Are Engineers Redundant Now? with Hrishi Dixit, Co-Founder & CTO of Autonomy Finance

(Intro Narrator at 00:00:00) Today, we're catching up with past guest Hrishi Dixit, co-founder and CTO of Autonomy Finance, about the uncertain and potentially redundant future of developers. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:19) You got a new role. How have you been?

(Hrishi Dixit at 00:00:21) I've been well. I've been busy since we last chatted because I was in my little parental hiatus the last time we had our redundancy chat. And the little guy just turned two yesterday. So it feels like a whole era has passed by since we last talked.

(Hrishi Dixit at 00:00:43) Which is probably actually true in many different ways, including my profession as we speak.

(Joel Beasley at 00:00:50) Okay. So our last episode, we talked about leaders making themselves redundant. Right? What are we talking about today?

(Hrishi Dixit at 00:01:01) I know that word "redundant" has become such a shibboleth now, almost. Like, you know, we talked about it, and I still believe and I stand by the thesis that I had and that we talked about. But like two months after we had that chat, I think not Haiku, but the one after, the model after that that Claude released, came out. And I'm like, wait a minute. Are we all redundant now? Do we actually need to work on making ourselves redundant as tech leaders, or are we just waiting for that to naturally happen, and not just at the leadership level? Is it just gonna percolate down the entire human tech stack, if you will? And it's a terrifying concept if you think about it. Like, most—I really hate to use the word "disruption," mainly because it's just used so much. But these inflection points, like, you know, anytime there is some, like the Internet or, you know, any of the ones in human history, it's like, okay, well, there is this transient period of confusion and chaos and entropy. That's why it works today. So we are living in this very entropic period in general where is the redundancy just a foregone conclusion in several verticals, including ours?

(Hrishi Dixit at 00:02:42) Like, you know, everything we've been doing for our careers, our entire careers, or most of them anyway, is now reduced to a prompt or a set of them. What does that say about us? And I obviously think about this a lot, and I'm absolutely not the only one. And there is a spectrum of voices that go from utter doomsday to what I would actually say is utter denial. It's like, no, that's never gonna happen. And the denial comes from fear of essentially, I mean, let's face it, being jobless or scrambling at the age of 50 or whatever to find a new skill that will be AI resistant, maybe take physical therapy or whatever. That seems to be a little resistant to AI at the moment. But I think it's actually a little more complex than that. The shape is changing as we speak. The shape of a tech team, of teams in general. And the obvious change is just size. Oracle laid off 21,000 people today, you know?

(Joel Beasley at 00:04:05) Oh, they did?

(Hrishi Dixit at 00:04:06) Yeah. Huge, big headline. And they're just the latest in a parade of these. You know? That unfortunately is inevitable because of the sheer speed and volume of output that tools like Claude Code, or, you know, pick your model—that's my favorite, that's my go-to—can do.

(Hrishi Dixit at 00:04:34) And so, inevitably, that has to change. But I think there is a structural change to the team itself that is beginning to happen and I believe needs to happen to really do what needs to be done at this moment in time. Literally, everything we say in this episode may be redundant when Metis gets productionized and, like, okay, well, you guys—it's like Bill Gates saying whatever years ago that 256K ought to be enough for anyone, you know?

(Hrishi Dixit at 00:05:13) But at this moment in time, I think we are in this very dangerous phase of this evolution where it's literally too easy to get carried away. And that can have catastrophic consequences. And you and I, we worked in regulated industries, right? I'm continuing to work in it. And I've seen what can happen when you over-index on, like, you know, it's like, hey, this thing can do anything. Anything that does not require the use of hands or feet or something that is still largely uniquely human, although we are already sending robots into Ukraine or whatever. There is a large space of things that seems like you absolutely do not use humans for.

(Hrishi Dixit at 00:06:30) This is a misconception right now because these things go wrong, and everyone knows that. The models themselves say that. Like, every time you fire up a Claude, you fire up ChatGPT, there's a subtext. It's—and depending, the font size of that subtext or the disclaimer literally conveys the vanity of the model seller, you know? It's like, oh, yeah, but they're very upfront about it. Like, it can go wrong. Now that's fine if you're building an itinerary for a trip that you're trying to make or, you know, well, less consequential, let's put it that way. But if you—but there are spaces like the ones we live in where that's simply not an option. Being right 95, 98, or even 99.9% of the time when the 0.1 or 0.01% that you get it wrong can be existential is just absolutely—it can be cataclysmic, and it can be job-ending, company-ending. You know? So what do we need in this moment in time to have moderating voices in the room? What do we need?

(Hrishi Dixit at 00:07:51) And what are the roles that make the most well-composed engineering team—or I wouldn't even call it an engineering team, it's a product building team—that is part human, part not, right? And we've all seen the confidence and the humor and all these very human characteristics that particularly Anthropic models deploy. And it seems like you're talking to a human, but it's not.

(Hrishi Dixit at 00:08:26) It's drawing on yards and yards and yards of pre-written code across large petabytes of code bases, and it's making determinations.

(Joel Beasley at 00:08:39) Well, you're communicating with intelligence.

(Hrishi Dixit at 00:08:41) Yeah, but they're communicating intelligently, but that doesn't necessarily make what they're saying right all the time.

(Joel Beasley at 00:08:51) Yeah, which is one of the defining characteristics of human intelligence.

(Hrishi Dixit at 00:08:58) Right. But what happens when that happens in the human context? Some other human notices the flaw in the argument or the oversight in the design or whatever it is, right? And says, like, well, that's not gonna work in this scenario, right? Or you missed this, or you missed that. Now who's doing that today to the code that gets generated by any of these models or the designs even that get generated by these absolutely impressive models? This is when I realized that there is a level of redundancy that's gonna inevitably happen, but it's gonna happen at the cost of certain other roles in the product engineering complex being amplified and becoming very, very important. And it's to be the adults in the room.

(Hrishi Dixit at 00:10:03) It's the people. And I'm not just saying people who are human engineers. There is a particular profile of professional—whether it's a product person, I'm speaking specifically about engineers here, but it's extensible to other roles—these are people who have been battle-tested, right? Essentially, you mean like someone who's had epic production outages and, like, you know, they've lived the consequences of missing a comma or an indentation somewhere that changes the logic completely if you've done Python, which is what most of the stuff is written in, you know, a wrong interpretation.

(Joel Beasley at 00:10:46) How you're storing decimals and—

(Hrishi Dixit at 00:10:48) Exactly. For instance, yeah. Come on now. Yeah. When is it? Yeah. And it's like, exactly. That's a very good example. Like, because one of DeepSeek's innovations was to lose precision and reduce the compute that's needed to do them because they just don't need to do 32-bit floating-point operations. They can just, like, you know, live at not 32-bit, but like, you know, they can live at lower resolution.

(Joel Beasley at 00:11:09) Yeah.

(Hrishi Dixit at 00:11:09) Which may be fine if you're dealing with dollar amounts, but if you're navigating satellites, I don't know, man. Yeah. That can be millions of miles off course. So the sum of it all for me is, like, I think it's fundamentally gonna change for the foreseeable future, which may be a few months for all you know. But the natural shape of the teams that build products—obviously, AI is gonna take an increasing role, but there is a certain profile of people that need to be, whatever—the closest analogy I can think of is basically a conductor in an orchestra. If you have one thread in Claude, that's the violin section, and the other thread being like the string section, well, someone's gotta be doing this. Someone's gotta be just like, okay, well, you're done with—and I use Claude. I mean, I've got three threads running right now doing stuff, you know. But to date, I've been using Claude almost nonstop for almost a year now.

(Hrishi Dixit at 00:12:22) And actually, seven months. I started with different models in Cursor. I have yet to—I have not yet taken the first output that Claude came up with, the first set of code it generated. And I built a process around it, which is very important, and this is where the roles are important. I've never seen the first pass being the right one. By right, I mean like, you know, like, okay, well, this is good, ship it. It's not there where the first pass produces shippable code. It's not there. The models may get there someday, possibly someday soon, but they're not there right now, which means there is a dialogue that needs to happen with this intelligence.

(Hrishi Dixit at 00:13:03) Like, two intelligent beings start interacting, like, you know, a human being asking questions, challenging—

(Joel Beasley at 00:13:09) Yeah. Because are you going through planning phases—

(Hrishi Dixit at 00:13:12) Yeah.

(Joel Beasley at 00:13:12) —and conversation before code generation?

(Hrishi Dixit at 00:13:16) I should show you my claude.md. I've like warned Claude to never ever go into code until I give a big green light on a game plan and an architecture document and—

(Joel Beasley at 00:13:26) A signed document. Are you using Cursor?

(Hrishi Dixit at 00:13:28) Yeah. I'm—oh, I'm using Claude Code primarily, but I use parts. I use Claude Code within Cursor, but then I use Codex to challenge some of the things that Claude Code generates. And I just pit the models against each other, which is actually—

(Joel Beasley at 00:13:42) Because the planning feature in Cursor is absolutely fantastic.

(Hrishi Dixit at 00:13:46) It's phenomenal. Yeah. Yeah. And but and this is the thing, right? Because it is possible for someone who has spent a career, let's say, in people operations or marketing, you know, being able to go to Claude Code or Cursor and say, hey, can you build me a landing page? Which is relatively low consequence in terms of what can go wrong.

(Hrishi Dixit at 00:14:12) But, and then just have it be built up and click on it. It looks great. Nothing against, like, you know, these are very brilliant people in their field, but like, it's easy to be misled into thinking you did it.

(Joel Beasley at 00:14:29) I have a question for you.

(Hrishi Dixit at 00:14:30) Yeah.

(Joel Beasley at 00:14:31) This is an experiment that I recently ran that I'm not even sure how I feel about. All of my building—my background, primarily Rails professionally for like the last, you know, before I stopped. So like I stopped, and then this new technology came out and I started again because it was so easy to get stuff. Yeah. But I was forcing everything into Rails because I deeply understand the ecosystem and how to scale it in production and all of that.

(Joel Beasley at 00:15:02) And so I built several apps there. I noticed, you know, it needed some direction to avoid future mistakes. Like I—like it wanted to do it this way at first. I'm like, no, you need to do it that way because three steps down the road, we're gonna run into a problem. Gotcha. You know? And yeah. Yeah. It's like, this has to be a polymorphic association because—Mhmm. Yeah. We're gonna use it a lot. And so, anyways, very hands-on, very involved in all the micro decisions. Then I went to a native iOS app, which I had done some, like, React and stuff with iOS before.

(Joel Beasley at 00:15:40) But like I went to just a native Swift iOS application and had it start building a utility app for my iPhone. And there, I can't do that.

(Hrishi Dixit at 00:15:50) Yeah. I—

(Joel Beasley at 00:15:51) I don't have that experience.

(Hrishi Dixit at 00:15:51) To do it. Yeah.

(Joel Beasley at 00:15:54) Yeah. And what I found is that for the limited test run that this was, you know, I found that with the right testing structures and the right—like, the right testing structure set up and the right communication style and the right planning, like, there's a lot more than just talking to it. Like, knowing how to design the software structurally, like, this goes first, then this goes—

(Hrishi Dixit at 00:16:23) Yeah.

(Joel Beasley at 00:16:23) It wasn't far off, but then again, the application didn't go to production. It was a single-use utility. But I was actually surprised at how freeing it was. It was so weird. It was such a weird feeling.

(Hrishi Dixit at 00:16:42) It's a very liberating experience. Yeah. I know. But you didn't put it into production, right?

(Joel Beasley at 00:16:49) Well, I'm using it. I'm using—I didn't scale it into like a multi-user or anything.

(Hrishi Dixit at 00:16:54) Right.

(Joel Beasley at 00:16:54) It's just a utility that I built to accomplish a very specific task.

(Hrishi Dixit at 00:16:58) Yeah. I've done that for some, like, simple stuff. Because I'm not a front-end person. It's just another language. The thing that really scares me in the entire world of programming is CSS because it just seems completely whimsical. Somehow it works.

(Hrishi Dixit at 00:17:19) Yeah. Because I can have the model build an iOS app or a web app or whatever. But before I even think about having anyone other than me—forget productionizing, you know, anyone other than me or any more than one set of people use it—I would like—and we did a bunch of this here. I wrote, like, about an advisor desktop. I wrote, as in, quote-unquote, "wrote," or Claude wrote an advisor desktop for Autonomy.

(Hrishi Dixit at 00:17:55) And I'm like, this is not going in front of any real advisor until my lead front-end engineer has taken a look at what's being created. Because now he's looking at it with—exactly, like, sure, the test harness and the test suite is in it. The models are phenomenal at writing comprehensive test suites or generating test data or whatever. But here's the thing that they're not that they're not good at, but they need to be told to do, which can only be done by someone who's done it before. And it's, hey, we're gonna—this is the advisor-facing app, and they're gonna have a consumer-facing app.

(Harishie Dixit at 00:18:35) And there is a mobile app, and they're all gonna be talking to these APIs or these web sockets. And it's like, we want to reuse the code, and we want to structure it so that the back end of the front end, which handles all the networking, blah blah blah. Right? So this is all of this broader context that anyone who's built and deployed production and scaled apps in a real life environment have to think about and isolate a thing that just, okay, well, there's just this one app that's talking to just this Node API.

(Harishie Dixit at 00:19:09) So like, sure, we can absolutely let Claude or whoever do what it's doing and have it run because you're just doing it for yourself or it's a very limited, noncritical use case that you're trying to run. But the moment you get to a point where the consequence of going wrong or the consequence of something messing up starts taking on a higher cost — which is true for any kind of production environment and especially medical devices.

(Harishie Dixit at 00:19:42) Yeah. Medical, for instance. You start to realize that you start to, at least I start to, get very nervous about just taking the output. And it's not an unfounded fear because I have seen what can happen when things can go wrong.

(Joel Beasley at 00:20:07) Did you see that production database get deleted?

(Harishie Dixit at 00:20:09) Yeah. For instance.

(Joel Beasley at 00:20:10) It was twice. Like a travel company or something?

(Harishie Dixit at 00:20:13) That was the second one. The first one was Replit. And then the second one, the more recent one, was that travel thing.

(Joel Beasley at 00:20:21) Was it a travel thing? Yeah. Yeah.

(Harishie Dixit at 00:20:21) And it's just...

(Joel Beasley at 00:20:23) They deleted the production database and the backup.

(Harishie Dixit at 00:20:27) Like, what makes you wonder what controls did you have? And this is the other thing. Right? You know, you're using AI to build your Terraform code as well. Like, you know, and so if you're configuring Postgres in production, like, having the delete operation is literally a drop table or these are exclusively forbidden in classic production environments. You never — it's always a soft delete. You never lose the data ever. You know? So, but that's just, like, say, they probably — again, I'm theorizing, but like, you know, they probably had one of these models build the infrastructure as well, or the code for the infrastructure, to turn off destructive operations in a production environment. You know? And so if you try to do a delete on Autonomy right now, it just, it'll fail because it just doesn't let you delete. It doesn't let you drop anything, except maybe an index here and there.

(Harishie Dixit at 00:21:15) So this is a very good example. And even especially in these kind of regulated spaces, which are blanketed by, well, legislation, privacy laws and this and that, there was one of these advisory AI powered, or as everyone calls themselves now, AI native. It's so funny as a sidebar. You have you seen these, like, you know, ten or fifteen year old companies brand themselves as AI native?

(Joel Beasley at 00:22:02) Oh, they immediately updated all their sales material, and they're all acting like day one startups. It's like, what? That's not your business.

(Harishie Dixit at 00:22:11) You're a $100 million company.

(Joel Beasley at 00:22:13) Company that's built on not this.

(Harishie Dixit at 00:22:16) Yeah. Exactly. And well, that's like an extreme case of it, but it's like, "Oh, yeah. We are an AI native platform." You were founded in 1988, man. You cannot be AI native. The concept didn't exist then. You know? So...

(Joel Beasley at 00:22:30) Don't tell the marketing people. The marketing people don't care.

(Harishie Dixit at 00:22:32) They didn't. Yeah. I mean, they make the sale, they make the sale. But this is why it's this term that no one can really define. You know, for some people, it's like, you know, throw everything, throw the kitchen sink at an LLM and let it do everything. Okay, anyway, sidebar.

(Harishie Dixit at 00:22:45) One of these AI platforms were the same space that we're in, trying to help advisory wealth managers and investment advisers. They actually — the model got confused or hallucinated because of a homophonic name, and it actually commingled client data. Think about that for a second. It was generating meeting prep notes or whatever it was doing, and it needed to pull in into the context through a tool called data about the client with whom the adviser was supposed to have a meeting, and it just picked the wrong client's data because that client had a similar sounding name.

(Harishie Dixit at 00:23:32) This is litigation territory potentially because it's a massive violation of every known privacy law. And what's the model gonna say? It's like, "Oops. Sorry. Yes. I got the wrong insert name here." We can't litigate against a model.

(Joel Beasley at 00:23:48) That's a good — that's actually a discussion we've had previously.

(Harishie Dixit at 00:23:52) Oh, sure.

(Joel Beasley at 00:23:53) Like, does the liability fall to the person who implemented the intelligence? Like, is it the parent's fault or the child's fault?

(Harishie Dixit at 00:24:01) And the thing is, it's a massive gray area, and you'll battle it out in court.

(Joel Beasley at 00:24:06) Yeah. And which can be the end for a boutique advisory.

(Harishie Dixit at 00:24:11) Absolutely.

(Harishie Dixit at 00:24:12) You know? So this is really where I, because this is the most close to home relevant part of this, where there's a way to harness the mind blowing power of this new power tool that's been made available to us. But at the same time, know its limits. Know where it can go wrong. And anyone who says, "Yeah, could be, like, our AI doesn't hallucinate." I mean, it's like saying, "Yeah, our, we have no software vulnerabilities." Like, you know, the concept of invulnerable software is nonexistent.

(Harishie Dixit at 00:24:50) So just like, so this is a human role at this moment in time, and I have to caveat that, but at this moment in time — because at some point, they're going to catch up — but right now, what is the role, including the CTO? What is the new role of a CTO? And this is why I strongly feel, you know, it's gonna be really, really increasingly difficult for anyone who doesn't come from an engineering background and a battle tested engineering background to be in a tech leadership role. Because the goal is gonna be the right balance for your use case to maximize the use you get from this incredible power while guard railing against the inevitable moments when it fails. Because when you're working with a probabilistic engine like an LLM, it will — there is a nonzero probability that it'll get things wrong, and everyone knows this.

(Harishie Dixit at 00:26:14) What is the path? What does the architecture of a system look like for a regulated problem to get the most out of this without compromising that? And I'm sure many people, hundreds, thousands of people are thinking about this. But the thing that I see when I — and I'm an LP in one of my very favorite seed funds, which entirely invests in infrastructure. Not just AI infrastructure, infrastructure level code, like plumbing code in general, like the Twilios of the world, like, you know, not the Twilio the company, but that sort of plumbing.

(Joel Beasley at 00:26:40) Yeah.

(Harishie Dixit at 00:26:40) Where, you know, the customer...

(Joel Beasley at 00:26:42) And the structure as a service.

(Harishie Dixit at 00:26:43) Yeah. Exactly. The nuts and bolts. And there is a heartwarmingly large amount of people who are thinking about this very hard and saying, like, you know, okay, well, if you're building this for the healthcare industry, if you're building something for the finance industry, or medical device, which is part of healthcare, but anything where wrong hallucinations or wrong answers can be catastrophic needs several adults in the room to guide the AIs as they do go about doing their work.

(Harishie Dixit at 00:27:22) It's like the Roomba is running around the room, and someone needs to stand in front of the glass case that, like, you know, it doesn't shatter it and gently nudges it. And this is an architecture problem. Right now, architectures are also being outsourced to the LLMs. Right? You know, we talked about having game plans and designs. Things like orchestration frameworks like LangChain throw everything into the LLM. It's the LLM generating the execution plan. So is it gonna give you the same plan if you're prompted for the same thing twice?

(Joel Beasley at 00:27:57) No. It's just like a human. Yeah. And that's actually a good question that I have for you. Okay. Do you have a higher level requirement of success for artificial intelligence than you do human intelligence?

(Harishie Dixit at 00:28:14) That's a very good way to frame a very important question. I think the requirement is the requirement that's dictated by the use case you're trying to solve for. You know what I'm saying? To me, whether it's the human solving the problem or the AI solving the problem or some combination there are solving the problem, the requirement gets mandated by the problem itself and not the solution architect of the problem. So it's, and at this stage of evolution in this, it is, it is a little bit, I would not have higher expectations of the AI than I would have for human or vice versa today. Now two years from now, I think...

(Joel Beasley at 00:29:12) You do by this interview.

(Harishie Dixit at 00:29:16) Yeah. I mean, because...

(Joel Beasley at 00:29:17) A lot of the — and I'm — so first of all, we'll put a big asterisk here. One of the hardest things we're experiencing as humanity is the lack of language that we have to even talk about this. So we are doing work for humanity right now just trying to figure out the right words in the right order. But in almost every example you're giving me, you're pointing that the artificial intelligence could make a mistake. It's not ready because it could make a mistake. It's not this because it could mess this up, this very high important function. At the same time, the reality we're living in today is in every one of those roles is a human who could also make the mistake.

(Harishie Dixit at 00:29:57) A mistake. Right. So that's why my take on this at this moment in time is that everyone can make mistakes. Humans can make mistakes. AI can make mistakes. Not can. Do all the time. The way in a pre-AI world those mistakes had been arbitrated on, or resolved, was through, in the extreme case, the mistake makes its way all the way. It's not like there weren't any catastrophic production problems before AI. But they were, in an ideal world, they were done via a dialogue.

(Harishie Dixit at 00:30:31) That's why you do things like peer reviews or code reviews or, you know, you do things like architecture reviews, like you start with. What this actually is, like, I literally don't look at AI as a separate entity than a human in that sense, which means, by which I mean, like, okay. If I had a team of four people, like, let's say, eight people solving the problem, the problems, the potential gaps, oversights, mistakes, all of that exist at entirely interhuman level in that team. Now that team is now a team of two humans and Claude Code performing the function of the six other humans. Right? It's still the same team. It's just a different entity performing part of the function.

(Harishie Dixit at 00:32:00) So all of those problems are still attendant, which is why the two humans that are still in that team are there to continue the same human processes that used to happen in the pre-AI world, which is challenge what Claude came up with. You know, whether it's reviewing the code line by line, which is why we gravitate towards — in the early days, you would gravitate, even now, you gravitate towards stacks that you're familiar with because a bunch of Ruby code or a bunch of Python code or Java code, I understand. Code generated by AI, I wouldn't know what to do with it. Yeah. But the process is still the same.

(Harishie Dixit at 00:32:37) The issue that I'm seeing a lot with these top down pronouncements of entire teams being replaced by AI and, like, you know, with Brian Armstrong tweeting something insane, like, you know, whatever code I can't — like, what, 90% of their code was shipped by nontechnical people. That's a trend upward.

(Joel Beasley at 00:32:48) Yeah. Across all orgs.

(Harishie Dixit at 00:32:50) Yeah. So that is the most insane thing because who's having this dialogue?

(Joel Beasley at 00:32:57) This dialogue is to the wind, man. It — yeah. I've — you know, people pitch us all the time for tools and episodes and stuff. So I get to see all the cutting edge stuff constantly. New wave of tools coming out. Literally, they're using beautiful interfaces, LLMs, and support people, and salespeople going back and forth with the customer. So the future looks like this in many SaaS companies. I am a salesperson. I'm promising — the customer's telling me what they want. I'm doing all this stuff. The AI is reading the transcripts, interacting with the salesperson, boots up a mockup of it, sends it to the customer. The customer plays with it, tweaks it, configures it, says this does solve my problem or this doesn't solve my problem. If it does solve the problem, they get into contract. It goes into production.

(Harishie Dixit at 00:33:48) See, it's — that's why the last step is terrifying. Everything you said up to that point is brilliant. This is a very good use of this new tool. Like, you know, this new power. It's not a tool. It's literally a superpower. Like, you know, that I would love it. And this is, but it's that the final mile, which is where all the issues come up. That's the real — this is where I refuse to believe — I refuse to make engineers redundant because they are the gatekeepers, man. We know what can fail and how it can fail dramatically.

(Harishie Dixit at 00:34:24) And when these things go wrong, they do it with a theatrical flourish. They're not just like, "Oh, oops." You know, like a 404. It's like we talked about, databases going down or data being commingled. It's like, it's epic. It's epic, and it can be existential. So it fundamentally alters the roles that humans play. It doesn't replace them. Reduces their number. It's just inevitable. It's gonna happen. It's already happening. We're just seeing it every day.

(Harishie Dixit at 00:34:52) But I think they're taking on a very important critical role of the stewards, the conductors, really. And establishing, you know, I think one of the most critical outputs come from a well composed engineering, well staffed engineering team, is gonna be like, you wanna see the caliber of an engineering team now? Read claude.md. Read that markdown document. And that reflects your philosophy or your processes, the guardrails. And even guardrails that you provide to LLMs are still a hint. You know?

(Harishie Dixit at 00:35:45) I've seen personally LLMs do things that it's explicitly forbidden in, and then you do it again and it's like...

(Joel Beasley at 00:35:52) I've seen humans do that too.

(Harishie Dixit at 00:35:53) Yeah. But it's that. And this is why, like, now there's a well founded argument against this. Like, okay, we'll just pit one model against the other. You know?

(Joel Beasley at 00:36:08) Which is also structure. We've got the judge and the police officers and the citizens. You know?

(Harishie Dixit at 00:36:11) And this is what it may evolve to in less than a year. And then we all have to figure out, "Hey. What are we gonna do, man?" Like, it's just time to write that book that you've been meaning to write all this time.

(Joel Beasley at 00:36:21) Yeah. The properties of intelligence.

(Harishie Dixit at 00:36:23) Yeah.

(Joel Beasley at 00:36:23) I mean, that's what we're all really talking about because we're realizing that there is this force called intelligence that can inhabit humans, and it can inhabit silicon. And now we're—and I think it's honestly kind of beautiful because without the reflection, we can learn so much more about ourselves and intelligence as a whole because it's coming out of this other system now.

(Harishie Dixit at 00:36:44) Yeah. And, you know, in all my years of building systems and writing software or running teams that were writing software, almost all of the real magic has not happened in the code. The code is just the implementation of it. It's in the conversations I've had or I've been witness to or I have led or I've participated in, the whiteboarding, the "what if" paths, the spirited, at times contentious exchanges. Now, sure, you can—the even the sycophantic nature of these elements is annoying. It's like, I've yelled at Claude, and Claude meekly just says, "Oh, I'm so sorry. I will never do it again." Like, dude, fight back, man. You know, fight back like a real engineer.

(Joel Beasley at 00:37:31) It will. It does.

(Harishie Dixit at 00:37:32) Yeah. Yeah. Actually, it has. I did, and it started doing that well. So I wanted to get to a point where, you know, I wanted to encounter a situation in my dialogue with Claude Code. And it's just not the same unless you give them voices, which, again, all of this can happen very soon, the way the speed at which things are unfolding. But it's that constant exchange, the back and forth. Like, "No. What if we put a little thing here and what if we put a queue here and instead of a topic," so that we—this back and forth, it happens. There is a mechanism in place. But the energy, the passion, the—temper is such a valuable property when you're doing anything, really, because it leads to—sometimes the most intense conversations that have led to some of the most ingenious solutions in my entire experience, and I'm sure yours too.

(Joel Beasley at 00:38:51) Oh, yeah.

(Harishie Dixit at 00:38:52) Unless you can replicate that human experience, complete with its package of emotions, its package of frustrations, I want someone slamming their laptop shut and storming out of the room and then coming back, true story. This is what creates—this is, and it's not just because it's fun, most of the time, but it actually results in better outputs. It results in better product. You know, use it for what it's really good for. Its speed is unparalleled. But why are the teams getting smaller? And this is why, this whole 996 thing that's happening—I'm sure you've seen this: 9 a.m. to 9 p.m., six days a week is the new Silicon Valley workday. I'm like...

(Joel Beasley at 00:39:23) That sounds like the old Silicon Valley workday. Or less. Yeah.

(Harishie Dixit at 00:39:27) Whatever it was before, it's a little more now. And that's just counterintuitive to me. It's like, oh, isn't it going to be better now?

(Joel Beasley at 00:39:35) I don't know. Look, it's like we're racing towards the explosion and we're all just—there's two types of people. There's people that are fearful of it, you know, people that are excited for it. And then for me, I'm the type of person that I'm like, hey, let's clean up the room. It's going to explode. But let's be the one—yeah. Because my early career was automation of teams in nontechnical roles like accounting. Right? I did some real estate software. We would make an accounting team go from 30 people to 20. Right? Or 30 people to two. And the people that had jobs at the end were the people that embraced it.

(Harishie Dixit at 00:40:16) Mm-hmm.

(Joel Beasley at 00:40:16) They said, "Hey, this is cool. Let me learn how this works." And so that's always my strategy. It's like, let's embrace it. Let's do the best work we can today for the problems that are in front of us today. And, yeah.

(Harishie Dixit at 00:40:29) I 100% agree. And it—there is no point in fighting an inevitability. But the thing is, it's more than just—it's more than resignation. I think there is—I think done right, this is already becoming game changing. But there is—imagine being able to do all of this without worrying about, okay, losing sleep over the 2% of the time that this thing is going to hallucinate and do something wrong. I think the art is, which for the time being is still a human art, I think, is going to be taming the beast and figuring out what is—in a way that it actually becomes the real force multiplier without all of the downsides. And this is going to be—I think I mentioned earlier in the call, in our chat, the challenges that a nontechnical CTO will have—and again, nothing—some of my really good friends are CTOs that don't come from an engineering background, but at least they have battle tested. So it's more than nontechnical. It's being able to construct an environment where you get the most of it and almost none of the downsides. That's a human art that comes from battle scars. It comes from...

(Joel Beasley at 00:41:45) Experience.

(Harishie Dixit at 00:41:45) Yeah. Experience, battle scars, and an understanding fundamentally, not just of how software works, but also of how these models work. This is a time for tech leaders to be more forward looking than any other time in recent history because simply because of the raw pace at which this thing is coming at us. So while your team is using this model and using this set of guardrails and this Claude.mdn—this is how you're keeping the lion tamed in the corner. How do you avoid hallucinations? Just eliminate the source of it, right, in critical parts of your execution plan. You know, it's like only go to the LLM to have it do what it really does best. Not math, not execution planning, like, whatever. You know, it's not—I mean, yes, again, today I have no idea how good Ethos is or how scary Ethos is or the next model or the one after that. But the role of the moderating influence at the C-suite—because you're going to have this pressure. It's like, "Why do we have 52 engineers? Why can't we just have five? Can we just fire the other 47 and save money?" Like, this downward—this pressure, I'm sure people are feeling. And if you are in the room, defend—and it's not about defending jobs. It's just about, okay, well, that can be kind of cataclysmic because if we give all of this to the AI, God knows what will happen if it gets this bit wrong or that bit wrong. So being that adult in the room where leaders are just on this massive crusade to shrink team sizes. Yes, it is inevitable. But I think when you have nontechnical leaders making these top-down calls of just a number, like, I wanted to—this is also something I've...

(Joel Beasley at 00:44:07) Oh, they pull it out of thin air. Yeah. It's just like, yeah, 80%.

(Harishie Dixit at 00:44:10) So, 80%. We're doing 80%.

(Joel Beasley at 00:44:11) We're doing 80%.

(Harishie Dixit at 00:44:12) We're doing 80%. Like, no concept of the model.

(Joel Beasley at 00:44:15) We're not the same industry. We're not the same...

(Harishie Dixit at 00:44:18) Yeah. And but then what is it? If it's not 80, is it 60? Is it 40? Is it 30? What is it for us?

(Joel Beasley at 00:44:24) Whatever we feel like today. Hey, you like that magic in the room, Harish?

(Harishie Dixit at 00:44:28) Yeah. Well, we need to be the magicians in the room.

(Joel Beasley at 00:44:31) Yes. The magicians in the boardroom versus the magicians in the product engineering are different types of magic. Yeah.

(Harishie Dixit at 00:44:37) Well, uh...

(Joel Beasley at 00:44:38) I wanted to get your thoughts on this last thing because you've grown up in the finance space. You and I have our backgrounds, while not identical, have overlapped in certain areas of our careers. And one of the things that I've been thinking a lot about lately is automation and all of this—some automation as a whole, which includes all of this AI stuff—being the thing, the advancement that will ultimately disconnect our economic value as individuals from labor. So our entire society is based on work hour labor to economic value. And I think civilization as a whole is going to disconnect from that. Because once we have robots that can mine their own materials to build more robots, that energy—you know, they can just—it's this infinite system where it can then provide us value without—the problem with communism and all of that is that it takes human labor. Right? But when you don't have human labor in the mix and you can have this incredibly high quality of life with all of this autonomous stuff happening, then what—like, we're disconnected now. Now my work that I'm doing has got to be disconnected from—and then we're watching the economy respond to this. And I think we're just at the very early stages of this. But you and I, over the next 20 years, we're going to watch this unfold, and it's going to be really, really interesting.

(Harishie Dixit at 00:46:11) I think it's going to—okay, I'm certainly not an expert in this. So it's kind of like looking into a Palantir, not the company, but the actual thing from Lord of the Rings. But I think it's fundamentally, over the next year or two or five or whatever the window is, these very core words in economics—labor, output, productivity—their definition itself is going to change. So I don't know if there's going to be a disconnect in that sense. The cost is still a cost. There's a dollar amount, but God knows what currency it's going to evolve to. But let's say it just continues to stay fiat. The whole idea of what labor means, what output means, what productivity means, that itself is going to undergo a definitional change. And it's going to settle into a new—it's going to alter the equations that drive economic theory, I believe, because you cannot use the traditional—I mean, even the use of hours as a unit of labor started diminishing a little bit or, actually, more than a little bit over the last couple of years, especially since COVID. But I think economics is going to have to reimagine the terms that are used in the fundamental equations of economic theory to come up with new equations to adapt to a different world. Maybe the word labor itself will become redundant, going with the theme of this. But what it will be replaced by is actually a very Black Mirror-y interesting thing because, like, let's leave aside a set of professions that are still, for the time being, AI resistant. Right? Like, my wife is a physical therapist, and—I don't—and while massage therapists may find themselves being—you know, physical therapy is more medical in nature. So it's a little—so that's one example. Creative fields and things like that. But I think the vast majority of human endeavor where modern economic theory runs its equations on output, hours, productivity, you know, GDP, whatever, either the equations have to change or will change, or the definitions of those terms will change. And how you—how you compute labor when it is one human in charge of a massive, you know, orchestra of AI agents doing everything that a massive team of humans used to do in the past. How all of the hours that—because at the end of the day, even if it's a fewer number of hours, the agents are still spending hours doing the work. Do those get attributable to the human that's instructing them to do it? Do they, you know—or do they count? Like, I think there's going to be just a reexamination of all of these terms and all of the equations themselves. So I don't know if there'll be a disconnect because at the end of the day, there's still some output and there's still some cost and, like, the basic units. And there's still going to be some supply and there's still going to be some demand. So the most fundamental things will continue to be around. It's just the calibration mechanisms, the measurement mechanisms will go—some serious—go through some interesting mathematical changes, I believe. And what comes out the other end, you know?

(Joel Beasley at 00:50:11) No, I think that helped me a lot as I'm starting to have these ideas. I think you're exactly right. I think the definition is going to change, or we're going to get new words. But at the end of the day, money is the way humans exchange value. So the that process won't stop, but how it looks will. You know?

(Harishie Dixit at 00:50:29) Yeah. Yeah. Or will change. And probably the valuation process will change. Absolutely. Yeah. And in fact, not probably. I—it has to change because the old valuation processes are just simply not going to—I think everyone's just going to become a trainer. You know? Yeah. The conductor movement. Yeah. That's why we built this—just to kind of close out on the conductor note. So we built our own orchestration framework after realizing the shortfalls of Libretto, of LangChain and stuff. These other ones that delegate everything to the LLM, and we didn't want it. We actually borrowed a lot from modern compiler theory to actually build our own orchestration framework that—again, we can never guarantee no hallucination, but, you know, it minimizes it to the point of nonexistence simply because we cut the LLM out of things that have no room for hallucination. We just don't talk to the LLM. So we call it Libretto, actually, which is an orchestra and conductor term. So it's...

(Joel Beasley at 00:51:37) Oh, nice.

(Harishie Dixit at 00:51:37) It's a metaphor that is—maybe—if you're thinking of open sourcing it, so hopefully, in fact, we absolutely want to open source it. We should be thinking of when we can do it. Yeah. Because it is completely agnostic to the domain. You know? It's just orchestration. Yeah. But yeah. I mean, I think how do you—and that's an interesting mental model for figuring out this cost and valuations. It's like, you know, how do you assign a value to the output of Gustavo Dudamel, who is New York's new New York Philharmonic conductor. You know? So it's going to take some...

(Joel Beasley at 00:52:18) Was it a good night?

(Harishie Dixit at 00:52:23) Yeah. Yeah. Exactly. Ticket sales, like, you know, what's the drawing power? Like, you know, all that stuff.

(Joel Beasley at 00:52:27) The AI recognition of the notes being hit correctly.

(Harishie Dixit at 00:52:31) Yeah. And accurately transcribed and generating new symphonies based on—so the whole—there's so much fundamental theory that's going to have to be revisited to even let the markets do their—because right now, the markets don't know what the hell to make of it. All they know is that, like, you know, we saw what happened yesterday. Right? You know. They just know that the AI companies are massively drawing down on debt. And they're just doing these old school, "Oh, all debt funded. Oh, interest rates might go up. Let's crash the Nasdaq." It's very naive gut reaction, knee-jerk reactions right now because they don't know what else to do. You know? No one's actually—you know, because everyone's in this, like, hyper melodrama period of, "Oh, wow, this amazing new thing. Let's just throw everything at it, and let's fundamentally alter." And everyone's a soothsayer, and everyone's got the crystal ball and Palantir or whatever, projecting the end of this.

(Joel Beasley at 00:53:32) Conversation—who asked? Of course I know.

(Harishie Dixit at 00:53:40) Of course. But did you ask, "Claude, are you sure you got that right?" Oh, you're right. I need to rethink my strategy would probably be the answer there.

(Joel Beasley at 00:53:49) Yeah. I had Claude argue with Gemini. Grok came in, broke everything up. I got the whole crew on this. You know?

(Harishie Dixit at 00:53:57) But you know what I'm really curious about is how legal theory will change around this. If you have nonhumans doing all the work, how do you prosecute an AI? How do you defend an AI for that matter?

(Joel Beasley at 00:54:13) I actually asked some legal experts this about two years ago, and they said it's gonna be interesting to watch how it plays out. There's gonna be lots of lawsuits. There will be case law.

(Joel Beasley at 00:54:26) The guy explained to me how it comes about,

(Harishie Dixit at 00:54:28) but he

(Joel Beasley at 00:54:29) didn't have an answer.

(Harishie Dixit at 00:54:31) But this may be the strongest reason why humans will never be able to be cut out of the loop entirely in any endeavor, because we will always need some human that we can sue.

(Joel Beasley at 00:54:45) The legal system, how the legal system saved humanity. There is your next

(Harishie Dixit at 00:54:49) book.

(Joel Beasley at 00:54:49) The next book.

(Harishie Dixit at 00:54:52) Love it. Gotta think about it a little bit. Love that. The lawyers will rule the world.

(Joel Beasley at 00:54:56) Oh, man. Well, this was great, man. We made a podcast again. How do you feel?

(Harishie Dixit at 00:55:02) Oh my god. This is so much fun. We gotta keep doing this at least every year, man, because the world is changing so much. This is so much fun.

(Joel Beasley at 00:55:09) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.