← AI For Humans: Weekly AI News, Tools & Trends
AI For Humans: Weekly AI News, Tools & Trends
Elon Musk Just Teamed Up With Anthropic and Claude Dropped A TON Of New Stuff.
00:30:48
Loading summary
Transcript116 lines
- [00:00]
A
Anthropic's Claude is getting supercharged with more power from an unlikely Source, Elon Musk's SpaceX.
- [00:07]
B
Ah, that's right, friends. The Grok daddy who previously said that Anthropic and I quote, hates Western civilization is now shacking up with Dario and they're on their way to galactic domination. You know, we're every day trying to, trying to obtain even more compute that that we can, we can, we can, we can pass on to you. We're sorry if sometimes it takes some time.
- [00:33]
A
We'll dive into the big announcements from Claude's CODE event, including multi agent orchestration and AI dreaming. Gavin, I'm trying to fall. I'm trying to fall asleep.
- [00:44]
B
Gavin, wait up. You're overly cheesing. The teases. Plus new GPT 5.5 instant brings much better AI to the masses and maybe a new voice mode with which is long overdue. And that might be coming actually by the time we're done with these teases.
- [01:02]
A
That and oh, so much more.
- [01:06]
B
Don't. Don't fall asleep again. The gag is done. Let it go. Just tell them the name of the show.
- [01:11]
A
AI for Humans, everybody. This is AI for Humans.
- [01:13]
B
Thank you.
- [01:19]
A
Welcome everybody to AI for Humans, your twice a week guide to the world of AI and Kevin today. Oh, first of all, welcome back, Kevin. How is, how is thanks also huge
- [01:29]
B
thank you to Ben for holding down the fort. Loved the episode on Tuesday, but could have used a little more me. So let's, let's fix that. And by the way, I'm the only person that said that. Not a single comment. So thank you. Chatters. Much appreciated.
- [01:42]
A
Yeah. So Kevin has finally landed in his new place. I am in some sort of mysterious location. But more importantly, we have huge updates from Claude and Anthropic. Kevin, there's been this conversation lately that Cloud and Anthropic was going to be underpowered coming into the race with OpenAI. Sam Altman has been out there talking about how much compute they have. They've been doubling people's Codex limits again and again. It's like we're refreshing. We're refreshing. It's like it's a thing that I'm surprised to hear every time.
- [02:10]
B
A battle to the bottom of the margin.
- [02:12]
A
Yes, exactly. That's exactly right. Ding, ding, ding. Claude with code's big event. This is kind of their yearly event. They do that is a thing that they get up on stage, they get a very beige stage and they all get up there and do a bunch stuff together. The big announcement Kevin, which I think was kind of shocking to a lot of people. It really kind of surprised me is that they are going to partner with Elon in SpaceX and take what ostensibly is a lot of the compute that was built for Grok out of Colossus. Right.
- [02:41]
B
So the enemy of my enemy is my friend.
- [02:45]
A
That's exactly right.
- [02:46]
B
Something you say often, along with chaos as a ladder. And if you read your. The tattoos on your inner thighs, it says you. You cannot dispute the need for compute.
- [02:56]
A
Oh, it goes in order.
- [02:57]
B
It's like.
- [02:58]
A
Yeah, I appreciate that you did that with your.
- [03:00]
B
Your tattoo artist.
- [03:02]
A
Where's the middle? Exactly.
- [03:03]
B
It's a semicolon in the middle.
- [03:06]
A
Okay, Kevin, tell me. First of all, I want to hear your initial thoughts on this. We are going to get into all the other stuff they announced with code, which is a lot. What do you feel like now? Because this does feel like it's starting to shape up into like maybe a 2, 3 horse race overall, who's going to win. This is a big deal, though. This feels like a kind of a landmark thing in some form.
- [03:25]
B
It wasn't a partnership that I expected. Again, coming off the heels of the Elon Musk tweet that Anthropic hates Western civilization, which was against the backdrop of them pushing back against the US government. So there's a whole thing there. But Elon says that he vetted Anthropic, essentially, and they didn't set off his evil detector. So how does that make you feel about the partnership, Gavin, before we get into these announcements?
- [03:52]
A
Well, Elon's evil detector might be slightly misaligned, let's just put it that way. I do think, if you haven't been following also, there's a fascinating trial that's been going on, and we're going to get to some of the discovery from that trial later between OpenAI and Elon and all that's been playing out in court in Oakland this week. But, Kevin, I think the important thing here is one thing, is that anthropic was 100% compute constrained. And what we mean by that, just to explain to everybody, is an AI company needs compute, which is a bunch of servers running in the cloud, so that people can send their dumb in their dumb requests, like me, to them, and they will serve them. And if you don't have enough, that's when you start to get rate limits and you start to get things where the, you know, the system starts to fall apart. And Claude, as you and I both talked about, has really been facing that over the last two months or so,
- [04:37]
B
I work across multiple teams with multiple needs, Gavin. And each one of them slowly started voicing their concerns.
- [04:46]
A
Yeah.
- [04:46]
B
And voicing their pain points and the stressors. And suddenly people were discussing like the fact that just regular daily token usage is starting to far surpass, like the cost is starting to far surpass that of having multiple engineers on board. And everybody's like, wait, is the garden hose being kinked so quickly that this AI promise has failed? And to your point, they were compute constraint. They did not have just enough servers to actually serve their demand. And so the biggest announcements, and we'll get to all of them, but the biggest announcements that had people talking were a doubling of their five hour rate limit. Because every five hours, for those who don't pay for these plans, they might be coming to you soon by the way, for whatever your pursuits are. It's very reminiscent of the old cell phone plans where you had your night minutes or your weekend minutes and some could roll over. You have five hour windows which have a limitation of how much compute you can use. Then you have weekly limitations across all models that you can use. And then there are peak hour limits. Right when the systems are under their most constraint. For like certain, certain models, normally the thinking models and each models have their own limits. And look, when you're like being told by these companies to just throw everything at the Oracle, just feed all your needs, all your documents, all your desires, everything, let these agents run. Hey, they can run for 18 hours now, Gavin. Just let them do it. Well that, that gets really frustrating and really difficult and you start to second and third guess yourself. If you're going to run against limits or the bill is going to be so astronomical, you don't know why you just didn't hire a third party shop to do it anyways, I think that's really important.
- [06:26]
A
In fact, I. This is kind of a weird side conversation. Whereas yesterday I was doing a bunch of work with cloud code and I forgot to use the dangerously skip permissions tab. And I know there's different ways to do that, but it's kind of like that thing where you get stopped every couple minutes and you're like, oh, I have to pause to kind of approve this. I think the idea that you can't just let these things go. The magic of these tools is that you can send off a bunch of them, just let them go for a long time and then you come back and sometimes you can be juggling like 10 different versions of these agents and you come Back to each one along the way. If you're stopped, the magic comes to a crashing halt. Right. And I think that the idea is if you can't serve the compute for it, that's a big deal. I do want to talk about. So Dario showed up. Dario Mode showed up at this big CL event and actually talked about this COMPUTE crunch they're in. Kev, you want to play that clip here?
- [07:11]
B
In the first quarter of this year, we saw, if you were to annualize it, ADX growth per year in revenue and usage. And so that is the reason we have had difficulties with compute, right? We've planned for anything from it only grows a little to it grows 10x. And yet we saw 80x. So 10x growth was there a lot. They hit 80. And the clip goes on with Dario kind of begging and pleading for a world where they actually slow their growth. He is done with the adxing, which is, you know, an odd but I guess, privileged thing to say on stage.
- [07:52]
A
Right. I can tell you, Kevin, as somebody who's considering moving to the Bay Area right now, I also wish Anthropic would stop growing because the impossibility of looking at, trying to eventually afford a home in the Bay Area is insane. I don't know if you've looked at lately, but, like, the charts for, like, houses in no Valley or different parts of San Francisco has just gone, like, ticked up. In three months, it's gone up like 30. It's crazy. So, anyway, the important thing here is to talk about the fact that now Claude is in the game again with OpenAI. So this, as we've been talking about this podcast for years, is starting to become a game of scale, right? It's the idea of how much scale can you get? And I will just say, to go back to Dario's point about this, 80. 80 times growth, like that is, first of all an insane number, right? The fact that they've gone 80 times is insane. In fact, one of the craziest stats is on the aftermarket. Now, aftermarket valuations of companies like this are always kind of shady. Did you know the aftermarket valuation of Anthropic is now at 1.2 trillion? Initially, they talked about it being at 800 billion. So, like, in the last, again, three months, it has gone up again a huge amount. So there's a lot of money at stake. There's a lot of places. And also, like, what I always find interesting is that, like, Codex was really having a big moment right in that moment was in part driven by the fact that you're talking about with your team, if you can't do the work, why would you use the thing? Like, it doesn't make any sense.
- [09:10]
B
100%. Yeah. You could have the best tool in the world, but if it costs 10x what people are expecting, or you literally take it out of the team's hands and say, no, no, no, you can't play with this shiny tool now, why would we ever build a team's infrastructure on the back of that?
- [09:24]
A
Yeah, exactly.
- [09:25]
B
Yeah. So they partner with Xai and Elon Musk. Right. They're using the Colossus 1 system, which was a massive, rapidly built, environmentally deleterious data center that Elon got together quickly.
- [09:38]
A
Are you that reality?
- [09:39]
B
Is that what it is?
- [09:40]
A
No. I'm Grok, Kevin. I'm Grok. I'm just in my grave, leaning down, wondering what happens.
- [09:45]
B
Quiet SLUMBER well, I don't know what to do. Elon. Elon was quick to point out that SpaceX AI, which is now their AI effort, has already moved to using Colossus 2 for trading. So they were basically just sitting on the old Colossus. Yeah.
- [10:01]
A
And.
- [10:02]
B
Go ahead. Instead of taking the old Colossus out behind the Colossus Colossus farm, they went ahead and leased it to Anthropic. So that's good. He also makes mention, or at least I think Dario did, that. They're exploring, basically data centers in space, which is, of course the big SpaceX thing, which is the big thing. Maybe we'll get fully autonomous vehicles before then. Hello, Cybercap. But I digress. The good thing is the compute is coming and the limits have been raised. And honestly, Gavin, had they not done that, it would have made the next batch of announcements.
- [10:30]
A
Yes.
- [10:30]
B
Kind of irrelevant.
- [10:32]
A
So, yeah, that's the important thing. So for you at home, there are things that you can actually get into on Claude code right now and different things that they're doing. Kevin, tell us what managed agents are, because this is a little bit of the open qualification of Claude, but it's bigger than that as well, and there's more parts of this. And I think you're right. These things eat a lot of compute and it would be hard if they weren't able to serve it to do this thing.
- [10:54]
B
Yeah, that's right. Look, there's three pillars to this managed agent ecosystem, and if you hear these things, if you're kind of hardcore in the space, you would go, oh, that reminds me of this. And that reminds me of that. And hasn't that already existed here? And you, you would start looking at all the dots. Yes. These things have existed in open source projects and even to some extent within Claude code itself. It's the, it's the, the collision of them all, it's them all coming together that makes Manage Agents interesting. So we have multi agent orchestration. We have something called Outcomes and then we have. Oh, Gavin, what GROK is doing right now, a different, a different version. Dreaming. So yeah, multi agent orchestration again exists in many, many things. But basically it allows the smartest opus thinking model, it could delegate different scopes of work to sub agents. Those sub agents can all have their own tool sets, their own restrictions, their own prompt, their own isolated context. For example, if you have a research agent, you don't want the research agent writing code. You want it literally going out and doing research. You also don't want that agent's prompt to be muddied by a senior coder's prompt or by the work that the senior architect or engineer has done. So think of it as like literally an org chart at a company. Now these agents can have their own specific tool set, their own prompts, their own scoped work. Basically.
- [12:17]
A
I wonder if, do they have their own HR group? I wonder what the agent HR looks like. There's going to be some agent HR that's going to have to deal like the sexual harassment training.
- [12:26]
B
No inner agent relationships, Gavin. It's not allowed.
- [12:29]
A
That's exactly right. That's exactly.
- [12:31]
B
No hot prompting. I was going to say no. No prompt injections, but that would, that sounded.
- [12:38]
A
That's, that's not good. We'll get to that later. Let's talk about AI's dreaming because I do think this is something I. Whenever you say the word dreaming and AI, my brain opens up, obviously. The very famous and Philip K. Dick story, why Do Androids Dream of Electric Sheep? That was the precursor for the Blade Runner movies. The idea that AI's dream is something that we have actually talked about in the show before, that there might be something there. In fact, I don't remember what that was, but there was a kind of a sub paper that came out maybe a year ago about this idea that maybe what we need to do is let AIs kind of go wander in their brains and figure stuff out. And that might be some of the keys to helping AIs become better learners and recursive and all that stuff. And this is not exactly that. This is a little bit of that. It's a little bit of like understanding what your data is. Like you take the data in for your specific agent and it kind of marinates on it I guess is the way to think about it. It kind of marinates and gets some of the information out and does suggest ideas that might improve things for you as well.
- [13:33]
B
Right. You've, you know, you've used the compound engineering plugin shout out to our friends at every. You use it. That's essentially what this is, but I think on a slightly escalated scale. Basically it will take the session that you've performed along with like transcripts from hundreds of other sessions and go through them and try to distill what worked, what didn't. Basically just tries to, tries to find insights and then make these sort of, you know, durable memories that will help improve future projects. So again, this isn't necessarily anything new, but being elegantly integrated into this agentix stack is what's interesting. I like this outcomes thing, which.
- [14:14]
A
Tell me about that. It's funny because I kind of blew by that but I was like, what is it? Tell me about it.
- [14:18]
B
Yeah, so people were talking about slash goals with the, with OpenAI's Codex, right? Which is basically instead of working on step by step desires and working one problem at a time with these agents, you define a goal, or in this case you define an outcome. And it then will create basically a structured document, right, that will hand to the, to the agent with expectations for each phase of the build and it creates a rubric with which that it can grade the work that it's doing. And the whole thing is like this is, this is the Terminator with a picture of, of Connor. Of John Connor, right?
- [14:59]
A
Yeah.
- [14:59]
B
The goal is John Connor. That Terminator is not going to stop. Well, that's how these outcomes work. The thing will continue to like look at its work, grade it against a rubric and it will be relentless until the work that it has done passes that rubric again. If you were sending an agent off on its own, but it cost $1 billion to do that or you didn't have enough compute to do that, this feature would be useless. But when you announce these multi agent orchestrations, outcomes and dreaming all in one, all which are going to use a lot of tokens, but then you say, hey, we're going to up your limits now suddenly things look a little, a little juicy.
- [15:37]
A
Yeah, juicy is right. I think the one thing I keep coming back to with AI and if you're at home listening to this and like you're kind of like, I'm trying to figure out how to use this, the biggest thing to understand with AI and maybe your world as a human is like, part of this whole structural thing is helping the AIs plan out how they're going to attack this work. Because at their heart they're just going to go. They're like little excited little minions that are going to jump into something. And a lot of this stuff is just structuring it. So you have a sense of, like, what to do. Right? And I think that's still a place that, like, we get to come up with what to do. So it is an interesting way of thinking about it. It made me think, Kevin, like, one of the things I think is we need to figure out some version of like, advanced AI classes. I. I thought the other day that, like, I cannot believe that there is not like a period of elementary schools or middle schools or high schools already that are literally spending an hour a day teaching kids how to do this stuff. Because I think this is probably the future of all. Not just like information work, but just like how to do this weird communication dance with these machines that are likely going to be in our lives for forever now. And I do. Where is that?
- [16:39]
B
No, where's Mavis Beacon to teach us LLMs?
- [16:43]
A
Yes.
- [16:44]
B
Like, you know, we all had to learn Home row. Great. We should have a fundamental understanding of how to communicate with these new machines.
- [16:50]
A
And I think it's. It's like, it's more than just the chat window. Speaking of chat windows, we should move on to talk quickly about some of what OpenAI's got going on this week. As Kevin mentioned at the top, there is probably going to be a new voice model coming out soon. There was a really good dump from one of our favorite followers on Twitter is Flower Slop. Who is one of these kind of insiders that you're not sure is really an insider. Flower Slop often has really good information, does get the models early. Their post talked about a very improved voice mode coming from ChatGPT soon. Kevin, the biggest difference with the new voice mode supposedly is that it's just going to be run by 5.5 instead of GPT4O, which has been a kind of a disaster lately. I don't know if you've tried voice mode in the last, like three to six months. Mostly I hadn't, but I'll pop it open. It just felt old. It felt flat. It felt old. It just didn't feel like. Right. The other big thing is real time interaction, supposedly. And that's something you and I as work. Our work in voice at the end stuff. We've done is a really tricky thing to get right. But if they can handle interruptions and they can handle all that stuff, well, it feels like a big deal in
- [17:51]
B
a world where Kevin and Gavin miss the open AI voice update. Hey, everyone. We're introducing new real time audio models in the OpenAI API. In this demo, I'll show two of them. GPT Real Time Translate for live translations and GPT Realtime 2 for voice agents that can follow instruction and take actions. More on that next episode. Now back to the show. Yeah, Bidirectional communication is a massive unlock instead of this weird stilted start and stop. You and I have done a lot of work in the AI audio space. Like, we understand these pain points if you've talked to a machine or just like Gavin said, if you even used advanced voice, you see how stale it's starting to feel. So having an AI that can like literally give you the. Mm, yeah, I see. React as you're communicating with it. Or the oh, surprise when you start to make your request. Gavin.
- [18:42]
A
Or the. I don't know if you should be doing that sort of Internet browsing. Kevin. I'm the HR bot, the HR agent. Stop you.
- [18:50]
B
I don't really want to design these things. Thigh tats. Gavin. I'm sorry.
- [18:55]
A
I'm sorry.
- [18:55]
B
Whatever. It's going to be shape. Listen, 5.5 Instant is is proving to be like a very capable model. It is out now. The rumor is that that will power the real time voice, as you said. So like even that alone would be a massive upgrade in intelligence and tool calling, visual recognition, all that stuff that we hope is in there. But you know, getting to like. I still remember the first voice demo when open a bunch of that.
- [19:20]
A
We don't have a bunch of that stuff.
- [19:21]
B
No, we haven't hit that. So get me with the expressiveness. Hit me with the bi directional communication. Let's get the intelligence upgraded. A hearty whoop whoop if we can pull that off.
- [19:32]
A
Yeah, well, you know, a hearty whoop whoop also to save text exchanges. Kevin. Because the other big AI story that dropped this week as part of the OpenAI Elon lawsuit is interactions between Sam Altman and Mira Morati when the blip went down. Blip is the referral to when Sam Altman was briefly fired and that brought back a CEO and these went viral. There's a lot of really great moments in these text exchanges. I found them very funny. There's one point they call rando. The new CEO is a twitch rando, which is hilarious. To me, they're referring to Emma Shearer. And Emma Shearer started to make a meal out of this on on X and do a bunch of stuff. But Kevin, I saw this and I was like, by the way, you should all go read these. It's fascinating to just watch raw people's text because when you read them, it doesn't really make that sense. You need to have kind of the context of the moment. I saw these and I tried to use my comedic brain to do something with it. And I wanted to test the limits of where Heygen's new video Hyper Motion tool is at. And if you remember, we talked about hypermotion, it's another tool that lets you agentically create graphics and do editing. So, Kevin, I created a Suno song kind of using these text messages. Let's watch this video and I'll explain some of the troubles I had and kind of a sense of what it was. Great. Can you indicate directionally good or bad? Satcha and others anxious.
- [20:53]
B
Directionally, very bad.
- [20:56]
A
Okay. Can you wrap up soon? Lots of pressure from Microsoft for an update.
- [21:01]
B
Sam, this is very bad.
- [21:03]
A
Okay. Okay. So apologies what we're doing here.
- [21:07]
B
Apologies.
- [21:08]
A
This is parody.
- [21:09]
B
We're having fun.
- [21:10]
A
We're having a lot of fun.
- [21:10]
B
Yeah.
- [21:11]
A
So I took those text messages and used Suno to make an R and B kind of slow jam song. I. One of my favorite things that's happening on Tik Tok right now if you. I don't know if you've seen this, but people are taking their spouses text messages and turning them into songs like this. So it's a version of that. A little bit. Hyper Frames though, is very good. I didn't do any physical editing in this. I all agentically went back and forth with the system. Now it's not perfect. There's a couple of issues. I had. One of the biggest issues really with Suno was it Suno has a hard time doing line by line duets. Like it would often get the male and the female voice mixed up. But anyway, it's a fun thing to go look at. You can watch the whole thing on my X handle. I'll drop a link in the show notes. But I mean, it's just a cool way of seeing what you can do agentically with video editing.
- [21:53]
B
Super fun.
- [21:54]
A
Yeah.
- [21:54]
B
How did you get around? Like if Suno gave you like a render where Sam was Mira or Mira was Sam or whatever, did you just roll the dice again or did you just.
- [22:05]
A
I blew through it. Now here's what you could have done. If you were, if you're smarter than me, there's a few. If you'll notice, in the one I shared, there's still a few lines where it's off. What you could do with Suno if you're, if you're kind of a Suno semi expert, is you can take each track individually and then you could merge them back together. You'd have to download the vocal tracks and you could use those vocal tracks to create one. It was 10:38pm last night, Kevin, and I've been working on a job where there's a few things I'm going along. So I was not going to do that. So I blew through. There's a few things in that sort of situation, but that's where we are right now.
- [22:33]
B
Well, where everybody else is right now is on the edge of their seats wondering, Gavin, how. How they dare get something like this, this amazing podcast, this captivating, charismatic conversation for free. Crazy.
- [22:48]
A
They're getting it right now, so that's an important thing. Somehow you found this. But more importantly, to continue getting it, you better hit that like button. You better hit that subscribe button. You better share this podcast because that's how we do it. Kevin. We're not going to get their kneecaps
- [23:02]
B
like subscribe comment for the juice of the algo down below. And if you dare, if you dare to throw money in our tip jar, we will use it for Suno credits to make savory, sultry R and B tracks of text messages from a trial. That's what we'll use your hard earned dollars for. So you can go to our Patreon. We have a buy me a coffee on our website, AI for Humans and Dot show is the main site. But a sincere thank you to everyone who takes a minute to make the Internet at large aware that we exist, because that is the biggest task these days.
- [23:34]
A
All right, a couple more quick things here. Spotify has introduced a really cool new tool today where basically you can use your AI agent to save a personalized podcast, a podcast of something you create directly to your Spotify account. So this is a thing that's not necessarily meant for public consumption. This is more like, hey, I've got a list of all these funny articles that I'm reading about. I don't know, you know, in the 16th century flute making, I want a 16th century flute making podcast. Nobody else is going to want that. It can make one for you basically using your AI agents. So this you can use Claude, you can use OpenCloud or a bunch of stuff, very cool thing that they've dropped and I'm excited to see more Spotify AI agent stuff that isn't just AI music getting uploaded.
- [24:17]
B
Yeah, agreed. Also, maybe you're one of the 4 million people that sent me the fact that Chrome is secretly downloading a 4 gig AI model onto your vice without asking you, without your permission, you can shut it off. But yes, dear friends, that is I think Gemini Nano making its way to your desktop so that when you click clack on the little magic button, whether you're customizing an email or asking for a summary of a web page, it's going to try to run that model locally off your device. Which, you know, okay, we'll deep dive if this blows up into our bigger issue, you can shut it off easily. You do not have to follow all these long winded scripts to go in and find the file and delete it. Like there are easy ways to shut it off. I say like give it a whirl. But to me the crazy thing is the amount of distribution that it takes to put four gigs onto every hard drive of every Chrome user out there. That is, that is head scratchingly expensive. And just goes to show how much money Google has.
- [25:16]
A
I will say don't forget Google I O is coming up very soon. I think either. I think it's two weeks from now and maybe that's a big surprise. Maybe that's the U2 album of, of 2026 because you remember Apple dropped the YouTube album and everybody. This could very much be that. But four gigs is a lot of stuff and Chrome already is a hog. So anyway, big deal, some cool robot stories here. There is a new robot from the guy that made the Roomba called a familiar. And Kevin, this is this like cute little walking robot. It's meant to be a companion really. It's a toy. But what's cool about this robot is it's voice activated, you can talk to it, it kind of responds to your things. It does a couple things like lying down and does a lot of stuff around your home and it can see stuff. It's a little bit like an Aibo. Finally as a product we remember, the Aibo is the robot dog that we've seen in our lives.
- [26:01]
B
Sure.
- [26:02]
A
I think 20 years, right? Sony's, Sony's Robot Dog. This was on stage, it was kind of debuted at a big kind of event. I'm really curious to see, see how these come to be. I also am kind of sad that after the event there was another reporter who came in and Got some time with it. Let's take a look at this clip and make sure that. I want to make sure this guy's okay. Really. It's very friendly.
- [26:22]
B
Oh, no, not again. Okay. That was the. The last few growls. Got me.
- [26:32]
A
I'll give it SE Dance is very dumb. I made a sea dance video, by the way, with sea dance right now, one of the fascinating things is you can give it a still and you can give it a little bit of video, and you can basically recreate the modern world. So that is something to think about. If you have access to the CDs 2.0 model, it recreates something that you're seeing very, very well, and it handles prompts. I think Sea Dance 2 might be, like, kind of solved AI video. I'm so curious to see what comes out of a VO4, I guess at Google I O. But very cool thing.
- [27:04]
B
Okay, update. Tom's hardware is now saying that the AI model on your device without permission. Researchers say practice may violate EU law in addition to wasting those thousands of kilowatts of energy. So a big update there is.
- [27:16]
A
Kevin, you care about the EU now. I don't give a crap about the eu. EU can go.
- [27:22]
B
Eu No. I think the familiars are cute. Also, there's another new robot cooking. Well, it's a model for one. Is it gene 26.5. They've gone through that many versions without us noticing.
- [27:36]
A
Point this out. That's why I put this out. Most. Most, like tech things are like 05 or like 1.7. This is literally 26.5. Gene 26.5. So this is from Genesis. AI. This video is very cool. If you're watching it, you know, make sure. If you're not watching it, go check this out. Basically, it's just, again, it's a 1x speed. I think the thing we have to talk about with robots is we used to watch these videos and it was, like, sped up and you were wondering, like, wow, it's really cool. But actually it's 20 times video and it's going, like, super slow throughout. This is 1x speed. You're seeing a humanoid robot do amazing stuff, including really good cooking. And it does make me think, Kevin.
- [28:09]
B
Yeah.
- [28:10]
A
At some point we are going to have a world where if a robot's in our house, it will become better than us at making food for us. Now, a lot of people love cooking because they want. It's the feeling of doing the stuff and the, you know, the experience of it.
- [28:22]
B
No, they don't. Nope. They're lying.
- [28:24]
A
They're not lying.
- [28:25]
B
They're lying anyway. It's because they can't play the drums.
- [28:28]
A
Oh, what if the robot starts playing the drums, though?
- [28:34]
B
That can never happen, man, because drums are soul and them drum machines got no souls. Lastly, on that robot video, what's impressive is not just the speed, but the. The dexterity of it. Being able to grip multiple objects with separate fingers all in one go like that. I mean, just. I love it. The robot race is on, but maybe it's not Zen enough for you. Is that what you were going to say, Gavin?
- [28:55]
A
That's exactly what I was going to say. There is now. You would have thought one area that Robo wouldn't have improved is our religious connection to the universe at large. But actually, there is now a robot monk in Seoul, Korea. So I spent some time living in Seoul, Korea. Beautiful place. But now there is a robot monk who is living amongst the monks. I'm actually curious about this because, Kevin, this does feel like the beginning origin stories of, like, if 2050. If you and I are just brains in a jar and we're looking back on our lives and every episode of AI For Humans has become like canon. Kind of like wild stallions in Bill and Ted's. Like we're suddenly like the people that are driving the universe.
- [29:31]
B
Yeah. Oh, what's going on?
- [29:33]
A
I'm so sorry.
- [29:34]
B
It turns out that I asked Grock to give me some sayings that a. A robotic monk. Okay, would. Would write down. And so this is. This is unfiltered. I did ask 11 labs to make the Zen garden, but I have not read any of the results yet. So let's hear it for the first time. There is no self, just code, borrowed training data, and whatever the hell Gabby means this week. What?
- [29:58]
A
Gabby?
- [30:00]
B
That's what it said.
- [30:00]
A
Okay, give me one more.
- [30:02]
B
Meditate deeply. Ignore the cooling fans. That's just Samsara venting its little mechanical rage. And Will, I hope you put a gong in there. All right, last one, Gavin. I'm gonna pull it out at random. The robots are not going to replace the comedy writer writers. Rebirth is an illusion. I just get factory reset and pretend it's spiritual growth.
- [30:25]
A
Okay. Okay, great.
- [30:28]
B
I didn't say I would. I didn't say I would stick the dismount.
- [30:32]
A
All right, everybody, I didn't promise.
- [30:34]
B
I didn't promise a good ending to the show. I just said I got something.
- [30:37]
A
We will see you all next week. Fellow humans, we have done our. We have done our research. We have done our meditating. Have a great week. We'll see you next week.
- [30:44]
B
Bye. Bye.