greater and greater and so personalization for consumer is actually a huge piece of that.

24:53 What about counterpositioning the other moat on the book?

24:56 The definition of counterpositioning is doing something that is difficult for the incumbent that you are competing with to copy because it would cannibalize their business. I think there's a couple of ways that this plays out. In every category, there is a Darwinian competition between the existing SAS incumbents building their own AI agents and the new AI native companies building AI agents on top of the existing SAS companies. So like for customer support, the existing SAS incumbents like Zenesk and uh Intercom and front are all building their own AI agents. But then we have like a new wave of companies that grew up in the last couple years that are building AI agents that interface with with those systems.

25:42 I think it's like I don't know this could be a topic of a whole like Lite Cone episode which like who will win in in in each of these fights I think is really interesting. Um

25:51 unstoppable force meets the movable object.

6 Highlight · Per-seat pricing traps SaaS incumbents

25:54 One way where this is playing out in the counterpositioning is that all almost all these companies their pricing model is they charge per seat i.e. per employee. And this is I think a very big Achilles heel that they have strategically which is that if their AI agents do a good job and actually work those companies will need fewer employees doing this work because they're like the work will be automated by AI agents and in a and in a simplistic way that will just actually reduce the more successful they are the more they will reduce the revenue. My guess is like some of them will be able to navigate this like especially if they're still founder controlled. I think like intercom for example like the I think the founder controlled versions of these companies are smart enough to recognize that this is existential and they may be able to cannibalize themselves. I think the ones that are not founder controlled I don't have a lot of hope for it's super hard to cannibalize your own revenue.

26:45 The alternative as we're seeing is so much of the startups um pricing models are around sort of like work delivered

26:51 or tasks completed. I think it's it's exactly what you said, but it's also that then switches the product towards having to actually be able to complete the work. And um something I actually repeat at the last YC batch um at the end as closing advice is that I wish the founders in a batch could just somehow go spend a month at some of the latest stage companies. Um uh cuz the top thing we hear from the founders running those companies is how hard a time they're having sort of resetting the engineering culture in their org to actually embrace AI to use the tools to want to do like context engineering and prompt engineering and and the the net result of these teams not actually being able to be AI native one of a better term uh is that they just can't deliver the products that work right and so like they both don't want to switch from Percy pricing because like that's what they're used to um uh and in a world of AI being able to do the work, there's going to be less seats to sell to, but they also just cannot deliver on products that can do the work. And so they like they they wouldn't that that

27:51 pricing model is not going to make any sense for them either.

27:54 Yeah. It's it's like the process engineering part. They're not good at the process engineering part for this new kind of engineering.

27:59 I mean something sort of uh emerging that's very interesting in a bunch of YC startups like uh Aoka for instance, they're doing customer support software kind of like Service Titan but for um HVAC. So literally like people who help you with heating and uh air conditioning and uh you know I think service titan has something like 1% wallet share 1% of the gross transaction value of like a given HVAC company um which is very small right I mean people don't spend that much money on software because these are relatively low margin service businesses but the wild thing that Aoka discovered is that you know they can come in as software but then over time they're actually getting a bigger and bigger chunk of the wallet share because they can get the HVAC people to pay them uh actually for the customer support piece which is not 1% of their spend but four to 10% of their spend. So what you may well find is that uh this new breed of AI startup will actually have more

29:01 growth uh and uh higher wallet share. So, you know, actually, we may well be all uh undervaluing how powerful and how big the vertical SAS uh AI companies will actually be because you're not like 1% of wallet share. You can get to 10. That's what we talked in that episode where vertical AI SAS agents will be 10 times at least 10 times bigger than SAS because it's really to your point Gary tapping into a whole different part of the spend of the companies is not the wallet of software where you're kind of at this point I suppose is a bit of a a finite budget but is really new space where with things that were not possible and it was mostly workflows from from people

29:49 and I you know I know that people are like pretty sensitive about uh workforce displacement but you know customer support for an HVAC services company is not a fun job and you can tell because all of these customer support jobs actually have like 50 80% annual attrition rates. like they're just such torturous, not fun jobs that uh the companies themselves and the call centers themselves spend almost all of their time trying to vet and bring in more people to work on these terrible jobs. And so when you have better software, what's sort of happening is that instead of like people aren't losing their jobs, these people are quitting their jobs anyway because it's terrible job. And then if anything uh what Avoka has told me is that many of the people who were in those customer support uh you sort of roles uh now they're actually having more fun jobs because instead of like managing a whole set of people who don't want to be there uh they're actually managing AI agents

30:50 and then handling the interesting weird cases. The coolest part of it is like they actually can go in and sometimes alter the prompts and sometimes you actually have an imp direct impact on uh both the experience of the customer but then also their own day-to-day and that immediately is like a 10 times more interesting job like wrangling a bunch

Second movers win by building better products

31:10 of AI agents and making uh the support process better and better over time. Like that's you know as knowledge work goes like way more interesting than follow this script and read what the computer says.

31:23 So Harj you you had a really interesting point about a second form of counterpositioning. this space has moved so quickly that in every vertical um or many verticals there's sort of early on emerged one company that's seen as the early winner in the space and often it's actually like the second movers at least within the YC context we have seen over and over again that like there's advantage to being the second mover in a space like stripe came after uh Brainree and authorized.net and then a bunch of things and was able to like actually win by just building a better product. Door Dash came after Grubhub, Postmates, various other delivery services and eventually went on to win. And so I think it's interesting to sort of just consider about if you're entering a vertical where it's already feels competitive or there are already there's already seem to be like a early winner in the space. How do you counter position against them? One thing I think is really interesting here is Legora versus Harvey. Lagora is obviously uh both in the legal AI space. Harvey was the early winner. The counterpositioning

32:24 that I see from Lorraa is Harvey came in early and maybe got early sales um but focus a lot on fine-tuning and sort of like their product differentiation when over time it's seen that that was actually probably not the right move. You wanted to actually focus in on the application layer and actually just sort of building a better product and and Lora has focused on that. That's what their branding and positioning is and it seems to be working really well for them as a second mover into the space. A company that I've worked more closely with, Giger ML, entered the customer service space and they're competing with Sierra and Deacon, like really well-known customer support companies and from having seen their sales motion, how they've been able to sign up some big customers. They're I think their counterpositioning is their product fundamentally just works better out of the box and as a result they can have a much faster sales and onboarding process. So it's like their counterpitching is if you want to sort of get your customer support working as quickly as possible um you should go through like the Gig ML onboarding process versus like the decong and I think that's actually worked quite well

33:25 for them.

33:26 Yeah, Giga ML is an interesting example of how to your point about like hybrid displacement. It's clear that an AI agent can do this job not just as well as a human but actually much better than a human. like the Door Dashers that the Giga ML agents are talking to, a lot of them don't speak very good English. They speak all kinds of languages. You can't hire a customer support person who's fluent in 200 languages. Um but

33:49 but LMS are actually out of box.

33:51 Out of the box. Um and they're infinitely patient if like there's a bad connection or so that's pretty interesting.

33:58 I think you have another example where to your point of superhuman abilities is where the AI version of the product actually works. I think Hargie you had the example of a Dualingo versus speak.

34:11 Dualingo is obviously the biggest language learning app I think um most consumers know. The emerging criticism of it I would say is that um what it's actually just sort of like a gaming app versus a language learning app that like the way the app works is orthogonal to learning a true language. And then you have speak um which is a uses LLM like uses voice to actually like help you practice and actually learn the language. Um, and that counterpositioning is working really well for them, right? And sort of speakers got explosive growth and it's not trying to compete with Dualingo on the we're we've got like lots of gamification and points and sort of like a great game mechanic. It's competing on hey, we're actually just a place you should come if you want to learn the language by speaking it. I think the counterpositioning mode is very um sort of close and overlaps with the branding mode idea. I think in the book he talks about you like brand is it's essentially a moat when you become so well known that even if you have an equivalent product um consumers will still choose

35:12 you um because the the brand effects I think the the example uses like Coca-Cola in the AI context I think it's probably harder to apply brand as a moat directly to startups it just takes time to acquire brand um but you can certainly see its effects like the thing that still stuns me is open AI chatbt has more consumers is using it per day than Google's Gemini. I think anyone who understands the models and uses them um daily would say that Gemini Pro 2.5 and Gemini Flash 2.5 are like equivalent models

35:42 and Google also had all the users like basically everyone in the world is a user of Google.

35:47 OpenAI had no users initially.

35:50 Google was already one of the biggest consumer brands on the planet. It was almost certainly the biggest consumer brand on the internet and yet somebody else came along and built the brand as the consumer AI app and Google is like playing catch-up.

36:04 If someone had try had told me in 2022 that that's how it would play out, I would have been fairly incredulous.

36:10 It's also a perfect example of counterpositioning. Again, I mean, this is Google had a uh a business model that required it to continue to support ads and an organization that uh they shipped and so you have the greatest cash cow in the history of man. So, why would you disrupt it um even at the cost of setting back uh human access to knowledge by a few years even if that's like the core stated goal of Google itself to organize the world's information? There's also the untold story of how uh the origin story of ChachiBT how it came to be which is really the original mode for startups with speed. It shipped very quickly in a matter of months with a very small team of a couple engineers. I mean, it required uh you know, Sam Alman and YC Research and Greg Brockman to go uh hire Ilia Suskgiver out of DeepMind because he was there and you know he all the people a lot of the people who went on

37:11 to help create OpenAI uh they came from DeepMind like it was already in the right place. It's just that that place didn't nurture exactly the thing that society really needed

37:22 for speed.

37:23 So there's that mode again speed number one. Do you want to talk about network economy Diana?

37:28 Yeah. On the book a network economy is

Data flywheels drive network effects in AI

37:31 described as uh where the value of the product increases as more users or customer get and use the product and everyone deres more value as a effect of more people using it. And examples that were given in the book are uh Facebook where as you use it and your friends use it is more fun for me to use Facebook because all my friends are in there. as more users come in then is the social network becomes more valuable and this was very much the era of uh the internet where people talked about uh network effects that came to be and the other example he gives is like visa the visa network where the more merchants are using Visa the more value the consumer gets because you're can swipe the Visa card in more places then that becomes the the moat because it's harder to then acquire and amass has this number and large number of uh users or merchants in order to to win. So that becomes very defensible. In the current era for AI, the shape of uh

38:33 network effects is different. It really comes into the shape of data. I think a lot of uh the data that a lot of AI companies get access to becomes the mode where the more data they get, the custom models they build become better and the better models, it becomes a better product for users. And there's lots of examples of these and um besides like the big foundation lab companies where they probably use some of the data. I don't know. I mean they probably use some of the data from the users. They probably do.

39:04 Chat GPT almost certainly like feeds a lot of that back because you have a certain reward function for right each training run, right? So all the history of every chat from chat GBD 1 2 3 4 5 now goes fed into GPD6 and then so on and so forth helps create the the next

7 Highlight · Keystroke data flywheels compound product quality

39:23 model version and there's uh even smaller versions of this for example cursor they have probably one of the best uh tap tap autocomplete because one the the free version of cursor they actually say it when you sign up that they they will use the data and they use that to drain it and the more users I think it's like all the data like I think it's like quite literally like every mouse click and every keystroke that you that you emit when you're using cursor like is fed into a model which is like kind of crazy

39:55 which then the more developers use cursor the better the product gets and then they compound a lot of the a lot of

40:01 the wins with that and the version where this applies to AI startup is when they go work with enterprises and large companies they get access to private data. I mentioned earlier salient or happy robot when the employees of the companies where they become customers as they use their product they have a lot of that private data that makes a lot of the workflows better and the way they improve that which is the second way of having modes with networks is really evals we we talked a lot about evals being the key mode for AI startups is evals is where you get a lot of the this workflow work or didn't work and then take that back and iterate and improve your context engineering. And that is a flywheel that you can only achieve when you get more and more usage of your product whether being in a consumer or a or a AI vertical SAS agent. So now the

Scale economies favor capital-intensive infrastructure

40:57 last mode in the book is uh scale economies. Jared, do you want to tell us about it?

41:03 Scale economies or economies of scale. you've invested a lot of money to build something that's really big and as a result you have economies of scale and you can offer the service cheaper than anybody else. So like the the classic example would be like UPS or FedEx or the Amazon delivery network. They built like massive like physical infrastructure and as a result they have like a lower cost per unit um compared to a smaller competitor. Um I think the way this has played out in the AI world I don't think it's actually played out that much at the application layer. It's really played at at the model layer, right? Like training a state-of-the-art LLM is very capital intensive. Only a few companies can afford to do it. Once you've done it, you can afford to like let people do inference on that model very inexpensively. This is why the DeepC announcement was so um was so earthshattering last year because it seemed like it might be a lot cheaper than people previously thought to train a Frontier LLM which would greatly diminish the power of this like

42:03 economies of scale mode that people thought the the AI labs had.

42:07 The key thing about Deepseek was they figure out and made public this new unlock for models which is uh how to do RL. They still built on top of one of the large foundation models so it's still expensive. the rail part is cheaper, but you still need the very expensive big foundation model. So that's one of the things that the media got wrong.

42:25 There's a separate question that people talk about, which is like how will the foundation model companies be defensible against each other? And like this is certainly one way, right? It's just like it's it's very hard to be a new entrant into that game now because of this economist scale. And we were we were thinking earlier about like how this had played out with startups and there's not that many examples, but I think a couple of good ones. Well, one one good one is is a company of yours, EXA. Harge, do you want to explain what what Exa does?

42:50 Yeah, Exa is essentially search for AI agents. Um, it provides an API for anyone building AI applications that wants to search the web.

42:58 And the way I I think this is playing out for Exa is in order to provide that service, they need to crawl the web. Not the whole web like Google does, but a big chunk of it. And that's very expensive to do. It requires like a large like fixed capital uh investment. But then once once you crawl a big chunk of the web, you can reuse that same crawl for for many different customers.

43:17 I think what's interesting about X the parallel to the model companies is that they they had invested in that like sort of before agents had really taken off like they were fairly early to this. I think they were working on this actually even prehat GBT launching. So they made the investment early on took a bet same way that the lab companies took a bet on like transformers and um uh and scaling laws.

43:38 Yeah. And there are two companies in just the most recent batch, Channel 3 and Orange Slice, that are both doing exod.ai like plays where they crawl a big chunk of the web, have a big like static crawl on their own servers, and then have agents that run on top of those of that crawl. So, I think we're going to see more and more of this, especially as the web agents work better.

43:57 You need to mainly focus on uh the first moat that isn't even in the book, which is speed. like you know if you're really breaking your brain about like oh well are we going to be a cornered resource or not you're just thinking about it in the wrong way like you should not start there you should start with do I have a specific person who has some sort of pain point and it's pretty painful it's not like a oh it'd be nice if I could do this it's a oh I am not going to get promoted this year maybe I will get fired like this is so painful that I don't want to go to work today Like that's sort of the type of pain that you're looking for. And if you can write software or build things that actually alleviate that pain, like existential pain, like the business is going to go out of business or oh my god, we could totally take over everything next year.

44:47 Like that's sort of the feeling that you want in your customer. Uh if you can find things like that, go go Z, you know, go find that and go zero to one on that first. With that, see you guys next time.