This transcript is generated with the help of AI and is lightly edited for clarity.
//
JOE:
I can outsource production, but I can’t outsource my taste. Like, that has to come from the engineer or the creative professional.
I’m actually learning now how good of a director am I actually? Because it’s going to be crap if I give it a half-hearted direction.
It pulls me to something that you write about a lot. Limits of my language really are the limits of my world. Except now language doesn’t just describe the world—it kind of summons it.
PARTH:
In creative, the model can generate any video, but everyone has a VFX team in their pocket with, like, image and video models today.
What a lot of people are missing is, like, not everyone has a storytelling skill set—they can create a story that actually someone wants to watch a second time. I feel that the writer’s skill set is actually the most valuable thing that I was undervaluing.
//
PARTH:
The token grantee program gives $1,000 a week in tokens to high-potential creators already deep in AI, across film, gaming, comics, print, and digital art. There are no tool restrictions. The freedom to choose is the point. Grantees also get access to my own custom fleet of agents and an ongoing collaboration with me. The goal? To close the gap between an idea and the world.
REID:
Joe, awesome to have you here. Parth, tell us a little bit about Joe. Like, how did you guys meet? Why did you select him for the token program, et cetera?
PARTH:
I met Joe about a year, a little maybe like almost a year and a half ago now. We met at this creator studio that we work out of here in Venice in LA. We work out of a place called The Lighthouse.
REID:
Oh, I think I’ve been to that one.
PARTH:
Yes, yes, you’ve been there. Yeah. So Joe and I work out of the Lighthouse, and he’s one of the regulars. Definitely, like, you know, you go in, it’s like, I see Joe. But when I first met Joe, at the time, I felt like I was maybe one of the few AI creators, and in a sense, like, definitely the first vibe coder at the Lighthouse—like, one of the few. And then I met Joe and I realized Joe was deep in—you know, his background is more in visual art, visual storytelling, and, like, more brand-related asset creation. And so Joe was very deep in the image and video model space. And, you know, you can tell because you ask someone what’s your favorite model? Why? What are your favorite workflows?
PARTH:
How does it fit into, like, the way we think about these new formats? And Joe would be the one that I would go to with my questions about the current state of image and video and what the third-order implications are. You know, it’s one thing to generate one thing, one picture, but for it to fit a more cohesive kind of brand thesis. So I’ve learned a lot from Joe over the last year and a half. And I also introduced—he’s one of those people. It’s like, well, what if I give him an agent? You know, what might happen to someone who’s not an engineer at all, but, like, when you give them coding agents, when you give them these types of systems, can we turn it into—
PARTH:
What are we going to—what will we learn? So that’s kind of, like, the story behind me and Joe. But what would you say, Joe?
JOE:
Yeah, I mean, you nailed it. Like, I had been experimenting with the tools for probably four years. I saw a group of filmmakers in, like, 2021. They made a really cool AI trailer that was—it was basically a mashup between The Lord of the Rings and Wes Anderson. And I saw it and I was like, that’s just really cool. I want to understand how to do that and just kind of have that creative weapon available to me. So from that point, I was just digging in and learning the tools and, like, playing that game of trying to keep up with all the releases. And then when I met Parth, you kind of opened me up to—
JOE:
The agent side of, like, hey, why don’t you try to take a step back and be an orchestrator of these things, rather than staying so in the weeds of making one piece at a time? Which was, like, a really important shift for me. And, I mean, you know, Reid, I know you talk about augmented intelligence versus artificial. Like, that’s exactly what I’ve been personally experiencing. It’s like I’m tapping into areas of my thought process that I actually didn’t really know were accessible to me before these tools. So it’s just stepping out. Larger picture for me has been kind of the story.
REID:
So what was the first thing you saw? The movie? What was the first thing you started diving into, like, on your AI journey?
JOE:
Yeah, image generation. I would say Midjourney, the creative tool, was the earliest thing. Just throwing things out there. It was purely creative curiosity. There was no other reason for me to be using it other than that. And, you know, as I dug into that, I wanted to know the other tools that were coming out. I wanted to tap into the video capabilities. So it just—it was a sandbox for me, and now it’s to the point where I can functionally apply it to my career, which is, you know, I make content for brands. So now I’m at the point where that’s evolved to: I am running automated static ad creation through Codex. Thank you, Parth. But that was the start.
JOE:
Image generation was the start. Yeah, it was very much, like I said, just keeping up with all the tools. What do these image models do? What do these video models do? And it’s really only been recent where I’ve been able to say, okay, the real skill set here is my articulation of my creative ideas. And I’m actually learning now how good of a director am I actually, because it’s going to be crap if I give it a, you know, a half-hearted direction. You know, I think it built out—it really did build out from how do I prompt for a specific image that I want, and how do I control the image, and then how do I make a group of images, and then it went into, okay, how do I make 200?
JOE:
So, you know, being able to set those structures of here is the creative process that Codex needs to go through. It’s a brief. We’re extracting all the different pieces of the brand-important elements. So it’s kind of, like, it feels like it was building a house for me. Like, the bottom floor was just do one, and then do two, and then expand from there. So I don’t know if that answers the question, but it was a gradual process for me.
PARTH:
Yeah. So you feel a lot—it’s like starting off and, like, generating one image, but then moving toward creating systems that are cohesive, or, like, a consistent kind of aesthetic, a consistent universe, is kind of where you’ve moved toward. Is that the kind of director approach—is like having a broader vision across the set of generations?
JOE:
Yeah, yeah, I think so. And, you know, to even answer that question deeper of, like, okay, how do you get to the automations? I think there are really good tools available to see what that workflow behind the scenes looks like. Like, node-based workflows like Figma, Weavy. I think that is the foot in the door for a lot of creative people, marketers in general, to understand. Like, okay, if I can see that this node goes here, this node goes here, and I end up with 50 different images coming from one input. If I do that node-based workflow enough, then I can say, okay, wait a second, I don’t even really need the node-based workflow anymore.
JOE:
I can just do that in Codex or Claude or Perplexity, because now I understand in my brain, I see the path there, and I know what it’s supposed to look like and the points it needs to hit on the way to my output.
PARTH:
Right. And this is, like, chaining multiple models together and multiple feedback rounds together. So it’s more like the final output has a bunch of intermediary steps. It’s not just one prompt or one generation. It’s, like, an orchestration of, like, this many of these different models. Right?
JOE:
100%.
REID:
And in terms of being a director, a conductor, what are some of the skills and insights that you’ve gleaned that you would give as, you know, prompts—deliberate pun, somewhat deliberately intended—to other people who would start this journey?
JOE:
Well, I think, like, macro-level constraint in prompting, I guess, produces better results. The more room you give it, the more frustrated you’re probably going to get down the road. You know, for someone building it from the ground up on their own, I would say lay out specifically what comprises a good image for you. And that’s the base-level starting point. I don’t know if that answers your question.
REID:
Well, it’s a beginning. It’s the first part of the prompt. It’s also kind of, like, you know, for example, when I’m answering it, but I’m curious—very curious about your answer—is I go, well, part of the thing is to think about, like, these AIs are kind of good role takers. So prompt it to a role. Like, say, you know, like when I was writing Superagency, it was like, okay, you know, be a deep historian of technology. What am I saying that is right and wrong? What would you elaborate or argue against? You know, like that kind of thing, as a way of elucidating better, deeper points.
REID:
But you’re kind of an artist, you know, kind of starting as an artist and a creator more than a coder. You know, this kind of conducting and directing and orchestration, I think, is exactly right. But it’s kind of, like, how do we get people into that mindset? That’s the reason why I was, you know, kind of prompting you on this.
JOE:
Yes, yes. No, it’s funny you mentioned it. I forget about those steps of the process sometimes, and it’s become intuitive. But yes, anytime that I’m going to go and build an image set, you know, if I wanted to make headshots, like, you are an expert editorial photographer in the style inspired by the style of Platon, like famous magazine photographer. And then, yeah, it’s basically—I think what we probably call that is meta-prompting. Right? Like, giving it high-level directives, what to do, what not to do, creating context for it. You know, if there’s a product, creating context for that product. So it really is, in a sense, just a creative brief document, whether it’s making imagery or writing a book.
JOE:
I think that high-level direction of here’s who you are, this is what you must do, this is what you can’t do—that, like, is the right term—latency—that decreases the latency of the potential spread of information it’s going to give you. So we’re just compacting with that creative brief. The more specific we can get with our language upfront, then everything else is downstream after that.
PARTH:
One of my favorite things about working with Joe is that he exists in this creative space, in a creative headspace, in a creative domain. And most of, before I met Joe, most of my work and exploration has been in the objective spaces. Like, you tell AI to build an app, and it’s like, okay, does the app run? Okay, it’s pretty obvious if it succeeds or fails. Did it build the feature right? It’s like a very—it’s like a yes or no. But that’s not the case in creative work. It’s not, like, one thing is better than another. It’s subjective. It requires someone who’s seen many things to know what good looks like.
PARTH:
And that’s—this is something that I wanted to unpack with you, because I work with agents, and a lot of times I can just bark at them and be like, oh, this doesn’t work, go fix it. But it’s different in creative, right? Because why something works or doesn’t work is unique to the person who’s creating it. So what would you say is counterintuitive about working with agents, copilot-type systems, on creative work in subjective spaces where it’s not obvious what good or bad is?
JOE:
Yeah, yeah, yeah. I think the agent has no idea if the work is actually good. Like, if it’s making a picture or a static print ad, it’s going to hand you the output and be like, this looks great, with the same confidence of, you know, whether it nailed it or it produced slop. It’s interesting to see that and realize that, okay, like, I can outsource production, but I can’t outsource my taste. So at least for now, that has to come from the engineer or the creative professional, or, you know, we have to intervene in that part of the process.
PARTH:
And what would you say—I know a lot of people are talking about taste, but what would you say is how you develop taste? Like, how does someone develop taste?
JOE:
Yeah, I think it’s really just, like, creative repetitions. Like, the pre-production process for me has become more important. Looking at Cosmos, different designers, looking at magazines, looking at old print ads, you know, looking at Pinterest—like, getting a clear idea and cohesive theme in my inspiration, being able to communicate that so that that’s the output I’m getting. So it really is making me go through my already-existing production process just even more precisely.
REID:
So Joe, you’ve worked with all of these tools. Leading edge—Codex, Claude, Hermes agents. So walk us through, you know, a project of your choice where this created an entirely new workflow, and go into some of the details of it.
JOE:
Right now, I’m working for a brand called Ketone-IQ. It’s an exogenous ketone beverage. We have—
REID:
Tried it.
JOE:
Oh, really? What do you think?
REID:
It’s good. It’s kind of a good, like, brain energy shot without sugar.
JOE:
Yeah, that’s it. Going to use that as an ad read. Right now, I’m set up with Hermes agents, and I use Codex for creative work, or Claude. But for this scenario, we’ll say Codex, Hermes. And I haven’t fully tapped into the capabilities of it, but I use it for research. So that would be step one. I went into Hermes and I said, hey, I want to pull deep research on 20 different successful static or print advertising styles that have proven across time, from the ’40s, ’50s, ’60s, ’70s, ’80s, ’90s, and on to social media marketing today. What kind of information can we use as case studies? Pull and prove that these templates will work in the sale of a product today.
JOE:
So I want to generate 20 templates that I can look at and use for Ketone-IQ’s sale coming up. So basically what it did is, it went and it did research and answered the questions of, here’s why this might work, here’s why it won’t work, here’s what you should change, and here’s the template description. So I then took that into Codex, where I already am connected to creative tools through fal. I like to use GPT Image 2 right now. I think it’s one of the better ones. And I give it my product—so the little shot bottle—and I also, in Codex, create some copy, some marketing copy for that campaign. So we have our templates, we have our product, and we have our copy for the sale.
JOE:
And basically, I start off small. I have 20 templates, but I want to just do a run of two or three. Like, let’s select template one through three. Let’s use some variation of this copy that I’ve supplied to you, and let’s go through our image generation workflow, because I already have a skill set where it’s basically a creative image director. So it’s going to go through my skill. It’s pulling all those elements—template one, product, copy, image. And I just let it go and give it a grade at the end when it delivers it to me. Once I got a couple rolling, I said, okay, let’s run the full flight of templates, and give me 10 different variations of each template. And then I just let it run overnight.
JOE:
And it was finishing up its generation when I woke up in the morning.
PARTH:
Yeah. That’s incredible. You’ve scaled a system around your direction of how you think about the brand media creation process.
JOE:
Yeah, yeah. And that just comes out of, again, creative curiosity. I’m not driven by fear around this stuff. And I know that’s strange in the creative space because there’s a lot of pushback on, you know, creative tools and AI. I think it’s just cool. Like, I haven’t been doing it out of, like, the need for survival. Like, I’ve just been doing it because I think it’s fun.
PARTH:
Yeah.
JOE:
So if I can add value because of this creative exploration, awesome.
PARTH:
One thing I’ve noticed about how you interact with—I mean, of course, you have Codex, and that’s one piece. But Joe and I are essentially in a Telegram chat with one of our Hermes agents. Actually, Reid, you’ve seen that. You’ve talked to this agent, but it’s one of our most creative agents. Why do I call it the most creative agent? I think it’s because of the few people it’s interacted with. Like, it works with me on graphic novels. Even Joe has had it produce a graphic novel. But now it’s gaining a bunch of abilities around how Joe thinks about ad creation and asset and media positioning.
PARTH:
And so I feel that this agent is a reflection of all the people it’s interacting with. You know, it’s learning some stuff from me, it’s learning some stuff from Joe. And that feels like the beginning of this kind of collective intelligence. Like, the agent that is at the core of the team of Parth and Joe kind of gets superpowers from both of us, but then can bring that third lens to the team. And then you’re able to port those skills and abilities to other agents in your network. Right now it’s manual. Like, he’ll be like, okay, you know, package that as a skill, send it to me, and then he can work with Codex using that new baseline ability.
PARTH:
I really feel that this, like, the diffusion of skills between people and agents is something that I’ve learned working with Joe over the last year.
JOE:
Yeah, dude. I mean, it’s been quite a resource to be able to just open up Telegram, chat, and just brain-dump and get really, really anything that I want from the brief that we’ve created—a morning brief. Another example, Reid, is like, I wanted to create a newsletter for Parth and me. It’d be delivered and scheduled at 8 a.m. every morning. And the prompt was something along the lines of, hey, tell us what’s been going on in the AI space and tech space in the last 24 to 48 hours. Not only that, but your goal as a newsletter for us is to be a future-scaper. So I don’t want you to give me information that is, yeah, it’s the news—but what does it mean? What is it? What’s coming?
JOE:
And how can we, you know, look at this information as something that’ll inform, you know, future releases or things that are coming down the pipeline or trends. So it delivers us that at 8 a.m. every morning. I think working on projects like that has just been—that’s continued to unlock my language further of, like, oh, here’s how I can think differently.
PARTH:
No, it’s incredible. It’s like every day I just open my phone and it’s like, oh, we got the latest news report. And I’m reading it and I’m just like, this is crazy. Like, you’re just passively more amplified, and you’re like, okay, what does this mean? Like, the agent doing a little bit of that pre-thinking of, like, where does this mean? Where does this go? What does this mean? Helps us think about the future and plan. Especially, like, I don’t think you can keep up with AI without AI. Like, we’re at that point now. So, Joe, what is one creative skill that you think matters more now, not less, because of AI?
JOE:
Yeah, I think it’s clear articulation of ideas, whether you’re in a creative field or a tech field. It’s—I think it’s a linguistic skill. Yeah, just being able to speak in a way that my ideas are clearly communicated. It’s interesting, because the more reps I get doing that, the more I see my communication is clear in, like, photo shoot settings or in settings that I’m directing. It’s having the reverse of what people might think it actually does, which is, it’s actually increasing my efficiency in human communication. And, yeah, also just giving me those repetitions day in and day out with getting better output with these tools. So, yeah, I mean, it’s what creative directors always have to do with humans anyway.
JOE:
But AI just made the gap between a sloppy brief and a tight brief much more visible.
PARTH:
Right.
REID:
So I have two—which may not seem apparent, but they are related—questions. So one of them is, you know, the tool habit or workflow that actually you had to kind of push yourself to get into, and didn’t kind of come naturally from step one, but was really important to do. And then the other one is, how does this play in the multiplayer game? Like, when you’re working with other people, like, not just the amplification of yourself, but, like, in, you know, AI in the multiplayer game?
JOE:
Yeah. I think for the first one, it took a while for the running automations to click for me because I’m very visual. That’s how I’ve always learned. And I needed to see it first before I could go in and just tell an agent, take me through this process. So I really do think that was an evolution of using the tools first, seeing what they’re capable of. Okay. Now I can see it in my brain—now I’m going to go here and do it, and I can see it happening, but it’s kind of behind the box. So it was automated workflows, and now I’ve kind of unlocked that by doing the repetitions up front. You know, in terms of the multiplayer, I’m dealing with that. I’m figuring that out right now.
JOE:
And, you know, I’m doing some consulting for brands that want to learn about these things and bring them into their marketing process. And for me, Reid, it’s showing them Weavy workflows. That seems to be where the light bulbs go off, because I think a lot of people have the same position as I did. It’s like, yeah, automations sound cool and everything, but I don’t understand how it gets from point A to Z. So I feel like Weavy is more of a teaching tool for me to open up the door for automations for these brands.
PARTH:
Right. Being able to see the system before you sell the system. Like, showing them, here’s what’s happening under the hood, here’s what we’re going to be doing. Gives people more confidence in the process. Brings them along. Yeah, yeah, yeah, we have time. Joe, if you have some questions for Reid, I would love to give you an opportunity to bring it to the table.
REID:
Notice how he emphasizes me versus him. But, you know, that’s part—
PARTH:
’Cause Joe and I chat every five minutes.
REID:
Yeah, exactly.
JOE:
I could talk to you guys both all day about this stuff, I’m sure. But, yeah, no, the one thing that I’ve realized doing this work every day is, my articulation is the differentiator, like we talked about. And it’s the specificity of how I express the idea that’s making the world that we’re creating land here. Which, you know, it pulls me to something that you write about a lot, Reid, and I know that you studied a lot, is Wittgenstein. Limits of my language really are the limits of my world. Except now, language doesn’t just describe the world—it kind of summons it. My question is, like, what does someone in today’s professional world have to practice to play this game well?
JOE:
And, you know, going forward, like, what is the skill that needs to be built up to be a contributor?
REID:
Well, it’s one of the reasons why I’ve had you on. I think you’ve given a bunch of good starts. And I think it’s an infinitely long path, which is good for us as people and as creators and as innovators, workers, et cetera. I think there’s a couple of things. One, on the Wittgenstein side—he was trying to figure out how it is language actually works. That we’re, like, fish in water. We communicate with each other, but we don’t really know how it works. We don’t really know how the thinking works. And so that was kind of his philosophical project, to kind of understand that. There were a lot of philosophical problems that were misunderstandings of actually how language works.
REID:
Let’s try to figure it out, since we don’t actually understand it now. On the, as exactly as you say, that it doesn’t just limit the world, it now creates the world through AI. It’s kind of the imagining what’s possible. I think it’s work to get better, not just in your imagination, but in your expression of your imagination. So, and by the way, AI can be useful to that, because you can say, hey, I’m trying to be more expressive, and so forth. So what I would like you to do is run me through exercises, and ask me about this, and interact. And when I say this, what could I have said better? What could be better? And kind of explicitly go through that loop.
REID:
And it’s part of the reason why, like, you know, I was kind of prompting you back to your early days from Fortune for kind of most people. Because it’s like, well, don’t just go, so give me a script. Like, well, of course you’re gonna get a pretty vanilla dumbass thing on that. But he said, well, you know, what if we were looking at the script for this film, but we were thinking about it the way that Matisse thought about Impressionism, and we thought about the pointillism and the way that hit the eye. What might be the way that this script might be reinterpreted in that way?
REID:
And by the way, what might come out as garbage—that’s fine, because then you go, well, all right, let’s try it with Dadaism, or let’s try it with, you know, if Dostoevsky were writing the script, what would Dostoevsky, you know, kind of—and by doing that, and by summoning, you know, kind of characters, agents, colors, synthetic alternatives, that’ll expand your own kind of language expressive palette, and will teach you to be more specific. Because part of the reason—like, for example, we would say something like, you know, Derrida says, in language, there is no there, there. I used to pun, you know, for, you know, in Derrida’s case, he’s right. I’m not sure it’s universally true. But, like, the reason why you summon specific people like Derrida, or an aphorism, or a quote, or—
REID:
Or a parallel, is because that allows you to suddenly get much richer context much faster. And I think that’s the thing that we all need to be learning. Our language doesn’t have to be that we’re all poets, but we all have to have a much richer language of gesture, expression, tools. And, by the way, I think that makes us more human. It’s actually part of the reason why I think AI and Superagency is kind of a more human story. And I think you’re living it.
PARTH:
Yeah, it’s like when you ask the AI to go back into the ’50s and the ’60s and draw inspiration on visual techniques that worked a long time ago. And I think it’s, like, it’s us connecting with the people that were before us to attack a new modern problem, even though they’re not here anymore. So I think that’s amazing. I mean, it’s a form of—it’s a form of time travel. Right.
JOE:
It’s like, if you have any—if you could ask anything at all, like, what are you going to ask? And I think that paralyzes a lot of people.
REID:
Yeah. But just start and iterate.
JOE:
I don’t want to ask about, like, what jobs are going to be the most sought-after jobs, but I am curious about that from your perspective, Parth. I’m very much a bloomer, as Reid describes. But what do you think the creative space might look like, or even the software engineering space, in, I don’t know, two to three years?
PARTH:
When Fable 5 was out for three days, I didn’t sleep for three days, and I felt like I saw a clearer picture. I thought there were things that I saw that I thought were five, six years out that I think are now two years out. So it has updated my notions here. I feel—I’ll say, first, I’ll talk about technical creativity, then I’ll talk about the expanded, regular, other forms of creative that we’ve been talking about today. In technical creativity, we are going to have the ability to build any software that you need on demand at a very low cost. Because it’s low cost and very quick to create software, we’re going to create more software, but also we will modify that software. So I feel that software becomes more liquid.
PARTH:
So when we’re interacting with Codex, it builds a little app—I didn’t show you, but this weekend, after you showed me Weavy, I had Codex build me my own version of Weavy for my video model pipeline. So I gotta show it to you tomorrow. But basically, like, the interface that we need will show up when we need it, because Codex and these systems will build the interface that we need. Right now, Codex is speaking at 50 tokens per second. Imagine if it’s at 1,200 tokens per second a year from now, running on a different inference chip. Then everything you ask for can be modified even faster. Right? Like, then I feel that software becomes—at that point, software becomes liquid. It becomes what you need.
PARTH:
It takes the shape of whatever interface that you need in the moment, and the agent kind of brings that to you. I’m very excited for this future. I think people are sleeping on this realization. And then I think that also blends back into creative, which is that, like, we use all these tools, but sometimes we prefer working with the agent, and the agent builds the tool that we need for the job. And so we have ephemeral software. We have things that are—we have the tools that are created just in time for us, for our unique project. And then, in broader creative, I feel that there are things that you can’t automate that I realize now, these are the most important remaining skills that I’ve now placed a huge premium on.
PARTH:
Like, technical skills are kind of being commoditized. The model can kind of generate code, so you don’t get points for memorizing code. You get points for designing the right system. You get points for architecting. In creative, it’s like the model can generate any video, but it’s like everyone has a VFX team in their pocket with, like, image and video models today. But actually, what people are missing—a lot of people are missing—is, like, not everyone has a storytelling skill set. That’s like, here, they can create a story that actually someone wants to watch a second time. And I feel that, like, the writer’s skill set is actually the most valuable thing that I was undervaluing in my past life.
PARTH:
It’s like, I wish I took writing as a more serious skill myself, because now I feel that’s the thing that’s missing when I’m trying to create a show, trying to create a piece of content. So then I go back and I study the good stuff, right? Like, I watch the good movies now with a different lens. I pay attention to every dialogue, every frame. You know, we’re working on graphic novels. So then I take my best friend—and, like, my friend who reads graphic novels—and he makes a recommendation, then I go buy it, right? And then I’m like, I gotta go study what’s good. I gotta figure out what good is, because we can ask for anything, right? So I think writers become world builders.
PARTH:
Soon we’ll be generating game worlds and interactive experiences, not just broadcast media. So it’s gonna get very interesting very quickly.
REID:
Yeah.
JOE:
It’s almost like people that wanna stay relevant, at least in the creative field—they think they need to learn how to be a director.
PARTH:
Yes, exactly. Yeah, yeah, yeah.
REID:
I think everyone’s going to become a director. And not necessarily all directors of, like, movies or video. It’s directors of lots of different kinds of work.
PARTH:
Yeah.
JOE:
I had another thought. I didn’t think I’d have time for another Wittgenstein, if I’m pronouncing it correctly.
REID:
Yes. Yep. There’s a quote.
JOE:
If a lion could speak, we could not understand him.
PARTH:
Huh.
JOE:
I was just kind of, like, bouncing back and forth with Claude on this. Like, when ChatGPT says, I understand, or I think, or I feel—is it really saying the same thing that you mean when you say those words? Or is it, like, the lion—like, it looks like English, it sounds like English, but it’s only coming—it’s coming from a different form of life. So is it actually understanding the words that we’re saying, or is it something else entirely? Is it just mirroring you?
REID:
Well, so, funnily enough, I spent a couple of weeks in Oxford in April and went to go talk to my old Wittgenstein tutor. He’s not old. He’s actually only a couple of years older than me. But of old is probably Bill Child, who’s one of, I think, the world’s leading thinkers on Wittgenstein. And it was precisely this. Because it was kind of the question of what Wittgenstein was trying to do, is say, hey, I’m trying to figure out, like, things have meaning in part because they have truth conditions—with the truth, that is, that the language is appropriately applied to the circumstances you’re in. And he’s trying to understand what that is. And first theory was kind of this picture theory, and the second theory was following-a-rule theory.
REID:
And it was like, well, but, you know, if a lion could speak, no one can understand him, because that seems like a different form of life from how, you know, we speak. So does that understanding work? Well, now we’re running this real world—it’s almost like this invention of the world that you’re mentioning, you know, kind of experience where we’re inventing, like, kind of language stuff here. And you’re like, the fact that there isn’t some kind of understanding going on between us and, you know, ChatGPT, Claude, Copilot, Gemini, et cetera, seems to be an odd claim. Maybe there’s an understanding there and there’s an understanding here, but there’s a very robust set of language games. So maybe the sense of what of this is, is interesting.
REID:
Now, in part, one of the models—frameworks—that I use to think about these AI models is that they’re kind of like alien life forms of sorts that have been intensely trained to be human, right? And it’s that—’cause the alien and human, that’s the interesting thing. And then what would a Wittgensteinian philosophy about understanding, truth, epistemology apply to this? And, you know, we’re like, well, someone should write that up, right? So, you know, who knows? I probably won’t ever get to it. But it’s a super fascinating question. And it has all these things to do with, you know, our world—to jump to a different metaphor, like, you know, the forests and mitochondria.
REID:
Language is how we interface with each other, how we interface with the world. And now that is all in a state of amazing evolution.
PARTH:
So much fun talking to you, Joe, over the past years, but also very glad we could bring you on to this episode with Reid today. So thank you. Thank you so much, Joe.
JOE:
Thank you, guys, for having me. That was fun.
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, Danny Garrison, Trent Barboza, and Tafadzwa Nemarundwe.
PARTH:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.