The State of AI: Models, Moats, and the Consumer Renaissance

Key Takeaway

Create a small “agent loop” for one recurring task today: define the trigger, give the AI access to the needed context and tools, require it to propose or execute a result, and add a human approval step for high-risk changes. Start with inbox triage, research, expense review, or a bug-report workflow. The goal is not a one-off prompt; it is a repeatable cycle that improves through memory, feedback, and clear boundaries.

Episode Overview

Anirudh Chakravarty discusses the AI market’s shift from a small number of general models toward a multi-model ecosystem, specialized applications, and AI-native consumer products. He argues that durable value will accrue to products that package intelligence for specific workflows, aggregate the best models, and build compounding user context, memory, and distribution.

Full Transcript

Transcript of The State of AI: Models, Moats, and the Consumer Renaissance from A16Z. Auto-generated from episode audio; may contain minor errors.

To help me understand all aspects of this incredible abundance, I'll invite Anirudh Chakravarty. And great. Perfectly. Perfectly. Hello, Anirudh. Anirudh, I was at a GP roadshow earlier this week and he told me that he was already using Grok Bot and had bought himself a bunch of jeans. So, Anirudh, do you want to tell me what you bought? The truth. The truth. Yes, I'll reveal an important secret, an intellectual property—I mostly wear Frame jeans. Frame is a great brand. And Grok Bot is a truly amazing product.

I would say that the defining characteristic of Grok Bot is a certain ingenuity. You know, a few nights ago I went to bed and said, "Hey, buy me a pair of jeans inspired by these." I took a photo of my current jeans. I said, "Don't spend more than $500 and do it." I woke up in the morning and he had already done his research, found a pair, same cut, different color, used my credit card, bought them, and they were on their way. So I think we're going to see this more and more often.

We already have these capabilities, and much of the potential will be unlocked through ingenuity, as well as product architecture that is understandable to most consumers . Perfectly. Perfectly. Perfectly. Yes, I told my team that I would finally set up my bot to sort through the pile of stuff I had been promising my husband to sell for the past 2 years. This is a project for this weekend. So, we asked the question earlier: which of the current AI leaders will be the clear winner in 3 years?

What is your opinion? I'm a fan of the fact that there will be many winners, and I see that... yes, I'm in good company with many of you. I mean, if you look at what's happened in the last 2 weeks, xAI has gone from a company that wasn't even a real contender in the modeling space to one of the top three. So we went from a two-horse race to a three-horse race extremely quickly. And, you know, overall throughout the year we saw Anthropic seem so dominant that they couldn't do wrong, and then OpenAI, who had a great 3 months.

You know, the new models are exceptional. The new Codex set and the ChatGPT desktop application are made with very high quality. And we see how these laboratories specialize in different areas. They are both growing incredibly fast, despite each other's successes. Um, xAI also has open scales. So I definitely belong to the camp of supporters of the idea of ​​multiple winners. Yes, it's interesting to watch sentiment on X, which isn't always a perfect weather vane of the future , but often serves as an early indicator of developer sentiment .

There seems to have been a lot of criticism of Claude lately regarding the use of tokens and such. And, you know, developers are prone to change. They believe they will go where the newest, best, and most advanced model is. And especially in the last 6–8 weeks, I think we'll see a very interesting progression in the dynamics of activity. But it is clear that Anthropic is going public later this year. And also, you know, there's a lot of interest in this. So, this brings us directly to the topic of today's discussion.

So, where and what awaits us at the next frontier of intelligence? Perfectly. Thank you, Jen. So let me set the stage for everyone. And please join in if you have any questions. So, first, let's look at the macro situation and what's happening at the market level. Then we'll move on to the application level in general and talk about why applications are the productization of an intellectual primitive. And finally, let's talk about the consumer sector. You know, with the launch of Grok bots and a few other products, it's been a very interesting few weeks in the consumer space.

OK. I hope our dear friend Leopold doesn't take offense to me making a little joke about situational awareness. Please continue. OK. Look, I think the " bubble" arguments have been rehashed, or at least fully discussed. In fact, the topic that is less talked about is the question: what if we are not optimistic enough? If you look at the underlying indicators, they point to excessive demand and extremely limited supply. For example, prices for B200, which are not the most modern GPUs, are increasing by the hour — this is very strange.

We usually see these things as being deflationary in nature . This indicates a very limited supply and virtually unlimited demand. We've been thinking and talking a lot about what reasons there are for optimism given some of these secondary indicators. The SaaS bubble, or rather, the “ seesaw” in the SaaS space, was an interesting look at market psychology. You know, back in February, when we saw a 30-40% drop in a number of SaaS companies, we said that the software market was oversold. And here we are, please, and many of those companies have grown by 40% again.

So I'm not sure what we've all accomplished together. But I will say what we said then, and it remains true today: enterprise software costs are 8-12%. This is a small fraction of the costs, so the benefit of writing your own code for payroll or CRM is not that great. At the same time, the risks here are essentially limitless. Of course, there are all sorts of consequences for compliance, for example, if you make a mistake with payroll. Therefore, most enterprise software today requires accuracy that code writing agents cannot yet provide.

However, one thing did happen —the tide receded. Many SaaS companies had a bunch of stock-based compensation (SBC) and other things that distorted their financial performance . I think it's very noticeable now, and they'll either have to speed up or disappear. So the situation for the SaaS market turned out to be not as bleak as we all thought a few months ago, but certain existential questions still remain. You know, there has been a lot of discussion about the "moats" around business. Do they exist at all? There are no more ditches.

It's very funny, because if you study the concept of moats, most famously described in the book "The Seven Powers"—one of my favorites. The vast majority of moats do not actually depend on available low-cost intelligence. When you think about network effects, economies of scale in distribution, or brand, which we in Silicon Valley often neglect, these things are working better than ever. No amount of coding agents will make Nike anything but Nike. Instagram's strength has never been in the complexity of building the app itself. Of course, it lay in the network behind it .

Therefore, I believe that most of the ditches are as effective as before. And, of course, they remain critical to creating long-term value. Although there are several ditches that are still under threat. For me, the moat of integration is the most obvious. You know, SAP is so notorious for its integration complexity that even switching from one version to another becomes a kind of existential risk. Coding agents make this much better. I think it's really an existential question for system integrators and global system integrators. So, what will be their value when historically they have been points of integration?

I do believe that this moat is somewhat under threat, but other traditional moats survive and remain as important as ever. And I think that's a very important concept. When you start thinking about what functions in an enterprise create value, it's usually product, sales, engineering, research. Conversely, which functions in the enterprise are, say, administrative— it's too gloomy, but they support other areas: legal, HR, finance, etc. We truly believe that the rational architecture that is currently taking shape is that to work with unlimited potential, like sales or product, it is always worth using advanced tokens.

The reason is that you simply don't know the value of a new product feature or a successful customer deal. This is virtually limitless potential, so it makes economic sense to pay almost any price for a model that is at least one IQ point smarter. Let's say your Fable 5, Grok or GPT-5, 6. However, when you're talking about something like finance, the best way to close the report is to do it accurately. You can't do it 10 times better than exactly. As a result, you are dealing with limited growth potential, so it makes sense to use models with open weights and reinforcement learning to achieve Pareto efficiency.

Maybe before we move on from this, because this is a great discussion, when Kimmy came out a few weeks ago, there was a lot of concern about this topic, given the relative cost that was at the center of the discussion. But our founder Jesse at Decagon published a great post about how in some ways, and for many companies like Decagon, open source is effectively the only option. It's not just about cost. The point is that they can localize, train, and fine-tune it. So, maybe let's look at this configuration in a little more detail .

Tell us about the nuances and why people shouldn't worry, despite the fact that there is a real abundance of opportunities for startups in this area right now . Yes, one of the main themes or trends we see is that different models have their own comparative advantages. So, models often have certain areas of focus that almost contradict each other. You see a set of models that have high levels of, let's say, neuroticism. For example, these are some kind of "autistic" models. The GLM 52 and GLM 53 are great examples of this: they are very literal, they only do what you tell them to do and nothing more.

And then we see models like the K3 , which are much more open. They are very confident and creative. There is a place for both types of models in an organization , although often their “mind-shapes, ” if you will, contradict each other. So that's one reason why you should actually use multiple models. Reinforcement learning is a very important point. If you have a task around which you can specialize a model using reasoning chains, you can create a cumulative advantage in your niche for your customer base, where you build intelligence that is better than any general intelligence for your task.

I don't know, Harvey had great results with that too. But the trade-off with this kind of reinforcement learning is that you lose generality. So if you have the best model tuned to solve legal problems, it may not be very effective in theoretical mathematics, and that's fine for Harvey purposes or, in the case of Decagon, for technical support. So, this property of specialization through open scales is unique and is one of the reasons why our startups choose them. This is also a big topic. We have learned so much since January.

We really should do this every month, Jan. Yes, to be honest , everything changes too quickly. So there were a lot of discussions in January-February , and it's very unique and interesting. Anthropic has released Claude, which they call a legal plugin. You know, plugins are just collections of skill files. You can think of them as a ZIP archive of skill files. Skill files are just prompts. These are just long prompts. And there was this huge panic, and suddenly Thompson Reuters and a bunch of other, you know, big law firms plummeted in value.

But these were just speculations, and there was a lot of discussion about whether the labs were going to vertically integrate up to the application level. Instead, we saw quite the opposite: yes, they are integrating vertically, but downwards , into the realm of data output and computation. And now, looking back, it seems logical because the data output workloads are very uniform. So you can scale to huge sizes in one part of the value chain. Whereas if you think about the application level, you know, there are so many specifics and unique needs around pricing, packaging, product strategy, how the market wants to buy.

Therefore, moving up to the application layer is a much more complex and costly proposition than moving down to the data output layer . And this is the point I mentioned earlier, a kind of discussion about the commodification of models . You know, if you use models every day, like I do, and I set a standard for myself to create something, small or big, with every new model, you start to understand that these are not just products, that they have a comparative advantage at the domain level.

A great example is OpenAI with their new GPT models, which are simply incredibly good at doing intelligent work. Their shell is also perfectly tuned for intellectual work. If you 've used the ChatGPT desktop app, you know what I mean. If not, please install it. He is very cool and interesting. And it's a perfect, well, when I say "shell" I mean a kind of product container, like a browser . This is the perfect grocery container for working with spreadsheets, presentations, written documents, and all that kind of work.

If you look at Claude code, which I'm sure many of you have used, it is strongly focused on software development . You know, it works in a terminal interface. Everything from small design decisions to areas of specialization like code planning and testing is focused on the software developer. And there are many compromises that both products make for the sake of this respective specialization. So, first of all, you already have this specialization at the domain level that's already happening. And secondly, as I mentioned earlier, you're dealing with something like the "big five" personality traits, if anyone has studied them.

You know, you can't be very open and very neurotic at the same time. Um, you know, sometimes when you apply intelligence to an accounting problem, you need neuroticism. When you apply it to a design problem , you need openness. So you actually need both types of intelligence in the organization, which is why you choose something like GLM 53 over Kimi K3. So, in our opinion, these are definitely not consumer goods . This is an important point. You know, there are many product categories where the aggregation of models produces a result that is greater than the sum of the parts.

And, you know, a good metaphor for this is Expedia. You know, it's much more useful to use Expedia than to go to the United site, then Delta, then Southwest. You just want to have one place where you can see the inventory of all airlines. Similarly, you know, in programming we see this with Cursor, where you want to use an advanced model for planning, but a less powerful one for execution. And you really need a single product shell or architecture that allows for multiple models. Creative tools are another great example where there are models that specialize in different modalities.

For example, there's 11 Labs, which is, of course, incredible at working with voice and music. And there's Black Forest, which does a great job with video and creative direction. And the right product is to combine all of this into one shell. And finally, research and decision-making. We constantly see models trained on datasets that often do not overlap. So you can get more information by passing the same request through many models and then using a separate model to achieve agreement. This is an area where the application layer really shines, as labs are motivated and structurally only able to provide their own models.

You, as an app aggregator, can provide the best of what is available. Okay, let's move on to the application level. The key point about the application layer is that intelligence is primitive. Just like buying a cloud—I noticed that. And what does Salesforce do? It takes the AWS cloud primitive and transforms it into CRM software that delivers cost-effective results for all customer segments. The same applies to the AI application level. You know, it 's great to have a basic intellectual tool, but you really need someone like Harvey to turn that into an economic outcome for the legal industry.

The same goes for credit unions: this is a very interesting market segment where they have their own unique approach to buying products, their design, and strategic ambitions . Most credit unions don't want to cut staff in half, they want to double it. Isn't that right ? They want to double it while maintaining a cost- effective business. So this is a very specific vision of how an intelligent tool works in their segment, and the task of the application layer is to provide exactly that. This is a somewhat complex concept, but I think it is important.

If you look at the evolution of AI usage, it has gone from creating queries against models to using models in loops. The term " agent" is often misused, but an agent is simply a model in a loop with certain tools, memory, and other components. A great example of this is programming . We've all seen it in software companies : they report a bug, reproduce it, create a fix, and then test it. If it is a low-risk fix, it is integrated and released, and the customer receives an email stating that their bug has been fixed.

If it's a high- risk change, it may be reviewed by a human. But thanks to this, every bug reported by companies is now automatically fixed through this programming loop. When you start applying this idea to other parts of the business, like pricing optimization or procurement, you see natural business cycles that can be fully automated by these models. Perhaps the most ambitious type of cycle is the business cycle , where you make changes that affect the entire business and the model says, "Hey, I think we should open a branch in Tijuana." Of course, the model cannot do this on its own, but it is capable of proposing changes at the level of the entire business, and that is extraordinary.

This is how enterprise automation using AI will happen. And for me, programming becomes the best illustration of this process again and again. The legal field is another great industry, not just a market. This is what Mark says, and he's right: if we consider intelligence as a primitive, let's now think about encoding intelligence as a primitive. All of these products work in their respective areas of the stack. You know, Quad Code does a great job of showing the "raw" hardware to the developer, and Replit is a great level of abstraction for the average small business owner who isn't familiar with code.

These are variations of pricing, product creation, packaging for the coding primitive, and intelligence, and they all work. So I think an important shift in thinking for us is to evaluate them as industries, not just as markets. Okay, and consumers. Consumers have had a really cool week, even a few. We've been saying for 3 years that this will be the consumer quarter, and I think this quarter could become it. Let's dig into this. The things that have really held back the consumer sector so far have been driven by several factors.

First, consumers don't like to pay for software. We have learned this lesson over and over again. And unfortunately, unlike the software magic of the past, AI software has marginal costs of distribution and engagement. And these marginal costs can sometimes be very high. I created an app that helped me browse my feed in X, and it cost me $250 to acquire one new user. So, as a startup founder, looking at that $250, even with zero acquisition costs ( TAC), it's very difficult to make a free mass product viable.

This is now changing thanks to open-scale models, which are significantly cheaper and more efficient. Secondly , we have never had an AI-native distribution channel. There is no app store for AI. So this real-world consumer product cycle is more like Web 2.0, where you have to build a channel with the product, and less like mobile technology, where there is a central distribution point for the entire ecosystem. And one last important point: we are now in the “DOS era” of AI, and for consumers to fully embrace this technology and its capabilities, we need our own “Windows.” So you think there's a huge amount of product and design work that needs to be done so that consumers know how to use all these magical new capabilities.

Two things work. Um, so programming agents—they're really amazing. I know they've already been discussed. I'm interested in thinking about how they work for consumers. You know, if you think about the concept of a digital entrepreneur , if you're not a programmer, historically you become a YouTube creator. Ten years ago, there was a whole moral panic about kids wanting to be YouTubers instead of astronauts, but I would interpret it this way: kids who grew up on the internet want to build businesses on the internet, and the only way to do that is to be a creator.

Now, thanks to coding agents, it is possible to create a software product that generates $ 100,000 or a million dollars in revenue per year . This is not the kind of business that venture capital funds invest in, but it is a unique opportunity for small family-owned SaaS businesses that are emerging now, and I think it's really cool for the country. Personal agents—in January we had a collective moment of excitement around Open Claw, it was an extraordinary composition of primitives, but it never made it into the mass consumer segment.

You know, it was a developer thing, more of an energy thing in the spirit of the Homebrew Computer Club. Um, we're starting to see with the advent of Grok Bot and ChatGPT how personal agents are turning into software that consumers can use. Anisha, could you actually pause before the demonstration? Because you were a founder who built a company in the bygone era of consumer apps. And when I even think about it, the question arises: "God, how do you define a consumer today?" Because the plumber who is now using Grok Bot to completely transform their business from start to finish...Is this a consumer or a corporate customer ?

Because it almost looks like a PLG-oriented movement . Yes. But it starts at the consumer level and then moves to the corporate level. And especially the latest era of consumer apps has been more focused on entertainment as a way to monetize. So maybe you'll look into it in more detail, especially in the context of what you spent time on. Our simple rule: if you can't justify acquiring a customer through sales, which usually means an average check of $15k, you should acquire them through marketing. We consider them consumers, as do most small business owners.

So I believe that a plumber is definitely a consumer in our investment sense. Entertainment is a huge field, and there will be many AI-based entertainment companies emerging. You know, I would say that Character is kind of an entertainment company . There's a big trend right now for short dramas, mostly in Asia, and it's starting to spread. Many of them are generative or created using generative technologies. So, listen, I think the entertainment industry is going to be huge. Most people want to waste time, not save it.

And the consumer is not very interested in performance. So it will definitely happen. Hmm, and this is probably worth a separate in-depth discussion. Okay, I think for people who use Town, it's just a magical experience. And you know, that's the number one piece of advice I give to everyone—friends, family, industry colleagues: "Please, just use the products." Because it's so easy to develop intuition when you see how they change every day. And the investment in Town, made by our partner Alex Rumpell, is a truly extraordinary productivity product.

And you see how the cumulative improvement of a product through memory gives it advantages over time. So, on the first day of use, the product does n't know you very well. It's like a new employee who is just starting to get into the swing of things. By day 30, it can make great guesses on your behalf because it has absorbed 30 days of contacts, memory, and skills. And this is a pattern we see more and more often: cumulative value for the end consumer is manifested in customer retention levels and pricing policies.

Yes, this is a great example because people can use Town for personal purposes, and it's a free trial. They give, I think, about 40 credits to start with or something like that . And you can see by connecting it to your personal email how effective it really is . Regarding the professional sphere, I always have " zero" in my inbox. And in my personal account, there are about 20,000 letters. David George is probably internally shuddering right now because this is unacceptable. However, you know, everything is difficult in personal life.

So, if you write to me personally, I will never reply to you. However, I connected Town, and now I don't even check it. If there's something important, Town will highlight it for me, and also clean up subscriptions and anything that can be optimized, and now it's starting to improve on its own. He sends you emails or says, "Hey, this subscription costs so-and-so. Here's how you could actually save money on the credit card itself." This is kind of the key to where it all starts in productivity.

And as you mentioned, maybe people won't pay for it personally, but once it starts to take off and expand in scope , you'll think, "Okay, I'll pay this money, just because it helps me manage my life and I can put everything on autopilot." Yes, that's a great point, Shannon. My mental model for this is an experienced worker with experience versus a newbie. You know, a newbie can be brilliant and even cost less than an experienced worker, but we all know the value of a specialist with experience.

They are simply capable of making great assumptions on behalf of the organization and yourself. And you know, it's a little philosophical, but I think that's what it's all about . As we talked about coding cycles and business cycles for enterprises, we believe that there are cycles that are informally defined and that actually, let's say , shape the life of the consumer. Think about family, friendship, money, health. These are all areas where you share information, decisions, actions, execution, and then the cycle continues. So, we're starting to see some of these cycles emerging around self-improvement.

Health and finance were two areas that Open AI focused on. We've seen a lot of startups working on shopping. But we believe that how this ultimately unfolds will lead to a significant improvement in the consumer's quality of life. And this really fits the mold of past product cycles, where 80% of the surplus goes to the mass market. Do you think that in each of them, all of these... sorry, maybe we'll just go back to the previous slide. There is a question here. You know, when you think about these examples of personal agents , like Town or Ethos, they all point to one context-aware assistant, but there seem to be so many options.

Do you think there will eventually be one dominant platform for such a personal aspect of your life as time management? Or will it be like an operating system, where many agents interact with each other and configure themselves on the backend? The question of competitive advantage comes to mind. I think the characteristics you expect from your financial advisor are different from those needed for a party planner. But the scope is so broad that yes, certain elements of the context will overlap. I think Graph Bots demonstrated this beautifully in a product where many bots, pointing in slightly different directions, coordinate to achieve a global optimal outcome.

I have a few questions, I want to go back to the topics you've already discussed. So, if the application level captures economic outcomes, how do you assess the competition from model companies? Will they allow value creation at the application level, and will companies at this level be able to compete with advanced labs entering this market ? I think so. Again , I think we underestimate the complexity of the product, pricing, packaging, and how the end consumer wants to buy. The way a teenager wants to use an intellectual primitive is different from the way a credit union marketing director wants to do it.

This is a very heterogeneous field. Therefore, in my opinion, it makes less sense for laboratories to rise to the level of applications than to descend to inference. So, the moment with permissions is quite interesting. I think if we lived in 2023, where there was one model to rule them all, it wouldn't matter if you had a permit, because the labs would eventually just take 100% of your gross profit. But now, because there are many options at all points on the Pareto frontier, it's harder for labs to do these things.

Perfectly. There were questions about success and development dynamics. So, are you funding something where, well, there's no revenue yet at this point , given how quickly people are making progress now, or is that extremely difficult? We try not to do that. I, of course, spent less time on such a strategy. Look, I think most of the investments in our portfolio are showing some signs of success. Certainly, in terms of product development speed, this used to be something we evaluated quite carefully. Today, it's a disqualification if a presentation doesn't show a live product at any stage, because creating something is elementary.

So almost everything we see shows signs of, you know, some kind of breakthrough. My model is pretty simple: I look at when statistically significant sales and product emerge, and extrapolating, are we comfortable with the price we have to pay to get involved and the risks we are taking? And that, I would say, is the main part of the work we do. Listen, for very talented, experienced people, we sometimes take a small option that looks like a pre-seed round, but that's not the main part of our business.

Yes. Yes. But if you think about the competitive environment, the consumer sector has been neglected for so long . Do you see this reversal now, given that it's clear that apps are the next level of value creation? The model level is already partially defined, and I say this with a big asterisk because there may still be new algorithmic breakthroughs . You know, people who appear out of nowhere, like in our portfolio. But do you feel that the competitive dynamics are shifting more towards applied solutions? 100%.

It's kind of a renaissance for consumer developers because you have this extraordinary toolkit to work with. By the way, we now have technology that can work in the emotional, interpersonal realm. You can chat with Claude, OpenAI or K3 and experience real emotions. We've had 40 years of technology that advanced intelligence and productivity, but nothing that addressed our humanity. So, this is a completely different technological plane . It is very wide. I think there are a number of products that large labs and tech giants are culturally not ready to take on.

You're thinking about launching, you know, a companion product at Google that might not align with your opinion or might contain sexual innuendo. These are the things that Google has created a thousand committees to prevent. So, startups have areas in which they are uniquely capable. And ultimately, consumers are happy to download new software and pay for it. It's like Christmas 2009 with the release of the iPhone. People want to try new apps, but unlike the 99 cent days, they are willing to pay 200 a month. So it's a kind of renaissance for consumer product developers.

And yes, I think things have changed. I'm trying to come up with a joke. But artists in San Francisco have been eagerly, very eagerly waiting for this moment for a very long time. Here's a good question from Michelle. How should we evaluate the new economics of companies creating AI applications, because there is a big debate about the unit economics of such applications, right? For example, they may have lower gross profit figures. They are under more pressure, simply because they have less access to computing power. Capital is a huge moat in this environment.

It's hard to stay competitive . So, how do you approach the economics of ensuring company profitability today? David wrote a great post about this. I believe that the issue of profitability is much more complicated now than it was before. I think that in many cases it is quite rational to sacrifice margin for the sake of broader product functionality . I think a very positive part of this cycle is the extraordinary willingness of users to pay . That’s why we often advise founders, for example in consumer products: if $20 was the historical ceiling, what would your offering be for $200 a month?

And really, what would the offer be for $2,000 a month? What is the "Birkin bag" of the software world? I think we will have such luxurious software. We already see a willingness to pay for it. So the margin issue is more complicated, but the willingness to pay and buy is higher now than ever. So it's a bit like the "fog of war," but we take all of these things into account. He mentions Birkin bags, Frame jeans. I had no idea you were such a fashionista. It's your bot helping you sit here, my friend.

For the guy I only see, I'm just here as a responsible money manager , okay? That's all I know. As for the guy I only see in zip-up sweatshirts, I'm just saying. Uh, okay. Maybe one question for you about the founders, because I don't know if you remember that conversation. This was probably about 5 years ago. Uh , when most of the founders you saw had more diverse backgrounds , partly because the software and technology was much more complex. So, you had a lot of program managers leaving Google, for example, and starting companies and so on.

What type of founders do you see building apps today? Do they lean towards more technical, more research-based directions? Are these product managers? What type of archetype do you see at least in the early stages of new applications emerging? Yes. Yes, fewer MBAs, more researchers. Uh , and both have their strengths and weaknesses . I think the business experience of the founders we see today is lower, but the technical training is much higher. And technical training is kind of the foundation for all the good things that happen.

You know, business thinking can be learned, but technical thinking usually cannot. So, we 're definitely seeing a lot more technical founders in the early stages of their careers . But the things they do are extraordinary because they don't have any preconceived notions about what's possible. And a big part of what holds back experienced founders from quite making it to the other side of that product cycle is that they're not close enough to the technology and have an idea that's rooted in a past notion of the limits of what's possible.

And I think the best thing about these young founders is that they believe that anything is possible. We were at an offsite meeting where Ben said that the biggest risk in the past was that ideas were too big, and now the biggest risk is that ideas are too small. But I think it illustrates the different archetypes of founders . Aha. Aha. And perhaps in the same vein: it used to be believed that giving a founder too much money would ruin a company, because the founder almost always has too many ideas, he is a visionary and does not have the talent to implement them all.

Uh, and we see a whole new paradigm in this. Can you elaborate on this idea a little more, because it was a very big topic at that meeting. Yes, I mean, that was indeed historical wisdom. You know why we didn't give every seed company $20, $ 50, or $100 million ? It wasn't just a matter of risk-reward ratio, the limiting factor was usually that they simply didn't have enough talented people to work on a $20 million product at once. They really had to focus on one idea, and capital was a great way to provide that focus.

Now we see that different compromises can be made on the product and model depending on the amount of capital. It makes sense that a company that raises a hundred million dollars uses it productively and purposefully and is able to offer a different value proposition than the same team with 20 million. So I think that just as we talked about the " fog of war" around margin, the question of the optimal size of a seed round and how much capital can be effectively deployed is much more complex.

It's kind of a nice problem, but I'd rather have it than the one I had five years ago: where a fintech company is indirectly subsidizing its customers through weak underwriting , and we don't see a path to profitability. Yes. Of course. Yes, Chris Dixon's model: you always want to have a supply problem, not a demand problem, right? Now we need to fix the part with the proposal, right? When demand is so huge , it will undoubtedly lead to the supply issue being resolved. Maybe I'll end with this last question from Mosfa.

So, more details about the implementation of AI in the small and medium business sector. Unlike large enterprises, there is significantly less friction during implementation as they require less change management. I agree with many aspects of this, but not all. Small and medium-sized businesses sometimes need more changes in the habits they have to work with. But the question is, how do you see the go-to-market strategy for startups targeting small and medium-sized businesses, and has it changed in the AI ​​era? I think for existing companies, it's mostly the same channels through which you've historically reached them.

In fact, one of the interesting things about marketing in the AI ​​era is that all the existing networks are so trained in the methodology of building new networks that they are very vigilant in making sure that no one does it on their platforms. Therefore, on Instagram, TikTok, X, it is very difficult to build a new distribution channel based on an existing one. So, founders need to create a product that has the original network effect, i.e., "word of mouth." So we definitely see a greater emphasis on "word of mouth." Yes, the old channels for reaching small and medium-sized businesses have not disappeared.

In fact, the most interesting segment of the market is the creation of new businesses, which, by the way, is currently at a record high. I think this is the highest figure, if you don't take into account the peak during the COVID pandemic. These are people who, under other circumstances, would never have become small or medium-sized business owners. This is not that 55-year-old plumber. This is a 25- year-old guy who could have been a YouTube blogger before, and now he's creating SaaS solutions for his neighborhood, city, school, or whatever.

Yes. Yes. Yes. Perfectly. Well , thank you very much for that, Anish. It's always nice to see you here . And now I know you're a fashionista and we'll endlessly chop this up for social media. But thank you, and if the audience has any questions, you know where to find Anish, and we'll also answer those questions that we didn't have time to cover today.

Go deeper

• All A16Z episode guides