Leaked AI secrets at Anti Fund Summit ft: OpenAI Chief Research Officer Mark Chen

You can gather around for a little chat. Well, welcome to day two of the summit. Mark needs no introduction. I think when Sam Altman was introducing Mark to us, Sam Altman calls Mark Chen the most important person in OpenAI. So I think the theme of this conversation is AI is something that we're all thinking about, whether we're actively working on, investing in, building around. But I think there's a lot of misconceptions and also just hot takes that I think there's very few people other than someone like the people in this group, but also Mark, to give hot takes on. So excited to have this talk as we just warm up our brain for a really awesome day. Any other welcome thoughts?

No. Yeah. It's been great. Thank you, Mark, for doing this. And obviously, hope everyone's having fun. If you guys want to do anything specific or any other activities, just let us know. But that's all I had to say this morning.

Do you think people will fall in love with their AI and is it already happening?

Maybe I'll switch to more of a business question. Like there's so much capex going into data centers. I think that there's the big data center deal that popped Oracle stock like crazy and that was because OpenAI has a big multi-hundred billion dollar deal with Oracle. I think when there was that tech dinner with the president, everyone was talking about multi-hundred billion dollar data center buys and then you look at the revenue numbers of even you guys are growing super quickly but it's like tens of billions of revenue. So how does that check out? What are the people not understanding in media that you guys are seeing where the smartest people are deploying hundreds of billions of dollars into capex?

So, do you think it's fully priced in or are these things underpriced still?

How important do you think physical embodiment and 3D environments are to understand reality versus a language approach?

How far out are we from an affordable, to the average person in the United States, an affordable robot task doer in the home?

Would you trust a robot in your house to just, you know, farm off physical tasks to it?

I think everyone having a house elf, like I think if it's a 7-foot tall humanoid, it's too scary. But I think if you have like a cute 4-foot tall humanoid, so like everyone has their own Harry Potter house elf, that's probably the winning form factor.

You couldn't put the cups up.

Maybe like the legs unfold.

They can jump up onto it.

Yeah, there's a real point to what you're saying. When people are thinking about development of robots, there's a difference between robots with tendons and just fully mechanical robots because you're not going to be able to push back against the fully mechanical thing but if there are tendons that can break or snap you feel at least like hey this thing's not going to be able to overpower me or something like that. So I think there's a lot of design elements here. Actually, yeah, like when you guys talk about this kind of wearable or something like that, I think that's really interesting because the way AI is going, I feel like it's going to look like your ChatGPT is just going to be your personal assistant or kind of like a digital house or something like that, right? And you want it there to be some way where it's seeing the same world that you're seeing. So, I think what you guys are thinking in that world makes so much sense, right? You should just passively be observing all the things that you're going through and it should learn about you, right? I think the models today, it's like every time you open up ChatGPT, it's like you rebirth a whole new thing. It doesn't have any real context on what you've been doing all day, right? And you have to prompt it and tell it all the things. But just imagine it's just soaking all this stuff up and it just kind of knows what you need, right, at any given point in time. I think that would be a really cool world. Just circling all the way back to the first question, just do I think you guys actually have more insight into this kind of stuff. Like, do you think people are going to fall in love with these things? Like, are they going to replace their human friends with AI friends? Is that what the world... I feel like you guys know the social world so much more than I do. So, I'm just curious how you think it's going to play out.

Yeah. No, I think for a pretty large percentage of the population, I think specifically males that are seemingly in a weird place in life. Even more so, you know, more women are graduating college, etc. And so I think there's this male crisis going on a bit. And I think that they will find love within the AI sphere. And especially if there's physical companionship, I think it's inevitable for sure. I don't know how big it will be and I'm sure society will maybe frown upon it, but I think it's definitely still going to happen and people will have AI girlfriends in their house that they're in love with for sure.

Yeah, I would double down on that because I think when the next generation of people grew up in their formative years in COVID, their main human interaction is through text or video, which is pixels, which language models can mimic and be slightly more sycophantic and friendly and encouraging than any possible human because they have infinite patience and infinite context. I think it goes back to something like we were exploring yesterday, which I think there's just going to be... I think there's already in modern economies like producers and consumers and I think the consuming category of humans is going to be like 99%. I think it trends to infinite basically and there might just be one productive human left in like a hundred years and everyone is just consuming the labor of a superpowered running 50,000 agents superhuman or like that's the limit right? So I think the aggressive prediction is that that is where we're trending where there's fewer and fewer productive people that are creating experiences and products for everyone else that is turbocharged by AI agents and then more and more of us are consuming and maybe that's scary or maybe that's okay or maybe that's inevitable. Like I'm actually curious, what do you have a hot take of what human society looks like in a hundred years?

Yeah, that reminded me, I think ChatGPT has made me more creative. The Paul Reserve logo is actually made on ChatGPT. I actually made it myself. So, that's actually a fun tidbit I forgot about. So, you can sue me for ownership rights. It's technically Mark's logo.

Yeah, let's... I know there's a couple questions. Steve. Yeah, but I just Googled it. Says like 34% of US adults have used ChatGPT and 58% of US adults younger than 30 are using ChatGPT now. So I feel like that's a lot of power that OpenAI has for the public benefit. Let's say maybe there's a declining fertility, would OpenAI ever tweak the models if they think it could benefit the public?

When you also think about output as OpenAI now the company's about 4,000. I think a lot of even sentiment in SF is that the company's a lot more bureaucratic and slowing down now. So people go to other labs or other places. As a leader, how do you think about if that's true? What are you doing about it? Are you just going to remove half the people, focus people, do something else?

You think more at the IC level. So the people that are actually doing the work in a lot of cases which is those that have been asking and a lot of those folks even outside of the Meta raids are like hey now I have the OpenAI brand I can go start a company. It's becoming like the new one year out coming live. How do you deal with that?

Greg, how do you fight the evils of AI and the controlling of information being delivered and who is handling that? How does that get managed?

Yeah, that sounds like an ambiguous answer to me. And you know, I see a George Soros type going in and paying you a billion dollars to control the narrative and what information is shared or available. So, how do we... who's watching you? It's like the police are policing the police and we all know how that works. Same thing here.

Fast forward 5 years, you've still been transformative or new data for Snip to do you think GPT 7, 8, 9 will be what it is?

I have a question. You can't put OpenAI in this category. What company or research direction are you most bullish on and what do you think is most overrated or you're most bearish on?

You think all the RL stuff is overhyped.

Aren't they doing that because of compute constraints? Like you literally can't scale through training beyond the 50 point if you don't have the capacity.

Okay. So, that's kind of both answers which is more long pre-training than median expectation and more short RL to summarize.

Hey Mark. So the models especially with pro versions are really really good at critical thinking. And so now for a lot of people maybe outside of research that are doing work they're relying on some of these answers to make decisions. And what I've seen in interviews is that a lot of people just lose all critical thinking ability. And especially with the new version of GPT-5 that's like oh and by the way do you want me to suggest the next thing? And so you're like okay yeah yeah. So I'm kind of curious like how you think about what is going to be a great human critical thinker with a super powerful GPT-5 Pro all the time? And then when you're thinking about hiring maybe other than research or I don't know how the models work or those types of things, how do you vet for good people? Are you looking for people that are just cracked at using the models? Are you looking for people that show good growth rate? Like what does that look like to find the right talent?

That's like assembly to us is like the cherry on top.

Steve, I have a fun question for all the panelists. What is your favorite prompt? And like it could be like in a research way or maybe like on a daily use way. Curious like how you all use AI.

I think the most helpful has been like I've put in, I've been taking notes for the past couple of years. I mean years and years and years of just everything in my mind, breath work, meditation, business, everything. And I just put it all in there and I said basically help me think better and tell me how I can be better and be more productive and where are my weak spots and all that stuff. And I thought that was really powerful to do it. It almost made me cry the response. So I think that was very useful and very helpful.

I think for me, it speeds up my normal critical thinking in the sense that like okay if I'm thinking reasoning through a problem I'm just like arguing against myself in my own brain right like what's the best arguments for doing X what's the best arguments against doing X and do Y. So, I think a lot of what I prompt is like, yeah, just dump as much context in. I think people don't dump enough context into their ChatGPT instance. I think people are too worried about privacy and stuff. I'm like, I don't care. Like, dump as much information in there. I trust Mark's not going to be reading my stuff. And I just have it argue against itself. And like I try to argue and you go back and forth. And I think I feel like that accelerates like probably 3 to 5 hours of thinking to like 30 minutes or 50 minutes of work, which is great. Like I'm 10x, you know, that's like a 5x order of improvement of efficiency of decision-making.

Okay, last two questions. Do you have a question? All right, last question here. This is two. One, because an LLM is like middle distribution. As people are using it more, people are going to become more middle distribution. And so people are going to be less original and more average. And then on top of that, do you see any functional use of 99% of human names intended?

Yeah. What's the second? Well, when you say shape it as well, what do you mean?

Yeah. And then the other question was do you see any functional use for 99% of humans and feminists? Just the previous question of what do you use ChatGPT for and I use it for a lot of work and selfishly my ambition is to automate my job at AI research as a professional boy. I think a lot of us and I feel the same way. So we're trying to put as much of the hairy problems that we encounter day-to-day in research and in our jobs into the model so that we can get much better at doing that automatically. I think once you get to a point where you would rather spend compute on an AI version of a researcher than a human researcher, then I can retire.

All right, we're retiring and all becoming professional boxers. Thank you. Yeah, go link up with your guides and get ready to work out. Go over to the gym. Brandon, thank you. That was awesome. You're addicted. You're a woman, kid.