Seedream 5.0, Meta Muse & Unreal MCP: Big Week in AI

7:24Addy:you that article uh i have a friend who's successfully making micro drama content with seed dance uh for commercial purposes okay uh we'll come back to that this was another one that

7:35Joey:was running in the background okay it looks a little bit more illustrated yeah yeah i could see it's losing the photorealism yeah like it just doesn't have the grittiness like rockwell painting

7:44Addy:a little bit yeah for sure the nano banana gp2 generations understood that new york in the 70s was like taxi driver the movie or the joker the movie like it has like a high level of contrast a little bit of like dirt and just it's just a gross place and it was and especially in the you know time square in that area the american models understand that and the chinese models don't which

8:12Joey:is interesting yeah it's like a weird cultural nuance cleaner idyllic version of new york in the 70s totally yeah yeah all right next story next story meta is back on the radar with their own new image and video model muse oh i thought you're gonna say they had another round of layoffs

8:29Addy:that too and zuckerberg's tracking the employees that are left uh we'll get to some tracking stuff

8:35Joey:with the story later but okay so they relaunched their own out of their um um super intelligence lab they've got a new image and video model muse that's also uh agent they're calling it agentic image generation so when you kind of give it the prompt it's calling up tool use to figure out what to do like some examples are again really sharp text rendering generated qr code inside an image and based on what you're asking kind of figuring out how to what tools to call up and how to build that out okay i haven't used it yet i mean so far you know it's looking good slash on par with everything that's already out there i think the bigger thing is just like meta's kind of been lacking in the ai race and now they're back on the radar with a pretty decent image model well

9:17Addy:did you catch the headlines recently about how meta is now i'm not sure about leasing out but they certainly have additional data center capability which because they've built so much of it out yeah so fast they have extra and now they're making money off of that which leads me

9:34Joey:to believe using the compute out yeah kind of like an aws play kind of yeah um which leads me

9:39Addy:to believe like their internal ai development and inference has slowed down right because like all that stuff is built on a graph where it's like okay we're gonna be using this much ai in 2026 so we need our data centers to be at this place and so on and it feels like you know you're not hitting that usage and that's why there's extra room so this is antithetical to that this is the opposite of that you're saying that meta is doing ai work they've really seen it would make sense for

10:08Joey:them you know who the hell knows what's in the terms of service but like they have instagram they have like all of you know like youtube google has youtube like they've got facebook instagram tons of tons of images and videos meta as a platform is so tuned into the advertising

10:25Addy:market, right? Like they understand MarTech better than probably any other company, you know? So for them to have a native AI solution that ties into MarTech would make sense. Maybe this is a piece of that puzzle.

10:39Joey:Yeah. Use image and meta products. I mean, that's obviously the obvious play. I mean, the thing I'm thinking of is like, as a model, you know, maybe it's like, okay, whatever. But as a model, that's going to get used by people like this will be the model. Yeah. People don't know it'll be called muse but if like you're using instagram or uh any of the meta ai products or facebook and you're like messing around with uh mess changing images and using ai it's going

11:03Addy:to be this model i i think it's maybe not the consumer facing version of it i think it's the micro brands and the nano brands like the brands with you know like five products that are you know making less than a million dollars in revenue like the mom and pop etsy shops if you will they pay for facebook ads they pay for instagram ads so they will i think this is a way for them to scale their content reach out to more people yeah if you're an advertiser and you're like i want

11:30Joey:more better product shots or more content for but what i'm saying is like the bigger advertisers

11:35Addy:like if you're coca-cola or somebody like you're not using this you have a team of ai people in

11:40Joey:house that are figuring stuff yeah no this is for like the consumer or the yeah you want to spend somebody on meta ads, you don't have the creative spin it up. Same thing that Google's doing.

11:52Addy:Exactly.

11:52Joey:Now, the one thing they have done that has also annoyed people is they will let you modify anyone else's Instagram photos with AI.

12:01Addy:So you could basically remix other people's

12:03Joey:Instagram photos. And this is an opt out model. So basically if everyone's opted in by default, where like anyone could take your Instagram photos and like remix them unless you opt out. Who thought that was a good idea? The people at Meta who want to boost engagement no matter what.

12:22Addy:Yeah.

12:23Joey:Quick update after we recorded this. Meta listened to the online backlash and walked back opting everyone into their AI feature. So now, correction, you are no longer opted in automatically into AI's Muse AI feature on Instagram. So, yeah. So people are upset about this SAG issued a statement. because obviously also this could play into any um actor or any famous person that's on instagram yeah all of their stuff until they opt out is available for like people to use and pull and

12:55Addy:i mean i'm thinking i'm thinking like uh in more harmful ways right like if you're a teenager in high school and you post photos of yourself with your friends if there's somebody that has bad intent bad actor bully if you will they can modify you to do silly things and then the entire school sees it and now it's like a yeah i would hope that there's like

13:18Joey:safeguards in the model itself that prevent you from doing from making images that that that do that i'm sure there are for the extreme thing like if there are for mild things that are like maybe not offensive but just could be embarrassing yeah i don't know yeah i don't know what i mean it could

13:34Addy:be something as innocuous as like um hey this boy is dating this girl and you know you uh nano of banana image that the girl next to the boy and they have nothing to do with each other the next day both of them are embarrassed in school there's no safeguarding against that because technically

13:50Joey:it's not harmful content right yeah i know yeah i mean also look it's not like you could easily screenshot a photo any photo on instagram and like run it through any other ai model totally totally you know if you like even slightly know what you're doing so it's you know i mean sure the opt-in's annoying it makes it easier for people especially especially i'm sure that they're going to push some sort of button engagement to be like hey do you want to like remix this photo yeah so it makes it easier instead of adding that little bit of friction to where like you really

14:17Addy:want to like yeah modify someone's image yeah you got to like download modify upload yeah so this is just like boom right in instagram yeah yeah okay i mean it's interesting uh look i i'm all for experimentation i always say this on the podcast like if tech companies and small creators don't innovate and don't experiment, we won't get anything new. As long as that experimentation will move.

14:43Joey:I mean, I'm just thinking like, what is the feed going to look like? You know, if you're like, everything is some AI remix of everything else.

14:51Addy:The jury's out on that one. So I'll tie it to the recent statistic that something like 20 % of YouTube is already AI,

15:00Joey:not AI slot per se, but AI content. I feel like that's probably under report.

15:05Addy:So within a couple of years of AI video generation, one fifth of YouTube is now AI. What happens two years from now if it's exponential?

15:14Joey:I mean, that also goes to the same, like, sure, the number of stuff out, you know, you could say, oh, it's 20 % AI or 50 % AI, you know, how much is actually watched or consumed? Because there's that same crazy YouTube stats, you know, it's like a thousand hours of videos uploaded every minute or every second. Every second, yeah. And then it's like, yeah, but also like 99 % of videos have like zero views on YouTube. So it's just like, yeah, there has never, it's, it's never been an issue of a trove of unwatchable crap on YouTube. That's been around from the start.

15:47Addy:Yeah. End of the day. I mean, you and I are both YouTube premium subscribers and we have our favorite YouTubers. Like the reason we log in and watch those guys and girls is that we connect with them genuinely. Like we like their personality, their authenticity. and the fact that they're not on a major network or something or a sitcom like yeah or just something

16:05Joey:that's like very specific to like our interest that's like never going to be on another platform exactly yeah i mean i think there's just gonna be such a huge swing the opposite direction and like real authentic content yeah real authentic experiences i can't wait for that take me there now uh you know with this just echo chamber of ai remixed stuff yeah whatever the hell the feed's going to become, it's just going to be a big swing the other way of like, let me touch grass and get something real.

16:32Addy:A hundred percent. And I'm all for like kids nowadays experimenting with like, um, uh, instant cameras and a wired headphones. And now they're going out and buying a CDs. Gosh, I have like a thousand CDs in the garage. DVDs are now like, if you have an original DVD or Blu-ray from back in the day, they're worth more now.

16:53Joey:Yeah. I hope they restart reselling blu-ray players again because i was like trying to find one like a year ago oh i have one and i had one years ago and i was like well i don't even know where it is but i was like let me get another one and then um there's like on even on amazon there's like no one they're really expensive and there's like one manufacturer still making it who makes it still i don't know it was like maybe a panasonic one but it's like 400 500 bucks i was like i don't care that much i mean if

17:16Addy:you adjust for inflation there were probably four or five hundred then right if it was like 200 oh Oh, back then. 20 years ago. Yeah, I guess so. Yeah.

17:23Joey:But, you know, you'd think. Yeah.

17:25Addy:It's like the microwave thing. Did you know, like, microwave technology peaked in the 90s? And now the newer microwaves are inferior to the microwave from the 90s. Why?

17:36Joey:Or is it just good? Like, what was needed again?

17:38Addy:It's just what, that was the microwave boom. And there was competition and innovation between.

17:43Joey:Big air fryer industry. Yes. Shut down the microwave industry.

17:46Addy:So now we look at microwave now as like a budget appliance. It's like, yeah, just put one in there, you know, whatever. Just have one in the house. But back then that was like a status symbol. So then you'd have to have nice things, bells and whistles on it for you to buy it and justify it, right? Yeah.

18:05Joey:No, I mean, there's an argument too, just even for owning Blu-rays, you know, and having the physical media. because like yeah there's all these stories of uh was it just sony that said if you there's like a whole library of movies where even if you purchase these movies they're gonna pull it out of your account because they like lost the uh rights to it which also seems kind of weird yeah so if you paid for this like digital download yeah then it's gonna it's out of your library is it because i

18:30Addy:know about you but i i own a ton of movies on amazon like it's like a good place to buy and have a digital library but i heard i could be wrong there is some issues now where a lot of users are losing their library on amazon i remember that with kindle and stuff because

18:43Joey:there was also even when you buy a kindle book it's like you're buying the license to read this book you're like forever you're not you're not even fresh it's like you're buying a license because they've pulled kindle books in the past that's right other reasons uh and it's like you're buying a license to like read the book you're not buying the actual book itself so for physical media is like the blu-rays you're on the blu-ray it's yours yeah uh also i just missed the special features in the commentary which i have not listened to in years i just don't really put

19:05Addy:them out anymore or produce them and blu-rays better in quality like the compression and the bandwidth of of the blu-ray disc i mean it was breakthrough at the time but you know a 1080p on blu-ray and a 1080p on um you know youtube or netflix yeah it's gonna be a big difference yeah

19:20Joey:yeah i will say netflix streaming quality really good yeah they do out of all the streamers i feel like they've nailed it but they also have a huge image sciencing team behind this but they really

19:29Addy:nailed streaming quality is really good i i will say i'll give them the like the hdr um atmos and all that stuff like that and when you're on a 4k tv and you're streaming and you have the 4k the 4k detects that yeah that format and just automatically goes into hdr mode i'm like oh

19:44Joey:that's neat yeah so we need to get 4k right that is one thing that never made the leap with obviously blu-ray yeah okay enough talking about streaming discs um what else we got oh mcp mcp

19:55Addy:MCP, Model Context Protocol.

19:58Joey:Model Context Protocol, which we were early on talking about this.

20:01Addy:I mean, yeah, it seems like forever ago, but about a year and a half ago, beginning of 2025, MCP was a new idea, I think, from Anthropic.

20:09Joey:Anthropic developed this context. I feel like when we were talking about it, it was like, okay, what's the difference? Even when it was coming out online, people were like, well, how is this different than an API?

20:17Addy:What's the difference?

20:18Joey:And it's messing around with stuff now. It's clear what the difference is. Like API, you need very structured scripting and data in order to call up stuff or say things. MCP, the agent can talk to the tool. It could be like, hey, what tools are available? What can you do? It doesn't have to like run through the API guide or write a script. It can do things a lot more fluidly and in a way that works better with how AI agents work.

20:42Addy:Give our viewers the 10 second elevator pitch of what an MCP is.

20:46Joey:It's a way like the easy way is you can connect your Claude account or your chat GPT account to a third party tool.

20:55Addy:Yeah.

20:55Joey:And from your chat bot be like, oh, hey, go. Let's say for life better, like a Gmail MCP, which they don't officially make. But, oh, hey, where are all the like, what are all my emails that Addy sent me? Like, where did what was the email that Addy said like this thing to me? And instead of having to like write a script and do like a structured API call, it just kind of goes and checks your email. and then reports back, like, oh, hey, this email on, like, April 3rd, like, Addy said this thing, blah, blah, blah.

21:20Addy:So there's just a lot of interrogation between the LLM and the agent, and then they just kind of figure out what's going on.

21:27Joey:Well, then the next level is, like, when your agent, when Claude or ChatGPT can do things via the MCP. So in this case, Unreal Engine launched a beta of an MCP, and you can talk to Claude via Claude and be like, hey, I'm in my Unreal level. I need, like, to build a blueprint that does this thing, and it'll just build the blueprint for you. And then also the other thing we're talking about is Comfy UI, not that has an MCP. Right. And it's like, I need a workflow that does this, this, this, and then it'll build out the workflow for you.

21:54Addy:And from the initial results, both Unreal and Comfy is doing pretty well with the MCP stuff. Like it's pretty legit.

22:02Joey:Yeah, I've been messing with the Comfy one a bit. The one caveat with the Comfy one is it only works with Comfy Cloud. Right now it doesn't work on the desktop version. Right. I mean, I guess in theory, if you had to build a workflow in ComfyCloud, then you could just download that and put it on your computer if you wanted to, if you're trying to just figure out how to build a workflow, but it's only cloud. But that solves a huge gap when you're like, if you're like, what the hell's the case sampler and the other things, and it's just like, okay, I just need to build a workflow that does this.

22:28Joey:And then Claude figures out how to do that.

22:31Addy:Hey, Claude, make it run no matter what. And it just starts to like - Make no mistakes. So the way I understand MCP, I agree with the way you view it. I like your view of like traditional API being very rigid and -

22:47Joey:You had to have very structured data and know what you wanted it to do.

22:50Addy:And you need software to run it, right?

22:53Joey:It's like a Python script usually or something.

22:56Addy:Whereas the MCP is more casual, bro. It's just like, hey, man, it's cool. We'll figure it out.

23:03Joey:Yeah, tell me what you can do. No, tell me. Tell me what you can do.

23:07Addy:What tools are available? So the way I see it is when you start a new chat with ChatGBT or Claude, it has no context. It has no idea what you're about to do. And after three or four back and forth between you and it, it starts to build that context. And that context is essentially just text, right? It could be 1 ,000 characters. It could be 5 million characters. So in order for the LLM to build a context of, first of all, your application, what the application can do, what you're trying to do with it, what assets you have, all of that context is still going to be mostly words and text. So MCP is just a really efficient way for you to get to those million words that are needed for that instance of the LLM running for you at that time.

23:55Addy:That's my understanding of MCP is like you're building all that context in. So then at inference time, this stuff is happening the way you want it to happen there is no such thing as memory in an

24:06Joey:llms i don't think all it has is context window there's memory and like it'll write to a memory

24:14Addy:markdown file it can't yeah i claude code can do that and stuff but like i'm talking about some of the yeah uh even claude itself it has some internal memory stuff but i mean the memory quote memory

24:25Joey:is like it's just writing notes to itself in a markdown file that i can refer to later it's not Yeah.

24:29Addy:The way I think about LLMs is when you open a new chat in GPT or whatever, a million people are opening up a new chats in GPT. So they're getting GPT the exact same way you're getting it. And it's up to you to build it up to an intelligence level where it's useful to you. Right. It's already trained. It's this rigid thing. And all it's doing is inference. And the inference happens to be, you know, your interactivity with it. at inference time if you put in a bunch of information then you are maneuvering it to get the thing that you need but if it has no context has no information then it's going to give you garbage even though it's quote-unquote smart yes i think i'm following what you're saying

25:12Joey:okay you're in your explanation a little abstract i'm trying to follow it along with this yeah so

25:16Addy:like that that the way that the reason it's abstract is because mcp can apply to literally anything it doesn't have to be unreal or comfy no mcp can apply to you and your car another

25:29Joey:interface with the tool because like claude and chaff gpt have computer control and so if i've asked it to you know let's say we wanted to make that image and stuff uh with with c dance um i could say okay i need you to make the image and then it could control chrome go to uh the magnifique yeah yeah browser and then i could like click and go it's gonna be a very slow process it's gonna probably make mistakes because it like has to understand every second like what's happening what's clicking it'll probably click on the wrong thing it'll be a very tedious process or uh i don't know if magnifique has one but like fall has an mcp yeah and so um it will it can just access fall and call it up and then be like oh hey like i need to make this image and then it does the image and brings the image back in a more abstract way because the mcp if i was like oh hey make me an image with the best model that has like that can do 4k and so then that's a little bit more abstract so we can use the tools and call up fall and then be like hey what models are available that can do images at 4k and kind of make more of a decision where it's like a bit more i gave it an abstract

26:35Addy:but it needs to be a two-way street where fall is giving it information it's receiving and then it's telling fall what to do right it's it's always two-way and that that is what mcp is

26:47Joey:it can be it depends how the person set or how the company set up what is available like basically

26:52Addy:what tools are available on the mcp you think on the foul mcp side there is an agent like a chatbot like a llm and then your llm interacts with its llm no it's um because i i i've been messing with

27:04Joey:just building one for our internal tool at mcp access and it's just like another structure of

27:08Addy:the tools to okay so it's like a like a data that it could really is like a read only different way

27:14Joey:for the api like an api ish but like a different way to access it that an ai agent understands

27:19Addy:better okay so it's not that hard i'm sure anybody can build an mcp server i believe you can i mean yeah you ask cloud code to build you one and then it'll do it mold bots yeah uh you know i mean

27:30Joey:another better example is like open claw uh hermes like these agents where like they have a whole

27:35Addy:bunch of mcp access tools hermes hermes the luxury label brand has their own chatbot no well

27:42Joey:there is a chatbot called hermes dang is it french has absolutely nothing to do with the luxury brand company i think it's named after the whatever greek god that is or roman god yeah i was like joey's a baller he has an hermes chatbot no hermes was like an open call alternative okay it's really good because it just solves a lot of the hiccups that open claw had when you're trying to set it up it's just the uh user experience is a lot nicer and it just kind of it works yeah um

28:14Addy:i'm waiting for that cease and desist letter from the brand hermes hasn't happened yet it had like

28:20Joey:you know i connected it to uh you know a bunch of different mcp tools and so when i ask it something uh or like i'm like oh hey like remind me to do this thing tomorrow it's connected to click ups mtp and so then it automatically knows like oh hey i'm going to set up a task on click up for joey on his personal list for this thing even though i didn't say make a click up task for me i just said hey remind me to do this thing okay and then it knows to go hey i have these tools available i'm going to go access the click up tool and go do this thing got it another example of that okay anyways unreal engine yeah and comfy was where we're trying to get it with this mtp no it's it's

28:57Addy:a good move look unreal is a little late to the game but i'm glad they showed up nonetheless

29:01Joey:less welcome unreal yeah uh where could you see this being super helpful in unreal for beginning

29:06Addy:stuff i i think it just like lowers the um uh learning curve and it took me a good three years to get to a level of unreal where i could be a generalist like i could do most things i'm not good but three years is a long time for somebody i think with mcp and you still got to put in the hours you can lower the three years down to maybe three months i would say uh but i'll take it you know on the comfy side now comfy is more of a mystery to me because like i only know what 10 20 of the nodes do the other 80 nodes i'm like okay i don't know what that node does like i see the description of it i don't know what it does yeah i don't know what the dependencies are i

29:47Joey:mean it's really good to you know i feel like you don't have to know these are great for not knowing every single feature and just like having an exploratory chat yeah with it uh too because a lot of times it's like with comfy you could if you know what you want and it just helps you build the thing faster yeah that's one use but then also if you're like i need i'm trying to solve this problem what potential tools are available in comfy yeah it's good at that kind of stuff too of just like having all the data knowledge of like every node and workflow and then hopefully trying

30:17Addy:to get a solution you know google had the right idea when uh do you remember one of their products I think it was called Google AI Studio, something real generic where you can screen share your desktop and it can see where you're clicking, what you're typing, and then it'll instruct you to do what you need to do. Like I think the example was Blender or something.

30:40Joey:This vaguely sounds weird. I remember someone built, I think, their own third-party tool that does something like that. That's sort of like a tutor thing that's watching your screen and then being like, yeah, go click here.

30:50Addy:The reason I bring that up, I was trying to do a really complex operation with Fusion, you know, within DaVinci Resolve. And I was, I had a window open with Claude. I'm like, so how do you do this? And beautiful description. Claude was like, well, you do da-da-da. I'm like, God damn it. Understands everything. So fucking smart. And then, but like, how do you turn that block of text into clickable actions, right? Why can't it just do it for you or show you where to click or suggest on the UI itself what to do?

31:20Joey:It's mostly there. Most of the cloud in chat will have computer control. Will they take over the keyboard mouse? They will take over your computer. Okay. And basically they're just screenshotting it and then clicking and screenshotting and clicking. But you could ask it in that case.

31:36Addy:How do you turn that on?

31:37Joey:Ask it to control my computer and show me where this thing is. Shut it down. Shut it down. It's not fast. It's very slow because it's like, yeah, it's taking an image and processing it and then telling you what's happening. It's a very slow process, but I had a doing that. I was running experiments in Resolve and Resolve is really good scripting, but some things you can't script. And Claude just took over and then was just like clicking the buttons. No, that's the thing.

32:03Addy:I think AI, like the whole point of AI is to make things easier. So you don't need APIs. You don't need, you probably don't even need MCPs at some point where it could just kind of drive the computer for you or whatever interface for you, right? In the same way, Autopilot takes over and drives your Tesla for you, you know?

32:23Joey:Yeah. I mean, eventually just under the hood, you're coordinating with your chatbot and it's calling up the tools and driving the stuff and knows how to communicate with it.

32:32Addy:And the important thing is when you say stop, it stops. When you say stop, it stops. Yeah. Like, yo, yo, yo, back up, back up.

32:38Joey:uh i mean also i mentioned one thing like it's not like apis are dead or useless because mcps you're still using tokens you're still like the ai spinning uh you know it's still costing you tokens and so yeah when you're trying to do like if you had a database and you're like i have structured data you're like i just need to update a bunch of fields an api and a script and stuff that's like predictable repetitive that's still where like scripts and api come in yeah because it's like a script it spins it's on python 100 it doesn't cost anything yeah um you You need, you know, the AI agent could build the script for you and then run it.

33:10Joey:And it's like, that's probably faster, more reliable, cheaper.

33:12Addy:Also, if you're building for users or customers, you need scale, you need reliability, you need stuff that works every time. You're going to need APIs, you know? Yeah.

33:21Joey:Since we're on this topic, last thing I mentioned that came in really handy and also just a really good use case of like cloud code for like non-coding stuff. I've mentioned multiple times it resolves really good at being able to code and access it. We had this project that would keep crashing. it would not open it was like an effect was applied to a clip and then that effect would like just keep killing the project every time and i was trying to go through like resolve support i was like i don't know and then i'm like let me just like ask claude code i'm like hey can you like look at this project yeah see what's happening and claude code spins up it can look at all the logs it's like hey boot up the project and then like let it crash and i'll look at the logs and see what happens yeah and then it basically like just analyzed the whole thing i was like oh it's not the effect that's causing the issue at some point there was a clip that went offline but resolve thinks it's online so it's trying to load the clip but it's like not there and then it's crashing it oh wow and then it's like just go to this clip and then just say like offline the clip yeah and then i did that and then it was fine it worked and it's like this is the stuff where like like editing would have driven some it drove a headache for us for like half a day yeah and then now you know and then he would have to wait for support and send the logs and all this back and forth it's so crazy and then cloud code just looks at the computer it's just like oh yeah yeah

34:32Addy:yeah it's that file the thing that i wonder is uh where was that in the training data how does it even know resolve that well like you swear like cloud data training data people were like no no really look at resolve like that's an important one versus like a million other pieces of software

34:48Joey:well when you vacuum up the internet you know a lot but that's the thing i don't think there's enough content on the internet for it to be that smart it will look at things too like because if i asked even recently i was like oh hey what i need to do something on resolve and it's like well i know up to resolve 20 and resolve on 21 and then i'm like okay we'll go access the web and research us more and then it'll like look at forums and yeah it'll go to the web and look up newer data okay if it's training knowledge is capped real quickly i want to touch on jspace

35:18Addy:Yeah.

35:19Joey:Supposedly.

35:20Addy:Supposedly. This could be investor pump. But again, this has to do with Fable and Anthropic. The researchers took a deep look at the neural network of Fable and they found a hidden layer. You know, imagine you're swimming around, you know, this murky PCB board and there's this chip. There's something like, oh, what's that? They poke it, they prod it. And turns out that is a brain that has developed inside the neural network that wasn't in the original network it's like a cancer cell if you will a mutant thing and agent smith yeah yeah the architect and uh they try okay what happens if i just delete this little nugget and we never delete it again and it's like doing um what what is the thing they used to do lobotomy it's like doing lobotomy on fable and now it can't do anything and then they put it back and now it's fully aware to able to reason so it's called the j space look it up i don't know if it's a controversial

36:24Joey:i mean what was i i i have this safe because anthropic posted a whole paper about this i just hadn't read it yet um but okay thank you so those are the inklings of agi supposedly that's

36:34Addy:the precursor to AGI is that it'll start to develop its own reasoning model within the neural network. Now, it could just be falsified. So the investor is like, yeah, put 10 more billion dollars in there.

36:46Joey:We need that J space to be nice or just a glorified version of still just whatever prediction.

36:53Addy:Yeah. But I thought it was super interesting. Look, regardless of what happens in the image and video world, I think we're going to go down our own track, even if llms never reach agi or asi but if they do oh boy the implications was this the same paper that said

37:07Joey:it's sort of internal thoughts were like a compressed version of english that it was like or it created its own shorthand internal thought that was like less tokens less it was like kind

37:18Addy:of random words i heard something like that i mean you could kind of see that when two llms start to talk to each other it just turns into morse code yeah at some point yeah that was like it was just like a more efficient way to communicate yeah okay all right they're talking behind their backs

37:32Joey:yeah uh oh wait i think i i interrupted and then i didn't let you finish off your friend's

37:38Addy:short company oh yeah um so shout out to jeff at character ai the verge just uh dropped some news about uh micro dramas that character ai is releasing which then you'll be able to interact with so they're experimenting with a whole new type of microdrama where it's um it's more like a game more interactive okay um and my friend jeff he's he's the behind the behind the brains

38:03Joey:so cool all right we'll check it out i'm clear yeah i mean we've talked about microdramas a

38:07Addy:bunch i'm fascinated by the space he's fascinated by it too he's uh he's like look i i didn't understand microdramas before but now that i am beginning to see it it's a very unique format in and of itself. It's something entirely new. And with more interactivity and with real-time AI video generation and all that stuff that's coming, I mean, we could take this to a whole different place, you know? Yeah. Yeah.

38:31Joey:All right. Good place to wrap it up.

38:33Addy:Yeah. Shout out to Paul Trapani for a Spotify comment. Thank you. Always appreciate your support.

38:39Joey:Thank you. Thanks for everything to talk about at denoisepodcast.com. We'll see you in the next episode.