AI, Robotics & the Future of Manufacturing

Marc Andreessen
The only prospect for rebuilding us manufacturing is advanced manufacturing. The only potential is to climb the tech stack and build new kinds of factories that are fully robotic and fully AI enabled and where, you know, they are extremely advanced, sophisticated systems. I think there's a path here. You might almost call this like, you know, sort of like an operation warp speed for manufacturing, where you just basically lean hard into this and you say, look, we're going to be, America will once again be the number one manufacturing company in the world, but not, not because we are opposed to automation or robotics or AI, but precisely because we're going to embrace all of this new technology as hard as we possibly can. The content here is for informational purposes only, should not be taken as legal, business tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investor or potential investors in any a 16 Z fund.

Please note that a 16 z and its affiliates may maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a 16 z.com disclosures welcome back everybody, to the Ben and Mark show. We are getting underway with part two of our two part series on AI and startups, talking about all the new developments in AI and everything that they have to do with company building. But I wanted to start with a we have a nascent meme from the most recent show we released, which was a conversation that Ben and I were having about the topic of Boeing's travails came up and Ben sort of asked the question of, like, what? Basically, if you pick an accountant for the CEO of an airplane company, what exactly is it that you expect to happen?

And it cut to a shot of me drinking my tea, trying to not burst out laughing. So in the spirit of. It's a nascent meme on X. And so in the spirit of seeing if we can fan the flames of a new meme. And it's also a very interesting topic, because the topic of who runs these companies is like an incredibly important topic, both for the companies and shareholders and for the public in terms of the products that get built.

So maybe let's do a slightly longer version of that discussion, because Ben, as you know, or I'll steal man it. As you know, it is very common in the Fortune 500 for the CEO's of companies like Boeing to be people with a variety of backgrounds, by the way, drug companies, another example. So car companies, accountants, CFO's become CEO. Lawyers, general counsels become CEO. Marketing people become CEO, by the way, also operating executives, become CEO who have not necessarily designed new products or created new content or whatever, but have run the production lines.

And then every now and then, somebody who actually came up actually creating the product is the CEO. Although for the big companies that it's a very minority position. It's very rare in american business these days that you'd have a drug designer run a drug company or a car designer run a car company or an airplane designer or an airplane company. And so, Ben, maybe give a longer form kind of version of like, from your perspective, like, what issues does this raise? And then, you know, kind of what, what questions does it provoke?

Ben Horowitz
Well, I think that, look, I mean, the big issue or the big question is, like, what is the, what is the kind of core thing that the company does in kind of management speak? What's their core competency? And if Boeing's core competency is not building airplanes, then what is it? Is it really just keeping the cost of the airplane down? That's what they're really good at or that kind of thing.

And I think that if that's what you're optimizing for, that's what you get. And it seems like really exceedingly obvious. And by the way, in airplanes, it's not like there's not a lot of new technology and new ideas that can be applied. And, in fact, I mean, I think the first Boeing incidents were on, like, the autopilot technology, which, of course, is, you know, kind of computer technology, AI technology, these kinds of things. And so if you don't understand how, if you don't even know how to build a plane, if you have no idea, if you've never done it, if you've not even been in those meetings, then the decisions you make as CEO are very likely to be not only wrong, but potentially dangerous.

And I think this is, you know, kind of certainly true of a lot of these businesses, can imagine businesses where you could take a finance person or a legal person and have them run it. But if you want to build new things and the things that you're building are complex, then it seems quite obvious that you should have somebody who knows how to do that. I feel like a positive example of this was Microsoft eventually got to set you who, look, his background was building stuff. A lot of decisions that I think had consultants completely baffled about Microsoft were obvious to him because he's like, well, we can't build that, but we can build this. Just that kind of simple ass fucking thing.

And so one question that I had, which maybe you understand because you've been on more big boards than I have is like, how does a board make that decision? Like, how am I on the board of Boeing? And I go, well, you know, like we're building airplanes and we're building new airplanes, but that's not really the tough thing that we need the decision maker in charge to do. The chief executive has to really know, you know, public accounting. How do you get all the way to that point?

Marc Andreessen
Yeah, so there's several answers to this, and let's walk through each of them, because this is precisely for people who most people have not been in the room for these decisions. And so what I'm going to represent to you is the discussion that actually happens because I've seen it, I've been part of it a few times. So one is just like, look, and I'm going to steal. I'm going to steal man, these, Ben, so that you can respond. It's like, look, these are not just product development things.

Like, yeah, when Boeing started, it was in the, you know, the main thing was building airplanes. But like today it's basically like a nation state. Like it's this incredibly large, sprawling operation with like, you know, every many different aspects of the business. Many different kind of plates have to be spun to keep the thing running. It's in then, you know, the CEO has to kind of be everything all in one.

But a big part of the CEO is they got to be like a diplomat. You know, they've got to be able to represent the company. They've got to have like, you know, extraordinary general management skills. They've got a, you know, whatever, probably, I don't even know, 200,000 employee, you know, workforce. They've got to understand how to navigate everything from employment law to safety law to this to that.

They've got a. And then the shareholder thing. Look, the shareholders that we have are not interested in our tenure. We don't have Elon shareholders. They're not interested in our ten year product roadmap.

They're interested in our quarterly annual earnings. And the company has to be optimized financially. And so we need somebody who's going to be able to do all that. And the kind of person who can do all that is somebody who has come up in a business or general management or maybe finance role. The problem with putting a product guy in charge is they've basically, like, they spent their 1st, 20 years in a lab and then, you know, they've spent maybe 15 years or something being in charge of product development, but they just don't have the breadth of what's required to run it.

Run a company like this. Yeah. So this actually, this is a really interesting point because this is one of the things that we get into a lot in the firm and that it's probably the core thing that I advise CEO's on what to do and what not to do. And it comes back to this thing that you and I talk about all the time, which is do you hire for magnitude of strength or do you hire for lack of weakness? And I think in hiring executives or CEO's, you really have to start with, okay, there's 30 things they need to be able to do, but what do we need them to be world class in?

Ben Horowitz
What do they have to be better than anybody else in the world in? And then, like, how do we mitigate the things that aren't that? And I think if you don't start from there, you always get it wrong because if you don't start from there, then you end up in exactly what you're talking about, which is, let's find a guy with no lack of weaknesses. Oh, he knows how to do this. Oh, he's talked to Wall street.

Oh, he's done this. Oh, he's done that. Okay, but how hard is it really to talk to Wall street? Can you not hire somebody to help you do that? Or is that really got to be the CEO?

And I think that if you don't, or is what you really have to do, because the output of a CEO is decisions. And so what decisions can you not get wrong? And, you know, and then that really come to the CEO. And I would say top of that list has got to be like, the airplane doesn't fucking fall out of the sky. That would be number one for me.

Marc Andreessen
And so, by the way, I've got my tea ready for comments like that. So, yes, and so if you don't start there, then you're just going to end up with this kind of lack of weakness. They're not really great at what you need them to be great at, et cetera, et cetera, et cetera. And this is, you know, I find this so often, and it's not just in CEO's, it's really in every executive position, but we do it a lot. Like in our business, we look at companies this way.

Ben Horowitz
You know, if a company has nothing wrong with it, that's not a reason to invest. It has to have something truly great about it. And I think that the fact that, and I think part of it as a board isn't any competent person who kind of has experience in hiring would think about it this way. But I think when you get into a committee, the problem with committees always is this gets to another important point about hiring is they always go for lack of weakness because it's the easiest thing to say, oh, I want to talk. I know what I can say.

I can see this thing that's wrong with that guy. But it's much harder to be so expert that you say this person is better than anybody in the world at this. And this is the thing that we actually need. Which is also why I think that for CEO's you can't do consensus hiring because you end up with lack of weakness. That is the output.

Like at the end of the day you can take input, but there's got to be a single decision maker who knows what the person has to be world class in and knows how to assess that and then can build the plan to mitigate, complement, deal with the things that the person isn't world class out, but those have to be the things that you don't need them to be. You can hire a CFO. So that actually goes to a second thing I would say directly on point with what you just said. So the second thing is incentives at the board level. So the incentive at the board level, the sort of board decision matrix for this kind of thing kind of looks, I won't draw it out, but it's basically, you know, I make a good decision, bad decision.

Marc Andreessen
It goes well, it goes poorly kind of thing. And so it's like basically if I go for somebody who has no weaknesses to your point, you know, I'm optimizing for lack of weakness then I'm making a quote, safe choice that, you know, they may not be great, but you know, they're probably not gonna, you know, probably not gonna destroy, you know, there's, there's no clear weaknesses then they're not gonna destroy it. So I'm, you know, and you know, you, sometimes you hear a term for this like steward, right? It's, I'm gonna, I'm gonna get somebody who's gonna steward the thing, right? Yeah.

Ben Horowitz
Right. Well that's a big part of it, right. Because it's like, okay, look, our days of like, you know, being that our days of inventing the next, you know, widget are over, it's, it's now a question of like we gotta keep the wheels on. And by the way, look again, steel manning it. We have responsibilities.

Marc Andreessen
Like we have responsibilities, right. And we got like regulators and investors and all these people. And, like, we have responsibilities. The, and then, like, because if we, if we take, if we take a little bit more risk, like, if we take a little bit more risk and we get somebody who's spikier on the strengths and weaknesses, you know, kind of, kind of grid, and they have stronger strengths, but maybe they don't have, like, for example, they haven't been in a finance line job before, and then they, like, screw up the Wall street part. Like, then we're all just going to look like complete losers.

Yeah. Right. And so you see, what I'm saying is it's a very strong orientation towards risk aversion at that point at the board. Of course, you also look like a loser if the airplane, you know, the doors fall off the airplane while it's in the sky. Yes, correct.

Yeah, I mean, or if, yeah, or, you know, if you, if you double down, you know, if you, if you do the various moves, you know, if you do the boneheaded strategic moves. Yeah. So this is like, I actually think this is a general kind of organizational psychology problem. And Andy Grove kind of remarked about it, and I think it was high output management when he talked about, well, there's, you need people with ambition. Ambition is critical to get anything important done, but they have to have the right kind of ambition, meaning they have to be ambitious for the enterprise itself.

Ben Horowitz
It has to be like, you need people who want Boeing to be the best airplane manufacturer in the world and do important new things. You can't have it. I want prestige for myself for being on the Boeing board, and I don't want any kind of, you know, heat about that. And I don't want to get criticized. Like, once you get into that kind of thing.

And I think this is, you know, true in any organization, you really end up with bad decision making because now you're no longer optimizing for the thing you should be optimizing for, which is a company. You're optimizing for the kind of individual ambition of the various members of the board or the kind of members of the executive team or whoever it is. This is probably one of the most dangerous things in business is when personal incentives start to override the goal of the organization. That, and, like, it always happens to some degree. Right.

Because nobody is like a perfect, you know, all for the team in a sense. Right. Like everybody cares about themselves to some degree. And when those interests start to diverge, it's problematic. But I think the job of leadership, the job of whoever the chairman of the board is, whoever's driving.

Who goes on the board has to be. How do you really get alignment between whatever anybody's ambition is and the ambition of the company? Because when that gets misaligned, you get exactly what you're talking about. Right. Right.

Marc Andreessen
Okay. So then the third thing I would say, and again, this is a very explicit conversation that will happen in the boardroom. It's almost, by the way, this is almost a given that just everybody just kind of accepts the following at big companies, which is basically, look, we have a career path for CEO's, and it starts with MBA. Right? Like, we have this concept in american business that basically, being a CEO, being a general manager is.

It's a general manager, you know, masters in business administration. You know, business. Like, these are generalized things. These are generalized skills. Yeah.

And if I know how to run a soup company, I know how to run a car company, I know how to run a plane company, I know how to run a computer company. These. These are because it's the running the company part. It's the generalizable skill, what the product is, whatever. Like, the whole point of being an MBA or a general manager CEO is you can adapt to those things.

And then, I would say, linked to that specifically, you asked the question up front, which is, like, if Boeing's not in the business of making airplanes, what business is it in? And, like, Boeing is actually, for the most part, arguably not in the business of making airplanes. Right. Like, any more than, like, you know, pc companies are in the business of making PCs. Right.

We're car companies in the business making cars, which is. It's a supply. It's more of a supply chain integration company at this point than it is a primary airplane maker, by the way, just talk about the car companies are like this. Like, most of what goes into a car is coming from what they call their tier one supply chain, which are these integrated systems of everything in the car. The engineering design is done by other people, and then you kind of put it together and put your name on it.

The pc industry, as you know, works the exact same way, which is you're buying parts from the half the time pc companies. Now, when they ship a pc, they haven't even designed the pc. They actually had what's called an ODM design, the pc, which is sort of outsourced design. And so these companies are basically supply chain integration and then financing, sales and regulatory machines. But that.

Ben Horowitz
Well, this is precisely why GM is totally vulnerable to Tesla and BYD. This is why, you know, HP ended up being totally vulnerable to Apple, because Apple was still building computers and HP was assembling computers and GM is assembling cars. And Elon completely kind of re engineered how you build a car. And, you know, the product difference is dramatic if you are actually doing the thing as opposed to kind of, you know, basically milking the cow. So, yeah, I mean, I think that's right.

I, by the way, I think that general management, the way you describe it, is fake. And I don't think that, in my view, like, if you can manage a soup company, you cannot manage, you know, meta. Like, that's just not true. And by the way, if you can manage meta, I'm not sure that, like, you can or would even have enough interest to manage a super company. Like, those are different things.

The products really, really matter in terms of how you organize the operation, how you, you know, how you run it, the kinds of employees you have, how you are, how you can and can't talk to them. You know, there's things tech workers, you know, won't live by because they're in very high demand that, like, you know, you know, maybe a manufacturing employee would deal with, you know, some companies have unions and, like, union negotiations are a very different kind of skill set and so forth. So I think that the whole idea of that kind of general management is wrong. I think that there's one true kind of part of general management, which is, you know, you have to learn how to manage if any CEO has to learn how to manage people who are doing a job that they haven't done. So, like, okay, now I, you know, I'm a product person, and managing an HR person or a salesperson.

So, like, that's a skill you have to learn. But the decisions that come to a CEO always kind of relate back to what the company does. And so very, at a very deep level, the CEO has to understand what the company does. Like, what, what the hell does it do? And look, you could be coming from a finance background and understand, like, at a deep, deep level, what the company does.

But that's the requirement. Or making a high quality decision is very difficult because, yeah, you do run into these things. Okay, I'm running GM, and I'm going to run it like a car assembly plant, and I'm going to be caught with my pants totally around my ankles when, like, electric cars come and there's nothing I can do about it. I'm stuck because I don't even know how to build cars, really. I just know how to, like, run this thing that assembles cars and by.

Marc Andreessen
The way, look, and again, I'm still a steel man. But, like, look, the new CEO might say, look, this company stopped making cars 30 years ago. Yeah. They stopped designing cars 30 years ago. Like, I can't, like, I can't transform it into, all of a sudden a different kind of company.

I can't take. I can't build a time machine. Take it back to 60 years ago when it used to design cars. Yeah. I mean, I think that, you know, at that point, look, that's what John Madden used to call brown, brown shoes and black socks.

Ben Horowitz
You know, it's a give up. You know, you were getting dressed and you just gave up.

Marc Andreessen
Of course, these days. These days, brown shoes and black socks would be a real overachievement in the, in the fashion arts. But to be the opposite of that. But the metaphor is still good. So, okay, then the fourth, there's five of these.

So the fourth of these is, it's called the people issues. And there's two, and they're often related. So I'll just hit the two. So one is you often have a long suffering number two. And so you've got, you generally, you're often in a situation where you've had, if you've had a successful CEO and now you're doing a handoff to somebody else, that successful CEO has had a key lieutenant.

And that key lieutenant is kind of the person who kept the trains running on time. And so, you know, in the best case scenario, that's a Sheryl Sandberg or a Tim Cook. Right. Where it's somebody who's basically been the partner to probably a superstar, highly visible CEO. They've been in the back.

They've been in the back. They've been in the back office. Yeah. They've been keeping all the trains running, and they've been doing all the work, and they've been doing that for, you know, some cases, 10, 15, 20 years. And so it's, quote unquote, it's their time.

Ben Horowitz
Yeah. So there's, there's some pressure there, and then there's a related thing because everybody's like, okay, why don't you just wave that off? There's this related thing that happens in the case where you have that person, or even if you don't, or even if you go, sometimes this pushes you to recruit from outside the company when you shouldn't, which is there's the phenomenon where if we appoint anybody on the current executive team, the peers will all quit. Right. And so if we don't appoint.

Marc Andreessen
Let's take the long suffering number two. If we don't appoint the long suffering number two, then we have to. The internal candidates are all then peers on the executive team, and there's probably ten or 20 of them running all the different functions in the company. So then if we elevate one of them, all the other peers are going to be like, hey, why not me? And they're all going to punch out, and then the company's going to collapse.

And by the way, that's also the reason often you go outside is because you may actually have candidates inside at that level, but if you appoint any one of them, the team's going to quit. The theory is if you hire from the outside, you get somebody in who's right, who's not, and then all of a sudden you can kind of keep the, keep the team coherent. So, yeah, how do you. Yeah, how do you think about that? Yeah, so, you know, it's interesting.

Ben Horowitz
I wrote about this. I have a thing that I wrote called ones and twos and the challenge with often, not always, but like, often, like, if you are a, like, super high qualified CEO who makes great decisions and knows how to run the company, then, you know, you're particularly good at a lot of the creative aspects and particularly setting the direction of the company and figuring out, you know, where it needs to go. And then you may not be as interested in, like, keeping track of all the OKRs and KPI's and designing all the detailed processes and other kinds of things that, you know, that just end up happening in a company. And so you'll have somebody very close to you who can do that. The problem with that model is when you get to succession, that person does not have the qualifications to run it.

And it's hard to get people who are good at running the company enthusiastic about working for a person like that because they're not setting the direction in interesting ways. It is a trap. Look, I think the thing about people threatening to quit or quitting actually is true. And I also think that them quitting is a little overrated in the following sense. I think that we're talking about big companies in this scenario.

I think very large company. There's a couple of things about big companies. One is if you have the right leader, they'll be able to identify talent at the next level. They'll be able to recruit in talent, so they'll be able to deal with the fact that people, his feelings are hurt and that kind of thing. And I also think that, like, if you pick the right one, everybody knows they're the right one.

And, you know, that. That mitigates it as well. But I. I definitely think that, you know, that causes people to make that mistake, that dynamic. Yeah, that's right.

Marc Andreessen
That's right. And again, it's sort of scary for a board, right. Because it's just like, wow, we're the board that unraveled the management team. Right? That's like a.

Yeah. And the board has such a, like, interesting view of the company, right. Because they only know the company through the management team. So to them, it's like, oh, if an executive leaves, oh. But, you know, like, probably there's some key engineer somewhere in the company who's actually way more important than, you know, whatever the head of partner channels or whatever the hell you're worried about leaving.

Ben Horowitz
And so, yeah, I mean, I just think it's a little bit of a misunderstanding of how a company works. Right. And you're right. The way the boards get information is. It's.

Marc Andreessen
The information to a board is really stovepipe through the management team and in particular the CEO. And so it's very hard to. It's very hard for boards in practice, nearly impossible to have an independent kind of read on things happening inside the company. Which actually takes me to the fifth and final thing on this topic, which is board composition. So.

And this is not a steel man. This is not. So this is the thing that's very important. That will never be discussed in the boardroom, but it is very important, which is big companies, the boards are not selected for. Who knows?

Who understands the business? Like, that's not really very high up on the priority list, what they're selecting. It is on the, you know, the board's on mom. Like, this is always the thing I push on is like, do they understand what we do? Or am I just gonna be on this board talking to myself, you know?

Ben Horowitz
Cause I do find, like, I am on enough big boards that, like, I do find myself in that situation time to time. But, like, if you go on the website, pick a typic, pick any fortune 500 company, basically pick essentially anyone. And they all public. Board members are public. Just look at the board members.

Marc Andreessen
They've all got the bios, the smiling photos and the bios. And you just read the bios. And basically what you see for basically every single one is the board is selected for. And you write down the list. You know what they all are as well.

First, we need ahead of the audit committee. So we need like, a CPA or CFO. Right? And then we need to staff the audit committee. We need finance experts for that.

So we've got three people on the board who are mainly there for their finance skills. We've got to have. And then you write down the list we got to have of the customer. So we got, you know, maybe it's for a Boeing or something. You know, best case it's like the head of an airline or something where, you know, good news is their customer, bad news is they don't make, they don't build planes.

And then we've got to have, you know, I mean, you just go right through it. We've got to have, you know, we got to have CEO's. We have to have CEO skills. You know, people who can advise, mentor the CEO, especially if it's a younger CEO. So we need retired CEO's.

By the way, we're under incredible scrutiny by the SEC and the FTC and DOJ on antitrust. We can't have conflicts. I was on a board once where we could not add a board member who was otherwise completely qualified and would have been great for the board because what was it? They had a line of business that 2% of their total revenue was advertising, which was the business of the company that was recruiting. And so, like, that small of a little, like nascent, like, stub business on the other side basically meant that it was a conflict and it'd be an antitrust Sherman act, antitrust violation, to add.

So for legal, regulatory reasons, we actually can't have, like, other people from the, from the, you know, from the, from the airplane industry, by the way. We can't have, like, somebody who grew up inside one of the other airplane companies anyway, because, like, you know, that would be resented inside our company and would they even want to do it or would it be an act of betrayal? And then let's just say we've got other, let's say, political diversity requirements. Right? Like that are now legal requirements and depending on the state.

Correct. We have requirements on board composition that are imposed on us by, you know. Yes. By the stock exchange imposes requirements on us. Our bank imposes requirements on us.

Ben Horowitz
Independent. Yeah. We haven't talked like, you need to be independent from ownership stakes and these kinds of things. Exactly. Like, if you're a big shareholder, that's a potential problem.

Marc Andreessen
By the way, if you're a former executive at the company, it's a problem because you might not be independent. If you're not independent, you can't be on committees, you can't actually staff it. You can't necessarily be on votes. Yeah, so anyway, like, there's this long laundry list of like the. Basically this.

It's like a django puzzle that you're trying to like fit and assemble the board. And then basically by the time you get all the pieces together, you have like two or one or zero people on the board who actually are actually, like from the business. Yep. And I think this is just, this is just. This is.

I was just saying, this is just reality. Yeah, yeah, no, it's funny because it's actually the only reason that I ever end up on public boards because, you know, like as a venture capital firm, we're kind of better off not being on the public board. But what keeps happening to me is the CEO will go, but Ben, we kind of need you because you're the one who understands the product, so can you stay on? And even though it's not really in my interest, you know, I end up doing that from time to time. Yeah, yeah, no, like, I do think it's a very hard process.

Ben Horowitz
I think it's a real problem with what's happened with corporate governance in the US. I mean, I think that, you know, particularly like the ownership requirements I want, I think is actually pretty stupid in that, you know, as. Cause the board is there to protect shareholders and this kind of thing, at least that's one of its main kind of functions. And as a shareholder, I'd rather have somebody on the board who owns a lot of the company and has skin in the game with me and is representing me. Then somebody who owns no shares and is like literally a professional board member that sits on eight of these things and it's kind of yappity.

Yep. And then write the product things and so forth. But like, it is what it is. I mean, I think the laws have evolved in a way that, you know, is really like, you know, detrimental to the shareholders. We have poor laws around boards in this country for sure.

I don't know, you know, I don't know enough about kind of international, various laws on this to know who's better. But we're certainly suboptimal. It's generally worse. That's what I think overseas because you end up with. Especially when you end up basically with you essentially bring the unions onto the board and a lot of like european multinationals and then you basically bring european.

Marc Andreessen
You bring, sorry, union politics onto the board. Yeah. And that is, let's just say that, look, maybe the Europeans, they think that's important and I defer to them on that. But let's just say that's not going to enhance the discussion. For example, on product design.

Ben Horowitz
Yeah. Like if you want the company to make great products and kind of make a lot of money and grow big and innovate and, you know, and we always think in those terms. Cause that's, I guess, our job. But like, that's what you're optimizing for. It's not a good board structure.

Marc Andreessen
Yeah. So, and then, you know, just, it's just worth noting, Ben, on your point of laws, like the laws and the regulations and the pressure and the activist campaigns, like, it's all. And that's not, I don't mean shareholder activists, I mean like the social activists who come in and, you know, bear the governance activists who come in and really bear down on these, these policy issues. You know, it basically all pushes you further and further, sort of away from people who understand the business. Yeah, like that's where all the pressure is headed.

Ben Horowitz
Yeah, it gets into politics. You know, it kind of is. I think the whole, there's this much larger issue, which is, you know, what's a better idea for humanity to grow the pie and create more resources and more abundance, or to focus on dividing up the pie and making sure everybody gets an even slice. And I think that we're obviously like way in the camp of creating more abundance, but as soon as you get into politics, it seems to always flip the other way and being much more about like, okay, can we make sure whatever we have is divided fairly? Come get a slice.

Marc Andreessen
Okay, good. Well, let's get to part two of our AI discussion. And we have some fantastic questions from x, as usual. So I have actually four questions in one that'll read that are all related. So our friend Beth Jizos asks, do you see a resurgence of hardware startups on the horizon given that the current scaling of AI is going to be limited by the scaling of compute and energy?

And of course, our friend Beth Jesus has a hardware startup directly trying to address that. A block asks, is energy production a limiting factor in the future of AI? Jeremy asks, as AI, crypto and electric cars rise, tackling energy challenges is key. How do you know, how do we address those? And then Tara asks, energy being a likely bottleneck, how much more importance do you see in the more algorithm aware hardware innovations like thermodynamic computing, which is the kind of thing that Beth Jesus is working on?

So, Ben, AI, energy, and then hardware, are we going to see basically, is it boom times for new AI chip startups? Is it boom times for new energy startups is it boom times for entirely companies doing entirely new computer architectures? Yeah, and data centers, by the way, another hardware thing. Yes. So I think the one that is probably most likely just because I think we're stuck without it is on the energy side, which is we've funded, of course, portable nuclear energy.

Ben Horowitz
And I think both on the fission side and the fusion side, if we don't have energy innovation, I think we have a huge problem with AI because when you look at, you know, once we get by the chip bottleneck, like, if we had unlimited chips, the AI power consumption would be like over 10% of global power consumption, I think is pretty clear just on AI, which people were like, you know, campaigning against crypto because of the, like, bitcoin mining, which is like tiny, tiny, tiny compared to what people are using on AI now. And that's tiny compared to what people want to use on AI but can't yet because there aren't enough chips. So I think power, hardware power for sure. And then you get into, okay, well, what's the next thing that you run into? And it's, well, if you are having like a gigawatt data center, cooling becomes a very hard problem.

Like even if you like, you know, stick it in the middle of the ocean and use the ocean light, like you're starting to boil all the water around the thing. It is. So, I mean, this is really, really, really high power consumption. So then you say, okay, well, chips that are kind of that throw off less heater or consume less power and so forth become like super, super interesting. For sure.

On the chip side, there's been a lot of, we've looked at a lot of things that are, okay, can you be like, AI optimized? And, you know, it's, I think on a general chip, it is really hard to comp. Like, the current chip makers. Like Nvidia is really good and innovating and AMD is really good and innovating and so forth. So to start a new chip bank maker, you need a pretty novel angle.

We've seen ones that are, that are quite interesting. Like, one is like, okay, we'll make a chip for a specific, like, model. Like, so we will take a model and we will like. And the whole process is designed to make a chip for that model. And given the cost of training that you have to amortize across all the inferences that you do over time, the math actually works for that idea where you would have, you know, basically a model on a chip.

But then the question is how long in the market does that architecture last and how long can you count on it? And then does that work? So, so that's another point of innovation. And then, you know, like, you know, data centers, we have not built data centers like this that require this level of power and cooling. And, you know, I think this is something that every, I know, you know, on, on the metaboard, I know they've been talking about this and, you know, Zuck is very focused on it and I know that, you know, Setya is very focused on it at Microsoft and I'm sure Sundar's focused on it.

So I think thats going to be another big area. So there is a lot of hardware opportunity. What goes to new companies and what goes to big companies is still probably an open question, although I think energy is very likely to go to new companies, particularly in the area of nuclear. So weve talked a lot about in the past, and ive talked a lot in the past about we have funded many hardware companies of various kinds, consumer hardware systems hardware. And as you said, were now doing.

Marc Andreessen
We have many drone investments. We have self driving car investments, we have energy investments. Now, you know, energy systems companies. We talked in the past about like, you know, hardware companies are harder than software companies. Yep.

And, you know, the way I, you know, I always just describe this is there's just like a lot more ways for a hardware company to fail because like, you know, not, it's not only can you, like, if, can you design the thing, can you manufacture it? Is it going to work in the field? Is it going to get recalled? You know, can it work in harsh conditions? You know, is it safe?

You know, is it regulated? You need a way better CFo if you're building a hardware company, that's another. Yeah, yeah, exactly. And then, yeah, and then also your cycle times are lowered. Software cycle times are usually very fast.

Hardware cycle times are slow just because it takes a long time to get, you know, something all the way through the production process. And then if something goes wrong, there's a recall. The recall can kill the company. Supply chain issues. You know, to make a hardware thing, you need every component all the time.

If you run short on one component, you're stalled out. So there's like ten different reasons that hardware, hardware companies are harder. Like, but however, like, look, when one. Works, like less competition, get it to work. Yeah.

So if it's like, if it's an important problem in a big market and it works, like, you know, you can really have something amazing and you can really change, not just the world of, you know, bits, but the world of atoms. And like, you know, many of the legendary companies in the history of the world have been the ones that have been able to do this. And so anyway, like, how, how should entrepreneurs think about us as a firm? Like, in terms of, like, what we will do, what we won't do, like, how we process through that thought process as we think about it internally? Yeah.

Ben Horowitz
So we have been like, we invest in hardware companies. As I mentioned, I do think the bar is higher. And some of the things, so some of the things that wed look for in a hardware company that we dont own, a software company, is the first thing I think is we dont take the financing risk for granted at all. So if we invest in a software company, whatever we do in a round at, you know, ten on 40 pre or something like that, whatever it is, we're not really, like, if it works, it's definitely getting financed, right? Like, that's no problem.

With a hardware company, even if it's working, you're going to likely hit multiple valleys where you're both low on cash and it's not working well enough yet to justify the valuation, and that becomes existential. And if you're not familiar with this, you can read, like, Elon's, the book that Walter Isaacson wrote on Elon because it kind of goes through a lot of the crises he ran into in both SpaceX and Tesla. And that's like Elon Musk, who's like, as good at hardware company building as anybody in the world, but they all have that characteristic. And so the CEO has to be, as one of the things that they do has to be, like, a world class fundraiser. Like, they can't be, you can't be, you know, like, I don't like to raise money.

You know, like, look, we have a lot of software CEO's who are like, I don't like to raise money. I don't want to talk to investors. Like, f them. Like, that doesn't really work on a hardware company. I think that, you know, getting back to, like, what are the world class strengths you need in a hardware CEO?

Like, one of them is like, you got to be able to get the money because it's going to happen. And, you know, we've seen that and, like, we try and be helpful, but, like, we can only help to, you know, a certain degree. Like that. The CEO has to be, like, massively compelling on that. And then I think, like, the precision with which you run the company, you know, like, quality ends up mattering a lot, you know, the details around the metrics, the optimizations, the, you know, and this is, you know, again, this is where I think, to me, the most interesting thing in the Elon book was just like, how focused, psychotic, creative he was about, like, ways to save money, ways to make things more efficient, et cetera.

It's so, so critical in hardware to have that attitude. And you can't just, you know, build things, you know, fat and happy, the Silicon Valley way. You can't have, you know, like, you can't focus on free lunches and organic juice and all that stuff. You really have to focus on cost and, you know, and then, like, inventory things and that kind of stuff becomes just fundamental to the business. And so you have to be focused on that.

So it's just a, I would say a more complicated thing. You have to be a great recruiter because you have to bring in people who are world class at things that, you know, at a really broad set of skills. As you said, mark, you know, like manufacturing, like, you know, you need a team of finance people that that's really, really good, all that kind of stuff. So it's a similarly higher bar in terms of, like, just the complexity and the competency that you need in the company for it to work. Yeah.

Marc Andreessen
This is sort of the thing that I try to tell entrepreneurs. Maybe important for people to know is, you know, there's this meme. There's this meme of like, you know, basically VC's are scared of hardware. You're not. Not bold enough.

True meme? Well, partly true, yeah. I mean, we are scared of it. Like, not that we won't do it, but like, it is scary. Yeah.

Okay, so, okay, but then why is it scary? So. But I want to double down on something you talked about. So a big difference. So basically, the way an entrepreneur.

Right. The entrepreneurs are trying to get through the next phase, right? And so when an entrepreneur talks to VC, the entrepreneurs thinking, I need to raise this round from this vc right now. And if he tells me yes, that's great. And if he tells me no, that's a pain in the ass.

It was waste of time. And like, you know. Right. And so it's like this. It's, you know, this.

I'm focused on this round. And of course, entrepreneurs have to think that way because you have to, you have to raise money. You have to get to get, get, get to the short term, to get to the long term. The C's, the way we think about it is. Okay, Ben, to your point like, okay, we fund you for this round.

What happens in the next round? Yep. Right. What happens, we fund you in the series a. What happens in the b, and by the way, what happens in the c, the d, the e, the f, g, the h, and the I.

Right. The more then the more complex thing that you're doing. Right. And the more, the more of its hardware, the more of it is complicated. The more of it is like fundamental advances, the more of it is complex integrations, the more of it is all these things.

It's just like, okay, where is that money going to come from? Are there going to be other capital partners down the road for this company? Or, by the way, are we it? And if we're it, we need to go into that with our eyes open. And by the way, what does it say about the company if we're the only possible kind of funder of it?

So we're thinking, as we're sitting in the meeting, we're thinking in our heads, like, okay, what happens next? And I think it's important for entrepreneurs to know that. I think as an entrepreneur, if you really wrap your head around that, it's like, okay, that needs to be part of my plan. Ben, this goes to your point of the entrepreneur has to be a great fundraiser. Like the entrepreneur, as the entrepreneur, you have to take it upon yourself to be like, okay, I really have to, like, I really have to think this through beyond just this round, but to the construction of the company over time.

And, like, is my story and my plan and my team and everything else that I'm doing, is it good enough and does it hit the bar? And is it, you know, kind of organized, orchestrated in the right way so that it's going to. I'm not. So that I, not only am I going to be able to raise an a, but also all the following rounds, and you're not going to have all the answers for that, but you can have a plan. We find entrepreneurs who have really thought this through, and then this goes to what I always say about the venture process, which is getting, yes, on a venture series a is the easiest thing you'll ever do.

We're in the business. Well, that's, by the way, I think we cause problems because that is true. So if you're an entrepreneur, your experience in fundraising is you raise your series a and you're like, oh, that was easy. And then you think the next rounds. But like, yeah, that's always the easiest round you raise.

And by the way, everything else you do, recruiting is harder and sales is harder and partnerships are harder and government relations is harder and marketing is harder. Right. Getting a good press story written, you're not getting your ears ripped off by the press is harder. So, like everything else that happens, basically everybody else you deal with is harder than we are. And the reason for that is just very straightforward, which is like our entire business is to like, sit there and let people sweet talk us into giving them money.

It's like our entire existence is that. Yep. Everybody else you deal with down the line, like, that's not their job. Their job is something else. And then, you know, whether they are going to go to work for you or buy your product is like they've got many options of who they go to work for or what product they can buy or whether they buy any product.

Ben Horowitz
Yeah. And so if you can't clear the venture series a test, if you can't get like, through us and firms like ours, like maybe we're the idiots, like, you know, maybe we're making a huge mistake, but also maybe. Right. It's because your plan does not yet hit the bar that's necessary not just to raise from us, but to do all the downstream things that you're going to have to do to succeed. And I think the, I say the smartest founders understand this and they think ahead on this and they use this as a catalyst to make sure that their plan is good enough to succeed throughout the company, throughout the life of the company and with all the constituents that theyre going to need to appeal to.

Marc Andreessen
And they understand that were just a small part of that. I think thats right. And I think that now that were talking about it, I think that if we werent a multistage firm, I probably wouldnt want to do hardware deals because I think that they do get into these states where I'm thinking, you know, we've got a space company. I won't name the name, but like, you know, they just got into it's a great company. They've got great contracts, they're growing.

Ben Horowitz
They got their stuff to work. But look, they hit a point where everything was working and they had one bad part from a third party. And by the way, that would have ended the company or potentially could have ended the company. But because we knew so much about the company and our multistage, we could step in and go dip. Like, we'll do this round, give us a good deal, we'll do the round and we'll keep pedaling and be on to the next thing.

I mean, knock wood but it looks like this company will be a great success. But if we weren't multistage, we would have lost the company, I think. And I think that is often true with hardware companies. And so that doesn't mean we would always be the funder of last resort. But if the thing is kind of genuinely working and they hit one of these incidents, which hardware companies always hit, I mean, you're like, you seem like Tesla had to go back to its like existing investors, right?

Like they could get new money. And that seems to be, I don't think that we've done anything in hardware where that hasn't been the case. I mean, the one, one was, was Oculus, but Oculus got, you know, they sold relatively early to meta and so. But had they not, they definitely would have been in that situation at some point. Yeah, that's right.

Marc Andreessen
That's right. Okay, good. And then let's go to related topic. Sunshine Sunshine vendetta asks, by the way, that's a great name if it's a real name, and a great name if it's a, if it's a pen name. Will nations specializing in low cost power data centers for AI emerge similar to oil rich countries?

Are these nations good investment areas for buying land to expand or build such data centers? I think nations are definitely looking at this, and I think it's not just so there's a regulatory aspect, there's an aspect of, you know, do they have like oils popping out of the ground? You know, so there's a lot of pieces to it, I think. I don't know the answer to that yet, because I do think it depends on like, okay, how fast does, how fast does the nuclear options, funny how that sounds, emerge? And then, you know, how do countries think about the regulatory environment?

Ben Horowitz
And does the kind of climate agenda kind of help the countries that are more flexible on that? So we're getting into, you know, one of the things that I think you and I learned not to predict is things that involve, like, large governments jumping into the economics of it. Because we learned this in the financial crisis of 2008, where I think when we were looking at it, we're like, well, if this plays out, were in a pretty serious depression. And of course, what happened was the global government just poured enormous amounts of money on the situation, which is causing, by the way, all kinds of problems now. But it was unpredictable in how it was going to unfold.

And I think that this one is similarly unpredictable because it involves government policy in a really major kind of way. And then also the arbitrage of that, you know, government policies against each other. But I think it, it certainly, you know, if I had to bet on it, I'd say that probably will happen. Yeah. Yeah, that makes sense.

Marc Andreessen
And then there's, you know, for people who haven't tracked this lace, it's an interesting thing where training particularly, can be placed sort of anywhere. Yeah. And that's, that's the biggest cost in terms of power, for sure. Yeah. So for people who haven't tracked this, Ben, see if I have this right.

There's two, two parts to AI the way we have it today. There's the training phase and the inference phase. The training phase is getting the dataset kind of assembled to be able to answer questions. And then the inference phases, when you actually ask a question, providing the answer. The training phase is what's, what used to be called a batch process, where you kind of do it all at once.

And, you know, and by the way, it may take months and it may, you know, some of these training runs for these big models now may be three months. Yeah. Oh, yeah, yeah. Three month run. But you're not responding to user queries in real time.

And so it kind of doesn't matter where that happens. It just needs, you need a data center big enough with enough power and cooling to do it. And then the other interesting thing that has come up some of these discussions is training runs can be paused. Right. You can actually, like, you can shut a training run off for two weeks and then start up again, which, you know, you can't do if you're running a Gmail service or a search engine or something.

Ben Horowitz
You don't need whatever. Five nines. Yeah, exactly. Yeah. And, or you could run it with intermittent power.

Marc Andreessen
And so, like, for example, if you had a nuclear reactor that needed to be shut down every now and then for maintenance, that would be fine. Whereas you couldn't serve like Gmail off of that because you can't handle, you can't have the service interruptions. And so there is a lot of flexibility on the, on where the training runs go. Having said that, Ben, there's also a fair amount of scrutiny. Governments are putting a fair amount of scrutiny.

Number one is some governments don't want training runs to happen other places or will have concern about that. Yeah. That's another regulatory layer. Where are you allowed to train? Who's an enemy?

Ben Horowitz
Do we allow export of AI at all? Or is it some kind of munition? Yeah. And the more you're talking about these giant runs that are in you're talking about these deployments of capital and these what are viewed as strategic assets. And then of course, God knows there's also plenty of regulation, regulation around nuclear reactors and all the other energy permitting kind of everything around all that.

Marc Andreessen
Do you allow civilian nuclear power? So there's a lot of this would intersect with government policy. Even if it wasn't considered to be. AI wasn't a topic that was itself relevant to government. But then on top of that, AI is itself a topic that's now relevant to government, which makes this something of heightened interest.

So it's a complicated, super complicated. Super complicated. It's a great question. So anyway, the reason to go through that is. Yeah, are such nations good investment areas for buying land?

Maybe, but you want to consider the economics of that. But you also want to consider the geopolitics of that. And if you're going to be an investor like that, you want to think in both layers. You almost have to be like a Stan drunken Miller type, like super macro mega type investor to figure that one out, I think. Yeah, well, the, the other thing is, right, like there's this whole AI alignment issue where like, you know, different governments have very different requirements on that.

That's right. Yep, exactly. Yeah. So, like a lot, a lot of countries, like, from a us perspective, like a lot of other countries may not want, you know, the AI that runs in their country to be an AI that has been, quote, aligned by, let's say, a crop of, you know, millennials in San Francisco. Typically, if your country follows, say, sharia.

Law, for example, or by the way, vice versa. Right. We may not, we may not want that in reverse and so, or in any given direction. So there's, there are. Yeah, there's an increasing number of questions on that front.

Okay. Ori asks, I love this question. Ori asks, I love this question. What do you think of selling work produced by AI versus selling software itself? Will AI enable service businesses to become the new norm?

Will their margin structure converge with that of software only businesses? And so, Ben, this would be the idea, which this idea has been around for a long time, including before AI. But it's a particularly potent question now with AI, which is, okay, are you going to build, if you want to do legal AI, are you building AI software that you sell to law firms? Or are you trying to run, literally, a robot AI lawyer service? And then same thing for accounting, same thing for Dot, dot, dot, every other aspect of any sort of services business.

Ben Horowitz
Yeah. So it's interesting because, like, it's already emerging as a real pricing model in many ways. So examples being like, so Waymo sells an AI driver that is basically the product that's an AI driver. Turns out the AI driver is, like, reasonably expensive compared to, like, a human so far. So we'll see.

This proper thing is all this hardware is real, these GPU's and whatnot. But Elon, to your point, Elon has declared, I think, that the future of Tesla is that he's building robotaxis. Yeah, like, that's the, that's the end state for the Tesla car business. Well, I mean, I tell you, Waymo takes it to further thing because they're not selling. They're like, we'll sell you the driver for your car or for your truck.

Right? Like, well, we'll install it. We'll, we'll fit it to it. And then, you know, Devon, the new kind of programming tool that just got the wonderfully high valuation they're selling, you know, they're, they're basically kind of charging you for an engineer, is the way they think about it. And by the way, that engineer at least, you know, looks to be a, like, she may be fairly expensive or he, I guess, Devin.

Marc Andreessen
No, could be either. Could be either. It's a dual. Dual use name. Yeah, that's right.

Ben Horowitz
That's dual use name. Could be. Could be he or she, and it could be a human or a robot. And then, you know, we have a company, hippocratic AI, which is a really interesting situation because it's sort of an AI nurse, and they similarly price it like a nurse. And it's not, you know, you may think, oh, my God, an AI is going to replace a nurse.

Well, you know, like, obviously an AI can't really, at this stage, replace a nurse fully. You know, nurses do a lot of things like, you know, wash you and stuff. There's no AI that can do that. But there is a massive nursing shortage. And there's a lot of nursing work that, you know, nurses don't necessarily want to do.

Like, ask you or, you know, what medications are you on? You know, like, when were you born? Or, like, call you later and say, did you take your medicine? These are important tasks, but things that can easily be done by an AI and Hippocratic Scott and AI that can kind of augment your nurse workforce, which is, as I said, I have a huge shortage of nurses right now in their model for charging. And it works because the people who hire nurses where you would sell this product into, are used to buying that way.

They're like, okay, this is what we pay for. Nurse. Okay, you have like a, you know, a very tireless but less functional nurse. How much, you know, does she cost or he cost and away they go. So I think that idea is a good idea.

I don't think it works in the way or it's not yet working in the way that kind of people fear in the dystopian sense, where like this thing's going to come in and replace my job. It's more, this thing is going to come in and replace the parts of my job that I really dont like doing. And then its not going to be as cheap as running a software program over and over again because the power cost of these inferences and to get them right and so forth is actually still pretty high right now. By the way, the economic numbers came in today and actually productivity growth in the US economy was super low in the last quarter. Yeah, always.

Yeah. Like, I mean, I think that's the most overblown fear ever. Right? Like, so we've now had this revolution started in 2018 and I think unemployment has gone nowhere but down since then. And, you know, like they're going to take all our jobs.

Every automation has this like total crisis of fear and, you know, we'll see. Like it may happen yet, but it certainly hasn't happened yet. Well, what's happening right now is the number of people being employed building AI is and deploying it. It's like rising really fast. So like, AI developers, AI consultants are like exploding as a job category, I think, far faster than any work is actually being replaced by AI.

Yeah, yeah. And I think you and I know why because like, in order to actually, if you're really going to replace actual work and people, you need to start a new company, basically. I mean, it's very, very hard to go in and re engineer the way, you know, good luck re engineering the way Boeing works, given it's run by an accountant. Sorry to go. Or good luck re engineering a european manufacturing company that has the union occupying half the board seats.

Exactly. So, yeah, good. Let's jump right into the. And I think we're already over an hour, so maybe this will be our final topic. But it's a big one.

Marc Andreessen
So three part question. Doctor Ali asks, how long before AI is fully integrated in robotics? Polynumera asks, do you think this latest robotics hype cycle is going to generate enough data? This is a complicated question, so I'll go through it. That the bitter lesson will overcome Moravec's paradox.

And so what that means is basically effectively, what that means is the AI is going to get good enough where the robots are actually going to work and create generally useful robots, or there will be at least one robotics winter death zone down cycle before they achieve general usefulness. And then Neg asks if data is the only thing that matters, which we don't say. But the premise of the question then why aren't we seeing more AI agents and robotics companies going end to end ML like Tesla? And the reference there is Tesla. Every Tesla car, whether you have the autopilot turned on or not, it is gathering sensor data from the cameras that is being brought back to Tesla HQ and used to train the Tesla self driving capability.

AI. So why aren't we seeing more AI agents and robotics companies going end to end ML like Tesla, and building data flywheels by acquiring large amounts of data. So, Ben, AI and robotics, yeah, this. Is a super interesting question. So just kind of to frame it.

Ben Horowitz
So if you, the successful robots today, you know, in the kind of large scale like the auto manufacturers and so forth, are, they're kind of very rigidly programmed to do very specific tasks. So, you know, put this bolt on this wheel or whatever kind of thing, and then if the wheel is off by like, you know, 2 mm, it'll like, you know, put the bolt right through the wheel or whatever. Like, it's not, it's not very adaptive AI or like the kind of current deployed, state of the art. And so there's been kind of a movement for a while to do kind of intelligent robots where the robot can learn the task and then do it more flexibly. So if something's out of place, you use computer vision to see it and then kind of plan the task and execute the task and so forth.

But that's only worked okay so far. You know, it's not been like the big breakthrough where, like, all of the sun, you can have, you know, basically robots walk in and substitute for humans. Like, you know, the Optimus prime thing that Tesla's building hasn't really worked actually yet. And so then, you know, with llms and these grade generative models, like, why is that? And it turns out, like, there does seem to be missing a missing technological piece, which is that to build an effective robot that can just operate in the world like a human, you need, it needs to under, it needs to really understand physics, you know, very well, you know, to not, you know, end up being either dangerous or just clumsy and wrong and these kinds of things.

And there's varying theories on how a robot might learn physics. You know, one being, well, if I just watch enough video, then I'll basically inference, I'll just figure out how physics works, because in order to predict what happens in a video, videos of real world all follow the laws of physics. And so I'll just follow the laws of physics, too. That's been kind of a big theory, and that would kind of enable also generated video and robotics and all these kinds of things. Things.

Now, there's a different theory that says that's actually never going to work because you're never really going to get physics down. And the only way to do it is you have to build a new kind of model that, you know, may have some transformer capabilities or whatever, but also has a fundamental understanding of how the real world works in three dimensions, you know, with full, you know, physics and that kind of thing. And that is work that is being worked on by, we've got a cell startup that's doing that, and then Elon is certainly doing that in exit AI, pursuing that kind of venue, whereas OpenAI with soar is kind of going the other route and kind of not necessarily explicitly teaching physics, but implicitly learning physics through video and video games and these kinds of things. It'll be interesting, but I do believe, I think the age of robotics ought to emerge, certainly in the short to medium term, as these things are worked out. But it's not tomorrow or in three months, I don't think, because I think the AI doesn't work well enough yet.

I guess that would be my short opinion. So the theory for optimism, I think, would be basically what's happened at Tesla. So let me, let me kind of describe that for people who haven't studied it. So there was a reference in the question to this thing called the bitter lesson. So the bitter lesson is this famous thing in AI that somebody wrote a paper on years back, and basically it said, the bitter lesson of AI is basically the thing that works is more data.

Marc Andreessen
And so basically, over the course of 80 years of trying to get AI to work, basically the efforts to do what you might call kind of top down, directed, explicit AI, where you're like trying to teach a machine, you know, the laws of physics or, by the way, common sense or language, and you're doing that in a top down way, basically, that just doesn't work. And the reason that doesn't work is because there's just, there's always edge cases, and the edge cases always get you and, like, it just, you never actually finish the job. It's never actually that useful. And, you know, Ben, you'll remember in the eighties, there were all these attempts to do, like, expert systems for medicine and so forth. It didn't work for, you know, very similar reasons.

Like, you know, it could diagnose if you have the common cold, but, like, if you had, you know, I don't know, diabetes and a broken foot, like, it would get very confused, right? And there might be consequences of the combination and so forth. So, so anyway, so the, but then, on the other hand, basically the breakthroughs that have happened, you know, especially since 2012, have, you know, essentially all been breakthroughs, or, you know, most of them have been breakthroughs where basically you just get like, these giant data sets and, you know, you, you. How. Why does it, why did image recognition start working?

The big thing was this. You could, you could train an AI off the giant data set of images off the Internet that didn't used to exist. Why did, you know, why did self driving cars start to, you know, why did, why did. Or I'll come back to that, but why did, why did text, why did LLM start to work? Is because you could train them on the entire corpus of text on the Internet.

And then basically what Tesla has found, they're very public about this. What they found is the self driving car capability that they had when it was a top down system sort of worked. But the one that they have now that's working really well is entirely based on basically neural networks and large amounts of data. And the more data they get, the better that performs. The whole thing on the giant data is somewhere in the giant dataset.

If you have a sufficiently large dataset, somewhere in the data set is every single scenario you could conceivably imagine running into. And so a construction zone with police cars and flipped over in a crash and a fire. And like, okay, if the data set is big enough, it has examples of that. And so you can kind of train it on what to do, and so it can adapt to the real world in a way that you just can't if you're trying to anticipate everything. And so that, that's the bitter lesson, the good, the bad news.

The bitter lesson is there's lots of AI techniques that don't work. The good news of the bitter lesson is more data does seem to work. It does seem to solve these problems. And so what, what Elon set out to do with Tesla is basically, just, as we said, turn every Tesla, it got like a million cars in the road. Turn every car into basically a roving sensor that's basically pulling in all this data and sending, you know, and this is all right, this is all imagery outside the car, so there's no privacy issue.

It's just you're driving around in the world, so you've got a million, you know, rolling sensors with, you know, whatever, 32 cameras or whatever pointed out whatever it is. And you just. You basically stream all that back and you just keep training the self driving car algorithm and it gets really good. And so in theory, you know, if you. If you believe all that, like, in theory, that then is the answer for how to train humanoid robots and how to train, you know, every other kind of thing that coming along.

And we actually kind of, you know, we have a large part of the problem, at least in theory, solved that we didn't realize before. Yeah. So, yeah, yeah, I think that. I think that's right. And then there is a kind of a bit of a question on sort of how.

Ben Horowitz
Well, two things. One is like, okay, which neural net, like, which architecture is going to work for this particular class? Or do we need, like, some change in the architecture to make it really work? Because we had image recognition working really well before we got to language kind of prediction, which required not a dramatic new architecture, but new architecture. And there's a question on, okay, do you need a different model for this?

And then I think the other thing with robots, as well as with self driving cars and why we don't have, you know, why all cars aren't already self driving cars is the edge. Cases end up being really, really important when it's life and death. And so, you know, an LLM hallucination is really cute and funny and a kind of a Tesla auto driving hallucination would be really horrifying and deadly. And so the amount of engineering to get from, you know, 99% working to 100% working is a lot. So those would be the things.

But I think you're generally. I think what you said makes. I think that's right. I think that's exactly right. Let's say it's reason for optimism that maybe we didn't have five years ago.

Yeah, yeah, yeah. No. Like, I do think robotics are going to work. And look, it's going to be amazing. I mean, you know, one of the things that I think people, like, get into fear about these things because they underestimate, like, what's possible in terms of work for humans.

And look, the original manufacturing jobs, which we all like, those were good jobs. They really, in some ways, were very, very bad jobs in that Henry Ford very famously doubled the minimum wage. And everybody, it's funny, he's celebrated by socialists now because he doubled the minimum wage voluntarily, on his own, for seemingly no reason at all. But there was a very core reason, which was all his workers were quitting because they hated going like this all day and they wanted to go back to the farm and be like, let me milk a cow and wake up at like, crack of dawn and that kind of thing. That's way better than coming here for 8 hours and numbing my mind.

And if you look at Detroit, Detroit quickly became a drug capital, and a lot of it was like, so many of the people on the assembly lines were just on drugs because the job was so damn boring. And so I think, you know, there are more interesting jobs for humans than jobs that robots can do. And so it should be a real breakthrough. And like, a lot of these jobs we do for free, like, you know, the laundry. Right, right.

Marc Andreessen
Exactly right. And then actually, Ben, that takes us to the final question, which is a big one and something I feel strongly about. So Dominic asks, could AI plus robots plus automation be used to reboot us industry and also reboot old industrial towns with capacity for factories? So let me frame the answer negatively first, which is the only prospect for rebuilding us manufacturing is advanced manufacturing. The only process is, the only potential is to climb the tech stack and build new kinds of factories that are fully robotic and fully AI enabled.

And where they are extremely advanced, sophisticated systems with that run in a very different model than a manual labor system. And you pour as much technology in them as you possibly can. And the reason for that is just very straightforward, which is the big thing that caused us manufacturing to a lot of it to move offshore was cost of people. Yep. It was just flat out cheaper to hire people in Korea or Vietnam or China or India or other, wherever you're at Mexico.

It's just flat out cheaper. And it just, the cost of labor was just too much. It's just a high, it was a high percentage of the thing. And as you said, ben, like, these aren't, these weren't. A lot of these actually were not very appealing jobs.

And so you had to pay people a lot of money for them. And there were lots of issues involved, and people were getting mad at you for all kinds of things. And it just was like, it's just difficult. I mean, Elon, quite frankly, deals with a lot of these issues in his business today. He has a large number of manufacturing workers in the US, and a lot of people are really mad at him about it.

In a sort of, for these kind. Of perverse reasons, he probably hired more new manufacturing jobs than anybody. Yeah, exactly. And he just. He gets no end of pain because it's like, is it unionized?

Is it not unionized? Is it? This is it. That is it. You know, isn't the politics.

Ben Horowitz
We need more manufacturing jobs. Elon's like, okay, not only create manufacturing jobs, I'm going to give them all, like, stock and Texas line. A lot of them are going to become millionaires. Screw you, Elon. Yeah, exactly.

Marc Andreessen
Yeah. Basically, f you get out of California. So it's just like, it's a lot to, like, have a manufacturing company that is based on basically human manufacturing labor, traditional blue collar labor. Like, it's just a lot to do that in the US. And it's just been more cost effective for 30 years for manufacturing companies to do that offshore.

Those jobs are never coming back. If there's not a step function change in technology and in the ability to use technology to get leverage and to be able to transform those jobs and transform from the economics. Like, they're just, they're never coming back. It doesn't matter, by the way. It doesn't matter.

Trade policy, whatever tariffs, like, whatever, whatever. It doesn't matter. Those jobs are coming back. And so the only way that those jobs come back is if you basically close your eyes and imagine, actually, I'll use an Elon phrase, the alien dreadnought. Right?

Fully automated, fully, like, AI enabled production facility factory that you just go in and it's just like a marvel of technology. And, like, it's just got, like, incredibly smart robots running around doing things, and it's like, got all AI everywhere, and, like, everything is instrumented and being run through software, and it's just like, you know, it's just amazing. You know, the control system for, you know, that factory is going to look like somebody playing a video game because it's going to be just like, it's going to be just like this amazing marvel of sort of fused technology and then real world actions. And so, like, that is how you could get these things back, by the way. It is.

It is. The US is in many ways the best place to build that kind of factory because the US is the technology leader of the world. And so we have the best engineers here building all the enabling technology for that, including all these great AI companies. And they, you know, they're going to do their best, you know, working, working here. And so just because of proximity and so, like that, you know, like, that's the only way it's going to happen, by the way.

Like, you know, those jobs would not have as many assembly line jobs, but they would have many, many, many, many jobs around that factory, right? And in fact, in it. And so, like, you know, building all these, you know, building these factories, running these factories, fixing everything when it breaks, you know, optimizing everything, improving everything, upgrading everything, like, and then all, and by the way, all the downstream things, like, if you have, you know, if you, if you have, like, an alien dreadnought advanced AI manufacturing facility and, you know, Alabama or something, you know, you're going to have tons of jobs that are going to be around that, right? And you're going to have, and this goes to the thing I like factory towns. Factory towns that, you know, that whole town that is going to have an ecosystem of service providers, everything from restaurants, hotels, everything else that comes with that, to be able to service that.

And so I actually think there is a path. There is a path here towards revitalizing and rebuilding the american manufacturing sector. The sort of geopolitics of our world kind of point us in that direction because for a variety of reasons, like, we've discovered that, like, it would be a good idea to make, you know, more stuff around China. Yeah, for example. Yeah, exactly.

Like, not. Yeah, right. Or by the way, did not, you know, earlier conversation of Boeing to not have american, quote unquote, manufacturing companies actually just be supply chain integration companies that they may actually need to get good at building? You know, that the world's leading airline company in ten years used to be good at building planes. And by the way, maybe that's a new company, but maybe those planes are made in the US in this new method, and they're made in a way that is better from a cost standpoint and from a quality standpoint and from a speed of improvement standpoint and, you know, having integrated engineering.

So you have engineers on site at the factory constantly making everything better as opposed to just having the factory off in some remote location that your engineers never go. And so I think there's a path here you might almost call this sort of like an operation warp speed for manufacturing, where you just basically lean hard into this and you say, look, we're going to be, America will once again be the number one manufacturing company in the world, but not because we are opposed to automation or robotics or AI, but precisely because we're going to embrace all of this new technology as hard as we possibly can. Yeah, I think that's right. I also think we underestimate how good robotics might be for job growth in the sense that robots are like, where do jobs come from? They come from companies.

Ben Horowitz
Where do companies come from? They come from entrepreneurs. For entrepreneurs to have a whole new tool set like, oh, I've got robots that can do things and that can create new products and services is a huge boon. And, like, you know, I think possibly a boon for things outside of Silicon Valley where, like, look, if software is the tool and that's the tool you've got and that's it, then that's a bit limiting to the, you know, you have to be able to have those kinds of people to build those kinds of companies. We know that better than anyone.

But, like, what we're seeing with hardware manufacturing and robotics and so forth is like the, those companies are much more spread out, you know, than the software clusters that we see. So I think, you know, generally it could be very good for the country and very good for jobs, provided we don't outlaw it all before it happens, which certainly a possibility. Well, that's the current plan. The current plan is everything I just described as illegal. So we would need a national campaign.

Marc Andreessen
We would need a major, major initiative on the part of both government and industry to do this. But I do think it is a possibility. It's more. I was supposed to say it's more a possibility now that it's been for at least 30 years. And so it is worth pulling on that thread, and maybe we can spend more time on that down the road.

Ben Horowitz
Aperture is reopening and really, I think, exciting way because it's always much better when you have kind of innovation that can be done, you know, by a broad swath of skill sets and people than just like a very narrow base. And I think we're kind of getting back to that through AI and robotics. Yeah, that's right. Okay, good. I think that's a great note to end on.

Marc Andreessen
Benjamin, thank you again. Yeah, that was fun. Thank you, everyone. Great to see you.

Ben Horowitz
Great to see you.

AI, Robotics & the Future of Manufacturing

Primary Topic

Episode Summary

Main Takeaways

Episode Chapters

1: Introduction

2: The Role of Leadership

3: Corporate Decision-Making

4: Conclusion

Actionable Advice

About This Episode

People

Companies

Books

Guest Name(s):

Content Warnings:

Transcript