This transcript is generated with the help of AI and is lightly edited for clarity.
//
REID:
Parth, welcome back to Possible. It’s been, you know, only a few months, January, since we’ve had you on Possible.
PARTH:
That’s five years. Yeah, that’s five years in AI time.
REID:
Or 100. Who knows? What’s up, Doc? What’s changed?
PARTH:
It has been a ride. I was actually just thinking about this just yesterday. Like, OpenClaw came out and went, I mean, OpenClaw went viral. Came out in November, but went viral really in February. And I was just looking back at my own logs, and I realized I installed OpenClaw on February 15. And so it’s really only been, like, four months since I was on OpenClaw. But for me, it feels like I’ve had these systems around me now, working with me for years, with the amount of, like, skills and abilities that they’ve accumulated, OpenClaw, Hermes-type agents. I think that my concept of an agent has now moved to this point of, like, it’s the remote. It’s this remote worker that works for me at home 24/7.
PARTH:
And then I have a few of them running around. At the same time, tools like Codex and Claude Code have gotten even better. We’re now in the place, I think when we talked in November, I was like, yeah, I can get Codex to work for multiple hours in a row. Now we’re at a point where Codex, GPT-5.5 can work for multiple days in a row. It can work five days straight in pursuit of a goal. And that’s very interesting. It’s so hard. I mean, it’s a whole new game of, like, how do you evaluate something that works continuously for 50 hours? Right? But it’s also, yeah, you don’t know what the ceiling is. We’re still figuring it out. And the game is afoot. Yeah.
REID:
So first, you said a few agents, but actually I happen to know that’s a little bit more than a few. What’s roughly the, you know, the panopticon or the pantheon of, kind of, agents? And what is the kind of circle, the number of, you know, and it’s hard to be exactly ontological on this, but the number of agents that are running?
PARTH:
Yeah. So I finally have an interesting framework for this, which is there’s, like, different levels to the agents. There’s agents that you spin up when you need them arbitrarily, when you want to scale up your effort on a task. Right? I think of those as, like, the drones, like subagents, drones, kind of like the ephemeral swarm that you can scale up and scale down on an as-needed basis. And then I think there’s the other side, which is more interesting, which is kind of like the OpenClaw, the Hermes. Because these agents that act, that you work with over the course of several months on a particular problem, they tend to accumulate experience, memories, and skills.
PARTH:
And what I’ve noticed is when an agent accumulates an interesting set of skills and memories, you are inclined to give it a name. And it goes from, like, it becomes. It’s elevated from the rest of these, like, ephemeral drones that you spin up to, like, someone on your team. And in terms of the number of agents that I have actually, like, named, that I go back to for different problems, unique problem spaces, or, like, they have different purposes. Some of them support our team. I have one that just tracks my personal life. Those. I have eight of them, and I think there might be, like, a Dunbar number here of, like, how many you can actually manage that you actually want to interact with.
PARTH:
But even they can spin up drones under them to scale up their effort on different tasks, which is fascinating. Let’s say, if you count the drones, it’s on the order of maybe, like, 10,000, maybe. But, like, a few. Yeah, it just gets, you know, you get — But if you count the ones that, like, are recurring characters, recurring cast members, I would say eight or nine. Yeah.
REID:
And then, you know, on the 50-hour, you know, kind of tasks. What kinds of tasks are those?
PARTH:
Yeah. So the first test I gave that clued me into this, like, the new ceiling was I asked for Codex to build an Android app and an iOS app at the same time for a concert music discovery experience. Basically, I had an app in mind. And I’ve never, never built a mobile app. I’ve done analytics, but I’ve never built a mobile app. And it installed the developer kits for both iOS and Android and then started building the app and testing it in simulation. Because I basically said what the success criteria were, like, send me the working build and send me screen recordings as you build features that demonstrate that the features work.
PARTH:
It builds the app, it builds a feature, it pretends it’s a user, clicks through the app to make sure that the feature works. The whole thing is a screen recording. So for me, it’s like, I also get the app so I can play with it. But the bottleneck is now reviewing time. Watching a video of a feature working is actually a lot easier for me to review than reading every line of code in a problem space that I don’t have experience in. But the idea that you can turn, I mean, it would usually take multiple people several months to get something like that done. But the idea that you can just turn a prompt and an idea into the first working artifact in the span of, like, two, three days is amazing.
PARTH:
It’s a new level of prototyping. Definitely.
REID:
And what’s the most unusual thing you’ve — or a couple unusual things that you’ve done with, you know, kind of your agent teams?
PARTH:
I’d say the most — one of the more practical, unusual things that I like to do is: So we talk a lot about, like, today we’re working with language models and you get out what you put in. So if you’re an expert in a topic, you can usually articulate what you want with a pretty good sense of detail. And so your prompts are very, like, precise. You know what you want, and then you get a very high-quality output. But when you’re trying to get agents to work in a domain where you don’t have experience, you might not even know how to ask the right question, how to ask the agent to do the right thing. And so what I do now is I have some agents.
PARTH:
Their entire job is to just do automated research, to construct a knowledge graph of a vocabulary around a space. And then when I actually prompt for something, my prompt gets intercepted by the agent, upscaled because it’s run through this knowledge graph of expertise found from the internet. And then that prompt goes in. So I’m able to, you know, when I prompt a video model, I don’t have cinematography skills, but if I describe the kind of effect that I’m going for, if I’m trying to, the model will upscale my prompt using an accumulated knowledge base of tactics that it’s constantly researching. So even researching, like, what are the best ways to use coding agents, to having an agent that just researches that is very helpful instead of me relying on my own individual learning curve.
REID:
So given obviously how fast all this is moving, kind of who’s not only keeping up, but leading, and what happens if you give them more fuel?
PARTH:
This is the most interesting thing. I feel there’s only so much that I can explore by myself. And I already. That’s all I do all day long. And it’s hard to keep up with trends in AI. There’s so much happening across the spectrum in the field. What I found is that the people that are obsessive, that are very passionate about a problem space. These are the ones that are uncovering what the ceiling of AI is, what it’s capable of. A lot of the best tools that I use today are a result of talking to my friends that are very deep in these fields. We’re going to meet some of them, actually, and learn from them.
REID:
Well, let’s go into it. Tell people what the token grantee program is.
PARTH:
Yeah, so I’ve had this idea for a while, ever since I started playing with language models. But essentially, we created a program, this token grantee program, where we take a certain number of high-potential individuals and we give them $1,000 a week in AI tokens. And this is a recurring grant, not a one-time thing. And we don’t restrict what tools they can use. And that’s by design, because the best model on any given day changes from week to week. So we don’t want to be prescriptive in terms of what models we want people to use. We actually want people to figure out what the best models are for any given job and then to just use that. That’s why we’re giving them $1,000 a week in tokens: use them on whatever tools you want.
PARTH:
And then on top of that, I give them access to a couple agents of my own, which have been accumulating skills that are using a lot of the same models. And so you get this, like, remote, you get this remote agent support, you get that, you know, from, like, Hermes agents, OpenClaw agents that I’ve been working with. And then you also get a collaboration with me. Right? So I’ve been deep in these models as well. So being able to bounce ideas off of someone that’s also kind of going psycho using all the best frontier models, it becomes helpful for both of us. I mean, I get a friend that I get to work with and unpack the future, and then they also get, you know, more and more resources.
REID:
So what does this mean in terms of, you know, kind of how you balance patience and impatience? What does it mean for creativity? You know, what are the kind of the restructuring of what is possible and now, you know, kind of what’s available in terms of what that possibility is?
PARTH:
Yeah. So creativity has always demanded patience, and I think for the first time, there’s certain vectors of the way we work that we can actually choose to trade electricity for time. We can choose to, like, the fact that you can have an idea for a mobile app and within two, three days you have a working Android and iOS version. I mean, you can imagine going to the App Store within a week without having any engineers is pretty incredible. If that’s basically the path that we’re on, the idea that we’re closing the gap on the time between having an idea and getting it out into the world, I think that’s the patience gap. We’re hoping to figure out the mechanisms by which we can close that distance down as much as possible.
REID:
Yeah, yeah. Well, I do think that, you know, I mean, it’s not just because, you know, the podcast called Possible, but the notion of redrawing the boundaries of possibility of. Because, you know, part of what we do even in innovation is we constrain our imaginations to a kind of a range of what we think is possible. And when the possibility landscape gets redrawn, you have to retrain your imagination. You have to retrain what is kind of what is possible on it, and kind of this balance of patience and impatience, because you want impatience for driving where you’re going. You want patience to know that actually sometimes a long slog actually, in fact, gets you there.
REID:
And this is part of, you know, going back to Superagency and Impromptu, of kind of, like, AI being amplification intelligence is it now makes what is possible for people to do very, very different. And also, of course, you know, now scaling is not just blitzscaling a company, but scaling individuals, scaling imagination, you know, and all the rest. So let’s talk a little bit about your roster of grantees. Run us through them.
PARTH:
Sure. We have Will Weinbach, who is a New York social media lead who thinks in terms of distribution, content, and social. He thinks about narratives and turning experiments into stories that can travel. And he’s also one of the most prolific users of agents that I’ve ever met. We have Ben Hansford, an award-winning director and creative technologist who teaches film at USC, and he blends deep machine learning work with commercial creative for major brands. We have Joey Salvatore, an editor turned VFX artist turned AI creator who builds out loud, growing an audience and learning and teaching in public through his content distribution and culture.
PARTH:
We have Dozi Anazia, a security-minded builder out of Arlington, Virginia, with a CS master’s and a deep technical base. The kind of person that turns spare compute into applied experiments instead of passive exploration. We have Andrew Denta, an applied AI operator building the future since 2021, shipping production-grade agents at Mangrove Technology and an early force behind Reid AI’s language, voice, and avatar work. We have Katie Jiang, a systems-minded builder in San Francisco whose arc runs from finance and crypto payments to AI-powered accounting and robotics product ops. And she’s even taken a stab at building AI agents for game design. And we have Matthew Tiemann, a D.C.-based AI engineer with a winding path through robotics, living, logistics, finance, and data analysis. Now building agents at Foundry Logic to strengthen American energy infrastructure.
REID:
Yeah, so look, it’s a fascinating and diverse group. You know, it’s kind of like, I think part of the selection was people who not only have kind of raw, unique, and interesting mindsets and talents, but also kind of are exploring in different directions. It’s kind of like, you know, kind of AI pioneers. So how about, you know, how did you think about the selection of this group, and then let’s move in from that to kind of thinking about this as a potential new shape for entrepreneurship.
PARTH:
Yeah, I think, you know, obviously we’re seeing in AI right now the first major commercial application is in programming and coding automation. And there’s reasons for that. It’s a verifiable domain. It’s very easy to see whether code runs or whether it works or not. But I think what’s more interesting is the second- and third-order effects, the other domains that are less objective. Things like the creative spaces. And when I think about AI, I think there’s a massive amount of creative leverage that’s being created as a result of the automation in the subjective, in the, as a result of the objective, the automation in objective domains, we have a huge opportunity.
PARTH:
In creative domains where it’s not obvious what right or wrong is, what good or bad looks like, that requires more judgment, better understanding of people. And so I think about, when I think about who these people are, I think about people that are, yes, some of them are very strong technically, but more importantly, they have this creative. They don’t view themselves as purely technical. They have this creative ambition. They work outside of their lane in their hobbies and interests, and then that passion is taking them to new realizations in domains where AI is not nearly as good.
PARTH:
And so I think, like, figuring out what we should be doing when AI is getting good at automating and doing math and coding, a lot of that is actually, like, what about the things AI is really struggling at that requires human judgment? And so that’s kind of why we went with a much broader mix than purely engineers.
REID:
Yeah, and, you know, part of the thing, people naturally think, “Oh, it’s coding, you know, it’s use of agents, it must be AI.” But actually, in fact, many folks who basically have zero engineering background, zero kind of coding background, actually, in fact, are hugely amplified because now there’s a higher premium on creativity. Ability to synthesize, being able to ask the right questions, being able to see what problems are kind of interesting to solve, being willing to try something different to make it happen. Like, all of this kind of, this stuff echoes into some of the key skills around entrepreneurship. And so it’s kind of this notion of not being afraid of the tech, of being engaged in it.
REID:
I mean, it’s one of the things where, you know, whether it’s from new college grads or existing employees, like, “No, no, actually, in fact, dive into it and explore it to amplify yourself.” And again, that’s part of the reason why the closest parallel to entrepreneurs that I’ve kind of talked about over the decades is pioneers. It’s being willing to explore a new space. And that requires a certain risk tolerance, a certain boldness. But now the depth of the shape of technical skills is very different and include being very light. So it’s much more important to be asking the right questions and be creative the right way. Which is, of course, part of the reason why you and I, you know, those few years back, started working together, not least because you’re one of the co-authors of Reid AI.
PARTH:
That’s right. Yeah.
REID:
So what do we think, if we kind of pull the camera back, if this works, what does it unlock? And we know that access like this is going to broaden. What happens as it broadens?
PARTH:
Yeah. You know, I think just three years ago when we met, you met me at the tail end of my deep exploration. But at the time, it was not.
REID:
Your first deep exploration.
PARTH:
Yeah, my first deep exploration of language models. And. But seeing what GPT-4 was capable of back then and what I could do with it just eight months in, I realized pretty early on it was kind of isolating at the time. And I realized, like, what I lacked around me was a bunch of other people that were also in the same headspace at the time. And over the last few years, what I’ve been doing is working with a lot of the people that we’re about to meet. I started meeting them back then, and I started introducing them to these technologies. Some of them had jobs, some of them were working on other things. But I basically clued them in back then.
PARTH:
And the interesting thing is, like, when you drop GPT-4 into a network and then you come back eight months later and you see how people are using it, how it changes their worldview, how it changes their career arc, their view of everything. That, I think, is the beginning of this idea of Superagency that you write about, because it’s not single player, right? I exist in a network, I exist in a community. And once I learn something, I can teach a few other people, and then we start working together, and then we can start collectively learning the next game, the next problem space. And I think that’s where we are now. We’re in this next phase of this community, exploration of many domains at the same time.
PARTH:
And AI is getting better, faster than you could ever imagine. So no one person can explore it alone. And so now we’re at the community level. And then I think it’ll be interesting when it starts getting to a society level, right, where each person is now amplifying 10 people around them, their families, the problem spaces that they work in, and it starts getting into different metros, into different countries. I think that’s very exciting to me, being able to show up somewhere on the other side of the world and have an interesting conversation about the frontier capabilities and then what it means for everyday folk. So I’m looking forward to the diffusion of the technology now more broadly.
REID:
Where are the areas in terms of human expression, and what is this going to mean? Because, for example, you and I have, along with a couple other key members of our team, worked on creating AI music: Christmas album, Valentine’s album, got some other ones coming up, you know, a pie song celebrating pie, April Fools. That’s one. You know, images, creating a variety of images, but also, you know, kind of like one-off books. Where is the kind of the range of human expressiveness? Because I actually think one of the things that people really undercount, part of the reason I did Impromptu first, is how much this amplifies the ability to express humanity, to do human connection.
REID:
Where do you kind of see the range of creativity and human connection going?
PARTH:
Yeah, so if we’re looking at generative media, and that’s whether that’s generative text, generative images, audio, video, as in its current form, and then where it’s going next. There’s also generative worlds coming online from Dr. Fei-Fei and similar teams. I feel that now any idea can be transformed into almost any medium for a pretty low price. I think the quality is still getting there in some of these domains, but it is rapidly accelerating. I mean, just two years ago, I was showing you prototypes of movie trailers with you and me in them, and it was like we didn’t have any dialogue and the max, the longest shot that we could get that was reliable, was maybe four seconds. But now, with Seedance 2.0, it’s clear that we’re going to start getting into long-form video storytelling that is generative.
PARTH:
And so it’s not just like an expansion of VFX, but that everyone has a VFX team, everyone has a studio in their pocket, is kind of where we’re going. And that changes the number of people that can create content, the number of people that are telling stories, and the range of stories that end up being told. So I’m very excited about that. I think ultimately I’m personally most passionate about. I think within three years we’re going to have a thriving ecosystem of vibe-coded games. And if anyone can make a game, we would have an explosion of games. And I look forward to that world.
REID:
Well, as a deep gamer, that is not a surprise. Have you seen any glimmers yet of, you know, technology starts with kind of the replication in the new technological medium of old, right? So it’s like when, you know, the first kind of online, like, “Hey, we’re going to have marketplaces for transactions,” it was like translating newspaper classifieds into, like, searchable classifieds. Right. And et cetera, et cetera, et cetera, as ways of doing this. And so in these media, you tend to go, well, we got music, we got video, we got games, we’ve got images, we’ve got words. Have you seen anything yet that is. And I’ve seen a little bit of art stuff that is super interesting.
REID:
But have you seen anything that kind of is adding to the human expressiveness, communication style? You know, a trivial one might be emojis, you know, kind of for where those are sort of getting added into the way that people would communicate emotionally, you know, kind of in quick form and is now part of a thing, you know, or acronyms from much earlier internet, you know, ROFL, et cetera, et cetera. Have you seen anything like that yet in kind of the various, you know, AI expressiveness?
PARTH:
Yes. I think we see it in images already, and we’re also starting to see it in video. But I think AI images, image models, GPT Image 2, which came out, I think, since the last time we talked, GPT Image 2 feels like, I would say, an AGI for images. If I were to say, like, what would that look like? It’s an image model that is both really good at rendering, you know, whatever you ask for, but also creating meaningful new infographics and stuff like that. But it’s also, like, deeply culturally aware, and it’s also capable of modifying images. So I think there is something there, which is the. There’s a new vehicle of creating memes. And if a meme is this.
PARTH:
This idea, this viral idea that gets remixed and broadcast and remixed and broadcast by many people, we now have the ultimate meme creator, which is this, like, image model that within a couple seconds you can edit the image into this new concept, bring in the cultural reference. So I think there’s new vehicles for memes to emerge, and I think same thing is true in video. And that’s really exciting because if creating something becomes very easy, then more people can do it, more people can participate. And creating memes used to be kind of a rare skill and kind of a slog. Like, you have to actually be good. You have to understand the culture and.
PARTH:
And then you have to go and have the skills to edit the meme together to make it and make something worth it, something that actually travels. And I think that more people will be able to create interesting memes, and more interesting new meme formats are coming out. Video meme formats. Yeah, I think that’s something to watch. We see it. We see the early forms of this.
REID:
Yeah, yep. And it’s kind of just beginning. And then one. In terms of, you know, one of the things that people kind of both wonder and worry about is kind of AI is the new companion. It’s a little bit like Inflection with personal intelligence and, you know, kind of other things. You started this very early, including, like, you know, helping your mom plan trips to Rome and all the rest of the. What is like, kind of a principle about, like, how to dive in and use this and welds for the wonder and any principles in terms of avoiding or being careful.
PARTH:
Yeah, I think the best way to start getting into these systems, getting into agents in a meaningful exploration, is to apply it to something that you’re very passionate about, something that you have experience in that you’re very passionate about. Because then no one has to tell you to do it. It’s not like, “Oh, you gotta use AI for your job.” It’s more like, actually, “Here’s what I wanna do. I wanna learn this new art style.” Like, my mom likes doing watercolor. She likes exploring art. But then she realized that GPT Image 2 becomes a really good copilot for pre-visualization for watercolor painting.
PARTH:
Because GPT Image 2 knows how to imagine the thing that you would need before you went to start painting that would give you a better shot of not making a mistake on the final output. But that kind of connection, it’s like, I know my mom cares about art. I show her GPT Image 2 as this, like, new tool in her toolkit for achieving her creative vision. I think that’s the advice I would give, is to tie it to your passion. That way, you know, you’re never going to feel like. It doesn’t feel like work. It feels like playing and exploring, which is really, really great. In terms of what I would say when I warn people, I would say that there are some mistakes that you just have to learn through experience.
PARTH:
Automating something that shouldn’t be automated. Automating something that represents your unique voice in a way that I wouldn’t necessarily show up on social media with totally automated language model text. Now, knowing how kind of average, I mean, it’s good. It’s good for transactional communication, but I don’t think that’s necessarily how I want to be presented on social media. So showing up as myself actually ends up being a more. It’s just a better. Like, I think I have a better take than the average take coming out of a language model. So knowing not to automate, things that are subjective, things that you really care about that you wouldn’t want to, that you think you’re actually better at. The language model app. Better than the language model app.
PARTH:
See, I’ll automate data analytics, but I wouldn’t automate the full range of observations that I’m making and the overall strategy I’ll have. Like, I consider AI a way to supplement my thinking rather than replace it in those lenses. And then the other piece I would say is it can be very. Especially this happens to a lot of my friends that I get into this space is. It gets very obsessive. People get very obsessed with their capabilities that are coming online. And so they get very excited. They start losing sleep. They start. It’s like, “Oh, one more prompt. One more. Oh, just like, if I figure out the next thing, you know, we’ll have this app done.”
PARTH:
And I think that mindset is useful for getting very deep into something, but it can be helpful to take a step back, go for a walk, you know, get some sleep. I think now that I’ve invested in an Oura Ring, mostly because I realize, like, the agents are very relentless. They’re constantly working, but actually, we are humans, and they need to meet us on a human level. So I’d say it’s very exciting. It’s very addictive. There’s some casino elements, I think, to, “One more prompt. We’ll get there.” But I think if you keep that in mind and play for the long haul, you should build a healthier relationship with these systems.
REID:
Yep. Well, I think this, you know, kind of sets up for what we’re doing because we’re actually having a series of these. And it’s, you know, partially kind of the question around things that people are passionate about reimagining, you know, kind of what’s possible in an era of amplification intelligence. And exactly as you’re saying, you know, kind of the question of, like, well, what are the places where we’re bringing, you know, kind of really interesting human capabilities, human perspectives, human questions and answers. Because, you know, frequently somebody who knows something with an edge is actually better than the kind of council of experts, you know, kind of call it opinion that, you know, kind of AI tends to generate. So, you know, we are.
REID:
The point isn’t the tokens, although one of our grantees burned through 20 billion tokens in a weekend. It’s what becomes possible when you rewrite the limits of imagination and possibility and when that becomes how you drive your creativity and your humanity. So, you know, I think it’ll be fun as people tune into this.
PARTH:
I’m super excited for this program. These are some of my favorite people in the whole world, people I work with every day, every week, and excited to bring them on and see what they’re working on and how they think about the future.
REID:
Parth, always a pleasure.
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, Aman Suri, Lexxi Kiven, Danny Garrison, Trent Barboza, and Tafadzwa Nemarundwe.
REID:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.