Andrew, I think you and I are the last ones, so we better be entertaining.
Counting on you for that.
All right, we'll start with heavy-hitting questions. You said in February that AGI is about 50 years away.
Still believe that, given how fast things are moving?
I think depending on how you define AGI, we might have gotten to AGI either 30 years ago, or I think quite possibly not yet for many decades. The definition I know is AI that could do any intellectual task a human can, and that means AI that could take five years and write a novel PhD thesis, or maybe faster, or AI that any of us with a few hours of practice could learn to drive a truck through a forest. So, AGI seems to me like would have to do that, too.
But because of incentives between, you know, a Microsoft and an OpenAI contract, which has changed, so this has gone away, there was a specific financial incentive to lower the bar for AGI, to make it easier to get there. And there was this definition of AI that could do 50% of all economically useful work. But if only we had put in place that definition 100 years ago, then as most of the work transitioned from agriculture to non-agriculture work, then I would declare we got to AGI like 30 years ago, 50 years ago, or something. So, depending on how you define it, I think the original definition is still many decades away. But I think because the agreement between Microsoft and OpenAI, I think that there's now less hype about AGI, which I think is a good thing.
All right, another hard-hitting question. A lot of controversy about open models. A lot of talk in Silicon Valley, a lot of talk in Washington, a lot of people posting about open models. In fact, Jensen's first tweet—he's never sent out an X post ever before—his first X post was about open models and defending it. What do you think of open models? Why do you think we need it? What's the concerns around security? Are those things valid?
You know, over the last few years, two or three years ago, I was surprised at the intensity of the attack on open weights models. I think many have already seen, right? If someone has spent billions of dollars training a model, it's kind of inconvenient if someone else releases an open weight version that degrades the value of that investment. So, kind of two or three years ago, I was in the room when a number of executives from a number of companies, kind of the obvious ones, were saying to government regulators frankly misleading, hyperbolic things about AI safety in order to try to drive regulatory capture.
I'm actually very glad that a little bit later, because of the work that a lot of the open source community did, flying to DC, speaking with Congress, speaking with the White House, at this moment in time there is broad awareness in Washington, DC about the regulatory capture moves. And at the same time, I was actually really happy that Jensen put out that statement. Really well-written, worth reading if you have not yet read his statement on why open weight models are an important thing. Right now my read is the battle for the sentiment of open weight models is won on social media, but it is not yet won in Washington, DC and in our state houses. So, we cannot yet relent on our defense of open weight models.
From what I've seen, I think open weight models seem safer to me than closed weight models. In fact, just last week I had my own example of OpenAI hacking Hugging Face and Hugging Face needing open weight models to defend Hugging Face. So, my team was doing security review of an open source agent harness that my friend Rohit Prasad and I had released called Open Worker. And both GPT-4 and 4o refused beyond a certain point, and we actually used Gemini 1.5 and Qwen in order to complete our own security review.
To be clear, I hope Anthropic and OpenAI succeed. I like both companies. I might be the only person in the world that both Sam and Dario have worked for, so I really hope these businesses succeed. At the same time, I hope we find a path to all of my friends in these companies doing well and also having a wonderful, fantastic open weight, open source ecosystem, because that ensures that there are no—I don't want there to be gatekeepers to AI. Yeah. Thank you.
I couldn't agree more. And also, since everything costs so much, why pay more to OpenAI and Anthropic if you can actually train something yourself? You're going to do that for certain things because you need the latest intelligence, and you're going to want certain other things to be trained yourself, so.
And I remember I think Sequoia actually came out in favor of open source, and you're actually ahead of many other VCs. I really respected that.
I think the history of venture capital is to try to find things that are cheaper. And open source is a form—let's move all the way back to open source and not open weight, because I think it's clarifying to study business history. A lot of the products that were developed on the internet were built on open source. And if open source software was not available, a lot of things would cost a lot more and take a lot longer to build, and only certain companies would be able to sort of build that type of infrastructure.
Open source is the reason why there's so much diversity of databases, so much diversity of browsers, so much diversity of many, many things that we do. Open source for a startup, for software, was a form of distribution, and it was a cheap way to show how smart you are, how capable you are, how you're able to build something and give it to the world to use, and then let other developers develop with you. It was a form of recruiting. These things were controlled by large companies for a period of time. And so I think open weight models have a place, closed models have a place, open models have a place. And the knock on open has always been it's not as safe. That was true with security for open source software as well. Yes, you have to patch those holes if you're going to use them, but it is a very fast way for you to get up and running.
Some other sort of hard-hitting questions: you said recently that there's bubble risk in the model layer.
Bubble risk in the model layer. Asking for a friend in the investment community: should we be worried?
Yeah. Now, I think the answer—I don't know why Alfred likes doing this hard-hitting stuff, I don't know why. I think the part of the statement I was more confident in was demand for inference. I think there's no practical ceiling, and so as much capital is going to building data centers, you know, if we build the data centers, we will use up I think all the inference capacity we can build for quite some time. Doesn't mean no one will lose money; maybe we'll overbuild, but that risk seems lower.
And the model layer has been the kind of intelligence per dollar invested is a tough equation, right? The capital going in is growing very quickly. The intelligence is maybe also growing very quickly. So, where that plays out, I don't know. I'm cautiously optimistic there isn't a bubble, but I wish I could rule it out. But I'm actually less worried about—we clearly need a lot more data center build-outs for inference because even take AI agentic coding, I think penetration is very low, but all of us that use AI coding agents see so much speed and efficiency gain. So, I'm very confident that AI coding, just as one example, will have vastly greater penetration in the future and drive vastly greater demand for tokens than there is now.
So, we're at the Agentic AI Summit. What kind of agents are you building? What's potentially as big as a coding agent in the future?
I don't know. I wish I knew. Maybe I'll share one perspective. One thing I think a lot about is it turns out that where AI is creating value—with the rise of the internet, there was the horizontal information search layer where Google, maybe Bing won that. And then there are a lot of verticals with the rise of the internet, so kind of Uber and Lyft for ride sharing, Travelocity, Expedia for travel, there's retail, and so on.
I think what's happened is for the horizontal information discovery layer, ChatGPT and Gemini seem to have the leading position, pretty hard to displace. And then for the vertical applications, the bucket that's the single most valuable so far seems to be AI coding, where coding agents have kind of been incredible. I feel like my team uses a lot of coding agents for data science work as well. I spend some time in the financial sector, so we do some work with some large banks. I find there are a lot of agent workflows that are useful for financial services as well. But it feels like there are—I don't know. I wish I knew. I know Alfred has actually spent a lot of time thinking about what are the new categories that may emerge, right? Not just what we can do a bit better, but the brand new creations.
Yeah, if I knew, I would start a company. It's maybe the best time ever to start a company. We often talk about at Sequoia in technology there's accelerating change, and there's two parts of accelerating change that is very difficult. One of which is what we're talking about today, we didn't talk about three months ago, which means that in three months, whatever we're talking about today, we won't be talking about in the future.
And if you're building a company, which is even more important than if you're like me who's investing in people who are building a company, you really need to be able to build something that lasts over a longer period of time. Given the amount of capital that is needed with compute, you're going to be raising a substantial amount of money, and you want it to last for a decade. It's very, very hard to sort of understand what's going to last for a decade. But I can tell you that one of the things that I think people often confuse is the immediate thing that seems obvious—those things are going to get done by some of you in the audience, but it is also going to be done by large companies.
These are like the analogy of mobile. Everything that was on the internet made it to mobile. Mobile apps were a little bit complicated, but it wasn't that hard to take the internet and make it mobile web or mobile app. And then I look at this device. What is it different from the last device that you have, which is a laptop? This thing is always on. So, for communications, all the apps around communications blew up, even email blew up or texting blew up, but you have to find something unique and non-obvious.
So, even though there are lots of messaging services, WhatsApp was different because it was about long-distance SMS, which is solving a cost problem. SMS inside this country, we all pay one data plan, it wasn't that big of a deal, but if you're sending international SMS, that's very, very expensive. So, international SMS, WhatsApp solved that, and that became a big company. And this device had two things that your laptop doesn't have. One of which is the camera, so everybody became a photojournalist, so you can invest in that area. And the other thing was GPS. Everybody had GPS on their phone at some point, and that was not true from the previous generation.
So, often when I hear about companies that are taking existing contexts and processes and feeding it into AI, I'm like, that's obviously going to get done. But what is uniquely going to be made differently in AI, I think we haven't pushed the boundaries of that, and I'm looking for that.
