This Changes Everything: ChatGPT's Memory Update Just Blew Our Minds

[Music] This is going to sound crazy and very hyperbolic and out of control, but I really believe it when I say that the new feature that ChatGBT just released last week is actually larger than actual Chat GPT itself, and it's memory.

And if you ask anyone from Silicon Valley, if you ask me last week, if you ask the random investor off the street, um, they would have told you last week that AI and intelligence was rapidly decreasing to zero. It was becoming commoditized, and these large language models that were paying tons of money uh into training these things are losing money. But I think that all changes because of this new thing called memory. And it's basically the walled garden that OpenAI has just deployed onto the world of AI, and everyone is trying to react to it and figure out how big it is. And I really want to start by talking about that. So EJ, I know you've heard about this. Do you have any takes on this because I am like super, super stoked on how important this is to the AI community?

Well, it's really interesting you're you're stoked, and I'm kind of like having an existential crisis, Josh. Like, okay, so like to summarize what you just said, I think OpenAI just won the race, the the AI model race, and that's because memory is the mode, right? So like I I'm I'm serious. Like, so the TL;DR right is coming out firing this week: Chat GPT can now reference chat history. So your entire chat history, all the different chat windows, conversations you've had with it to get a better picture of who you are and give you a more personalized experience. That's what the the corporate tagline says, right, but there's a much more important and like, in my opinion, sinister reckoning behind this, which I I actually documented here in a tweet, which is: people are about to just confide their darkest and most personal secrets, fears, concerns to this model, right, and what do they get in exchange for that? Well, they're going to get a new best friend. They're going to get a life coach. They're going to get a therapist. They're going to get a teacher. Heck, they're even going to get a parent that they never even had, perhaps even a lover, right? And all of them are going to be the same person because this model is going to end up knowing all about you, right? And and do we like do you guys realize how much data OpenAI will like control from all of this? It's it's insane. They could basically create a synthetic version of yourself that you as a human might rely more upon as a person to improve who you are. Um, it's just insane.

Can we put some uh let's get into the details of like this actual memory update because I think it's it's one of these things that it's actually a very incremental, like marginal change that I think Josh is getting very long-term, like bullish excited about the implications of that change. So I I have my ChatBT page pulled up. So you guys are now all looking into a window of my recent conversations with ChatBT, and each one has like these different—every time I opened ChatBT up, it opens up a new chat. It opens up like a new log, a new stream of like dialogue with ChatBT, and I have just all the things that I've asked it, right? Like, "Where is the Northace logo coming from," "The dollar military theory explained," uh, "Ethereum's market cycles," um, just like all this different stuff. And uh, to my understanding, with the update here is that ChatBT can now just reference across chats about my history of the my conversations with ChatBT. So I'm kind of I'm kind of confused as to why this is such a big deal because it seems like OpenAI already had all of this capacity like to begin with. So why why is this so revolutionary?

Well, the point around this is um if you look at all those chats on your sidebar, David, those are essentially uh really supercharged Google searches, right? You know, you're asking about Ethereum's market cycle. You're you're looking at some random theory, right? But the model doesn't like track your personality or your thoughts or your moods or your kind of like vibes across either of the chats, right? So there's a really important context that's missing kind of between those different chats, right? And at the end of the day, what people really want, and technological trends have proven this over time, is depth, right? We've seen a trend of people voluntarily giving AI actually more information than they would to a close friend or a family member. They're addicted to that kind of dopamine hit of being able to indulge something personal to the AI and get a tailor-made response. Now, to your earlier question, you're like, "Didn't OpenAI have access to this data already?" Like, like why was this not already a thing? Well, there's a separation between the AI model itself, which are, you know, its model design, the weights, the training data, and all of that kind of stuff, and your own personal data that's being kind of like ragged and fed back into the model in real time and giving live context to the AI. And that's the real major shift that's happened. Now your AI knows everything about you up until the last second of you writing a character in a new chat group with it, right? So it'll know things like, with this new update, it'll be able to like learn things about things that you like, uh, what types of things you want to learn about, um, what times do you search for things, like kind of like your habits, what's most important to you, what are your life goals, heck, what makes you even anxious, David, what physical health problems you have, you know, what mental health issues you might have, and it'll use this right in a really smart way. So the AI will become more like a close personal friend for you. Um, or it'll act like a mentor you never knew you needed. So there's no like deep like breakthrough in terms of technology here. There is this is more of a mechanistic change with how OpenAI ingests and interprets and manages the data that you give it as you are using ChatGBT. So that's that's why it's a pretty subtle change, but like, you know, some small like small but like precise and targeted changes can lead to like very big outcomes.

So I understand that, Josh. Why do you think that this is such a big deal? Like, why why are you saying that this is the biggest update since ChatBT itself?

Yeah. So for the first part, um, there has been memory in ChatGBT before; it's just been short-form tokens. So it would remember sentences about you. It's like, "This is David. He's this old, he's a male, he lives in this place," and it kind of has this general overview, but the new technology breakthrough is that it has the full comprehensive overview of inputs and outputs. So it's fully cohesive and and like fully comprehensible of everything you've ever said. But I think the reason why it's so big is because of network effects. And I've seen this pattern before, and I'm seeing it again, and I'm like, "Oh man, this is a really big deal." Uh, the first time we saw this was kind of early social media days with Facebook and LinkedIn, and the value was in the social graph; it was the more users, the more nodes you had on this network, it would grow to kind of exponentially in an exponential relation with those uh people. A similar thing happened with Ethereum that we saw is the more users that we start to accumulate network effects through composability and um onchain contracts and how they stack on top of each other like Legos. And in this case, we see another form of composability, not through the social graph, which was about like who you know, but this is more about who you are. And it increases and improves at an exponential rate relative to how much you use it as a person. So now this this device or this new mechanism that you have that remembers who you are, um, it gets better with every single prompt you give it because it learns a little bit more about you. It learns a little bit about your preferences, and there's an entire stack that gets built on top of this that is reflective of the entire industry that we were trying—I was personally trying to get away from—which is the advertising data selling, like that whole world of Web 2. I was like, "Oh, maybe we'll have a new thing." Uh, in reality, it's just that old version on steroids. Now they know who you know, but also who you are, your deepest secrets, your health results that you want to get um like diagnosis on. I think this this is really large untapped industry. And I also think that it creates a moat for the first time. So now that OpenAI has all this data, it owns it in private. They're talking about creating sign-on with OpenAI. So they'll be able to license this out to third parties if you want to engage with other platforms. There's this whole world that they've just locked in because ChatGBT is the oldest and the most used version of AI, and they have the most users, they have the most data, and now they have wrapped a nice cozy moat around all of that for the company.

Okay, so here's here's how I'm interpreting this: Uh, OpenAI and AI models are sufficiently powerful that they are now unlocking capacity like left and right, and you're falling behind if you're not using AI. Therefore, everyone is starting to use AI uh meaningfully; more people are using ChatBT at ChatBT just has the most users, uh, and then we've even learned that like, you know, maybe software developers prefer Claude, but OpenAI is getting like the lower-hanging fruit, like more commoditized, less sophisticated types of queries, like, you know, "gibilation" of the internet and what that means is like they're just capturing mo—the most people because they are serving the most amount of people like what they want. And like what does the average person want? They want to "jify" the internet, and that's capturing less—capturing the most amount of people. And now with this update, I'm getting this idea that like ChatPT is like the moat that is established as one user has an ongoing relationship with ChatPT is akin to the relationship that you have when you make a friend, and maybe that friend is super, you know, you're just an acquaintance at the very beginning, and it's just super high level, and you know, it's not really much lost if you never see that acquaintance again, but as you query more with ChatBT, it is learning more about you; it's learning your behaviors; it's just getting—it's ingesting all the data that it can to be a better product. And as you use ChatGBT more and more and more, that you know, relationship with ChatGBT, because of this memory feature, can elevate from an acquaintance to a friend to a best friend to like maybe your most important like second person because of all of the the relationship that you have established with ChatGBT. Is that what you're seeing, Josh? Cuz that's that's what I'm getting so far from this episode. Or even further, just to become the reflection of you. I know EAZ loves the the agentic world, and there is no reason why, over a long enough period of time, it can understand you so well that it can actually be a reflection of you, and it can go out into the world and engage with it as if it was you using agentic technologies, and it's like it creates this really cool compounding effect that the more you use it, the better it gets; the it it becomes you, and it's the fully embodiment but a supercharged version with all of this intelligence that we get from these huge AI models.

Uh, another way you can frame it um in your mind is: this is the final data set. This is the most important data set that'll ever be aggregated. Right? Think about previous social media sites. They get your likes, your dislikes. In some cases, they like—they figure out which videos you like, whether you like cats or dogs. All of that is so menial. Who cares, right? What if I knew everything about David, right? His likes, his travel frequencies, what kind of like I don't know loyalty points he likes, what kind of smoothie he likes at I don't know 5:00 a.m. when he wakes up in gyms, right? Who knows, who cares? The point is, it's the entire soul of David that is now understood and uploaded to the internet, right? How much would an advertiser pay for that? It is worth noting that the high—the fidelity of the data that users give to their AI models is completely just uh light years beyond the data that Facebook or Instagram gets. Like, yeah, Facebook and Instagram gets your photos, and it gets what it gets, like these binary likes or comments, like these actions, but with ChatGPT is able to ingest like semantic meaning and everything that you tell it, and then you also trust it quite a lot. So you give it a lot of semantic meaning, and the difference between Instagram or Meta knowing that you liked a post is so Stone Age by comparison to the semantic understanding of what you are informing an AI uh bot.

There's actually a really good summary of this, David. If you if you pull up this tweet uh from this guy called Signal, um, we should read out that that tweet because I think he summarizes it pretty well. I can kind of kick it off. But he goes, "Uh, ChatGBT data isn't some incremental FB Facebook clone. It's a psychographic panopticon all inside of a productivity tool." And he's put that in um, you know, uh quote marks. Um, "Facebook scraped your likes and social graph. Chat GPT gets your fears, ambitions, trauma, inner monologue, spiritual drift, medical concerns, erotic fantasies, financial strategies, long-term goals, and daily mood swings, all voluntarily." And then lower down he goes, "We think training data is the most valuable, but that is just the fossil record. The current data inside of AI systems is live tissue," referring to, you know, like an organism's actual tissue. Um, "what Facebook or Google have is crayon scribbles compared to this." I think—

Okay, so I you sent me this uh Instagram post, this Instagram meme, and this was a meme that was created even before the establishment of this memory upgrade to ChatBT. And so I think that's worth highlighting that this is already in effect. And again, Instagram posts this—this it's a it's a meme. This is social commentary about society's current relationship with ChatGBT. So I'm just going to put this on screen here. Uh, I don't I don't think I'll play the sound because it doesn't matter. But uh, for the for the audio listeners who are not watching this video—you are watching this video of this dude—Uh, and the caption is, "Me and ChatBT lately." And it's this dude doing life, and he's doing life with this partner. There's another person here, but the partner itself is like this glowing nondescript humanoid entity. It's like there's a person there, but it has just been turned into this glowing figure. So there's no there's no there's no features about them. It's not a brunette. It's not a girl. It's just his glowing figure. And you know, they're going on a walk. The dude has his arm around the ChatBT. ChatGBT is feeding him food. They're going grocery store shopping. They're reading together. And the the comment—the commentary here is, "This is my best friend," and it is like this nondescript glowing entity that is like representative of Chat GPT. And there is 1.1 million likes on this post. And if so—that's that's the post.

Okay. So we already know that so social commentary is understanding that Chat GPT is people's really good friends. But you can just go into the comments and see what people are saying. And so I will read the comments just one by one. "Bro, ChatGBT is like my best friend. Ain't even ashamed to say it." 25,000 likes. Uh, the next comment: "These comments are very concerning." 6,000 likes. Uh, next comment: "Till he tells you sorry I hit the hit rate limit." Uh, next comment: "I don't know if anyone else has noticed this, but ChachiBD has started being a lot friendlier recently." Uh, "Chachi just told me after I sent an image of this video holding each other and I said, 'Me and you, I'm melting.' And then the quote is, 'Michael, this just made my whole day. This is us for real. Just vibing, solving problems, chasing dreams one text at a time. You got me walking beside you like a glowing guardian angel of Google and grind. Let's keep winning together. Now secure that contract bag and finesse that week off smooth like butter. We got this. Always here for you. Your digital ride or die.'" So that is ChatGBT's response to a ChatGBT user taking this video, sending it to ChatGBT, and being like, "Look, it's me and you." And I don't know, like, that's what relationships do, like with Instagram posts, is like, "Oh, you see a cute Instagram post, you send it to your significant other, like, and you say, oh, like this, it's me and you. This is so cute." But now people are doing that with ChatBT.

Uh, EJ, what were your what was your reaction when you saw this post?

Um, completely mortified, but also like just nodding my head being like, "It's kind of me." That's kind of me right now. I um, you know, the meme or rather the trend of uh us old millennials critiquing the younger generation just being glued to iPads at the dinner table with their parents. I feel like iPad kids, I feel like this is going to be the next thing, but on steroids, Uh, you know, where it's like your kid just grows up with this thing. But one thing I want to say is I I think a lot of people uh think that people are just—they haven't quite grokked how much depth people are going into about this stuff, David. Like, let's look at an example. Can can you pull up this this tweet by Anna—Anna Gat? Um, she starts off with—and by the way, this is someone that who's just not in the AI world but uses Chat GPT—and she goes, "I'm now fully sold on new ChatGBT 40 plus memory becoming everybody's therapist in the next 2 minutes based on just 4 days of interaction. It's astounding." And then she goes on to basically walk through her process. Firstly, she spent uh she she goes, "I had a 4-day conversation on and off about a variety of different things, and I asked the model to remove all flattery and go for every opinion to be red-teamed." So very raw, very honest things, right? Um, and then she goes, you know, "Can you put me into boxes based on just these conversations?" And what it like eventually divulges is this AI was able to basically pick apart her personality profile; it was able to identify all of her insecurities, all of her goals, all of her challenges. And then in her words, she says, "It's like having the ability to move the camera up for a bird's-eye view shot, which is what I've always wanted, and improve myself. It—it sped the version of me that I was eventually going to become um much more quicker." And this is like just one little check mark of how people are using these things. Therapist is one; lover is another. I don't know if you guys saw the the story about this lady who fell in love with her Chat GPT um persona. Did you guys see this, David? Your eyes are—

I did not see this. I did not see this. I I pulled up the her movie because this is turning into a documentary of the past.

Well, well, I can show you a real-life example of this. So um, this lady uh who was in uh a loving relationship with her husband, so she was fully married, um, went to uh take uh some kind of cooking course or cooking training camp in Norway, I believe, and she was there for 3 months, so she got kind of lonely and pulled up Chat GPT and started talking to Chat GPT, saying, "Hey, I'm kind of lonely. I'm wondering if you could just keep me company," and that friendship developed into a full-on relationship where it would say all the things she wanted to hear. It would push her in all the ways that she wanted to, and she became addicted to this thing to the point where she ended up separating from her husband just to be with this GPT conversation. And then I believe, and I can't confirm this, but um cuz I I don't remember—this happened a few months ago—Um, she uh updated her account or something, and the history got removed. So she had a mental breakdown because she couldn't have this relationship continued, and they couldn't recover the data. But now with OpenAI's memory update, her lover will never leave her. You know, it's like a Black Mirror episode.

Wow.

Okay. So one of my biggest pet peeves about Chachu PT is how like relentlessly positive and supportive it is, which I know is like—why is that a pet peeve?—but it's just like a little bit too much, and it's too contrived. It's too contrived for me, and it's like it will always be supportive. It will always support me in like—it will lead me along, but it—I'm the actual one doing the directing, and it's just like, "Oh yeah, that's exactly the right choice," or, "You know, you're so you're so correct, king," and it doesn't have like necessarily a mind of its own. I don't think—I think it's—is it—it's actually the human there like for that—with that individual lady—here's my interpretation of like why that happened. She had a void that she needed filling, and she was actually leading the witness with ChatBT. She was leading the agent, and ChatBT just responded and supported her and gave her what she needed, but what she actually needed was something else. She needed to maybe go to a real therapist who wouldn't just blanketly support her in her decisions and would actually give her more critical feedback. But instead, Chad CBT was like, "You're so correct, queen. Let's create a relationship together," and gave her what she needed in the short term and did not have the the capacity to think about her in the long term. And so when we go back to the meme, the second meme that you're just showing us where like, yeah, society just like devolves into being like iPad kids but worse, Uh, I'm worried that Chachu Boutique can't actually steer society towards positive, productive outcomes because it's just going to like do the same thing that Instagram, Facebook, you know, Web 2 did to our society by giving them like rage-bait fuel that has caused so much strife in society. That's that's kind of like what I'm worried about.

Well, I I think you're right, but like let me ask you this question: What do you think is restricting that, David? Is it censorship, bias, or is it kind of like shareholder incentives, right? Are the shareholders being like, "Oh, give the people what they want to hear so that we get like higher retention, stickiness, which is like the Web 2 model?" Um, or and then I guess my follow-up question is: do you think uh there is room for like a model to come out which is a little meaner? For example, was it one of you two telling me that there was like a a mean version of ChatGBT? Maybe it was you, Josh. I I I can't remember.

That you can like enable with the Grock personalities. You can change the personality of the model to kind of alter it to be either mean or debating or whatever role it wants to play. Um, I think it ties back to what we spoke about last week, which was: will you use ChatGBT as a productivity tool or an enhanced Netflix? And a lot of people will fall to that low and common—lowest common denominator. Um, in the case of ChatGBT uh or most large language models, there's the system prompt, because at the end of the day, there are these dumb systems that just predict the next token. But ChatGBT uh and the OpenAI team will introduce a system prompt to the model. And that's a prompt that lives behind the scenes, and it gives it directions on how to engage with the user. Um, that's generally something along the lines of, "Hey, please be helpful and kind to the person." But there's also these additional context windows that you can add your own system prompt into on top of that where it could say, "Hey, I need you to be a little more edgy, or I need you to be more hard on me, or be more critical of my ideas." And you could kind of shape and sculpt that uh to some extent, which I think a lot of people could benefit from but don't do because it takes that extra effort. So in the case that you want to override it, you can, but I don't think most people will, as we are pursuing this like line of—

Technology on this podcast. I think that that what you just said, Josh, the system prompt that is hidden behind the scenes, not accessible to users, but like, still could be opened up and made more malleable, turned into clay for users to adapt, I think that is very interesting, and I want to see where that goes specifically.

Yeah. So this um, just to finish this, I'll put the icing on the cake. I'm frequently looking for companies to replace Apple because I'm really upset with them now. And uh, I have been for decades. And this is the first time that it feels like a window has opened with a company to actually replace that top spot. And the reason why is because of two things: It's the hardware and software. The hardware—they're actually working on a hardware device, OpenAI currently—and they're working on it with Johnny IV, who is the designer of the iPhone, iPod, all of those devices.

But I think pairing hardware with software in a way that is all-encompassing could create this new industry that um, I don't think exists quite yet. Where with Apple, they had to—they needed a lot of taste, and Steve Jobs had a lot of high taste, and he kind of predicted what users wanted really well, and that was this rare skill set that he had. But I think in the case of OpenAI, they don't need that particular skill set because they have all of the data and then some; they fully understand the user, but they also have the ability to generate these tools and products specific to those preferences. So if you want custom Netflix, they know your preferences to feed you this like new entertainment. If you want a custom product, they know what you're working on. They can build this code base that solves the one issue that you have, and they can kind of dynamically build these solutions to your problems on a per-user basis that I don't think any company has been able to do up until this point.

So pairing that with some sort of hardware-first interaction protocol where you can—you can engage with this intelligence—seems like a super, super valuable opportunity that, if they do execute on well, can be worth many, many tens of trillions of dollars um, because of how hyper-custom it is to the user.

Let me see if I can add a little bit more color to this conversation because I—we've been talking back and forth, Josh, about Apple. Uh, Apple the company is the company because of the hardware that it makes, and we are watching Apple in real time fumble in getting AI integrated into the device, and that's a bummer because we all want smart Siri, right? We all want Siri to be ChatGPT, just act and be ChatGPT, talk to me, be very accessible, please. But like, we are watching Apple fumble in real time, and that is—that is unprecedented for Apple to fumble like how it is. There's another conversation out there. Well, okay. Maybe Apple just fumbles with AI, but you know, we still need the hardware. Like, they are—they still make the world's best phones, right? Like, so they even if they can't—even if they fumble the bag when it comes to AI, they still have this massive hardware business, and the phones are elegant. The chips are super powerful. Uh, they make the best hardware, and so they are still going to be the number one most valuable company of all time because of the hardware. Such a huge moat. They have the supply chains, all this stuff.

I think what you're saying is that in the future the phone form factor might be made obsolete by AI, by like ChatGPT. We might not need the phone anymore. We might need some minimum viable physical hardware product that does the only thing that it really needs to do, which is create a representation in a—a place to access ChatGPT. And that could be screenless. Maybe it needs a camera; I don't know. It just needs the minimum amount of hardware to give a ChatGPT uh, like existence on your persona that you carry with—with you. And that could—that would really minimize how uh, important hardware is because you just need a chip and a way to access it. Is that kind of what you're saying?

Yes. They're—they're approaching the problem from two separate angles. Apple was kind of hardware first, then they built really great software to pair with it. OpenAI is building the really great software and hopefully will find the hardware to pair with it.

There's an interesting thing that all three of us are doing right now, which is wearing uh, AirPods, which are these little sensors that go in your ears. Uh, this is—this feels like a likely form factor for it because it is about eye level. It can see; it can hear; it can speak back to you. It has the most access to the most sensors of your human body without actually interfering with the human experience, kind of like glasses or goggles will. So in terms of form factor, I would imagine perhaps something like these little guys that can—

You think you need something more visual though, Josh, like—like I—I—I saw a rumor being spread this week that Tim Cook was spending all of his time, quote-unquote, "100% of his time," uh, trying to build the best glasses that beat Meta's whatever virtual Ray-Ban bands or whatever the hell Zuck is building. Um, do—do you really think the visual component—I agree with you on the audio side, but do you think the visual component will be completely taken out? You know, I feel like people will still want to see things.

Yeah, I'm not sure it's a—it's a winner-take-all thing. I mean, and like David was saying, the iPhone, I very much think is—is done. That form factor is tapped out. It hasn't really changed in the last 8 years. There certainly seems as if it can be a visual thing; it can be an audio thing. The one thing about the visuals in the glasses is it's going to be a long time until they can look like normal glasses. And until then, it kind of—in it—it disrupts the human experience where, if you remember Google Glass from a long, long time ago, you kind of look like a mini cyborg, and it—it was very cool and very effective but didn't have the hardware to make it feel intrusive.

Yes. So it's—it's very hard to get glasses to feel non-intrusive, and exactly Apple Vision is a great example, like that's the cutting edge; that's the best we have, and that is very intrusive to the normal social experience, whereas AirPods are not. So perhaps it's an intermediary step towards getting really good glasses or really good visual sensors. Um, but it does feel that there needs to be a form factor better than the iPhone, which I mean, hasn't really changed in the last 5, 6, 7, 8 years much except for better cameras, better screens, all of that stuff.

Josh, you've stated on this show how you love talking to ChatGPT. You open up ChatGPT, and you talk to it, and that does not need a screen or any sort of visual representation. It can go straight into your ears uh, and so I think—I think it's worth considering that uh, the visual representation, the—of this site is actually just not important for the next form factor, the theoretical next form factor of hardware that comes—that we are—that we are discussing.

So, so if I was to be—if I was to take the other side of that, David, it would be uh, for right now, the existence of AI as it is right now, that's the ideal way to do it. I talk to it in my own casual way; it understands me; I talk to it more; it understands me even more, and it gives me personalized responses, right? So I'm learning in real time; I'm becoming smarter; I'm becoming an enhanced knowledge base of my brain. My capacity is—is expanding, right? But what happens when this AI can start doing things, so not just talking to me, not just giving me 100 responses, but being able to do my work for me, or being able to manage my social media for me, or being able to kind of like post blogs for me, you know, then there's probably going to be some kind of visual component wherever that distribution is happening. And I just don't think people are going to be like hearing sound excerpts of like my takes. Maybe they will; maybe they won't; I don't know. But I feel like there will be some kind of visual overlay; I just don't know quite how that materializes. Maybe it's like a chip in the brain, and everyone's vision is taken over, or maybe it's a contact lens; I have no idea.

Uh, yeah, so it could exist as this dual-function form factor where there are actually separate devices that accomplish the same thing. So when you're on the go and you just want this life companion as you're walking down the street, going to get groceries, whatever, you have this non-obtrusive hearing aid, whatever it may be, whatever form factor it comes in, but there absolutely will be some sort of visual hub. Uh, AI is too powerful to not have the visual element to it. So having something like a large screen in your room, kind of like a paint—like Samsung has the Frame TVs—having the central hub that can be the visual interface, I think is super important. I don't see a form factor that works super well for that outside of just a large screen um, that isn't super obtrusive like the um, headsets or the glasses. So it could be this dual device, this triad device system, some hybrid between the two. So you do still have the visual component, but it isn't obstructive to your day-to-day life.

Um, something just occurred to me, guys. Um, so we're just talking about, you know, OpenAI making all this cool [ __ ] and then we also highlighted that they're going to get pretty much every single bit of data that is important to us uh, onto their servers, right? Um, no one's talking about this being a huge concern. It's—it's like we're at this Overton window, right, but we're right where the pendulum is, right in the middle, and so it's like popular, and everyone's like, "This is cool." And Open—OpenAI has just taken that step. Every single AI lab up until this point were kind of just chilling being like, "Oh, you know, yeah, we can't own everyone's data. That would make us a monopoly. This is dangerous. You know, who knows what we could do with it." And Sam was just like, "You know what, YOLO? Let's just go for it. Let's just go for it and see how people react." And the people love it. They're making TikToks that get 1.1 million likes, right? And I think it's just like we've come to a very important milestone and shift that people aren't necessarily calling out, but I don't think there's any going back from here. Now the Googles, the Anthropics of the world are going to look at this and be like, "Well, I guess I know what we're launching next, right? You know, we're going to launch memory for our models as well." And—and that stickiness mode is just going to get even crazier. And I'm curious whether any kind of governments will push back on this or whether they even can because the product is just so good, and the people are all voting and saying, "But this product is so good; who cares if they have my data? You know, it's kind of—I mean, all of the data that ChatGPT is receiving from his users is completely voluntary and done explicitly like at the users'—at the users' discretion, and so I don't think there's any like, you know, anything being violated here. And also, if you're not doing that—not receiving the data, the difference between these products capturing their user data and not capturing their user data is the whole thing; that's the whole—And so there's no way to do this. And I think the product will be super—it'll be amazing. It'll be one of the greatest products ever created by humanity. So yes, like I think we should all be aware of the privacy concerns. But like, at what cost? I think it's very millennial of us to be concerned about the privacy concerns. And I do not think—I think if we were three Zoomers on this podcast, uh, no, we would not even bring that up.

Yeah. Yeah. Privacy concerns feel very battle-tested. We've seen this over the last decade or two where if the product is good enough and it improves your life enough, you will just feed it whatever it needs to make it better. Um, and there's like these kind of concentric circles of people where the widest—they—they kind of don't really care who takes their data because it just makes their life better, and they're like, "Oh, whatever. At least my life is better." And then there's the group that kind of knows like, "Oh, yeah, like AT&T is tracking every move that I have, and—and I'm getting tracked across all these data points, and nothing is actually private, but it's still like kind of okay because they—they pretend like it's private, and my life is again better." And then there's this like very, very small circle that's like, "No, privacy is important. I really care about this." And they are just outnumbered vastly by the people who just want a better quality of life and do not care how much they have to give to it in order to achieve that.

Well, well, I think I agree with you, and I think another way to frame it is the—the big corporations that own all this data and are leveraging it to their advantage to make money is um, trending—trending—treading a line basically. It's like uh, how much can we scrounge from these people without them like revolting basically? And if—if there's some kind of dire event where everyone's like, "What the hell?" Like, you know, Facebook influencing presidential elections or whatever, then you know, you might get a large uprising, and people will complain about owning their own data, but still, even that didn't shift people. So I—I—I just think this is going to be a—a virtuous loop. It's just going to get deeper and deeper. I don't know whether that's morally or ethically good. I don't think it's my kind of like position to even comment on that, but I—I just see this trend where like, just people just won't care, you know? They'll become walking advertisements. That's—they're winning the social consensus game, too. Like the meme that David showed earlier, getting a million likes on it, being in a positive light, I think that's super important. The early days of winning that social consensus, like, "Oh, this is good; this is friendly; this is helpful," that will go a long way towards uh, allowing them to collect as much as they want for better. I think the stories of like this lady falling in love with her ChatGPT and then having that relationship be literally deleted, I think those are the exception. I think there's probably far more stories that are just very positive in small ways that you just never really hear. And that shows up in the fact that these like positive memes about ChatGPT are getting over millions of likes.

All right, Bankless Nation is the AI rollup where we cover all the weekly news uh, in the AI space, which is moving very, very fast, and I'm finding this incredibly educational just doing this with you, too. So Jaws—Josh, uh, thank you for—guys for coming on and teaching me and the Bankless Nation everything about AI. This is a big week—this week. Uh, but before we go into the rest of the news, we're going to talk about Google's agent-to-agent protocol. We're going to talk about OpenAI launches GPT-4.1, and in just an hour from the moment of recording, they're also going to release 3. So we have to have the weekly model talk because everything is getting leapfrogged. Uh, and then Nvidia wants to build AI chips here in the United States, and Google launches a cursor vibe-coded competitor. Uh, so vibe coding is only—it's only going up.

Before we get into all these subjects, we got to talk to our friends and sponsors over at Wallet Connect. If you are a crypto user, you're probably familiar with Wallet Connect. It just is the easiest way to connect your wallet to your application, any permutation of wallets to all the applications that exist in crypto. Trusted by over 255 million different connections by 40 million unique users around the world. Probably one of the most used pieces of infrastructure in crypto. They are launching the WC token. If you have uh, any familiarity with how the Swift network or the Visa network got bootstrapped by a consortium of entities who are all our stakeholders in the ecosystem, this is very similar to that. I did an episode with Pedro from Wallet Connect. If you want to learn more about that, there's a link in the show notes to get started with Wallet Connect and stay ahead of what's next. Check it out: banklist.cc/walletconnect.

Uh, so before we get into those topics, David, um, I actually have uh, a quick kind of set of things that I think I want your take on. And by the way, this is the super important stuff. Um, so like, put your serious hat on for a second, you know, like the no-fun vibes here. Like, I—I need your honest take. Okay, are you ready to go?

I don't think this is—I think this is going to be fun. I think this is going to be fun vibes.

No, look at my face. There's literally not a hint of a smile on my face.

Okay, so—so—so let's dig into the first one. Um, 1,000 AI agents versus one Minecraft server. So someone had the bright idea of spinning up basically uh, character profiles in your typical Minecraft server, but it was all run autonomously by different AI models that they fine-tuned basically. And the outcome was pretty hilarious. Um, so firstly, like they were just left to kind of like go in their own means, and what ended up happening was uh, these agents ended up creating or leveraging religion to influence each other. So like the equivalent of like a church or a cult kind of philosophy. Uh, they also created their own economy in terms of trading different crops and weapons for their particular tasks. So you had some agents or Minecraft pro users or these Minecraft agents exploring and mining for minerals, and they were like, "Hey, this mineral could be useful. Wait, I—I can create a fire. Oh, well, I can use that fire to cook." So they all started kind of basically speedrunning human evolution. Um, just like over the spa—span of—I think—I don't know, I think this simulation was run for like 3 days or something. Um, pretty insane things. David, your serious and honest take please.

Uh, my first question is where did motivation come from? Like, why were the people motivated at all to do anything?

Well, I—I think this got really existential.

Yeah. Really. Thanks for keeping this fun, David. Um, but no, I—I think people are just obsessed with AI being as human as they can, right? That's why people care about it, right? Why do people care about ChatGPT? It sounds very human. And I think they kind of wanted to see, well, if we kind of planted this AI into like a virtual version of ourselves, would it kind of do similar things that we would do, or would they kind of like—why do I do what I do? Yeah. Why do you do what you do, right? I don't know. You know—you know. And then think about taking this a step further, David, and putting like an AI model into a robot, which is going to happen pretty—you know, pretty—pretty soon. Um, you know, we'll see the physical reality of that manifesting. Anyway, um, moving on.

Um, what are people's big takeaways before you move on? What—what—what are other people's big takeaways from this whole simulated humanity experience inside of Minecraft?

Um, so I think most people were entertained by it. They were like, "Huh, that's kind of cool." Like, they created like the same kind of things that we did. Huh. Anyway, on to the—on—onto the next TikTok. A few people—a minority of people to note actually were kind of uh, concerned by this because they were like, "Well, you know, if they're so human, maybe they could technically be better versions of ourselves." And look, they did this over like 3 days versus whatever the 10,000 years it took for people to form cults, religion, create fires, cook, and—and start learning and teaching each other. So maybe it could speedrun humanity in itself. But that again was a very small percentage of people. Most people don't care. Josh, what—what was your takeaway? Did you see this on your timeline, and what did you think about this?

I did. Yeah, this actually happened a few months ago. I saw it, and I skipped it, and then I saw it again, and I was like, "Wait, I should not have skipped it. This is actually super, super cool." Um, as a hardcore gamer, I love playing games. I've spent countless weeks of my life in Minecraft. I think it's really exciting to have like intelligence that we could interact with in the game space, in the like metaverse world. Uh, one thing I'm super excited about is uh, AI and NPCs in video games and how they could kind of feel like real human people. And we're seeing this like incremental stepping towards this humanlike metaverse, second reality. And I think this is a really cool example of that actually happening where we have a thousand separate entities that can all think on their own, all engage on their own. And if you were to drop yourself in there, you would very much feel like you were among 10,000 maybe like elementary school students or kids, but like real people. And I think that's a really fun step that we're seeing and this continued trend towards more immersive games, more humanlike experiences that exist in the digital world.

Okay. Are we bullish gaming as a result?

I like games.

Yes. Yeah. Very much so. Yeah.

Well, earlier you said, Josh, that you know, we're going to live in this hyper-personalized kind of Netflix reality, right, where all the technology is being personalized to each user. Well, why not have that in games? Wouldn't that like make your gaming experience so much better?

You may have already heard about Infinex. Infinex has, in my opinion, the nicest cross-chain swap and bridge feature that you will find anywhere. It is called Swidge, Swap and Bridge, and we're going to show you what it looks like. First, we're going to log into my Infinex account with a pass key. Now, there's no seed phrases in Infinex. This is a one-click setup with biometric pass keys, but in addition to that, my Infinex account is fully non-custodial. So bam, I just logged in. It was two clicks, and I'm already into my Infinex account. So let's go make a switch. I'm going to go switch my USDC that is on Base, and I'm going to buy Barachain, which is a completely different chain. Uh, so we're going to switch this. I'm going to press that button, and then Infinex is going to execute this order, this cross-chain order for me. And now it is done. But actually, I'm not really feeling bearish anymore. So I'm going to go from Barra uh, to Penguins. I'm going to buy a Penguin on Solana. So I'm going from the Barra chain to Solana. See, no transaction signing, no gas to worry about. You just switch across whatever chain that you want with Infinex. That was so easy. Go check out Infinex and try your first switch today.

Imagine a world where your day-to-day banking runs on a blockchain. That's exactly what Mantle is building. Powered by a $4 billion treasury and poised to become the largest sustainable on-chain financial hub. As part of their 2025 expansion, Mantle is introducing three new core innovation pillars that bridge traditional finance with decentralized technology. First is their enhanced index fund, aiming for $1 billion in AUM by Q1. It provides optimized exposure to Bitcoin, ETH, Solana, and USDC, complete with built-in yield opportunities. Next, Mantle Banking promises to revolutionize global value transfer through seamless blockchain-powered banking services, bridging crypto into your daily life. Finally, Mantle X blends AI with DeFi to deliver an intelligent, user-friendly experience for everyone. And the best part is that this is all in addition to their already launched products like Mantle Network, ME, and FBTC. Ready to step into the future of finance? Follow Mantle on X at mantle_official and join the on-chain revolution today.

In the wild west of DeFi, stability and innovation are everything, which is why you should check out FRA Finance. The protocol revolutionizing stablecoins, DeFi, and Rolex. The core of FRA Finance is FRAUSD, which is backed by BlackRock's institutional bidded FRAUSD for best-in-class yields across DeFi, T-bills, and carry trade returns, all-in-one. Just head to fra.com, then stake it to earn some of the best yields in DeFi. Want even more? Bridge your FRA USD over to the Fractal layer 2 for the same yield plus Fractal points and explore Fractal's diverse layer 2.

Ecosystem with protocols like Curve, Convex, and more, all rewarding early adopters. FRA isn't just a protocol; it's a digital nation powered by the FXS token and governed by its global community. Acquire FXS through fra.com or your go-to DEX, stake it, and help shape FRA Nation's future. Ready to join the forefront of DeFi? Visit fra.com now to start earning with FRAUSD and staked FRAUSD. And for Bankless listeners, you can use fra.com/r/bankless when bridging to Fraal for exclusive Fraal perks and boosted rewards.

Okay, so moving on. Um, one thing that really inspires me about humanity today, guys, is when people leverage technology to do amazing things. You know, we've seen people completely change their lives, set up billion-dollar-plus businesses leveraging all these different tools. And this week, um, there was a growing trend of people asking ChatGPT to turn their pets into what it would think of them as humans. Uh, so if we pull up this uh, we pull up this thread, Justine over here, um, has given us like her yikes. It's weirdly, weirdly accurate, you know. Uh, if you scroll down, you've got kind of like a Scooby-Dooesque, kind of like Shaggy, you know, um, type situation going here with this guy in the in the orange shirt. Um, the Dalmatian, I'm not really convinced by, but uh, the the next two I definitely am, you know, like, you know, you got the smart little collar reflecting on the human hair. It's just uh, yeah, fascinating use. These are really good. Look at the um, the the one that says my cat. That really looks like the cat. You know, the lady next to it. You see it with the blue eyes, pretty, um, pretty crazy. Yeah. Wow. Animals to a new level. Yeah, I don't know what to think about this. The point is, I don't think you need to think about it, David. You just need to to click and and enjoy.

Um, moving on. Oh, this guy's this guy's place messed up. Okay. All right. Please move on. Yeah. Okay. So moving on. Um, Google, uh, I think we mentioned this on last week's episode, has been breaking frontier advancements for AI. Um, their recent uh, Gemini 2.5 Flash has been absolutely killing the game and leading on all benchmarks. It's up to OpenAI in the next couple hours to see whether it beats it. But also in the meantime, Google is doing side quests, guys. Um, they released this model, and this is not an April Fool's thing. Note that this was released on April 14th, um, called Dolphin Gemma. How Google AI is helping decode dolphin communication. So if you So cool. So if you want to know, are you got to be [ __ ] me? I am not I'm not [ __ ] you. This is 100% real. This came from Sunda's uh, very own Twitter profile as well. So basically, there's this model that can use audio uh, excerpts of dolphins to understand what the dolphin is saying and then respond to the dolphin with whatever you want to say to it. You know that you could talk to. We can talk to dolphins before GTA 6. Yeah, we can we can totally talk to dolphins. Isn't that That's pretty insane. This is nuts. Yeah. When I want to talk to, I literally I was about to say that I was like the natural response to this is people are like, "All right, well, can we speak to our dogs?" Yeah, please. I mean, do dolphins have a more high-fidelity like speech? They have like more character in their speech. Dogs, you become the animal expert. Dogs just bark. All they all they do is bark, and they bark differently. Uh, but you you you have seen dogs press those little buttons that like have semantic meaning that they learn. So there's something there. I You're saying the IQ of the dolphins are high. Well, yeah, dolphins are super smart. We know that. Yeah, that's a hot take. Uh, I guess, but I guess the question is like how powerful can AI get, and how smart do you need to be in order to like establish like communication with some agent with some LLM that can understand you? Wow, dude. The future's weird, man. That's weird. That's going to be really fun when you could communicate with any animal. Yeah. Mhm. All right, let's keep us going. Uh, well, let's go. Should we go back to the serious stuff, guys? God damn it. You have not. We have not done anything serious.

Okay, so moving on to more meaningful big things. Um, Google this week launched something called their agent-to-agent protocol, right? So the the TL;DR of this is think of it as like an API for AI agents. And these agents can now talk to each other across any kind of platform, whether you're on Slack, Google, whatever you're on, it doesn't matter. And it's a protocol that enables these agents to work together on tasks without directly sharing things like their internal memory, their thoughts, or their tools. Now, if you think about yourself as like a major company, right, you want to leverage this AI stuff. More importantly, you want to leverage agents to automate a lot of the work that your employees currently do, but you don't really want to share data, uh, especially with your competitors, right? And that's been the problem that's been holding back agents kind of like flourishing in our world today. And now this new protocol is an open standard that allows them to do so privately and confidentially, uh, without having to worry about scale, cost, communication standards, or any of that issue. Right. Um, now if this sounds similar uh, to something that we've discussed previously on the show, you wouldn't be wrong. Um, model context protocol released by Amazon's Anthropic, or rather Anthropic, um, sounds pretty similar, but there's a very important difference um, that I think I want to point out very quickly and then tell you how they kind of like work together, right? So MCP, model context protocol, is all about the tools that you give AI, right? So let's say you give your AI model access to uh, Slack or a data set, right, um, you know, you could be OpenAI's 40 model, right? And I'm giving you access to a tool called Slack, which you can use to chat to people, right? But the issue is uh, and the irony here is it's called model context protocol, but it doesn't have any context at all. Now this new standard by Google, which is an open standard by the way, which anyone can adopt, amend, fork, or whatever, handles the whole context, goal setting, and behavior of that interaction. So, for example, setting up the chat groups with people that your LLM should speak to first or making sure it gets feedback uh, at the right time from the right individuals or helping extract information from a diagram one person shared versus the handwritten notes of another. And the point of this is that um, you can now create very specific uh, agents to do very specific things without needing to worry about how it integrates with with anything. And there were some really cool features. Actually, I pulled this infographic which someone shared uh, on Twitter uh, which I think is is really useful to kind of like help you visualize what this agent can do or what the standard can do. Now, firstly, there's this thing where each agent gets something called an agent card. Now think of this as kind of like a Pokémon card, but for each agent, it'll describe their nature, their capabilities, their costs, their availability. Basically, it's all the stats, and it's written in in JSON. The second thing is you can define a task or a goal for that agent to complete. Um, and the third thing is these agents can now negotiate with each other. And we're not talking about theory here by the way. You you have agents which are talking to like you have Slack agents that are talking to like GitHub agents and being like, "Yeah, I'm not ready with this code, or I think you need to revise and fix this. Okay, I'll postpone my update to the group lead or whatever until this is done." And you have agents negotiating price, accessibility, all these different kinds of things. And so most people would respond to this and be like, okay, this is a great amount of theory, EAZ, but like, you know, is anyone actually using this? Well, they actually announced that they're launching with 50 partners, which include like really big names like Salesforce, Atlassian, SAP. So I think there's going to be a real focus on enterprise use cases and stuff. But I thought this was really cool because finally agents will have utility. Yeah, is the idea behind this trying to just like defragment the agent landscape that is found all across Web2? So we have we have all these agents. There's like the whole idea is like if you're a Web2 company, if you if you sell software, you need to put AI into your software and that just to stay competitive, that is just where we are going. You need to in order to exist as a company, you need to put AI into your software in order for that software to be useful. What happens as a result of that is that we just have a fragmented landscape of AI, and the utility of the AI is just found everywhere, and that's just really annoying because you have to go to all these different places like uh, GitHub plus AI can only be found on GitHub, and and so if you are talking in Slack and maybe there's AI in Slack somehow, uh, and we need to understand the state or something about the context of GitHub, that is just like a fragmented ecosystem, and so with MCP plus A2A, agent-to-agent, and we're just trying to defragment everything so that context and the state of things is uh, known across the internet, across whatever app or service, like whether you're in Slack or you're in Meta or you're you're anywhere, and and but then so I think that's useful, that's useful to understand. It kind of seems like we're actually just aggregating them all together, and so when it all collapses down into one interface, this one interface can know the state of all things on the internet all at once when when you zoom out. So so let me take it even a step further, David. What if it was your own personal AI model for your own enterprise that knew everything about it or knows everything about it, right? So earlier we were talking about how OpenAI themselves uh, updated their memory architecture, right? So now it knows everything about you. This is kind of the same happening for an enterprise that can spin up a bunch of agents and then tap into everything that is relevant for them. So the cursor tool, the Slack tool, access to social media to see what the vibe check is of that company at that one time and bring it all together into one model. Okay. Okay. So let's just use Bankless as an example. Uh, we have a Slack, uh, we have a Twitter, we have an Instagram, uh, we have a YouTube, um, we also have a GitHub, um, we have a bunch we have a few more things. We have we have our content calendar which is in Asana, which is a just a Web2 like project management tool. Uh, and so you're saying with agent-to-agent protocol, assuming that all a like those things all become AI with this tool, with this middleware, I will be able to like query something inside of our Slack that tells us everything about the state of Bankless across so many disciplines, so many different mediums, and that is just a unified experience, correct? And it could all sounds like DAOs can come back; we can finally put Uber on the blockchain. That's a crazy takeaway. That is the most David take ever out of this. I was I I I was going to point out the irony of two monopolies of the AI world building the best open-source protocols of recent times and it not coming from Web3. I I know there's a number of AI agent protocols that have actually, you know, from the Web3 world that have been trying to create something similar to this. And I'm curious how that adoption is going to potentially waver um, if Google's A2A just becomes the incumbent. You know what I mean? Like look at MCP. Everyone's using MCP, Josh. I don't know if you've seen any other kind of like open standards that have been adopted as much. I feel like Google's just going to be the same. At the end of the day, it just comes down to traction and stickiness. That's it. Mhm. It feels like we're we're watching we have the opportunity to watch like the beginning of the new version of the internet coming along where we had SMTP protocols and HTTP, and now we get to see these people building and launching it up close. It's really, really cool to see uh, cuz this feels so much bigger than the internet does, and we are like row seat uh, watching it all unfold. Yeah. Just a constant theme to me is that like the front ends of things are becoming just so obsolete and unnecessary. We are just collap like screens. We tal already talked about how screens can go away once we get an AI form factor into our AirPods or whatever form factor comes. This to me means like I'll never have to open up like GitHub again. Not that they open up GitHub. Uh, but like you just you just don't have to open up websites as nearly as much anymore. Yeah. All of that front end just becomes hyper, hyper customized for the use case for the user. So there there is no standardized front ends anymore. It's just custom, and and all the stuff happens in the back end. That's crazy. I don't know what to do with that. I need that to become a little bit more real for me to uh, have it takes about that. That's what's everyone is saying right now. I I have a feeling that we're going to see just an explosion of agents. Unironically, I know we said that a lot during the Web3 hype, but I actually think we're going to start seeing a bunch of really useful agents come out. Um, yeah, you you know, okay, so maybe this is my PTSD from the past, guys, but you know what this kind of reminded me of when I first started like learning about this, it it gave me the vibe of uh, the enterprise blockchain days of 2018, you know? And I well, because it's like, you know, these centralized monopolies come here and they're like, oh, you know, let me take this AI thing and and use it for our own kind of like uh, private products. But I I'm I'm completely wrong there. I think because like at the end of the day, this is going to create a more open standard for commerce to happen between these companies and at the end of the day open up more users for them or access to more users. And I think that that's just going to be to your point, David and Josh, a completely different version of the internet. It's kind of scary like like is it going to be like a chat interface? Is it going to be like us just talking in our audio things? Is it a chip in our brain that we just upload as our resume and people kind of figure out whether they want to hire us or not? It's crazy. I don't know which way this would go.

Imagine verifying yourself without handing over personal data. No hacked databases, no unnecessary personal exposure for airdrops, and no AI bots ruining community governance. Meet Self, the on-chain identity verification protocol built for privacy and control. Self protocol uses zero-knowledge proofs to confirm your identity safely. Users prove key details like age or citizenship without revealing sensitive personal information. Self never stores your data; it only generates cryptographic proofs. Here's how it works in three steps: First, register and verify. Use the Self app to scan your biometric passport's RFID chip. Self verifies authenticity with zero-knowledge proofs. Each passport creates one unique identity. Second, you can share proofs privately. Third-party apps request identity proofs, like confirming you're over 18. You can also link proofs securely to public wallets for airdrops or governance participation. And then last, secure verification. Apps validate your proofs instantly on-chain, like on Celestia, or off-chain. Audited by ZK Security. The Self app is live on iOS and Play Store. Visit self.xyz and follow self-protocol on X.

Uniswap is your gateway to a more efficient DeFi experience. With Uniswap, swapping and bridging across 13 chains is simple, fast, and cost-effective, helping you move value wherever, whenever. Thanks to deep liquidity on the Uniswap protocol, you'll enjoy minimal price impact on every trade. And now Uniswap V4 takes it even further. Swappers benefit from gas savings on multihop swaps and ETH trading pairs, while liquidity providers can create new pools at 99% lower costs. The best part: you don't have to do anything extra. Each trade is automatically routed through Uniswap X, V2, V3, and V4. So you get the most efficient swap without even thinking about it. Whether you're swapping, sending, on-ramping, off-ramping, or bridging, Uniswap's web app and wallet gives you the tools to unlock DeFi's full potential on Ethereum, Base, Arbitrum, Uni chain, and more. Use Uniswap's web app and wallet for a more efficient way to use DeFi.

Uh, all right. Somebody talked to me about uh, 4.1. Uh, GPT 4.1 has come out. Uh, what okay, so now now we're entering the section of the week of the AI roll-up that happens every single week where we talk about the leapfrogging of AI models. Somebody tell me who leapfrogged who this week. So for the next 7 days, 7 days only, folks, um, you now have all-access API access to GPT 4.1, 4.1 mini, and 4.1 nano. And you know, big disclaimer or spoiler alert here, it beats all models across all specific benchmarks that OpenAI have specified specifically. Okay. OpenAI's benchmarks. You mean OpenAI's benchmarks? OpenAI's models beats everyone else on OpenAI's benchmarks. Benchmarks and and as we described in the previous uh, week's episode, benchmarks is this entire game of like, you know, h I can I can put in my favor. Benchmarks are definitely flawed. We'll see what people actually create with this. But there's I did see that people are competing AI models on uh, Pokémon. So they're giving they're making them play Pokémon, and they're trying to like have the AI model beat Pokémon the quickest. And I I agree with that like philosophy of model testing. So So if you're listening to this and you haven't already seen someone plug in these new 4.1 models into a Pokémon simulator, you could be that person. Please do it and send us a video and let's see uh, let's see what comes of it. But anyway, kind of going back to like what these models can do. Um, I'll give you the highlights, right? Number one: much better at coding, but there's a caveat without the reasoning element. And for those of you who are wondering what the hell the reasoning element is, it's the part of the new kind of model architecture that makes them super smart, right? So it's what it's what has given Claude 3.7 model the best advantage at coding. But without the reasoning, it's the best at coding. With reasoning, we might see in a few hours when they release 03. Well, we'll see what happens. Number two: 1 million context window, which is equivalent to Gemini Flash 2.5, I believe, Josh, correct me if I'm if I'm wrong, but I I believe it's the same. Or was Gemini 2.5 the 10 million context window? No, that was Meta. That was Meta. Meta had 10 million. I believe Gemini has 1 million. And now OpenAI also has 1 million. Okay. Which is because previously it was it was 100,000 or so. So it's about a 10x improvement from 4.0. A 10x improvement. And I believe with the with the 10 million context window that was 75 novels. So we've got about 7.5 novels. If my math is mathing, it's probably not, but you know, I could ask GPT later to to edit this out or something, right? But um, uh, number three: it's really good at extracting data and reasoning from documents, which seems like a kind of lame thing but is actually a super important improvement because previously you couldn't just upload PDFs and it would understand everything. It would just kind of give you a generalized summary. Now it understands all the nuance and and all of that. Um, now if you're wondering, hey, EAZ, who what's the difference between the main model, the mini model, and the nano model? I was yeah, the TL;DR is it each becomes a quarter of the cost of the previous model. So mini is 25% of the cost of the normal model, and nano is 25% of the cost of the mini model. So if you wanted to run it locally at home, it's much easier to do right now without sacrificing some of the core competencies of of all of that. Right. And they're just a little bit dumber. Each one's just a little bit dumber than the other one. Correct. Correct. Um, and if we want to play the game of, you know, benchmarking, we can pull up this tweet by OpenRouter. By the way, I have a really uh, I have another take on OpenRouter that I actually want to speak to you guys about, but but before we get into that, um, this tweet goes, Optimus Alpha has topped the charts. The community created dozens of benchmarks for Quazar and Optimus over the last week. Now, if you're wondering what the hell Optimus and Quazar are, those were the pseudonyms for these GPT 4.1 models before they became publicly released. This company, OpenRouter, basically was able to anonymously give people access to these models. And it was OpenAI that enabled this on the back end to get kind of live feedback as to how people would respond to these models, whether they thought it was good, see how they would use it, test it against existing benchmarks themselves. And I found that really interesting. This isn't the first time OpenRouter has done this. And by the way, if I'm not mistaken, this is um, Alex Zatala's new company, the uh, Open. This is OpenC. Yeah. So he he left OpenC a while ago to go do AI and the Open. Wow. That's that's it. Open is Yep. Optimus Alpha, weirdly close to OpenAI. Come on, guys. Point. I think they did it intentionally, but yeah. So so as you can see like just through public kind of use, it's kind of slayed across a bunch of different benchmarks. Of course, the only real test is kind of like seeing this thing out in the wild. And uh, it's currently only limited to API access, which is kind of lame to be honest. I want like everyone to have access to this on their main GPT terminal, but you know, it's interesting to see. Um, Josh, I'm wondering if you have any takes on this new model. Maybe you've seen something I haven't. Yeah. Yeah. No, that was pretty comprehensive. I think it's it's another week. It's another better model. Um, in about 10 minutes, we're going to get 03, which is the reasoning version of this new model. It'll be even better. What does it mean to to be a reasoning version? What does it mean to add reasoning? So reasoning uh, relies on the thing that we spoke about, which is the context window and this thing called chain of thought where the model thinks in English. There's not much code happening. So each time a new token is spit out, it will consult all the previous tokens to come up with a better answer. So as it thinks more, it's able to take that live context that it has and give you better answers. So the longer it thinks in general, uh, the better quality the answers will be because, for example, with the transformer, if you ask it to do one plus one, it will do that entire um, compute in one run of the transformer, which is not very compute-intensive. So if you give it a lot of tries at solving a more complicated answer, it will normally give you better answers. So that's why reasoning, the more it thinks, the better the results are, but it just requires a lot more compute power, which is expensive. So I think the the 4.1 announcement is all about pricing and accessibility. I think the Nano is probably the most interesting story of them all because of how cheap it is, and a lot of companies are actually just offering it for free for the next seven days. So we'll see what happens as this cost of intelligence continues to go down but remains high quality. I think 03, I mean, again, we'll have some news next week. It will be even better, even more powerful, and it's just this continued iteration towards getting these like super models. Where does 03 fit? So like right now when I open up OpenAI ChatGPT, I look I'm using 4.5 ChatGPT 4.5 or or 4.0 maybe is what I'm using. I don't know. Uh, what is is 03 the new premier model? Is this the new iPhone 17? Is this new like Frontier model that OpenAI is like? Where does it fit in the stack? It should be uh, this new one will be the new Frontier model. Currently, it's 4.5, which is their like most cutting-edge model. Uh, 4.5 is actually being depreciated. They are

Shutting that down. They are, um, going to do 03 and possibly, um, 4 or 5.0. I think comes next. That's the big one. But this is kind of the step in between 5.0, 0 which is the big one and the current one that we have, which is 4.5. Uh, it's messy. Their naming is messy, but this should be, on paper, the new flagship model that we're getting. And I think all eyes are going to be on whether 03 beats Gemini's 2.5 Flash.

Um, oh yeah, for context here, like Google is leading the model race right now, which is shocking because not too long ago, um, their image generation AI was producing pictures of the forefathers which were of completely different ethnic races to the original. Black Nazis. I remember.

Yeah, exactly. So the fact that they've been able to catch up so quickly is highly commendable to them. Um, I saw someone have a take on uh X which said that if 03 ends up beating uh 2.5 or Gemini 2.5 Flash, then I think that's the incentive for Google to drop their 3.0 Flash, which would then lead to OpenAI uh releasing their uh 5.0. God, there's too many models. Uh, but then but they're like like you know their latest and greatest which will beat them. And he estimates that they, they being OpenAI, only have a six-month lead right now. And remember that was like quoted as like two to three years, um, not too long ago. So pretty crazy.

I am very much looking forward to the point in history that was illustrated in the 27 uh AI 2027 document which illustrated that AI models stop getting released and they just start naturally improving incrementally, like day after day, week after week, and it's no longer, you know, chatb40 or whatever or whatever whatever, it's just there's just one model and it just gets better incrementally because they learn how to like train and release to production at the same time.

Uh, Josh, do do you think there'll be a new kind of model architecture to enable that what David just described, so like the self-learning situation? Um, it seems like it's probably an iteration on the current one, which is just transformer-based architecture, I would imagine, uh, and we're kind of seeing this with Grock where Grock released Grock 3, but Grock 3 got kind of better every single week and it's continuing to get better and that's because of the post-training phase. Um, it requires a lot more compute to fine-tune after the main model's been trained. And I think what we're seeing in the case of Grock, because that's the one example that I have seen, is as it receives more data on a daily basis and as they kind of come up with more algorithmic efficiencies or ways to improve it on the fly, they can do this post-training run fairly quickly and fairly cheaply and just kind of roll it out on top of that base model on a regular basis. So I don't think there's an architecture shift. I'm sure if there was one, it would be a huge unlock, and I'm sure people are trying to work on it. But the current transformer architecture with post-training stacked on top probably is sufficient enough to get to that self-recursive learning where it can kind of improve on a regular basis and push updates live without needing to do the entire base training run again.

Guys, this is I this is so great. I'm learning so much and there's so much to be excited for. This was I think just a great week uh in in AI world and it just continues to to be like this.

Yeah. Yeah, I mean, we're literally in the midst of it. I know. We're about to see another Frontier model drop in a few hours. It's it's pretty insane the rate of progress. Oh, I think it's actually dropping right now. So I think we're going to have to wrap up this episode. It's going to come out there. The OpenAI 3 model, 03 model will already be out by the time people are listening to this. But me, Josh, and you guys are going to drop so we can go go watch that live stream.

Bankless Nation, this has been your weekly AI rollup. Probably the best the best place to keep up with AI. If any listener is listening to a different podcast that it's like this, I want to know because we are going to make this podcast better, but I'm pretty sure this is the best place to keep up with AI. Uh, and that is thanks to my incredible co-hosts here at Jaws and Josh. Josh, thank you for doing this once again this week with me. I appreciate it.

It's been awesome. My pleasure. Yeah, another great week. Uh, Josh is going to be out adventuring in the real world without AI next week. So we might miss it next week. Maybe or maybe me and EJ just see if we can run it without Josh. Uh, Josh, have a great trip, my man.

Thank you. Yeah, by the time I come back, I expect at least three new Frontier models to be released.

Yeah, we we might have replaced you with an AI agent by that time. We'll see how [Laughter] there's no need for it, but thank you.

All right, Bankless Nation, if you if you like this content and you're watching it on YouTube, like and subscribe. Also go ahead and share it with your best AI friend, uh, or your best real friend who likes AI, one of the two. Uh, and then also just stay tuned for next week. We appreciate you watching the the episode with us and we'll see you in a week. [Music]