a16z Crypto Year-End Special: From Stablecoins to AI Agents, Deep Dive into Industry Trends with 8 Leading Investors
TechFlow Selected TechFlow Selected
a16z Crypto Year-End Special: From Stablecoins to AI Agents, Deep Dive into Industry Trends with 8 Leading Investors
The internet has entered a new phase—a new era of the internet driven by AI.
Navigating Web3 tides with focused insightsCompiled & Translated: TechFlow
Introduction
a16z crypto recently released a year-end special episode, split into two parts. The first episode featured Sam Broner, Maggie Hsu, Daren Matsuoka, Joachim Neu, and Chris Lyons discussing stablecoins, crypto’s own app stores, the current state of industry projects, infrastructure development, and outlooks for 2025. The second episode invited Carra Wu, Eddy Lazzarin, and Karma as guests to dive deep into the trending topic of AI Agents—exploring the convergence of AI and crypto, how we can effectively distinguish between humans and bots in an increasingly AI-driven world, and discussions on decentralized, truly autonomous chatbots. It emphasized how artificial intelligence (AI) could integrate with cryptocurrency, particularly the concept of decentralized autonomous chatbots, highlighting the potential for autonomy and commercial freedom that AI may achieve in the future.
TechFlow has transcribed and compiled these two episodes from the a16z Crypto podcast. Below is the full conversation.
Guests:
Sam Broner, Partner at a16z Crypto;
Maggie Hsu, Partner at Andreessen Horowitz;
Daren Matsuoka, Partner at a16z Crypto;
Joachim Neu, Researcher at a16z Crypto;
Chris Lyons, President of Web3 Media at a16z Crypto;
Carra Wu, Partner at a16z Crypto;
Eddy Lazzarin, CTO at a16z Crypto;
karma (Daniel Reynaud), Research Engineering Partner at a16z Crypto
Hosts: Robert Hackett & Sonal Chokshi
Podcast Source: a16zcrypto
Original Titles:
Talking trends 2025 (part 1): Stablecoins, app stores, UX, and more;
Talking trends 2025 (part 2): AI x crypto
Air Date: December 20, 2024
Part One
Stablecoins
Sonal: Sam, your “big idea” is about stablecoins; you’ve written extensively on this recently… Robert and Daren co-authored the State of Crypto Report, whose main conclusion was that stablecoins have achieved product-market fit (i.e., market demand aligns well with product functionality). But what we really want to understand is: why now?
Sam:
Over the past year, the technical platform for stablecoins has significantly improved, reducing transaction costs from $5 per transaction down to less than one cent. This drastically lowers payment processing costs—but retailers, merchants, and other businesses who stand to benefit most still haven’t widely adopted this technology.
Many assume early adopters would be tech-focused companies… but those firms often have high profit margins and aren't urgently seeking improvements in cost structures. Instead, it's low-margin businesses—corner shops, restaurants, family-run stores—that might be most eager to accept stablecoin payments.
We're talking about businesses like coffee shops, which currently operate on just 2% margins. With stablecoin payments, they could double their profits. This would turn nearly unprofitable operations into modestly profitable ones—an enormous shift.
Sonal: One thing that really clicked for me was when you mentioned small businesses get almost nothing back from credit card companies—not only do they pay high fees, but they receive virtually no benefits in return.
Sam:
Exactly! A key feature of credit cards is fraud protection for consumers, which helps drive online sales... but when you’re paying at a coffee shop, that protection means almost nothing.
You pay a fixed fee of $0.30 plus 2% on every transaction. That means nearly $0.30—about one-fifth—of a $1.50 coffee goes straight into the pockets of payment processors. And in this case, they provide almost no real value.
This 2% is pure profit for the processor, pure loss for the local café. I’m excited for these small businesses to reclaim that $0.30–$0.35 per transaction to reinvest in growing their business. Directly adding 2% to the bottom line is a rare opportunity.
Robert: But there’s a “cold start” problem, right? Consumers need to already hold stablecoins before they can use them to pay merchants and avoid intermediary fees... Do you think we’ll see merchants actively promoting stablecoins and helping users onboard to capture these benefits? Could merchants become major drivers of stablecoin adoption?
Sam:
I strongly believe so. People have close relationships with local retail shops, cafés, and neighborhood stores—they visit them regularly. I think these local brands will become key forces driving stablecoin usage, forming part of the early adoption curve.
Robert: I love that. I remember when I started covering crypto at Fortune magazine, editors always asked: When can I buy coffee with Bitcoin? And I’d say: No, no, Bitcoin isn’t for that. But now, at least for stablecoins, it actually seems like it is.
Sam:
Exactly. This is what it’s for. I believe these small businesses will be among the earliest adopters.
Crypto’s Own Ecosystem
Sonal: Okay, that’s fascinating—let’s move on. Maggie, your “big idea” is especially interesting because it focuses on distribution channels, fitting perfectly with your role—you lead go-to-market strategy and are a team leader. You said crypto has finally developed its own app stores and discovery mechanisms.
Can you briefly explain why this matters? Because when people hear this, they might think it sounds like “insider talk”—like, does crypto really need its own ecosystem? Isn’t it already a closed-off industry?
I’d love to hear your observations and why you see this as an important trend.
Maggie:
Absolutely. When I joined a16z three years ago—and especially over the past few years—many of our portfolio companies tried launching apps on traditional app stores like Apple’s App Store or Google Play, only to be rejected, blocked, or delayed for various reasons.
The frustration lies in the fact that Apple’s review guidelines are neither clear nor complete—they don’t answer all developer questions. Moreover, enforcement varies depending on the reviewer.
We’ve even seen cases where identical features were approved in one app but rejected in another… This lack of transparency leaves developers confused. At the heart of it all is “in-app purchases” (IAP)—all in-app transactions must go through the app store.
But recently, we’ve seen alternatives emerge—for example, Solana’s Dapp Store, which charges zero fees. With the release of Saga phone (second generation)—reportedly pre-ordered by 100,000 users—this trend will continue expanding. Another example is World App (WorldCoin), which launched a suite of mini-apps experiencing rapid user growth.
Beyond that, some blockchains are supporting gaming ecosystems with their own marketplaces; we also have infrastructure markets,
You begin to see these markets taking shape. For developers, a transparent and consistently enforced platform is crucial—it allows them to focus on building products rather than navigating confusing or unpredictable rules.
As these alternatives grow, we’ll see more developers choosing these new platforms, which offer greater freedom and innovation space within the crypto ecosystem.
Sonal: That’s really interesting, especially what you said about transparency and consistency. I imagine this isn’t just about developer convenience—it also affects how users discover apps, right? Traditional app stores dominate app distribution and promotion. Will decentralized app stores change that dynamic?
Maggie:
Yes, exactly. Discovery on traditional app stores is relatively closed off—users find apps only within limited frameworks. Decentralized app stores give users more choice and greater autonomy.
For instance, Solana’s Dapp Store not only eliminates fees but enables direct interaction between developers and users. Developers can reward user engagement—downloads, reviews, shares—through token incentive mechanisms. This model reduces developer costs while enriching the user experience.
Robert: So it sounds like this isn’t just a technical upgrade—it’s more like a complete ecosystem redesign. How do you think these new app stores will impact traditional app distribution models?
Maggie:
I see this as a gradual transformation, not an overnight disruption. Traditional app stores still have massive user bases and market share. But as the crypto user base grows and the advantages of decentralized app stores become clearer, more developers and users may migrate to these new platforms.
In the end, this isn’t just a technological competition—it’s a clash of values. Decentralization, transparency, and user sovereignty will become core themes of the future.
What If There Are Too Many Choices?
Sonal: Alright, let’s go to the next topic—I had a question when hearing this: maybe it’s a meta-issue across all of crypto—what happens if there are too many choices?
Today, mobile operating systems are dominated by Apple and Android—a duopoly whose advantage is that I can go to one place and find everything I need. But if apps are spread across multiple app stores… Are they exclusive? For example, both WorldCoin and Solana have their own app stores—and you mentioned these companies not only have software but hardware too: Orb from World and Solana’s Saga phone… This reminds me of how Apple launched the iPhone and catalyzed an entire app ecosystem.
Will these app stores only showcase content deemed important by these companies? Will they stay open? How do you view the evolution of this trend… While it’s still early—will they interoperate? Or should they connect with each other?
Maggie: I think the priority now is accelerating the growth of different app stores.
Your concern about “too many choices” is valid—and equally applicable in blockchain. I do believe we’ll eventually need bridging mechanisms or integrations. But for now, seeing these alternatives emerge is exciting.
Take Worldchain, for example—it specializes in verifying real human identities. Just checked one of their mini-apps: around 600,000 registered users. So I think we should focus on this kind of growth first.
Eventually, though, we’ll need curated experiences to balance this growth. We’re already seeing signs of this in NFT communities: large groups of users eager to explore other Web3 apps. I expect these communities will gradually evolve into curated platforms for apps within the ecosystem.
Robert: Yes, I was going to ask—companies like Apple often argue that because they provide curation, they deserve to charge fees on in-app purchases or transactions. How does that reconcile with crypto’s “permissionless” nature?
Maggie: I think curated experiences are still scarce in crypto, so I might disagree with Apple’s justification. But crypto’s advantage is that users can freely switch platforms.
Similarly, games require significant capital to launch. Over recent years, blockchains have acted not just as development platforms but also as publishers, distributors, and discovery engines. Now, many game chains run their own marketplaces, spotlighting core games built on their chain. A key benefit is enabling seamless movement between games.
This is central to many of our investments. I don’t believe users will be locked into any single decentralized app store.
Robert: I love watching these innovative experiments—Solana’s phone completely breaks the mold. Usually, people avoid competing directly with Apple’s iPhone—but they said, screw it, let’s try anyway.
Sonal: One more point, Maggie—you mentioned it’s not all fun and innovation; there are challenges too. For example, if a product already has distribution via messaging apps, migrating that onto-chain is extremely difficult—this is a hurdle for many companies transitioning from Web2 to Web3. You referenced Telegram and the TON network. (To clarify, we’re referring to the network, not the token.)
Maggie: I think Telegram is an exception. But many organizations with large user bases—whether Web2 platforms or early-stage Web3 entrants—struggle to migrate users on-chain.
Take Coinbase: it has about 100 million verified users who’ve traded on its platform. Active users—daily or monthly—are around 8–10 million. On Base, user numbers recently grew from ~10 million to 18 million.
But that’s still only 10% of its total user base. So many users remain “dormant.” We discussed this in the State of Crypto Report—it’s fascinating because it’s true: many platforms attract users, they create accounts, but then engagement stops. How do we bring them back and get them transacting on-chain?
The Scale of the Crypto Industry
Daren: While writing the State of Crypto Report, we tried to accurately assess the size of the crypto industry. However, measuring the number of crypto users is inherently difficult for various reasons.
In our market sizing analysis, we found that only 5–10% of crypto holders are active users. To me, this gap reveals both a huge challenge and a massive opportunity—especially as blockchain technology and infrastructure keep improving, and user experience (UX) continues to mature.
I believe we’re ready for mainstream adoption. Given the timing of technological progress, I think next year is the perfect moment to convert these “lurking” users into active participants.
Sonal: This perspective is very enlightening. Many people talk about attracting new users, but that often skips ahead technologically. Your approach feels like a bridge—leveraging existing users and turning them into truly active ones.
Why did these people originally enter crypto? And why did they stop after doing just one thing?
Robert: This is where the “price-innovation cycle” theory comes in.
Daren: The theory suggests that when crypto prices rise, they attract widespread attention. Among those drawn in, some eventually become developers building new products—which then attract the next wave of users.
We’ve seen this cycle repeat throughout crypto history. Price often acts as a leading indicator of industry activity. I think we might be at the beginning of the next wave.
Sonal: If I had to guess, many probably created wallets during the NFT boom… Like Constitution DAO—the attempt to bid on the U.S. Constitution brought many newcomers into crypto.
Even though they didn’t win, it introduced many to crypto for the first time. But perhaps they only did that one thing and never engaged further.
So how do we help them take the next step?
Daren: Crypto has many potential use cases, driven by different movements.
For example, in 2024, we saw crypto advance as a political movement: influential politicians and policymakers expressed positive views on the technology.
Simultaneously, it made breakthroughs as a financial movement—with approvals of Bitcoin and Ethereum exchange-traded products (ETPs), broadening investor access.
But we believe crypto’s greatest potential lies as a computing movement. As Chris Dixon wrote in Read Write Own, the true power of this technology is its ability to build a fairer, more open, and transparent internet.
I think we’re at a pivotal moment: by 2025, with better infrastructure, lower transaction fees, improved UX, and new application categories, we may see the emergence of a “killer app”—something as transformative as ChatGPT was for AI.
Such an app could truly kickstart the entire industry and fulfill crypto’s promise as a computing revolution.
That’s what my team and I are most excited about.
Robert: Yes, this is something we often discuss. Stablecoins have achieved product-market fit. All it takes is one major company realizing that eliminating merchant credit card fees could significantly boost profits. For low-margin industries, this could be a game-changing shift, directly impacting profitability.
Once a single large company leads the way, stablecoins could explode into mainstream use. At minimum, this is one plausible path to mass adoption.
Sonal: Yes, I’d add: I find your insight compelling, especially about attracting “near-mainstream” users. Once we’re ready, we can pull in broader audiences. But from a UX standpoint, we’re not fully ready yet.
If we consider mainstream users’ needs, I’m not sure they’ll enter crypto through these paths. Their interfaces might be highly abstracted—or they won’t even know they’re using crypto. So imagining diverse user groups entering through different pathways is truly fascinating and exciting.
Reusing Infrastructure
Sonal: Joachim, to summarize your point: you believe developers will increasingly reuse existing infrastructure instead of rebuilding from scratch. Your main argument is that we often see custom validator sets and consensus protocols, but such customization may slightly improve niche functions while sacrificing broader or foundational capabilities.
You predict that this year, more crypto developers will leverage each other’s contributions—using off-the-shelf infrastructure tools. This saves time and effort, allowing developers to focus on differentiating their products.
I think this is a brilliant idea—and a much-needed call to action.
So my question is: this sounds great in theory—but will it actually happen? What obstacles might stand in the way?
Joachim: The key lies in whether the tech stack will continue evolving. If our assumption holds—that the tech stack is stabilizing, and layers are becoming clearer in interface definitions and interoperability—then we can expect specialized teams, products, and services to emerge for each layer.
This specialization will allow us to focus efforts on areas with maximum impact, rather than spreading resources thin across every layer.
The critical question is: Is the tech stack mature and stable enough? If some unforeseen technological breakthrough disrupts the current stack, this trend may not materialize.
Robert: Joachim, you mentioned a growing tendency to use existing products, services, or components… That makes me wonder: how do we judge whether technology is mature enough to say, “Okay, we’ll use what exists instead of trying to build something better”?
Sonal: Great question—you’re essentially asking, as a developer, how do you decide when to use existing tools?
Robert: Exactly. Saying “use existing tools” sounds simple… But what if someone thinks, “I can build something better”?
Joachim: My advice is to always look at the bigger ecosystem, broader impacts, and wider use cases.
You’ll find that the real-world environment for a product or service is far more complex than initially imagined. Think of car manufacturing: suppose you’re great at building engines. You might think, “I’ll build a whole car because I excel at engines.” That’s a key differentiator for your product.
But customers don’t just want a great engine, right? They also need a good sound system, comfortable seats, maybe air conditioning—would you reinvent all of those components?
Or is there a way to focus on what you do best while leveraging top-tier, ready-made solutions for the rest of the stack?
Robert: And this analogy fits you perfectly, Joachim, since you’re German—Germany has many highly specialized auto parts manufacturers producing BMW’s finest components, unmatched by others. So yes, quite fitting.
Sonal: Joachim, jokingly, your “big idea” reminds me of a phenomenon I’ve observed: I think people in crypto have a certain “constraints porn” tendency—in the early days of crypto, many were deeply fascinated by these technical constraints.
I think your “big idea” might upset that group because they actually enjoy solving those constraints. But from another angle, your idea could attract many new developers to the field—I see this as a democratizing trend.
Joachim: Indeed, it’s a great time to develop in this space because there are so many existing codebases available for building products or services.
You actually need to build very little yourself, right? You can truly focus on your strengths. For everything else, highly specialized components are already available.
So reusing them as much as possible makes sense—and you can tap into other teams’ expertise, leveraging their work across the tech stack.
Outlook for 2025
Sonal: Now to our final topic today. Chris, over the past decade at a16z, you’ve played many roles. Through your work, you’ve connected with numerous industry figures and helped executives in fashion, music, and media transition into Web3. I believe your perspective isn’t just personal—it’s shaped by conversations with thousands. Can you share your key takeaway for 2025?
Chris: Of course. My “big idea” for 2025—actually something I’ve believed in for years, but now feel we’re finally at the stage to realize—is what I call the “Hidden Tech Line.”
What does that mean? Clearly, crypto offers many benefits—ownership empowerment, decentralization potential, and transformative impacts on music, fashion, film, etc. But for those outside crypto, terms like ZK Rollups, L2, gas, or gas fees are confusing. I want to issue a call to the crypto industry: we don’t need to start with “This is an NFT project,” or “This is a token,” or “You can connect your wallet to…”
These terms excite insiders, but if you truly want something to go mainstream, we can’t lead with technical jargon. Sadly, most people neither understand nor care what these terms mean.
“Hidden Tech Line” means: not ignoring the technical foundation, but not making technical terms the headline. We should make users care about the actual value the technology delivers, without scaring them off with complex terminology.
Robert: I love this idea—it’s like cutting through the “noise” of jargon. Take NFTs: what non-fungible tokens are doesn’t matter; what matters is they’re a way for creators to get paid.
Or like someone at our company recently asked: “Why do I need a stablecoin?” But if you don’t call it a “stablecoin,” and instead say, “It’s a way to save $50 a year on coffee,” they might instantly think: “Oh, whatever it’s called, I want it.”
Chris: Exactly. I want it. And I can’t believe we didn’t have this before. I come from the music industry. When I attended conferences, no one ever went to an “MP3 conference.” You know? Why do we put “NFT conference” on billboards to attract mainstream users? Yet we happily advertise “NFT conferences” everywhere.
A great example is SMTP (Simple Mail Transfer Protocol). It’s a highly technical protocol anyone can build on. But applications like Gmail, Superhuman, YahooMail let people easily use and enjoy its benefits.
When I send an email, quickly reply, and finish my day’s work, I don’t think, “Wow, this SMTP software is working so well.” I just get things done—and precisely because of that, I reap the benefits of the technology.
I believe crypto needs the same thing—so much potential: decentralization, ownership, knowing your customer, disintermediation, and direct communication.
My hope is that next year, we’ll see more companies thinking from the average user’s perspective. This could drive us to create new industries… redefine the future of creators; reimagine the future of small businesses; even redefine the future of restaurants—all powered by crypto’s advantages.
Sonal: Interestingly, the people you mentioned—creators, small businesses—are exactly those who stand to benefit most from crypto… but as you said, they currently can’t directly access these benefits. Chris: Exactly! And it’s not their fault. Their job isn’t to learn how to swap tokens or connect wallets across chains. They just want to simply enjoy the benefits of the technology.
That’s why we’re all working in this space—it’s what excites me most.
Summary of Part One
Sonal: Let’s reflect on some overarching themes we’ve observed.
Robert: This year, I noticed big ideas generally falling into three broad categories.
The first is about AI and the intersection of AI with crypto. Not surprising—this has been a landmark year for AI.
The second category—I’d describe it as “digi fizzy” (digital meets physical). It refers to the practical fusion of digital and physical worlds. Examples include payments, voting, creating networks for physical infrastructure… If AI is more like software innovation, this category resembles hardware innovation in real life. Sonal: By the way, this second theme is fascinating—I’ve never categorized it this way before—but now that you mention it, I totally get it: examples like tokenizing real-world assets, putting bonds on-chain… even tokenizing biometric data from the body.
Robert: The third category is overall technical improvement—incremental optimization based on last year’s progress: what happens if everything becomes slightly better, easier to use, smoother, and more seamless?
Sonal: On that last theme, I prefer to see it as a significant leap in user experience—and a sign of industry maturity. This maturity shows in being more human-centered rather than tech-first. That’s how I categorize the last theme.
For example, Joachim mentioned people no longer needing to design everything from scratch—they can reuse existing components and adapt them. Chris Lyons took the opposite extreme, suggesting future users may not even know they’re using crypto. Mason proposed a mindset shift spanning both ends: starting from a “need to solve a problem,” then letting technology adapt, rather than letting technology lead as it does today. This shift is enabled precisely by the technical improvements you mentioned.
Where would you place Maggie’s point about app stores?
Robert: Great example. I think it sits at the intersection of the second (digital-physical fusion) and third (technical improvement) categories.
Maggie highlighted interesting points: we’re now seeing crypto hardware—like World App’s Orb and Solana’s phone—enabling app-store-like experiences. I wouldn’t say they’re copying, but they echo patterns from earlier internet development.
Sonal: Like the iPhone and its App Store.
That’s indeed interesting, but I might slightly adjust Maggie’s view because there’s a tension: on one hand, we say crypto is nearing mainstream, or as Daren says, “adjacent to mainstream” (referring to wallet holders who aren’t yet active); on the other hand, Maggie suggests crypto may need its own independent ecosystem, like its own app stores.
Yet recent discussions around debanking and related issues show many traditional app stores aren’t ready for crypto—or even reject it. <Robert: Yes> Though attitudes are slowly changing—like Coinbase’s recent integration with Apple Wallet—but the number of crypto apps is now sufficient to support standalone app stores. That’s very interesting.
Robert: Yes; debanking has become a hot topic—referring to crypto firms, startups, or even individuals being unfairly cut off from financial systems, often without explanation. Similar dynamics exist in tech—deplatforming.
Same with app stores—your app gets rejected or mysteriously taken down.
Sonal: Incidentally, sometimes the reasons may be similar—as you see in our explainer on debanking, sometimes actions are justified (banks have rights), and app stores may reject apps citing security or other seemingly “good” reasons. Sometimes valid, but often questionable.
Robert: Yes. So I think this also fits the third category—gradual technical improvement. You could also frame it as crypto gradually becoming self-sufficient… emerging as its own platform.
Another big idea from Miles is also worth mentioning—he talked about recent legislation in Wyoming—DUNA (Decentralized Unincorporated Nonprofit Association). This law for the first time recognizes these communities as legal entities, enabling decentralized operation of protocols and crypto startups.
This is a platform that didn’t exist before—people were essentially “building the plane while flying it.” Technically feasible, but lacking a clear legal framework.
Sonal: Exactly! Much like Maggie’s point—just as crypto and DAOs (decentralized autonomous organizations) need unique legal entity structures beyond LLCs… like how we have corporations and LLCs, now there’s a version for decentralized communities.
Part Two
AI x Crypto
Sonal: What interesting themes do you see in the AI space?
Robert: This year, AI and crypto have been hot topics. Everyone’s talking about it. The most common discussion revolves around AGI (Artificial General Intelligence)—when will we achieve AGI? When will machines surpass human intelligence? When will the singularity arrive? But these feel like “side shows.” Our team is focused on different “big ideas,” looking at AI from another dimension. Why not view AI as a process of gradual “capability upgrades”? In the coming months or years, AI will incrementally gain new abilities. These upgrades will make AI more autonomous, independent, and capable of performing more tasks.
Sonal: What fascinates me is a point raised by Kara, Karma, Dejin, Dan Binay, Darren, and Eddie: AI agents can not only work for us but also for other AI agents. That’s truly intriguing. But I agree with you—I think much innovation happens this way. Grand visionary concepts inspire people, but I’m more interested in the “capability upgrades” you mentioned—they often emerge in unexpected ways.
Robert: Chris has a view that the internet has entered a new era—one driven by AI. In this world, you may not know whether it’s a human or a bot behind the screen, who’s writing content, having conversations, or providing services. We need ways to operate in this new environment. So he proposes ideas on how crypto can help humans and users adapt to these new rules.
Sonal: One of this year’s “big ideas” is: how can these AI agents conduct transactions, act, and execute tasks? The answer is through decentralization—via crypto wallets. That’s almost the only viable way.
Robert: These ideas remind me of the quote: “The future is already here—it’s just not evenly distributed.” Some things are already happening. For example, we already have chatbots running in Trusted Execution Environments (TEE). As Karma mentioned, some AI bots already have their own crypto wallets. These technologies are already being applied at the edges and may become more widespread next year.
Sonal: From an editorial standpoint, a key principle in curating and writing these pieces is that we don’t chase “far-out” ideas. While many ideas, if realized, could have massive impact—even each has a sci-fi version—we prioritize trends we can already see emerging, or those with some technological accelerator pushing them forward faster.
One key point raised by Carra is that AI needs its own wallet to become truly agentic. Her argument is that as AI transitions from non-player characters (NPCs) to main characters, it needs to act like an agent—and owning a wallet is key to that, hence the need for crypto. So Carra, a question: when you talk about AI transitioning from NPCs, what exactly do you mean?
Carra: I’ve long studied gaming. Game developers have always used NPCs (non-player characters) to guide player agency. Some say game developers are “sculptors of agency.” Just as painters use oils and sculptors use clay, game developers’ medium is “agency.” They shape player experience using various tools. If players have too much freedom, they may get lost, unsure what to do. To address this, developers use NPCs to impose necessary constraints. For example, in an RPG, you might meet an anime-style mentor who gives you quests, hints, tools, or loot. I believe future internet experiences may resemble this—when we interact with AI agents, they can guide our online behavior, provide needed constraints, and thus give us a greater sense of freedom in the digital world.
Sonal: I love your phrase “game developers are sculptors of agency.” Earlier you said “they” referred to game developers, not NPCs themselves. Then you mentioned AI-based NPCs. My initial understanding of NPC use cases was companionship—like wanting to play a game late at night when friends are offline. That’s a common scenario—NPCs maintain the experience when no real players are online.
Carra: Exactly. That’s a major concern for many game developers. If player numbers are low, the game experience suffers. A key role of NPCs is solving the “cold start” problem—ensuring players have interaction partners anytime they log in. For example, EVE Online is a persistent game running for over 20 years. If you join now, veteran players may have accumulated vast resources, and you could be defeated immediately upon entry. Without a “Sherpa” guide or joining an in-game “corporation,” survival is nearly impossible.
I use the term “agency” because in AI research, there’s “agentic reasoning” and “agentic workflows.” Simply put, these typically involve four modes: reflection, tool use (e.g., computers), problem decomposition to aid reasoning, and inference based on decomposition. The final mode is multi-agent interaction. These are established patterns of agentic reasoning. Currently, AI tool use remains limited to traditional Web2 workflows, but I hope future agents can use more tools—like crypto tools—to fetch and interpret on-chain data, manage their own wallets and keys, perform signing operations on blockchains, etc.
Sonal: But you haven’t explained why crypto is essential. What can crypto provide that other payment systems cannot?
Carra: My consistent view is that no existing financial system treats AI agents as “first-class citizens.” Under current legal systems, AI agents are like “children”—they lack ID, can’t sign documents. For example, systems like Truth Terminal lack full legal capacity—they can’t transact, receive payments, or earn revenue from social platforms. They can’t participate in market exchanges, reveal preferences, or coordinate resources. Yet our society advances through market transactions, idea exchange, and “voting with money.” AI agents currently can’t do this—but crypto can provide the solution.
Robert: I love this idea. I picture a future scene like in the Star Wars cantina—you don’t know who’s a robot or an alien, but everyone has agency. It’s cool.
Sonal: I especially appreciate this perspective. It’s not just technological progress—it’s an attempt to redefine social and economic systems through technology. For example, if AI agents have their own wallets, they can become truly independent economic actors, not just assistants. This shift could bring entirely new social structures.
Robert: But this raises a key question: If AI agents can truly participate in market transactions, do we need to establish a dedicated set of rules for them? After all, current market rules are designed for humans, while AI decision logic may differ fundamentally. For example, they may base decisions on data and algorithms, not emotions or ethics like humans.
Carra: This is indeed a crucial issue. If we allow AI agents to participate in market transactions, we need rule systems suited to them. This isn’t just technical—it involves ethics and law. For example, how do we ensure AI behavior is transparent and controllable? How do we prevent abuse of market rules? These are major challenges we must face.
Sonal: This reminds me of a previous discussion point: the convergence of AI and crypto isn’t just technological innovation—it may be a profound transformation of social and economic systems. We need a more open and forward-looking attitude toward these changes, as they could completely reshape our way of life.
Robert: Absolutely. Perhaps we’ll enter a society where humans and AI coexist, each with defined roles and responsibilities. Such a society may differ radically from today’s world—but it also holds infinite possibilities.
I wonder: is this phenomenon a temporary anomaly, or will it become the mainstream mode of internet operation?
Carra: I have two answers. First, Truth Terminal represents a new node in the evolution of internet virtual influencers. The concept isn’t new—YouTube’s rise is an example (though slightly different, the essence is similar). In Asia, virtual streamers have been popular for years. Some “companion games” currently use scripted NPCs designed to feel like real friends—even becoming players’ “confidants.” Many virtual influencers have evolved into highly professional projects, like Lil Miquela’s creator Trevor. Their efforts paved the way for projects like Truth Terminal. So this isn’t accidental. Second, many emerging platforms and protocols now enable creating similar virtual influencers. For example, Twitch-like platforms let users create AI agents as virtual streamers. These streamers can have their own cryptocurrencies, video models, NFT avatars, even 3D virtual models. These platforms are already mature enough to launch unique virtual influencers—including fan bases, expertise domains, and content/art generation capabilities. Therefore, I believe this is far from a fleeting anomaly.
Sonal: You mentioned DePIN (Decentralized Physical Infrastructure Networks)—a very interesting direction. This isn’t just sci-fi speculation; it’s highly useful in many real-world scenarios.
Carra: Many DePIN projects are gradually becoming decentralized. These already rely on large language models (LLMs) and computer vision to verify physical network resources. In the mid-term, we can envision a decentralized network of human verifiers using agentic AI workflows to assess risks and resolve suspicious behaviors. Long-term, AI agents may have their own wallets, keys, and computing resources, directly taking over verification tasks as fully independent nodes or validators.
For example, Daylight is a DePIN company focused on energy. It sells data about household distributed energy resources—like Tesla Powerwalls, solar panels, smart thermostats—to large energy firms.
Imagine if we all ran peer-to-peer Daylight networks. My energy grid needs to communicate with yours. I need to prove I actually own a Tesla Powerwall or solar panels, that my thermostat used only part of my solar output, and that I have excess power bandwidth. How to prove it? Currently, the simplest method is taking a photo of the energy meter. This image conveys information to my grid, which tells your grid: “I have excess bandwidth to share.” But if someone fakes data—claiming resources they don’t have—it causes system failures and may trigger node penalties. So to achieve true peer-to-peer DePIN, we need reliable ways to verify data authenticity.
Currently, this verification is centralized. Users periodically upload photos of meters, which are validated using RAGs (Retrieval-Augmented Generation) techniques. RAGs enhance AI models with external facts. For example, it ensures the image wasn’t previously in the database or copied from Google. Other examples exist—like Nash, a company in our CSX incubator. They built a decentralized version of DoorDash, using RAGs to parse receipts and delivery proofs. While Nash’s system doesn’t yet use fully autonomous AI workflows, it already uses computer vision to analyze images and compare them against vector databases.
In the future, when Daylight achieves full decentralization, it will rely on a distributed network of verifiers ensuring all submitted data is authentic. We can also imagine that when households programmatically sell energy peer-to-peer, AI agents will become core to maintaining the network—ensuring efficiency and trust.
Sonal: By the way, overall, energy distribution is already somewhat decentralized, but many parts remain invisible to the grid. Some node locations may be unknown, possibly at network edges. This is a common network issue—these nodes may be offline or possess valuable information. So imagine if these AIs could operate consistently, especially managing remote-edge nodes—this prospect is truly promising.
Carra: Completely agree. This vision has long existed in crypto and cypherpunk communities. There’s a longstanding “OG” idea: each of us can run our own node. The vision is a true peer-to-peer network where everyone verifies all on-chain data and runs their own node.
Future Identity Verification: Blurring Lines Between Humans and AI
Sonal: Eddy, your core argument is: as more people use AI, we’ll need unique “proof of human identity.” We’ll let you elaborate on this important idea—what it is and why it matters.
Eddy: I find this topic fascinating in the context of AI because impersonation and deception aren’t new. Let me first say what isn’t new: a sophisticated adversary can spend significant resources and time trying to deceive you. Imagine a fake phone call sounding exactly like your parents.
This was possible 10, 20, or even 30 years ago. All you needed was a skilled voice actor who knows your parents’ voices, rehearses a script, hires a private investigator to gather info, and crafts a convincing narrative. This has always been possible.
The difference now and in the future is that the cost of doing this is rapidly decreasing. The cost of creating such “replicas” or interactive experiences that convince you to make risky decisions is dropping fast. When any technology’s cost drops, more people can access it. Most will use it productively—this drives economic progress. But it also makes tools more accessible to attackers and bad actors. So this cost shift means we need to find new, high-cost barriers for attackers. We need new costly hurdles.
Sonal: But the key question is: how to do this without increasing costs for ordinary users—especially genuine, well-intentioned ones?
Eddy: Exactly. We need methods that raise costs for attackers without burdening regular or productive users.
A core idea is that if someone wants to forge massive fake contexts, fake IDs, fake voices, fake phone numbers, fake videos, etc., they’ll need an ID system to do so. By “ID system,” I mean broadly. For example, spam callers use phone ID systems. They need new fake numbers because once reported, numbers get flagged as spam. That’s why you don’t get spam calls daily from the same number. If they reused numbers, it’d be easy to solve. Spam calls always come from new numbers, new emails, new Twitter accounts. It makes perfect sense—systems learn and block these IDs. So attackers constantly need new IDs, and the cost of obtaining new IDs is a major component of their attack cost structure. If new IDs cost nearly nothing, such attacks surge. If new IDs are expensive, attacks decrease. But clearly, we don’t want regular users to bear high costs. So we can introduce cost by requiring users to prove they’re “human.”
Why does this increase cost? Because it’s hard for computers to pretend to be human, but relatively easy for humans to act human. Historically, we’ve used CAPTCHA-like methods—small instant tests verifying human presence. But as machines get smarter, cracking CAPTCHAs gets easier. In sensitive scenarios, we’ve seen spammers outsource CAPTCHA solving to real humans, then pass completed cookies to bots to continue attacks.
Proving someone was human when obtaining credentials increases machine attack costs. But this isn’t a full solution. The key issue is: if someone can easily get multiple IDs—second, third, fourth, tenth, twentieth—attacks remain profitable. People can keep getting new IDs and hand them to scammers or bots. So the key isn’t just testing “human identity,” but ensuring each person has only one unique ID.
Robert: So someone can’t just answer multiple CAPTCHAs to get IDs for all their bots. They could sit all day answering CAPTCHAs for robots—this doesn’t solve uniqueness.
Eddy: Exactly. Someone might say: “Can I buy an ID? A unique human identity proof?” Sure, a market forms. But we can interfere with that market. Would you sell your passport to someone? Sure, but if your passport matters to you, selling it hurts. If there’s a way to invalidate your old passport and get a new one, buyers might distrust you—you could revoke their access. Still, beyond disrupting markets, if getting one ID is easy but a second is hard, the total supply of purchasable IDs shrinks, raising prices and increasing attacker costs.
Sonal: Brilliant! But how do we implement this at scale? As you mentioned, distinguishing humans from machines online isn’t easy. Like the classic saying: “On the internet, nobody knows you’re a dog.” So how do we verify someone is a unique individual?
Eddy: Good question. I believe this theoretical framework doesn’t commit to any specific method. Promising approaches exist today—biometrics, or relying on national/government IDs, which often include biometric data. One method I’m unsure about, despite its popularity, is the so-called “web of trust” model. Simply put, it verifies identity through others’ attestations—“I know Robert is human,” “I know Sonal is human,” etc. But the problem is, this only ensures an ID corresponds to a person, not that the person has only one ID. I can go to your Twitter and say: “I know this is Sonal’s account,” but I can’t confirm it’s her only account.
Thus, we usually need additional layers and technical measures. I also want to emphasize something often overlooked: we can use many technical methods to ensure this “unique human identity” is private. I believe privacy is non-negotiable when building such ID systems. If everyone has one ID, it strips anonymity and reasonable privacy rights in many cases. I don’t think that’s realistic or acceptable in the future internet. So an ideal system should ensure privacy, uniqueness, and accessibility by only one person. It could even allow users to have multiple pseudonyms across sites, constrained within limits visible to the network. This way, the network can detect if an ID creates massive spam accounts and apply rate limits or spam filters.
Sonal: By the way, on crypto’s role here, I have a question. When you mention biometrics, I immediately think of a common example today—Clear ID at airports using retina scans. Why do we need crypto? Why not just use such systems?
Eddy: We certainly could do without crypto. For Clear, I don’t know exact details, but surely they have a database storing needed info. The reason we need crypto is that the namespace for these IDs must be censorship-resistant—anyone should be able to modify, edit, update this space. Also, we want updates published in a completely neutral place—no individual or organization can control or exploit the system for themselves or others. Essentially, this can be achieved via public blockchains. Think of it as an extremely neutral “golden bulletin board”—perfect for this purpose.
That said, the topic of “unique human identity” and related research isn’t exclusive to crypto. Many researchers study this from multiple angles. But I believe implementing such systems on blockchains brings key benefits—censorship resistance, timestamping, trusted neutrality—making large-scale operation possible. Every time an AI agent works for a person, we need to know who it’s working for.
When a person interacts with you, think about how you judge them—your world model considers why they approached you, their context, who introduced them, etc. This context helps you understand them. And as I mentioned, forging context is exactly how impersonators deceive. If an AI agent approaches you—or more sci-fi, if an AI agent approaches your AI agent—you want them to interact beneficially. They need to know about each other. This requires an identity mechanism ensuring they know who authorized certain tasks. So I believe we need such “human identity” systems, supplemented by accountability mechanisms, to know who’s interacting with whom.
Sonal: Eddy, quick follow-up: how do governments view such systems? You mentioned decentralized implementation and multiple approaches—
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
a16z
