与 Olivia Moore 一起解读 a16z《Top 100 AI Apps》报告

Olivia, welcome.

Thanks for having me.

It’s the most exciting time of the year: the Top 100 report is coming out today, I think. Is that right?

Yep.

It’s been 6 editions over 3 years. Talk to us about what’s the same, what’s changed, your excitement level, and what’s up with the report.

In many ways, so much has changed, and there’s been an incredible amount of growth since the first time we put out this list in 2023. On the other hand, from a macro level, we’re still so early. ChatGPT is by far the biggest global AI product, and still only 10% of the global population is using it on a weekly active basis. So there’s a lot more to come.

I do think this past 6 months has been maybe my favorite and most exciting time because of the shifts that we’ve seen. One of them has been that the race for the consumer is really heating up. ChatGPT, of course, but also Gemini and Claude are doubling down on their own ICP within consumer and prosumer. I think we’re starting to see how these platforms might have compounding advantages over time, and that makes it especially existential and interesting to see who is acquiring the most users.

On a related note, this was actually the first issue in which we included products that were not AI-native but are now majority AI-enabled. Things like Canva, Notion, and Freepik. Notion actually announced that they now think half of their new ARR is driven by AI-first features, which is very cool.

Lastly, I think we’ve seen a big expansion of AI outside of just the website or app prompt box. We have all of the browsers that have come out, like Dia, Comet, and Atlas. We have Claude in Excel, PowerPoint, and Chrome. And then we have desktop apps like Cursor, Wispr Flow, and Granola. There’s been a really exciting explosion in the ways that people are using AI.

So exciting. There’s a ton to cover here. Let’s start with the big foundation models. Can you talk a bit about what you think are the respective areas of specialization for Gemini, Claude, and, of course, ChatGPT? It feels like it’s been a rising-tide story more than these models trading off with each other.

Yes, I agree. Despite the drama of the past week, where we have Katy Perry taking sides on Twitter in the LLM war—which is something that I never saw coming—I think at a base level, if you look at usage, ChatGPT is a very clear winner.

On the web, they’re 2.7 times bigger than Gemini. On mobile, they’re 2.5 times bigger than Gemini. And then, despite the tech Twitter discourse, Claude is almost 30 times smaller than ChatGPT on the web and almost 80 times smaller on mobile.

We saw that Sam Altman tweet back in the Super Bowl ad wars era.

Tweet.

Yes, he was like, “We have more people using the ChatGPT free version in Texas than Claude has all users globally,” which is true.

Mhm.

That being said, I think we are seeing—I don’t think bifurcation is the right word—but maybe an expansion in the number of products people are using and what they’re using different products for. That has changed the market share a little bit.

Claude in particular has really doubled down on prosumer with things like Cowork, Claude Code, and Claude in Excel and PowerPoint. If you look at the app stores that are emerging on Claude and ChatGPT, they both have 200-plus apps, but there’s only 11% overlap. Claude is very much doubling down on premium data sources, research tools, science tools, and financial data. ChatGPT is really doubling down on consumer marketplaces, travel, nutrition, consumer finance, and things like that.

Gemini is in its own little corner as well. (Laughter.) The traction has largely been driven by creative tools there. If you look at their active users and paying users, it’s nearly perfectly correlated to releases of Veo 3, Nano Banana 1, Nano Banana Pro, and Nano Banana 2. They’re doing a little more on prosumer, adding AI to Gmail, Sheets, and Calendar, but that’s all being captured by their existing products rather than a net-new experience.

Maybe let’s dig into the app-store dynamic a little bit, because that’s so fascinating. Can you talk about the bull case for ChatGPT with what I think they call the apps directory?

Yeah. I think the approach we’re seeing with ChatGPT—and Sam said this himself on Twitter—is, “We want to be the AI for everyone.” That means they’re trying to acquire every consumer and monetize them in different ways.

I think Claude has been very clear that they’re going to monetize via subscriptions, which is great for people and companies who can pay for subscriptions, but it won’t be everyone.

Things like that.

The plugins they’re leaning into are paid, high-ACV work-data tools or data sets, like PitchBook—things that you’d use if you’re an investor, a scientist, or a mathematician.

ChatGPT, I think, is taking more of a Google-type approach. They’re building things that the average person will want to use. Maybe a smaller percentage of those people convert to subscriptions right now, but they’ll be able to monetize those people through ads and, I would guess, eventually through transactions.

If they’re building the gateway to book a trip or make all of these other long-tail consumer purchases, hypothetically they should eventually be able to take some kind of cut, at least for the traffic they’re driving. I think that’s the bull case for the ChatGPT app store. It isn’t yet showing up in the data, but it will probably become even more evident in the next year or 2.

Yeah, it’s really interesting because it touches on your point in the report about compounding advantages and how context compounds. Can you talk a little bit about that concept and then what your proxy is in terms of a metric for it? Is it session time? Is it number of sessions? Is it the amount of data you’ve provided, or is there something else?

Yeah, this is a really exciting question to me because, thus far, with these horizontal LLMs—ChatGPT, Claude, Gemini, and Perplexity—we’ve lived in a world where the context and the memory are somewhat easily exportable. Claude ran a campaign around this recently.

I think there’s going to be increasing lock-in, and I do think that probably actually benefits the broader, more horizontal tools like ChatGPT for a few reasons. First, we’ve already seen ChatGPT focus on, or start to build out, products where you interact with other people through the platform: the group chats. Imagine if there’s an even more successful version of ChatGPT group chats and all of your friends are on there. If you wanted to churn from ChatGPT, you’d also have to convince them all to go over to—

Exactly.

I would say the second one is also like an Apple-Google comparison. As these app stores emerge, developers might start to concentrate their time and effort on who they build for in the most sophisticated way and who they ship to first, depending on who has the most users or, in some cases, who’s the most willing to pay. For a lot of these consumer tools, it’ll be who has the most users. I think that also benefits ChatGPT.

The other thing that I’m probably most excited about this year is a layer that Sam Altman hinted at: authentication with ChatGPT. Essentially, you’d be able to log in with your ChatGPT account and take your memory and your tokens with you. That other product would be able to borrow those things to be even more powerful and helpful for you. If that’s the case, then you’ll want to have more of your core identity live on ChatGPT because it can lend that identity to other tools that are even better for you.

It’s so smart, and it really plays to their advantages. They have sign-ups for 900 million people, and the third-party developer ideally would not want to pay for the inference.

Yes.

So if the user can bring their inference capacity with them, there’s an advantage for the developer, ChatGPT gets the lock-in, the user gets the benefit of personalization, and it all kind of works.

Yes, I totally agree. The one question mark I still have on this, which I think could play both positively and negatively in terms of increasing lock-in for the consumer product, is what your work goes with—what your enterprise contract is.

For example, in some ways it’s good for me if my company uses ChatGPT for work because then I know how to use the product. As a normal consumer, I might have tried 1 or 2 AI products, so I’m more likely to be comfortable with and keep using something that I’ve already used.

Mhm.

On the other hand, some people might not want to mix identity and memory across their personal and work use cases.

True.

So I’m really interested—I think OpenAI hinted at this recently—in how we segment memory across different personas within yourself that are using these products.

Don’t cross the streams.

Yeah, exactly.

Well, maybe switching gears to Gemini for a moment: I think about the vibes around Google with their early AI products, Bard, which they’ll never live down.

Some tough times there.

To where we are today with products like Nano Banana. Even naming it Nano Banana is such a perfect microcosm for how far Google has come.

Yeah.

And it seems like they have a lot of intentions around multimodality.

Yeah. What's your assessment of their approach?

I've been impressed. I think they have been hesitant—maybe in some ways more hesitant, and in some ways exactly what we would expect—to bake AI into the core features because there's a risk of either cannibalizing their own product, or scaring users who have used these tools for 10, 20, 30, or 40 years. The switching cost there is a little bit high. They don't want to scare users when AI is suddenly popping up in everything, which I understand.

What they've done a really, really good job at is building these new creative products that are basically very model-driven, from the DeepMind team, which I think is generally fantastic. I think NotebookLM was actually the first look at this, and that was something truly new in consumer AI audio. Now we have the image and video models. In some ways, with a big company like this, they kind of have to get out of their own way.

In terms of being able to actually innovate, it seems like they are. You also worked at Google, so I'd be curious for your take.

It's interesting. I'm glad you brought up NotebookLM, because NotebookLM is sort of a greenfield product area within the company, so you don't have 10 VPs fighting over it. As a result, I think the progress at NotebookLM has been tremendous. They just launched a video-generation feature that helps visually demonstrate all the concepts in your workspace, which is cool.

Conversely, when you look at the existing product surfaces, like Sheets or Docs, there's so much momentum and inertia from the past, along with management overhead around them. It's harder for them to do anything other than the most obvious incremental thing.

Yes, I agree. We'll see what happens there in the next few years. I feel like they're going to put up a fight on some of those products because they don't want to lose that user base. But to your point, they're already locked in with so many enterprises that they might not have to do that much, at least in the near term, to keep up.

You know, implicit in this conversation is that we experience and talk a lot about AI in the West.

Yeah.

Talk a bit about the sort of global AI trends. There were a few surprising things I saw in there.

Yeah. We expanded our scope in terms of what we looked at for this report, which ended up being very fun and interesting. Two things that are probably obvious in terms of how they differ from the rest of the world would be Russia and China.

In China, everyone knows that a ton of AI products are censored or banned, and so almost all of the usage is on domestic products. They actually have the lowest combined ChatGPT and Gemini usage of any country. It's only 15%. They're mostly using Doubao, which is made by ByteDance, DeepSeek, Yuanbao, Kimi, and those kinds of models.

Mhm.

The somewhat surprising thing to me was that Russia is actually a very, very similar story. They also have their own kind of parallel AI ecosystem out of necessity because they have some level of sanctions and things like that that prevent them from using all the U.S.-based tools.

We've seen products like GigaChat and Yandex, which are Russia-specific and built by Russian, often state-affiliated companies, have big usage there, along with DeepSeek. Russia is the number 2 market for DeepSeek after China. If you look at the per-country adoption data, there are some blips where one country uses Claude a little bit more or another country uses Gemini a little bit more. But the 2 huge outliers are Russia and China, and those are big, big markets. I think it's worth watching what's going on there.

It's interesting, though, because both Russia and China are outliers because of restrictions around how models can be used and maybe cultural preferences. Are there any other countries that have geo-specific trends, or is this a sort of global AI behavior set?

Yeah. I would say in terms of model development—proprietary model development that allows you to deploy proprietary AI products—most of that research is coming out of the U.S. and China, maybe a little bit out of Russia.

I think we are seeing a few native ecosystems in other places. I would say Korea has a couple of its own products, like Naver and Kakao, that have built out nice LLM interfaces.

Mhm.

India is probably the other one that I watch really closely, just because there are so many people. You can have standalone big companies focus on India. The other interesting thing about India is that there are so many different languages—a huge range—that both LLM products and even voice products don't necessarily support very well. It's a worse experience if you're a primary user of one of those languages and you're trying to use something like ChatGPT.

So far, we haven't seen a huge amount of variance there yet, but I wouldn't be surprised to see more founders, even from the U.S., targeting the Indian market for AI. The other thing I wanted to mention is that, for the first time, we also did essentially a heat map of which countries are adopting AI the most and the least on a per-capita basis.

We looked across the 10 biggest LLM products, on web and mobile, to see what this might look like. Singapore is number 1.

Crazy.

Yes. Then Hong Kong, the UAE, and South Korea. The U.S. is down at number 20—not super low, but not incredibly high. Russia and China are very far down the list, below 50. There are a lot of interesting stories that live in that data.

Yeah.

The first one is that, if you think about those top 5—Singapore, South Korea, Hong Kong—the demographics of the workforce are very tech-first, white-collar, and high-skill. The U.S. has a giant chunk of jobs where AI hasn't really touched them yet, like retail, transportation, and some of these other areas.

I think the cultural norms around AI are also shockingly diverse. If you're in the U.S., you've probably internalized this ongoing angst and questioning around, "Is AI going to take my job?" or "AI is terrible for artists," along with all of these other things that make people pick up or not pick up AI.

Yeah, I was going to ask you about this.

There was actually a big survey last year from Edelman, the global media company, and the U.S. had a fairly low rate of trust in AI. It was 32%, while most of these other countries that are high on the list are at 50%, 60%, or 70%. That has also held the U.S. back, despite the fact that the biggest products come from here. Our per-capita usage is lower than a lot of these other markets that may have smaller populations but have embraced it more.

I think that's exactly right. I was reading that in China, favorable views on AI are at 80%—80% hold a favorable view. And I know the UAE and Singapore, I think, are culturally wired to be tech-optimistic, which is an advantage.

Yes. Yes, definitely. It's interesting to see some of these smaller countries and their per-capita adoption rates. In the U.S., it's around—probably a third of people are monthly active users of something like ChatGPT. In some of the European countries, or Eastern Europe, it's 50%, 40%, 45%, or 60% on smaller bases, but they've embraced it more quickly than we have here.

Yeah, really interesting. One thing that I'm watching and am interested in is that, as you look at the spectrum of AI—from the most functional, almost like a Google Search replacement, to the most cultural, creative, and personal—we should see more divergence country by country. Obviously, the culture and the movies they make in India couldn't be more different from the movies they make in China or the U.S.

Yeah.

So why wouldn't their use of creative tools be different?

Yeah. This is honestly part of the reason why we started looking at geographic segmentation in this report. For the first 2.5 or 3 years of generative AI, the vast majority of consumers were maybe interacting with 1 product. Now it's broadening quite a bit, and I think we will see more market-specific tools.

If they capture enough of that market, like some of these Russian or Chinese companies, they can actually surface on the global list if the market is big enough.

Talk a bit about the evolution of creative tools and how much you think that is a reflection of culture or is driving culture. When do we cross that threshold?

The creative-tools trend has been fascinating. Obviously, the first big generative AI product was actually Midjourney, which came out before ChatGPT.

That's right. Yeah.

In our first few editions, it was very much dominated by creative tools. I've said this before, but the creative tools benefited from the hallucinations of the early models because they produced things that were more surprising, beautiful, or original. For a while, those were the only things really working in consumer AI.

Now it's shifted a lot. The creative tools are still a huge chunk of the list, but the type of creative tool that is a standalone big business has changed. I would say the biggest change is we're seeing fewer standalone image generators.

Mhm.

A lot of this activity—if you're making a basic commodity image, like a meme, a basic marketing image, or an infographic—the core models in ChatGPT and Gemini are quite good at those things now.

Yeah.

So the products that are still surfacing on the list, like Ideogram or Midjourney, are either very aesthetically opinionated or have much more sophisticated workflows that you can't get on something like ChatGPT.

By contrast, music, voice, and video all seem to be areas where the biggest model companies have invested less. We've seen players like Suno in music and ElevenLabs in voice completely break out and rise to the top 20 or top 15 on the list, and then hold their spots there over time. There's also a compounding lock-in from the community, the large base of enterprise customers, and all of that.

Video is where I have the most questions. OpenAI has been investing in it with Sora, and of course Google with Veo, but the Chinese models are so good because they can train on any data.

[Laughter]

Seedance 2.0 is probably the best example of this. It's, in some ways, head and shoulders above what the U.S. companies have thus far been able to do. I think we'll see. I think this actually benefits platforms like Krea, where you can use all the models in one place, because my sister Justine wrote an article about this. The way video is shaping up, there's unlikely to be one model to rule them all, so you need to be able to switch between them.

That seems true of most of the model spaces: chat models, creative models, and even code models have their areas of specialization. People talk about the ergonomics of Opus versus the accuracy of Codex, and that's a trade-off. You have to choose what tool you want to use for which problem.

Yeah, absolutely.

Sora's really interesting to me because it represented both a big step forward in the model and a really ambitious experiment around social. There was data in the early days of Sora that the percentage of people who were creating was dramatically 10 times higher than we'd seen before. What's your assessment of the Sora social effort versus the model effort, and where do you see that going?

Sora is so fascinating, and I think it was a very interesting release experiment that taught us all a lot about both creative tools and, perhaps more importantly, what consumer social in the AI era might look like.

Mhm.

By the numbers, they had a massive launch. They were number 1 on the U.S. App Store for 20 consecutive days, which is very hard to do. It means you're—

Getting to be number 1 on the App Store, you probably have to get, these days, 150,000 daily downloads. So it's a high download volume.

They actually hit 1 million users faster than ChatGPT itself, so it was a huge launch. What I think a lot of people underestimate is that it still has very significant usage: 3 million dials, per Sensor Tower, which is not bad at all.

What's dropped off about Sora is the new downloads. Maybe they peaked at 6 million downloads a month in November. It's looking like 1.5 million now. I think what has really worked about Sora is that it's a very good video model—

Mhm.

—and they innovated by introducing this concept of Cameos, where a real person can grant their likeness to Sora so that they and others can generate videos of them.

Yeah.

A lot of people in the early days were making meme videos of their friends. Jake Paul went viral because he was the first big celebrity to lean into Sora, so you were seeing insane Jake Paul videos—

Paul.

Yeah. I mean, honestly, good for him.

Yes. Yes.

I think what worked less about Sora is that, because the content was exportable—

Mhm.

—people would take it to TikTok, Instagram Reels, or YouTube, where it competed against the best human-made content.

Right.

The overall feed experience was just better because you were seeing the best of both, not just the best of Sora.

Right.

I don't think we've seen a social product succeed yet that's entirely AI content. The emotional stakes just feel lower in some ways.

Right.

I would imagine we'll see more examples like these, where Sora still has clearly very significant usage and revenue as a creative tool, but not so much as a social app.

Right.

I don't know if there will be—there probably will be—a massive AI-native social network, but we haven't seen what it looks like just yet, I would say.

Right. It'll be interesting. We discuss this frequently, but every social product has a status game. On Instagram, it's maybe being the hottest, and on X, it's being the most interesting. It felt like the emerging status game on Sora was being the funniest.

Yes.

I think this is one of the reasons why it's hard for the content to cross over, because it's just two different ways of judging what is interesting and great.

I agree. What they might do, if I had to imagine where they might find more of a niche, is that they've now inked a bunch of deals with big media companies like Disney.

And so if Sora is the only place where you can make licensed fan videos of beloved characters and entertainment figures, that's very interesting.

Totally.

But we're early, I think, in how that plays out.

It's so early.

We keep saying it.

[Laughter]

Yeah.

We can't have this conversation without talking about agents: OpenClaw, Manus, Genspark, Moltbook. Give us an overview of what has happened in the last 60 days in the world of agents, and what does the report tell us?

Yeah. I think this is mostly why I say the last 6 months—or actually even the last 2 months—of this report have been the most interesting that I think we've seen.

OpenClaw, as you'll see, is not on our rankings because it blew up in February. Our data ends in January. But we did pull the data for February, and if it had been eligible, it would have been number 30 on our web list, which is a pretty big debut.

I think the really interesting thing about OpenClaw is that usage has just continued to accelerate in the technical community. Now I think it's number 1 in GitHub stars of all time. It passed React. It passed Linux.

Wow. Passed Linux?

Yes.

Holy cow.

Very impressive.

Yeah.

But in terms of overall new users, it's kind of plateaued. We looked at visits to the get-started or sign-up page, and that's been flat week over week since early February. I think that indicates that it's an amazing product if you're technical—

Mhm.

—but it has not yet fully escaped containment to nontechnical people, which is, of course, a bigger population.

Mhm.

They were acquired by OpenAI, so if I had to guess, what I would love to see OpenAI do is productize OpenClaw into something that's usable for a mainstream consumer.

I think we've also just seen the ideas behind the OpenClaw architecture inspire so many other founders. How many pitches do we take a day where the founder says, “I want to be OpenClaw for this,” or, “OpenClaw made me realize this was possible”?

Yes.

I think OpenClaw itself will continue to succeed and be a massive product, and I'm guessing we'll see more verticalized, focused versions of OpenClaw for different use cases.

Yeah, it's so interesting because it feels like one of the things that makes OpenClaw work so well is that it can operate across all models and in all directions. I wonder if it dilutes the value of OpenClaw to have it be a sole model provider, and therefore it's counter-positioned against the labs.

Yeah. They've kept it, I think, multimodal for now, at least in my usage. We'll see how it trends. I think it would be smart to keep it that way for usage, but—

Yeah. Is Manus the consumer-grade OpenClaw, or how do you distinguish the two?

Some might say that. I do actually think Manus made our web list, and of course they had a $2 billion-plus acquisition by Meta during the period covered by the list.

The growth was incredible. The ramp they reported, from $0 to $100 million or $200 million in ARR in the span of, honestly, 6–9 months, is really best in class.

My view on why Manus was so successful is that it was really the first consumer-grade agent that could operate fairly autonomously across products and platforms. You could connect email, have it browse the web, and it could make slides and spreadsheets.

I spent a lot of time in the early days trying ChatGPT Operator and Google's Project Mariner. This was a year ago, and none of them were reliable. Manus was a breakthrough in agent reliability and agent accessibility for the consumer.

I think the fact that they did the acquisition is interesting in terms of where this is going: once everyone has that agentic capability—and you might imagine they will if it's based on the core underlying models—

Then, if you're such a horizontal product, you may be better off with the distribution forces of Meta, Google, or something like that versus a standalone company. That's definitely not true if you're building something more vertical. But if you imagine that Google now has the resources to create a Manus, that's a really hard thing to keep fighting against as a startup. Obviously, the big companies have a billion different priorities, so they're not going to do everything best-in-class.

That's why I've generally been a little more cautious about the very, very horizontal consumer AI apps, because it's probably both in scope for the bigger companies, and they have the advantage of already having IT approval, enterprise contracts, and all of that.

Right, right. It is interesting that we sort of crossed this cultural threshold where Manus seemed like a non-obvious bet in terms of just the breadth of the offering, and now it seems like they're living in the future a little bit.

Yes, absolutely. They have an incredible engineering team. The quality of the product was 3 to 6 months ahead of the rest of the market, which is not easy to do when you're competing with teams of thousands of researchers.

Totally. Let's use this to segue into a conversation about other horizontal AI products, things that live beyond the web window.

Yes.

What are you seeing there?

That has been a massive theme. When I think about the products that I interact with on a daily basis in the AI world, quite a few of them are actually desktop apps—things like Granola, voice dictation tools, Claude Cowork, and those kinds of things.

It becomes a methodology problem for our report, because we can track website visits very well, and so we can track the first time that people download the desktop app. We can track mobile app usage very well. We cannot track desktop usage that closely, and I think that, increasingly, as AI products become more sophisticated, having them live in their own dedicated application—much of which will run on desktop because it can interact with your files and can be more ambient—is going to happen more and more.

Moving forward, finding ways to parallel-track rankings of these products by web and mobile usage, but also by revenue, is going to be a pretty good idea. If you think about things like Cursor, some of the consumer and prosumer AI apps that are generating the most revenue have very little usage on the web. It's almost all in a dedicated app.

Yeah, it's really interesting. It also feels like the fact that OpenAI released Atlas—

Yeah.

—and Anthropic released Claude Cowork—

Yes.

—shows you where their priorities are.

Yes, definitely. I fully agree. The AI browser debate is its own interesting thing. I feel like we're still in the early-to-mid phases of how that's going to play out.

Yeah.

I think the instinct behind an AI-native browser is right, in that if you can have AI be always on, always available, and ambient where you're spending a lot of your time online, that's a good opportunity. Perplexity's Comet, I think, actually led the way there.

Great product.

The interesting thing is, if you look at the highest spike for Comet and Atlas in terms of visits to the download page, Comet is 5 times ahead of Atlas, which is wild because ChatGPT's audience is so massive.

I think what we've seen is that Comet and Atlas still have very dedicated, excited user bases, but for the average consumer, the switching costs of a browser are nontrivial because you have workflows set up and you naturally just open this one app.

Yeah.

It not only has to have feature parity; there have to be 1 or 2 features of the AI browser that are really killer and easy enough for the average person to set up and access. I don't think that we've seen that quite yet.

You know, it's really interesting, because Sam said, I think 6 months ago on a pod, somebody was asking him, “What has surprised you the most?” And he said, “It's that the world hasn't changed more.”

Yeah.

If you look at the trends around how people are using ChatGPT at scale, it's still homework and Google-like queries, with a little bit of companionship. In a sense, something like a browser gives you an opportunity to point the user in a different direction. What's your view on how the average person is using AI today?

Yeah. I think a couple of things. I feel like teenage girls are the best source of what is happening in consumer and what will be happening in consumer. If you look at all of the biggest consumer outcomes, they were the early adopters of all of these products.

There was actually a Pew Research study fairly recently on how teenagers are using AI. Now, finally, I think for the first time, over half of them are admitting to using it for their homework. The real number is probably 99.999%, but some of them didn't want to get in trouble with their parents.

Uh-huh.

Thirty-eight percent are now using it for creative tools, so editing images, editing video, and generating images and video. Then there's this emerging, slightly longer tail, but I think it will ultimately be among the biggest behaviors: 16% are using it for casual conversation—not the intense companion products, but just having someone to talk to. Twelve percent are using it for emotional support and advice.

I think all of these use cases will ultimately asymptote around probably 100%. Those are behaviors that maybe have been less well served by products so far and will be going forward, whether it's on ChatGPT or on a standalone product. The other big thing that I'm looking out for is agents.

Are teenage girls going to use agents?

Come on.

Here's the thing. I think that agents are similar to how, in 1990, an internet company was a dot-com company, right? Or a tech company—dot-com was its own designator. I think this is what's going to happen with agents. Ultimately, every tech company was a dot-com company, and I think ultimately every AI company, and then every tech company, is going to be an agented company because that's just where the models are headed.

If you can deliver outcomes, and not just inputs, to your users as a software product, that's so much more compelling. So, yes, I think 13-year-old girls will be using agents, but they will not think of them as agents.

I think it does unlock a lot of these other consumer use cases of AI, like finance, health care, travel planning, and complex shopping. Before agents, there was just so much data you had to go out and grab, do reliably, and do across systems that it wasn't really possible. Now it is. I think we're going to see an explosion of those other use cases in the next few months.

How long do you think it takes to play out? Is everybody using their own OpenClaw in 12 months? Is that 5 years away? Is that the wrong mental model? When we have this conversation, perhaps in 6 months at the next Top 100, what does the world look like?

I feel like every time I predict something, it happens much more quickly than I would have thought, which I think is what we're seeing every day: startups are growing faster than they ever have. I think the cultural change and cultural adoption will be slower than the technological change and what's actually possible.

What we'll continue to see is this early wave of often technical, sometimes not technical, AI adopters leading the charge on a behavior that then, 6 months later, everyone else is doing. One good example of this that I'm very, very excited about is voice, which we've talked about a lot.

We have talked about voice.

To me, it's the most information-dense, high-quality source of media that we have.

Mhm.

So much of what you do every day is actually downstream or upstream of what you say.

Mhm.

For the first time in the past 6 months, I think we've seen first engineers and now other people within tech companies adopt things like voice dictation. It's now almost a norm at many companies that your meetings are going to be recorded and transcribed by AI.

Yeah.

Whether that's voice dictation or a voice AI that answers questions or does tasks for you, I think that is going to spread to the mainstream consumer in the next 6 to 9 months.

Really, really interesting. Maybe to close, can you talk a little bit about memory?

Yes.

And where you see that going?

Yes. Memory, as we mentioned earlier, can be a little bit jarring right now. Claude and ChatGPT in particular are very good at this. Even Gemini—Google has launched something called Personal Intelligence, where it can now pull information it knows about you from your docs, email, and so on to serve you better with AI across all of the apps.

As I said, it can be a little bit jarring now because many people are talking to AI about everything, personal and professional. It can sometimes inadvertently cross the line of what it knows about you to try to help you better, but in the wrong context. I think there's a lot of work to do, almost on the infrastructure side, around how we sort out who someone is in every context.

Once that is settled, I think that memory will be one of the core advantages for AI products. Whether it’s their own memory or ChatGPT lending memory, any product that you start to use 2 years from now, if it doesn’t immediately feel like it knows you, it will feel broken.

The concept of onboarding to a product should not be something that exists in a couple of years, and I think that is something memory is really going to enable. I see it for myself where I talk to several AIs all day. The way that they interact with me and the kind of value that they’re able to provide has been so much higher 2 or 3 months in than it is when you start using it.

Incredible. Well, I don’t know what the future holds, but it’s going to be weird and wonderful. I’m certain of that.

Yes.

Olivia, thank you so much. It was super fun to actually have this conversation today and go through the report. Any closing comments?

No, I’m just excited for people to read it. There’s a lot of interesting data in there. Next time, I’m sure it will look wildly different 6 months from now. So we’ll be back then.

Really exciting. Well, tell us what you think, and thanks for checking us out.

Thank you.

Mm-hmm.