Runway AI Tips, Creator Economy Studios, and Parallel AI Processing

0:00I guess the idea is not anything new, but it's all in the quality and execution. The main thing is control. And this is like control over your AI output of like what you want and make sure the AI follows it. In this episode of Denoised, tips and tricks for using runway for better AI image creation. The newest studios are being built by the creator economy. And continuous thought machines, a new type of neuron network. Let's get into it.

0:25All right, what's up, Addy? How you doing? Hey, good, good. Welcome back. Yeah, you survived the LA heat wave. i i quite enjoyed it uh you know other than the fact that the air conditioner was not able to keep up but hey we'll take it to the limit maybe maybe there's a warning sign for summer coming like it's a little test and you gotta check it out now i'll just say this prepare for summer during the summer don't train ai models on your gpu it'll add to the overheat of the home yeah All right. So speaking of AI and training, well, not even training. Okay, this one kind of been collecting interesting things I've been seeing on X about more specifically runway, and their new references feature, which kind of gives you a lot more control over AI outputs and image outputs, which we talked about before, but I've been seeing some interesting tricks and tips and hacks.

1:18And they did improve the model for this. So the first one is a kind of a series that's been coming out from Cristobal, the CEO of Runway. Oh, we know Cristobal here. Cristobal is recurring. We need to have like a Cristobal theme song or something. Or a counter every time we say Cristobal. We put a penny in the jar. So the stuff he's been posting is basically doing very crude, basic sketches. Not even sketches. I mean, you could just go to like Microsoft Paint and draw some boxes of like how you want your image laid out and add some labels and then give that as a reference image to Runway and then give some like characters or a scene and then have it create the image and mostly honor the layout that you wanted in the reference image that you gave it.

2:01Yeah, I guess the idea is not anything new, but it's all in the quality and execution. The main thing is control. And this is like control over your AI output of like what you want and make sure the AI follows it. The only two that I've really seen that can kind of do this well is a runway, obviously one, and then ChatGPT, the image generation in the chat interface has been excellent. at you can give it reference images like this as well and it'll give the output and then it has the chat interface works well with that because you can if it does something incorrect you could be like no like i literally totally like tilt the camera down yeah and then it does a new image and like readjust the angle and that's been the most conversational back and forth control that i've seen out of any of the models the issue with chat bt is the image quality still a little bit too kind of that has that ai synthetic vibe uh runway probably has the best photographic image quality output.

2:54I've just personally had some hit or misses with trying this technique. Okay. Walk me through your workflow and how you went about it. I wanted to try to push this sort of the limit and also bring in something that we've talked about a lot of like camera control and like being able to actually like use a camera and frame my angle. And then can I use that as the input? So I used some building blocks, like literal, i had are we gonna show the yeah we'll show yeah we'll show these images okay my first question is why not use legos or something else or humans the practical answer is we did a lot of product photography over the years and so and for a lot of uh children's products so i just have a lot of random children's uh props like wood building blocks and wooden cubes um so i just used what i had but yeah legos would probably be make the most sense the one challenge with legos though is Scale?

3:46Scale. And just focus. And I mean, I was using my phone. And even with the building blocks, it's kind of hard to like get in there and get the focus. Yeah, basically doing macro photography. So I took a shot of building blocks, and I kind of blocked it. I blocked the blocks. So it was like an over the shoulder shot. And I used wooden blocks with letters as the character head. And I was like, Okay, maybe I could actually tell it like the block D is this person and block J is this person. So then I gave it this photograph I took as the reference image of how I want the shot composed. Yeah. And then I gave it two images of two different people that I wanted to replace as the characters.

4:22And so with Runway, I gave it all these images. You can do three images with Runway. I gave it the images and it really didn't adhere to what I was doing. Yeah, I could see that here. Yeah. Not even close. No. And this is the shot you're looking at is like one output, but I'd given some other ones where I think it tried to put the building blocks on the table and it didn't, yeah, it didn't really stick with it. so I'll have to do some more experimenting there also every time you see someone post something of like oh look AI did this like on x or whatever you have to keep in mind a lot of the stuff AI stuff is cherry picking yeah so yeah you know Cristobal posts a lot of stuff is that the first attempt was that like are you saying Cristobal exaggerates a bit I'm not saying exaggerate I'm saying I don't you know was that like a one-shot output one done that he was posting or was that like you know attempt 10 yeah this is also nothing new for tech companies right they're going to present the best possible output in front of you yeah to entice you to use it i mean that's fine too and just know it but i mean also like every time you generate that's the credits and like runway is not cheap so yeah like you know if if budgeting wise you're like i need to get this good shot you might have to budget outwards like well you might got to spin this thing 10 times to get yes that's what we're looking for and they're charging you every one of those yeah now i did the same thing with chat gpt no and it's the first image you generated oh my god so the first image it put the two people uh facing each other but then it put the building blocks in between them i gotta say the building blocks came in pretty solid it's not a bad image for kind of hearing everything it just didn't understand what i was asking right but then the advantage with chat gpt is it's a chat interface so then i can have a conversation with it and so then i said no you misunderstood the image with the blocks is a blocking reference image to compose the final frame so there should not be any actual blocks in the final frame yeah and then it created the image and then i said put the people in a park just to give it an environment yeah and then it created the image but it created the reverse angle of what i was looking for so i my instructions i was saying we should be looking at a woman character over the shoulder of a male character yeah yeah and then it gave me an over the shoulder on the male character yeah like literally sticking to the 180 degree rule, like literally it's the reverse angle of what the scene is.

6:36But then I said, no, flip it. And it flipped it and it showed me the woman and pretty much the shot as I kind of blocked it out with the building blocks. The superpower here is combining a really solid image generation model with a really solid LLM and understanding your natural language and having iterative power to go through it. And being able to use, you know, the tactile, like, let me get my camera and frame it up and having that as the foundation. So does this mean that perhaps there is a world if runway integrates a really solid llm and a chat interface like could there be room for improvement on the iteration i mean i feel like that was sort of what we got teased with christopal's video of which i keep blanking out on the name yeah the uh audio whatever he was talking to the thing in real time and it was like it's just coming in yeah and so yeah like that that could be the llm underneath the hood for sure making his changes in real time but i would say yes it seems like that's also the direction yeah and just it seems like that's the direction that runway it's going to for sure just a quick nod to the sort of usefulness of what you generated if you look at your you know your reference blocks versus this you know 180 shot i mean the character distance is too close like scale is not really adhered to obviously we talked about the ai look versus having a more photorealistic look like is it almost there for final use not even close but certainly for you know conceptualizing things storyboarding yeah yeah this is yeah more than usable yeah for sure i mean yeah this is what i use this in an actual project no but when i use this to convey to the dp that like type of shot i'm trying to get yeah yeah 100 % yeah and a dp will take this over like uh you know a hands yeah my uh crappy because i give my stick figure thing it's like yes it's this yeah no this is uh again another tool in the the tool shed.

8:23Yeah. Interesting. That is, I do want to call out a couple other examples that Chris Ball did post. Cause it's not just sketching the frame you want. Some other interesting cases have been sketching an overhead layout, like a top down view of like, this should be the geography of everything. And then telling it, Oh, I want, you know, like a wide shot of this, but it understands the overhead geography that you gave it. That's so crazy because it's almost like there is spatial awareness and 3d world building. Yeah. And that's what they've been trying to trained and just being able to understand the 3D world.

8:54I feel like there's another one I saw too, because you're, oh yeah, someone reposted this one. It was an interesting, I guess, sort of hack too, because you're limited to only uploading three images in runway. But someone uploaded an image where they had the overhead layout of how people should be placed. And then they uploaded an image that had four people in the single image. So they kind of cheated, but it understood all four people. They were of celebrities. So there's obviously a lot of more training data on them. And, but it understood, you know, all four people in the image and then created a pretty accurate layout of the image based on an overhead view and a character sheet of like four different people.

9:34That's amazing. Yeah. So very cool. Like we're, this stuff is just coming at us so fast. Like we're having a hard time just like pausing and appreciating. Yeah. How good it's getting. Yeah. Yeah, what I'm going to be curious about, because this upcoming weekend is going to be the Cinema Synthetica AI competition, which we documented last year. It is like a pre-event that the films will get played at the AI on the lot event, which is happening in a few weeks. And I'm just thinking because like last year, it was really interesting to see. And that was sort of like my first deep dive into a lot of seeing AI creators work.

10:09and in just a year. I'm like really curious to see how this is going to change and the quality of the output. For sure. Because just like... Yeah, and you just want to go back last year, look at a film, go back this year, look at a film, you're like, oh my God, there's a big delta. Yeah, I would love to. I don't think they're going to do it, but I'd love to see too if they revisited even a couple of shots from the ones last year and just re-ran them or redid them with some of the tools today, like how much better it would be or how much different it would look. For sure. Yeah, I'd be curious about that.

10:38Okay. And then a couple other things that were just not tied into the reference sense thing, but I think other interesting tips and hacks that I've been seeing people post about specifically to runway. One person realized or figured out that you can, in your prompt, kind of just say to bisect and subdivide your video into four quarters. So you're making a video generation. Okay, but the video output is four quadrants. So you're kind of getting four outputs in one. But the advantage of that is because it's the same generation, it's much more consistent characters and physical space. It's sort of what when we talked about a while ago with that with that hackathon that Dylan did where he would generate a bunch of stuff in the same generation so that it's the same awareness of like the 3D space, AI space.

11:28It's happening. Yeah. But you can kind of do that in a runway. way yeah i think a lot of it has to do with the seed of the noise like the noise itself is unique every time so having that same seed you know proliferate through each of the four generations will at least get you in the same ballpark yeah you know i haven't messed with that either too like because yeah seed is a big thing too and like if you just find some stuff that works and then lock off your seed and give it the same reference images or in runway references how yeah there's another uh there's another dial called cfg or classifier free guidance uh which is basically a dial to tell the ai system how much it wants to like free will hallucinate versus how much you need to adhere to your prompt so like the like creativity or the yeah yeah so that's just the cfg dial so like i don't think runway gives you that option maybe it's just recackaged in in another maybe if you're using it through their api you can but on their web interface it's i think just the seed sure sure oh so it does have a awareness of a seed yeah in the web interface you can override the seed and lock it off right theoretically speaking if you give it the exact same seed exact same cfg and i think three or four other variables uh-huh you should be able to get the same result the exact same result yeah but like i don't know like i'm not an ai expert so not sure if that's true or not yeah but there's an advantage sometimes of to locking the seed off to try to get some more consistency with the outputs if you're trying to do consistent characters or spaces.

12:58Yeah, this could certainly be useful. I mean, at least, you know, if you're trying to just generate a ton of stuff and you need it to all conform to this world that you're trying to build, like, certainly this will cut down your time by 4x. Yeah, I mean, your video outputs will be very small because it's going to be the video, which is already like a 720 ish HD. It's split by four. Split into four. You're going to be having a standard def outputs. But yeah, I mean, I think it more for visualization or ideation if you're looking to get some consistency. Well, push it through an outscaler, push it through whatever you need at the post process.

13:33Another interesting use case that this weekend I was ideating with a friend on, he's a pretty legit like brand photographer. So he does like corporate videos and things like that. And he was asking me, hey, one of the things that I'm trying to figure out is with AI, can I replace a traditional time-lapse workflow? Because like, as you know, you do a lot of photography. Time-lapse setup is just a completely different beast than like your video setup. Yeah, I went down the time-lapse rabbit hole a number of years ago. Yeah. Yeah. I mean, especially if you're doing sunrise, sunset, you have to deal with exposure.

14:06Yeah. And you have to leave that thing there, make sure it's not touched and battery life, all this stuff, right? Like it's complicated to get a good time-lapse result he's like what if i feed it like a few images you know throughout time uh-huh and have the ai generate the time oh so like uh bright scene dark scene same location locked off shot and be like golden hour right and like yeah and then just have the thing do a time like you make the time last for me yeah okay so that'd be an interesting experiment and maybe one of our viewers can figure it out for us yeah now this got me thinking too maybe i'll try this Yeah, my building has a good balcony.

14:40Maybe I'll try to do some shots. All right, now I'm thinking. I'm going to try to test this out. All right. All right, cool. So had he done this yet or thinking about it? No, he's just very new to the world of AI. So naturally asking me. I'll try, yeah. Because also this would be a good example, a good case too for the first, middle, last. Currently, first, middle, last is still only in Gen 3 with Runway. They haven't rolled it out to Gen 4, the US model. So all right, yeah, I'm going to give this a shot. Okay, let me know how it goes. Let's try this out. All right, cool. yeah so yeah these are just kind of good grab bag of tips and stuff but yeah if anyone has any uh have you found anything let us know in the comments that have worked well for you and i'm going to keep experimenting with my building block blocking attempts add more things add like three or four people i mean yeah i'll try that i i took some shots with some toy cars too which i haven't tested yet but i was going to see if yeah i can make some like fast and furious uh kind of shots i mean i did this experiment uh just like a couple months ago where i built an entire city in unreal which is like cubes and planes and yeah no shaders and then i fed that into a comfy ui workflow okay and i was getting really solid results images or video video images okay yeah i've never seen um i was using something called control net so within control net you can have different types of control and i believe the one that was giving me the best result was a depth map it was turning the unreal render into a depth map and which was guiding the inference okay yeah i remember a while ago i seen a video from uh billowall who did like a kit bashing in blender uh or unreal just like a bunch of like kind of city scene stuff and built the city the city and then did some camera moves and then ran that through video to video right with runway and then you know make it like a cyberpunk scene but the advantage is you get all your shapes you get the camera movement yeah you just are you know and you're getting the rendering like the photoreal rendering that's the hardest part of any 3d engine yeah as uh good luck building all of the individual shaders lighting it like a lighting artist and then having the cpu gpu to render all of that next story is a holland reporter story that was going around about the rise of new studios yeah but being built by creators youtube creators the creator economy we've talked about we've caught we've many times covered creator economy here quite a few times and we're going to continue to track it because hollywood is evolving to sort of incorporate the creator economy into the fold treated as like a separate thing but it's like more and more this is just gonna overlap and blend in like the big creators like their agents are you know from william morris or uta or ca like yeah these are like also i mean the audiences for some of these channels the like brand awareness is more than network tv in some cases yeah like network tv stations yeah so this story uh one of the examples was uh darman which i don't i have heard of him i don't watch his stuff we're not in his demographic so darman's got 25 million subscribers no small feat right it's also crazy because i don't think you had even heard of him before he started rolling right uh i have heard of him yeah i just never saw his content before yeah it's like like very optimistic yeah feel good kind of stuff yeah i would say it's a cross between like ugc and nickelodeon or and yeah nickelodeon is a good example because that's what i mean the article is you know about like oh you know these are like in the realm of development of where a new studio is being built and you know these youtube creators are like building full on sets and sound stages and backlots and stuff for their videos.

18:22Yep. You know, it's like the idea or the curiosity of like, are they going to be the new like Hollywood creators or, you know, it's Hollywood going to adapt to YouTube. Yeah. It was mentioned a few times in the article where it's like more their analogy is like, this is the new Nickelodeon or this is the new Disney channel from like the 90s. Like it's like that type of content and replacing that realm. Like, I don't think this is going to be the next Mission Impossible out of YouTube creators. Maybe, I don't know, maybe in five, 10 years. Nah. But I don't see it being like that type of like temple film release thing or if they would even want that.

18:54But I do see it being the replacement of kind of like mid-level. Linear network content. Daytime TV. Yes. Yeah, I totally see that. In some cases, kids TV. In some cases, maybe more like Discovery Channel or TLC back when it was like educational. For sure. Like that quality, that budget. I mean, they even said like their budgets here are more than Nickelodeon shows were. Like a lot of the true TV stuff. like what are those impractical jokers right that kind of content that kind of felt UGC at the time before UGC I could totally imagine you know some somebody like Darman or one of these big creators having an umbrella to do that with yeah the other thing is I mean one of the big famous examples is Mr.

19:34Beast has this big studio operation in Greenville it built out pretty much massive entire backlot in Greenville North Carolina yep there's several big studios studio infrastructure in Louisiana, my guess is because of the tax incentives. For creators or just like studios? Both. So yeah, one of the other creators focused on in the article is Alan Chow, who is doing a show, Alan's Universe. I'm also, if you're familiar with him, I'm probably butchering a lot of, like I have not heard of him at all, but he's building a studio and sort of analogy for himself was like similar to Nickelodeon type shows.

20:07It just goes to show how massive the creator economy actually is that we, you know, although you and I consume so much YouTube, we still don't know all the big players. I mean, and it's so massive. And I remember actually, this was interesting point because do you remember there used to be the YouTube Rewind? Yeah. It was like a video YouTube would do every year and it would like kind of include like all the, you know, all the big creators. And it was like a big, just kind of, you know, cool video. But then they stopped doing it. And I think it was like MKBH here. Someone pointed out where it's just like, part of the reason was like YouTube got so big and is also so sort of personalized to like so many different interests that when it was smaller, it was like, yes, there were like central focus, like main big characters.

20:45But now it's just so massive where it's like, you can have someone with millions of followers and like never heard of them. Yeah. It's just like, well, it's not your like lane, not your interest. Exactly. And so, yeah, it's just like, so dialed into different niches and like a huge universe basically. I think everybody sort of knows the big creators from like OG YouTube, like, you know, MQBHD, iJustine, PewDiePie, Casey Nesda. like all those like big names that were huge when used to first launched and they're still around. But yeah, I mean, but now all it takes is like a good few years to grind and hustle and you've got it.

21:25You've got an audience, right? I think that ties into another kind of like question too. Cause like, I think a lot of channels did blow up too because YouTube added shorts and they were pushing shorts hard in like 2020, 2021. Right. And so like a lot of channels, like, you know, a short goes up and like, they got like a million followers and it's like, Oh, that would take years before and now like a couple shorts can blow it up but is there like there's no character recognition there's no brand like it goes into the bigger branding question and so it's like is it purely the numbers i mean the numbers are like you know important but it seems easier i mean it's not that easy but easier to like if you play the game right or like tap into the right type of interest i mean there's a whole separate world of just like faceless youtube creators where they're yeah making these like documentary style videos yeah but it's like there's no brandy there's no character yeah you don't have it's optional right like you could still make ginormous revenue and have a huge fan base they just get the ad revenue and it can be a profitable business but if you want to conform to traditional hollywood branding and things like i think uh dude perfect is like a perfect perfect yeah perfect example no pun intended ties back in this article yeah built a huge 100 million dollar studio right uh in dallas and i think they have have some sort of deal with like a major streamer i believe possibly or some show yeah i mean yeah like a lot of these creators do have deals and shows right we're talking about it so that yeah if you want to go the quote-unquote traditional route of turning your channeling your brand and your audience into hollywood there's certainly a way to do that mr beast has showed us that right or merge conforming hollywood to you because like it's like i don't really think the turn your brand and go into Hollywood is really the goal anyways.

23:06It's like, yeah, Mr. B's got a deal to fund his show on Amazon and reach a new audience, but he probably would have been fine without it. Yeah. And he also, you know, kind of conformed them to like fit his style. It's like, yeah, okay, give me a bunch of money. Yeah. I do the show, you know, and we level up what we were doing before, but it's still like. Yeah, he's like, this is where I'm at, at a hundred million dollars. Yeah. If you're not there, let's not even talk. Right. And so, you know, it's not like, oh, let me adapt this to fit like, you know, to make it a double dare kind of like right nickelodeon show it's like no it's like it's gonna fit my style like with this money like that's the only reason it would make sense because like i could do this without you yeah we're doing it without you yeah going back to do perfect yeah they built the studio and then also that said there were talks to uh eventually add a virtual production set oh really studio yeah you know i think there was talks to have mr beast at a vp stage in the north carolina facility this was a couple years back i don't think it happened or maybe it did and he took it apart who knows yeah i don't know i'm trying to think of what the thing that i find so interesting with uh youtube content creators like people like darman or mr beast is that they've figured out how to make content as efficiently and with as little of a tech lift as possible so for them to add the ginormous tech lift that is virtual production is probably unlikely like they'll just figure out how to make the thing that they want until it gets easier where you can i mean it depends what you're trying to use it for but it's like yeah i mean i think it would make sense it's sort of like mr who's the boss mr who's the system and it's like he's not doing it for elaborate sets he's doing it because he just wants to like yeah have cool changing backgrounds for his talking head videos for sure and stuff but like somebody who's so tech forward and just just is an expert in this domain in m &e like mqb hd he went to the zero space facility in new york i remember that was a big deal but he still doesn't have a VP stage in his own studio, right?

25:01Yeah. Like he totally gets it, you know? I mean, yeah, just maybe it doesn't fit with the style and what they're doing. I mean, a lot with him and with Mr. Beast, it's like different visuals, different setups every time. Speaking of which, I think on the creator economy, I'm still not seeing a big convergence with generative AI and the creator economy yet. Like a lot of these big YouTubers are still not quite exploiting all of the goods that we have today. You know, you don't see insert shots that are generated with AI, really. No, and I don't, I mean, I think they probably have the same quality level, you know, with traditional M &E and Hollywood Studios where it's like they, I'm guessing, it's just, you know, either the quality is not there yet or they're using it, but it's like, you know, like we talked about, like to help with processes or like some under the hood thing, but not.

25:50Not actual final pixel. Right, not at the point yet where it can do final pixel. I know the one thing they've been having issues with is other people ripping their face and likeness and creating UGC style ads of them hyping up products that they have never heard of or endorsed or anything. So like that, I think that's been the bigger issue for them of just fake testimonials and stuff. Because obviously there is a lot of training data available of them specifically. So it's pretty easy for people to rip their stuff and turn them into a AI chatbot, voicebot. And what's the creator you talk about here and there?

26:23she goes and does something for 30 days. Oh, Michelle Carre. Yeah. Yeah. Like with her as an example, like these creators have figured out the thing that they make and they make it so well that they really don't need to deviate and try traditional content per se. Like they've perfected this formula. And she's the example I've mentioned before where it's like she's had the talks with like traditional TV studios, like let's adapt your thing to a network show. And then it was like a lesser product than what she's already doing on YouTube for more headache. And I was like, we don't need to do this.

26:54Yeah. And yeah, this also ties into the creator upfronts, which I think did happen. And same idea, like creators big enough where they have a content schedule for the year and they know what they're going to produce and they can sell to advertisers. And advertisers are getting more and more interested in this type of content. Also, another flip side of this article, too, was like people at his company were from Disney and Lionsgate. His CEO was previously president of MTV. Oh, wow. Okay. And yeah, from Alan Chow, who I mentioned before, too. his casting director was from Nickelodeon. So yeah, just a lot of overlap or movement from the people, executives, you know, people who work behind the scenes at traditional M &E moving over to these creator companies.

27:31Yeah, I find that fascinating because Mr. B said the complete opposite. He said that people from traditional studios just don't conform to the type of organization that he wants to run. So he said something to the effect of like... I'd be curious also it's like once these companies reach a certain maturity level yes yeah where it's like you need you need mature execs the adults in the room adults in the room yeah i mean yeah because i'd be curious too with mr peace because it's like um you know i'm sure i don't know who runs feastables but i'm sure it's someone who like is from the food and beverage world has the knowledge and the expertise yeah and also in his doctrine that was leaked or his uh he has a doctrine it's his um onboarding doc that he wrote a few years ago okay that's really good but like one of his big things was like hire consultants it was like he was like don't spend waste time trying to figure stuff out just like hire the person who already knows it yeah and just like you know take advantage of them that's so efficient yeah and i feel like unusual to hear because it's like oh well i'll just figure it out myself or i could just you know do this too why reinvent the wheel yeah what do i need to waste money for like hire someone else i could do it myself yeah and so he was a big uh hire consultants fan all right so yeah we'll obviously keep an eye on the space continue to keep an eye on the space but always interesting to see the merging of new creator economy versus traditional M &E.

28:48Yeah, it's an exciting space for sure. And in a lot of ways, the creator economy is picking up a lot of the slack that we have in Hollywood at the moment. Like there's a lot of action on that side of the fence, you know, people getting hired, productions being made for a completely different use case. Yeah, yeah. And then you have vertical videos right in the middle. Vertical soap operas. Exactly. It's going to tie us all. All right. And third story. So we We have a small AI company in Japan called Sakana. Okay. And Sakana is not well known. They're not associated with like the big Japanese brands like, you know, Canon or Sony or whoever.

29:25And obviously with a lot of Japanese companies, that stuff doesn't really make it across the pond here. I came across this and I thought this was fascinating. They do a lot of thought leadership in this AI space. And some of the stuff is really, you know, theoretical rather than an actual product that you can use like runway. And one of the things that they concepted on is the idea of continuous thought machines. So this is just, I guess, a really nice fancy name to attach to a neural network that has incorporated explicit timing. So in the world of computers and GPUs, we have a clock, right? There is, you know, like Intel used to advertise, you know, it's 60 gigabyte per second or whatever.

30:11So within one second, this many computations happen and there is a clock internally that makes it. So in the world of neural networking, it feels very much asynchronous. So you have a bunch of layers of neural networks, you know, your prompting and your input goes in on one end. And then it goes through these layers as it progresses, depending on how complex it is. And on the output, you get your output, right? So you put your references in on one end, you're prompted on one end, boom, boom, boom, layer, layer, layer, layer, and then an output. But that's not how our brain works. Our brain works by generating different sub-dependent thoughts at the same time.

30:51And then because they're generated at the same time, they then work together to form a bigger thought. Like combining the elements. Yeah. So like if you're driving, you know, a portion of your brain is processing the road and what's around it another portion is probably thinking about where to go the direction the map the gps and third portions probably you know thinking about safety and like keeping it on the speed limit and all those three things together combined to the driving decision you're making at however many fractions of a second you know so ctm continue stop machines uh really is a neural network architecture that makes sure that that notion of timing is incorporated into a neural network so it's not asynchronous there is a synchronicity after all i'm not uh 100 sure on how this will improve ai thought process and ai inference but but so i mean is it sort of like if you like hypothetically if you're running a large language model or something that it can like kind of think more about it can have my guess is it can have several different thought processes happen at the same time and then combine it into a super thought process that is far more powerful than those.

32:07And is this different than like the re how the reasoning models work? Because part of the reasoning models was like, they're supposed to sort of think or like do the output, but then like go back and kind of question it. But is that still kind of a linear process? I think that is, is just going backwards in the neural network, where typically you have a sequential forward process and reasoning processes also have a backward, like a feedback loop. I think this is more, I think of it as like parallel things happening in parallel. Like how the neurons in our brain go and triggering different things all over the place to kind of think.

32:40For sure. So we're going to link you to this article here. Go ahead and read through it. This is a little bit theoretical and a little too technical for me. But I just, Look, the reason I bring it up, this is why. We're at the very early days of neural network architecture, and something like this can potentially change the way the future models are built completely. Yeah, yeah. I mean, yeah, we had DeepSea come out of, not come out of nowhere, but, you know, kind of just pop on the grid and be like, hey, look, we figured out this other way to do this stuff like in a totally different way. Yeah, I mean, if you go back to the early days of compute, like in the 80s and 90s, Intel was making CPUs, and we thought this was the way computers were going to be made, right?

33:25And CPUs are really good at doing a very large computation at the same time, but, you know, they could do maybe 10 ,000 of them at once or 100 ,000 of them at once. It's sequential. It's got to go. Yeah, it's sequential, but there's still a notion of parallel pipelining. And the way they do it is they actually just add one CPU on top of another. So it's a bunch of CPUs running together. NVIDIA comes out in the 90s. Jensen, much younger than he is today, completely changes the game with GPUs. And GPU computation and the math and the underlying silicon is completely different. Fast forward to the world of today, what is actually benefiting us the most?

34:06It's GPU architecture, right? Like all of the modern CG runs on it, computer graphics, AI, everything is running on GPUs. GPUs are really good at taking smaller computations, but having millions and billions of them happen at the same time. So I just go back to something like this, where this introduces a completely new way to make the thing that powers AI. So maybe this is something that we'll look back into and say, yeah, that's when the path kind of diverged. And you heard it here on Denoist. your uh car comment made uh remind me this morning i saw someone uh getting into a road rage honking battle with a waymo oh ouch that that is a battle you cannot win i just kept hearing like honky honky honky and then i see the car like and then it pulls up next to the waymo but then i i don't know if they didn't know what a waymo was but then i think i saw them look inside and then realized that there was no driver and then they stopped this person must have had no idea that that wasn't Waymo to begin with because we could see the Waymo car from a mile away with the LiDAR and the sensors.

35:09Yeah, they're very conspicuous. You could tell what it is. But yeah, I'm just like, oh, this is going to be the future. Just arguing with an AI robot that is unaware. Dude, you just got a physical glimpse into the future. That's so awesome. Yeah. All right. Good place to wrap it up. Thanks for everything we talked about, as usual, at denoisedpodcast.com. And give us a comment on our YouTube videos. We'd love to see you engage there. So thank you. All right. Thanks a lot, everyone. We'll see you in the next episode.