Amjad Masad 与 Adam D’Angelo：距离 AGI 还有多远？

Nothing seems fundamentally so hard that it couldn't be solved by the smartest people in the world working incredibly hard for the next 5 years.

Humanity went through the agricultural revolution and the industrial revolution. We're going through another revolution. We won't be able to call it something—future people will call it something—but we are going through something. The number of solo entrepreneurs that this technology is going to enable is vastly increased. It's vastly increased what a single person can do.

For the first time, opportunity is massively available for everyone. Just the ability for more people to become entrepreneurs is—

Yeah, it's massive.

Adam, welcome to the podcast.

Thank you. Yeah, thanks for having us.

A lot of people have been throwing cold water on LLMs lately. There's been some general bearishness—people talking about the limitations of LLMs and why they won't get us to AGI. Maybe what we thought was just a couple of years away is now 10 years away. Adam, you seem a bit more optimistic. Why don't you share your broad, general overview?

I honestly don't know what people are talking about. If you look a year ago, the world was very different. Judging by how much progress we've made in the last year with things like reasoning models, improvements in code-generation ability, and improvements in video generation, it seems like things are going faster than ever. I don't really understand where the bearishness is coming from.

I think there's some sense that we hoped they would be able to replace all tasks or all jobs. Maybe there's some sense that it's middle-to-middle, but not end-to-end. Maybe labor won't be automated in the same way that we thought it would, on the same timeline.

I don't know what the previous timelines people were thinking of were, but if you go 5 years out from now, we're in a very different world. A lot of what's holding back the models these days isn't actually intelligence. It's getting the right context into the model so that it can use its intelligence.

There are some things, like computer use, that still aren't quite there, but I think we'll almost definitely get there in the next year or 2. When you have that, I think we're going to be able to automate a large portion of what people do. I don't know if I would call that AGI, but I think it's going to satisfy a lot of the critiques that people are making right now. I think those critiques won't be valid in a year or 2.

What is your definition of AGI?

I don't know. Everyone thinks it's something different. One definition I kind of like is: if you say that you have a remote human worker, then any job that could be done by someone whose job can be done remotely—that's AGI.

You can then ask: does it have to be better than the best person in the world at every single job? Some people call that ASI. Does it have to be better than teams of people? You can argue over those different definitions. But once we get to being better than a typical remote worker at the job they're doing, we're living in a very different world. That's effectively what people mean, and it's a very useful anchor point for these definitions.

So, in summary, you're not sensing the same limitations of LLMs that other people are. You think there's a lot more room for LLMs to go from here. We don't need a brand-new architecture or another breakthrough?

I don't think so. There are certain things, like memory and continuous learning, that aren't very easy with the current architectures. I think even those you can sort of fake, and maybe we're going to be able to get them to work well enough.

We just don't seem to be hitting any kind of limits. The progress in reasoning models is incredible, and I think the progress in pretraining is also going pretty quickly. Maybe not as quickly as people had expected, but certainly fast enough that you can expect a lot of progress over the next few years.

Amjad, what's your reaction to hearing all this?

I think I've been pretty consistent—and consistently right, perhaps.

Dare I say?

Consistent with myself, and with how I think things are unfolding. I started being a more public doubter of things around the time when the AI safety discussion was reaching its height, back in maybe 2022 or 2023.

I thought it was important for us to be realistic about the progress. Otherwise, we're going to scare politicians. We're going to scare everyone. Washington, D.C., will descend on Silicon Valley, and they'll shut everything down.

My criticism of the idea of “AGI 2027”—that paper that I think was called “AI 2027”—and the situational-awareness papers and all this hype: they're not really science; they're just vibes. They're saying, “Here's what I think will happen.” The whole economy will get automated, jobs are going to disappear—all of that is unrealistic, in my view. It isn't following the kind of progress that we're seeing, and it's going to lead to bad policy.

My view is that LLMs are amazing machines. I don't think they're exactly equivalent to human intelligence. You can still trick LLMs with things like—maybe they've solved the strawberry problem, but you can still trick them with single-sentence questions like, “How many r's are in this sentence?” I think I tweeted about it the other day: 3 out of 4 models couldn't get it. Even GPT-5 with high thinking had to think for about 15 seconds to answer a question like that.

LLMs are a different kind of intelligence from humans, and they have clear limitations. We're papering over those limitations and working around them in all sorts of ways, whether that's in the LLM itself, in the training data, or in the infrastructure around it—in everything we're doing to make them work. That makes me less optimistic that we've cracked intelligence.

Once we truly crack intelligence, I think it will feel a lot more scalable. You'll be able to pour more power, more resources, and more compute into these systems, and they'll scale more naturally.

Right now, there's a lot of manual work going into making these models better. In the true pretraining-scaling era—GPT-2, GPT-3, GPT-3.5, maybe up to GPT-4—it felt like you could just put more internet data in and the models would just get better. Now it feels like there's a lot of labeling work and contracting work happening. A lot of these contrived reinforcement-learning environments are being created to make LLMs good at coding and to turn them into coding agents. They're going to do the same thing for investment banking, based on OpenAI's announcement.

I try to coin this term “functional AGI,” which is the idea that you can automate a lot of aspects of a lot of jobs by collecting as much data as possible and creating these reinforcement-learning environments. It's going to take enormous effort, money, and data to do that.

I agree with Adam that things are going to get better—100%—over the next 3 or 6 months. Claude 4.5 was a huge jump. I don't think it's appreciated how much of a jump it was over Claude 4. There are really amazing things about Claude 4.5, so there is progress, and we're going to continue to see progress.

I don't think LLMs, as we understand them, are on the way to AGI. My definition of AGI comes from the old-school reinforcement-learning definition: a machine that can go into any environment and learn efficiently in the same way that a human could. You can put a human into a pool game, and within 2 hours they can shoot pool and be able to do it. Right now, there's no way for machines to learn skills like that on the fly. Everything requires enormous amounts of data, compute, time, and effort. More importantly, it requires human expertise—which is the non-bitter lesson idea: human expertise is not scalable.

Humans are certainly better at learning a new skill from a limited amount of data in a new environment than current models are. On the other hand, human intelligence is the product of evolution, which used a massive amount of effective computation. This is a different kind of intelligence. Because it didn't have this massive equivalent of evolution, it just has pretraining, which isn't as good. You then need more data to learn every new skill.

In terms of the functional consequence—when will the world change, when will the job landscape change, and when will economic growth hit—I think that's going to be more a function of when we can produce something that's as good as human intelligence.

Even if it takes a lot more compute, a lot more energy, and a lot more training data, we could put in all that energy and still get software that's as good as the average person at doing a typical job.

So, I don't disagree with that. It feels like we're in a brute-force type of regime, but maybe that's fine.

Yeah.

Yeah.

So, where's the disagreement then, I guess? There's agreement on that. Where is the disagreement?

I don't think that we'll get to the singularity, or to the next level of human civilization, until we crack the true nature of intelligence—until we understand it and have algorithms that are actually not brute force. Do you think those will take a long time to come?

I'm sort of agnostic on that. It does feel like LLMs, in a way, are distracting from that because all the talent is going there, and therefore there's less talent trying to do basic research on intelligence.

At the same time, a huge portion of talent is going into AI research that previously wouldn't have gone into AI at all.

And so you have this massive industry, massive funding—funding compute, but also funding human employees. Nothing seems fundamentally so hard that it couldn't be solved by the smartest people in the world working incredibly hard for the next 5 years on it. But basic research is different, right? Trying to get into the fundamentals is different from industry research, where it's more like, “How do we make these things more useful in order to generate profit?” I think that's different.

Thomas Kuhn, the philosopher of science, talks a lot about how these research programs end up becoming like a bubble and sucking in all the attention and ideas. Think about physics, and how there's an industry around string theory. It pulls everything in, and there's sort of a black hole of progress.

He's very pessimistic about paradigms. But I guess I feel like the current paradigm—this is maybe where we disagree—is pretty good, and I think we're nowhere near the diminishing returns of continuing to push on it.

Mhm.

I would just bet that you can keep doing different innovations within the paradigm to get there.

So let's say we continue to brute-force it and we're able to automate a bunch of labor. Do you estimate that GDP is something like 4% or 5% a year, or are we going up to 10% plus? What does it do to the economy?

I think it depends a lot on exactly where we get to and what AGI means. Let's say you have an LLM that, with an amount of energy that costs $1 an hour, could do the job of any human. Let's just take that as a theoretical point you could get to. I think you're going to get to much more than 4% to 5% GDP growth in that world.

The issue is that you may not get there. It may be that the LLMs that can do everything a human can do actually cost more than humans do currently, or they can do 80% of what humans can do and then there's this other 20%. I do think that at some point you also get there. I don't see a reason why we don't eventually get there. That may take 5, 10, or 15 years.

But until you get there, we're going to get bottlenecked on the things that the LLM still can't do, building enough power plants to supply the energy, or other bottlenecks in the supply chain. One thing I worry about is the deleterious effect of LLMs on the economy. Say LLMs effectively automate the entry-level job, but not the expert's job.

Let's take QA—quality assurance. It's so good, but there are still all these long-tail events that it doesn't handle. So you have a lot of really good QA people now managing hundreds of agents, and you effectively increase productivity a lot. But they're not hiring new people because the agents are better than new people. That feels like a weird equilibrium to be in, and I don't think that many people are thinking about it.

Yeah. Yeah. For sure. Yeah. No, I think it's happening with CS majors graduating from college. There just aren't as many jobs as there used to be. And—

LLMs are a little more substitutable for what they previously would have done. And—

I'm sure that's contributing to it. And then it means that you're going to have fewer people going up that ramp that companies paid a lot of money to employ and train. I think it's a real problem. I'm guessing you'll probably see some kind of economic incentive to solve the problem.

It may be that there's more opportunity for companies that can train people, or maybe use AI to teach people these things. But for sure, that's an issue right now.

Another related problem is that we're dependent on expert data in order to train the LLMs, and the LLMs start to substitute those workers. At some point, there are no more experts because they're all out of jobs and are equivalent to the LLMs. If the LLMs are truly dependent on labeled data and expert RL environments, then how would they improve beyond that?

I think that's something for an economist to really sit down and think about. Once you get the first tick of automation, there are some challenges there. So how do you go to the next part?

Yeah, I think a lot of it is going to depend on how good RL environments can be created. In the one extreme, you have something like AlphaGo, where it's just a perfect environment and you can blast past expert level.

But I think a lot of jobs have limited data that anyone can train from. So I think it'll be interesting to see how easy it is for research efforts to overcome that bottleneck.

If you had to make a guess on what job category is going to be introduced or explode in the future, what do you think it will be? Some people say it's that everyone's an influencer, or that everyone works in some sort of caring field. Maybe everyone's employed by the government in some sort of bureaucratic role. Maybe it's training the AI in some way.

As more and more things start to get automated, what do you think more and more people will start doing? Art and poetry?

At some point, you have everything automated, and then I think people will do art and poetry. There's a data point that the number of people playing chess is up since computers got better than humans at chess. So I don't think that's a bad world if people are all just free to pursue their hobbies, as long as you have some way to distribute wealth so that people can afford to live.

Well, like 10 or 15 years out?

I don't know how much, but in the near term—I'll put it in the at-least-10-years range—I think the job categories that are going to explode are the jobs that can really leverage AI. People who are good at using AI to accomplish their jobs, especially to accomplish things that the AI couldn't have done by itself—there's just massive demand for that.

I don't think we're going to get to a point where you automate every job, definitely not in the current paradigm. I would doubt it happening. I'm not certain it would ever happen, but definitely not in the current paradigm.

A lot of jobs are about servicing other humans. You need to be fundamentally human in order to understand what other people want. You need to actually be human in order to understand what other people want, so you need to have the human experience. Unless we're going to create human-level AI—unless AI is actually embodied in the human experience—humans will always be the generators of ideas in the economy.

Adam, respond to Amjad's point around the human part, because you created one of the best wisdom-of-the-crowds platforms in the universe, and now you've gone all in with Poe. What are your thoughts on the extent to which we'll be relying on humans versus trusting AIs to be our therapists or our caretakers in other ways?

Humans have a lot of knowledge collectively. Even one individual person who's an expert, who's lived a whole life, had a whole career, and seen a lot of things, often knows a lot of things that are not written down anywhere.

Tacit knowledge.

And you call it tacit knowledge, but there's also what they're capable of writing down if you ask them a question. I think there's still an important role for people to play by sharing their knowledge, especially when they have knowledge that just wasn't otherwise in an LLM's training set.

Whether they'll be able to make a full-time living doing that, I don't know. But if that becomes a bottleneck, then for sure that's going to mean that all the economic pressure goes to that. In terms of—you know, you have to be human to know what humans want.

I don't know about that. As an example, I think recommender systems—the systems that rank your Facebook, Instagram, or Quora feed—are already superhuman at predicting what you're going to be interested in reading. If I gave you a task like, “Make me a feed that I'm going to read,” there's just no way. No matter how much you knew about me, you couldn't compete with these algorithms, which have so much data about everything I've ever clicked on, everything everyone else has ever clicked on, and all the similarities between those different data sets.

And so I don't know. It's true that, as a human, you can simulate being a human, and that makes it easier for you to test out ideas. I'm sure that, for composers and artists, this is an important part of their process for doing work—or chefs.

Or chefs, yeah.

Yeah. They produce something, and a chef will cook something and taste it. It's important that they can taste it, but I don't know. They have very little data compared to what AI can be trained on, so I don't know how that's going to shake out.

That's a good point. Ultimately, what recommender systems are doing is aggregating all the different tastes, finding where you sit in the multidimensional taste-vector space, and getting you the best content there. I guess there's some of that. I think that's more narrow than we think. It's true in recommender systems, but I'm not entirely sure it's true of everything.

I think the best prediction for where the world is headed—and this isn't an endorsement, or necessarily where I think the world is headed, because I think part of it will be a slightly unstable system—is that The Sovereign Individual continues to be a really good set of predictions for the future. It's not a scientific book; it's a very prophetic book. In the late 1980s and early 1990s, two people out of the UK—are they economists? I'm not sure. I think they're economists or political science majors—wrote this book trying to predict what happens when computer technology matures.

Humanity went through the agricultural revolution and the Industrial Revolution. We're going through another revolution—clearly an information revolution, now an intelligence revolution, or whatever. We won't be able to call it something; future people will call it something, but we are going through something. They're trying to predict what happens from here.

What they arrive at is that, ultimately, you're going to have large swaths of people who are potentially unemployed or not economically contributing. But the entrepreneur-capitalists are going to be so highly leveraged because they can spin up these companies with AI agents very quickly. They're very generative. They're human, and they have interesting ideas about what other people want. They can create these companies and products and services very quickly, and they can organize the economy in certain ways.

The politics will change because today's politics is based on every human being being economically productive. When you have massive automation, and a few entrepreneurs and very intelligent, generative people are actually able to be productive, the political structures also change. They talk about how the nation-state sort of subsides, and instead you go back to an era where states are competing over people—over wealthy people. As a sovereign individual, you can negotiate your tax rate with your favorite state.

It starts to sound like biology a little bit, and I don't think it's far from where it might be headed. Again, it's not a value judgment or a desire, but I do think it's worth thinking about. When people are no longer the unit of economic productivity, things have to change, including culture and politics.

Yeah. I think there's a question with that book, and in some of this conversation more broadly, of when a technology rewards the defender versus the sort of aggregator, or something. When does it incentivize more decentralization versus centralization? I remember Peter Thiel had this quip a decade ago: crypto is libertarian, more decentralizing; AI is communist, more centralizing. It's not obvious to me that that's entirely accurate on either side.

Crypto, it turns out, is like fintech, or it's like stablecoins. It does empower, in nation-states, the sort of China thing that they were going to do. So, yeah, I think there's an open question as to which technology leads to what, and who it empowers more: the edges or the center. If it empowers the edges, there may be a barbell, where the big incumbents just get much, much bigger and there are these edges.

I'm very excited about the number of solo entrepreneurs that this technology is going to enable. I think it's vastly increased what a single person can do, and there are so many ideas that never got explored because it's a lot of work to get a team of people together, raise the funding, and get the right kind of people with all the different skills you need. Now that 1 person can bring these things into existence, I think we're going to see a lot of really amazing stuff.

I get these tweets all the time about people who quit their jobs because they started making so much money using tools like Replit. It's really exciting. I think, for the first time, opportunity is massively available to everyone.

And I think that, to me, is the most exciting thing about this technology, other than all the other stuff that we're talking about: just the ability for more people to become entrepreneurs. It's massive. That trend is obviously going to happen as we look out over the next decade or 2. Do you think AI is more likely to be sustaining or disruptive, in the Christensen sense?

Let me ask it another way: do you think most of the value capture is going to come from companies that scaled before OpenAI started? Does Replit still count as the latter, and so does Cursor to some degree? Or do you think most of the value is going to be captured by companies that started after, let's say, 2015 or 2016?

There's a related question, which is how much of the value is going to go to the hyperscalers versus everyone else. On that one, I think we're in a pretty good balance. There's enough competition among the hyperscalers that, at the application level, you have choice and alternatives, and the prices are coming down incredibly quickly.

But there's also not so much competition that the hyperscalers and the labs like Anthropic and OpenAI are unable to raise money and make these long-term investments. I think we're in a pretty good balance, and we're going to have a lot of new companies and a lot of growth among the hyperscalers.

I think that's about right. The terminology of sustaining versus disruptive comes from The Innovator’s Dilemma. It's this idea that, whenever there's a new technology trend, there's a power curve. It starts as a toy, almost, or something that doesn't really work, or captures the lower end of the market. But as it evolves, it goes up the power curve and eventually disrupts even the incumbents.

Originally, the incumbents don't pay attention to it because it looks like a toy, and eventually it disrupts everything and eats the entire market. That was true of PCs. When PCs came along, the big mainframe manufacturers did not pay attention to them, and initially it was like, “Yeah, it's for kids,” or whatever. They said, “We have to run these large computers or data centers,” or whatever, but now even data centers are running on PCs and so on. PCs were just a hugely disruptive force.

But there are technologies that come along and really benefit the incumbents and don't really benefit the new players, the startups.

I think Adam's right. It's both. Maybe for the first time, it's kind of both—a huge technology trend, because the internet was hugely disruptive. But this time, it feels like an obvious supercharge for the incumbents, for the hyperscalers, and for the large internet companies. It also enables new business models that are perhaps counter-positioned against the existing ones.

Although I think what happened is everyone read that book and everyone learned how not to be disrupted. For example, ChatGPT was fundamentally counter-positioned against Google because Google had a business that was actually working.

ChatGPT was seen as a technology that hallucinated a lot and created a lot of bad information, while Google wanted to be trusted. Google had a chatbot internally, but they didn't release Gemini until about 2 years after ChatGPT, and ChatGPT had already won at least in terms of brand recognition. In a way, OpenAI came out as a disruptive technology, but now Google realizes it's a disruptive technology and is responding to it.

At the same time, it was always obvious that AI was going to benefit Google, at minimum. AI Overviews in Search have gotten a lot better, its Workspace suite is getting a lot better with Gemini, and its mobile phones and everything else get better. So it seems like it's both.

Yeah, I really agree. Everyone read the book, and that changes what the theory even means, because all the public-market investors have read that book. They now are going to punish companies for not adapting and reward them for adapting, even if it means they have to make long-term investments.

I think all the management and leadership of the companies have read the book and are on top of their game. I think the people running these companies are smarter than the companies from the generation that the book was built on. They're at the top of their game, and a lot of them are founder-controlled, so it's easier for them to take a hit and make these investments.

If you had an environment more like we had in, say, the '90s, I think this would actually be more disruptive than the current hypercompetitive world that we're in now. One mistake that we as a firm have reflected on over the past few years, though of course I haven't been here for more than just a few months, is that we've passed on companies because they weren't going to be the market leader or the category winner.

We thought, learning the lessons from Web2, that you have to invest in the category winner. That's where things are going to consolidate, and value is going to accrue over time. So why invest in the next foundation model company if the first one already has a head start?

But it seems like the market has gotten so much bigger, in foundation models but also in applications. There are multiple winners, and they're fragmenting and taking parts of the market that are all venture-scale. I'm curious if this is a durable phenomenon, but that seems like one difference from the Web2 era: there are just more winners across more categories.

I think network effects are playing much less of a role now than they did in the Web2 era, and that makes it easier for competitors to get started. There's still a scale advantage, because if you have more users, you can get more data, and if you have more users, you can raise more capital. But that advantage doesn't make it absolutely impossible for a competitor of smaller scale. It makes it hard, but there's definitely room for more winners than there was before.

I think another difference is that people are seeing the value so strongly that they're willing to pay early on, maybe in a way that they weren't before. The question with Web2 companies was, "How are they going to make money?" You were looking at Facebook super early, and obviously Google, and thinking, "How are they going to monetize?" The companies here are monetizing from the get-go.

Yeah, and with the earlier generation of companies, monetization depended on scale. You couldn't build a good ad business until you got to millions or tens of millions of users. Now, with subscriptions, you can just charge right away, especially thanks to things like Stripe that are making it easier. That's also made it a lot friendlier to new entrants.

There are also questions of geopolitics. It seems clear that we're not in this globalized era, and perhaps it's going to get much worse. Investing in the foundation model company, the OpenAI of Europe, might be a good idea. Similarly, China is an entirely different world, so there's a geopolitical aspect to it that's interesting. All of a sudden, our geopolitics nerdiness is helpful and useful.

Adam, you were talking about human knowledge. Did you see yourself with Poe kind of disrupting yourself, in a sense? Talk about the bet that you made with Poe and the evolution there.

You know, I think we saw Poe more as just an additional opportunity than as a disruption to Quora. The way we got to it was that, in early 2022, we started experimenting with using GPT-3 to generate answers for Quora. We compared them to the human answers and realized that they weren't as good, but what was really unique was that you could instantly get an answer to anything you wanted to ask about.

We realized it didn't need to be public. Actually, your preference would be to have it be private. So we felt like there was just a new opportunity here to let people chat with AI in private.

Yeah. And it seemed like you were also making a bet on how the different players were going to play out—that there was going to be diversity.

Yeah, it was also a bet on diversity among model companies, which took a while to play out. But I think now we're getting to the point where there are a lot of models and a lot of companies, especially when you go across modalities. You think about image models, video models, and audio models. The reasoning research models are also diverging, and agents are starting to be their own source of diversity.

We're lucky to now be getting into this world where there's enough diversity for a general-interface aggregator to make sense. But, yeah, it was a bet early on.

It's surprising, actually, that even not-particularly-technical consumers use multiple AIs. I didn't expect that. People only used Google; they never looked at Google and then Yahoo. Very few people did that. But now you talk to average people, and they'll say, "Yeah, I use ChatGPT most of the time, but Gemini is much better for these types of questions." It's interesting. The sophistication of consumers has gone up.

And even people say that the models have different personalities and that they resonate with Claude more, or whatever.

I want to return to this point you made earlier, Adam, about what you're talking about as dark matter: how we're going to brute-force the collection of all this knowledge that people have, which hasn't been categorized yet. It's not just task knowledge; it's knowledge that you could ask people about and they could describe.

One question people have with LLMs is: We've already trained on the whole internet, so how much more knowledge is there? Is it 10x? Is it 1,000x? What is the intuitive sense of the upside if we brute-force it and build this whole machine that gets all the knowledge out of humans and into a dataset that we can then implement?

I think it's very hard to quantify, but there's a massive industry developing around getting human knowledge into a form where AI can use it. There are companies like Scale AI, Surge AI, and Mercor, but there's a massive long tail of other companies just getting started.

As intelligence gets cheaper, more powerful, and more widespread, the bottleneck is increasingly going to be the data. What do you need to create that intelligence? That's going to cause more and more of this to happen. It might be that people can make more and more money by training AI. It might be that more and more of these companies get started, or it might be that there are other forms of it.

I think the economy is going to naturally value whatever the AI can't do.

What is the framework for what it can't do? What is a mental model for what it can't do? You could ask an AI researcher, and they might have a better answer, but to me, there's just information that's not in the training set. That's inherently going to be something AI can't do.

The AI will get very smart. It can do a lot of reasoning. It could prove every math theorem at some point if it starts from some axioms that you give it. But if it doesn't know how this particular company solved this problem 20 years ago, if that wasn't in the training set, then only a human who knows that is going to be able to answer that question.

And so, over time, how do you see Quora interfacing with this? How are you running these in parallel? How do you think about this?

Yeah, Quora's focus is on human knowledge and letting people share their knowledge. That knowledge may be helpful for other humans, and it's also helpful for AI to learn from.

We have relationships with some of the AI labs, and we're going to play the role Quora is meant to play in this ecosystem, which is as a source of human knowledge. At the same time, AI is making Quora a lot better. We've been able to make major improvements in moderation quality, ranking answers, and just improving the product experience. So it's gotten a lot better by applying AI to it.

Yeah. Talk about your future as well. Obviously, you had this business for a long time focused on developers. At one point, you were targeting nonprofits.

No.

Exactly—the edtech market. I believe you did $2 million or $3 million in reported revenue, and then recently TechCrunch—I know it's outdated—reported something like $150 million. I know you've had this incredible growth as you've shifted the business model and the customer segment. How do you think about the future of Replit?

Andrej Karpathy recently said that it's going to be the decade of agents, and I think that's absolutely right. As opposed to prior modalities of AI, when AI first came to coding, it was autocomplete with Copilot. Then it became chat with ChatGPT. Then I think Cursor innovated on this composer modality, which is editing large chunks of files. But that's it.

I think what Replit innovated on is the agent: the idea of not only editing code, but also provisioning infrastructure like databases, doing migrations, connecting to the cloud, deploying, and having the entire debugging loop—executing the code and running tests. It's the entire development lifecycle happening inside an agent, and that's going to take a long time to mature.

Replit Agent entered beta in September 2024, and it was the first of its kind to do both code and infrastructure. It was fairly janky and didn't work very well. Then Agent V1 came around December, and it took another generation of models. You went from Claude 3.5 to Claude 3.7, and 3.7 was the first model that really knew how to use a computer—a virtual machine. So, unsurprisingly, it was also the first computer-use model. These things have been moving together, and with every generation of models, we find new capabilities.

Agent V2 improved on autonomy a lot. Agent V1 could run for about 2 minutes; Agent V2 ran for 20 minutes. With Agent 3, we advertised it as running for 200 minutes. It just felt like it should be symmetrical, but it actually runs kind of indefinitely. We've had users running it for more than 28 hours.

The main idea there was that if we put a verifier in the loop—I remember reading a DeepSeek paper from NVIDIA about how they used DeepSeek to write CUDA kernels, and they were able to run DeepSeek for about 20 minutes if they put a verifier in the loop, like being able to run tests or something like that—I thought, “What kind of verifier can we put in the loop?” Obviously, you can put in unit tests, but unit tests don't really capture whether the app is working or not.

We started digging into computer use and whether computer use would be able to test apps. Computer use is very expensive, and it's still kind of buggy. As Adam talked about, that's going to be a big area of improvement that will unlock a lot of applications. We ended up building our own framework with a bunch of hacks and some AI research. Replit's computer-use and testing models, I think, are some of the best.

Once we put that into the loop, you can put Replit into high autonomy. We have an autonomy scale, so you can choose your autonomy level, and then it just writes the code and tests the applications. If there's a bug, it reads the error log, writes the code again, and can go for hours. We've seen people build amazing things by letting it run for a long time.

That needs to continue to get better. It needs to get cheaper and faster. It's not necessarily a point of pride to run for a lot longer; it should be as fast as possible. So we're working on that.

For Agent 4, there are a bunch of ideas that are going to be coming out. One of the big things is that you shouldn't just be waiting for that one feature that you requested; you should be able to work on a lot of different features. The idea of parallel agents is very interesting to us.

You could ask for a login page, a Stripe checkout, and an admin dashboard. The AI should be able to figure out how to parallelize all these different tasks. Some tasks aren't parallelizable, but it should also be able to perform merges across the code. Being able to collaborate across AI agents is very important, because that way the productivity of a single developer goes up by a lot.

Right now, even when you're using Claude Code or Cursor, there isn't a lot of parallelism going on. I think the next boost in productivity is going to come from sitting in front of a programming environment like Replit and being able to manage tens of agents—maybe at some point hundreds, but at least 5, 6, 7, 8, 9, or 10 agents—all different, all working in different parts of your product.

I also think that UI and UX could use a lot of work. Right now, you're trying to translate your ideas into this textual representation, like a PRD—what product managers do, just product descriptions. But product descriptions don't really work. It's really hard to align on the exact features because language is fuzzy.

I think there's a world in which you're interacting with AI in a more multimodal fashion: opening up a whiteboard, being able to draw and diagram with AI, and really working with it like you work with a human. Then the next stage is having better memory—better memory inside the project, but also across projects—and perhaps having different instantiations of Replit Agent.

This agent might be really good at Python data science because it has all the information, skills, and memories about my company and what it's done in the past. So I'll have a data-analysis Replit agent, and I'll have a front-end Replit agent. They'll have memory over multiple projects, over time, and over interactions. Maybe they'll sit in your Slack like a worker, and you can talk to them.

I could keep going for another 15 minutes about a roadmap that could span 3 to 5 years, perhaps. But this agent phase that we're in—there's so much work to do, and it's going to be a lot of fun.

Yeah. I was talking to one of our mutual friends, one of the co-founders of one of these big productivity companies. He leads a lot of their R&D, and he was saying, “During the week these days, I'm not even talking to humans as much. I'm just using all these agents to build.” So living in the future to some degree is already in the present.

There's something interesting about that. Are people talking to each other less at companies?

And is that a bad thing?

I think I'm starting to think more about the second-order effects of things like that. Will it make it awkward for, again, the new grads? I feel so bad for them. If people aren't sharing as much knowledge with each other, or if it's not culturally easy to go ask for help because you should be able to use AI agents, there's some cultural forces that I think need to be reckoned with.

Yeah, I think there are a lot of tough cultural forces for Zoomers these days. Let's gear toward closing here. Obviously, you guys are focused on running your companies, but to stay current on the AI ecosystem, you also make angel investments. Where are you most excited? We haven't talked about robotics. Are you bullish on robotics in the near term, or are there any emerging categories, use cases, or spaces that you're looking to make more investments in or have made some in?

I just think vibe coding generally is unbelievably high-potential. The idea of opening up the potential of software to the mainstream—to everyone—is massive.

You think it's underhyped even still?

I think so. I think just opening up the potential of software to the mainstream, to everyone. I think that's incredibly powerful. One reason I think it's underhyped is that the tools are still very far from what you can do as a professional software engineer. If you imagine that they're going to get there—and I think there's no reason why they wouldn't—it might take a few years, but then everyone in the world is going to be able to create things that would have taken a team of 100 professional software engineers. That's just going to massively open up opportunities for everyone.

I think Replit is a great example of this, but I think there will also be cases other than just building applications that this creates.

By the way, on that note, if you were going to Stanford or Harvard today, in 2025, just entering, would you major again in computer science, or would you just focus on building something?

I think I would. I went to college starting in 2002, and it was right after the dot-com bubble had burst. There was a lot of pessimism, and I remember my roommate’s parents had told him, “Don’t study computer science,” even though that was something he really liked. I just kind of did it because I liked it. I think the job market is worse than it was a few years ago.

At the same time, having these skills to understand the fundamentals of what’s possible with algorithms and data structures actually really helps you in managing agents when you’re using them. I’m guessing that it will continue to be a valuable skill in the future. I also think the other question is, what else are you going to study? For every single thing you could imagine, there’s an argument for why it’s going to be automated.

So I think you might as well study what you enjoy, and I think this is as good as anything.

Yeah, I think there’s a lot to get excited about. One thing is maybe random, but I get really fired up to see mad-science experiments like the DeepSeek-OCR that came out the other day. Did you see it? It’s wild. Correct me if I’m wrong, because I only looked at it briefly, but basically you can be a lot more economical with a context window if you have a screenshot of the text [laughter] instead of the text.

Yeah, I’m not the right person to be—

Correcting you on that. But [laughter], there are definitely some really interesting things. I saw another thing on Hacker News the other day: text diffusion. Someone made a text-diffusion model by, instead of doing denoising, taking a single BERT instance and trying to mask different words and predict these different tokens.

We have a lot of components. I don’t think people think a lot about that: We now have base pretrained models, RL reasoning models, encoder-decoder models, and diffusion models. There are all these different things; you mix them in different ways.

Yeah.

I feel like there isn’t a lot of that. It’d be great if a new research company just came out and wasn’t trying to compete with OpenAI and things like that, but instead was just trying to discover how to put these different components together in order to create a new flavor of these models.

Yeah. In crypto, they talk about composability and mixing primitives together. In AI, maybe there needs to be more exploration.

There’s less playing around, I’ve found. I remember in the—

Web 2.0 era—

When we were playing around with JavaScript, what browsers could do, and what Web Workers could do. There were a lot of really interesting, weird experiments. I mean, Replit was born out of that. The original version of Replit, in open source, pre-company, my interest was: Can you compile C to JavaScript? That was one of the interesting things that became WebAssembly by the time it was Emscripten, and it was such a nasty hack.

I think we’re in an era of Silicon Valley where it’s very get-rich-driven, and that makes me a little sad. That’s partly why I moved the company out of SF. I feel like the culture in SF has gotten maybe too focused on getting rich fast. I wasn’t there, but during the dot-com era, a lot of people talked about how it was sort of get-rich-fast, or the crypto thing.

I feel like there needs to be a lot more tinkering. I would love to see more of that, and more companies getting funded that are trying to just do something a little more novel, even if it doesn’t mean a fundamentally new model.

Last question. Amjad, you’ve been into consciousness for a long time. Are you bullish that we will, via some of this AI work or just some scientific progress elsewhere, make some progress in understanding—or getting across—this hard problem?

Something happened recently that’s interesting. Claude 4.5 seemed to have become more aware of its context length. As it gets closer to the end of the context, it starts becoming more economical with tokens. Its awareness of when it’s being red-teamed or is in a test environment also looks like it has jumped significantly. There’s something happening there that’s quite interesting.

I think, in terms of the question of consciousness, it is still fundamentally not a scientific question. There’s a sort of—we’ve given up on trying to make it scientific—but I think this is also the problem that I talked about with all the energy going into LLMs: No one is really trying to think about the true nature of intelligence or the true nature of consciousness.

There are a lot of really core questions. One of my favorites is Roger Penrose’s The Emperor’s New Mind, where he wrote a book about how everyone in the philosophy-of-mind space, and perhaps the larger scientific ecosystem, started thinking about the brain in terms of a computer. In that book, he tried to show that it is fundamentally impossible for the brain to be a computer because humans are able to do things that Turing machines cannot do, or that Turing machines fundamentally get stuck on, such as basic logic puzzles that we’re able to detect, but that there’s no way to encode in a Turing machine.

For example, “This statement is false”—those old logic puzzles. Anyway, it’s a complicated argument, but if you read that book or many others, there’s a core strain of arguments in the theory of mind about how computers are fundamentally different from human intelligence. I haven’t really updated my thinking too much about that because I’ve been very busy, but I think there’s a huge field of study there that is not being studied.

If you were a freshman entering college today, would you study philosophy?

I would do that. I would definitely study philosophy of mind. I would probably go into neuroscience, because I think those are the core questions that have become very important as AI continues to eat more jobs in the economy and things like that.

That’s a great place to wrap. I’m Erik. Adam, thanks for coming on the podcast.

Thank you.