The Grittiest Conversations of 2025: AI, Business & Beyond

0:09Today's episode is a bit different. We've curated highlights from some of our favorite conversations this year with founders, operators, and builders across finance, software, AI, and media. From the future of work and AI to the changing economics of the web and culture, these clips capture how the world is being reshaped in real time. Let's dive in. First up is David Rubenstein, co-founder of Carlyle, on why he has stayed out of partisan politics and how remaining neutral has allowed him to work across administrations and bring people from both parties together. I stay out of politics. I don't give money to politicians.

0:45No money to anybody. I've never given a penny, never one penny to anybody running for president of the United States. Why? Well, the reasons are this. One, the way our system works is if I give$100 to somebody running for president, every mistake that person makes, they're going to be blamed on me. If this person comes out with a position that is antithetical to my own views, they're going to say, well, you believe in the big green deal. I had nothing to do with the big green deal. I just, have to give this person$100. So that's one problem. Secondly, as Carlisle became bigger, if I gave$1 ,000 to somebody's elected president, and that president has a defense department, and they give a contract to some company that Carlisle owns, the headline could be, Rubenstein gets a contract because he gave$1 ,000 to Joe Blow or something.

1:33I don't want to read that. Now, obviously,$1 ,000 is probably not the same as giving$1 million or$10 million. But rather than have anybody even think that I'm trying to buy access or I'm politically getting things, I just don't give any money to politicians. Well, also being political automatically ostracizes 50 % of constituents that could potentially want to fundraise with you. Like Michael Jordan famously was like, you know. Republicans buy shoes too. Exactly. But for example, I, for the last 11 or 12 years, I've hosted a program at the Library of Congress. I am the chair of the Library of Congress support board group called the Madison Council.

2:10And I chair the National Book Festival. And I'm pretty involved with the reading world and books. And I started this program to educate members of Congress about American history a bit. So once a month, I host a dinner at the Library of Congress with the Librarian of Congress, where I will interview Doris Kearns, Goodwin, or the equivalent, about American history, one of their new books. And so I will host and pay for the dinner and the reception. Members of Congress come, Democrats and Republicans. We ask them to sit with people from the opposite party, which they rarely get to do in Washington, and sit with people from the opposite house, which they rarely get to do anymore either.

2:50And if I was an ardent, ardent Democrat or an ardent, ardent Republican, I'm not sure I could get people in the opposite party to show up at that kind of event. So by being right down the middle, I think I can get more people to show up and maybe educate people better. I've been the chairman until President Trump fired me. I was the chairman for 14 years of the Kennedy Center. I was the chairman for four years of the Smithsonian. And I'm chairman now of the National Gallery of Art and the Library of Congress Board. So all of these things I think I can do more effectively by not being involved in politics.

3:21Did Trump actually fire you? I read that headline, but I was going to ask you if that's true. 14 years. Yes. I was put on the board by George W. Bush. and President Obama reappointed me and then President Biden reappointed me. And I know President Trump reasonably well and I've talked to him about many different things over the years and I think he likes me for whatever that's worth. He was not happy about something at the Kennedy Center. Somebody told him there had been a drag show there and he was upset about that. I think some people told him that the Kennedy Center honors weren't going to people that were politically conservative of, whatever the reason, he decided to make a change.

4:04And he wasn't happy. I think that President Biden had put some people in the White House staff on the committee board right before he left office. Whatever the reasons, I don't really know. He decided to change the board. He told some people he thought that I'd already retired because I had announced I was retiring in December and it was a big farewell party for me. But the board couldn't find somebody to replace me. so they asked me to stay a couple more years. So I was still the chair, but he says to some people that he thought I'd already retired. Whatever the case, he replaced me. So in the long history - With himself.

4:40Yes. In the long history of the United States government, 250 years we're about to celebrate, I'm the first person ever to be fired by a president and succeeded by him. Next, we have Yamini Rangan of HubSpot on cutting through AI hype and grounding the moment in real customer value and outcomes. The jobs that we all did as a marketer, as a salesperson, as a support person is now completely getting rewritten and we can do more. And that is what is so exciting about this. While there are some parallels in terms of the ethos of like how crazy people are growing and all the conversations that are happening.

5:18this is fundamentally exciting because you can do better things with your time and you can re-imagine amount of work that can be done. And it is no longer about software helping you do tasks. It is about software doing tasks to help you that then maximizes the results and the outcomes that you can achieve in the time that you spend. I totally agree. I mean, when I was starting out my career, like I was probably making 100 cold calls a day and writing literally four to 500 emails a day and copy and pasting them. And we're investors in a company called Nooks. It's just going to do it all. It just does it all.

5:59It literally just does all of that work. I didn't need to be doing that work. What a waste of time. I still needed a job. There was a job for me to be done there, but it was not that job. Well, and think about this, the job to be done for you as that sales rep, and this is how I started my career as well, is to close deals and is to win business, right? Is to be in front of customers, is to understand your customer's pain point and to win them as a customer. That was the job to be done. In order to do that job, you researched companies, you called, cold called, you wrote a bunch of emails, you followed everything up.

6:37And then you did all of that to get a sliver of time in front of the customer. And that was like the paradigm then. And look, I've been in sales and go to market for a very long time. You know, the really incredible statistic for 20 years, I've been tracking this, which is what's the time a salesperson spends in front of a customer? And this has not moved. It's like 25%. And if you got your act together and you have great operations, you have amazing data, you have great systems, it's 27%. So you literally have like a 2 % improvement if you get everything right in terms of the time in front of customer.

7:17And everything else is the prep that you're talking about. It's the research. And then, you know, when I first moved on to operations, I would like literally beg the sales team to be like, can you please complete your notes in the system? because the next person, the next person needs a better handoff. So that is changing. And it's so cool. It's so exciting to see what, you know, companies are doing. You know, HubSpot is completely there. And that's what we're trying to imagine for our customers. Like, I agree. I think this is a real thing. Yeah. Okay. Yeah. But like at the time of crypto. Yes.

7:55Or at the time of COVID. Yes. Everybody also thought it was a very real thing. And in the days of everyone being remote and it's going to last forever and this digital demand is going to sustain for decades to come. The problem is, which is similar to now, whether it's true or false, if you're wrong, if it's true and you're wrong and you didn't make the right investments at that point, you're screwed. Yes. Meaning like if you're growing 100 % at a billion dollar base and you're just like, oh, no, this isn't going to last. And you're like, and it does last. Yes. You just left like hundreds of millions of dollars.

8:35Exactly. On the table. Exactly. Right. Or like even these, these, these, these big tech companies now like Meta and, and, and Google and, and, and Microsoft, the investments that they're making in infrastructure are insane. Like insane. It's insane. Like we've never seen dollars spent like this. Exactly. But like, what are they going to do? Like, what if they're wrong? What if like, what if this is maybe not even enough investment? Yeah. And Microsoft decided not to do it. And Meta and Amazon did. Yeah. Like they missed the whole thing. Absolutely. Absolutely. And so. This is why it's hard. This is so hard.

9:12So I think like, you know, one of the things I feel is there's a lot of hot takes and hype and all of these. And every cycle, whether it turns out to be real or not, has like hot takes and every day, like some crypto this and Web3 that and now AI this, right? It really comes down to customer value. And you got to go back to, you know, it doesn't matter what all this technology talk does. You got to go back to who's your customer? How, what is their job to be done? And can you improve that and add value? That's our, you know, that's how we look at it. We have like a North Star, which is all for the customer, SFTC.

9:53We talk about this all the time. And then it comes back to great, forget all the hype and forget all the cool technology. Can we take a neat feature of AI and make it a necessary feature in AI? If we can prove to ourselves with customer usage and repeat value that we deliver, then there's something in the cycle. If it's not, if it's the cool thing that you do today and then next week it's like churned and everybody is left, then there is no value. So I think you got to like ground yourself in customer value and truth. And if you do that, then you can sift through it. But, you know, at the beginning of the cycle, you just have to like, you know, use some kind of pattern recognition to say, is this going to work or not?

10:37And make a bet. And you cannot be wrong. And, you know, that's what happens. Like I would say that, you know, SAP, I used to work at SAP. And, you know, back in those days, there was a lot of question of, is cloud real? You know, will on-premise ever go away and is cloud real? And would, you know, CIOs ever trust data that's not in their data center and is in some kind of public cloud? These were real debates there. Now it's not the same thing. I think all of us, you know, who've been established for a while, or if it's a startup, we know and see the value of AI. And we're kind of like really focused on what is that value for customers.

11:16Next is Ben Chestnut of MailChimp, reflecting on how fast technology cycles have sped up and why he's relieved he's no longer running a tech company in the middle of the AI wave. I'm so glad I'm not in business. Are you? Yeah. That's very atypical from what I hear. Really? Because AI is disrupting everything. And then as soon as there's an AI disrupting something, it gets disrupted. And you know how, so early on in MailChimp, I realized in the tech industry, we needed to reinvent ourselves every three years. Customers' tastes change and their technology needs changed every three years. I had to keep MailChimp very nimble.

11:57I would never commit to anything that would take longer than, let's say, one year because then I couldn't be able to turn on a dime. So, you know, that worked for like the first 10 years. And then like you're at year 15, it's like, it's accelerating. I can feel it accelerating and my company's getting bigger and harder to turn. I have to start making bets. And these are not big bets. These are just, you know, medium-sized bets for what I can afford. And then now with it accelerating even faster, there's, I don't see how you could, how you can keep up with that. And this is coming from a guy, we invested in AI really early, very early.

12:37Big data, all of that stuff. Using NVIDIA like 2008. We were very early in that world because we were processing abuse data, email data, spam data. We needed to protect our server. So we were doing it for abuse prevention. And then we got into email design, copywriting with AI. and then I sold. And now I look back and just say, oh my God, I'm so glad I'm not a CEO of a tech company anymore. At Harvey, Winston Weinberg is focused on the hardest problems in legal, building AI that can survive years of model advances by earning trust first. How many companies are there that are going after this AI for legal?

13:22I think I saw a market map a long time ago and I had to zoom in like a lot. Yeah, you guys are just, But, you know, I mean, I'll let Winston opine on this one. But, you know, the markets, it's not just like there's going to be one legal tech company. It's not like one market necessarily, right? There is segments of that market that are really important, impactful, and a good place to start. And I think, you know, where Harvey is, which is big top law firms, big enterprise in-house legal teams, that's like the juicy, the right entry point. And that's a pretty interesting one because, again, it's very collaborative.

14:00And so I'm sure there's going to be, you know, AI for legal in a variety of flavors, obviously for lots of parts of the market. But to me, at least, this is kind of the biggest, meatiest one because it really drives everything else. There's trust, there's brand, there's reference customers. That's like the place. I think like one thing that I try to think about a lot is just like, what can the models do in like 10 years? Right. And how do you build a company around that? Right. And, you know, the term mode, et cetera, all of these things. But I think like one interesting piece about the kind of like, let's go back to like the Microsoft Activision merger, right?

14:42This is work that is at the complete tail end of the models can just automate this. In other words, if we're in a world where GPTX whatever can just automate that, most companies are gone. Like, right? Like, we'll just be honest, right? And so I think that what's really interesting about going after that type of work is it is incredibly messy, is incredibly context specific. The accuracy of it really, really matters. And so we are trying to basically put all of our effort into that because we think that that is not something that is going to be solved by the next generation of models. I would say that when I think about competitors, I think that the largest competitor for us is just not moving fast enough, by far.

15:32I think that indirectly, you want to make sure that you are building a business that is going to survive the next 10 years of model releases. And that's how I think about the product. That's how I think about GTM. That's how I think about basically everything. Is that the correct play? I think it will be in a while. But we've had a lot of revenue growth. To be honest, we'd have much more if we made it self-serve and we kind of did a different type of route. I think that that stuff ends up getting kind of cannibalized by the models. When Glean was growing in its early days, there was so much pressure on that team to self-serve it, just give it away, right?

16:15Yeah. and they did it the hard way similar to how you're doing it tops down enterprise get deep into the organization and um that's paying off really well for them yeah and it's it's interesting too because we have a very like land and expand motion too and so we have this weird like we have plg it's just you have to get into the organization for the plg to start and you know i think there's a lot of advantages to that. And I think like the nice thing too, and we talked about this kind of like early on, is we have the option because we have Fortune 500 and you talked about these like three parties collaborating, you have internal virality.

16:55In other words, like at a law firm, you have different lawyers, you know, spreading it to each other across different groups. In an enterprise, you know, we'll sell something to the legal team and then they'll tell the tax team about it as well, right? And so you have virality there internally. And then you have external virality of if you have a bunch of law firms using our product, they tell their clients. Remember I said that they mostly were referrals for the enterprise customers that we had in the beginning and vice versa, where the enterprises will tell their law firms like, hey, we're doing a deal.

17:25Are you using Harvey? We are, right? And so I think that's another way to kind of go about expansion. It's just slower in the beginning, but I think you can still do PLG really fast. It's just a different kind. I think the other part here is really trust at the end of the day. Because if you sort of release something into the ether where people can self-serve, it's bottoms up adopted, whether it's in a big company or a small company, but it's not hardened yet and it's not fully trusted yet and you haven't been sitting next to the customer using it, you're potentially going to miss out. Because if you get something wrong, especially in something as critical as legal, that's kind of a third rail.

18:03And so I think that the approach here is really smart because you're going to build something that is hardened enterprise grade, works super well, and then you can kind of let it spread on its own over time. And here and with Glean, I think that's the right. Yeah. Garrett Lord explains why the next gains in AI come from expert human data and how frontier models rely on domain specialists to train, validate and correct complex reasoning. So, for example, OpenAI or whatever one of these labs would come to you and their model is spitting out some algorithm, something around some scientific discovery.

18:43And they need somebody to look at that and basically train it and say, is this right? Is this wrong? They need somebody with a specific set of expertise to basically look at whatever that model is producing within a given domain and then ask the expert. Yeah, one thing is a great, very simple story, which is like, you know, people talk about agents all the time, right? Like agents requires an ability to follow step-by-step instructions or, you know, fundamental idea behind agents. You understand like a reasoning and a reasoning choice need to be correct. So if you ask any one of the large frontier models questions and you ask it to break down its thinking of its step-by-step instructions, too oftentimes these models are not getting the answers correct.

19:26So when you think about a model, it's like you have compute, you have algorithms, and you have like data. Those are like the three things that matter with models. Everyone's applying a ton of compute. Everyone's constantly evolving algorithms. And people are obviously chasing after one another in terms of algorithmic prowess. And you have data. And data is what really fuels the gains on what are called the post-training side of the house. Post-training is human data. And what we do is, you know, the world's evolved from like generalists, where you were predominantly leveraging like international skilled, lower cost labor, taking advantage of an arbitrage between that to now these models have kind of sucked up the entirety of the entire corpus of the internet and every book and video.

20:08They've also sucked up and they've gotten good enough where like generalists are no longer needed. The miles are actually better than generalists. And so what they need is they need experts. They need experts in accounting and law and medicine. They need experts in coding. They need experts in STEM domains like physics, math, chemistry. And designing a network that's focused on experts and treating those experts well and making it an amazing community experience with high retention where people can, you know, we're connecting to like badges on your profile and leaderboards by school. You know, we have access to millions of experts.

20:42So these experts, it's pretty cool. You can make like 60, 80, 100. You know, I think our average bill rate's over$100 an hour, like, you know,$125 an hour applying your expertise in the domain that you're studying. You know, you're getting a master's in accounting and you're getting paid$100 an hour to provide data and correct models. And yeah, we're doing really also wild things like, you know, audio, like a ton of audio work, like these models. I can't get, I won't get too far into it. But, you know, in the areas of trajectories and, you know, tool use and text data and coding data, like Handshake has, in our opinion, like one of the largest networks of experts that are captive and trust Handshake.

21:27And so we're trying to build a great experience for them to produce data for the labs. And can you just walk me through like a how like, OK, they find an expert on math or science. What is that expert actually doing? An average frontier lab of which we work with six of them, most of the frontier labs, let's pick something like mathematics. So they all care about and let's merge mathematics with educational design. So an analogous type of project that we would do for a lab is actually explaining and breaking down the decisions around how you would teach a particular math problem to a student. So you'd be applying mathematics experts who would actually agree on the step-by-step, like here's how I would solve the problem.

22:14And you'd merge that with educational design PhD around whether you like option A or option B more. And you might even get more specific, like you might provide like hints or you might provide data around students in the Southeast prefer this educational design pattern, whereas students in the Northwest prefer this educational design pattern. So it's a public paper. So you go in, if you're a physics PhD, you break the model. You identify a reasoning error in a step-by-step instruction flow. You correct the reasoning flow. And then you provide the ground truth right answer. And then they do things like, you know, pass at K1, pass at K3, pass at K10, which is like, does the model, because they're non-deterministic, does the model get it right one out of three times, you know, three out of three times, get right one out of 10 times.

23:00So we'll actually go in and provide distributions of data around how many times does the model passing this question, here's what the right answer is. And that would be an example of like a physics PhD focusing on like their area of research, asking, you know, questions the model cannot do today and providing the actual right answers. Most of the data we produce, if you were to read it, like, you know, you could lock me in jail for like 30 days and ask me to solve one of these physics problems. Like, you can't even like, I can't even barely read the type of data that we create because, you know, it's really experts in PhDs.

23:39Next, Aidan Gomez reflects on co-authoring the Transformer paper and why an open, scalable architecture sparked the wave of AI progress that followed. Listed as one of the co-authors of what has now become maybe one of the most consequential papers in our industry. Totally, yeah. Obviously, I don't think you had any idea then. I don't think Google had any idea then. Otherwise, they probably wouldn't have given it away. Yeah, yeah. Tell me the next time. Well, I mean, would it have been big if Google didn't give it away? Would it have blown up like this, right? You're saying, would Google have known what to, like, would Google have created the chat GPT moment?

24:22Yeah, yeah, exactly. Or would have the, you know, academic community created a different architecture under a different name that looked very similar and felt very similar that carried the weight instead. And so maybe Google had its proprietary transformer model, which nobody could use because it was IP and, you know, didn't become very popular. But the rest of the world created something slightly different, which, you know, had enough of the same properties to drive all the stuff that we've seen over the past 10 years. I think that's probably more likely because at the time there were like the transformer was built off of a set of ideas like uh what was it called uh byte net uh wave net um whatever it was like seek to seek there were these um series of papers which had these ideas of auto-aggressive models which were much more scalable for training and the ideas were out there in the ether and we just sort of pulled them together in a particular arrangement.

25:27I think if we hadn't have done it, someone else would have within the next 12 to 18 months. So if Google didn't tell anyone about it, kept it secret, I think someone outside would have created something very similar under a different name. Why do you think that? Because all those ideas were like in the ether. It was just building towards that. Yeah, totally. Totally. Like it was necessary. Like something like the Transformer had to be tried by someone. It's just that, you know, we were the ones that tried it first. And we got enough eyes on it to catalyze this snowball. And once it started rolling, once people started...

26:09This is once you wrote the paper. Yeah, once we released it, published it at NeurIPS, it got enough eyes and momentum that the community figured the rest out, right? Like we sort of like started the seed, like, you know, packed a little snowball and then gave it a little nudge down a hill. And then the community just poured in and started implementing all the different frameworks, you know, testing this thing on all the different benchmarks. And the community carried the work. It was less like us at Google pushing this thing forward, showing the world what it could do. It was everybody else taking this little seed and growing it into what the Transformer is today.

26:50And in your words, like, what is the core unique insight in that paper that propelled so much of this forward? A focus on efficiency. So it was a very simple architecture, like very minimal. And it was extremely well suited to scaling it up across many GPUs. at the time we built it like many gpus was tens of gpus right like that was like and what's many gpus today oh tens of thousands you know maybe hundreds of thousands um in a couple years maybe millions but yeah so like the the core insight was hey we're gonna need to scale this thing up and for us that meant instead of training on one gpu we would like to train on 32 um and so let's make sure we build a training framework and an architecture that's compatible with that.