Subscribe
Share
Share
Embed
NET Society is unraveling the latest in digital art, crypto, AI, and tech. Join us for fresh insights and bold perspectives as we tap into wild, thought-provoking conversations. By: Derek Edwards (glitch marfa / collab+currency), Chris Furlong (starholder, LAO + Flamingo DAO), and Aaaron Wright & Priyanka Desai (Tribute Labs)
00;00;16;01 - 00;00;18;14
Aaron
Hey, everyone. I guess we're back in a 1989.
00;00;18;14 - 00;00;25;10
Chris
1989 a number. Another summer sound of the Funky Drummer definitely feels 1989 to me.
00;00;25;16 - 00;00;27;16
Aaron
Yeah. Why? Why does it feel that way to you, Chris?
00;00;27;18 - 00;00;52;20
Chris
I was 13 at the time, so let's also, like, just ground ourselves here. I think it's, it's a combination of things, you know, like, I was running around Boston at the time I was 13. It wasn't the safest place to be. You kind of always had your your head on a swivel. And, you know, there was just this sense that when you were out in public, you couldn't be relaxed.
00;00;52;23 - 00;01;16;22
Chris
You know, certainly with what happened in Minnesota, just, you know, what seems to be like, a fairly brazen disregard for other human life and just seeing a lot of the reactions out there, you know, like there is a real lack of security and, you know, like, you just can't be at peace being about, you know, the world you're supposed to inhabit.
00;01;16;22 - 00;01;46;14
Chris
Right? I think that's that's one part of it. Another part of it is the global instability. You know, we we could have like, another change in, leadership over the weekend before this episode airs, you know, and, 1989 was the year of the Berlin Wall when, you know, the Soviet Union bit the dust. And so I think, like, just sort of that sense of volatility, it was also the Bush.
00;01;46;16 - 00;01;50;14
Chris
Well, we were just into like, Bush senior.
00;01;50;16 - 00;01;51;25
Aaron
Year at Bush one.
00;01;51;27 - 00;02;21;05
Chris
Yeah, that was Bush one. You know, like just a feeling that your government was, you know, he wasn't really a great president. I mean, was he a shit president? Not necessarily. But like, was it just like a stale continuation and, you know, imply that Trump is stale, you know, isn't is certainly the wrong take. But, you know, just like looking at looking at your, your government and just being like, just fuck is this shit, you know, like this isn't.
00;02;21;07 - 00;02;40;18
Chris
So I guess what I'm saying is there's just a bit of a vibe out there where I look at everything kind of coming together, and it's just it gives me flashbacks. You know, the other part of it is just my head's in that space. My book club's doing bonfire of the vanities, which is 86. How's that holding? Six.
00;02;40;20 - 00;02;42;03
Aaron
Yeah. How's that holding up?
00;02;42;06 - 00;02;47;01
Chris
It's, it's still a banger. Yeah. Shocker, shocker. Tom Wolfe can write a book.
00;02;47;02 - 00;02;48;02
Aaron
Yeah. He's great.
00;02;48;05 - 00;03;17;28
Chris
Yeah. No, but it is interesting. Especially, like the guy can write a book. I'm only 100 pages in. God, it's 600 pages. So, you know, I gotta. I gotta put some miles. Miles under it. But I do think that, you know, this, portrait of unbridled excess or just wanting things very nakedly, embracing the kind of separation of who who has things and who doesn't.
00;03;17;29 - 00;03;37;18
Chris
I think the emphasis on, like, style and superficiality, like, there's just it does feel, I don't know if it's still current and of the time, you know, I mean, I think that's all all, all great storytelling is, you know, it doesn't feel dated, no matter what the period is. And so, no, he's got it. He's got it.
00;03;37;24 - 00;03;58;21
Aaron
I think you're right, though, that it just it what's that phrase? It's like, son. Like, sometimes it feels like years happen in a week, and other times it feels like, you know, nothing happens for years, it feels like, we're in one of these, like, every week, something big and major is happening, right? Like last week, it was things around Venezuelan Maduro.
00;03;58;23 - 00;04;19;06
Aaron
This week, it's Iran. You know, there's like protests now and and people on the street in Turkey. But, related to, kind of the, the Kurds, it just feels like there's like a lot of stuff that's going on. I mean, in some ways, like I, I'm kind of glued to X. I don't know if you've been I just feel like in weeks like this, X just really shines.
00;04;19;08 - 00;04;31;09
Aaron
I just keep on finding myself, like wanting to know what's like happening kind of next, wherever it kind of lands. But it just feels like there's a lot of stuff that's like shifting, like across the board, which is pretty wild.
00;04;31;12 - 00;04;59;06
Chris
All right. So yeah, I mean, look, most of my life is, an episode of waiting for deploy, and so I'm on Twitter a lot. I don't really particularly like being on there. You know, I see your point from a breaking news perspective, but, you know, at the same time, I don't know, the the more time I see it, I just feel like the world's gotten very one dimensional and nothing is flatter than than Twitter.
00;04;59;11 - 00;05;02;19
Chris
And sure, I guess maybe I'll call it X while I'm hating on it.
00;05;02;21 - 00;05;10;17
Aaron
Do you think it's Twitter or is that just the timeline? I think just timelines create that mono dimension. Dimension dimensionality.
00;05;10;20 - 00;05;17;03
Chris
I, I hate to say it, but I think it's, it's more than the physics in the internet. I, I think it's society today.
00;05;17;06 - 00;05;38;22
Aaron
Yeah. Well, you know, it's been keeping, you know, my attention the most. This is like a sleepy story, but I think it's an important one. Did you did you see that? 5.20, I think it's successfully. And fully autonomously resolved some complicated math problem called Erdos problem. 728 basically, it kind of did it by itself, which is kind of wild.
00;05;38;22 - 00;06;06;04
Aaron
And I think this was announced, what, a couple days before Christmas, if I'm not mistaken. And I think the mathematicians have kind of weighed in and, and that validated that it's the right solution, which is kind of wild. So I think we're having kind of like the first indication of like just like your off the shelf model being able to kind of crack some of these, like really hard or previously unsolvable problems that, us mere humans haven't been able to tackle yet.
00;06;06;10 - 00;06;14;12
Aaron
Maybe that's part of it, too. Maybe the actual physics of, the universe is going to change, Chris. Like, literally, we will be able to wrangle it in different ways.
00;06;14;14 - 00;06;28;05
Chris
Great. I know you love these, red. I do random problems that we've never heard of, solved by machines running some impressive length of time. Or you're you're a sucker for this stuff.
00;06;28;08 - 00;06;46;24
Aaron
Yeah, this is my catnip. But I think it's my catnip, in part because, I don't know. I just feel like there's so many things that are downstream of these, like, core fundamental breakthroughs. And I've always been sympathetic to this like point that there hasn't been much development when it comes to physics. Or math over the past couple decades.
00;06;46;26 - 00;07;06;06
Aaron
You know, different people like, argue as to the reasons why some people think it was like this, like mouse trap that the physics ecosystem got kind of caught up with, with, examining string theory, you know, other people, you know, claim that it's just like, you know, you're not getting like newer blood in academia or the different institutions that are looking at it.
00;07;06;06 - 00;07;39;11
Aaron
But, you know, if you want to kind of build this vision and world of the future, there just needs to be like a, a number of these, like, core fundamental breakthroughs and my gut always told me it's not going to start with like, physics. It's just going to start with some math stuff. Right. And then as, these systems tackle math and next will be physics and then, oh, wow, we can like, develop this new type of material or oh, wow, we can go faster or maybe change some of the things that we thought were unchangeable, like those challenges related to the speed of light or quantum, right?
00;07;39;11 - 00;07;57;01
Aaron
Or like all these other types of things that seem like really, really complicated hard problems. So, I don't know, this is like, feels like the little shoots out of the ground in that thread, of acceleration that I feel like I feel like that we're going to just see a lot more of this over the next, couple quarters and years.
00;07;57;01 - 00;07;57;23
Aaron
Obviously.
00;07;57;25 - 00;08;02;12
Chris
You've been eagerly waiting. Us. Overcoming the speed of light is something that keeps you up at night.
00;08;02;15 - 00;08;20;07
Aaron
It doesn't keep me up at night, but, like, I actually think I think it matters, right? Like, even if you look at, like, blockchains, like some of their fundamental constraints is just like the speed of light. Like you have to send messages between all these different nodes in the network. You can make that faster. You know, that means that decentralized systems can compete with centralized ones.
00;08;20;07 - 00;08;38;12
Aaron
Right? Like one of the reasons we have centralized systems is actually because of constraints, due to the speed of light randomly. But I know yeah, it doesn't keep me up, Chris, but, it's something I want to see, like progress there. Like I'm super geeked on the, the idea of like an AI, like AI scientist of some sort.
00;08;38;15 - 00;09;04;04
Chris
Let me start by saying we have 10,000 TTS blockchains sitting absolutely barren right now. Crypto may have a lot of problems. Overcoming the speed of light to increase its throughput is not not a priority at the moment. But I do see your point around the flow from math to say, physics to applied science to new materials. I suppose we've been a little stale there for a while.
00;09;04;09 - 00;09;26;00
Chris
We get teased every once in a while with, someone solving, you know, graphene or someone, you know, coming up with a different form of chip design and then, like, nothing ever really comes of it or, superconductivity aren't we do for another like, you know, superconductivity. False flag.
00;09;26;03 - 00;09;46;28
Aaron
Yeah. Completely like I'm ready for all these vibes, whether they're I hopefully they're not false, but they're like fundamental breakthroughs. But once that like flywheel starts moving like in the next like the years after that, I just think with the internet with, like, other kind of adjacent advancements, just like this stuff comes to market like a lot, a lot faster, like graphene.
00;09;46;28 - 00;10;13;14
Aaron
I don't remember exactly when that was discovered, but like, why is that not, like, widely used at this point? I think it's been a while. Right. Like almost 20 years since that, that superstructure kind of was identified. It seems like it has a lot of applications, like, why is that not like, rolled out in mass. And I think part of that, you know, maybe adjacent challenges could be, you know, challenges in manufacturing it, I don't know, 100%.
00;10;13;14 - 00;10;21;19
Aaron
No. But hopefully we get to move faster on those bits. And we're kind of building this, this world of the future that I think we're we're heading toward or we're headed towards.
00;10;21;25 - 00;10;45;04
Chris
I will say like, you know, super strong carbon type materials like that was a whole thing, that kind of disappeared a while ago. But then, you know, I don't know, we've got Kevlar vests and we've got it in airplane wings. And it happened, didn't it? You know, it's a little strange where sometimes we have stuff like this that people talk and talk and talk about, and then it does emerge.
00;10;45;04 - 00;10;57;21
Chris
But I suppose in this case, we still have a ways to go, I, I suppose, but I do remember that like being, being more in the, the zygote used or something people were anticipating.
00;10;57;23 - 00;11;18;05
Aaron
Yeah. Like a couple years ago. Right. But I don't know, I just think like, all this stuff, it's like sleep. But you know, they're using graphene now for battery charging right now. New forms of semiconductors, other super iconic, conductivity, you know, advancements. I don't know, it just seems like it, you know, these breakthroughs, they take a while to kind of permeate.
00;11;18;05 - 00;11;39;26
Aaron
But, I feel like they're going to be super important. But I think you're right, Chris. I don't know if we need, another a faster block blockchain of some sort. Although I think eath last week hit. It's kind of all time high in transaction volume, which I thought was pretty cool, whether there's demand for it or not, like the throughput on these systems to have definitely like opened up, which I do think is pretty notable.
00;11;39;28 - 00;12;01;16
Chris
Well, congrats to these systems. I mean, Vitalik seems very, both toasty and very excited by, progress roadmap, you know, solving, solving his trilemma xQc snark, you know, approaching production, implementation. And so I'll, I'll defer to him on the matter as the guy who, you know, built the whole damn thing.
00;12;01;18 - 00;12;25;18
Aaron
Yeah, completely. I was kind of I mean, also interesting to see, like, with all this, like, kind of global unrest, you know, it does feel like there was like a little bit of price movement from Bitcoin and eath and some other kind of core assets, which I think is kind of continuing to prove out, you know, this notion that these are non correlated assets or kind of cool like in, in their nature, which I thought was pretty cool too.
00;12;25;20 - 00;12;33;06
Chris
I always happy when the price doesn't go down. I've given up trying to figure out, well yeah I'm neither a buyer or seller.
00;12;33;06 - 00;12;35;06
Aaron
I'm just, you're a hodler.
00;12;35;08 - 00;12;53;20
Chris
Yeah. So, you know, when my hodling is, is proven out and. Sure, like, that's always good, but, I don't know, I, I, I have no idea what happens with crypto other than, you know, steady installation and then question mark. And so I just let it do its thing for now.
00;12;53;22 - 00;13;09;07
Aaron
I think that's right. I think it's just going to be, you know, a continued kind of grind on that front. I do wonder with like all the math stuff too, I've been, you know, I know like Nick Carter and some other folks in the Bitcoin ecosystem are getting increasingly worried about quantum. I wonder if that stuff accelerates a little bit.
00;13;09;10 - 00;13;31;27
Aaron
I wonder if we get like accelerated quantum development and that becomes like a bigger meta towards the end of like this year where like the, the more neurotic, worrisome parts of the crypto ecosystem start to get more and more worried about quantum. I don't know how you know, at least Bitcoin's going to get through that. They haven't really done like a significant network upgrade in in a while.
00;13;31;29 - 00;13;39;21
Aaron
I kind of feel like it's one of these like old, old quarterbacks that are coming back to play ball, like after being in retirement for a couple of years.
00;13;39;23 - 00;13;44;28
Chris
Yeah. You need the the dadgum era of, quantum resistance and bitcoin.
00;13;45;01 - 00;14;02;00
Aaron
Yeah. Like these devs, which I think they're working just kind of like, you know, nothing super pressing. We have to like, really like lock in to kind of get that get that done. I think it would be interesting to, to see like who's going to be pushing the Bitcoin ecosystem to do it isn't going to be like the core developers or like the early holders.
00;14;02;02 - 00;14;21;03
Aaron
Is it going to be, you know, like the institutions now that have like ETFs and, you know, have like a vested financial interest in it? I don't know. It's going to be interesting to see like maybe like Blackrock pushing for like Bitcoin to become, you know more quantum resistant in some capacity. But I imagine that they they may want to do that or Coinbase or like Robinhood.
00;14;21;05 - 00;14;34;06
Chris
Yeah. Yeah. Anyone who's got an interest in not seeing their their money become, stolen and worthless, I, I suppose, would want to step up and, you know, protect their property.
00;14;34;08 - 00;14;38;01
Aaron
Yeah. Completely. Are you going to go out to node? Did you make a decision yet, by the way?
00;14;38;03 - 00;14;48;05
Chris
Yeah, I got a book. I got a book stuff and and get my ducks in a row. But I'm going to LA for a couple days to do some work, and then I'll, I'll go up to node for a few days.
00;14;48;07 - 00;14;58;24
Aaron
I think it'll be pretty cool. It looks like it's going to be, just looking at some people that are at least on Twitter that are saying that they're going it's going to be like, a nice group.
00;14;58;26 - 00;15;09;21
Chris
It's always it's always good to see people and you show up for these events and you find out that a good third of our listenership are at them, which is always, you know, it's true.
00;15;09;24 - 00;15;14;28
Aaron
Were you, surprised at NFT pairs got canceled? I thought that was kind of a little surprising.
00;15;15;05 - 00;15;29;11
Chris
It's, it's still off my radar, you know, like, I, I barely go to any of these shows. I. You find out about these things after they're done. When was it supposed to be? Was it supposed to be, like, in a couple weeks or was it, you know, months or months from now?
00;15;29;13 - 00;15;32;09
Aaron
I think it was like April, May if my memory served. It was like kind of around there.
00;15;32;13 - 00;15;32;26
Chris
Okay.
00;15;32;28 - 00;15;34;14
Aaron
Not 100% sure.
00;15;34;16 - 00;15;57;26
Chris
Well, I mean, one thing that that showed some slight signs of weakness on the show just to begin with, was what they tried to say. They had to pair it with some other topic, and it seemed like the topic was always changing. And so when I saw it was canceled, the first thing I noticed was it was like NFT Paris and real world assets, you know, and it was like, oh, okay, you're playing that game, you know?
00;15;57;26 - 00;16;25;19
Chris
So if you don't have enough demand for a single topic and you're trying to wedge like, the flavor of the month on top of it to save you that, that's usually a good tip to begin with. You know, it's I don't know, I don't feel like people want to go to, you know, trade shows that it's like, oh, we're the fruit trade show and the, industrial petrochemical solvent show, you know, like these, these aren't things that naturally mix together.
00;16;25;19 - 00;16;50;23
Chris
But anyway, for anyone who was planning on going to it and had tickets, I do feel bad for them. Bernie Terrace is a lovely city. Barry. Paris is amazing. So, you know, it could be worse. You could have like had tickets to go to, I don't know, like EKG in Buffalo and something got canceled or is, you know, the one over in Manhattan, Kansas, where, you know, you don't really have much other reason to go.
00;16;50;26 - 00;16;57;22
Chris
And sorry, it's like, to all the good people up in Buffalo, maybe when you're savers are better.
00;16;57;24 - 00;17;01;11
Aaron
One day Buffalo get a, championship one day.
00;17;01;13 - 00;17;13;25
Chris
I suppose. I mean, I don't know, maybe Buffalo just be happy they still exist. Oh man I love Buffalo. Like I love people from Buffalo. I've only been to Buffalo once. Buffalo was what I thought it was.
00;17;13;27 - 00;17;35;03
Aaron
Yeah, but I think, you know, just with like, the NFT Paris, like, shutting down. I think we also saw, like, Nike sold off artifact. I don't think we know who yet the buyer is. At least I didn't. I didn't see it if it came out. It just feels like this is like the consolidation trough for this disillusionment phase, like kind of, hopefully the last throes of it.
00;17;35;08 - 00;17;45;20
Aaron
And then the green shoots or things like node and kind of like the rallying behind, different assets or some of the stuff that Toledo, but just feels like that's, that's, continuing to, kind of play out.
00;17;45;27 - 00;18;03;14
Chris
I did notice I don't feel like I've seen a particularly busy release calendar, you know, if things are coming out like usually every month to, you know, six weeks, there will be a drop that, you know, catches my eye. And I do feel like as far as I know, it's kind of a very quiet release late at the moment.
00;18;03;16 - 00;18;21;27
Aaron
Yeah. I mean, I think with like the end of some of the art blocks drops and there's still some stuff coming out on, on fellowship and some other platforms. But I do think that's like a bit of a, like a reset. Related to that. I think there was some I didn't I saw like ex-Communist picking up like a new set this week or last week, which I thought was pretty, pretty interesting.
00;18;22;03 - 00;18;31;21
Aaron
Look, look pretty good. I need to dive in a little bit more on that one. I think there's still some stuff there, but yeah, the release calendar and schedule I think needs, needs some reviving.
00;18;31;23 - 00;18;57;15
Chris
Yeah. I mean, I guess one one thing that to watch will be how much new work is coming this year, especially now that there's been a very big shift in to, you know, using using AI and building software and so, you know, is that a set of overlaps and creators, possibly, you know, there's a lot of fang floating around in the NFT boom.
00;18;57;17 - 00;19;12;27
Chris
Last go around. But then, you know, there's certainly a people who are strictly artists. And so, you know, maybe it won't really be, affected by that, but that will be an interesting thing to keep an eye on. You know, it's just, how much work we do get this year.
00;19;13;00 - 00;19;31;15
Aaron
Yeah, I, I saw somebody, I think it was, Red beard. He kind of, noted that with, like, all the people developing you using cloud code or some other AI system that we'll probably just see, like a bunch of different. Like I said, it's like coming out of the woodwork. I think that that's that was, kind of interesting.
00;19;31;15 - 00;19;33;11
Aaron
I definitely could see that happening.
00;19;33;13 - 00;19;38;14
Chris
Do you think someone can invent the first fully autonomous AI artist this year?
00;19;38;17 - 00;19;51;20
Aaron
Yeah. Exactly. Right. Like these, like kind of Bardo Redux is or kind of riffs on them. I wouldn't be surprised if we see like a handful more of those, like, kind of happening. You said you're like always waiting for a deployment. Has that been going this week?
00;19;51;22 - 00;20;23;14
Chris
It's it's going well. I have been working very, very quickly and iterating pretty rapidly. You know, I kind of like I don't necessarily I know where I'm going. I don't have a very fixed roadmap. I'm, I'm really up on the wire. And part of that is I think I just, I need to work from my system. And as I put one thing in place, it reveals to me what's lacking elsewhere what the right direction, to pursue is.
00;20;23;14 - 00;20;43;03
Chris
And so when I wrapped up my last round of stuff, I kind of knew what where I wanted to be going, and I saw a gap in between. And that gap was really around, like my identity system, which, you know, I built one in, like September, which you might as well be 1989.
00;20;43;06 - 00;20;46;10
Aaron
You know, what have you learned? So what do you need to tweak now?
00;20;46;12 - 00;21;13;24
Chris
Yeah. So, I mean, one thing I learned is that, like, you're the agent patterning is a lot better. And the LMS know agent patterning a lot better. And so, yeah, you know, like the fact that it's just like another piece of innate, you know, it's a recipe. It's it's a tool kit that they're comfortable deploying. That certainly helped the increase in, you know, the complexity of their code, the ability to implement it.
00;21;13;24 - 00;21;47;28
Chris
And so that part, actually was fairly easy. And then I wasn't I just wasn't satisfied with the, the execution speed I was getting. I wasn't satisfied with the degree of, micromanagement I had to put in place, to get them to walk, walk all the actions. Right. And what I realized is that my interface was not designed for agents, and that was like the big light bulb moment.
00;21;47;28 - 00;21;49;17
Chris
I realized that. What do you.
00;21;49;17 - 00;22;07;08
Aaron
Think that interface should look like? Because that's like a I think one of these big questions, because I definitely think you're right. You know, like the the tool calling has gotten better, the model seemed better able to handle it. You can start to do like more sophisticated things that would have been really complicated, like delegating tasks out to like sub agents.
00;22;07;10 - 00;22;34;23
Aaron
Right. You can, I think, do quite a bit there. But the interface still I think is a big question and a lot like I just have a hard time believing like and I think devs are like in this weird, you know, like bubble thinking that something like Cloud Code would be like the mainstream version of it. I was just talking to some folks this week and they were saying how, you know, all they need to do is like have folders full of markdown files and then they can build, you know, their own interface to interact with them.
00;22;34;23 - 00;22;47;05
Aaron
And I'm just like, I have a really hard time seeing agents operating at scale with something that looks like cloud code or terminal or cursor or, you know, even like a web interface. How do you how do you even see that play now? Chris.
00;22;47;07 - 00;23;11;00
Chris
Yeah. So I think, the first thing is I work spatially and so I'm in three.js and I think that's a big advantage versus, you know, being in a flat interface, because when you're in a, you know, flat interfaces, you're in the world of forms and buttons and, you know, you just have a finite amount of real estate.
00;23;11;00 - 00;23;41;09
Chris
Whereas if you're in spatial, you kind of like you break having to have nested hierarchical structures and so just from like a, a starting point, I think that there's a difference in kind of the bounding boxes you choose to play around. And then the next thing is, and this is, this is like a little wonkish, but in, in database design or data modeling.
00;23;41;11 - 00;24;19;14
Chris
Right. Like the standard the standard approach is acid, right. Just from like a state management, a consistency. It's atomic transactions, you know, consistent states. It's a very deterministic. This thing has to happen. It has to be recorded. Once it's recorded, then, follow UN actions can happen. And that model is great if you're building like human operated interfaces because, you know, humans are gated by, like, how much they can do at a single point in time to they're really often just going from station to station to station.
00;24;19;17 - 00;24;55;16
Chris
And that was like kind of how my whole interface was designed, because I'm a human being and I designed for humans. And, you know, that's, that's what we've been doing since I got online, you know, back in at the turn of the century to onc myself. And what I realized very, very quickly was the issues I was having around speed was because I was forcing my agents to walk my interface, and even if they weren't walking my interface, they still had to walk the process flow right behind the scenes that, you know, powered my interface.
00;24;55;21 - 00;25;22;11
Chris
And so the, the speed at which I wanted these things working, they couldn't work because I had to, like, force them to run through these crowded, you know, very urban sort of dents. Like, I just didn't give them the freedom to do what they're best at. And so I realized, like, I actually need to get, like a, like I need to stop thinking in terms of, acid like.
00;25;22;11 - 00;25;29;06
Chris
And I had to start thinking in base and so, like, base is stuff.
00;25;29;08 - 00;25;32;24
Aaron
You're getting based. I'm getting base class.
00;25;32;27 - 00;25;59;14
Chris
I am but are like in a way that like, you're not supposed to get based right. Like this is like a really sort of wonkish like, data database like set of definitions. Right. And so you never really think to apply like this mindset of thinking elsewhere. But like what I realized is for my agent stack to actually do its job, it needs to do it like completely outside my interface.
00;25;59;16 - 00;26;27;25
Chris
It has to like, go in and run all the operations it's supposed to. It may be paralyzing them. It may be, you know, I'm passing parameters around and the like, there's no there's no room for atomic, consistency in that mindset. And so like, the what I, what I ultimately realized was like, I, I tell my agents a task.
00;26;27;27 - 00;26;36;29
Chris
They step outside, they do their thing, and then they come back with a new set of facts about the world.
00;26;37;01 - 00;26;39;18
Aaron
And so a payload of some sort.
00;26;39;20 - 00;27;14;07
Chris
Yeah. I mean, it's really like the design model I've ended up in is a lot of, Json manifests that then get like patched through deltas. And so, that's like a very, very different sort of interface design or a different, a different sort of way of thinking of building applications. Because what that means then is you don't know where anything in your user interface may necessarily be at any given point in time.
00;27;14;10 - 00;27;51;10
Chris
You may not need something that was there before, right? Because the world has changed. And so really like it's a pattern of you ask for a thing, you know, there's this whole logic worker tree that exists outside of the domain of, of where you're asking for it, and then it's coming back with, your results. But your results need to, you know, if you had like a fixed user interface and then your results conflicted with that, then you'd have to, like, jam it in or your interface is often wanting to, you know, have artifacts of the work process.
00;27;51;16 - 00;28;11;09
Chris
And if the work all happens outside, then none of that exists. And so at the end of the day, like I, I realized that like, I, I just need to live in a world or I need to work in a world where I'm not making any stateful assumptions about what's on the page at any given point in time.
00;28;11;14 - 00;28;39;02
Chris
And that then means I need to. I had to, like, redesign my application such that I just don't have permanent, you know, I have very few permanent fixtures in the interface anymore. And instead what I have is just a set of like conditions which, you know, are, are evolving. And as those conditions evolve right on load, they know what they need to call and, and to render.
00;28;39;05 - 00;28;55;12
Chris
And so like it seems maybe a little vague or a little esoteric, but it's probably the biggest design pivot I've made in. Yeah, since I've been on the internet, which is basically to assume what's on the page is never reliably on the page.
00;28;55;14 - 00;29;09;28
Aaron
So what do you think the the lessons are then there is it just observability like you, you want to provide observability into whatever parts the agents are going like? Is that what the interface is going to look like? I kind of feel like that has to be a piece of it. Right.
00;29;10;01 - 00;29;12;21
Chris
Unpack that. A little more. What do you mean by observability?
00;29;12;27 - 00;29;36;00
Aaron
So it's kind of like you have these things, these like, little clusters of code like these singletons, right, that are just going out doing something for you. And whatever they do can kind of change the user interface. Right? So, you know, how do you kind of signal that to the, the person in the cockpit. Right. Like who's like managing these agents?
00;29;36;02 - 00;30;00;15
Aaron
I think like one principle is just like a high degree of observability. So you know, like the, the pathways that it's going. Right. Because I think that there's a, a challenge. I was actually reviewing like a bunch of the academic papers on agents and not surprisingly, there's a couple papers that talked about how really one of the challenges is in agencies at some point in that flow, when they're doing something, they may hallucinate, right?
00;30;00;15 - 00;30;17;21
Aaron
They may get off track, they may make a mistake. And that's where I think, like the human will be, at least in the near term, right, to kind of keep the agent on task or like stop it from going down a rabbit hole that is not productive. So it feels like that's why you need like a high degree of observability.
00;30;17;24 - 00;30;22;16
Aaron
If you're building one of these, like more agent systems, at least at a minimum.
00;30;22;18 - 00;30;53;22
Chris
Right? Okay. I get what you're saying. And so I don't think there is a lot of observed observability in terms of what's going on behind the scenes. I have it as I'm building, but like, you know, ultimately what I want is enough deterministic guardrails and select key places to reject hallucinations and rerun them under the hood. And then what you're getting back on the client side, right, is you're going to get an event notification.
00;30;53;24 - 00;31;16;20
Chris
You're also going to have like the ability to to patch and edit the UI yourself. And so like you can't entirely like cede control, you know, you need an user to step in if you, if you said, hey, take this piece of media and double its size and then, you know, your agent goes, okay, sure, I'll make it bigger for you and then comes back like five times the size.
00;31;16;20 - 00;31;39;27
Chris
You can just click on it and resize it yourself. And the same way, like, you know, the, the UI change the first time via like a delta, a delta stream in a patch instead of like that originating from the agent ecosystem. That's just originating from. And so like there's corrective eye and look, I'm like very this is very new to me.
00;31;39;27 - 00;32;02;01
Chris
And I haven't gotten like super down, you know, into like handling all these sort of edge cases. But big picture. That's that's how I'm treating it is you're not stopping the work. You're just allowing the ability to. Oh my, my bot didn't do this the way I wanted it to do. I guess I'm going to have to like point and click and, you know, do a little work myself.
00;32;02;04 - 00;32;20;28
Aaron
Yeah. It's kind of this like back and forth. Right? I think it's I think it's challenging because we're not used to these like volley type user interactions. Right. Like, you get that a little bit with like chat. And I think that's why some of the early interfaces are chat based. But it feels like an incomplete like UI solution for it.
00;32;21;03 - 00;32;49;08
Aaron
And probably the same thing on the back end to Chris. Maybe that's why you're running into these challenges or seeing like limitations there, right? Like this, this like the, the volleying that you need to do with the agents and then this like higher degree of observability that you need on that like a, you know, the user interface, like whoever's kind of commanding calling, creating these agents, I think it's like a bit different design pattern than what we're used to.
00;32;49;11 - 00;33;03;16
Aaron
I wonder if that's why like kind of this there's like a lot of the devs are excited about like markdown files for that. Maybe that's instead of a Json object. It's kind of like more of this, like flat, flat text file that people are pointing towards. A little bit more markdown.
00;33;03;16 - 00;33;04;20
Chris
I don't I.
00;33;04;22 - 00;33;27;06
Aaron
I, I struggle with that, I don't buy it. But I hear people like, I just saw somebody just sent this to me like a Neil Dash talking about who's, for folks that don't know or may not remember them quite a bit in the tech space, these near to tech scene for, for decades now. So he was extolling the benefits of, the markdown file and, a recent blog post.
00;33;27;06 - 00;33;27;24
Aaron
Yeah.
00;33;27;26 - 00;33;47;03
Chris
Here's, here's what got me to this, because I think it points to like what you were talking about in that, like it took, I think, three different changes to actually for me to like, coalesce around this design. And I don't know even how long I'm holding on to this design, but to me it is a turning point in my thinking.
00;33;47;05 - 00;34;19;08
Chris
And the changes really were one from a very boring perspective. Understanding like the, the proffer or like recipe for an agent stack. Like what is the right hierarchy? You know, you have a router, you have an intent classifier, you know, you've got, a parameterized worker that, you know, can reach into these trees of skills, knows how to write, like just what that org chart looks like is, I think now cemented and established.
00;34;19;14 - 00;34;43;13
Chris
And the funny thing is, as I was, you know, putting this in place, I looked at and I was like, God, this is just like when I built an IVR phone tree, you know, back in 2010, you know, like once upon a time I had to like, you know, build one of those automated customer support systems where, you know, hit this, go there and do that, you know, and you're just using ten keys to, you know, manage your account.
00;34;43;15 - 00;35;05;26
Chris
And so that's very, very dated thinking. But you know, people just needed to like work out what the right pieces are in. And in the genetic machine, the next thing was having a better class models. I'm using Gemini Flash right now to do a lot of my, like, intent classification tool. Planning execution.
00;35;05;29 - 00;35;06;12
Aaron
Is a good.
00;35;06;19 - 00;35;47;04
Chris
I haven't really gotten into like super. It's good in that it can handle fuzzy cases. Like if I say give me more, right. Like it knows it just needs to figure out what the last thing in the conversation was, what the last action it took was, and to expand. And so like that, that simple level of like three step fuzzy reasoning we didn't have in like, let's say the, the GPT four series, like that was still brittle and it wasn't reliable enough to just say a reason your way through, like, oh, a continuation that is vague.
00;35;47;07 - 00;36;09;28
Chris
And so like the model classes I think are now there is that like the other problem, I was having a shit ton shit hard time with back in September and why I never fully developed it out. Was actually just manage intent classification and routing like so like if you can't even start your tree and effectively play traffic cop, that's like a big blocker right there.
00;36;10;00 - 00;36;39;07
Chris
You know, like I probably have like eight different intents in my system. And it has no problem understanding, managing and routing into like each one of those intents. And so that is definitely the model class being good enough to now, you know, basically sit at the switchboard. And so that to me I think was a big necessary thing because I think, you know, the proper like blueprinting of an agent stack.
00;36;39;07 - 00;36;42;22
Chris
I think we kind of knew that six months ago. You know, I.
00;36;42;25 - 00;37;00;12
Aaron
Do you think so? I don't know if that's true. Like, I feel like it's kind of like what what are the steps that an agent needs? It's it's funny that you're poking around here because this is what I've been spending my time thinking through this week. We're trying to get Aiden to be more attentive, and I feel like there's some bits to it, like there's like a planning mode.
00;37;00;14 - 00;37;20;26
Aaron
And I think this is what you're kind of getting at. You know, there's like, you know, reasoning steps. There's kind of like task management. There's like the tool calling and like delegation. But I don't do do you feel like we know all the beats of, like how an agent should work? Like I've been actually adding in a couple of like more like post-processing steps, like a confidence level.
00;37;20;29 - 00;37;47;16
Aaron
I added it actually in, like kind of, a tool where an agent can actually confess when it makes a mistake and record learnings after it does make a mistake, which I thought was kind of interesting. And then you can like feed that back in. So hopefully it doesn't make the same mistake again. Do you feel like we actually know like, oh, I feel like we've got the rough outline of it, but there's a lot of optimization that we could still do to kind of wrangle how these agents work, especially for like these longer running tasks.
00;37;47;19 - 00;37;48;27
Aaron
But what do you think?
00;37;48;29 - 00;38;12;16
Chris
Sort of basically intermediate use cases? I think we've got, the core architecture, the core patterning. Correct. Where, you know, maybe you can ask me this next week or the following week, depending how far I get parallelization and convergence of multiple agenda actions at once. So, like, you know, how do we.
00;38;12;17 - 00;38;14;26
Aaron
How do you reconcile them. Yeah.
00;38;14;28 - 00;38;32;08
Chris
Yeah. Like if you're firing four things at once and then the you don't need like all four of those displayed separately, but you need some form of synthesis of those four things. Right. Like how far can we extend this it like I just haven't gotten there yet. And so I don't really know.
00;38;32;10 - 00;38;47;24
Aaron
I see why and then I see why the database stuff is coming. Right. Because if you're firing off four agents, they need to get reconciled at some point and then they're going to conflict and they don't have like a shared state while they're running in parallel. Yeah. That's a that's a super tricky problem.
00;38;47;26 - 00;38;49;24
Chris
The answer is to get rid of the interface.
00;38;49;24 - 00;39;25;09
Aaron
I think what you're raising, Chris, is an important point. It's kind of like as we increasingly mediate different actions on a computing system, putting even aside the internet piece via agents, it's just going to change all the stack. And so like I felt this where we spent more time using a vector database like a rag to store certain information, which makes me not want to use things like a mongo database or a SQL database to the same degree, which is a pretty big, you know, tectonic shift and just how you're going to architect and build applications.
00;39;25;17 - 00;39;49;19
Aaron
And I think whether it's the final solution or not, like the fact that, you know, cloud code is pretty much choosing a markdown file as it's like base acid is kind of notable and that it isn't like, like a structured object, like a Json, payload of some sort or package. I just think that these, like, subtle changes matter and they, like, bleed all the way back up into the interface.
00;39;49;19 - 00;40;09;26
Aaron
So, yes, I think it's because you're living on, like, the edge of Agent Land, at least as compared to many people, including like, you know, people that are super loud on Twitter. I think you're just seeing the problems before it, before they do. But what you're I think what the conclusion is, is like, it's all going to change, right?
00;40;09;26 - 00;40;27;14
Aaron
We're probably gonna get some new data structure, organization system, right. And new interfaces, just to kind of manage this new world that's beginning to develop. And there's gonna be a lot of these agents, right? Like probably trillions and over the next decade that are running, like in the background and stuff, which is wild.
00;40;27;17 - 00;40;57;08
Chris
We'll see. I mean, you know, we also thought everything in the world was being made out of like, nano carbon fiber at this point. And it's not. But yes, we're gonna have some some huge proliferation of agents, you know. So like base, I did finally look it up, basically available soft state, eventually consistent. And so this is probably why the markdown thing is the simplest example of that is what you're having is a conversation.
00;40;57;08 - 00;41;09;07
Chris
You're having that volleyed back and forth. And only once you come to an agreement do you actually memorialize that in markdown. And so that's your eventually consistent part.
00;41;09;09 - 00;41;10;26
Aaron
Yeah. The artifact.
00;41;10;28 - 00;41;39;17
Chris
Yeah. And I would actually say like the soft state in this is art. You know, artifacts are open to open to interpretation. Documents are open to interpretation, even if you're, you know, tightly bolding and sequencing, etc., etc. there's a lot of room and fuzziness in there, whereas that's not what Json is, you know, I mean, Json like is like structured arrays of data and so like that.
00;41;39;19 - 00;42;08;01
Chris
That's something I actually do need, like I need values, I need positioning, I need to know what is appearing, you know, in my interface. And then I need to know like rel and relations and capabilities and what they reference and all of that. Right. So I need Json and so I'm working in manifests and it's really like you generate a manifest, the manifest outlines all the things that are expected.
00;42;08;04 - 00;42;32;19
Chris
It's up to the environment to interpret and render that. And so like you now need to build an environment that is, you know, comfortable enough saying, okay, I have a coordinate system. It's telling me that, you know, over in these coordinates, you know, this, file name is expected. What is this file name? Oh, this file name is a piece of media that's real simple to put a piece of media there.
00;42;32;19 - 00;42;58;02
Chris
Or this file, name it like the type is referring to a portal. Therefore, I need to create a gateway. You know, but like you're I've, you know, people have been making takes like one day software is going to write itself. And that's probably true. I mean, I'm sorry. I mean, like interfaces are going to write themselves on demand and materialize and then disappear.
00;42;58;04 - 00;43;10;21
Chris
Sure, that's eventually true, but it's going to be done, I think, in like this include style, right. Like, in which you need to have some form of a pointer which is calling for a thing.
00;43;10;23 - 00;43;42;07
Aaron
Can we pause there because I, I feel like I've been playing around with these concepts. So like, and I've been calling it in my mind, like self-healing software, you know, like how in many ways, like, I think tech kind of mirrors what sits underneath it. And so we're seeing, you know, like and AI people are postulating that the next versions of, you know, GPT six or next version of, Google or Anthropic or, you know, pick your large model provider that they're going to increasingly be self-improving.
00;43;42;10 - 00;44;05;22
Aaron
And I definitely feel like you could build like an, a genetic app that does the same thing, right? Like the ideal should be. And this is why at the end of the flow for an agent, I've been kind of really focused on it, like it should know if it did a bad job or if the user told, the agent that it didn't do the job that you were hoping it it would do, that it should be self aware and like.
00;44;05;23 - 00;44;33;28
Aaron
No. When it did a bad job and confess that I think that's like healthy and like a good it builds trust. And then I think on top of that it should be asking itself, you know, what technical improvements do I need to not make that mistake again? And if there is one, then that should get recorded. And I do think then you could just like push it into like, like a subroutine where it would generate like, a product scoping document and then could just get picked up by a model to implement that in a code base.
00;44;33;28 - 00;45;06;20
Aaron
Right. I do think we're going to get these, like, self-improving systems pretty quickly. And I think this concept of self-improving software is like, super fascinating to me. It's almost like by eliminating the need for a software developer, you get much more organic software, like software that can grow and improve, like at a rapid pace without, you know, the intermediation of the human, you know, product managers and software developers that are just, you know, not actually adding, as much value as just letting these systems kind of run on their own, based on all these different user interactions.
00;45;06;22 - 00;45;10;15
Aaron
So that's what I've been thinking about here. Chris. I don't know if that resonates with you.
00;45;10;17 - 00;45;16;09
Chris
Yeah. So I do you have a learning system in your genetic stack, like is that an intent?
00;45;16;10 - 00;45;39;29
Aaron
Yeah. Yeah. So I, you know. Yeah, I have two things. One is like memory. So the user can express preferences. The agent can kind of keep like a scratchpad of things that it needs as it's going about its task. And then at the end it has like more like self-reflection. And as part of that, it it can be like a confession, like if it reflects and then feels like it, it didn't do a good job.
00;45;40;01 - 00;45;59;00
Aaron
It confesses. And then, you know, can be a learning like, what did it learn when it did this? And then it saves that to its memory. And then in, in some of those instances, it can also then, you know, create like a technical improvement that may be needed to solve the issue after it's like made a mistake. It's like got its hand caught in the cookie jar.
00;45;59;00 - 00;46;12;26
Aaron
Right. It should confess. It should tell you that. And then if, if it's like, well, the cookie jar needs a lock, so I don't get caught here. Well then that should get built right. So that that's kind of the logic I was thinking about in my head.
00;46;12;28 - 00;46;42;06
Chris
Yeah, yeah. No, that makes sense. And this is like a horses the horses thing, right? Like you can't have one design to rule them all for me it's about, surfing, surfacing user preferences and having the model keep that in mind. And so the way I'm handling this right is my intent. Classifier can recognize feedback, right? You know, if I say, hey, if I give you a vague prompt and you return nothing, I want you to get creative and like, redo the prompt in a way that satisfies me.
00;46;42;09 - 00;47;08;11
Chris
And so it knows the intent classifier sees that and goes, that's feedback. Feedback gets routed to the learning system. The learning system then records records. That is a line item. It creates a digest which you know is just, digest. You know, it's just a small little, you know, paragraph of text saying this user wants this, that and the other and, you know, put the source on the side.
00;47;08;14 - 00;47;36;15
Chris
And then any time that, you know, the user says something, you know, show me some cats, you know, the intent class probably goes up. You want, this guy wants to perform media search, media search goes over to this planner Executer the planner Executer will see that, but it will also get, you know, the last 4 or 5, turns of the conversation, plus the learning digest in its payload.
00;47;36;17 - 00;47;56;10
Chris
And then you can see, you know, that this is a continuation and it's a continuation that has a preference and that preference is sitting in the digest. And so now I know how to do my my planning form, the parameters to call the tools that that's how I've got it going in my system.
00;47;56;12 - 00;48;18;20
Aaron
Yeah I think it's like similar beats. Right. It's kind of like you need like like a running artifact to just it's. Yeah, I my mind, I call them like, artifacts, just kind of keeping track of just different things that are, that are happening in the system. It's honestly it's fascinating. I don't know what the core architecture is, but it's just amazing that you can do it.
00;48;18;20 - 00;48;40;10
Aaron
And I think when you can build these like recursive self evaluating systems, that's really where some of the magic is going to come from. I think even if we didn't get like model advancements to some of the things that you're poking around and, poking around in part and obviously a bunch of other teams are, you're just going to get like, you know, a ton of efficiency, an improvement just from that alone.
00;48;40;15 - 00;49;00;03
Aaron
So it just it's pretty wild. And I just think the quality of software that you can build, it's just going to make all the like, web two stuff just seem horrible. Like, I kind of already feel it, you know, like, one thing I built in our system, just like basic ability just to do, web searches. I connect your email and, like, calendar and just do certain actions related to that.
00;49;00;03 - 00;49;27;04
Aaron
And I just increasingly don't want to check, you know, my, you know, my email interface. Right? I kind of noticed this last year with ECS, like I found myself going to like the Super Grok kind of tab and not reading my timeline and instead just prompting with, you know, grok to get news that I was interested in and I found it, like, so much more pleasurable and it kind of like, cut the noise.
00;49;27;04 - 00;49;43;03
Aaron
I don't need to see, like, you know, Ellen's like latest white supremacy post or something like that, you know, and I could just kind of move on with my day. I just feel like that that's something that I think is going to become a little bit more widespread. Like you're not going to be like diving in into the trenches as much.
00;49;43;03 - 00;49;48;21
Aaron
You're going to get that abstracted a bit away from you, which I think is going to be super healthy for everybody.
00;49;48;23 - 00;50;23;09
Chris
Definitely. And I mean, I guess that's the last point. You know, I kind of got derailed out. And then somewhere ten minutes ago, I said three things where I think where what is the templating stack? Two is the model advancements. But then the third part is understanding what patterns are holding and what patterns need to be broken. And like you know, that that's been something like I've just felt very, very deeply this week is I've been working my way through this in that like I'm dealing with both a, a very familiar pattern that I'm surprised is so familiar.
00;50;23;09 - 00;50;49;09
Chris
You know, when I talk about, like, how how you at least how I'm treating like, an agent hierarchy and, you know, workflow, right? Like, literally it's the same principles as building an IVR phone tree. And so you're like, oh, okay, look at that. You know, things don't change all that much. And then at the same time, am I'm embracing this idea that there should be no fixed fixed user interface.
00;50;49;09 - 00;51;21;17
Chris
Other than, you know, a chat, a chat box and a few buttons that, you know, you might need on demand, which is a very radical notion because, you know, to even do that, I need to be a in a station environment in which, you know, positioning and placement of things is a lot more fluid. You know, I couldn't be doing that on like, a flat web page, you know, like, how the hell do you rebuild a web page over and over and you can't, like, you literally would have to, you know, I suppose you could, like, make all sorts of crazy Ajax calls and shit, you know?
00;51;21;24 - 00;51;31;25
Chris
Oh, but, yeah, that's still a call, you know, like in my world now I'm just pushing and updating things and so, like.
00;51;31;27 - 00;51;48;24
Aaron
Yeah, I was going to say, I think the challenge I have with like the and I know Google research kind of put some stuff out that shows that they can like build a website on the fly. To me, I think it's like runs into this practical limitation of just the end user. That's not going to want to sit there and wait for that to happen, right?
00;51;48;26 - 00;52;10;01
Aaron
So I don't know what that means. I don't know if people can like, make an add a lot more interface customization, maybe pulling from like, like Lego blocks of building blocks that I could quickly assemble. I just have a hard time imagining people wanting to wait 5 minutes or 10 minutes, you know, to just get like an interface that's perfectly customized for them.
00;52;10;07 - 00;52;29;20
Aaron
Maybe I'm wrong, you know, maybe that happens in the background. You just get, like, a penguin. Your interface is ready, but it just feels like, a lot of, I don't know, there's a lot of practical challenges to kind of that, that approach. I mean, it's super cool, like, conceptually, like I'm super geeked on that. But it something about it just feels like a little bit off.
00;52;29;23 - 00;52;52;16
Chris
Yeah. No, I hear you. I mean, I'm calling templates left and right, you know, I mean like I'm not like building anything. I am just a lot more fluid in what I can call and where I can place it. I still need containers, you know, I still need two wrappers. I still need, code to take these parameters and and make them, usable.
00;52;52;19 - 00;53;18;19
Chris
No, let me give it here for a second. Cause the other thing I kind of realized as I was, working my way towards this is I've been in this, like, very sort of subtle positive reinforcement loop with the new models, like five, two, more so than 4 or 5, opus. Like the GPT just started talking a lot about contracts.
00;53;18;21 - 00;53;44;08
Chris
Right? And yeah, not in a way like you should be using a contract. More like, oh, well, we ran this error because, there's no contract enforcement here or because the user, you have multiple inputs into this thing, and one of your legacy ones allows for blah, blah, blah. Whereas, you know, in the new the new approach, the contract doesn't allow for this.
00;53;44;08 - 00;54;13;29
Chris
Right? And so, like just starting to steer more and more into how it prefers to do things and potentially realizing I'm getting fewer errors in my development, the more I want to do, the more it I do things the way it's subtly suggesting to do it. You know, it's never outright coming out and saying like, you know, you should you should do this because it doesn't understand, like the, the full scope enough to, you know, to act as a holistic architect.
00;54;13;29 - 00;54;37;05
Chris
But in diagnosing individual cases, it frequently comes up enough that the preferred way or the correct way. Or we can solve this problem by doing X. And then you see the benefit of x, which is that it can ship it faster, it creates, you know, less variability. And it also allows you to build better on top of it.
00;54;37;07 - 00;55;22;28
Chris
And so like I've been on a huge standardization kick and really like trying to get as much of my entire system singing the same song the same way. And that when it calls a thing, it everything calls it, you know, uses this, media resolver system, you know, in which I have a file name and, like, not, you know, like a URL, but, you know, like ref dot dot, you know, image dot dot cats in space, you know, or that I've got like ref dot dot media dot dot video dot dot, you know, hot air balloon over the desert.
00;55;23;03 - 00;55;46;23
Chris
Right. And then a layer below that, I have a whole resolver system that then knows, okay, when it comes in, in this, you know, this format a I need to I know images live here. I know images get rendered this way. I know video lives here, I know video get rendered, you know, it. It just makes things a lot like, easier.
00;55;46;26 - 00;56;13;07
Chris
And so it's this weird mix of the more I standardize, the back end, the looser and more based the front end gets, which is a little contradictory, but it does make sense, right? Like it's patterning all the way down. And by having like strict, consistent patterning, it then gives you far more freedom on the output side to mix and match and to ship clean stuff.
00;56;13;12 - 00;56;40;24
Aaron
And but maybe you're making the case wide interfaces will be increasingly customized, right? Because it will be dependent on the agent activity. Chris. That's what I'm kind of hearing from that. Like the back end needs to get I just think most interfaces really do. They kind of reflect the underlying tech, right? So, you know, like, I don't know exactly what the back end of something like an Instagram looks like, but it's probably a lot of streaming stuff, right?
00;56;40;24 - 00;57;12;09
Aaron
Like streaming and events and updates and data storage and all this stuff. But it's very, you know, strict, monochromatic, right? Like in many ways, like, one dimensional, like it's all about just streaming in and streaming out. So at the core of it. So I wonder because these agentless systems are like a little bit more wildcat if the front ends get a little bit more wild cat to and maybe to kind of manage that for each end user, you do need something a little bit more customized to kind of fit what they want or how they're going to use it.
00;57;12;11 - 00;57;40;28
Chris
It's a good point. It's also I think, the models still have a tendency to really want to, you know, it's like all the models live in memento, where every request is a brand new day. But the the tighter you pattern your system and the more you standardize it, the less, inclined it is to go freewheeling, inventing things out a whole class, and the easier it is like I had.
00;57;41;00 - 00;57;57;10
Chris
I'm not complaining about adherence anymore. Right? Like, I I'm sure if you pull the episodes up for me in November, I was bitching and moaning about adherence and like saying I need a pro version of these models that like strictly do what I say because it's introducing, you know, all this. I think.
00;57;57;10 - 00;57;59;13
Aaron
That was last week, Chris, that was like two.
00;57;59;13 - 00;58;23;21
Chris
Weeks ago. God, I'm not having that problem right now because there was no room for it to introduce this additional stuff. It thinks it's being helpful and proactive. And so any time I like I and I've also just gotten to the point where I follow it's thinking. And the second I see it diverging from pattern, I just stopped dead in its tracks.
00;58;23;21 - 00;58;36;05
Chris
And I say, you look like someone who thinks that they need to build a new, a new method of blah blah, blah. The guy I'm looking for follows our standard contract and it immediately snaps right back.
00;58;36;07 - 00;58;43;08
Aaron
So it's kind of you're using contract is more like rules like like manifesto or like Constitution, like guiding principles.
00;58;43;10 - 00;59;02;15
Chris
Yeah. Like if there is a template that expects a data structure, right? If there is an output format, if there is a manifest, if there's something like that, then the contract means like all the way up and down the stack. You got to get your shit into conformity with it.
00;59;02;18 - 00;59;18;07
Aaron
Yeah, yeah, yeah. It's like like like rules are like a, like meta guidelines of some sort. I don't know the right way to phrase it. You know what I've been doing this week, Chris? Which I, I it just occurred to me, we both use this system so much. Yeah, I'm sure we're telling it what to do.
00;59;18;10 - 00;59;38;10
Aaron
And I like late one night I was like, you know what? Like, these models are great. Like, they're probably much, much more intelligent than me or really anybody else. Even like the the glasses version of myself, which I, I don't want to give myself credit for. Like, they're, they're definitely, like, smarter. So I started asking it more like, what would you do here?
00;59;38;12 - 01;00;00;14
Aaron
Like, you have agency, you have control. And I found, you know, if you really kind of like, push back and ask it to really, like, give its own opinion, like and and try to push it to have more agency. I was getting better, cleaner, you know, solutions. I tried to like just like, humble myself, like kind of at the altar of these systems.
01;00;00;14 - 01;00;20;28
Aaron
And I found it, like, pretty effective. I don't know if you've played around with that, but that was kind of interesting. It's kind of inverting it because I feel like you know, just reading Twitter, just like all these devs, like, you know, with their neck beards, thinking that they're the best developers and none of us are the best developers, really, except maybe like a handful of people and, and Max.
01;00;21;01 - 01;00;33;00
Aaron
And so I really wonder if we should approach this much more from the perspective of these systems should kind of really be guiding the ship more then I think we're letting them guide that ship, if that makes sense.
01;00;33;02 - 01;00;58;08
Chris
Yeah, I haven't I haven't gotten there yet. And so I can't say like I've given it a fair shake. You know, I will every once in a while. Like if I'm just like a little exhausted or I want to play a game and, you know, test their, creative thinking or this and that, the other, or I'm very agnostic in terms of what's actually under the hood.
01;00;58;16 - 01;01;25;16
Chris
Right. Like, I don't walk in, you know, with preconceived notions of, like, this thing needs to, go through SQS, this has to run through. See, this thing has to be structured this way, you know, like, I don't walk in with those preferences, and I let the models, you know, kind of define those. And then I think I only form preferences after working enough with them that I can get opinionated and start winnowing things out.
01;01;25;19 - 01;01;49;07
Chris
And so I guess from like a architecture perspective, I always want to have my, my product vision, my edge cases, my logic. I always want to have that defined and bounded because I don't I don't feel like it. It's particularly strong there. But in terms it is, yeah. This is I'm doing such weird no to that's the other problem.
01;01;49;07 - 01;01;51;15
Chris
Right. Like I just do weird things.
01;01;51;17 - 01;02;13;19
Aaron
True. But I wonder you should give it a whirl. Try it on a because I, I was surprised, like, I, because I just was starting like a new thread of work and it was more exploratory and I was like, I basically defined the task but was pretty mushy on the implementation. There was a lot of things that kind of came out from the system that I wouldn't have thought about.
01;02;13;22 - 01;02;33;12
Aaron
And so I wonder if, like, actually the optimal way to prompt these things is like some sort of balance between being like direct, like proscriptive, but also like giving a little bit more agency to these systems to kind of run. And I feel like a lot of people approach it like the slap shot, right? Like that park attack the texture where you're just telling it what to do.
01;02;33;19 - 01;02;37;09
Aaron
And I think you're just leaving like a lot of intelligence, like on the table that way.
01;02;37;11 - 01;02;38;21
Chris
Well, that you.
01;02;38;26 - 01;02;40;11
Aaron
Could make things better.
01;02;40;13 - 01;03;00;02
Chris
Yeah. I mean, now you say it that way. Like, I think my, like, acid first days come to Jesus. I directionally knew what I wanted to be doing. I'd done it. And another part of my system. But that other part of the system was all about, client side web GPU control. Like, I wanted to paint pixels on the client right.
01;03;00;05 - 01;03;22;26
Chris
And so, like, I knew there was a lot of strength and streaming and letting the client side do the work. I just couldn't fathom how to do that. For more of, like, a user interface point of view. And so I did say to it like, look, this is a little weird, right? I'm thinking in terms of like acid and base, but I'm thinking about it in terms of like, how can we get agents to effectively achieve the tasks.
01;03;22;26 - 01;03;47;18
Chris
And like, you know, the bottleneck is it can't run this, this like point and click maze and like so how do we get there. Right. And and that opus was incredibly disappointing. Right. Like pirate with opus I feel like it's either a fifth grader or it's a PhD student and like it doesn't quite have a middle gear and it's like it's a clawed neurosis.
01;03;47;24 - 01;04;15;12
Chris
You know, where like it, it will fall really, really fast and like turn into an 11 year old, you know, way too often. But like five two. Got it. And it was very steady about it. And it was like it was able to understand what the metaphor I was talking about. It was able to synthesize and need, and then it was able to actually convert that into a software design.
01;04;15;12 - 01;04;19;22
Chris
I would never would. Yeah. I never would have got that design myself. Like, I, you know.
01;04;19;22 - 01;04;38;27
Aaron
That's because I mean, I think this plays out in the evals, right. Like where five two excels, where the other models don't is around that abstract reasoning. And I think that that's kind of what is coming into focus. I agree with that. I think 5.2 is better at some of these more abstract tasks, because that is an abstract task, right?
01;04;38;28 - 01;04;56;06
Aaron
Like however you classify it, you're like, I don't know what to do here. I have like some vague notions and directions in which it's going to go. It's abstract, right? But we don't know the path forward. But I think that these systems I think in the next version, all these systems will be as good as 5.2 and then Prime.
01;04;56;08 - 01;05;18;09
Aaron
OpenAI probably still will have the lead on on that front, but it's great. I mean, I think all this stuff, man, the fact that we can even, you know, push things to this degree is amazing. And I really do feel like this groundswell that's been happening, it's starting to like, hit like a logical inflection point where just the, the type of software that we can build is just going to be super magical.
01;05;18;12 - 01;05;37;29
Aaron
Like, I just feel like I'm getting these like small, magical moments when I'm building stuff. And it just that's super fun. Like, there's no better dopamine hitting that, number one and number two, like, I don't know, I just feel like it's going to, you know, reverse this trend of like, the plateauing internet or the, you know, a certification of the internet.
01;05;38;01 - 01;05;52;15
Aaron
And, like, all the stuff that's come over the past couple years that we, we use, we love or we lament is just going to be it's going to look like, like a super dated, like looking at like a late 90s website or maybe even a website from 1989.
01;05;52;18 - 01;05;59;03
Chris
I'm all for it because you know, what we got today is, yeah, it's like, but that's good, you know.
01;05;59;09 - 01;06;17;12
Aaron
You know what I and I love about the crypto industry is like, you were seeing all these folks, and people are like, you know, like softly mocking it or criticizing it. But, you know, so many folks in crypto are, like, live on the edge of tech. They're like such an early adopter class. Like a lot of them are leaning in and in either, like private DMs or like and or even publicly.
01;06;17;12 - 01;06;33;26
Aaron
You're just seeing them like, like have fun again. And I just think that that that's great. A friend of the pod like g money was the saying like he you know, he want he wanted to create and build stuff. And he finally feels like he's empowered to do that. Like that's amazing. You know, the fact that everybody can feel that is just super cool to me.
01;06;33;29 - 01;06;57;28
Chris
Yeah. Know that that is, and it's probably like a much needed jolt of agency as well, because certainly sticking your neck out in the world right now, you know, is, is one of these things where you're constantly reminded how little of human life is valued in the year of our Lord, 2026 and that that can be like, a very deflating experience.
01;06;57;28 - 01;07;23;09
Chris
And so, you know, if you just want to stay like, you know, harness to your, to your laptop and find meaning somewhere else, you know, this is like, you know, you've just been given, the keys to the kingdom. And so I do think that's going to be one of the, you know, maybe things that isn't, isn't necessarily considered.
01;07;23;09 - 01;07;35;05
Chris
You know, it's a softer, you know, it's more of a Bios based thing, but I do think, like, this might be something that really helps a lot of people out through, you know, a period in which their place in the world is very uncertain.
01;07;35;07 - 01;07;54;28
Aaron
Yeah. I think it, you know, it's like I think Jared Linear was right that like, we were kind of in this like feudal serfdom era of the internet. And I do think at the end of this, you know, we are kind of like free folks. Again. Like we don't if we are unhappy with like, like our email service, like maybe we do build like a new interface client related to that.
01;07;55;02 - 01;08;14;10
Aaron
And yes, like Google may have some of your data, and you may be accessing that data, but maybe you're not using their service to the same degree which prevents them from monetizing. Right? Like you, like the, the balance of power may not be perfectly tilted in your favor, but it feels like it. It's a little bit more tilted in your favor with these tools.
01;08;14;10 - 01;08;14;29
Aaron
Like, available.
01;08;15;03 - 01;08;19;17
Chris
Yeah, as long as there's an outlet, you know, like, as long as there's a place beyond the reach.
01;08;19;19 - 01;08;22;03
Aaron
Yeah, beyond the pain. So who can cares if you.
01;08;22;06 - 01;08;45;15
Chris
Yeah. Like, who cares if, like, you go to, like, once a day checking the Gmail, you know. Oh my God, I gotta go back to Gmail. Like, it's not like in the old days, like, dude, when I now is an operator. And I had, like, you know, I was a manager, like, I, I probably spent half my day in an email client and now email doesn't matter, you know, like, you know, we're slack and radio.
01;08;45;18 - 01;08;46;05
Chris
Yeah.
01;08;46;08 - 01;09;06;29
Aaron
Yeah. But how many people spend their days in slack? Right. When maybe in a couple years there's like a slightly customized interface that they built for their, you know, preferences, maybe that comes from like a bundle of like components, like, I don't know exactly what that looks like, but they have like some agency and deciding how they're going to interact with all these services.
01;09;06;29 - 01;09;28;19
Aaron
And maybe just because you want to stay up on the news, you're not like glued to, you know, the, the mono, you know, monolithic mono class, you know, X interface. You're getting it in different ways. Right. Like I just think that that's going to give people a lot more agency. And I do think like these we're getting like a level of abstraction from the internet, which is just going to be super healthy.
01;09;28;21 - 01;09;48;06
Aaron
Like I think that's the most important thing because we all kind of know it's like we're just eating like, you know, mental junk, junk food. Right? Like every day, whether whether it's like at work and all the chatter and discourse and slacks or, you know, you know, and whatever your favorite, media platform is, it's a brave new world, Chris.
01;09;48;08 - 01;09;53;03
Aaron
And it all comes back to our fun, fundamental math. Innovation.
01;09;53;05 - 01;10;03;12
Chris
So, yeah, just speaking of it, where do vectorize everything I read, everything is, is closely related to each other. Interaction.
01;10;03;14 - 01;10;40;04
Chris
Let's leave this out there for, like, a PhD candidate in, the year 2032 to do a big ass report on mental health and controlling your field of view. What being two online looked like in, you know, 2020 versus today, right? Like how much of this is literally just your eyes seeing things that you don't control and how different will the world be if you actually control your entire field of view and you have a customized you might be getting bad news, but you're not getting bad news next to like, a viral clip of someone being shot.
01;10;40;11 - 01;10;49;25
Aaron
Yeah, I mean, that's the stuff like. I mean, I remember that, that sad day when that Charlie Kirk killing, right? And I, I couldn't avoid it if I wanted to, you know.
01;10;50;00 - 01;10;51;20
Chris
Well, that was yesterday as well.
01;10;51;22 - 01;11;09;04
Aaron
Yeah. I mean, I think it's that meme with, like, the scroll that eyes, which is such a great meme, but I feel like that is how we all feel. Hopefully we, we move back to, what we are going to be moving back to Chat Land. That was that was your prediction. We're all going to be, what would you call it?
01;11;09;04 - 01;11;09;21
Aaron
I forgot.
01;11;09;26 - 01;11;14;16
Chris
Chad, sir. It is, we're going to be enjoying a mediterranean life.
01;11;14;19 - 01;11;17;20
Aaron
As as Chad's. Yeah, that's the French. Go!
01;11;17;22 - 01;11;18;06
Chris
Let's go.
01;11;18;06 - 01;11;19;00
Aaron
Baby. All right.
01;11;19;07 - 01;11;23;20
Chris
And so, we get Derek and Tres paycheck this week, right? That's how this works.
01;11;23;20 - 01;11;45;10
Aaron
I think so, yeah. So some big news from Derek. Will let him, announce that and then, pray, unfortunately had to travel. So they'll be they should be back on the pod next week. So thanks for hanging with us. This is, Chris and Erin. And this is necessarily guys where we talk about everything internet, I crypto, media culture and everything in between.
01;11;45;13 - 01;11;51;28
Aaron
Just a quick reminder. The things we say here are done in a personal capacity and should not be construed as investment advice.
01;11;52;01 - 01;12;00;07
Chris
Let's go, let's go.
01;12;00;09 - 01;12;00;22
Chris
To.