[0:00:00] What's it going to mean to be human like 50 years from now? I think the world is finally taking seriously, and we've been waiting for this for years, that social contract is going to have to evolve and maybe sooner than people thought. What are your predictions for when this breaks 85%? Yeah, I think it really depends how much efforts the frontier labs are putting into targeting ArcGIS 3 explicitly. With the new set of models that I think will come out, the impact that the technology will have on accelerating science and accelerating the economy.
[0:00:28] We may be wrong, we may be over-optimistic, but it feels like it's about to be absolutely massive. There are a lot of interesting projects, and yet there's this one new thing that's reworking. Hi guys, I'm D, I'm a partner at Menlo Ventures. I'm one of the early investors in India, and we are investors in Anthropic. So, not to say that we'll color the discussion at all. But, yeah, let's get started. The first thing I wanted to ask you guys is, you know, both of you,
[0:01:00] Our parents, not just founders and CEOs of companies. And today, when I go and talk to a lot of parents, they're quite concerned about their kids and the future. As fathers, what do you think about the most when raising your kids today? Is it different from before? What really matters to you when it comes to raising your kids? Maybe we'll start with Francois. Yeah, no, it's not my lot. It's not my lot.
[0:01:30] Clearly, they're going to grow up in a world that's different from the world I grew up in that I experienced as a child. And yeah, think about it. Obviously, their relationship to technology is going to be completely different. They're going to grow up in this Star Trek world where they can just talk to the computer and have the computer do things. While back in my day, we had to write code like character by character. Yeah, and sometimes I'm a bit concerned as well because...
[0:01:56] I'm wondering what's going to happen to human agency and our ability to think independently. Are we just going to outsource our minds, our thoughts, our cognition, and all our abilities to the machine? What are we going to retain? What's it going to mean to be human in not even 10 years from now, but 50 years from now? My son, my daughter, they're going to live to see the 22nd century. That's something I think about a lot.
[0:02:26] What about you, Sam? Intellectually, I think it'll be great. I understand that throughout history, parents have always freaked out about the rate of change, wondered how life is going to be for their kids. I think it is awesome that our kids are going to grow up in this Star Trek world where you just ask a computer for anything and it does it. And I think they will, you know, look at us like I kind of looked at my parents with this, like, oh man, you really have to do things before computers?
[0:02:55] And my parents worried that I wasn't going to learn anything either because I was just going to look everything up on Google. It turns out I was able to have a fulfilling career of a kind they didn't have a model for. And I assume that our kids will have the same experience and they will do things that absolutely astonish us, that they will look back on us, look back on our lives with some like...
[0:03:22] Pity is maybe too strong of a word, but man, you all had it hard and bad and miserable. And the quality of life will be so incredible. Intellectually, I think all those things. I think it's amazing that my kids will never be smarter than a computer. And they will never expect that a computer can't do these incredible things. And they will achieve, and they will grow up expecting to achieve these amazing feats.
[0:03:54] I think they'll have more agency and more will than any of us do. And they're like, they'll grow up expecting that it won't be weird. And they'll develop with that way. And then we develop with computers and thought of all the things we could do that our parents didn't think of. Emotionally, I'm like, well, you know, what's this really going to be like? Do you think just a quick follow up on that? Do you think in that world, you know, when we grew up, intelligence probably mattered a lot. You know, our parents probably were like, be smart.
[0:04:24] You know, you guys went to some pretty good schools. You learned a lot. Do you think that still continues to matter in this world or not so much? Or maybe a different kind of intelligence? What are your thoughts? So intelligence is really the ability to adapt. And the more rapidly the world changes, the more important it is to be able to adapt to it. So personally, I trust my kids that, you know, they will see a level of change, technological change and social change that I haven't seen in my life. But I trust them to adapt.
[0:04:52] Kids are the most susceptible minds in the world. I'd love to get a peek into your lives. What's something that's happened in the last week or two? Like a small anecdote you could share that... Just tell us about what goes on behind the scenes. Sure. I mean, I can... Yeah. So one thing I've been doing for the last few weeks with my son. So my son is a huge Minecraft fan.
[0:05:22] He's turning five soon. He's playing so much Minecraft and watching Minecraft videos on YouTube and so on. And what they've been doing is that they've been vibe coding effectively a sort of like simple, pretty feature-complete Minecraft clone. And in fact, using Codex. So thank you, Sam, for that. And it's been so fun. And basically, so my son is telling me what he wants to see in the game, which features he wants. And he's pretty creative about it. And then I tell Codex to do it.
[0:05:51] And I haven't written a single line of code. So it's been a very fun experience. What about you, Sam? I think one interesting... I'm very grateful to get to work on so many interesting things and build... My kids can't build the codecs yet, but we can put blocks together and stuff. That's the most fun. But the most kind of intellectually stimulating thing is I think the world is finally taking seriously, and we've been waiting for this for years, that...
[0:06:23] The social contract is gonna have to evolve and maybe sooner than people thought. And it finally feels to me like there is a moment where in many different countries, many different political parties, many people who normally would think very differently about how the economy should work and how power structures should be developed. There is a real, finally, I think it's real.
[0:06:48] Desire to sit down and say, let's start talking about ideas. Let's not just keep saying, man, stuff is really going to change and this is really going to be a big effect on the economy. But the Overton window, the interest, sort of willingness to get to work on this with the people I think can actually make it happen. This last week, few weeks, whatever, first time I believe like, all right, maybe the world's really going to do this. And that's been an exciting new development.
[0:07:14] That's awesome. That's awesome. Let's move into some AI questions. This is a question I just have to ask. I know Sam probably just saw this benchmark. Francois, you probably have a good answer to this, or at least a prediction. You know, Arc AGI, Greg covered this, right? November 2019, five years later, O3 preview crushes it 88%. Arc AGI 2 launches about a year ago. Right now, I think the top score is Gemini DeepThink in 84.6. Today, Arc AGI 3 is 0.37.
[0:07:43] What are your predictions for when this breaks 85%? Yeah, I think it really depends how much efforts the frontier labs are putting into targeting ArcGIS 3 explicitly. Like for Arc 2, for instance, a lot of the very fast progress we saw in late 2025 was really because there was just so much effort, so much compute that was thrown at the benchmarks.
[0:08:08] I think if the frontier labs decide that they're just not going to look at Arc3 whatsoever, they're not going to target it, they're going to completely ignore it, and they're going to let the scores increase organically just as a function of the G of the model increasing. I think in this scenario, it would take several years to saturate Arc3. But what's actually going to happen in practice, they're going to throw a ton of compute at it. And I would say at least a year until saturation, because...
[0:08:38] It's significantly harder than Arc 2. We were deliberately trying to make it, you know, like out of distribution, like the concepts, the dynamics you will see in the test environments are very different than the ones that you see in the public games. It's going to be much, much harder to put forth for sure. So at least one year. Thankfully, we have a big lab CEO right here. So I guess this is a kind of freebie question, but does Arc AGI 3.7 and...
[0:09:07] How do you think about Arc AGI 3? Yeah, of course. I think any benchmarks that are interesting and real and not currently saturated matter to us. There are many left. I mean, there are metrics about discovering new science. There are benchmarks about discovering new science. There are benchmarks about economic acceleration. Those, I think, really matter. But there are not a lot of benchmarks in this sense. So, of course, it matters.
[0:09:37] I thought that was a very good and very nuanced answer, which is how long it takes will depend how much people target it. I hope people don't target it that much, but I'm sure people like scoreboards and they will. AGI wasn't supposed to be numbered. In the way that people first talked about AGI, it was you have AGI or not, you're not supposed to have an AGI-1, AGI-2, AGI-3. And yet, we learn how hard it is to precisely define
[0:10:10] What AGI is and how we're going to test for it. And I bet there will be a 4 and a 5 and then a ASI 1 and ASI 2 and then it'll keep going. We're working on it. There we go. But I think the main thing that points to is not that this team doesn't do incredible work, which obviously they do, but the goalpost move, AGI is hard to define and great benchmarks are hard to make. Another benchmark that's talked about a lot that sort of has no ceiling is this meter benchmark.
[0:10:43] Which is how long of a task can an AI agent autonomously do with some success rate? So right now, there's 12 hours with above 50% success rate. There's a model called Opus that does that. What do you think of that benchmark? Is that not valid? How is that different from Arc AGI? Francois, any thoughts? I like me too. I think there's good signal in it.
[0:11:06] It's also a bit of a noisy signal. I wouldn't take the claim that current AI can do like 12 hours of autonomous work. And I mean, as a counterclaim, I mean, look at any Arc3 environment. It doesn't take 12 hours to sort one of them. It takes about five minutes, like five to 10 minutes for a human. And current frontier models cannot do that. So I don't think we're at 12 hours over arbitrary tasks. So you should look at Meteor, I think.
[0:11:33] Some of the issues is if you're looking at longer time horizons, and it's definitely the regime that we're in now, it's less reliable because they don't have a lot of tasks for long-term horizons. So it's small sample size. They don't have very good human baselines as well. So yeah, let's not just take it at face value. A big difference between their approach to benchmarking AI and our approach is that we are really targeting...
[0:11:57] General Intelligence. And I feel like Meteor is more of a domain-specific or even task-specific benchmark. Like, for instance, they're benchmarking the harness. I mean, if you start in your harness, it means that you're expecting a certain kind of task. Your system is already task-specific, like it's domain-specific. And it's very heavily indexing on software engineering. And meanwhile, we're really trying to benchmark the G, like the T in AGI. So, in principle, you know, if you have a GI,
[0:12:25] It should be able to look at any benchmark and it should not need, and your harness should not need, you know, repurposing with a human in the loop for the new task. You should just, you know, go and do it the way humans do. I think it's an interesting thing to measure for sure, but I think it's inadequate in a few different ways that are very important to understand. One is that as these tasks get to five-year or 50-year tasks, it's like very hard to evaluate and we don't want to wait.
[0:12:55] Five years and you know, that's going to be weird. And then the other thing is there are things that I could work on for years and not solve and that a genius mathematician could work on for a few minutes and solve easily. There are things that no one on earth can do or haven't been able to do yet without the help of tools, no matter how much time they have that we'd really like to get done. There's something that matters about
[0:13:26] Like, you know, can we discover new science that humans have not been able to discover? And whether that takes a huge amount of compute and a small amount of time or a small amount of compute and a lot of time, like we want the new knowledge. And very soon, I think we'll be talking about things like that. Can we do things of real value to people, new scientific knowledge, new economic acceleration, and, you know, time on task is an interesting but very perfect proxy for that.
[0:13:56] Thanks. Guys, I would be remiss if I didn't ask you about, you know, both of you are apart companies that have sort of different approaches to achieving this goal. So namely, in the OpenAI world, at least what is publicly known is, you know, pre-train a big model, you do some post-training based on RL data, and you get the model to be really good. And we had reasoning happen along the way. It sounds like we got really, really good.
[0:14:22] Just doing that. And that paradigm has worked so far. Obviously, NDEA, you have a slightly different approach, Francois. You want to do deep learning based on program synthesis. So maybe the way I'd want to frame this question is, Francois, can you talk a little bit more about your approach? I'm sure more of the audience knows less about the approach. And I'd like to pair and contrast that to sort of, Sam, what you're seeing at OpenAI. Sure.
[0:14:47] So, you know, when I mention program synthesis, lots of people ask me, are you doing like code agents, are you doing code gen? And that's not at all what we're doing. We're like adding a new layer on top of the LLM stack. We're actually trying to rethink the foundations. So we call our approach symbolic learning. And it's basically a new branch of machine learning. Like if you go back to the machine learning problem in its simplest form, you have some input data.
[0:15:15] You have some target data. You're looking for a function that will map the inputs to the targets, and hopefully it will generalize to new inputs as well. And if you're doing parametric learning, which includes deep learning, then your function is a parametric curve, and you're trying to find the right parameters via gradient descent. And all of this, all these pictures, basically what we are doing, is we are replacing the curve with a symbolic program.
[0:15:41] It is minimal. It's like the shortest program that can accurately explain the data. And when you're doing this, you're much closer to producing a form of optimal machine learning, optimal learning. And there are a few consequences when you get closer to optimality. If your model is genuinely the simplest possible to explain the data, then it's also going to be the model that will be capable of generalizing the most strongly.
[0:16:08] It's going to be the most compute efficient at this time, and it's going to need the least data points to find that model. Because, you know, if the model, if you buy it, you don't need trillions of tokens to find it. So that's basically the story. We're trying to rebuild learning from the ground up on top of a symbolic substrate that will bring us much, much closer to optimal AI. And eventually, you know, we're going to rebuild the entire stack. Like I expect in the future, we'll have
[0:16:37] Very large scale foundation models that will be symbolic in nature. We want our approach to be as scalable, as compute efficient as deep learning. Sam, what do you think about that approach? And, you know, OpenAI obviously takes a very different approach. Is that going to continue to be the case? How do you think of different ways to get to this amorphous, somewhat amorphous concept of AGI? What do you think? I think that is a great thing to try. I think that people should try many approaches.
[0:17:07] We try to bet with conviction on ours, but we could very easily hit some sort of unexpected roadblock. And I think, you know, when we started with ours, it was very unpopular and turned out to have some value. And I'm sure other people will start out with things that, well, this one isn't even very unpopular. This is a great thing to try. Our approach is to push on the paradigm we understand and try to build.
[0:17:34] Towards an aligned AI researcher that can discover whatever remaining stuff is to discover. And, you know, we're like seeing promising signs, but still a ways to go. There's a specific question I got from Twitter about your timeline specifically for AI research, Sam. You at some point said automated AI research in turn would be done by September 2026 and fully automated AI researchers by March 2028.
[0:18:02] Do you have an update on those timelines? I didn't say would be done, at least that was our goal. But I feel like we are on trajectory for those goals. Makes sense. Sounds good. I also want to ask a question that's come up to me a lot, which is so many people in the research world specifically feel like they've just been dominated by AI. And they're really looking to people like yourself to say, what are the kind of research directions that are,
[0:18:32] Less explored that you wish people would do more of. Almost like a call for research. Do you have anything top of mind that you think, yeah, I wish more people tried these approaches or tried more experiments in this direction? François, do you have any thoughts? Like outside of AI? No, inside of AI, different than the current models or the current approaches that we're taking in some of the labs. Or it could be outside AI if you have something interesting. I think there are so many directions that people have not tried or not tried seriously yet.
[0:19:01] And in a sense, you know, I believe compute is a great equalizer. Like if you throw a ton of compute at one approach, you will see some really cool stuff happen. I think that's true of evolutionary algorithms in particular. I think that's been very under-explored. We have not tried to scale them, but we could. I think we would see some really cool stuff. That's true of program synthesis, and that's precisely what we are trying to scale right now. So yeah, these two, I think, are pretty good candidates. But in general, they are just this infinite supply.
[0:19:29] There was a time right after we had put out GPT-4 where I was having dinner with one of the great OpenAI researchers and he said, you know, the thing that is most sad to me right now, there are all these brilliant people in academia, there are all of these great ideas to try, everybody just wants...
[0:19:58] to reproduce GPT-4, which at the time was, you know, still a difficult thing to do. And he's like, there's this incredibly wide space, how we convince every academic department around the world to stop saying like, you know, we just need to, we just need to train a big model in the same style. And we never found an answer to that. But I think finally people are saying again, let's try some of these very different ideas.
[0:20:25] I'm not sure what's changed there, but it does feel exciting that that's happening. However, I will give an answer outside of AI, which is for the first time in the last few months, I think you can use the latest frontier models, hugely accelerate some other kinds of research. And, you know, like AI research has no lack of smart people working on it and no lack of funding. And I think it's great. But the fact that other directions of research, you know, physics, mathematics, bio, whatever you'd like.
[0:20:55] That have been difficult to make fast progress and all of a sudden feel like they might get burst open and all these factors. That feels very underexplored relative to the potential. Makes sense. I have one last question for you, Sam, and it can do the right fire. I know it has a hard stop at 730. And he asked you the OpenAI question, Sam. OpenAI is doing a lot of different things, right? So you have different modalities. You have Sora, you have Operator, Atlas. There's a bunch of different experiments you have.
[0:21:24] I think everyone's grateful that you're running all these experiments. But very recently, you said you'd shut down Sora. There was a leak about this new pre-train that you have and you want to focus all of your effort into sort of coding and the enterprise. Can you sort of throw some color on that and talk a little bit about are OpenAI's top priorities? With the new set of models that I think will come out, the impact that
[0:21:54] The technology will have on accelerating science and accelerating the economy. I mean, we may be wrong. We may be over optimistic, but it feels like it's about to be absolutely massive. And we think that will be the most important area for us to dedicate our compute, our resources. I think things are going to get pretty heated. And we, I mean, I think there's many great things, but...
[0:22:25] But several times in OpenAI and history, we've said, you know, there are a lot of interesting projects. And yet there's this one new thing that's reworking. This happened in the original GPT days when we shut down a bunch of other projects to concentrate our compute. And it feels like it's really time to concentrate our compute. I mean, to give like an obvious example, if we will be able to put...
[0:22:49] A million GPUs towards curing all cancers or a million GPUs towards making videos. I love the videos. I pick the cancer research. And as we look at the next couple of years, I think it is a time, like we've had a couple of times in our history to say, okay, we see a really big thing on horizon. Like let's point all of our firepower at it. That makes total sense. All right, let's do the rapid fire. Let's start with AGI. This is Arc AGI launch after all.
[0:23:17] I want you to answer, on a scale of 1 to 10, how far are we to your definition of AGI? 1 being nowhere close, 10 being probably going to happen this year. Francois, maybe we'll start with you. Are we talking a linear scale? Linear. I'll go with 5. 5. I think so. Obviously, we have models now that have genuine fluid intel.
[0:23:43] They are intelligent in a sense. They're not human level, and we're still in this regime where it's very easy to find things that are easy for normal people, for regular people. You don't have to look for super geniuses to be the models and things that are basically impossible for current AI. So we're still in this regime where it's not hard to find these tasks that humans can do, AI cannot. I think as we move closer to HCI, it's going to be harder and harder.
[0:24:12] And at some point, you know, it's going to have to be like, you know, PhD level tasks only. And it's going to be very, very short list of them. It's going to be very difficult to come up with them. And that's basically when we have AGI. Makes sense. Sam, what do you think? It feels very close to me. I think the only thing I'd say is really missing is some sort of continuous learning or long-term memory. But other than that, it feels very AGI-like to me already. Awesome. What is one product or experience that you've tried other than your own that has sort of blown your mind?
[0:24:42] AI or otherwise in the last month or so. So it's probably, if I can answer, for me, AT coding, like Codex and CodeCode, are really, really impressive and really useful as well. That's awesome. Sam, what do you think? These magnetic tiles for small children. Every face connects to every other face. Okay. I had to cut one open to see how it worked, but kids really love them. They're awesome. Do you know what they're called?
[0:25:11] The tiles, they're little cubes, like full cubes that every side can connect. Those are going to be sold out after this. That's awesome. Okay. What is one industry or asset class that you are super bullish on and so bearish on in the next 10 years? I'm very bullish on hardware, especially robots and drones. I think that's going to be more and more important, you know, as AI capabilities advance.
[0:25:38] We're going to want to put more and more of these capabilities into the real world, and we're going to need support for that. We're going to need robots, we're going to need drones, some very bullish drones, very bullish hardware in general. As software gets easier and easier, with like agent decoding and eventually AGI, the importance, the competitive importance of hardware is only going to rise. I could give like a three-hour answer to this question, and I had so much fun. I will just...
[0:26:04] Let me myself to a single observation. There was this quote that was bouncing around Twitter for a while, which is all of the things we've learned about society, about managing scarcity are not really going to apply to managing abundance. And it seemed really smart and really profound. And then I was like, after a day, I was like, yeah, that's kind of like not that interesting. And then it kept like rattling around my head. And I was like, you know what? There is one thing about this that is blowing my mind, which is countries like us have a great deal of debt.
[0:26:32] And the plan on this is we are going to inflate it away as far as I can tell. You know, we have 130%, someone will know the number, something like that, debt to GDP ratio. But, you know, we're like a powerful country and inflation is going to keep going and it's not that scary. If we have a forever deflationary world, I don't know how you're supposed to think about that as an asset class and what's going to happen. That's a great answer. I want to do favorite book or a book that's impacted you most in life.
[0:27:04] Field for you. Wait, anyone. Let's go first. I'm going to go with Asimov. Okay. Robots, the foundation. I've wanted to build human-level AI since I was like 16. And I think one of the big influences on me at the time was Asimov's robot stories. What about you, Sam? That's a great choice. I was going to say the beginning of infinity, but the Asimov stories maybe did have more of an impact on me when I was younger.
[0:27:33] In general, I think about this now with kids. Books that you read as a kid where you're like, I can go do anything. I can be a high-agency person. I can dream about the future. Those are really great. Was that the book for you? I mean, it wasn't one. I was like all of the young adult fiction that had some vibe of that I loved. The last one. What is a person alive today that you respect the most? Could be family, could be a famous person. So many great names, you know.
[0:28:07] Okay, let me go with a really random answer. I want to highlight a researcher whose work has really influenced me, and he's not well-known at all. His name is Pierre Voudier. He's a French researcher, so AI researcher. He's been doing AI for a very, very long time. And he was one of the pioneers on AI curiosity, like artificial curiosity, which is really a fascinating concept. So he's done a lot of great work. He's very successful, but he's also not super well-known.
[0:28:36] So, yeah, I respect him a lot. But, you know, I could go with so many different people. That's awesome. I like that. You can answer well-known or not well-known. Up to you. My mom. Love that. Love that. Well, with that, I think we should wrap. Thank you guys so much for being here. And thank you all for attending.