Why Anthropic is watermarking every word Claude writes | E26

31:49And it's very hard to post something that would actually hurt your business and your tribe, which is ultimately what business leaders care about if that vision came to light. So that's the healthy skepticism. I happen to completely agree with everything he wrote. So I really like the same as Josh. And by the way, I think he completely believes it as well. So I think in this case, what Meta is building in AI is actually very aligned with what the vast majority of people want and what's good for the world, right? They're like, the fundamental part of this piece is like centralization versus decentralization.

32:26And centralization financially benefits a couple companies, whereas decentralization benefits millions of companies, right? And humanity. Yeah. Well, okay. Well, that depends. I mean, there's a counter argument, which I agree with you, Jason. But like I think the counter argument is a one of this is like optimism. And then the counter argument is like one of fear. It's like, oh, no, if it is decentralized, then people will create cyber weapons and bioweapons and it's uncontrollable and we need some centralized power. And there's an argument there. I disagree with it. I think. Why do you disagree with it?

32:57Explain why you don't need to worry about the cyber hacking, the biological weapons being created with this, because he does address both of those. He has his own thesis. But let's unpack that piece. Well, your opinion matters to you. Yeah, if you jump into the details, okay, play it out. How does it actually work if there's a cyber, like some cyber threat or biological threat from decentralized models? Like, can we handle it? And if you play out those details and you could just, oh, anyone can make like Ebola that spreads at the rate of COVID like in their, you know, microwave, whatever it is, like nuclear weapon in a microwave type of analogy, and then actually just destroy the world, that's very bad.

33:38I don't think that's the case. I think the biological weapons are are limited by physical resources. So it's actually really hard for someone to just like really just like have a humongous effect on the world. Just just, you know, in their in their house. I think it's still possible. But I think that the positives outweigh the negatives there. And there's always a risk of anything. Right. And so on the biological weapons side, I think it's mostly protected. And yeah, one of his arguments is that if people have all these different defenders, it could actually help. and on the cyber side i think it's it's even a less of a risk in the sense like yes there will be huge damages there's some mass like um spreading of uh some some virus some intelligent virus but that that one feels like the kind of thing that we will learn from handle it and be better off because of it yes i like this like this hugging face attack scary but amazing that it happened right like there was no no one was hurt and we learned a ton from it and now we're going to prevent against swarms this is how humanity learns to prevent disasters we unfortunately that we have to make the mistake.

34:35And then - Yes, Fukushima was a huge mistake. They built that seawall too low, in the wrong place. They were told, this is in the wrong place. You're going to get a hundred year storm and this is going to be under salt water. And they said, YOLO, we don't think so. And 50 years later, literally almost like clockwork, the tsunami hits. And we're left here like - why are we so stupid as a human species? And he points out two really important things, Will. Well, if you really care about the biologics, like, well, why don't you just not let people buy the source materials and regulate that a bit more?

35:17And then on cyber, which is just the general term for cyber hacking, for cyber warfare, for hacking, he's like, well, the codes are going to be two things. He doesn't say the code's going to be written by AI and be perfect, I'm adding that piece, but he kind of insinuates that. That software is going to be kind of perfect because it's all going to be vetted this year. If it keeps going at this pace, what software is not going to be vetted? You'd have to be an idiot to not have your software vetted by these latest models, even if it's cobalt. And if the majority of software is going to be written by these systems, hey, it's going to be pretty strong from the beginning.

35:56Yeah, Will? Yeah. I mean, I agree. I think you would have to, to be against this, you have to make the case that like, no, like the one event is going to kill us all. Like it's like the first biological, the first time. And so like, I think the analogy of like, can you, if you can make a nuclear weapon in your microwaves, would we regulate that? And there's an argument that you would, because then one person can actually do so much damage. But I think that the AI, like local open source models, at least right now are very far away from being as the blast radius being as significant as a nuclear weapon.

36:22So, and way less than that. Right. So, so like, you know, I think that the positives of decentralization are way overpowering the positive centralization, at least right now. Maybe if the models get way better, they might change, but at least right now. And interestingly, Will, if someone can cook up next Ebola in their microwave, if everyone has access to the same tools, can someone also cook up the vaccine cure in their microwave if AI is decentralized? so I completely agree with you on that part that the open source and decentralization of AI is very positive, net positive for the world and more people having access to these tools are going to increase the capacity I guess, and just playing devil's advocate the only action that remains is a rogue state that has way more resources in the physical world also.

37:26Yeah, but Kash, aren't they all been trying to build bad things in the world from Al-Qaeda to North Korea to Iran? Like, they've been at this. Already, yep. Luckily, terrorists and evil people tend to be stupid and bombastic and they telegraph what they're going to do. and Israeli, US, UK, 007, all of our intelligence agencies bring the best and brightest. So to your point about the white hat, black hat, the black hats in this regard tend to be dumb and they signal what they're going to do and they're easily identifiable. And with LLMs, you can be certain if somebody's typing in, how do I make a fertilizer bomb?

38:09The first thing Claude is doing, Dario has like a bunch of speed dials And he's watching everybody trying to do something bad. And he's got like this war room that he sits in. He's got like a red phone that rings. Totally. He's like, oh my God, there's a 14-year-old that just said, how do I make the Oklahoma city bomb? And he just calls, you know, Emile Michael or who's the guy in charge of the Department of War now who does a lot of pushups and drinks. A lot of beer. Yeah. He's like, Pete, Pete, come regulate me. There's a 14-year-old who just said, how do I make a bomb? Let's knock his door down.

38:42It's a radar-dade 14-year-old doing what 14-year-olds do, which is look for boobies and type in things that send off alarms on the internet. Like, come on. Maybe he asked Andy Jesse to make the call. He doesn't even make the call by himself either. He does a conference call. He's like, you know when you get the FaceTime call from the group chat and somebody asked, dialed the start of Zoom or whatever. is like, does this guy even know he ass-styled us all? That's Dario. He just has like Trump, Besant, everybody. Pete, they all get on the same call. We did something bad. Jason, I think a huge important piece, just back to the Zuck, like that we went over.

39:29I think one of the things that stands out the most is in the ensuring American leadership section. And so what he says is a very clear stance against what some of the labs are taking. So he says, and this is a direct quote, which is, I propose a collaboration model based on sharing intermediate training checkpoints and technical staff. Given how dynamic and competitive the AI landscape is, I think it will be more productive to have a close, proactive collaboration between frontier labs and government rather than a rigid process and review timeline. And so this is very clear against some of the labs, because what some of the labs have proposed is, hey, we need a government committee that basically we release the models to.

40:18They have 30 days to review these models and then they can approve or reject it and then we can release it. And what he's saying is absolutely not. Right. What we need is just close collaboration. We do not need these committees, et cetera. And then I think the second point, I mean, you know, Jason has been a beacon for open models. models like for i think for longer than almost anyone at this point it's been two we pulled a clip from all in from two years ago where i said listen i think open source remember that clip long that you found yeah it was like episode 50 or 100 i was like you know and i think open source might just run away with this and they were all the other buzzies were like crazy yeah it was from like summer 20 you're an idiot yeah remember jake how was right about everything it's honestly it's It's been Jason and then Mark Andreessen.

41:03I think Mark Andreessen for two years, he's been saying, I don't think that the open race is over by any means whatsoever. And Jason, these have been the two voices on the side. And so now Zuck is coming in under this American section. And he's saying that foreign labs have a whole bunch of competitive advantages against us American labs. namely with all these regulations that American labs are some are proposing have to be complying with. But he says, quote, for the U.S. to lead an open source, we need to rethink our policies in several areas, including distillation, data use and training. So I think this is a really big point.

41:49Curious to hear what you guys think. It is a big point. And it also I it's so interesting, Josh in that 6 ,500 words. That was the second thing I teed off of, which was him saying, he says, if you can observe it, Lon, see if you can find this. He says like, if you can observe it, then you can learn from it. That was his way of saying, I can steal any data I want to be competitive, but he embedded it and he framed it as competition versus China, which is like the best straw man ever. It's like, I get to steal because they, the Chinese steal. It would actually be better if you licensed the stuff and just paid the New York Times for their content.

42:25That would be reasonable. Just because China steals the New York Times doesn't mean you have to. Just because they steal the Lord of the Rings doesn't mean that you get to steal it in the West. I thought it was like a really important point. Could you find what he said? Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. This is how the world works and the U.S. will not be able to lead if we restrict ourselves on this front. Okay, so I can observe all the accounts on Instagram and make a database of every Instagram account.

42:55Not like that. Not like that. No, no. You can't distill Instagram or the social graph from Facebook. He stopped that and he sues every startup. I know because I've invested in a couple of them. He sues every single startup. This is the hypocrisy of Zuckerberg. Now I'm getting fired up again. If Zuck really feels that way, let us distill the social graph. The number of times Zuckerberg has sued Israeli, specifically companies, and companies in India and the Philippines, those three. He keeps trying to sue them for creating databases of influencers. And I know this because we had a startup, I won't say which one, that had a heck of a time with the threats from Facebook, which was their partner.

43:44and then they kept threatening them like, oh, well, we're going to pull all your access and we're going to sue you into oblivion because you have a database of influencers. Like, well, F off. I can distill your database. Give me a break. Jason, a question for you on the same mindset. If they were one of the leading labs and they were part of the duopoly, let's just say, will this be Zuck's take? This is a great point. What you're pointing out, Cash, is a truism we've had in Silicon Valley for over two decades since open source actually worked. When open source first started to hit the consciousness of enterprises in the early 2000s, that's when MySQL, Hadoop, and a bunch of these databases started to go from, that's farcical to think that an enterprise would use them, to enterprises saying, well, we don't want to be locked into some big tech service that they can rug pull, just like people don't want to be locked into Fable, which was actually rug pulled multiple times.

44:52That's why enterprises eventually said, hey, this is cheaper, better, and more stable. Think about that. Open source, controlled by nobody, was better and more stable in the mind of enterprises. So we came up with this truism. When you're behind, go open. When you're ahead, be closed. So he is behind in terms of models. So he's for open distillation. He's ahead on the social graph. So he's suing everybody into oblivion when they try to distill Instagram. I'm going to ask that somebody come up with a startup that distills Instagram and Facebook. And I would love to back that or somebody come up with an open source product that does that.

45:41Wouldn't that be an incredible thing, Josh, if you could go to GitHub and there was an open source project, I would put a bounty out for this that just said, here's the social graph as we've read it from a bunch of browser, you know, extensions or, you know, whatever that scraped it. Oh my God. I mean, imagine if you could really like search and and fine tune your searches of who you were looking for in terms of influencers across Instagram and Tik TOK. I mean, you had, you were locked into their algorithmic system. That would be amazing. You know, we'll on, uh, on, on the call, like if, if X had the social graph, right.

46:19Of, uh, if it was open source. Um, so I, that would be curious to hear what your perspective is there, but just quick remark on my side, I think, yeah, I mean, I think there's obviously the Catch-22 from his perspective, where if you're going to take this position, then hopefully the actions will reflect that this is the position. But overall, his point on American open models, I think, is important. And hopefully his actions in broader perspectives will reflect this. Because this is new. This is a new race for America. We have Mira's company that's firing, you know, seems to be starting to start firing now.

47:00We've got a new company that was just funded River. You know, hopefully it's a good thing for humanity. If Zuck, if they can figure this out, if they can come out with an incredible open model, this is going to be a fantastic thing. Right. For the world, actually. So hopefully, you know, his points on openness and American, you know, open source, you know, I'm rooting for them. I hope that they can come out with something amazing. It'll be good for democracy. So it'll be good for humanity. And then to your point, Jason and Lon. Yeah. Right. So then fine. If you believe in open source, you know, help the community and right.

47:35And open different assets that other people can use. Right. To bolster some of their models. Right. So curious, Will, like your perspective on that. Yeah. I mean, everyone wants to think that they don't care about to be open. It's exactly what Jason said. Like every at least like it explains everyone's position. It's like we have people in the lead want to be. Yeah, I mean, of course, like, we really care about giving people access to information. And so like, it's almost like we're kind of enabling all the open source models to have a fantastic search engine, right? Like every open source model, there's no open source search engine.

48:08It doesn't make sense. You could run a model like pretty cheaply on your laptop, but you can't run a search engine. So you kind of need to use some search engine. We want to enable all these open source models wherever you are, whether it's in a company or on your local laptop to like have access to whatever information you need. But yeah, I mean, within the limits of what companies are willing to make publicly available. I think the balance is correct. Like people, like maybe there's an interesting thing where like between like access to data and then distillation, they are different. Like you know, lots of companies have all sorts of data that they rightfully so don't want to be out there.

48:47But then if you do, if you have a model, like, can you, are you, are you allowed to, so basically if you have access to data, you know, you should be able to control whether that data is available to others. But if, but if you're a company that has access to a model, should you be able to control the way people use the end results of the model? I think it actually is kind of different. I could, I could see a world where like data is more controlled, but model outputs should be less controlled. This is a great episode because it got me inspired to like, because, you know, if you want to put out your manifesto and there's items in there that are incongruous with your past behavior, Mark, somebody clip this and send it to Mark Zuckerberg at his email.

49:30And I would sincerely like a response from him. And so I just tweeted, serious question, Fink D. If you're pro-distillation, does that mean that Facebook and Instagram will allow scraping of the social graph now, yes or no? And here's a link to him saying, we're going to sue you into oblivion, which he's done over and over again. I think that's an important question. You know where I stand on this? If you're breaking the terms of service of LinkedIn and of these social graphs, it's their right to their data. They should be allowed to get a licensing fee. So if Claude or OpenAI want to use this hard-earned data from Instagram, they should get permission, just like they should get permission from Reddit or the New York Times.

50:13So this is like a very important moment because he has legal cases where what he said in the manifesto works against his legal cases. The same thing with Dario and Claude saying, oh my God, you're distilling, you know, fable, you're distilling Claude and then running to, you know, daddy president Trump and saying, protect us and give us some, you know, protection. They had that database of like hundreds of thousands of pirated books. And they, you know, that they they were scanning like that's OK. But taking their models and studying them for your own model is not OK. It is it is a one point five billion dollar fine.

50:54The largest copyright infringement in the history of copyright infringement paid by anthropic just in the last 30 days. So take from that what you will. Let's keep moving through the docket. There's another couple of big stories here. Yeah, we like the end of software stories. Really good. You got some other ones you like, Lon. Let's do it. Sure. All right. Well, I do. I do want to bring up Texas, the great state of Texas, where you and I are right now, J. Cal. They're temporarily halting new data center approvals. On Monday, Governor Greg Abbott ordered the Public Utility Commission of Texas and the Electric Reliability Council of Texas, ERCOT, we call them here, to audit all proposed data center projects in the state that connect to a public power grid.

51:34Governor Abbott has issued a moratorium on approval for all data center projects until the audit is complete. It's unclear how long that's going to take. He's asking for basic information in the audit. Who owns it? Will the data center receive tax breaks? How does it use to generate power? How does it use water for its cooling operations and so on and so forth? ERCOT currently tracking 1 ,800 data center projects representing 474 gigawatts of electricity or more than five times the grid's record for peak demand. 90 % of all new power requests right now in the Texas Q are data centers. Here's a quick quote from Abbott before I throw it to you guys.

52:12Our top priority is to protect Texans' safety and quality of life. Any project that fails to comply with the requirements set forth by the P-U-C-T and ERCOT and by state law must be denied connection to the Texas grid. Simply put, Texans must come first. So I guess my question to you, Jay Cowell, and the rest of the panel, I thought Texas was supposed to be the, like, come build anything here pro business. Is it a bad sign that even Governor Greg Abbott is saying, hang on a second, we got to hit pause? There's probably, like, two cases for this. You can take the positive, optimistic one, and then you can take the more maybe cynical, like political one.

52:47So we can start with the cynical, political one. So I think from a population base, like we have to remember, OK, when you go, I'm in San Francisco. I, you know, here, right, like everyone is pro AI and all of this stuff. But when you step out, OK, it's this is not necessarily the sentiment. even in America, everywhere, everywhere in America. And so I think from a cynical perspective, you know, of maybe gathering more votes, right, like this data center is going to become this like lightning rod issue, I think, of a lot of these campaigns that are happening, states that are pro building it, states that are against building it.

53:31And what he's asking for is information for things like, you know, okay, right, how's this going to impact electricity bills? How is Is this going to impact water use? How is it going to impact noise and, you know, local jobs and all of this kind of stuff? I think the second thing inside of here, which is interesting, is the leveraging and gatekeeping. So he's saying that an audit must be completed before any project moves forward. So it kind of turns the governor's like appointees into basically these huge regulatory cap. They have so much power now, right? Every hyperscaler has to pass these people that he appoints.

54:08So who are these people? What do they know about data centers? What do they know? Like, this is going to become a pretty big thing. And I would say, like, on the having it both ways side, right? Like, he made Texas kind of like years of marketing itself as the place for data centers, right? Like, this is a huge position that they've said. And so I think the other side is that this is an audit, perhaps, where like there's maybe there's more fluff here and it's like a nice political stance to take saying, hey, we're going to take a little bit. A reasonable one might be my take on it. If I'm going to give the most charitable, you gave the cynical, which is great.

54:49I think the most charitable one is people in Texas are very sensitive about the brownouts we've had. People have died, you know, when this happened during the two weeks of unbelievable winter we have surrounded by 100 degree temperatures. We have these like two frosty weeks. And listen, people's homes, the plumbing, the electricity is not designed for ice in Texas. So if you do get an ice storm during that two-week period, which happens every couple of years, it's cataclysmic. So people are sensitive to it. Yeah, there should be an energy audit. Obviously, there should be. And it sounds quite reasonable to me.

55:28The question is really, I think, cash going to become very quickly pollution from these. In order to get these up and running today, you need to burn natural gases. You need to burn fossil fuels in whatever form you want to do that. And that's why people are buying turbines. Elon personally bought a turbine company that was like very under the radar. He bought it for like a billion dollars really quick. The boom supersonic making that jet plane decided, hey, let's do a mini pivot to do the engines because there's so much demand for these. These things pollute and they raise the temperature of areas.

56:08So now that's like a very nuanced cash thing that people have to, I think, contend with and answer a very valid question, which is, hey, you want to go fast. You want to burn fossil fuels. Is it going to just be dirty and people are going to lose years of their life and we're going to have a smog issue? Or is that going to raise the temperature in Utah? There's issues in Utah where like certain valleys, the temperature raises two degrees, three degrees on average for the miles around it. And what impact does that have on the environment? And so you don't have to be like a climate nut to not want to have this kind of pollutants.

56:47What's your take on that, Cash? Yeah. I think I agree with Josh from pitching this a little bit of a lightning rod issue. when you go into a little bit more details of this story, they are tracking 1 ,800 projects. 1 ,800. And that demand of electricity will be 474 gigawatt, which is five times what taxes already produce. So, I mean, if you want to see, take a look, give them a little bit of a favorable understanding, what they were trying what they are trying to understand is what is going to happen if all of these 1800 projects go through and we have we are producing five times more electricity what will be the um risk to the pollution water all of those so i think that's uh that's how i understood it they are trying to audit it but from the ai company's perspective when i say it okay texas used to be like you can always go build there then where else you go build um where's the home where you you can actually like scale and everything so um let's see how that's audit go but i think you're um it's not the uh not in my backyard type of uh mindset it's actually true it can be a true crisis for the quality of life on people um 1800 projects going on and five times the electricity and i i didn't read what time frame that these uh the projects are looking at but it it has to be a little bit more thoughtful approach from the and i think government is trying to do that audit um but from the ai perspective company's perspective where do you go from there yeah i think a couple things just to start so first is a lot of this centered around something called ERCOT.

58:48So what is ERCOT? So it's the Electric Reliability Council of Texas, which is quite interesting because it's a nonprofit that actually operates 90 % of Texas electricity grid. And it's about 27 million customers, right? So you have this ERCOT committee, right that is a non-profit it's actually not run by right the texas uh government and they're responsible for 27 million residents and so what they're saying is um okay hey they're regulated you go up one notch which is where he's at governor abbott is at by this thing called P-U-C-T, which is the Public Utility Commission of Texas. And so the Public Utility Commission of Texas, that's where Abbott can appoint a commissioner to basically oversee kind of this ERCOT, right?

59:48And what ERCOT is saying, to your point, Cash, is, hey, we have 474 gigawatt connection requests. This is five times higher than the highest demand that Texas has ever hit. Like we need to ring the alarm. Like what are we doing? We need to actually like not pause, but we really need to get this under control. And as far as the data centers, what they're saying and them being ERCOT is they're saying out of the 474 gigawatt of requests, they don't know how much of that is coming from the same data center so they actually don't have visibility into that so if they don't have visibility into it they're saying how can we plan the grid for the 27 million customers that we support if we have no idea basically where the requests are coming from where are their duplicate requests coming from how can we better route this traffic, right?

1:00:51And so ultimately, I think there's a lot of stuff here that, again, I think that makes a lot of sense, right? Especially for, to your point, Cash, you know, the residents of the state of Texas, right? I think that they have the right to know these details for, and I don't think this necessarily goes against actually Abbott's, you know, in Texas's vision of, hey, come here and build. I don't think anyone's saying by any measure, hey, no, no, no, we're closed for building at all. I think what they're saying is like, hey, these are two things can be true at once, which is, hey, you know, we're going to be the place to come here and build.

1:01:30Right. And we're going to have the granular details to know, right, how we can make these projects a reality. Yeah. I think my fear is that a lot of blue state can take this and misread it and say like well texas did that we need to figure out on our own and that kind of become a de facto ban overall building new data centers and these new regulations and new structure emerge by that time yeah the only thing i would say like to add to that the only like i i would say in my view the um area that is like interesting for discussion curious what you guys think is the way this is written is it he's freezing everything so the question i think here becomes okay hey um right it it literally halts every single shovel okay in texas and so i think the question here is like okay well um there's been a bunch of these that have already been funded they've gotten the capital needed for this they've gotten the approvals ready for this and they're ready to go with the shovels and now all of a sudden hey you know what happens there?

1:02:39Do they move the capital to other states? So the devil's in the details. Basically, if the audit is, you know, you fill out a form and say, we're getting, we're bringing our own energy, BYOE, then we're done. You know, like I'm doing like Crusoe Cloud is doing a nuclear power plant. Other people are going to the Permian Basin and putting their data center next to Nat gas fields. It's like, yeah, we're good. And we all know Elon has access to solar and batteries. Like people are going to be able to just check a box. Yeah, we're good. And I bet you Abbott is so pro development that he's just going to like, I don't want to say rubber stamp, but be like, okay, great.

1:03:22You're good. And if you said, I don't have a plan, then it's like, okay, come to us when you have a plan. And maybe this will actually accelerate the ones with the plan and then penalize the ones that are YOLOing it and want to start building without a plan, which would be good for the ones who do have a plan. Yeah, it actually seems reasonable if it's executed well. Worth noting that this only applies to data centers that want to connect to the public grid. If you come in and you're like, I got a plan, it's on site, we got our own power, you're still good to go. Yeah, very important distinction.

1:03:54Yeah, that's right. All right, let's keep moving. Yeah, well, I wanted to talk a little bit about is software dead? Are we at the end of software? where everyone from Wall Street to our own panelist, Josh, has been suggesting conventional software and coding days are numbered. Soon, we'll just all be building things for ourselves directly from prompts. I want to know what everybody thinks about that. And Josh, because you sort of made this prediction, what's the timeline? What's the horizon? Like, at what point will everybody be putting their coding away and their software away and just start asking a chatbot to make things for them?

1:04:26Yeah, so I think it's already happening. and then like the caveat is I use the analogy you know when we used to listen to music someone you know we used to get a CD and you'd have to go to this record store and buy the new album and we can listen to songs and then today obviously you just click a button and you stream whatever song that you want the analogy I think in software is I think that a human would log into a piece of software is going to be the same thing as using a CD. I think in a couple years, if anyone says, oh, I'm logging into, you know, Salesforce, or I'm logging into, you know, this business application, I think will be like, that's crazy, right?

1:05:16So for software, two things will play out. Like one is, hey, if I have a CRM or this big enterprise application of my company, um the question is do i need humans logging in and accessing the interface so what i mean by is it dead is i think that a human logging in manually manipulating records inside of crms or erps um i think that is dead right and the question that a lot of the companies we work with um is hey you know i have 2 000 licenses for a crm i have a thousand licenses for the crp why can't i just have you know five licenses and then why can't agents just access all of the underlying data and go and uh do this record so it's not it's not uh it's a little different than the hey i'm gonna vibe code my own enterprise crm i don't i think that that's a little bit away um but it does challenge the per seat model and moves you to consumption which cash you have a per seat or consumption or both model because that i think what what the AI startups are deciding to do is what's going to impact the people prior.

1:06:24And I'm sure there were accounting, not co-pilots, but accounting software that were on a per-seat basis that were extremely expensive. And then you have, I guess, either a choice to pick a business model here. What have you chosen and what's worked and what doesn't? So we have the hybrid model where per seat basis you bring your team in and then you pay for the outcome based pricing based on outcome so if you need to prepare a return that a human uh prepares will take four hours you can pay thirty dollar one credit and get done in 15 minutes and our our whole goal is to add productivity into the teams uh so imagine a very complicated return from 10 65 11 20 that a human preparer reviewer takes the whole day or maybe even the whole week to prepare all of a sudden they can they their productivity is three to five x and they are able to do three to five x more and people happily pay for that because um what we are selling to are financially very smart people they see the ROI and for us our software is impacting the human cost the the resource cost we are dipping into that so this hybrid model works for us and I mean we've seen it within six months we're like a third of our revenue is agent actions already within six months um so working very well but taking a step back what josh was talking about um and the the greater debate of is software dead or or is it going to be dead i i uh i think um my opinion will be against it i think um all of these and enterprise is still in the decision paralysis, build versus buy, safety, change management, so many of these issues, they are not moving at all or moving very slow.

1:08:32Everyone is confused, internal politics, all of these. They look at the internal team and they say, oh, we can build it in our house, in-house. Why do we need to do that? So just keeping enterprises on side, there are 33 million small businesses in the country. That's 46 % of US economy, between 1 to 25 million in revenue. I don't think they're going to code all of their own CRM. I mean, you can, technically speaking, you can build your own email server, but we don't have server rooms in our basements or in our spare room. you can actually have your own protocol of private messaging. But if we don't build it, we can biode code it.

1:09:27We all can biode code it. We are happy to give our data to WhatsApp and Signal and all of those things. I mean, Signal is different. So I believe people like packaged software where the maintenance and the cost are someone else's problem. and it becomes truly economy of scale. People like more productivity cheaper and faster. If I want to, let's just say I'm a small business owner and I want to code my own CRM or anything, what will be the token cost of that? What will be the maintenance, electricity? There's a lot of hidden costs on maintenance alone and then there's business disruption. I think the SLA is kind of what you're alluding to here, Cash, And I think that's where I get concerned.

1:10:17My team has been vibe coding incredible software projects. And I go to my team and I'm like, who's going to maintain this? It's fantastic that you've built it. Who's going to maintain it? Tell me about the security. Lovable does a really good job because they have a perspective on security databases built into it. But if you're just vibe coding with the coding tools, like I find them quite random in how they apply stuff. And then the people who are building, they don't know code. They don't know what's brittle and not. But I do think software as a problem set is about to be solved. What do I mean by that?

1:10:56I think the concept of writing code is going to be abstracted to the human language. And it's starting to happen even with developers who are pushing code. and there's an agent that checks it, and are they actually looking at the code well now? Are people who are even able to look at the code, are they even bothering to look at it because they trust it so much? I guess, is it the end of code if developers describe what they want, somebody else in the organization checks it in and uses AI to proof it, and neither party looks at the underlying code? Does that mean the end of code? Yeah. I have a lot of thoughts here.

1:11:42I'm trying to organize on my head as you guys are talking. No, it's kind of mind-blowing when you think about it. Like, at what point do you just say, you know, it's kind of like music. At what point do you say there are no more musicians? Music is solved. It's just people pressing buttons, beep, beep, boop. I sampled a bunch of stuff, and nobody in this quote-unquote band actually knows how to play the songs. So when we go on tour, we're also hitting the play button like a DJ does. Yeah, okay. I give the perspective from XS. So we have, you know, a lot of engineers and they're all just using agents all day.

1:12:11They're very rarely actually looking at the code itself. So to your point, it's like, no, people are not writing the syntax of the code. I feel like this has been true for some time. I don't know how, like, XS has always been very much on the forefront of these kinds of tools, but we're like, you know, purely agents. Agents are looking at the code. Agents do all the checks before the code goes into production. Yeah, it's just rare to actually look at the syntax anymore. So in that sense, if you define software engineering as looking at the syntax and writing the actual language. And yes, it was dead maybe months ago.

1:12:40But at the same time, we're still hiring like crazy for software engineers. So how do you explain that difference? Well, I think software engineering is more becoming like software managing. Like no one's really, I mean, it's semantics, but like everyone is managing their agents and using their judgment and conversations with customers to decide what to build. That's not going away anytime soon. So anytime like the agent has to rub up against the physical world, whether it's customers, like feedback, or like, where's the world going and building for that, like, then you need human judgment. And we want to hire the smartest people to manage all these agents.

1:13:13So that's like, one one example, you know, one direction to take the discussion. But what else is interesting? Yeah, by the way, yeah, it also I think you mentioned on all on some mentioned on all on podcasts. Now, it's a lot easier for someone else to come in and just immediately get started. Which is which is cool. Now it's like, it's like, you know, humans can come and go and just all be operating on this soup of like, uh, of like, it's like, it's almost like, uh, you're operating in product space. And so no one actually needs to know what like the underlying code is. It's happening in venture as well.

1:13:45So when somebody joins our company, they can say like, you know, Josh or cash might send us an update on, you know, how the business is doing every quarter or something. We set it up now that lawn, or let's say an associate in training at our firm. We have a training firm for associates. They can say, tell me the investment history and summarize everything going on with Aragon or, you know, tax GPT. And it will just bring it all back to them. And it will search everybody's email, everybody's notion in the entire Slack. So it'll go back to when these guys were in the accelerator and what their presentations were, what notes I gave them on their presentations, what venture firms we invested in.

1:14:26So it's kind of like that, Will. All the institutional knowledge is now freely available in the corporate Oracle. And that's the same thing in code, right? Or the sales team. The sales team now, it's like, you just asked Claude Tagg. I don't know if you guys have set that up. Tell me the history of LinkedIn advertising or Squarespace advertising with us. And it's like, yeah, Squarespace was on for these five years. The CTO, the chief marketing officer left. They took a two-year break. Now they're back. They're advertising this. And here's the codes they used. And these are the products and the promotions they did over the years.

1:15:01It's kind of mind-blowing for anybody to be able to come into an organization and be up to speed hour one. Just quickly on that note, it's really important to have all the, if the model has all the context that you have, then sure, it arguably might have really good judgment. But the problem is there's so much context that's missing. Like any conversation you have in the office that's not recorded, if it doesn't have that crucial information, it's just going to have really bad judgment. So for example, we tried to get the Frontier models to do our Q3 planning, and they were terrible. Why is that?

1:15:29Because they're just missing all the little nuances and even emotional states of people as I walk around the office. There's so much context in an organization that's not just like, what's recorded in your notes? So I think until we somehow fully give these models context, there will still always be a role for human judgment. Just two things. So first, like on is software dead? I think it's like good, you know, Jcal to define this. So if something's not growing, in my view, it's dead. And when you look at the SaaS era, CRMs, ERPs, you know, all of these kind of business apps, the thing that will remain is the data that they have.

1:16:05So actually the underlying DBs, I think, are really, really, really valuable assets here. but as far as humans interacting logging in taking actions like getting more licenses all of this stuff um it is so dead in in in my view over there and second i think as far as um the actions that will start to happen um i was just talking with someone you know that is about to raise a bunch of money for this idea to actually train small models, you know, a CRM model, an ERP model, you know, so on and so on and so forth. And how do you take all of the context from the underlying DBs and put it actually even into the weights of a small model?

1:16:47This stuff is also pretty interesting as well. So where all this data goes, whether, you know, to Will, to your point, to Jacob's point, whether it's like this memory, whether it goes into these like small company models. This is like TBD. But I think the fact that a company would need to ever spend like tens of millions of dollars on some enterprise piece of just software, I think is over. It's not a question of if, I think it's just a question of when. I'm going to agree and say, and then what are you paying a software company for? I'm paying for them not to quit making and supporting the software.

1:17:25And that's worth paying for. And I saw this in startups where people would be like, oh my God, I'm going to make the 17th meditation app. I'm going to make the 27th personal CRM system. And it was like, yeah, you could raise 500K, make it. And are you going to be here in four years? Because some business is going to be dependent on that product or some individuals going to be done. So longevity and the founders promise that they will not quit. Like, say what you want about the valuation of SpaceX or Tesla. Is there anybody in their right mind who thinks Elon's going to suddenly quit trying to get to Mars or he's suddenly going to quit solar and batteries or suddenly going to quit self-driving?

1:18:11Like those projects, he's been laid on, but never wrong. And so that's why people buy those stocks and that's why they pay 30 or 40 or 50 times revenue for them. Same thing with Palantir. Do you get the sense that Alex Karp is ever going to leave that company and stop jumping up and down on his chair and speaking faster than me and being able to run circles around me and the other besties at the All-In Summit and not let us speak? No. Those people are lunatic, committed founders. That's the new moat, is having a founder who is so deranged, Will, Cash, Josh, sound familiar, and this isn't mentoring advice, but are you so f***ing deranged that your customers believe you're never quitting, you're never leaving?

1:19:03It's like the wolf of Wall Street. I'm not leaving. That's the moat. I'm not leaving. The show goes on. We're going to f***ing bars. We're going to solve tax. That's what people want. They want a driven, absurd founder, and that builds the moat in your business. Every founder, every VC, oh, what's the moat? You know, anybody can write code. The moat is f***ing you. You're the moat. Will you quit? And I can f***ing tell when founders are going to quit. I can tell because then they start asking for like a crazy comp package. They want to work on a side hustle. they want to start an incubator or there'll be a scout.

1:19:42Like when I start seeing that and founders like, Hey, I'm a scout. And I'm thinking of starting a micro fund and a podcast. I'm like, is the podcast about tax? Yes. Cash do it. Or is the podcast about you trying to recreate all in or whatever? Like, let's stay focused on the mission. How obsessed are you with the problem, the product, the market, the solution that you're building. That's the new mode. Thanks for coming to me, Tato. Exactly. I mean, on a side note on this software thing, I think what an incredible job Cloud, OpenAI, and all of these coding tools has done. Cursor has done to make everyone dream.

1:20:25We went to Cello Shop, and the celloist is building their own CRM to manage their customers better. An Uber driver is building something. My CRO and his wife is a psychiatrist and therapist, and she built a bot that goes into other psychiatrist videos and give her inside topics to create videos about. Amazing. So she created a research assistant, a chief of staff, that would normally cost$100 ,000 that doesn't sleep. Yeah. And so it's such an incredible explosion of ideas people can dream and build. So I think, I mean, just a year or a year and a half ago, the dominant conversation was, oh, what is the time of coding agents and a generation of code?

1:21:14Nobody's questioning that, that they put a software engineer in everyone's pocket. And our goal is to put a tax advisor in everyone's pocket. Done. I mean, you know what the TAM of software development is? The population of planet Earth, which is the population of food, electricity, smartphones, and now software. Like, everybody wants that. Yes, there are some people who don't have it right now. You know, they're still working on food, clean water, et cetera. But the number of people living in abject poverty is dwindling. And the number of people connected to the internet on high speed with access to these tools is skyrocketing.

1:21:55There's reason to be optimistic. Um, I'm sorry for getting a little crazy, but you know, I just did, I don't know what you guys are on, but I'm, this is not a commercial for breakers, uh, coffee for milligrams, but boy, does this get me active? What are you, are you on coffee? We should reach on coffee. I'm on four. If they want to come on board as a sponsor. All right. Cash is on Zin. He's got a, he's got, you know, I got to tell you as your investor here and as your guide on the side here, I'm going to tell you, you got to get rid of the vape. Okay. No vaping. I don't want any popcorn. If you're going to do nicotine, my professional opinion is do it with that.

1:22:32So you got this in. What do you want, Josh? You want to get, you want to give it? Okay. Celsius. Very nice. Blood type Celsius. Oh, plus a little sesh. All right. So we look at that. We got Lucy sesh and Zinn fired up after that. Now we're fired up. We all have to go back to work. I have a funny story about the vapes are banned in San Francisco. So I had a guy. So I went to his shop. Wait, what do you mean banned? Like you can't buy them? You can't buy a vape. No. Yeah. What? Vapes are banned in California or just San Francisco? San Francisco particularly. So I went to - You go to San Mateo? You go to Milbra?

1:23:11Last time I bought it from Monterey. I mean, we were on trip at that area. So I bought it from Monterey. so funniest thing happened um i went into the shop and they said they banned it and the city uh the official when they came in they had the pictures of someone vaping outside of their store i was like all right cool and i didn't have the zin either so i went came out and someone offered me to buy a m*****g right outside the aisle and said this is the most san francisco thing ever you can't buy a vape and i was a journalist jason as you might know and i used to uh smoke two pack of cigarettes every day and seven six seven years ago uh it helped me quit uh juul and all of that and this common sense regulations like it's my choice i am choosing the lesser of the two evils, you can buy the cigarettes, you can get offered, you know, like just on the market street in San Francisco, but we're going to ban the vapes and the zin.

1:24:18I think, don't hold me on this, but my understanding is that you can't buy flavored zin either in San Francisco.

1:24:29This is a very unique San Francisco moment. It's crazy. It was the funniest also. I mean, I just, I got like three great tweets and clips from this episode. Another amazing episode of This Week in AI. Go to thisweekinai.ai to sign up for our newsletter. In about two weeks, we're launching a new paid product. Every week, we're going to profile two AI startups that I pick. And my team deep researches. We write a deal memo. We do a mini interview. And you're going to get that. It's going to be$300 a year. We haven't started it yet. But if you're interested, sign up for the email now and get the free product.

1:25:04We'll put the first couple of profiles up for free, I think. But that's a new product that's going to pair with what we're doing here on the show, which is if you're a venture capitalist or you're a business development person, you're doing M &A, you want to buy companies, you're going to want to go to thisweekinai.ai to get those 100 profiles a year. You guys are hiring. Will, where should people go to find out about an incredible career? Or just DM me on Twitter. All right. Right. There it is. Josh, where should people go if they're interested in joining you and your company? Aragon.ai slash careers or email me, josh at aragon.ai with anything exceptional that you've done.

1:25:41No resume needed. No resume needed. Yeah. Okay. Cash, what do you got? Go to taxup.com slash careers. We're hiring for incredible exceptional sales hunters and customer success managers. Is that remote or in the office? Will you take somebody in Arizona or in Toronto if they're a killer? Anyone, exceptionally anywhere in North America will take them. Okay, there it goes. You might want to look in Tokyo too. There's some great expats living there. I mean, the base case is you make at minimum 200K. And plus, if you - I mean, listen, in San Francisco, hard to get salespeople, too much competition.

1:26:21What I learned from Jeremy at Yelp, they moved to Arizona, had this other startup. This was one of the best sales hacks I ever had. ever heard from a sales startup. They were like, we're going to start a sales office. So they had their team make a list of what they thought were the five best places to do it. One of them was Arizona. Somebody look up on my team where Yelp's sales offices. They put out a bunch of job descriptions. They did a bunch of interviews. They went on a tour for 10 days over two weeks, two days in Phoenix, Arizona, two days in Salt Lake City, two days in bowl one day in boulder two days in austin one day in dallas whatever it was at nashville and they just set up um interviews for these like very high paying jobs and they were just like yeah you know we went to salt lake city and so many people showed up we got all these mormons these people were like knocking on doors to tell people they need to convert to christ and i was like and they just person had this epiphany like if they can do that they can sell my CRM system or whatever sales automation software it was.

1:27:28And after that test, they decided to just hire like five people there. And then they looked at where they lived and they just picked an office equidistant from those five. That was the most equidistant popped up a sales office. Anyway, I thought that was one of the great tactical, practical hacks I'd ever seen. And that's what great entrepreneurs do. Uh, well, we got to talk. I got, I mean, how am I going to get on this cap table i gotta figure out how to go in a time machine and go back and make up for lost time we gotta make it three for three today's panel we only went two absolutely absolutely we'll see you all next time on this week in ai bye