#2307 Eric Ries: Why Anthropic Won and How To Build Incurruptible companies

Speaker A: I met the Anthropic team when they first left OpenAI. They were really committed to the idea that this new generative AI should be used and commercialized for the benefit of all humanity. The irony of this whole situation is one of their early backers was FTX Vibe Coding era is going to be remembered for a Chernobyl style disaster is my prediction. The transformer technology that is the basis of all modern LLMs was invented at Google. If you look at the co authors of that paper, they all, every single one, had to leave and do it elsewhere. I know when someone sends me an AI generated thing, I always know it instantly. It's always garbage. Too many people are excited about using AI to replace human creativity instead of augmenting it.

Speaker B: Eric Rees, who helped so many entrepreneurs build phenomenally successful businesses based on his Lean startup philosophy, is back with a new book called Incorruptible where he talks about why some companies succeed, do well over the long term and others just keep dwindling. And that's what we're here to talk about today, specifically related to AI startups, to see what's working, what's not, and also what he's building himself. Let's get into it. Presented by Zapier, the AI automation company. What's Anthropic's mission and how are they been able to do well because of it?

Speaker A: I met the Anthropic team when they first left OpenAI and they left over a dispute with OpenAI over um, you know exactly how to pursue the question of AI safety. So they were really committed to the idea that this new generative AI should be used and commercialized for the benefit of all humanity. And they were very worried about certain specific safety scenarios that at the time, time seemed like far out science fiction. I remember being like, whoa, are we really imminent? Is this, is this imminent or is this like a law? And they were like, no man, it's imminent. We need to be working on this right now. So they were farsighted in caring about that stuff. And again, everyone wants to be like, okay, AI safety is now a super polarized debate. And of course I have my own opinions about AI safety. I'm happy to talk about that. But I would ask for the purpose of this conversation, like put that aside for a second. They were committed to. So my position with them was like, let me help you encode these commitments into your company structure. Because that's really what they were worried about. It's like, well, we're going to raise all this money and I Remember, they first were like, we think the solution would be to raise money from really, really values, uh, aligned people. Which they did. They were in a position to really curate their initial cap table. They were selective about, uh, who they took money from. But I remember we had this conversation about, but what if you're wrong? Like, what if somebody who seemed really values aligned turns out not to be? And what if investors who are aligned, like, naturally, when the amount of money we're talking, because they were talking about AGI before that was in the news every single day. We're talking about technology that could be worth trillions of dollars, but maybe hundreds of trillions of dollars. And it was like, are you really so confident that these people will be able to maintain their principles even in the face of this overwhelming temptation? And, you know, they were, they took it really seriously. Okay, we need to figure out how we're going to build the structure that is, um, resilient, even if, even if someone tries to betray the values, even if someone turns out to be unaligned. Now, the irony of this whole situation is one of their early backers was ftx. So they thought those guys were super aligned because of their supposed commitment to effective altruism. And it actually turned out to be a total disaster when that company imploded. Huge chunks of anthropic stock were sold at auction to any investor who wanted them, including people who are super, super unaligned. So it actually turned out to be extremely important. And I, um, mean, we can talk about the structural stuff, you know that the elements of building a company that can be truly strong.

Speaker B: Did you help them do that? Did you help them actually codify? You did?

Speaker A: Yeah, but I mean, I don't want to take credit for what they've accomplished. I played a very big part.

Speaker B: Okay.

Speaker A: Uh, and so, uh, in this story.

Speaker B: Yeah, so you helped them codify their, their beliefs. They then said, we only want investors who align with what we stand for. They ended up with an investor. He did. Sam Bankman fried very famously. Turns out he didn't. And then random people were able to buy his shares. Why were they still able to maintain their, their. I mean, why were they able to still continue standing for what they meant for what they did before?

Speaker A: Yeah. So there's two components to it I would think of as now, like the inner part of it and the outer part of it. And I know, again, it's so natural that now, now people see anthropic as this mega company. I totally get it. And so it can seem inaccessible, but Remember, to me they're just a couple guys in a garage. Like I, I met them when they were. No, like, it was not a big deal except to people who are very, very in the know about this inside baseball stuff, um, related to AI and they did, they made the critical commitments then before billions and hundreds of billions of dollars were, um, were at stake, which is a really important part of it. So the inner part of it is really about alignment, coherence. Can we get everyone in the organization really committed to some set of principles or some kind of thing, some kind of vision? And then the second part is what I call integrity, which is not some vague thing. I know people hear these languages like they worry. We're going to talk about morality. What I mean is more like structural integrity. The ability to make and keep promises and the ability to be trusted by employees, by customers, by partners, and even by investors. So anthropic structure is not your typical. Investors run the show. What's called shareholder primacy. Uh, set of practices. Uh, their company, uh, is governed by something called the Long Term Benefit Trust, which is an outside set of trustees that are not like a vague advisory board or like, remember the Trust and Safety council that uh, Facebook famously had and until they ignored. Um, the LTBT has the actual power to appoint members of the directors of the board of the for profit entity of Anthropic. So they have real effectively veto power over things Anthropic does. And their job is to act as guardians, they call them mission guardians, uh, to make sure that Anthropic never deviates from that, that mission.

Speaker B: And so that's why when the government asked for carte blanche over their technology, they said we can't. This just goes against what we, what we believe.

Speaker A: Yeah. And you know, it's people now that now we know how it turned out, obviously, like the district court or the district court just gave a preliminary injunction in which they absolutely demolished, uh, the government's case, saying this was a case of overreach. Um, and not only that, but like they, they've reaped already, even though they gave up $200 million. $200 million contract, even for a big company like event, is actually really a lot of money. And you can tell because their competitors all race to go get this contract as soon as they gave it up. Um, but like Claude went to number one the next day. Like that was a huge boost. I saw a video. Some friend of mine sent me a video. People had come out to chalk up the sidewalk outside of their headquarters, thanking them for doing this. And believe me, like, that is not typical for tech companies in San Francisco right now. If you're being chalked up, it is not usually in a positive way, like it caught when you right thing, it causes the positive ripples sometimes, which were down to your commercial advantage. But the key to the whole thing is you have to be willing to do what's right, whether or not you know that you're going to be rewarded for it. You do the right thing for its own sake and you trust that, that the chips will fall where they may. And I think that kind of strength is only made possible if you're not always looking over your shoulder like, oh, uh, I'm worried I could be fired. Oh, I'm worried investors might not like it. I'm worried about the next activist campaign or whatever. Um, so I think, I think that that courageous stand that they took, again, leaving aside the politics of it, just looking at it from a pure business perspective, is an example of a very savvy business strategy that sees trustworthiness as an asset, like a business asset that can be intentionally cultivated and acquired. And in fact, I write in the book, I think it's one of the most underrated assets in the world today.

Speaker B: But are they doing well because of that? Or are they doing well because Claude code is good, because they kept shipping Claude, Cowork and so many other tools? Honestly, does that even really matter? Isn't it just is the product good at the best price possible and the rest is just, we like you?

Speaker A: Okay, yes, you can say that it's, it's, uh, because they have the best product. But the question you got to ask, I think is a deeper question of why do they have their best product? You look at who chose to work for them. Like, part of the reason they have such good products is because they've been able to attract such incredible talent. Why? And especially at the beginning, they've been the, the, you know, they were behind, perceived, widely perceived by a lot of people as being behind OpenAI for a really long time. And even in some cases I've heard people say that they've been behind, you know, behind Google, behind Grok, behind whoever. Like they, they were not the, uh, odds on favorite to win this race. And part of the reason they've been able to do such a good job is simply because they've been able to attract capital and um, uh, and employees. But then the other question you got to ask is they're ahead in enterprise.

Speaker B: Why?

Speaker A: People would say, well, they have the best product but like, is it your experience that enterprise procurement departments have a history of like, being really able to meritocratically judge the best product and it's so reliable. Like we, we kind of, uh, we overlook sometimes the facts that are staring us right in the face. I think it's extremely obvious if you look at the data that's been published so far that part of the reason they have an enterprise advantage is because they have a trust advantage. If you are worried about the liability of adopting a platform like that, it matters who your vendor is. And you know, the ones that have a kind of cowboy kamikaze ethos that's very scary for enterprise.

Speaker B: But let me give you, let me give you a counterexample from your book. You talk about Google.

Speaker A: Yeah, this is a real. Yeah, this is not my, like, personal judgment of Google. And listen, Google's a great company. That, that's not. Again, it was not about absolute right, absolute wrong. Um, I made a study of people who have blogged about what it was like to work at Google who had been there for 10 years or more and who left. So there's like, there's so many of these blog posts out there. Google people are very prolific in their blog posts, uh, after they leave. So it's, it's a unique. Only insofar as we have this beautiful glimpse into what it was like. And a very recurring theme of those blog posts is people talk about this creeping mediocrity and kind of loss of something that made the company special. They, they, they talk about it like a grief, like something precious was lost and they can't figure out where it went. Like, they, they don't say Google's leadership sucks. They don't say Google's a bad company. They said Google had great culture, great leadership, great int. And yet despite all of those advantages, plus Larry and Sergey have dual class shares. Remember, despite all those advantages, this thing was lost. One of the, um, former employees put it this way. He said, over my tenure, I think he'd been there like 13 years, decisions went from being made for the benefit of the customer to being made for the benefit of Google to being made for the benefit of whoever's making the decision. So yes, Google's still around, they're doing fine. But over time, this kind of corruption is corrosive to their ability to create value. And of course we're talking about anthropic and OpenAI on these guys. Don't forget that the transformer technology that is the basis of all modern LLMs was invented at Google and if you look at the um, co authors of that paper, has a ton of co authors. Not a single one commercialized the transformer at Google. They all, every single one, had to leave and do it elsewhere. So this is not just a matter of morality or virtue signaling. It has real tangible business consequences.

Speaker B: The reason I was going to use them is that Gemini is not a bad product. It's a good product. Notebook is something we use on a regular basis. I could keep on going through all their products that, oh yeah, I use

Speaker A: Google products every day. I guess Google's a great company now you have to, but you have to ask the counterfactual. Since we're, since we're saying, oh, it doesn't matter, we have to ask what would have been, like what, how would, how would the, how would Google be today if they had been the leaders in this category instead of allowing their own employees to leave and do it elsewhere? So how much was that head start that they gave to the rest of the industry? How much was that head start worth? Um, now I think if you do the math, that's a bigger corporate debacle from a value creation, value loss perspective than Kodak's inability to commercialize a digital camera just because the opportunity cost of that is so immense. We're talking about companies with hundreds of billions of dollars of valuation that were created by ex Google people that could have been created at Google but had to be created elsewhere. So again, we can't know for sure. It's hypothetical. You can never know with a hypothetical. But the good news is the book is not about hypotheticals. We have so many examples practically in every industry of companies that have bucked this trend. And yet when you study those companies, you'll notice they all, pretty much every single one, violate many of today's supposed best practices about how companies are supposed to be run, created and governed. And I basically think today's best practices are a value destroying mess. And part of my goal in writing the book is to replace them with better best practices. So we mentioned Costco. Costco is repeatedly attacked over its decades of life by activist campaigns that have accused them of having bad governance. But you also have companies like Patagonia and Vanguard and Novo Nordisk and Ikea and John Lewis Partnership in, uh, the uk. These companies that a lot of them have been around for 40, 50, 60, 80, 100 years and are still basically true to the ethos that remained, even though we're living at a time when average corporate lifetimes are collapsing.

Speaker B: So what else, what else goes into this.

Speaker A: So when we talk about an organization that does the right thing, we're talking about an organization whose character is consistent and aligned with human flourishing. So a big part of the book is, how do you create such a thing? How do you instill that deep down into the bones of a company? And, um, what you have to realize is that for the vast majority of the history of time, there have been joint stock corporations. It was seen as completely obvious that, uh, corporations should be incorporated to do a specific thing. And in fact, in the 19th century, to convert a company from a mission of doing any specific thing to just, I'm just going to enrich my shareholders would have been seen as a crime, and your corporate charter would be voided. Like, this is not. This is not actually what is the foundation of capitalism. So anyway, because of the shareholder primacy thing, Federal price basically says anyone who's rich enough can take over any company they want at any time. And most founders are hopelessly naive about this point. They think that they're in control of the company. And they're always so betrayed when they find out that their own founding documents basically say that they can be removed at any time. And they don't realize that that's a choice. But rule by the richest would be bad enough. This is actually kind of more like rule by whoever can borrow the most money, because you don't actually have to be that rich to take over a company if banks will loan you the money. And we've seen lots of examples in recent years of people borrowing a lot of money, taking over a company and doing whatever they want with it. And there are people who defend that on the basis of free market, uh, free markets. And you should be allowed to do any crazy thing that you want. Okay, this is not a book about policy and politics. To me, the question is, we who build organizations of all sizes, do we think that's a good idea? Do we think that's actually value creating? And almost every person I know who works for a living, who builds things for a living, has an intuitive sense that that cannot be right. And I think we have to, um, adjust our formal categories, our formal definitions, to bring them more in line with this intuitive understanding that all builders share.

Speaker B: All right, can I talk about minimum viable product for a moment?

Speaker A: Sure. Yeah.

Speaker B: Okay. It was the most exciting idea in the startup world for a long time because you basically brought us down from trying to build too much too simple. The thing that I wonder, though, Eric, is we talked about Claude. Is Claude going to take over Every little mvp. Is the idea of an MVP still possible? Will it still be possible a year from now to be able to do that? Or are we going to now every turn be competing with Claude Code and OpenAI and customers who are expecting perfection because that's what they're getting already?

Speaker A: Well, I have a real contrarian view here. I know this is not the hotness at the exact, at this exact moment, but I think too many people are excited about using AI to replace human creativity instead of augmenting it. And I don't think that's going to work very well like this. The Vibe coding era is going to be remembered for a Chernobyl style disaster, is my prediction. People are Vibe coding software that they do not understand. And it's only a matter of time. With all the hype that's being poured into this right now, it is only a matter of time before somebody deploys a Vibe coded solution to a mission critical system that not only wasn't reviewed, cannot be reviewed, because nobody, there's no human being who could possibly understand what the code does. I've had a lot of experience with Vibe coding. I've done a lot of stuff in cloud code and in other tools. So I know what I'm talking about. The um, the, the fallacy, the error that people are making. And this is a really important thing for people who want to use these tools for MVPs. If you focus your energy with these tools on artifacts, if you see like, oh, Claude code is awesome because it can make me a perfect artifact, then you're making a double mistake. First of all, you don't know how good the artifact is. We have a lot of good evidence now that when people get enamored with their own creations, they way overrate how valuable they are. So you don't really know what you're talking about. So the part of MVP that is about getting FE back from human beings is more important than ever. Don't delude yourself. I'm going to talk about the research that's been done in this area. But the second, I think even more important part is when you're using Claude code to create artifacts, you are actually causing skill atrophy in yourself. You have gone from the flow state of building stuff into what we call dark flow, which is the state you get into when you're using a slot machine and you just see it like I, I have apps where I'm just like, I'm hitting next, next, next on Claude code and I'm like, ah, uh, just whatever you do whatever. Yeah, it sounds good. Sounds good. Sounds good. And I notice over time, Claude will start to make design suggestions. Hey, I think this would be a good design. And you're like, oh, yeah, it sounds pretty good. And, like, when I review, where did the design end up compared to what I originally envisioned? I'm like, wait a minute, that's not the app I was even trying to create. So I think we have to be, um, much more careful about how we use these technologies, and especially for MVPs now, I think it's not the underlying models that are the problem. It's the way that there's harnesses that they're being embedded into. Now, I'm talking my own book here because I helped start an AI research research lab around this contrarian thesis. So obviously, you know, uh, I'm a believer in that thesis. But I think we have really good evidence that it's right.

Speaker B: Uh, because we're disconnecting ourselves from the final customer, and we're getting stuff that we feel we made, and anything we make, we feel more love for. And so we're twice delusional. Once for having m made, twice delusional.

Speaker A: I don't know if you've ever seen this psychology research is incredible. They put people in an MRI machine and scan their brain, and they ask them to do various tasks to see what lights up. And they'll have them do something like, um, read their favorite poem. People who like love poetry, and they read your favorite poem that you ever read in your life, and they'll have them read it in the mri, and you'll see the pleasure centers of the brain just go whoosh. You know, it's like your favorite. You're remembering your favorite thing. You're reading your favorite poem. Then they'll be like, great. Can you take two minutes and write a poem yourself? And the person will be like, I suck at poetry. I can't just, like. Just anything. Just whatever you can do. Write a poem, and they'll be okay. They take two minutes writing the crappiest poem you ever heard. They'll put them back in the MRI machine. I'll be like, now read the poem you just wrote. And the brain centers light up just the same as if we're reading Keats or whoever else. You can't imagine how addicted you are to the feeling of the thing is your precious thing that you created. It is. Causes such incredible delusion.

Speaker B: I find that myself, too. You're right. Like, if I create it, I design it. There's some beauty in it. Uh, like a Suno song that I made is just going to hit me so hard and emotional, might even make me cry. I thought part of it is because it's tapping into my own use. You're saying in addition to that, the fact that I created it makes me love it.

Speaker A: Yeah, the fact that you yourself created it, it makes it so much, uh, more valuable to you. And unfortunately, um, vibe coding gives you the simulacrum of having created it, so you feel the same sense of ownership over it. So it lights up the same pleasure centers in the brain. And even though in a lot of cases, you don't even understand what you've created. And how many of us have received a vibe coded proposal or email. I know the Claude cowork. I know all the major tools. I know their house style of how they like to style things. So I know when, uh, someone sends me an AI generated thing, I always know it instantly, and it's always garbage. So think about it this way. How can the things that other people are vibe coding be garbage? And the things that you're vibe coding are great. What are the odds, right? Like, what's happening is you're using different parts of your brain to evaluate. So, like, so at Answer AI, we've built these tools that are human in the loop, um, that are designed to help you learn how to produce the artifact. And this has always been the insight of lean startup going back many, many years. We called it validated learning. The learning is the asset, not the artifact. So this has been like, one of the biggest crusades of my career, is starting to get people to value actual scientific learning. And what's cool about LLMs, I think, is not that they're artifact generating machines, they're okay at that, but, like, take them out of their distribution and all of a sudden you're in big trouble. But they're incredibly good teaching machines. They're maybe the best teaching technology we've ever developed in history. And so if you just get into the habit, and of course, if your tools are designed for learning primacy, it's better. But aging any tool, instead of saying, make me an artifact, just say, teach me how to build the artifact. Teach me how to do it and everything it creates for you, insist that it walks you through, step by step, exactly what it does. Make sure, quiz you to see if you actually understand. And people hear that, they're like, but I have Claude code generating thousands of lines of code a second. There's not time for that. I'm like, right, that's What I'm talking about. So you want to be responsible, morally, ethically and economically responsible for deploying these armies of robots that you don't understand. You're much, much better. I think in the long run. People who invest now in craftsmanship, in the understanding of software principles, architectural principles, artistic principles, whatever the thing is, are going to be so much better off. They're going to wind up being like incredibly super productive cyborgs that are going to run circles around the Vibe coders.

Speaker B: Because they know how to code or because they understand how a Vibe coded. No, it seems like you're saying because they know how to code.

Speaker A: Well, I don't care if you can type the physical lines on your keyboard. Right? Like, so if you say I'm a great writer, you're like, really? You have excellent penmanship.

Speaker B: Okay.

Speaker A: No, no, you should see my hand. My handwriting was this terrible. Like, to be a great writer is to understand the craft of writing, and to be a great programmer is to understand the craft of software engineering. That's what we're talking about. So I think these tools will make, uh, make that, that elite level of performance available to far, far, far more people than is currently possible with the way we teach people programming. So I think a lot of Vibe coders can graduate to this skill is you just have to be determined to understand what you're doing. And if you have that hunger to understand, you will become far, far more powerful.

Speaker B: You know, let's take it to writing. Because writing is universally understood. I've never found that AI can write well for me, what you're saying is don't even try. Instead, maybe give it the transcript from this interview and have it help me figure out what I should write. Not even create a first version, but like, like asking me the right questions to understand the meaning. That's.

Speaker A: Yeah, you got it exactly right. And. And it's. It can critique too. So like, so I listen. I used AI I know this is like, not popular to admit this kind of thing, but it's funny. Vibe coding is supposed to be like, I vibe coded. I'm M5 maxing and whatever. Token maxing. But in writing you're supposed to be like, oh, no, God forbid anything AI helps. Like, I use AI a lot with the writing and research of this book. It was incredible because I had a. I had a fully custom rig that kept me in control. So AI did not write the prose for me. But like, when I would write something, it would have available to. It would create a context for me. So we had what we call shared context at Answer AI, where I and the AI have the same information and we see the same steps, and we can go back and modify any steps if a wrong turn was made. And it was. The context was not just what I had written my previous draft, but also I had, like, hundreds of test readers. I had 10,000 comments from test readers on this book, and I had access to all of them while I was writing. But that's too many to. You can't look at 10,000 comments. The AI would find the comments that are related to the thing I'm writing to make sure I have that context available. Same with research. I had this massive research archive, more than I could possibly keep in my own brain at any given time. So it would help me make sure I was aware of what information was relevant. And then it has. It's been trained on all the literature that has ever been written by a human being, unfortunately, speaking of someone who's in the training data without any of us being compensated by the way, which I think is atrocious. Um, even still, it has knowledge about writing. So it's like, you can be like, critique this for me. Is this really good? Is this 10 of 10 good? How could it be improved? And again, don't listen to its suggestions because it will often go back to, it's not X, it's Y. And all this other. You'll start loading it up with EM dashes and other garbage. You can't do that. Instead you could say, like, is this really good? And what I would find is basically I would get into an iteration cycle where, um, I would bring in the research, I would synthesize. Sometimes I would have it help me do an outline. Sometimes we would just go paragraph by paragraph. And after. This is the key. I would work with it for a while until it was convinced we had the best possible thing we could make. It's like, this is 10 of 10. When the AI says it's 10 of 10. Now our work. Now the work begins. See, people stop there. But no, now we have reached the limits of its training data. Now it's time to actually do the creative act of writing, which often would be like, okay, now I see how it's like. This is a very basic way of arranging this information. Now let me try to take it to the next level and bring my own unique skill and creativity to bear. But from as a writer. For me, the hardest part of being a writer is the blank page or the feeling like, what do I do next? And AI is just extremely Good at. You're like, look, I have a hundred things I got to do. Will you just pick one for me and let's do it and like, help me m. Like, uh, you know, I would get interrupted. I have young kids, I get interrupted. I was in the middle, I was doing great, the truly great writing, and I get interrupted. Now I can't remember. It's like, you could just be like, where was I? Okay, get me back up to speed on what I was just doing. Help me make forward progress. So for me, it was extremely powerful as a tool and I think we only scratched the surface of what it's capable of. But again, because I was 100% focused on having it improve my own craft and skill, not on having it do the artifact for me.

Speaker B: Can I see it? Can I see what setup you use, what tool you used, and like how you used it?

Speaker A: Sure, yeah, yeah. Um, I don't know if it's public yet. The tool is called Solve it@answer AI that I do to build all this stuff. I just don't know if we have made it public.

Speaker B: Um, I think the video is right on the page.

Speaker A: Oh, if the video is right on the page and it's already public and yeah, go to Solve it dot com. Sorry, I just, I don't want to, I don't want to over promise when I'm not 100% sure what's public yet.

Speaker B: But ultimately, when you were writing, what were you looking at?

Speaker A: Solve it is based on jupyter notebooks for those that know what that is, which is basically like a messaging system where instead of a chat, we create structured messages. And the messages can be content, like markdown, you know, just text. They can be prompts to the AI to do something, or they can be Python code.

Speaker B: Is this what it looks like? Yep. So this is what you are staring at as you are writing. Mhm. So what am I looking at here? The top is what and then the bottom is what?

Speaker A: Yeah, the red boxes are prompts. So the brain, uh, the little brain icons tells you that Claude is responding in thinking mode.

Speaker B: Okay.

Speaker A: Um, and so, um, the green boxes are notes. That's just raw markdown. That's just the information. Um, so like you were talking about, like taking the audio of this, um, interview and make a blog post out of it. Like many Answer AI blog posts are done that way. We just sit down, we record a discussion about it, and then we transcribe it and bring it into this, uh, into this thing. And the key is, so what happens is, um, LLMs, like people make AI into this magical thing. And I just really, really encourage everybody to, like, learn about how large language models actually work so that you can learn to reason about what the technology can and can't do. It is not a magic trick, although it is remarkable. So large diagrams are autoregressive, meaning they learn from examples. That's really all it is. People is just a token predictor. And that's true. It's just a token predictor. It's just trying to figure out what word comes next in the sentence. The fact that it evolved a world model is an incredible feat of engineering and pokes so many holes in what we usually used to think about our theory of mind. And what is intelligence like? It raises huge philosophical questions. But put that all aside because it's a token predictor. Like if you're in a, if you're noticed, if you're in a chat with Claude or OpenAI or any of these tools, if it makes a mistake and you correct the mistake. So you. Let's say you had say, here's the transcript. Can you write me a summary? It writes a summary like, oh, God, that summary is terrible. It doesn't include these important values. This is stupid. Like, please do it over again. It does it again and you say, oh, that's even worse. And you go back and forth, right? You might notice, like I noticed that chats with these tools, they either get better over time or they get worse over time. You'll just be like, God, it's like it's getting dumber the more it talks to me. And that has to do with the mechanics of the attention mechanism. As the context gets larger, you're literally spreading the attention over more and more and more data. But also it's learning from all the bad examples it gave. So it's actually much more likely. Even though you corrected it, the fact that the bad example is still there in the context makes it more likely to give you a bad example. So when we would take a, uh, transcript, we would say, look, give us a summary of the first section of the discussion or whatever. And it would write a paragraph. And instead of saying, no, that's wrong, this tool allows us to go into that box and change the AI's output as if it got it right. Ah, uh, okay, so now the AI thinks that's what it generated. Now it's learned the style of how you want the rest of the blog post to go. And the next paragraph is far more likely to be what you want. You do that two or three times. And now you can say, great, now please write up the rest of the blog post for me or like write it to this outline. Or you know what, you could kind of go back and forth with it. So, yeah, so that's, that's what we're looking at. You can see here there's like section headers and, um, text going on. And then the red boxes are the, the actual interactive bits with the, um, uh, with the, with the LLM and there's video.

Speaker B: Are you writing. Where's the actual writing that you write or that it writes? Where's the output?

Speaker A: Yeah, this is a technical, um. This is being used for writing code. So the writing is not. I don't think it's being demoed in this one. But yeah, a lot of the writing would either be in these boxes myself, like if I was just working on a small section, or I would keep a separate document with my markdown files per chapter. And if I just want to do raw writing, I would just go over there and barf a bunch of stuff on there. But once I had a complete first draft that was very rare, most often I would be like, even if I want to rewrite a section, I would do it in this environment because it's just so convenient once it's there to be like, wait, is this rhetoric? Like, I often use a metaphor or a certain rhetorical move, and I just be like, is that effective? Can you follow what I'm saying? Would my reader be able to understand it? Like, do people know? Like, just like, I could just ask it these questions in the process of doing the writing. It was exceptionally helpful.

Speaker B: Can I see it? Like, can you log in and do screen share?

Speaker A: It's funny because we were just talking about, um, uh, trustworthiness is an asset. See this?

Speaker B: Oh, yeah. So there's a section of the book as you wrote it. And so red means that the AI came up with the writing.

Speaker A: No, no, no, no. This is, this is at the end. I'll show you, I'll show you. This is just like. This is the end of the process where this is a diff. So there's the red are the parts that are being removed.

Speaker B: Diff.

Speaker A: So I had this section about, um, that was like trustworthiness research. It looked like it was a bit of a stubborn mhm. And I had written this section, which at the time was called Master, using it. And you can have this, um, for those that get the Zelda reference, although I think this is no longer in the final manuscript.

Speaker B: No.

Speaker A: Um, and it's like here is. Oh, uh, yeah, he's actually here. The metaphor I just gave you about the friend who's a drug addict. Oh, uh, that's funny. This is when I wrote this section and, um, it was riffing on the fact that, uh, Jim Senegal, the founder of Costco, made this quote about raising prices as a form of heroin. Once you do it and get away with it, you can't stop. Anyway, so if you go back, you can see my, my dialogue with the, uh, AI. It goes on for a while. So for example, here I brought in some research called trustworthiness is a catalyst for strategic benefits. So this is a note which I can pull up for you. You can kind of see it here. This is just like tons and tons and tons of research of different studies and information about the value of trust. I pulled in so much research for this book, you can't even imagine it.

Speaker B: No, I felt it.

Speaker A: Here's a whole bunch of it.

Speaker B: You're saving the note just so the AI has more research in this.

Speaker A: Yeah, yeah. So I was like, okay, good. So give me this. Consider this section of research.

Speaker B: Mhm.

Speaker A: What are some ways that we could use this specific research in this specific section? And so here it's making different ideas about what we could do.

Speaker B: Uh, what did you respond to it to? Uh, you said now

Speaker A: you can't actually see. Unfortunately for historical reasons, this isn't great because remember I mentioned that when it does something wrong, I rewrite it.

Speaker B: Mhm.

Speaker A: So what happens is we brainstorm and go back and forth for a while. I then delete all the brainstorming and back and forth and collapse it down to a revised version that it thinks it wrote so it doesn't know that we did. I erase all that from context because basically when you're using LLMs, context is everything I remember. This is old. This is old work. This is from last year when Solve it was very primitive and when context windows were crazy short. Now there are a million tokens. This is probably back when it was like 100 or 200,000 tokens. Um, anyway, so here you can see we've now added these different stats that are from that research document. Um, and here's some additional things. Anyway, so we go back and forth. I get an evaluation from it right away. I love this part of it. I find evaluating my own work extremely difficult as a writer because of course, think about the MRI thing. It's not that I want the, uh. It's not that I trust the LLM's evaluation to be correct. This is the hardest thing to understand. But just having something to react to when you're alone trying to write and you're feeling stuck is incredibly powerful.

Speaker B: And so you said to uh, it evaluate this new version. Is it better than the previous one? It evaluated and said here's a strength. Research based credibility. Added statistics and research finding provide concrete evidence for claims that previously felt more anecdotal. Okay, and it's giving you all of this. Where do you then take it and rewrite based on what it's given you in chat.

Speaker A: Do it again. Let's streamline this version for both length and tone.

Speaker B: Um, and so then it comes back with an answer with like what the streamlined version looks like.

Speaker A: And you go in. Well, every time you see a new version here, you're missing the fact that we went back and forth on it.

Speaker B: Got it. Meaning, like you went into this and you edited it. Came up.

Speaker A: Yeah, exactly. This is not. Yeah, ah, exactly. This is a mix of its writing and mine as we went back and forth and back and forth. Um, and let's see. Um, yeah, like it has noting things that it thinks could be elsewhere in the chapter. It's just like it's giving suggestions without telling me what to do, which is so valuable.

Speaker B: I see.

Speaker A: Yeah. And then like, okay, it's funny, like I, I know where this all wound up. Like this, this giant section. A lot of this is not, uh, is not in the book anymore. But like this quote from Fred that we added here, that was added for the first time right here. This, this sentence is in the final manuscript. I know this because I just read the audiobook. So like, this is extremely valuable, this quote that we, that we surfaced from the research. And like, it's funny, like I know Fred's work extremely well, but I didn't know this quote. This quote is from a, like a paper of his, I think, that I had not read at the time. So it's like surfacing things for me that like, I know a little bit about it or uh, I know adjacent to it. It's like, look, no, you should quote him with this specific quote because it's just the perfect one for this paragraph. Those kinds of suggestions are super valuable.

Speaker B: And the difference with solve it versus like if I'm using Claude or OpenAI with their artifact, uh, section I would be editing on the right and asking questions on the left and we might edit together on the right. But here it's conversation back and forth and the bad part just gets removed because you've edited the Answer.

Speaker A: Yeah, I always, I prune the bad part out of the dialogues with cloud code, with cowork these things. There's a huge amount of work happening behind the scenes that you can't see. And like how many times have you been through like a compaction context compaction in cloud code?

Speaker B: Yeah.

Speaker A: You have no idea what it did. And then now all of a sudden it doesn't remember what you told it to do anymore because it's decided what's important here. The human always decides what's important. The human is in control of the context. So it's slower but it's way better.

Speaker B: I love that you shared it. I love that you did the screen share.

Speaker A: Yeah, no problem. No problem. So you can see. I hope people will when they read the book, I hope they'll be able to feel the level of care and effort that has gone into it.

Speaker B: Here's what I liked about Incorruptible the book. I feel like you're on a mission and I can see it from even the long term stock exchange. I could totally see the mission from there through the book, the stories. And the mission is this. Hey folks, stop caring about the quarterly earnings. Stop caring about the person who's the loudest with a few shares in your company. Think about what it is that you stand for. And I'm not even going to take a position and tell you that you should be for or against anything. I'm just telling you earn or feel, feel good about standing for something. And once you do that, here's examples of all these companies that, that you're proud of because they did that. All right, thank you so much for doing this. It's a great book, well written.

Speaker A: Thank you very much. I feel bad now showing how the sausage is made, but for you and your audience, I think, I hope people will appreciate it. Yeah, it better.

Speaker B: I do. Thanks. Hey, my agent says that if you watch this far, you're going to want to subscribe. And Google thinks that if you watch this far, you'll want to watch that video.