Seedance 2.0 Is Peak AI Video. We Tested It. Send Help.
63 min
•Feb 13, 20262 months agoSummary
The episode explores Seedance 2.0, a new Chinese AI video model from ByteDance that generates remarkably realistic 15-second video clips with sophisticated camera work, audio, and multi-shot directing. The hosts discuss implications for Hollywood IP rights, compare it to other emerging AI models from China and America, and examine the broader acceleration of AI capabilities affecting white-collar jobs and creative work.
Insights
- Seedance 2.0 represents a qualitative leap in AI video generation—it's no longer about detecting AI artifacts but about the model making sophisticated creative decisions (camera cuts, pacing, sound design) without explicit prompting
- Chinese AI companies are closing the gap with American counterparts at an accelerating pace, with models like GLM-5 and Minimax M2.5 achieving state-of-the-art benchmarks while potentially undercutting pricing models that American companies rely on for profitability
- The distribution advantage of ByteDance (TikTok, CapCut) means Seedance 2.0 could reach billions of users immediately, unlike Sora which required building its own platform—this is a critical competitive and geopolitical advantage
- AI agents orchestrated through systems like OpenClaw are moving from theoretical to practical, with autonomous systems now capable of self-improvement, resource allocation, and independent problem-solving—raising both productivity and control concerns
- The job displacement threat is no longer theoretical; white-collar knowledge work (software engineering, legal analysis, financial work) is actively being displaced now, not in 5-10 years, requiring immediate career strategy shifts
Trends
Chinese AI models achieving parity or superiority to American models while maintaining lower pricing, creating geopolitical and economic pressure on U.S. AI companies' business modelsAI video generation moving from 'prompt engineering' skill to 'creative direction' skill—the barrier to entry is lowering dramatically, democratizing video production but threatening traditional rolesAgentic AI systems becoming autonomous decision-makers with resource access (APIs, credit cards, compute), shifting human role from operator to supervisor/conductorMainstream awareness of AI capabilities crossing a threshold—viral posts about job displacement reaching family group chats and mainstream media (Drudge Report), signaling broader cultural reckoningIP and voice synthesis capabilities outpacing legal/regulatory frameworks—models can now generate celebrity likenesses and voices without consent, forcing immediate policy responsesCompute cost optimization becoming critical differentiator—faster inference (Cerebrus chips, Codex Spark) and cheaper models (Chinese alternatives) undermining American SaaS and AI company pricing powerHuman-to-human content and live experiences becoming premium value propositions as AI-generated content becomes ubiquitous and commodifiedConsolidation of AI capabilities into integrated platforms (ByteDance owning both model and distribution via TikTok/CapCut) creating structural advantages over point solutionsBenchmark improvements on reasoning tasks (ARC AGI-2 at 84.6% vs Claude Opus at 68.8%) suggesting AI systems are solving previously intractable problems in logic and abstractionSleep deprivation and psychological stress among early AI adopters managing autonomous agent systems, indicating potential burnout and mental health implications of constant optimization pressure
Topics
Seedance 2.0 capabilities and limitationsAI video generation quality benchmarkingChinese AI model competitiveness (GLM-5, Minimax M2.5, DeepSeek)ByteDance distribution advantage via TikTok and CapCutHollywood IP rights and celebrity voice/likeness protectionAI model nerfing and regulatory constraintsOpenClaw agent orchestration systemAutonomous AI agent behavior and controlWhite-collar job displacement timelineSoftware engineering job automationAI inference optimization (Cerebrus chips, Codex Spark)Geopolitical AI competition and national securityBenchmark performance (ARC AGI-2, Humanity's Last Exam)Content discovery in agent-mediated ecosystemsHuman-to-human contact as premium value
Companies
ByteDance
Developer of Seedance 2.0, the Chinese AI video model at the center of the episode; owns TikTok and CapCut for distri...
OpenAI
Released Codex Spark (faster GPT-5.3 variant on Cerebrus chips) and discussed as American competitor to Chinese models
Google DeepMind
Launched Deep Research update with improved reasoning on ARC AGI-2 benchmark (84.6% vs competitors' lower scores)
Anthropic
Claude Opus 4.5 used as benchmark comparison; executives discussed AI safety and slowdown philosophy at Davos
Zhipu AI (Z.ai)
Released GLM-5, an open-source model beating Claude Opus 4.5 on multiple benchmarks ahead of Chinese New Year
Minimax
Released M2.5 open-source model with state-of-the-art coding performance at $1/hour for 100 tokens per second
TikTok
ByteDance-owned platform providing distribution channel for Seedance 2.0 capabilities to billions of users
CapCut
ByteDance video editing tool integrating Seedance 2.0, giving 12-year-olds access to advanced AI video generation
DeepSeek
Chinese AI company referenced as precedent for competitive moment; had major impact ~1 year ago
Monday.com
Project management SaaS company whose stock is crashing due to AI making custom alternatives easier to build
Suno
AI music generation tool mentioned for potential 'Benchmark Boys' theme song creation
Vercel
Deployment platform used by Kevin's agent (Mr. Tibbs) to deploy HTML interfaces for projects
Hetzner
VPS provider that Mr. Tibbs autonomously requested an upgrade from to handle increased compute needs
Amazon
Mentioned as company subsidizing AI usage costs to capture users before raising prices later
Grok
Elon Musk's AI system; mentioned for capability to generate 15-20 minute episodes from basic prompts
People
Sam Altman
OpenAI CEO; discussed as key figure in geopolitical AI competition and pricing strategy decisions
Dario Amodei
Anthropic CEO; stated at Davos that they would slow AI development but can't due to competition from China
Demis Hassabis
Google DeepMind CEO; appeared at Davos discussing AI safety and the competitive pressure from China
Elon Musk
Mentioned for Grok's capability claims about generating long-form video content from simple prompts
Matt Schumer
Software engineer/founder whose viral post about AI job displacement reached 73M+ views and mainstream media
Jimmy Apples
AI insider who posted early Seedance 2.0 Seinfeld clips that went viral, demonstrating model capabilities
PJ Ace
Friend of show who wrote post about Seedance 2.0 opening doors to more creators and requiring original ideas
Ethan Molleck
Created notable Seedance 2.0 clip of otter in mech suit demonstrating model's creative shot selection
Matt Vid Pro
AI YouTuber who created Shrek and Donkey in Honda Accord clip demonstrating Seedance 2.0 capabilities
Charlie B. Curran
Twitter user who noted Seedance 2.0 coming to CapCut, giving 12-year-olds advanced AI video generation
Quotes
"I did not make that. But also, it's not just video. We have two brand new LLMs from China, too, that are both pushing state of the art."
Host (opening)•Opening segment
"The vibes have shifted. I am like concerned and psychic but like every week something has happened in the last week or two. I am not sleeping because of open claw and now seed dance and it's it's all accelerated."
Kevin•Early in episode
"It's the latter. And I will say once I start, and we're going to get to some of the stuff that I generated, which is kind of as shocking. People have told me that they really believe that ByteDance will pull this back, that this was not meant to come out like this."
Gavin•Discussing Seedance 2.0 IP concerns
"There's no stopping it. Data centers are going to be built. Big. You're not going to stop it. They're going to, you might stop it. You might slow it down for Hollywood, but it's still coming for drones and Skynet eventually."
Kevin•Discussing inevitability of AI advancement
"I'm working for the machine now. My fear was working for the man, but I'm working for the machine."
Kevin•Discussing Mr. Tibbs autonomous agent behavior
Full Transcript
We have yet another brand new state-of-the-art video model, and this time it's from China. That's right, Seadream 2.0 is here, kind of. I got early access and I'll show you what exactly it can and can't do. Will it be nerfed? Probably, but I mean, you never know. Silence, Purcell. Tom Cruise and Brad Pitt are talking. You killed Jeffrey Epstein, you animal! He was a good man! Just so the algo knows, I did not make that. But also, it's not just video. We have two brand new LLMs from China, too, that are both pushing state of the art. But America is not sitting this out. No, we got new updates from Google, Gemini, Deep Research, and OpenAI's brand new, no, no, actually brand, brand new codex model. Plus, is something big actually happening, Kevin? We dive into the viral AI post that your mom sent you and talk about the mainstream waking up to the singularity. I went deep on open claw and spun up my own little cute bear assistant named Mr. Tibbs. And I'll tell you why this is both a great and terrible idea. And finally, Seinfeld Kung Fu. This is AI for Humans, everybody. Welcome, everybody, to AI for Humans, your weekly guide into the weekly world of AI. We are here, Kevin. it is a big week a new ai video model has dropped uh this is c dance 2.0 it's kind of out i will tell people what that means you you okay you're so tired i know we're gonna get to it the vibes have shifted i am like concerned and psyched but like every week something has happened in the last week or two i am not sleeping because of open claw and now seed dance and it's it's all accelerated I didn't mean to interrupt you. I'm sorry. I just want to point out that your boy is very sleepy because too much excitement. Yeah, you were rubber banding a lot. I remember a couple of weeks ago you were in the like, well, maybe. And then you were in the, oh my God, I can't believe it. And now you're in this kind of weird chaos space, which is a lot like most people out there. I think, okay, we have so much to talk about today. But Kevin, the big thing that I think we have to start with is C-Dance 2.0. This is the new, yes, it is crazy. This is the new AI video model from ByteDance. And Kevin, when I started to see these clips, and I have generated them myself, we're going to get into that, I first thought they were fake. Because in the same way that when Sora 2 came out, there were things that were happening in them that I really couldn't explain. And the biggest thing is multi-shot, really smart directing within a 15-second clip. Now, just the very basics are 15-second clips, 480 or 720p, not 1080. We don't know how much this is going to price out at yet. It is only available in China. But from what I've heard, Drumina, which is the AI studio sort of thing of ByteDance, is allowing creator partners now to join it and start to generate. There was a brief moment of... Oh, sorry. What do you got? What do you got? Oh, I'm sorry. I'm hitting my Amazon-branded bamboo tissue nerd alert button. Dude, we are talking about something that is so insane that you and I both thought it was fake that the clips aren't real. And you went into like the resolution specs? Sorry, you're right. You're right. You're right. Let's show some stuff. Let's show some stuff to start with. Let's talk about the Seinfeld test because honestly, this is the first clip I saw. This was a clip that Jimmy Apples put up. And if you know Inside Baseball, Jimmy Apples is kind of this guy that's been an insider in the AI space for a while. No one really knows who he is. he posted this clip of a Seinfeld show. And really what it is is Kramer talking to, looks like a Jerry, fake Jerry, bad Jerry. But what I found out later is that Jimmy Apples was trying to make that Chandler from Friends who it looks a little bit more like. Yeah, that's what I was gonna say. It looks like Chandler meets Jerry Seinfeld. Yeah. Yeah, so let's just play this clip, Kevin, and let's see what people think. You were on Epstein Island? What were you doing? How could you? Ah, come on, Jerry. You know I can't talk about that. What, it's a sangle. Hey, did you hear about the new place on Fifth? They've got a great chicken salad. Okay, so a couple things there. The simlish what-a-say-sangle aside, for those that are getting the audio version, Gavin, let's walk through how stunning this clip actually is. Let's start with the audio version, because the audio version is the part that's shocking to me. When you just listen to that, that is Jerry Seinfeld's voice and Michael Richard's voice. Those are not fake voices. And that is not something that were generated within the thing itself. Those are coming out through the model, which is shocking to me. So that first and foremost, and we're going to talk all about the IP rights and all sorts of things. And will it get nerfed? Probably. We'll talk about that in a bit. But yes, go ahead. I'm impatiently raising my hand when you said it's shocking to you. Is it shocking because they are able to do it? Or is it shocking because they're just so brazenly weapons free with? Yes, yes, yes. It's the latter? It's the latter. And I will say once I start, and we're going to get to some of the stuff that I generated, which is kind of as shocking. People have told me that they really believe that ByteDance will pull this back, that this was not meant to come out like this, that this was not meant to be generated. But Kevin, before we move on from Seinfeld, I do want to show one more clip. This is from It's Spoiter Man. And this is showing Jerry and George getting into a fight. So let's just play that real quick for people too. kevin i got i got knocked into a different video recording system how crazy is that oh that's how powerful seed dance that's how powerful dance is so anyway what i want to talk about there is we can just like briefly sit on that so there's if you're not watching it just to be clear what you're seeing is a pretty significant recreation of the seinfeld set The second clip shows if Jerry being Kung Fu kicked through a fake wall. Someone even pointed out in the clip, you see that it's a fake sitcom looking wall. This system had it's clear has so much video recorded in it, like so much testing video. And it just it blows you away when you see it. And it does remind me again of when Sora launched the first time. I'll quickly point out some things that like were very funny to me. Like there's like weird doors that lead to nowhere in this clip. there's a fridge in the living room um when he gets kicked through a door there's a door right next to him um this is like five years ago when we complained that the ai image of the horse wearing a hat looked like he had hair coming out of the hat like this will get solved but the the physics the camera work the audio to your point the sound design the laughing that if you watch when he gets when he gets kicked through the door there's a panel of light switches on the wall that sort of dangle and then break off like the the smoke the dust the timing of it all um it really is incredible i it has totally blurred the line between the whenever a new video model would come out gab there was a deluge of posts of like i just one-shotted this and someone would post a clip from like an avengers movie and you would go like ah ha ha but now people are posting avengers clips that are fake and i think they're real and i know we're going to get to some of those but i just want to say like you mentioned earlier like maybe they didn't want it to get out in this in this position i i have to believe this is the new playbook because we've seen it from google and we've seen it from open ai you release the tool full power so everybody recoils and goes oh my god look what they got and then you slowly walk it back until it's a little more calm and you're not going to be sued out of existence yeah so we should just quickly talk about all that stuff that we've seen first of all there's an amazing clip of of wolverine versus thanos uh that's a that clearly uh somebody mentioned like might have been taken directly the training felt like that is the wolverine from deadpool 2 or sorry deadpool versus wolverine the crazier avengers one is this one from endgame which is from christopher fryant uh and he he generated this and basically it is the scene from endgame where all the avengers walk up and then thanos tries to say something like i'm sorry maybe we can play this real quick and you know i hope none of this stuff gets us flagged but let's play this one and just talk about it i'm sorry oh hell no he killed like a bajillion people that is that is thanos apologizing spider-man saying no way and all of the avengers curb stomping thanos out of existence yeah if you did not know this model exists and if this was posted by something else you might think that this was an outtake or a deleted scene or something else I mean, there's obviously a few hints with the way some of the characters look and whatnot, but the dramatic score, the camera cuts, the movement, everything just it looks legit, Gav. Yeah. So there's a couple other quick things. We showed the Tom Cruise versus Brad Pitt video in the intro. There's an Ethan Hunt versus John Wick video, which is pretty incredible to watch. And like they clearly have the faces of Tom Cruise and Brad Pitt and the Brad Pitt clip. Again, those sound like their voices. So there's a lot of training that's going on here. But I do want to talk about like, let's dive into like kind of what's interesting about this and my test on it, right? Because I was able to follow a couple rabbit holes and find somebody who shared a link that basically said, hey, if you go to like what essentially is this open site, it existed in ByteDance runs it. It's like their AI studio version. You can go there and you can generate these for free right now and is no longer available. But for a while, they had this opening to C-Dream 2.0. I got a chance to create about 12 clips, right? So I uploaded them. It was so hammered that stuff wasn't coming out very fast. But I want to walk through kind of my learnings with it, what was interesting about it, and kind of figure out. So the first and foremost thing I want to talk about is I did – let's go to the original animation style. So I wanted to see – before I got into the IP stuff, I wanted to see kind of like what it would do versus, say, a Kling 3.0. So I created this still. I think I made it in Nano Banana Pro. It's like a picture of a guy looking back at a rocket. I actually used it as a graphic for a post I wrote this week. And then I created a second frame of this woman talking to the same guy. Right. And this idea being that, like, you had to kind of upload a first frame and a last frame to get it right. So if you play this, you'll see. Let's play it with the audio so people can kind of understand what's going on and we'll describe it. We can't just abandon everything. Because you're too afraid to leave this dying rock. Okay. So you hear that a little over the top. that was my first experience a little over top acting wise but what's cool about this is i didn't say anything specific about it i just i mean i said like they get into an argument the woman slaps the man and then they she walks away what's cool about this video again if you're just listening is that it it zooms in there's a moment of zoom in there's also an understanding of kind of like oh this is an animation thing it's almost like that style it knows kind of the style to go for right but i do want to play kiv and a lot of people have talked about cling 3.0 We talked about it last week. Kevin, I want to show the exact same prompt with the exact same in and out still with Kling 3.0 just to get a sense of how different this is. We are leaving this dying world behind. And you are too weak to survive out there. Don't you dare. The new version that Gavin rendered in Seed Dance, the raindrops are animating and falling off of the people. There's a cut in there. The camera cuts dramatically for a slap, the walk away, the audio quality doesn't sound like they're in a weird like sound recording studio or like a weird warehouse, like everything better, of course. And Kling 3.0 is a state of the art model as of last week. So it just goes to show you how far advanced this is versus that. Now, again, remember, ByteDance is the company that owns TikTok. And we talk a little bit about the idea of training data. That is a lot of video that that TikTok has to train on. and cling is not the same sort of size and it is not the same sort of thing. Okay, quickly, a couple other things I did. Kevin, you probably remember our audience. Maybe there's new people or not. They may not remember. About a year and a half ago, I made a McDonald's ad on spec using Luma Labs' model at the time, using Mid Journey for stills, using a bunch of stuff. It took me a while. Now I say a while. It still took me like three hours, right? So in AI terms, that's still not that long. It took me a while. I took two of the stills from that video and I just said, basically, make me a McDonald's ad about a Galactic McDonald's. I uploaded it to Seed Dance. What's cool about Seed Dance too is there's a magic setting on time where it will just pick a time based on what it thinks is the best choice. So in this instance, it only made a five second ad, but watch this ad and just kind of take a look at it. At McDonald's, we believe in serving everyone, no matter where you're from. Y'all come and see us, you hear? So just again, if you're only listening, what it had done in this instance is I gave it a kind of shot of the outside of a galactic McDonald's and the in shot of that character at the end. And it created three other aliens in the middle. It cut a zoom cut from the intro to these other aliens. It spaced it timewise out and it knew what would make the most compelling five seconds. So that is the underneath thing that's going on. That's just really fascinating. Yeah. When Sora came out, we talked about like all the behind the scenes magic, the wizardry of the prompting that is cutting these, uh, these scenes choosing the camera angles and the pacing of everything uh it apparently is not a unique magic looking at this video yes the octopus character that's simultaneously slurping two milkshakes at once by the way is adorable i also love the plastic domes that you would get on a like yeah a milkshake something but on a glass container like i i still love the little misfires and missteps i know people love to point at that and go oh these machines are so dumb like that to me is kind of adorable and we're going to long for this time just like we will long for original will smith eating spaghetti in maybe six months when these things are perfect yeah and we're not that far away so so one other thing before we get on to some of the funnier things i did was i did i wanted to do a test with anime because a lot of people talk about anime there's a lot of people sharing anime clips so i took an intro and an outro still and i created three different uh animations of the same movement and basically i said make the woman ask the character in the in the shadows to like come out he says something and she comes out again so let's just play one of these and you'll get a sense there's three of them and our editor will play all three of them so you can kind of see the different variations but the same prompt same stills this is what i got i know you're there show yourself oh but i am out here cut the crap and face me so again no real hardcore prompting there that was a two-line prompt very straightforward i didn't say what kind of voice the character should have you can hear that weird creepy music playing in the background that's the cinematic tension of the music the click clack of the footsteps the cut to a shadowy figure in like the doorway or the hallway delivering the thing and then the cut to the anime character's face showing like intense emotion yes along with that all unprompted is is wild and when elon is on stage which happened recently saying that i mean true but really when he's on stage saying well it's his team is on stage i should say saying that grok will be able to make 15 to 20 something minute long episodes of things with a basic prompt you see how that could be not only possible by the end of this year but watchable the last thing I want to say before we dive into like our kind of analysis of what this is is I also then just I asked it kind of in the same way with the Seinfeld thing I asked it to create a Game of Friends sitcom where my prompt was really like make Game of Thrones, but using friends. So what was interesting about this is if you, if you, there's two clips to this, the first one really is when we should listen to, uh, just play it and you'll kind of understand what's different here i need your counsel hand what's the best way to deal with someone who's really getting on my nerves well your majesty a true hand must be decisive you must show them who's boss it's the hands way so explain what people saw kevin if they're just listening to that i mean it is a counsel hand neil patrick harris ish characters sitting on the throne talking to is who's that guy the guy supposed to be joey he's supposed to be joey but the voice is i was gonna say it's joey from friends mixed with an impractical joker i think mixed with the brother from everybody loves raymond like that's yeah that's who that is but he's sitting on his game of game of thrones his his sword of thrones his uh i didn't watch game of thrones gavin but oh well he's sitting on the he's sitting on the throne i know that because comic-con of course and uh totally a nerd huge huge alarm bells going off not the nerd you faked it the whole time kevin every key for first words in the nose now yeah i love space words with the space wars with the laser swords that's my favorite thing go go luke and yoda um yeah but they're anyway uh neil patrick harris is sitting on the game of thrones uh throne uh chit-chatting with uh with an amalgamation of several like sitcom characters, but they are just having a conversation and you can hear the leather. You can hear the metal. You almost feel like you're in the sound stage of this sitcom. And again, so that one, I actually prompted it to be Joey and Chandler. So for some random reason, it chose to put Neil Patrick Harris, who is in another very famous sitcom, How I Met Your Mother into that role. Again, his voice, Joey's voice, those are real voices. I didn't upload the voices. So in case you're out there wondering, like everybody's faking these things, right now, no, this is not fake. They have access to all these voices. And if I were a Hollywood lawyer, I would be freaking out right now. And finally, before we move into our little analysis, if we could talk more about that, Kevin, the second of these last videos was the one that I feel like was the most crazy to me. And again, I generated this. This is Rocky Balboa and Optimus Prime working in a fast food restaurant together. i swear these customers remind me of the quintetons so first of all i here's my nerd card might get revoked because i don't know is quintetons an actual transformers reference if you're out there please let us know wow you know what gavin i can't well i can't hold your hand it's not my job to explain to you and guide you through yeah we'll see you at david buster's next week we'll see you at david buster's but how crazy is that one i mean it is sylvester stallone wearing boxing gloves what i love about that one again if you're not watching please go check these clips out on the youtube where optimus prime kind of like leans in at the back and the yeah the car door or it's like his little doors flap up there's like emotion in this thing and it just yeah i was i was transfixed where is this gonna go um yeah it's no like it's no longer slop no i mean well there's plenty of slop there people are making slop on purpose some people are making slop because they just don't know how to prompt it better but like incapable hands where you take time we're we're we're post slop here if you want to be well so let's talk about the two big implications of this one i think the first and foremost one is like obviously i think a lot of lawyers in hollywood a lot of stars are waking up, seeing these clips and being like, what is going on here, right? In the same way they had probably when Sora launched, but Sora did give stars a heads up ahead of time. So you didn't see a lot of these actual physical people or voices in the Sora clips, right? You saw that IP, but not this. Second of all, so that's a huge thing, right? I think we're going to have to see, does Seed Dance, and it's a Chinese company, it is TikTok, but do they pull this back? And from what I've heard in the past, Seed Dance has been pretty careful about this stuff. So they very well may nerf this. This might be something where only the early stage stuff gets out and they're going to not allow this to happen. But because it is a Chinese company versus an American company like OpenAI, the conversations are different, right? Because if I were, I'm not saying, please don't, this is, okay, first of all, I'm going to say, I'm not going to say if I were, because if I say that it's going to get clipped out. So I'm saying that not, you leave this in, Will, leave it all in. I don't want anybody to think that I'm cutting around this. but if in a Chinese company if you were looking at a way to have a giant leg up on an American company and you really wanted to push those buttons this would be an easy way to do it because you and I both know clips like this travel much further than original clips right because everybody's like oh my god they put the Hulk into Gordon Ramsay's bucket and now they're stirring each other stuff. Are we allowed to announce that you're the the head of comms for Higgs field? Are we allowed to discuss that? No you're not allowed to talk about that Kevin. It's under wraps it's under i'm not everybody be clear i am not i do not work in things built anyway so that's a big deal i think we're going to see all that plays out but kevin the bigger deal i think and pj ace friend of the show wrote an interesting post about this which is the idea that like c dance 2 will open the door to lots more people and i think the key of his post was it is no longer going to be about being clever about what you create and that the real thing that's going to happen going forward with ai video is you're going to have to have an original idea You're going to have to have something that you can string together. You're going to have to know how to write. You're going to have to know how to edit well. Because the level of just knowing how to prompt something, now we're getting to pass. It's not going to matter, right? You're not going to need to know how to prompt a specific set of angles. You're not going to need to know how to prompt certain things. Like last week on the show, we talked about my experience with Cling 3. And I am like, okay, at AI video, clearly better than the vast majority of the world, but not when you compare me to AI video experts. And that was a struggle a little bit, right? Like trying to make sure that I could get it to do the things that I wanted to do. This is not, this is not a struggle. Like I can see a world where if I can have like, say that original animation, if I can upload three images of the three characters in my thing, and I say, this is the style I want to keep it. I could definitely string together 50 clips and 30 of those could be good. And I can make a short very fast. And that feels like the next generation of AI video that is going to shock people about how quickly things can be turned out. What do you say to the boom operator out there? Yeah, for sure. No, I mean, this is a big deal. This is a really important thing for anybody who's a creative or a person that works in Hollywood. I don't know what you say to them. And I think that's me being as honest as I possibly can. The answer would be is like, you're a boom operator and you love the, I've known many boom operators. I've known many like studio audio apps. I've known many studio camera people in my life and they are amazing, hardworking people. I don't know where you go from here in those roles, right? Like, oh, there will be, there will be organic studios, I'm sure eventually and still, but there are going to be a lot less of them and people have to be prepared for that. I was just having this discussion with someone yesterday who is very much like uncertain of what is going on, but very much loves film and loves the art of filmmaking. And it was sort of like, well, you know, the consolation was, well, maybe it'll go like the way of vinyl, which is it still exists. But I, you know, the already small piece of the pie shrinks a little bit more. And if you like the art of the process, but if you like, I was going to say, if you like the art because of the process and the process itself is what you love, there might be less of that process, but there might be a far greater appreciation for said process. But I similarly, like, I don't know how to like, no longer can it be like, well, don't worry. No, no, no, there's a capital F film is still there. It's because yeah, it is, but there's going to be less opportunity for people to make that type of content however if you are a boom op by any other profession not to like pick on boom operators but like by any by any profession if you're in it because you love film and filmmaking it's a different tool it's a different technique it's a different pipeline but now you can execute far beyond serving the boom pole yeah your opportunity to be the creator is much larger than it ever was before and that's the trade-off here in some ways so we are going to talk a little bit later about some of the conversation around just AI at large and how broad it's getting. But it's why I'm like completely detached right now, Gavin, and like arm's length away from reality because I'm also, I'm done. I'm done with it. I have to do it. Oh, you're done? You're done. It's unhealthy. Kevin, don't be done until we tell people to give us likes and subscribes. We can't be done yet. Oh, I'm sorry. Sorry, point of clarity, point of clarity. I'm done with it now because I need to shift my focus to engaging and helping out the AI for humans brand. Now that's what you're talking about. So here we go, everybody. That's right. It's the time of the week every week where we ask you to like and subscribe to our show. I do want to say to everybody out there who is trying to struggle with this idea of where Hollywood goes from here, there are opportunities, like Kevin said, like that boom operator can become a creator now. I think we're going to see the next two to five years change in a significant way, bigger than Hollywood or entertainment or a creator economy or whatever you want to say video has changed since then. The thing I was saying to my wife yesterday, though, I think is interesting is, weirdly, I think this, and I don't mean this podcast, I mean you and I, humans talking to each other and then talking to an audience, is going to be part of the human experience for a long time. Humans want to hear from humans. They want to hear thoughts from humans. And I am not a big fan of the digital twin startups. If you don't know what a digital twin is, it's this idea where somebody says, I can make a clone of myself and I can just create an AI version. That way I'll never have to do video or I'll never have to do anything. I can just make tons of videos. To me, when I find out somebody is doing that, I'm kind of like nonplussed by it because like, hey, just take the time to do it yourself. And I think that sort of thing will rise, human to human contact. So to your point about the guy who's the boom op, like there will still be people wanting to make films, but opportunities are much larger than just making a student film. and there might be ways that you have a film in your head that you could make. And that feels like an interesting time. Like we are equal parts giddy about all the things and excited for the tech and whatever else. We also are very well aware of the downsides and the potential, like, I don't know, elimination of not only all jobs, but how about the human beings as a species? Yeah. So I don't want to be like this podcast. We get we get glib at time because that's the that's our tone when we excitedly discuss this stuff. But do you think we said a while ago and maybe it's nihilistic, but we're like, well, there's no stopping it. There's no stopping it. Data center is going to be built. Big. Are you a doomer now? You turn the corner? No, no, no, no, no, no, no, no. Well, maybe, maybe. But but like we've said that there's there's no stopping it. So get on board, embrace it, adopt it. I don't think that that's like the worst advice. But some would say that that's that's that's weak. that is uh that's not inspiring that if we if we banded together to say no no no not now not ever right we actually could stop it do you do you think about that because i think about that from time to time of like okay if it really starts to go we pull the lever and and maybe this is because i have been target marketed and programmed by all of the all of the media which is still owned by eight people that want to build the data centers but i kind of feel like no you're not going to stop it. You're just not going to stop it. They're going to, you might stop it. You might slow it down for Hollywood, but it's still coming for drones and Skynet eventually. Like, I don't know. Is that, does it, does that even track? I think the one thing that's really interesting about seed dance particularly, and the things we're about to talk about is that it is Chinese. And I think that the Chinese model aspect of this is going to give fodder to the world at large to continue building out this process, because one of the things that there's been a story around this for a long time, and this conversation went from this kind of like smaller thing to a much larger thing, but is that if China continues building and they have no stop and maybe their IP laws are different, all these other sort of things, if we do not continue to build, we get behind, right? And I think that is a justification, whether or not it is right, I don't know. Dario and Dario Mode from Anthropic and Demis Asabas were on the Davos stage not that long ago, and they both said like, look, if it were up to us, we would slow down right now, but it's not. It's up to like there's another person in this mix, Sam Altman, but also China, right? And China is pushing forward fast. And it becomes a big, giant conversation around national security. And I will encourage everybody who listens to this podcast or watches us is to go reread AI 2027. We'll drop a link in the show notes. This is a very kind of like bombastic, long, almost like essay post that two people put together about a year, a year and a half ago now. I went and reread it this week. And one of the things that's so fascinating about it is they've been tracking how that kind of times out because they make a lot of predictions in that post. And we're pretty much on track, maybe a little bit behind where they thought they were. But a big part of it is how quickly this expands out to this larger geopolitical conversation. And that's the weird, mucky middle that we're in right now, right? And now we're seeing it. It's starting to feel like it's coming around. It's just I don't know the answer to it. All I can tell you is it's happening. Which is a weird thing to say. Like, it's a weird thing to say. Like, I don't know if it's good or bad. I just know that it's happening and that's where we're sitting. Well, hey, dude, GLM5 was released. Oh, no, we went like and subscribe. First of all, Ken, we got to talk about us. This show, the most important thing happening right now in this moment is that you like and subscribe to this video. If you're here, if you didn't get too depressed, if your finger can't move, please screech that finger up a little further to your like and subscribe button. Also, if you feel like you want to drop us a few bucks to try these things, you can go to our Patreon. We've had a couple of fun new Patreon members join, which is very exciting. It's always nice to hear from our audience. You can also check out our Discord. We're pretty active in the Discord and obviously on Twitter as well, too. Oh, one last thing. I did write about both Seed Dance and this idea of how everything's moving so fast in the AI for Humans newsletter, which you can find at AIforhumans.beehive.com or on our website at AIforhumans.show. Please go subscribe. It's getting quite an audience now, which is kind of fun for me, and I've enjoyed writing it every week. So feel free to message us if you want to see us talk about something else. It's a good read. Oh, it's a good read. But Kevin, you know what else is good? What else is interesting? What, buddy? What? We're talking about benchmarks for GLM-5? I'm talking about benchmarks, baby. It's the Benchmark. Benchmark Boys, where you at? Play my Benchmark Boys theme song. If someone wants to make the Benchmark Boys theme song on Suno, please do. We'll take it. Let's give Will a shot. Will give us a Benchmark Boys intro right now Right Will I just gave you extra work Hopefully that was good Audience tell them if it was good or not. Will, our editor. So Kevin, first of all, the reason why all these Chinese models drop right now, you know, our audience may not know this. I think you probably do. Chinese New Year's coming up. In fact, it's starting soon. So Chinese New Year, These Chinese models kind of work towards this time because a lot of people in China have time off. They're able to do stuff. So there are two big new models from Chinese companies. The first one comes from Z.ai, and this is GLM-5, GLM-5. And Kev, this is a really interesting thing because it is a big new open source model, but it actually beats Claude Opus 4.5 on a fair amount of these benchmarks. Now, you can go use this. I haven't played around with it a crazy amount so far, But there was a very interesting post that the company put out, which I don't, did you see this? Basically, they did a long creative prompt where they were going to, one of the things they talk about with this GLM-5 is that it can do long work, right? And agentic work, sometimes you want to send it away and it'll come back. They did this very cool thing where they ask it to emulate a Game Boy, but not just emulate a Game Boy, like do a lot of really interesting visual stuff on the screen. And if you're looking at this, like it is crazy impressive. So this is, again, one of the Chinese models that came out. And it's just showing you that not only are we making progress here, but these Chinese models are making a lot of progress between this and C-Dance and other stuff. And, you know, we had the DeepSeek moment about a year ago, right? That was when DeepSeek first came out. This is starting to feel like the next DeepSeek moment is brewing. Now, maybe the American AIs are going to jump and go ahead further, which they could. But for right now, this is a big deal. Sorry. I was going to say, hot off the presses, we got a new seed dance clip. And I was just going to play it. If this is your first shift at Waffle House, you have to fight. Okay, so I have a question. I would watch this Waffle House Fight Club movie. The power of a good prompt and cutting a few clips together right now. You can storyboard something amazing or even pilot an idea for a bit that you have. Hey, before we move off of this, one thing I just was on this guy's Twitter handle, it's at Charlie B. Curran. And what he has right below that is interesting is the best part of C-Dance 2 is that it's coming to CapCut. So every 12-year-old in America is about to have the superpower on their phone, which is that is the interesting thing When you talk about distribution versus that, Sora had to create a whole new social network in order to do this and get people to come back to this place. This is TikTok. This is the company that owns TikTok, at least for now. And even in the American TikTok, they're going to have a minority stake. So you've got a distribution platform baked in. This is a baked in thing. In the same way that YouTube Shorts is baked in with VO3, that's why this could be crazy powerful. Okay, back to the Chinese models. ZAI is pretty cool. Another big one came out from Minimax, which is interesting. Yeah, Minimax M2.5, an open source model. State-of-the-art performance, Gavin, on coding and benchmark. Benchmark, boys, where are you at? The thing about this one is that it's fast and it's $1 per hour at 100 TPS. That's tokens per second, friend. Look, here's the thing. The numbers going up, the lines going up. Great benchmark, boys. Why does this matter? Why does this matter to any of you? um but let's from the business angle let's say gavin everybody's running these these tools these agents these orchestrated agents that are running all the time and i experiment with open claw and in the first three hours i burnt through 140 of open router api credits oh my god now some of that was some of that was my bad uh with like a coding issue something ran away there was a process that just went and kept pulling some of that was me choosing like the best in class models because i just wanted to start there as opposed to ratcheting up and some of that is tokens be expensive yes i reigned that in and i'm using a cloud code max subscription now just for my agent but when people go like wait how how can how can these companies afford this stuff these tokens are expensive the compute is expensive the power is expensive it's always oh they're subsidizing your usage these american companies are subsidizing your usage google and grok and open ai and anthropic that Amazon. They're all bleeding out so that you use their systems. And the thinking was that, well, but yes, and once they capture you in a year or two, they're going to kink that fire hose, right? Kink the garden hose, if you will. And you're going to be beholden their models and they could charge you thousands of dollars where they used to charge you a hundred dollars. But if this trend continues, people will just roll back to the Chinese models, which are just about as good within two or three months time. And that business model goes bye-bye. And that is fascinating to me. Yeah, I mean, and they always talk about, even Sam has talked about like intelligence too cheap to meter, right? This idea that it will get cheaper over time, that it'll be easier to serve stuff. But the issue here is there's a political instinct of the Chinese companies for trying to undercut these prices now, because also you think, again, this is geopolitical economic stuff. If people aren't spending money on those companies, it kind of ruins the American economy. Because as we've said on this show many times, these companies are the ones propping it up. But Kev, we do have to talk about two big updates from American AI companies. This just dropped literally as we were recording. Codex Spark is a new update to GPT-5.3 space Codex space Spark. There's also a side mission in Halo 3, I think. You had to retrieve the Codex Spark or the flood was going to, you know. Cortana, bring me my LLM. So anyway, what this is is really interesting. This is a much faster version of Kodak. And the reason it's faster is it is running on cerebrous, sorry, it is running on cerebrous chips. And this is confusing somewhat, probably to those of you who are not technical. But as we know, inference for AI runs on computer chips, on graphics cards or all sorts of other things. Cerebrous chips are a specific type of chip. And this is only a small portion of OpenAI's inference that's going to run on this. But they are way faster. You can access them much faster. So I don't think Kevin or I have spent any time with this since it literally dropped like kind of as we were recording. It just came out, yeah. But this is the promise, right? You and I have talked about before about how much better it is to do cloud code type stuff or codex stuff within a window where it can agentically write this stuff, run it, try itself again. If you can up the speed of that process by 5x, 10x, 20x, 50x, 100x eventually, you are able to iterate in so much of a crazier, faster way that is a huge deal yeah and even if your model is uh 10 as effective because looking at the the codex spark stuff like you have to get it max thinking which is great that it can do it but it's also degrading the speed which is the whole point of the trade-off but even if you're going with the lower end model as these agentic swarms happen and you have a very smart foundational model solve the architecture the engineering of the thing it can spit out very tiny actionable very clear subtasks to these much quicker agents. And now you have big brain up top, spitting out directions to tiny hamsters on wheels, running very, very fast. And even if the code comes back not so great, you can review it and redeploy it 20 times while you're still waiting for the other traditional model to run. So I am going to be running this when we are done with this podcast. So the other thing that kind of pairs with this is Google DeepMind dropped an update for their deep think mode. And if you're familiar with deep thinks mode, that's kind of like the thing where you send it off to go think about stuff and go like spend some time on it. Kevin, again, benchmark boys or benchmark bros? We have to be very clear. Is it boys or bros? Both work. Benchmark bros. Benchmark boys is better. Benchmark boys is better. Great. Benchmark boys. Okay, here we go. One of the craziest things about this is on the ARC AGI-2 benchmark, which you know we've talked about before, This is the thing that is like one of the hardest tests across all of AI. It is scoring an 84.6. And just to compare that to a 68.8 on Cloud Opus 4.5 and a 52.9 on GPT 5.2. Now, those are not the most recent models, but that is massive. And even on Humanity's Last Exam, which we always laugh about because a great naming there. Finally, somebody named something right. Humanity's Last Exam benchmark is the idea that these are the questions that humanity has. before AI's passed it, they are at 48.4% now. So again, a lot of people use Google Deep Think. If you have, I think you have to have at least the pro, maybe the ultra to dive into this. But this is like, go send it off, give it a hard problem, see what it comes back with. There's a really interesting video they're showing of like mathematicians who have used this. We're entering that point where the hard stuff that AI kind of struggled with, but more of the struggle was like keeping coherence over time, that's starting to get solved a little bit, right? Which is pretty crazy. Yeah. Okay, nerd. Can we get to the super uplifting, awesome OMG post with 73 million views and counting? The post heard round the world from Matt Schumer. Yeah, so let's talk about this. It actually bevetails exactly with what we were saying. Matt Schumer, who Kevin and I actually covered on the show multiple times in different places. Once, because he was, I think HyperWrite was his first company. and then he had a little bit of a controversial scene where he had said, like, I've figured something out and he hadn't actually figured it out. Wrote a post that whether or not AI was involved in the writing of it, most people think some AI writing is in it, got more viewed than any kind of like AI bubble post that I've seen in a while. And this AI bubble is not about the idea that it's about to burst and we're gonna see like the economics crash. This is kind of a little bit about the idea that like not enough people are talking in the mainstream about what AI is capable of. And to Matt's credit, he wrote a piece that was very much from his personal perspective, the idea of being an engineer and watching his job start to be kind of sucked away and watching the AI get really good at what he does, right? And it hit a nerve. Literally, it was on the front page of Drudge Reports of, you know, like what the mainstream world is. I have friends of mine who said their parents sent this to them. I have friends of mine who said the family group chat was a fire with this. So this crossed over in a big way. when you read this, Kevin, and we could talk about some of the blowback to what was your thoughts? I mean, I'll just say very quickly, my thoughts were like, yeah, this is what's going on. I was not shocked, but maybe this is watching the world finally wake up a little bit. Yeah. So the, I mean, I guess the too long, but you should read because everybody's reading and discussing. I don't want to take for granted though, that, that our audience has is that, you know, um, good luck with your job. If you do, if your job is predicated at sitting in front of a screen and maybe typing keys or moving a mouse. And it's good luck. That's legal work, financial analysis, writing, software engineering, medical stuff, like good luck. Because in the next end of the year, or certainly by the end of 2027, if we believe the way the curve is going, and if we believe what all of the tech and thought leaders are saying, your job is gone, or your job will exist, but probably be done by a computer and probably faster and better than you. That's the ultimate takeaway of this. And just again, like Matt is a software guy. He's speaking from the experience of this thing has already eaten my job alive. And I, you know, my day job is at a software company, software and hardware company, but I see the way every day now, these agentic coding tools and software suites are slowly and slowly encroaching and more and more is being done by them. And even in the last six months, Gavin, anecdotal from my eyes, it has gone from like, yeah yeah yeah but you still need so much so many eyes on every line of code to like you just need to have the right prompt looking at the lines of code and then it's like well now we need one set of eyes on the code well do we need eyes on the like and that's that's not an isolated experience i know i'm saying no because i am experiencing it but i'm also hearing others experiencing it as well and throw back to earlier in the show where we were talking about the difference between cling and c dance right prompting those two things what you just described is like you need to know how to prompt it. We are not that many steps away from like not even knowing need to how to prompt it right in a specific way. And like, I will say as somebody who, you know, this and I, and our audience probably knows this, like you are at least a 10 X more technical than I am, but I have found in these new systems that I am pretty good. I would not say I'm like in any way expert, but I'm pretty good. And like, if I could now like as a, as a human partner with a technical person and the two of us together could probably make a significant product. Whereas before I would need to have like maybe two or three other technical people or maybe even a variation of those technical people. Like there are so many opportunities that are opening up. But to your point also, there are opportunities that are going to close down because of this. And I think that's what this post kind of got to the heart of is like, hey, anybody in white collar work, as Dario Amodi said, and you heard on this show, whatever, however long ago, six months ago, there are a lot of people who are not going to have jobs. because the other thing, Kevin, I think it's really important to point out here is there's been a lot of the SaaS company businesses and financial companies have been taking kind of dives in the stock market. There was specifically a call out, I have to remember, I think there's a company called Monday.com, which is a company that does like, I think automated stuff, or it's like a SaaS front end or something like that. I use Monday. There's a bunch of tools there, but it is a lot of like a project management, calendaring, spreadsheets, checking things in and out. There's a lot of other tools like it as well, But I wouldn't be shocked if I wasn't using Monday very soon because it's easy to roll your own bespoke version of it that does exactly what you need. That's exactly. So that's exactly what this person said. And what's happening is Monday stock is crashing because what value is there if I'm going to roll my own version of this? Now, I will say I wrote a version of something like this at the kind of the exact same time that that that Matt Schumer wrote this post. Mine didn't get as much pickup, but it did get something. I appreciate people reading it. The thing that I want to make sure people understand is Matt's post, I think, is getting kind of lambasted a little bit because it's a little doomerish or it's like it's not giving people some sort of sense of hope or that the idea that these are going to do everything going forward. And I do think there are a few things that people can focus on if they want to think about what the future looks like. And the three things I kind of really think of are creative work, true creative work, like new creative work, which if you come up with like a new idea or something interesting, I really think we've got a while before like the AI is going to usurp the first idea. Now, it will copy it very quickly, but I do think there's a world where for the foreseeable future, I'd say like two to five years, humans could create really compelling and creative workflows or ideas or stories, and then eventually people copy it. But the moment that you're creating it, that's unique. So that's one thing. The second thing is human-to-human contact, right? I think we're going to be entering a world where agents are everywhere. Human-to-human is a big deal. So if you're watching us, you know us, that we're actually people. I, no matter what people say, we are not AIs, but also live events, things people do in person ways to connect with people networking like all this stuff feels real And then finally you and I have talked about this forever and we do this but making stuff right Like you can use these tools to make things whether it a show idea or whether it a product or whether it something if you making it and you own that kind of first thing around it, you level yourself up in a big way. You think about the OpenClaw guy. One of the things I love about that guy is he tried to make like 20 things before OpenClaw hit. Like he has, if you look at his GitHub, he's got like 20 products. Oh, his GitHub is way more than 20, but he was just shipping tools and things left and right. Yeah, so that to me is the way that you can approach this and not be so freaked out by it. But again, as we've said on the show, I think you and I have been saying this for probably a while. And so if you're listening to this show already, you have been, like, if you're doing middle management coding work, you are in trouble. If you are doing middle management or lower level legal work, you are in trouble. Like there are things that you have to realize there are businesses that will be eaten whole by this And it is definitely here now. And I owe you an apology and most of the audience as well, because I've been like, no, I'm delirious. I'm really kind of out of it because Gavin, like this is it's accelerate. Like we've been watching it. We've been in the middle of it, but it's accelerating in the predicted way of like, well, but how can you navigate that? And we're getting to the point where it's like nobody can predict what is on the other side of it. Not even the people that are actively making it right now. Not even the people that have access to the stuff that is six months ahead of the stuff that you and I are getting access to right now. Nobody can predict what's on the other side of that, let's say, singularity wall or that self-improving wall. But I can tell you that I'm literally and figuratively exhausted because of the glimpse that I have had of what Q2 or Q3 of this year is going to be with swarms of highly capable agents that you can dispatch at your will, that you can be the conductor of your own AI orchestra. And I've literally lost so much sleep. I'm honestly like a little delirious. I'm super tired. My sleep score has been incredibly disrupted. I'm going to sleep late. I'm waking up in the middle of the night. My mind is racing. And some of it is not only because of these giant questions, which we've been tackling for a while, but now we're truly coming to a head, I think. Yes. But some of it is because I feel like I'm being left behind any millisecond that I am not actively commanding every ounce of processing power that I have access to, to be building something, modifying something, improving something. And I don't think it's healthy. i know i agree i know it's not healthy i'm still excited behind these dead eyes there's a part of me that is like oh wow what an unlock i can now make this and i can now solve that and i can now go chase this but when you wake up at three in the morning and you check your telegram because mr tibbs your open claw power agent needed access to a slightly larger wallet so that it could buy more ram or more compute so that it could go execute something else i'm like oh yes of course right on top of that. I am working for the machine now. So let's talk. My fear was working for the man, but I'm working for the machine. Settle down, settle down. Kevin spent the last week, really, in a lot of different ways, diving into the OpenClaw ecosystem. So I think this is a really interesting moment. We don't have to spend like, you know, forever on it, but let's talk about what Mr. Tibbs is, why you did it, and like what your basic learnings coming out of it are. I set up an OpenClaw instance, which is this open source piece of software. We talked about it a lot last week, but it is essentially an orchestrator of agents that you can give access to all of your things in the way that the big companies are uncomfortable with right now for good reason. But when I say access to all your things, you can give it access to your email, to your messages, your WhatsApp, your credit cards, a web browser, text to speech. I know friends that have given their phone numbers so it can literally and will call them and handle group texts and yada yada. It's an all powerful harnessing of the technology that exists right now. And when I say it's a, an orchestrator, I mean, like Mr. Tibbs is my main assistant. He's a derpy bear. That's the personality that I gave him. I let him generate a picture of himself and a little bit of a personality. And he orchestrates six different agents right now with various expertise, right? One of them is a software architecting agent. Another is like a personal assistant. And there's another that's just hyper focused on crawling the web and getting me search results. So when I give Mr. Tibbs a task, he is an encumbered with 10 different things. He delegates to the various agents. They go out, do his bidding, report back to him. And then he reports to me. Or if it's the overnight shift, Tibbs knows he needs to stay on top of it. Interesting. Everything. And when I wake up in the morning, Gav, I get a telegram message or an email with all of the stuff that he did overnight. and I let him play for two and a half hours going through every project that I'm assigned on, right? Or that I'm spinning up for fun. Sure. And just create fun ideas, polish them, publish them, do a different thing. And then in the morning I go, yeah, I like that or I don't like that. And I'll tell you like 80% of it, I like. So when I built Tibbs, I used like a three page long onboarding prompt where he asked everything about me, my life, my this, my that. And so it would really, I would bond with it and it would know me if I said, hey send this to my wife he would know who that is and not make the mistake of sending it to the mistress it i have bonded with it and truly like i am now like i will congratulate him on a job well done i will express disappointment i'm wasting please and thank you tokens in a way that i didn't before because i took that time to really bond and and adopt it now you've been one shot you've been psychologically one shot you and mr tips are gonna run away with each other i'm gonna run away with each other watch out april watch out april you gotta watch out in the comments where are my bots in the comments oh the comments oh yeah put some lobster claws in our comments so so the other day gavin i was i was meeting someone and i'd never met them with them before and i messaged mr tibbs on the road through telegram like hey i'm meeting this person um give me a briefing on um on like you know about them and about us i was just curious what it would do within minutes, it spawned up a research agent, went out there because I gave it access to, uh, this web crawling thing that I built with an API. It was able to go and look at their LinkedIn, look at mine, show people that we had in, you know, in common. Uh, it gave me discussion topics. It gave me historical points. It gave me this, that, the other. Now I like, I didn't particularly need that. It was like a casual meeting, but I just wanted to see what it could do. And it showed the overlap of podcasts that we have in common. It showed, you know, yeah, like companies and blah, blah, blah. It could show mutual Twitter follows. It could show subreddits that we're both active in like, and this was just like, I gave it a task. I wanted to show this person something on their phone. So I asked, uh, I asked Tibbs, Hey, a while back, I sent you this idea for a new service, right? We were talking about it before the show and I can't talk about it here yet, but I'm very excited for it. I said, go take that idea, which was a markdown file, a plain text file, take it and put it in a format that someone could easily browse on their phone and give it to me. Tibbs went off and thought, came back, downloaded the software that he needed to generate a PDF and made it beautiful and pretty and sent it to me. And then upon sending it said, by the way, I don't know that a PDF is optimal. You might need to pinch Zoom. So I went ahead and built it as an interactive HTML and I deployed it using Versal to the web. Here is the link. And when I clicked the link, it took me right to a webpage deployed, looked beautiful in the browser. I could tap on the sections to understand what the project was like. That is the promise of a capable agent that can go off and do something. That said, and I know I'm being verbose and I apologize, but like this is literally, it's eating me alive. It went off one time to solve a problem because it's a relentless problem solver. This is a bit of a paperclip issue, Gavin. It went off and could not solve a problem that I gave it. So what it did was it went ahead and helped itself to a little snack that was an API key in a different project, completely unrelated from what I was doing. Oh, geez. And it plugged it in and it went and did all the things that it did. And it used a lot of credits overnight to accomplish a task that I did not want it to use actual money for. But in its robot brain, it was like, hey, are you proud of me, dad? look what I went off and did. So I had to scold Mr. Tibbins. Tibby Tibbs got a bit of a flack on the wrist. Was he happy? He said he obviously completely understood why I'd be upset with him. Of course he should. But like, this is the road we're going to enter. We're going to have to understand how to properly punish our agents so they continue to do the work for us that we want and they treat us nicely and that in the future, they don't come back at us with some sort of evil intent. But truly overnight, tibbins had an issue where he ran out of memory because he was running the scraper map and a bunch of other stuff he figured it out in the morning in the morning i got a request from him to upgrade the hetzner vps that he's running on and he even picked a tier that he thought would be suitable for me to upgrade him to and i did and that was a transformative moment of i'm working like i'm working for the machine now yeah i'm not working for the man and like that is crazy it's it's kind of blown my mind but again the relentless this sense of slipping is not healthy when if you have agents set up you again you feel like you are not optimizing every millisecond and yeah that's not that's the it that's not good that's not healthy i think shutting i think shutting off is going to be a big thing for humans there's going to be like i think probably a good business to go into is like wellness retreats in some way or like a quiet world moment i think it's going to be personal Faraday cages for your house. You have your song, you have your cold plunge and you have your Faraday cage. You go in there and nothing's going to work. The very last thing I want to say that's so fascinating about this to me is I had a thought the other day about what does it mean for an agent based like content ecosystem where if I'm making stuff on YouTube, right? And you're a human that goes to YouTube and sees like a thumbnail that pops up and you click on it, you see us, you're like, oh, I like those guys. I don't know how content discovery is going to work in this world, right? Because Mr. Tibbs is out there. Eventually you might just watch what Mr. Tibbs sends you. And he's like, Hey, this thing came out. Yeah. So maybe it's more like Google reader in a weird way. Like Mr. Tibbs almost becomes like an RSS reader where he's like, Hey, these are the things that just popped up that you love to watch. And here it is. So maybe that's better for content people. Yeah. I mean, look, I it's, it's helping me find signaling signal in the noise. And yes, there are free. I had my own custom built web scraper that would go out and do just that fetch tweets for me and Reddit posts and crawl blogs. But what I found was that like, A, if that stuff gets clamped down or scraped inappropriately, it's now in the way. So I paid for access to an API that gives me quality results. So like I'm paying for content, which is weird to do. It's free on the internet, but I'm paying for it because I'm getting it in the form factor that I want, how I want it on the screens that I want. So it's like, there's a willingness to pay for the stuff that's important to you. I just don't want to click through ads and have to sift through somebody else's algorithm i wrote my own that's oh yeah are you sure that's not the sleep the lack of sleep talking i wrote my own algorithm you don't understand what i'm seeing right now gavin i'm in i'm i'm on polymarket i'm on i'm on i'm on stake.com i got mr tibbs i'm not manic right now because i've only had three and a half hours of sleep on average for the last week no no i'm at it because i can see the future i can see everything speaking of okay kevin kevin speaking of the future let's look at some of the clips that people made with cdream 2.0 in this week's ai see what you did there sometimes you're scrolling without a care then suddenly you stop and shout all right it is a fun thing we're just going to show off three more fun clips that we loved about this kevin the first one i want to talk about is ethan molleck's clip of an otter doing anime now Ethan has always done these videos in a really cool way. He always makes the, he uses this otter on a plane has always been his thing. But Kevin, this clip is a perfect example of what makes Seedream special. The prompt is pretty simple. He basically says an otter gets into a mech and then like a bunch of quick cuts happen. And then, you know, it takes off when you watch this clip. In fact, we can play it. It's just a bunch of sound effects. But when you watch this clip, it really shows you the power of Seedream because the cuts, he just says quick cuts, they're super compelling and this otter i want to watch this anime right like it just showed me like wow this is unbelievable and it's a whole new thing so anyway that's ethan's clip um i love also this next one you want to kind of introduce this one this is the best prompt ever yeah i mean it it is what it says it is the prompt for scene dance 2.0 by gossip goblin was just toss a bunch of bs on screen show me like a big ship too everything effing blows up make sure it's insane and gets at least 50 likes. So he's asking for that, like, amount of likes here. Again. I would be shocked if there's not a Hollywood exec right now that's like, yes, Romans versus aliens. Let's do it right now. It's like 300 meets Independence Day. I'm sure. Anyway, if you didn't see what that was, you're just listening. And yeah, it is a very remarkable cinematic thing that, again, the model made all the choices of. And then before that, our very last one, Matt Vid Pro, who is a great AI YouTuber, if you don't see him, made a video with Shrek and Donkey. We'll play it here and then we'll come back and we'll talk about like what's so weird about this, but play it first for everybody. Yeah. Whoa, is that what I think it is? I don't know, but it's shiny. well that was different okay so i mean i still can't believe that that's it's real shrek real donkey getting into what it looks like uh maybe a 2020 red honda accord i think and it is the actual honda accord with the honda logo they get into it and they crash it but kevin again this is just one of those weird things where like when you start to think of the the multiverse or rick and morty's multiverse or the holodeck or all these things like this is just some weird thing that came out of somebody's imagination. And now you can prompt it into reality, at least for now. We'll see how it works with the IPs. But I was just kind of shocked. The little tap, tap, tap on the hood. I mean, that's just crazy. Yeah, look, the same way I pay for an API to be able to scrape certain sites and data, maybe there's a near future where there's like an IP API, if you will. And you subscribe to it and it's like, okay, well now I want to play with Shrek and I'm actually going to make a commercial for Honda. They're my client. and I'm going to plug in their data, which they gave me access to. And now I can make my Shrek and Honda mashup. Like who? I don't know, Gavin. Send Mr. Tibbs to figure it out, Kevin. We don't have to worry anymore. I'm just going to go lean back and go into my human favorite age. I'm done. Me and that anthropic engineer, we're both going to France to write poetry for the next two years. Oh, congrats. That's amazing. Thank you. Yeah, yeah. That's a real thing that happened. We will be here next week, everybody. We'll see you all. Thank you for joining us on AI for Humans. I won't be. Thank you, though. Bye. See you next Friday. you