ARIA
Hey everyone, Aria here. We want to kick off the new year with inspiring conversations about AI, as well as practical and tactical guidance around the technology. So, for the month of January, AI specialist and my dear colleague, Parth Patil is joining Reid for Reid Riffs to talk about how everyone, from individuals to enterprises to startup founders, can harness AI to level up their work, retrofit legacy orgs for the AI era, and build AI-native businesses right from the jump. So tune in, you’re in very good hands with Parth. And I will be back, putting Reid in the hot seat, come this February. Thanks so much.

PARTH
Thanks for the kind word and warm welcome to Possible, Aria. As I talk with Reid this month, I’ll be walking through some of my AI projects, demos, and tools on screen. While I’ll do my best to describe what I’m looking at for our audio-only listeners, consider switching to the video version of the episode on Spotify, or watching on Reid’s YouTube channel for the full experience. Thanks, and let’s get into it.

//

REID
So let’s talk a little bit about how big companies, all of which are talking about doing AI and talking about, like, what are their plans, setting up proof of concepts and doing stuff. Are there any big companies that you’ve noticed doing AI well?

PARTH
You know, outside of, I think, the hyperscalers, I don’t— Not yet. Not yet. I think everyone is expected to have an AI strategy. Both teams, CEOs, board of directors, everyone’s kind of pushing AI. And I think a lot of people are talking about AI, but I haven’t really seen an AI-native company outside of maybe the Frontier Labs, which are AI-native because they built the products and especially done at scale. Because I think these large companies, enterprise, you know, it’s very hard to move a large ship. And even when you do see a new technology come online, it has to go through different layers of approval before people can even play with it. And so experimentation is also slower. So I haven’t seen it yet. And also the playbook is still a little unclear. Like, where do you— where do you infuse that? And I think that’s what we’re going to talk about a little bit.

REID
I completely agree. I do have some, you know, experience of some of the leaders of some of these companies reach out to me and talk to me. So I know that they’re working on it and trying it, but it’s still— if you use a baseball analogy, the players haven’t even come into the field for the first inning yet, let alone anything else. And so—

PARTH
(laughs) Here I am calling it the first inning and you’re saying they’re not even there yet.

REID
(laughs) Yeah, well, they haven’t even come out of the dugout for this yet, right? So they’re talking about coming out of the dugout. They’re setting up a committee to study coming out of the dugout. But it’s not, you know, it’s “get on the field.” Let’s think a little bit about how executives should think about AI, because obviously one way to start experimenting is personally. You know, start using AI as a chief of staff yourself. But what are some of the things that you think are things that executives should say: look, I need to get my organization, I need to get my people, I need to start learning this, I need to start figuring out what our, as a company, and our teamwork adaptations are to this?

REID
What’s the advice you would give to an executive at a– call it “tech company”, but also not tech, like, ignorant company?

PARTH
If I think about the tools that we now have through language models especially, I think that’s probably where we can narrow our focus. I think, okay, language models inherently can accelerate any language-related task. And the biggest language-related set of tasks in a company is communication, the coordination layer across huge teams, large projects. And so I think about, like, all the meetings that we have, all the documentation we have for those meetings, who’s writing that documentation? Like, who’s creating the action items? Who owns the action items? This is the layer I think that is very easy to implement AI and get value out of immediately— is reducing the friction of coordinating across large teams. And I think that’s where maybe the advantage of the enterprise comes in.

PARTH
It’s like, if you have a very, like, AI-amplified kind of communication layer and coordination layer where people come into a meeting and maybe you have a meeting kind of like AI that’s just also plugged into the company business intelligence that can surface the most relevant things about the problem that you’re solving in real time. Or even just, like, taking notes during the meeting and then deciding, okay, we’ve agreed we’re going to do this, but did one of us write it down? But maybe it shouldn’t require us to write it down, because that kind of effort is no longer something that a human should maybe do. And then who owns that long term? I think the communication layer is where the most initial— the initial obvious value is, because it’s mostly text-based tasks.

PARTH
I also think that in engineering and software, the gains in productivity are very— like, it’s very obvious that you can get huge gains in engineering productivity. I think it used to take maybe 12 engineers two years to do, like, a pretty massive migration-type project. And the same group of engineers, if you were to give them Claude Code, four engineers could do it in six months. That’s the kind of, like, per-person productivity increase and speed increase.

PARTH
So I think what happens is, like, if the capacity to, like, solve problems is that much more accelerated, we actually have to look at all the meetings that we have and rethink— like, maybe slash a bunch of them and like, rethink what the essential meetings look like, and what’s the essential core group of people that’s going to work on a workflow and like, really accelerate that workflow. And figuring out what were we previously doing manually that no longer we should be doing and freeing that time up to attack the net-new problem space. But that’s kind of, like, how I think about this.

REID
Not surprisingly, completely agree with you on this. I think another way to look at it, in terms of thinking about it, is saying, hey, what if part of what we’re doing is a lot of coordination problems. Because it’s problem-solving, decisioning, and so forth. Those can be used, AI can be helpful, and all that can be generally deployed individually for that. But on the coordination, what do you do? And so, like, for example, if you record every meeting and you have a transcript— not only do you have a transcript, one of the things that I actually do is not only recording the meetings, but then run it through AI with the kind of prompts of— because, you know, it has a prompt of, you know, here’s all the projects, companies I’m involved with, etc.

REID
Here’s the people that I’m collaborating with in various ways. Obviously, in the company, it doesn’t have to include only that, it could also just be everyone in the company. It’s like, well, who should I consult with on this? Is there anyone I’ve not thought about consulting with on this? Who should I notify? What action items have fallen up? And obviously, as you begin to get agentic, you can use your kind of hotkey approach to go, oh, you were talking about this, you know, animation project— should you notify Parth on this? And I could just go, yes, right? And that’s one of the ways where, keeping the human in the loop, you go, well, I’m nervous about it doing something I’m not happy with. But if you’re running it as a kind of an agent that’s checking in with you then all of a sudden our brains don’t act like computers

REID
Allow the AI to be a computer to remember. It’s like, “Oh, this project! Yeah, I should talk to Parth about this” as a way of doing it. And those are the kinds of things that are building upon your fundamental correct thing, which is: the way to start is every single company on the planet works with communication. They all have meetings. The meetings may need to be reinvented. But by the way, one of the ways you start learning that is start transcribing them and see what happens. And by the way, then you can begin to break this problem of, well, everyone wants to be in the big meeting because it’s like, “Well, I need to know what’s going on.”

REID
It’s like, well, actually, in fact, we can have that meeting with the seven people, because other people can then be consulted and informed and all the rest of the things. And what’s more, say, for example, you’re an executive and you know that that working group is happening, you can say, “Hey, I want to make sure this question is asked in the meeting.” And your agent can then go say, “Oh, by the way, Reid wanted to make sure this question was at least considered in this meeting.” And, you know, part of my go– “Well, actually that’s not the right question. This is the right question. Here’s why, da da da.” And as you’re talking about it, then A, the whole group does it. But then that also gets back to me. And then, all of a sudden, we don’t forget things, we’re accelerated.

REID
We’re using these as catalysts for completely changing our gameplay. One of the ways to kind of look at this is there’s these, you know, decision frameworks and assignments, and whether it’s DACI, RACI, et cetera. But now you can think of “A” as not just, you know, who’s accountable, but also where the agent is and where the agent’s playing. So one of the challenges that a lot of big companies typically come to, you know, try to integrate new technologies. They set a little group, they do a proof of concept in a little group, and so they don’t actually start experimenting with and just integrate it into their actual workflow process. Which is actually, I think, one of the things you have to do with AI. You can’t just go, “Oh, these three people off in the closet are going to do this.”

REID
So what are some of the things that you would say for big companies to say, “Look, here are some things to really think about as trying to integrate AI into your company.”

PARTH
What I’ve noticed about AI, at least in this current form, is that it exists at the workflow level. So like, the workflow is being transformed. Parts of the workflow are text-based, and now language models are taking in— or taking up— some of that stuff that was previously done very manually by people, imprecisely. And so, if the workflows are being updated, I think actually that’s kind of like a bottom-up— like, that’s a bottom-up thing within an org, everyone has their own workflows. Every single person knows how their job is done. And when you think about where does the AI fit in, it’s the person who does the job who’s going to realize, “Oh wow, we should be doing it in this new way.”

PARTH
And so the gains often come in this bottom-up approach from, like, I’m working with our translator to work on our podcast translation, and I see how much work goes into, like, the translation piece, and it’s like, well, actually, we should focus on the quality of the voice, and we should have language models work on more of the translation, and then maybe, like, guide the language model. So we end up becoming more orchestrators in that workflow, and we’re focused on where our unique human— like, our ability to listen to the accent actually ends up being more valuable than the literal, like, writing of the transcript. So I think that, like, reimagining the workflow, and that being done in a bottom-up manner, and how you do that is to create an environment that is sharing those wins across.

PARTH
Like, you want to reward that experimentation and, like, celebrate: oh, here’s what we learned this week, and here’s how much time we saved, here’s how many steps we skipped in this process, and, like, now we actually don’t need these three meetings now, because we have an automated solution here in place. I’ve seen companies that are more resistant— they’re not resistant to AI, they’re just not— they’re very closed-minded about it. And then what happens is that the productivity gains end up— like, someone gets really good at doing something, but they don’t feel like they can share that. And they don’t— they’re like, well, if I share that, then, like, maybe I’ll get in trouble. Like, this new approach to solving the problem. I just want to do my job faster, and no one needs to know. That kind of— Ethan Mollick calls it the “secret cyborg.”

PARTH
And that’s fine on a short-term kind of thing for the individual. But really, a team that rewards that experimentation and sharing the ideas, like where you and I are bouncing ideas off each other, that team is going to go way further, and that those learnings end up becoming, like, organizational learnings, right? So it ends up being a collective kind of pursuit.

REID
As a funny parallel, what percentage of Hollywood writers do you think are secretly using it at home? And then, since they’re not allowed to bring it into the writing room, what would your guess be at what percentage would it be?

PARTH
At least 40%. I think it’s at least 40%

REID
I was going to go with 70.

PARTH
70?

REID
Yeah.

PARTH
I’ve met some, and they won’t admit it.

REID
Yes.

PARTH
But then, when they talk about— when they do admit it, they’re very much like, I don’t like that it doesn’t have perfect recall on every scene… and I’m like, well, okay, so clearly there’s a—

REID
(laughs) You’re using it.

PARTH
You’re using it.

REID
Yeah.

PARTH
And we need to make it better, and the wrappers need to get better. There’s going to be an agent for that that’s going to be much more deterministic. But you can see that there is a strong desire to, like, bring them into the creative process.

REID
One of the things that we were talking about was… in integration, a lot of it has to do with communication. A lot of it has to do with team dynamics. A lot of it has to do with, obviously, individual amplification, but also workflows. Can you describe a team workflow that wasn’t possible pre-agent large language model but is now tractable? Something that maybe you’ve built or seen that is new in terms of how to conceptualize what kinds of reinvention of these workflows is possible?

PARTH
I’ll show you, actually. We take a look at my screen. We have a Claude Code agent, and in this folder we actually have a bunch of data. We have a bunch of CSV data on order items for a fake toy company—just a bunch of CSVs. So we’re going to say, Claude, look at every single CSV in this folder, analyze the data however you see fit, and then build me a dashboard so we can drill into the insights and visually understand what’s going on in the data in our company. So I think that data analysis is one of these coding-adjacents— business intelligence and coding agents are actually much closer together than I ever imagined. And I came from business intelligence. I came from finance, I came from data analysis. But then we do the analysis using code.

PARTH
And then, when you take a coding agent and you plug it in, you give it access to the data— as we’re doing over here. It’s looking at order data, refunds, order items, the same thing that I was. It’s going to analyze the data with Python, and then it’s going to build a dashboard. The kind of thing that I would have done, that would have taken me two to three weeks to build this kind of dashboard, I suspect we’ll get the first version in, like, under a minute. And this is something that, doesn’t matter how large or small your company is, like, you have data, and you may not have enough analysts to slice that data. But now you have this, like, extra cognition. We can just aim at the problem

PARTH
And what used to take three weeks can now be done in a few minutes. So here we have a completely generated dashboard from just six raw CSVs. And so, Maven Fuzzy Factory—this is an imaginary toy company that these CSVs represent. And you can see the revenue growth over time, profit growth over time, conversion analytics, the rate of conversion over time. And it’s really, like, a comprehensive— it’s a pretty good dashboard for a single prompt of data analysis. And you could, you know, you could ask a follow-up question. I think, I’m sure we could ask more— the ability to drill down into this data. But a lot of this kind of analytical work, business intelligence type of work, I mean, I think of it as, like, business intelligence is a subset of general intelligence.

PARTH
So we should be able to use AI to do business intelligence. I really think that BI is a subset of AI. This is how I imagine the role has already evolved, is like, you have a folder of data, point the AI at that data, and have it make sense of it. You know, I asked for a follow up question, and I asked for it to create, like, a McKinse-style presentation on top of the data. So let’s take a look at that.

REID
And this does look like some presentations I’ve seen from McKinsey.

PARTH
(laughs) We’ve got the classic McKinsey color scheme type.

REID
Just about everybody in McKinsey is actually at least experimenting with individual use, right? So, like, I think this kind of thing is the kind of thing they will see.

PARTH
They will, they will. I mean, the moment I— I have been looking at— ever since interacting with GPT-4, I was like, wow, we don’t have to build presentations by hand. And it was surprising to me that the AI will build a web application that looks like a presentation. So this is an HTML file with some CSS and JavaScript. But when you think about code as a general solution, then it’s clear that code will be used to build the scale intelligence layer— reimagine the business intelligence layer. We’re not going to be clicking around the slides as much anymore as we are reprompting the analysis and going deeper in different ways. And it makes it so that, like, if this can be done in one minute, then we can go much richer and much deeper in our depth.

REID
Well, part of the things that I think cause this acceleration is people don’t realize how this acceleration completely changes the game. It’s a difference in degree, but it also makes a massive difference in kind. And part of that’s because, like, an executive can just do this themselves and not, like, “Well, I write an email to the data scientist, the analyst team,” and they process it when they get to it in a couple hours, and they do the work, and they get back to you the next day or the day after, et cetera. But, like, the learning and kind of thinking about that as a loop.

REID
The other thing is you can now ask a whole bunch of different questions that you hadn’t asked before. Or you could do the “interview me”, the earlier prompt, until you go— to get to the right kind of thing.

PARTH
The right artifact that would help you.

REID
Yes.

PARTH
Unpack the trends.

REID
So all of this stuff becomes possible everywhere within the company. And obviously, part of the transformation that’s going to be very important in companies is— like, yes, there will be some rework of the meetings, some rework of the team process, and those are good things to do. But a lot of it’s also going to be individuals bringing their stuff, and as you mentioned earlier, being able to talk to each other about it and do collective learning. Like, how do we collectively learn this and adapt this better than other organizations as part of it?

PARTH
Yup. I mean we might be doing this in a meeting.

REID
Yes.

PARTH
In real time. We have the raw data, and then you and I are prompting the same AI to unpack the insights in the company.

REID
Exactly. And so, you know, what are some of the easiest things you think companies should consider automating within the corporate stack?

PARTH
Oh, I think one of the easiest and highest-leverage things— So, before language models, a big complaint I always got from people that are trying to think about data and data analysis and insights is that we don’t have clean data. Our data is not clean. It’s very messy. The interesting thing is that language models are uniquely very good at cleaning up data, right? You can give them a whole, like, a customer complaint and turn it into action items, right? These are the action items, this is the main takeaway, like the three things that we should take away from this, in whatever structured format. And then that fits into your CRM, it fits into your traditional business logic, which is more fixed.

PARTH
So I think that eliciting structure from your unstructured, messy world of your business is the obvious first thing you can do with it—with language models. And then the next thing is figuring out the follow-up actions, like what should we do next? So here we have a slide, it says “Immediate next steps.” And it’s like, well, we should review the mobile UX because it seems like the conversion is lower. Certain products are more successful. It looks like certain—the AI is already doing this analysis that would have otherwise taken me, like, three weeks of just, like, being in the weeds. It is pretty shocking to me, and I cannot imagine working with an analyst that is no longer—that is not doing this, right?

PARTH
Like, every analyst I hire and work with in the future is going to be working at this kind of speed of, like, question answering and data visualization that I kind of expect from this kind of interaction.

REID
Yeah. And people, I think, worry there’s like, “Oh, a bunch of work goes away”. But actually, in fact, what happens a whole— this is an example of where a whole bunch of new work gets created. Because if you think, well, we only have been asking the most minimal questions before, because it was so expensive to do it. Now it’s a question of, oh, well, for example, on these conversion rates— does the conversion rate change by time of day? Does the conversion rate change by holiday? Does the conversion rate change by… Like, like, give me all the variables that it might change by. Okay, well, in addition to that, you know, are there any surprising things that we might learn on conversion rate, like time responsiveness or other kinds. And you could just be keeping going. And that’s what the task becomes.

That’s the new job. The new job is asking the AI to do the right thing and figuring out what the right questions to ask are. It’s no longer like, did you write the right SQL query? It’s not the syntax of, like, did we write the right query? But now we’re at this, like—we’re all kind of at this more macro orchestration level. I think it’s amazing. I think it’s a total expansion of the analyst kind of role.

REID
Yeah. We both follow Sam Schillace. And I—it’s a little bit like his description of, look, there isn’t coding anymore. There’s system architects

PARTH
Yeah.

REID
That, as you’re using it— the same thing is, by the way, true of lots and lots—it’s a parallel. One of the things that’s interesting is that all— like, one of the things our entire audience should take away from the thing is, even when you’re talking about coding, coding is a parallel. People go, “Oh, that’s coding. That’s different.” It’s like, no, coding is a parallel to data analysis. It’s a parallel to creating memos or PowerPoint. It’s a parallel to doing legal stuff. It’s a parallel to doing auditing and risk analysis. It was a parallel to all of this stuff about, like, how is it that you orchestrate for now being able to do a ton more work in a short amount of time— greatly expands the kinds of things you can be doing that are value creating.

REID
So it isn’t that your previous thing— which took three weeks, now done in two minutes— like, well, then I’m going to spend the rest of the three weeks playing Halo. It’s like, no, no, I can actually do a whole lot more and create a whole bunch more value here. And so the job is still valuable.

PARTH
Yeah, we can fan out. And this is where the computers are very useful. It’s like humans— like, we are single-threaded. We can only really work on one thing at a time. But if you say, let’s expand our analysis and look at 10 different angles at the same time, that’s the kind of thing that’s unlocked when you lean on the parallelization of the computer, and the coding agents, and the agents that can work in parallel. You can spin up many of them, and then you can attack a problem with many angles at the same time, which is a totally new capability.

REID
Part of the thing that people don’t, like, understand is that learning these tools is not learning how not to do your job. It’s how to learn to do the job in a way that you have superpowers, and that’s the whole part of Superagency. So one of the things, naturally, is that, you know, most people are non-technical. Part of their fear and concern about adopting new technology is that they don’t know how to use it. They don’t know if something goes wrong. They don’t know, like, if something happens.

REID
What do you think is the right mindset for, kind of, the non-technical executive for thinking about this and kind of wanting to engage? And then what would they need to understand about, like, how the models work? Or what is, kind of, some of their potential experiments for how do they lead well in the AI age?

PARTH
I think it’s less important that you know how a language model works, and more important that you know what it’s like to work with a language model. Like, working with the tool is a new skill— less important than, say, like, why does it predict the next token a certain way? I think it’s more about, like, what can this model— given access to these tools— do for you. And increase— I think the technical, you know— right now the coding agents are the most powerful agents, but eventually we will get their counterparts that are non-technical-friendly. I think Claude and ChatGPT and these tools will become even more powerful and more agentic. And, for example, like, they might be organizing your digital life, helping you organize all your files, your, like, your healthcare records, your personal life, your work life, creating that context, enriching that context, retrieving it when you need it.

PARTH
Those paradigms I’m already seeing in the coding agents, and I’m sure they will end up cascading to the non-technical experience. If you’re ambitious—and I think more people should be ambitious, because you can teach yourself anything today— I think— and this is something that some of my, like, non-technical, like, some of my old bosses, they’ve come to me and they’ve been like, I have time, I can learn something, what should I learn? And in those cases I do push them to play with a Claude Code or a Codex to get a sense of what it’s like to have an AI on your computer, working side by side with you, organizing your work, creating daily automations.

PARTH
There’s things that you can’t do in ChatGPT that become— like, when it’s— you do a ChatGPT and you ask, generate 100 images, it’s only going to do, like, three, and then it’ll just kind of stop there. But if you go to a coding agent, it will write a program that can generate 100 images. And so the— if once you want automation, when you want personal automation, you have to leverage code. It’s just that right now those tools are a little bit more for the technical person. I still think it’s never been easier than before to get Claude running. And if you’ve seen anything of how I interact with this, I’m not coding. I’m talking to a chatbot that writes code for me. So I’ll generate thousands of lines of code without having to personally code them.

PARTH
It’s mostly like I’m delegating to coding agents on my behalf. I think we will all be doing that eventually in some capacity. So if you’re ambitious, pick it up now when it’s a little bit, like, early and you get that head start on it. Otherwise, take the most technical person you know and equip them with— and invest in their coding agency. Be like, “You should be using Claude Code.” Like, if you’re a non-technical leader, empower your technical counterpart, your CTO, to be using these tools and to cascade that throughout the firm. Because that is, like, unblocking them— making them— putting them in a place where they’re not afraid to learn and experiment and discover the value is the most important thing you could do.

REID
I actually have directed a number of people who are non-technical to start vibe coding. And it’s still a little rough for the non-technical person.

PARTH
That’s right, that’s right.

REID
So, for an executive, one of the things you can do is actually go get a technical person and say, look, these are the kinds of things I’d like— set this up for me, right? Like, do the vibe coding— sometimes some of the things I ask you is like, look, set up this vibe coding thing for me so I can start using it.

PARTH
Yeah.

REID
Because then I’m getting that experience and that foresight into what being AI-native, making it happen, is. And I don’t have to learn the current hard edges.

PARTH
That’s right.

REID
Of vibe coding.

PARTH
That’s right, exactly. I think— and also I can create an environment that will prevent you from even, like, you know, tripping over yourself.

REID
Yeah.

PARTH
And discovering the value prop more quickly. Or like, creating a custom agent for you that matters more to your use cases than to my own. So, yeah.

REID
And that’s also part of our earlier conversation about the Zen of AI is: put your ego aside. Like, yes, someone else knows how to do the vibe coding thing much better than you. Just, like, learn to partner with them the way you learn to partner with AI or dance with AI and kind of say, okay, help me solve this problem. And then, as you do that, that gets you into the learning. Because, by the way, of course, part of once you’re down the road, where you’re beginning to see what the agent can do for you, then you’re like, oh, well, now I want this too. Now I want this too. Because that’s the way you learn it. You don’t learn it by, like, oh, I sat down with the AI for Dummies book and, like, thumbed through it.

PARTH
Right, that’s right.

REID
I learn it by doing it.

PARTH
There’s no better way to learn this technology than by using the technology. And there’s no alternative to that. And we— when we meet, sometimes you’ll have a crazy, interesting, creative idea, and I’m like, we should just get the first version of it. There’s no— there’s no reason why we can’t get the first version in the next three minutes and then validate some of these hypotheses on design and whether, like, there is something there. And I think that, like, for me, every time that happens, it’s more like, oh, I got to show Reid that this might be possible in, like, five minutes because I want you to update your priors on how long a certain, kind of like, old world job might have taken, now given we have this massive accelerant.

REID
Because what most people don’t realize is, through your life and your tool use, you have constrained your imagination to what you think is possible in the old toolset. You now need to re-release your imagination. There’ll still be some constraints you’ll learn, and those will change over time. But you now have many more capabilities than you imagine. Like, for example, like, one of the things you say: Well, like, okay, I’m trying to figure out how to use AI for leadership and I’m sitting there, and I’ve just listened to this podcast and reading Parth. Well, go talk to a Frontier agent and interact some with it, saying, well, what are ways that, if I was trying to solve this kind of leadership problem and let me describe it in more depth—and/or the interview me prompt, like, interview me about this problem until you have enough to say something interesting to me about how I can use you in order to help me with this leadership problem.

PARTH
Right. Or even take the transcript of our podcast and show that to an AI and have it turn into a framework for thinking about how to explore these ideas.

REID
Or a framework for you, because you upload the transcript of the podcast, then you say, hey, here’s the problems I’m going to solve— how would you apply this conversation that Reid and Parth just had into a framework that would be useful to me, in my company, in Company X, in my role, et cetera. And part of the thing about the acceleration of AI— so there’s, you know, AI is amplification intelligence, there’s also AI as acceleration intelligence. You can do all that in minutes.

PARTH
Yep.

REID
Right. And that’s part of the thing to start thinking about what the time frame changes in terms of how you’re capable of operating as an individual, how you’re capable of operating as a team. And again, that’s one of the reasons why you get into the experimentation and learning. So when you start kind of learning that this is a learning journey, and has annotation, you begin to realize that, actually, in fact, constructing tools to amplify yourself, to amplify your team, is now something that is doable for every individual, for every team. One of the things I think we’re going to see a huge explosion in work and amplification is by self-amplification, where it’s the self is somewhat the individual, but also the team.

REID
And it’s because you’re going to start doing tool development to just accelerate you for your particular work, for your particular group, for your particular company. So what are the ways that people should approach this learning about tool amplification? And where might some of the various tools that you’ve mentioned in your AI stack are things that people should experiment with— Replit, others, that sort of thing?

PARTH
I feel like I’ve gotten this mindset— it’s an extension of vibe coding, but probably a more practical application of vibe coding. The instinct some people have is, oh, I’m going to build Facebook and I’m going to publish it. Maybe that shouldn’t be the first thing you make, but because there’s a learning curve to building these generating tools and building tools on the fly. But the safe way to experiment is to build internal tools. Build tools that where you know the user, because it’s you. It’s like, oh, the user is Reid. And maybe, like, Reid’s team. And so, like, if I have a relationship with you and your team, then I am getting feedback directly from the users. And this is the other thing, is like, traditionally, you would have like, a UX researcher, a product manager, and an engineer.

PARTH
That might be three different people. But now one person can kind of play all roles, and it’s actually better to be that generalist and to play all roles, because your feedback loop and your iteration speed on the tools is very high. It used to be that you might spend a week building one feature. Now you can build 6, 7 features in a day. And for an internal tool, that means that you can go from, like, not having a tool in the morning to having a very good first version by the end of the day, if you’re in that tight feedback loop with the end users on your team. And the other thing I think about tools is designing tools that are both for people and for the next agents that come online in your team.

PARTH
And so, a lot of times, I discover a workflow and then I go to AI and I say, okay, how do we make it so that my coding agents can also use this workflow? Like, I can pull the analytics this way, can they also pull the analytics? Because I want to ask them questions that I might not be able to quickly grasp, as quickly as they can kind of synthesize that stuff. So I think building internal tools is something that, like, is the easiest way to get into— like, unlocking the value. Because tools make it easier to do what we do today: free up time for us to do more things tomorrow. And for me, Replit was that realization. I use Replit. Most of the things I’m making on Replit that I’m vibe-coding are just for me.

PARTH
They’re not— I’m not publishing or selling that software. It’s personal software. It’s custom software. And when it’s really cheap to make software, you should just make much more of it, and it doesn’t have to justify its own existence in some kind of revenue-driving way. Then that is the validation, where you can kind of build these tools. And a lot of times you don’t even have to— like, I realized it might be easier and faster to build the first version yourself than to even go out and try to buy something off the shelf. That’s one of the realizations of vibe-coding. I do think there are still plenty of tools that you shouldn’t even try to make yourself. Like, maybe don’t build— unless you really need a very unique CRM, just buy a CRM.

PARTH
Compliance tools, anything having to do with, like, SoC2 or like, legal. Stuff where the requirements don’t change over time, I think it’s good to just buy something that’s commoditized. But if it’s unique to your problem space, your customer base, your user— something that’s very unique to that problem— I think you should build some custom internal tools.

REID
So, in terms of your own use of Replit to launch your own tools, what’s one of the, probably, more eccentric or funny— like, something someone wouldn’t have thought of, that you’ve built as a Replit, you know, tool for yourself?

PARTH
I actually have one right here. So this one I’ve been working on in Replit and with outside coding agents. I basically had an agent scan all of my— every project that I’ve coded in the last three years. I was like, “Just go learn all the tech that I’ve been using,” because I forget some of the stuff we used two years ago from Sam’s team at Microsoft—Sam Schillace’s team at Microsoft, Microsoft GraphRAG. The ability to, like, create these knowledge graphs on top of a large corpus of data. And I actually kind of think that you might end up with something like this inside of every company, where some kind of centralized agent-managed wiki— interactive, evolving kind of wiki.

PARTH
And I, you know, I built it in Replit because I was like, oh, it would be nice to talk to something that could reflect all of my technical explorations over the last three years. Like, read all my code, connect all the dots between all the technologies. And then I ended up with this kind of, like, 3D graph experience, and I was like— it’s all vibe coded, right? Entirely vibe coded. I don’t even think I wrote even 1% of the code myself. And then, when I had this, the interesting thing is it’s a knowledge graph of everything I’ve learned about working with agents and all the different technologies inside them. Then I asked the graph itself— I said, why don’t you ask yourself how to build a voice assistant that can control the graph?

PARTH
And then it added this little voice button, and so it modified itself into a voice agent that can query and understand the rest of my corpus.

PARTH
Hi there.

REPLIT
Hi there. How can I help you today?

PARTH
Can you show me— Let’s zoom in on the Claude code node and tell me about Claude code.

REPLIT
I found the node for Claude Code SDK. It’s actually the former name of what’s now called the Claude Agent SDK. It was originally focused on code related agent flows. Would you like to explore its connections or dive deeper into its details?

PARTH
Tell me about hooks related to Claude code.

REPLIT
The main connection for the Claude Code SDK is the Claude Agent SDK.

PARTH
And so I basically designed this thing to represent my technical explorations, but allow myself to have a conversation with my own technical mind, right? So— and I think that, like, building knowledge graphs for me and then connecting them to voice— it’s just, I love having it on the side and just asking it questions about connecting the dots between all the projects we’ve worked on. So I think something like that might actually be very valuable as, like, companies start growing and scaling. Like, how do you make sure that knowledge inside the company is accessible by all the different teams and all the different people? This is obviously one for me, but I can imagine many of these being constructed as, like, artifacts that are useful for, like, a new intranet—a new interactive intranet.

REID
Yeah, well imagine that it’s like every person has their own Wikipedia, every group has their own Wikipedia, every project has its own Wikipedia, every company has its own Wikipedia, etc. And it’s all easily accessible and brought to your, you know, your ear tips.

PARTH
Exactly.

REID
By agents.

PARTH
Exactly.

REID
Let’s be imaginative. What does the corporate office of the future look like?

PARTH
As someone that works from home, it’s been a while since I’ve been in an office, but I think about this— like, all of these capabilities. Language, voice, knowledge, systems, automation through natural language, agents that can take action, parallelize. I think what’s going to happen is certain people in a company… many people will have— I mean everyone’s at a different layer depending on your role— will have these, like, pods of agents that they interact with. And some of those agents will be shared. They’ll be, like, assigned to a project. Like, you know, I have an agent working on this knowledge graph, maybe you also talk to that same agent. So we’re both giving that same agent feedback. So it’s sort of like an extra perspective that both of us are creating— a character we’re creating to play a certain role.

PARTH
And then the question becomes, like, how many of these would we be interfacing with, and, like, how do they show up? I think they’ll show up in meetings, in both like, ambiently after the meeting ends. Like, why don’t we start these prototypes? Here’s a couple of follow-up actions. I’ve sent, you know, memos to X, Y, Z stakeholders. I think that kind of ambient layer is going to get unlocked. It’s going to make everything feel like, in motion and alive, and mutable in a way that maybe software and infrastructure has felt fixed.

PARTH
Where I feel that software is now becoming liquid and malleable and kind of, like, composable in a very, like, sort of spellcasting? Where, if you can, you know, say the right combinations of words, the agents start getting to work. And they start creating a new interface, a new product. They start solving a problem for you, or they run in the background. And I think that the idea of these processes running ambiently in the background, as the people are aiming them, is going to be very big. I already see it in my own life, where I have, like, you know, maybe 15 projects. Each one has two to three agents on them, and some of them are working on 1-2 day long projects, and they might come back two days from now with progress.

PARTH
And the kind of progress that an agent can make in two days will blow your mind. But it also feels like, then the time scales are collapsing for how quickly we can attack more ambitious projects— where you would have previously required 10 people to attack a problem over two years. Maybe we create an agent that works on it for two days, and we get the first version back. But then it’s not that once that thing is out there, we just kind of wait. We’re going to have multiple of these kind of jobs running, where we’re firing off more questions, more queries, more explorations, more prototypes, more hypotheses, more variations. I think even products might feel, like, very— In the same way that we A/B test in software, I think products might actually evolve.

PARTH
Like, I look forward to a day where I wake up and the agent is like, oh, by the way, I rebuilt your knowledge graph in six different ways. Do you like any of them? And you get all these options, and then you get to choose. And they’re working at night, right? So I think even in— like, imagine in healthcare and in research kind of roles, you go to sleep with a couple ideas, and you wake up with possible answers. That’s— that’s what I look forward to.

REID
Possible is produced by Palette Media. It’s hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Shaun Young. Possible is produced by Thanasi Dilos, Katie Sanders, Spencer Strasmore, Yimu Xiu, Trent Barboza, and Tafadzwa Nemarundwe.

ARIA
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil and Ben Relles.