# AI's winners depend on unresolved technical and funding questions

Markdown copy of https://realalpha.app/s/ais-winners-depend-on-unresolved-technical-and-funding/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) · Marc Andreessen is an entrepreneur, investor, co-creator of Mosaic, co-founder of Netscape and Andreessen Horowitz. · Co-founded Netscape and co-created Mosaic, then co-founded a16z, a venture firm backing AI, crypto and frontier tech.
- Host: [Lex Fridman](https://realalpha.app/people/lex-fridman/) · Host, Lex Fridman Podcast · Robotics and AI researcher at MIT; started the show in 2018.
- Show: Lex Fridman Podcast
- Episode: [Marc Andreessen: Trump, Power, Tech, AI, Immigration & Future of America | Lex Fridman Podcast #458](https://realalpha.app/v/OHWnPOKh_S0/), Jan 26, 2025
- Moment: 3:17:22 to 3:25:10 of the episode (8 min), playing on https://realalpha.app/s/ais-winners-depend-on-unresolved-technical-and-funding/; the episode from there: https://realalpha.app/v/OHWnPOKh_S0/?t=11842

## Quote

> “And so logically, if the flip is going to happen, the for-profit has to go **raise $150 billion out of the chute** to buy the assets.”
> — [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) at [3:20:05](https://realalpha.app/v/OHWnPOKh_S0/?t=12005) · Lex Fridman Podcast

## Takeaways

1. Big versus small and open versus closed models remain unsettled.
2. Synthetic data and reliable reasoning could change model capabilities.
3. OpenAI's proposed transition could require a $150 billion raise.

## About this segment

Marc Andreessen lays out open questions that could reshape the AI race, from model size and synthetic data to hallucinations and financing. He notes that OpenAI's nonprofit-to-for-profit transition could require $150 billion upfront.

From “Marc Andreessen: Trump, Power, Tech, AI, Immigration & Future of America | Lex Fridman Podcast #458”, Jan 26, 2025. [Watch the full episode](https://realalpha.app/v/OHWnPOKh_S0/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[3:17:28](https://realalpha.app/v/OHWnPOKh_S0/?t=11848) like it's a big year so who do you think wins the race open ai meta google xai who do you think wins the ai race i would say i'm not going to predict i'm going to say there's questions all over the place um and then we have we have this category question we call the trillion dollar question which is like literally depending on how it's answered people make or lose a trillion dollars and i think there's like i don't know five or six trillion dollar questions right now that are hanging out there which is an unusually large number yeah and i just you know i'll just hit a few of them and we can talk about them so one is big models versus small models um another is open models versus closed models another is whether you can use synthetic data or not another is chain of thought how far can you push that and reinforcement learning and then another

[3:18:08](https://realalpha.app/v/OHWnPOKh_S0/?t=11888) one is political trillion dollar questions um policy questions which you know the u.s and the EU have both been flunking dramatically and the US hopefully is about to really succeed at. Yeah. And then there's probably another, you know, half dozen big important questions after that. And so these are all just like, say, this is an industry that's in flux in a way that I, even more dramatic, I think, than the ones I've seen before. And look, the most example, most obvious example, the flux is sitting here three, sitting here in the summer, you know, sitting here less than three years ago, sitting here in December of 22, we would have said that OpenAI is just running away with everything. And sitting here today, it's like, you know, there's at least six, you know, world-class God model companies and teams that are, by the way,

[3:18:51](https://realalpha.app/v/OHWnPOKh_S0/?t=11931) generating remarkably similar results. That's actually been one of the most shocking things to me is like, it turns out that once you know that it's possible to build one incredibly smart Turing test passing large language model, which was a complete shock and surprise to the world, it turns out within, you know, a year, you can have five more. There's also a money component thing to it, which is, um, to get the money to scale one of these things into the billions of dollars. There's basically right now only two sources of money that will do that for you. One is, um, the hyperscalers giving you the money, which you turn around and round trip back to them, um, or, you know, foreign sovereigns, you know, other, you know, country sovereign, sovereign wealth funds, which can be, you know, difficult in some cases to, for companies to

[3:19:28](https://realalpha.app/v/OHWnPOKh_S0/?t=11968) access. Um, so there's a, there's another, there's maybe another trillion dollar question is the financing question. Here's one. Uh, so Sam Altman has been public about the fact that he wants to transition OpenAI from being a nonprofit to being a for-profit. Um, the way that that is legally done is that, um, there is a way to do it. There is a way in us law to do it. Um, the IRS and other legal entities, uh, government entities scrutinize this very carefully because the us takes foundation nonprofit law very seriously because of the tax exemption. Um, and so the way that historically, the way that you do it is you start a for-profit and then you, you raise money with the for-profit to buy the assets of the nonprofit at fair market value. And, you know, the last financing round at OpenAI was, you know, $150-some billion. And so logically,

[3:20:13](https://realalpha.app/v/OHWnPOKh_S0/?t=12013) if the flip is going to happen, the for-profit has to go raise $150 billion out of the chute to buy the assets. you know, raising 150 billion is a challenge. Um, so, you know, is that even possible? If that is possible, then OpenAI, maybe it's off to the races as a for-profit company. If not, you know, uh, you know, I don't know. And then, you know, obviously the Elon lawsuit. So, so just because they're the market leader today, you know, there's big, important questions there.

[3:20:37](https://realalpha.app/v/OHWnPOKh_S0/?t=12037) You know, Microsoft has this kind of love hate relationship with them. Where does that go? Apple's, you know, lagging badly behind, but you know, they're very good at catching up. Amazon, you know, is primarily hyperscaler, but they now have their own models. and then there's the other questions like you laid out brilliantly briefly and brilliantly open versus closed big versus little models synthetic data that's a huge huge question and then test time compute with uh chain of thought the role of that and this is fascinating these are i think it's fair to say trillion dollar questions yeah these are big like look you know it's like here's a trillion dollar question which is kind of embedded in that which is just hallucinations right like so if you are trying to use these tools creatively you're

[3:21:17](https://realalpha.app/v/OHWnPOKh_S0/?t=12077) thrilled because they can draw new images and they can make new music and they can do all this incredible stuff, right? They're creative. The flip side of that is if you need them to be correct, they can't be creative. And that's, you know, the term hallucination. And these things do hallucinate. And, you know, there have been, you know, court cases already where lawyers have submitted legal briefs that contain made up court citations, case citations. The judge is like, wait a minute, this doesn't exist. And the very next question is, did you write this yourself?

[3:21:43](https://realalpha.app/v/OHWnPOKh_S0/?t=12103) and the lawyer goes uh i mean that's why what elon will grok uh yes looking for truth i mean that's an open technical question how close can you get to truth with lms yeah that's right and and i my sense um this is very contentious topic at the industry my sense is if to the extent that there is a domain in which there is a definitive and checkable and provable answer and you might say math satisfies that coding satisfies that and maybe some other fields then you should be able to generate synthetic data, you should be able to do chain of thought reasoning, you should be able to do reinforcement learning, and you should be able to ultimately, you know, eliminate hallucinations.

[3:22:21](https://realalpha.app/v/OHWnPOKh_S0/?t=12141) But by the way, that's a trillion dollar question right there as to whether that's true. But then there's questions like, okay, is that going to work in the more general domain? So for example, one possibility is these things are going to be truly superhuman at like math and coding, but at like discussing philosophy, they're basically as smart as they're ever going to be. And they're going to be kind of, you know, say midwit grad student level. And the theory there would just be they're already out of training data. Like they literally, you know, you talk to these people, like literally the big models, the big models are like within a factor of two X of consuming all the human generated training data to the point that some of these big companies are literally hiring people like doctors and lawyers to sit and write your training

[3:22:58](https://realalpha.app/v/OHWnPOKh_S0/?t=12178) data by hand. And so does this mean that like you have to, if you want your model to get better philosophy, you have to go hire like a thousand philosophers and have them write new content. and is anybody going to do that? And so, you know, maybe, maybe these things are topping out in certain ways and they're going to leap way ahead in other ways. And so anyway, so we just don't, you know, I guess this is, this is maybe my main conclusion is I don't, any of these, anybody telling, you know, anybody telling you these big sweeping conclusions, you know, this whole super, you know, all of these abstract, generalized, super intelligence, AGI stuff, like it, you know, maybe it's the engineer in me, but like, no, like that's not, like, that's not the that's too abstract like it's got to actually work um and then by the way it has to actually

[3:23:36](https://realalpha.app/v/OHWnPOKh_S0/?t=12216) have to be able to pay for it um i mean this is a problem right now with you know the big models the big models that are like really good at coding a math they're like actually very expensive to run you know they're quite slow um another trillion dollar question future chips um which i know you've talked a lot about um another trillion dollar question um yeah i mean all the global issue oh another trillion dollar question censorship right like and and and and um and all the as they say all the uh human feedback uh training process exactly what are you training these things to do what are they allowed to talk about how long do they give you these how how often do they give these incredibly preachy moral lectures how here's a here's a here's a good here's a trillion dollar question how many other

[3:24:19](https://realalpha.app/v/OHWnPOKh_S0/?t=12259) it. Countries want their country to run its education system, healthcare system, news system, political system on the basis of an AI that's been trained according to the most extreme left-wing California politics. Because that's kind of what they have on offer right now. And I think the answer to that is not very many. So there's massive open questions there about... And by the way, what morality are these things going to get trained on? And that one, we're cracking wide open with what's been happening over the past few months censorship on every level of these companies and just the very idea what truth means and what it means to expand the overton window of llms or the overton window of human discourse so what i experienced you know going back to how we started what i