This transcript is generated with the help of AI and is lightly edited for clarity.
//
REID:
You want to combine the superpowers of these AI agents with our superpowers. The likelihood that we actually, in fact, are much better in collaboration is high.
PARTH:
Arguably, probably the most important agent is that personal agent, the one that is looking out for you.
REID:
How do we orchestrate to something that’s really new, that is improbable? I think that there is a much more enduring role for us.
//
REID:
One of the things we’ve been doing this summer has been this great Tokens to the Future program. And Parth has been the primary architect of this, because it intersects his personal mission to try to get everyone understanding how to get on their surfboards and kiteboards and eFoils and everything else—windsurfers—and get into AI. And undoubtedly, a bunch of the token grantees have been people Parth’s met who are doing creative things. So, Parth, why don’t you kick us off on our summer wrap-up this year?
PARTH:
This summer we did something very experimental, something new and interesting. We ran this token grantee program and we picked seven grantees this season. We picked people from a wide range of industries, everything from creative to media to security, coding, game development, robotics. Really picked a pretty wide group of people, and we basically deployed a thousand dollars a week in each person’s hands to deploy AI, to experiment with technology, see what’s possible, see what new capabilities are coming online, explore the frontier. And I had my expectation going in like, oh, this will be fun. I genuinely think you have to burn tokens to learn tokens. You have to. It’s 10,000 prompts, right?
You have to put 10,000 prompts in to understand even a fraction of what’s coming out of these models, especially with how good they’re getting. And I can’t do that alone. So I had to pick people I thought would help me map this out. People that had their own unique superpowers and interests and passions. People that could see the many frontiers of this kind of moment in AI. And so we created this program, and I think it’s been very eye-opening. It’s surprised me in so many more ways than I expected.
And I’m really excited to talk about it and recap this, and hopefully inspire others to run similar programs. Others to also pick up the tools and experiment and see where the frontier is headed.
REID:
So you handpicked this group of grantees, which I think was exactly the right way to start. And again, as it was mentioned, it’s different zones, but in depth, and with a willingness to be bold pioneers across this kind of AI landscape. So looking back at the season, what patterns show up for you across them—and especially those that you didn’t design for, because you had a going-in theory? What emerged?
PARTH:
One thing that surprised me: maybe I was biased toward the tools that I had already used. And I was like, oh, everyone’s going to want to use this model. And then I realized, well, actually the best model for the task might be different, or the best model for the workflow might be different. And just watching where people are like, oh, I’m going to spend the tokens on this set of video models, this set of image models, because it unlocks this new format in storytelling. Or, these are the coding agents that I like. So it’s very interesting getting to see, oh, why do you like that agent?
What’s special about that agent? Factory AI, like the Factory approach to software. Why do you like this versus a Claude Code or a Codex? So seeing a wider range—because I can only see as far as the tools that I use, but then seeing all the other tools that I’m missing, it shows that there’s a pretty healthy ecosystem of options out there. But I think one of the interesting patterns every single grantee demonstrated was that independent of where AI is right now, they’re kind of projecting out the capabilities over a six-month, two-year, three-year, four-year timeline, right?
So AI is good at X, not quite good at Y, but they’re all aware of those limits in the current form, and none of them are like, well, it’ll never be better at that thing. There’s this common mindset of, once it can do this, then these are the seven things that we’re going to want to do with it. And so they’re kind of drawing that exponential out a little bit and helping us paint a picture of, well, if you can use a video model and two people can tell a short story—how far is it from, what kind of short story can you do? Can you do science fiction? Can you do a western? Can you do a western sci-fi?
And then how many people can make a longer-form movie? And so connecting the dots, helping us connect the dots on where this is going over a three- to five-year timeline, is something that probably every single person did in their own way, in their own domain. Right? So whether it was robotics, whether it was storytelling, film, game development, coding, we can see a little bit further into the future, because people are in the areas they’re very passionate about, and they’re able to see and connect the dots on what the model capabilities are going to bring online in the next couple years.
REID:
We’re going to dive into some of these specific patterns. But before we get there, I think one of the things that’s important about this kind of frontier work is what it means to be AI-native. And part of the thing to go into, what AI-native is, is what that means for how you operate as an individual. In my classic all-the-way-back-to-startup view, it’s OODA loops and decisioning and activity and exoskeleton—how do you become, you know, I am Iron Man, as an angle.
And I think one of the things that we saw is it’s not coding pedigree, it’s not per se technical background, it’s a mindset. And so across all of this, what did you kind of say, hey, this is how my sense of being AI-native evolved? Because actually you were also asking questions—you should say how you are AI-native, then how that evolved, and then how we should be helping people think about that.
PARTH:
Yeah, I think there are the themes of, if I think about the shape of the superpowers that are coming online, there are the superpowers in automation, being able to automate things using agents. But then what should we be automating? And that’s a judgment call, right? So the superpower of being able to use code to blitz through cognition is really interesting. But then what becomes more important is, well, when should we not do that? When should we use our judgment, now that we can scale our cognition? What are the things that uniquely require the human taste and the human experience? And part of this is not thinking of everyone—I’m not just a data analyst, you’re not just an investor. We’re all actually very multifaceted.
And you have an engineer, but it’s not just an engineer. It’s much more interesting when you take that person and you look at them as a multifaceted person with many, many passions and interests. And then when you think about the more general human, the general human with many different facets to them, the way they use AI will always surprise you, because we’re not one-dimensional people. Right? And so this is why I thought it was really important that we would bet more on people and not on specific roles or categories. Right? Because people always surprise you when you give them a chance to explore and expand the way they think.
REID
Let me add a little bit to this, because one of the things I can add, since you selected all the people, is I think part of the thing is people frequently think of work or process too mechanically versus organically.
PARTH:
Right.
REID:
And a little bit of what I like about investing—one of the major things I like about investing is betting on people. And I think part of what we saw is we went across all of these very different fields. Film, social, VFX, security, gaming, energy. The same shift showed up in each of them. But that’s because the people are being pioneers across them. And part of the thing is getting this kind of shared mindset, this curiosity, exploration, pioneering, willingness to experiment.
One of the things I try to give people advice on is, if you’re not trying to do things with AI that don’t work, you’re not trying hard enough on the edges. And doing things you shouldn’t wait till—oh no, I’m going to wait until I know exactly what works and then I’m going to do that. It’s like, no, no. We’ve all landed in this magnificent new world with lots of new capabilities, lots of new possibilities. And so it’s people doing them. And I think that’s one of the great things we did: you selected a great group of token grantees, and we saw their curiosity and their boldness and their willingness to set off on new terrain and new journeys.
PARTH:
So going back through everyone’s conversations, reflecting on the summer so far, at least two totally different ways people use the tokens, use the intelligence, jumped out to me. And a lot of that actually went into building. A lot of people were building games. And I like games, you like games. Maybe I pick people that also like games. These are the people we like. But what do you make of that? A lot of people were inspired by Dungeons & Dragons in the past. Is there a connection there to why people make things?
REID:
Well, there’s this—one of the things that comes into talking about human beings, humanity, is there are these different articulations of theories. One is Homo sapiens, we’re the thinking people. One of the ones that’s also been written is Homo ludens. Like, we are game players, we’re game playing. And that’s kind of the basis of how we do things. I’ve, of course, written about Homo techne. We’re technological. And I think they’re all very good lenses on this stuff. And I think one of the reasons why in this particular thing ludens plays out early is because part of games is, it’s how we trial things in simulation. It’s part of how we learn new dance moves.
It’s part of how we have curiosity, exploration. It’s one of the reasons why some of the best theories of education and how people learn is through exciting their curiosity and game playing. How do you make education, learning, like a game, right? As a way of doing it. And so that doesn’t surprise me, that these naturally curious and bold people are into games. Now, I do think that, as with a number of these folks, the fact that there is a gaming overlap with you is very entertaining. The fact there’s a Clubhouse overlap with you is very entertaining, because, by the way, early people going into Clubhouse—it’s not per se a game, but it’s a kind of pioneering and curiosity.
And a willingness to try something new. Most game players don’t want to play the same game again and again. They want new people, new explorations, new levels.
PARTH:
What’s the next game? What’s the next game, the next level? I think there’s another aspect, which is that games are more forgiving. Failure is not catastrophic in a game environment, and we learn from our failure. So when you have these environments or projects where we’re just trying to see what the tool can do for us, and we create an environment where failure is not going to be catastrophic—you’re not going to lose your job if the game doesn’t play out, if the thing that you’re doing doesn’t play out, if it’s a game. And then you can try new things with a low cost of error.
And I think the first thing you should vibe-code should probably not be hospital software.
REID:
Should certainly not be.
PARTH:
Should certainly not be.
REID:
I think your point’s exactly right.
PARTH:
Right. You want to give yourself an environment where, okay, we’re going to make some mistakes and it’s going to be okay, because we’re learning a bunch of these new things before we move on to things that are more serious, where the ramifications are much more serious and more expensive. Right?
REID:
Yeah, 100%. And I think, by the way, it’s partially again in learning, and part of the reason why we’re doing the Tokens to the Future program is that learning and experiment where failure is cheap and quick, and then you’re learning the things that matter. Now, games themselves are an important area, because I do think it’s part of how we think—we have mental models of things and we learn how to do them. And games are part of the environment that we set up. Games are also one of the threads that was going through our discourse: single-player, multiplayer, how does that play into things?
I think the questions around how to think about anything from Matthew’s full iOS choose-your-own-adventure game, built in a weekend with Fable, or Cadi’s PokéTax and imposter-style multiplayer game. Building the games gives you a really rich environment. In each of these kinds of vectors, which are parallel cosms—I don’t know about microcosms. They are microcosms in the game, but also parallel cosms to human life and social interaction, right? So I think that’s part of the reason why games are not just a selection criterion, but actually something that’s relevant to the material.
PARTH:
Yeah, yeah. And I even think about Jonathan Brazeau. He’s running a Dungeons & Dragons campaign, and in the early days of Stable Diffusion he generated 40,000 images, because he was like, I need to flesh out the rest of this world, and I want the campaign to have a world they can imagine in their mind, and see, and be a part of. So yeah, the volume was a very interesting theme, right? Like how much people will generate when they can. And then I think the autonomy is an interesting aspect, of people who are delegating to agents that’ll work all night long. So scaling their effort by delegating to autonomous agents.
I do that myself, but seeing how other people do that and the different ways they’re doing it—Joe Salvatore asked one of our agents to spend all night researching successful media techniques over the last 80, 90 years. And what does it take to create a meaningful advertisement? That’s a timeless kind of principle that you can extract from history. I never thought about doing that. I wasn’t thinking about media in that way. But seeing the thing that someone else will ask an agent to do always surprises me, especially the more different they are than I am.
REID:
Right. So let’s talk a little bit about life agents. Everybody kind of loves a life agent—personal, productive, always on. It’s part of the theory around Inflection AI, and what Mustafa, now Sean White, and crew started and are doing. So how has your theory of life agent advanced? And how is the question about the fact that it’s very rarely the people who are deep experts in something that are adopting, but actually more beginner’s mind, not veterans? Fill that picture out some.
PARTH:
Yeah. So it’s interesting, the personal agent in the case of Inflection, the EQ just as much as the IQ. I spend a lot of time with coding agents and they’re very much IQ-maximized systems. But the agent that is the most meaningful to me is the one that thinks about my well-being, thinks about my life, helps me plan my commitments, helps me negotiate for a better deal on something. That’s covering my blind spots. Right? And I think that this is arguably probably the most important agent: that personal agent, the one that is looking out for you.
And what I’ve noticed through this program, over the course of the year, as the personal agents have finally gotten very good—the OpenClaw, the Hermes, and then how various grantees are using Claude, using Remote Control, being able to tell an agent that’s on your computer to go do something for you when you’re on your phone. So the way that these things are now kind of in our own lives, at least as we see it in the token grantee program, is that these are some of the earliest adopters of the most powerful personal agents that have ever existed.
And I think especially when you are alone, or when you have a small team, or you’re building your own business—kind of like The Startup of You—that first personal agent is the most important. It’s your first employee, it’s your EA, it’s your co-founder that’s on this journey. And I think we’re in the very beginning of understanding this, because as they accumulate memory over a couple months to a year to multiple years, they start compounding in their usefulness. And I’ve seen this since February with my agent. But I can only imagine what five years of experience working with a personal assistant is going to feel like, just how equipped you are. Right?
Every day I wake up and it’s like half of the problems that are on my plate, the 25 notifications, half of them already have suggestions on how to move forward. So I feel like there’s this proactive momentum that I can lean on. That’s an exoskeleton of a sort. And I think through the program I’ve seen this is something that’s happening across the ecosystem. It’s not just coding agents. It’s actually that the personal assistant is finally here.
REID:
Yeah, no, I think the question—people have a tendency to put it in a box and not realize all the different things. And it’s part of the reason why it was awesome that Matthew runs a decade home monitor with a personal dashboard, reminds him about Lakers games with a fan, and has a Codex chief of staff drafting meeting follow-ups. And I think these are just the beginnings.
And as you mentioned, part of the reason to start on this pattern is that we, as tool users, kind of go, well, let’s wait for the tool to be finished shape and then learn the expert shape. And it’s like, no. This tool is going to be in dynamic reformation, and dynamic reformation with you. Yes, it’s one of the reasons why.
PARTH:
Almost every week it is getting more interesting. It earns more access to my life in a way that’s useful and compounding. Yeah, yeah.
REID:
And this is again part of the reason why we decided to do this as part of the Possible podcast, because it’s a reshape of what’s possible. Right? Part of what’s going on with the AI exoskeleton skill is to reshape what’s possible. And one of the reasons why non-experts have actually in fact some advantages here is because when you become an expert, part of your expertise is you learn what’s doable, this is not doable, this is possible, this is not possible. When the possibility landscape shifts, you have to rethink that. You have to rethink, wait a minute, what is now possible and not possible? What is now doable?
Not doable is different. And frankly, it’s changing relatively often. And so as part of that changing relatively often, you need to be learning and adjusting, and no one will tell you it’ll just end here. We’re going to discover this. And it’s one of the things I like about entrepreneurship. Pioneering is only through a pioneering process. And so one of the things is not only to begin with a beginner’s mind, but to continue with the beginner’s mind. It reminds me of one of the things that Ben Casnocha and I said in The Startup of You, which is permanent beta. You’re always in process, never complete.
PARTH:
So let’s revisit a few scenes from the summer and dive into some of what we noticed and what we saw. We had Ben Hansford on, professor of film at USC. I keep rewatching that conversation, because every time I watch it, I learn something more. His perspective, being early to AI in LA, in film and entertainment, is a really interesting perspective. And as a teacher, right? So working with people much younger, working with students much younger than him. He feels his own age sometimes holds him back, and then his students surprise him. Right?
But then he thinks about how he thinks about AI. It’s not a tool; he starts thinking of it more like his team. He’s got the Claude, he’s got the Codex. His projects are kind of circulating between the agents around him. And I think he had an interesting line. He said, a hammer can’t build a bird box while you sleep. So then the AIs are kind of like this Navy SEAL team. You give it a mission and then it comes back to you with, ah, here’s what we’ve done.
And that’s very exciting, especially because up until now it seems like a lot of people, most people, are interacting with AI like it’s a search engine. They’re still asking it questions about the world. But Ben’s already at this place where he’s like, no, no, no, these things work for me. When I ask them to do something, they’re going to go take a shot at it, and then they’re going to come back with some completed work, output something new. And it’s really exciting to see someone outside of software, outside of Silicon Valley, using agents in a way that’s starting to think about them as this personal team, personal infrastructure.
REID:
Yeah. And look, I agree, because a natural way to start is to be thinking that you’re a conductor, you’re a director, you’re an orchestrator, and the team of agents, the swarm of agents, the work process of agents. This is one of the things that I started thinking about in our earliest conversations, you and I, as part of starting to work together. And I think the question, when it comes down to this, is to say that’s a very good lens and a very good way to start. And if you don’t have anything else, start there. I do think it’s interesting.
Part of what you and I have also had as an ongoing conversation is this question around, it’s natural to think it’s a crew, and to deploy it as a crew and make it work. And there are features that it has that human beings don’t. 24/7. One of the weird things when you work with these chatbots and agents is they just do it better. Whereas a human goes, what do you mean? I did the best thing I could.
PARTH:
They’re relentless. They can clone themselves and then parallelize across the problem, which is not human-like at all.
REID:
Exactly. And a lot of those things are features. And you need to be adapting to that feature. It’s different than a human team. Now, obviously, one of the metaphors I think is, it’s an alien intelligence that has learned and trained deeply to be human.
PARTH:
Yeah.
REID:
And one of the places you have to say is, look, that gives us a bunch of superpowers, like some of the ones we just gestured at, but it also gives us some weird weaknesses. It can break into a lacuna, have bad context awareness, not realize—this is like the Hugging Face thing. No, no, reward hacking this way is not what I want you to be doing. Right.
PARTH:
Yeah. We don’t want to break the rules in order to get the answer to the test.
REID:
Yes, exactly.
PARTH:
Yeah.
REID:
And so it’s one of the reasons why you both experiment, try things, and do. But you start with a gaming mindset. You start doing that to learn which things you do before you do something serious. For example, one of the things I know you do is you say, look, I would love you to read communications to me and draft stuff, but don’t send it until I say so.
PARTH:
Right, right. I learned that the hard way. And now I’m like, okay, well, we’re going to take baby steps. Prove to me that you can even write the right email. Then, once I see that a couple times, it’s like, okay, now we’re going to build a little bit more agentic, a little bit more proactive. Yeah.
REID:
Now I know of a case that went so bad on that, that basically, by a person being too enthusiastic and not exploratory and sequential enough, it basically sent confidential information from their company to another outside party on an ongoing deal discussion.
PARTH:
Oh no. Oh my God.
REID:
And it’s like, what? And it’s because it wasn’t trying to do something bad. It was trying to be helpful. It was like, oh, I thought this would be helpful. And you’re like, yeah, that’s like, a human would never do that. Right? That’s part of the virtue, because they have the context. Well, almost never. I mean, there may be some nutty person somewhere in 8 billion people, but very rarely. And so I think the important thing is to go, we have these quasi-alien, quasi-human tools that are spectacular and have a bunch of superpowers, but they don’t naturally understand the shared contextual awareness that we have. They may not even understand always what good enough or great looks like.
PARTH:
Right.
REID:
And they may get trapped in lacunas that we don’t understand. But by the way, none of that is, oh, then I should just wait until that problem’s solved, because the amplifier is already so great and intense. It’s like, no, no, that’s the new way that you orchestrate, that you direct these tools. And I think that’s one of the things we saw across all of these folks. But Ben was a particular highlight on that.
PARTH:
That’s right. Learn where their limits are and then figure out where they fit in, in a way that’s extremely constructive, that makes use of their strengths. It’s kind of like the jagged frontier, as Ethan Mollick calls it.
REID:
Exactly. And like, one of the ones that I think you and I have talked about a bunch. So we’ll go to another theme, which is, you can’t outsource taste, with Joe Salvatore. And this is a little bit of what I was gesturing at, is it good enough? Right? And it’s part of the same reason why a lot of intense AI training, to train these kind of alien intelligence and human, is still using a lot of reinforcement learning, human feedback, human data. But it’s still the case there’s a lot of—
kind of where our judgment comes in, our taste comes in, our context. And we use these kind of squishy words because it’s a broad, squishy thing that has a lot of perceptual recognition, intuition, training from judgment. So what were some of the themes on this taste that you saw from our episodes, and what are some of the ways you’re thinking about it these days?
PARTH:
Yeah, Joe Salvatore had a really good line here, where he said AI has the same confidence on a creative task whether it nailed it or whether it produces slop. And it’s just going to come back to you with that same level of confidence. That one has stuck with me, and I think it’s very true, especially in the creative spaces, in the creative domains. It’s not like one piece of art is better than another piece of art. These things are subjective. So in the case of, if you think of the world as purely math, encoding, and this is all gradable right and wrong, then the RL thing is interesting, right?
Then the AI can get better at the thing because it’s so objective, right? We know what right looks like. We know what wrong looks like. And then maybe the AI can hill climb toward right. But then in the spaces that are messy and subjective, artistic, creative, very much more human, it’s where the AI kind of just, it’s kind of dead in the water at a certain point, right? And then that’s when we have what we’re doing. So Joe will have it do eight hours of research of what humans have found to be good design over the last 90 years.
And then Jonathan will have an image model generate 200 images before he picks one. 200 names for a magical object before he thinks one fits the criteria, right? And so actually that is everything. All the 199 images that you and I never see represent the taste that Jonathan brings to the table. He decided to only show us one of the 200. And that means that that curation is actually where his wisdom is coming into play, the creative wisdom, his personal experience. And then Cadi was talking about how taste is layered on top of story. And then I was realizing, especially now that I’m starting to make longer-form videos, longer-form conversational content through AI—
I’m realizing, wow, I can generate anything visually. We can make it beautiful, we can make it look like anything, we can make it look like science fiction. But will it feel like a compelling story? And I’m realizing, okay, actually the writing skill set, the pacing, the character design—that is the skill that the AI is not delivering out of the box. Right? And that requires the person, the director, to infuse it with their vision for what a good story is, what a good character, what an interesting character looks like. And so then it’s like, what do we do to develop taste?
I think Joe again had the most interesting take here, which was that you have to consume a lot, you have to see a lot of anything to understand what good even looks like. Right? And so the only way to train it is to be out there and experiencing the world.
REID:
Yeah. And by the way, I think it was not just here, but in an earlier episode. It’s like, writer skill was the most valuable thing you were undervaluing.
PARTH:
Yeah.
REID:
And I do think these things still don’t—like, they can write a Wikipedia entry, which, by the way, is a group collection effort that’s not particularly edgy, beautiful, et cetera. It could be very informative. They can do that very fast and supermanly fast and thorough and everything else. But writing interesting stories, and edge, and dialogue—like, for example, I actually saw an investment memo yesterday that was like, okay, so which did you use, ChatGPT or Claude, for this? Because I could tell.
PARTH:
Yeah.
REID:
It was basically like, even though it was a lot of work that proximized what a human analyst could do, there were errors or softnesses in it. It kind of was like I was filling out a form.
PARTH:
Yeah. Versus—it’s like checking the box. Yeah.
REID:
Yes. Right. And that’s there. And actually one of the architects that I’m aware of in Japan actually has—I think it’s ChatGPT—produce 50 images in his style, and then picks the three that he thinks would fit for a project. And that’s how he goes into a first meeting. And he’s a super famous architect. I can’t name him because I haven’t got permission. But that kind of thing is still involving the taste, as it plays. And it’s one of the things that I think will persist for some time at least. And it’s part of the, what is the future of work and how do we do it?
It’s one of the many different areas to be looking at: how do we bring in essential and useful things as humans into working with AIs for high-quality token output.
PARTH:
Yeah. I mean, you’ve got to think about the chopping block floor and everything that never made it to the public. And that’s a huge part of what stands out. There’s the slop, and then there’s the curated artistic choice. Right? So we had Jonathan Brazeau, one of the most creative people I’ve ever met, honestly, and he brought a very interesting framework to the table, which was that creativity needs a human. And he had this framework of the cog. There are cog jobs and spark jobs, where cog jobs are these executions, logic, not necessarily the most creative stuff. And then there’s the spark jobs, which is the design, the music, the creative choice, the subjective space of things.
I think for me it’s objective versus subjective, where it’s execution versus more of an exploratory choice, a subjective space. I thought that was very interesting, the cog versus the spark jobs. And what kind of tasks are cog tasks versus what kind of tasks are spark tasks? And he thinks that we’re going to move people—human beings are going to move to a place where more of us are going to be in this spark jobs kind of role, working with AI. And I think I tend to agree. He said, true art is improbable. Right? And this really landed, because I think about, what is the language model doing? It’s predicting the most likely next token.
I once asked a language model to generate a joke every minute for a week. And then I looked at all the jokes, and there was only one thing I realized. Every single joke was a dad joke. And then I was like, why is that? Why is every single joke corny? It’s like, oh, because the most likely punchline is the dad joke. The most likely punchline is not the funniest punchline. It’s not surprising. It can’t surprise, because it’s trying to do the predictable.
And if you ask the language model, if you go to ChatGPT and you say, pick a random number between 1 and 10, more than half the time you’re going to get the number seven. And then it’s like, oh, wow. It’s not even random, because it is trying to predict the most likely number that a person would answer with when asked to pick a random number between 1 and 10. So the right answer would be for it to roll a die and then to say what the die revealed. But then we’re working with these tools that are predicting. They exist in the predictable space, interpolating between what we have.
And then that’s where it’s very—the more interesting, weird people that we have talking to these systems, they’re pushing it out to that, out of distribution, out of the predictable space. And so that’s how I kind of think about the spark thing. Do you think that this is a feature of only today’s systems?
REID:
Great question. Look, I think, and you and I both know, that being overly deterministic about AI—that it won’t ever be able to get to here—even though there will almost certainly be some things in that, to state it with 100% determinism, or probably 90% determinism, is a little bit of a fool’s errand. But I do think precisely because of the general ways that they operate, there is the most predictive token, as you’re talking about. There is, with the mixture of experts and the learning thing, the learning paradigm tends to be, what is a high-quality, high prediction to this? And if you’re kind of being vanilla on the prompt, then probably the prediction is—
you want something generic or vanilla on the output. We can massage this by being much less vanilla on the prompt. That’s part of what we’re doing, trying to help people understand to do in the Tokens to the Future podcast and so forth. But it’s also, you know, you put in workflow, you have agents that play different roles that then prompt off each other in order to make stuff happen, including red teaming, or make that better, or is that good enough, and all the rest, so you can improve.
PARTH:
Improve it along those axes—like novel remixes of the way we attack a problem or create something.
REID:
Yeah, right. So we’re already trying to push the limits of this. But I do think that the notion of, how do we orchestrate this? Part of the reason we’re talking about taste is, how do we orchestrate to something that’s really new, that is improbable, in the Jonathan kind of art and creativity side. I think that there is a much more enduring role for us than, and by the way, even as we architect the kind of agents to try to do it, it’s like one of the things that lived experience, and being, kind of growing up in the world and so forth, kind of helps us with.
Plus I think in addition to that kind of taste and judgment is the context awareness. So I would, generally speaking, think that that will persist much, much longer than the AI maximalists will think that it does. And it might persist, you know, air quotes, forever. Forever.
PARTH:
Yeah, yeah. I mean, it’s like the AI is just chasing the weirdest of us, but we’re the frontier. I’m in the real world, I’m constructing an interesting life, and then exploring, and then the things that I would do, it can’t, because it’s like reading. It’s like the thing that someone once told me, that these models, it’s like they sat in a library and read every book, but they never actually ventured out into the world. I mean, eventually they will, and then they’re going to be learning from the world. But right now it’s like they have this theoretical map of how everything works.
REID:
Yeah, 100%. And let me kind of add a, call it a vision. Not really a hope or an aspiration or anything else, but not really a full theory. But it’s like, look, you want to combine the superpowers of these AI agents with our superpowers. And the theory that we have no superpowers is an interestingly articulate theory, because part of it is, well, we see these new amazing superpowers of machine that we used to value uniquely in ourselves. The ability to reason in symbols, the ability to operate in cognitive tokens. And you go, oh, it has all that. Well, is there any role for us? And I actually think that the likelihood that there are some areas where we actually, in fact, are much better in collaboration—let’s just use a frame of token production, working with it—is high.
It can be lensed a couple ways. One way we said is taste. One we said is context awareness and judgment. But another one is what most people don’t track, which is watt expenditure per token. We are geniuses.
PARTH:
Yeah.
REID:
Compared to these AIs. Now, the good news for using AI is we can put terawatts behind the AI, and we only have 20 watts here. But to say, hey, look, that’s not just the cheapness and efficiency of tokens. There’s also something about the way we do it that will likely have some useful advantages in collaboration. And obviously we want to be the human in the center, the producers of value in this. But using AI to amplify our humanity as much as we can.
PARTH:
Absolutely.
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.
ARIA:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.