zeke
/
zisper
My own personal copy of daanelson/whisperx
Prediction
zeke/zisper:d64f0cf8ModelIDedt7htdbdxi24uolokrzxcygbuStatusSucceededSourceWebHardwareA40 (Large)Total durationCreatedby @zekePrediction
zeke/zisper:d64f0cf8ModelIDa6gxr5tb4alxt3tyy6oh4u3go4StatusSucceededSourceWebHardwareA40 (Large)Total durationCreatedby @zekeInput
- audio
- Video Player is loading.
- debug
- only_text
- batch_size
- 32
- align_output
Output
[{"text": " Let me stand on the shoulders of giants. Yeah, that's what we're all doing. All right. Should we kick it off? Yeah. Cool. Hey everyone. I'm Zeke from Replicate. Today I'm here with my colleague, Charlie from Replicate and Bilal, who we just recently met online because Bilal made a really cool website called RepliCover. That is another way of exploring Replicate models.", "start": 0.008, "end": 27.127}, {"text": " So for people who don't know, Replicate is a platform that lets you run open source machine learning models using an API. And you can run models that other people publish, or you can publish your own public or private models also to run with an API. Replicate has this explore page, which is curated by the staff at Replicate. And we use it to feature models that are new or that are really exciting. The six models here on this page are", "start": 27.751, "end": 57.316}, {"text": " models that someone at Replicate thinks it's worth letting the world know about them. Of course, LLAMA is a really big deal right now. SDXL and stability, stable diffusion are really popular. We also have these collections, which are also curated by people who work at Replicate to help people discover new interesting models that fall into these different categories.", "start": 57.856, "end": 81.211}, {"text": " But we know that we could do better than just having a staff managed curated view of this content. So recently, about two weeks ago, we created a new API for Replicate. It's a public API and it lets you list all of the public models on the website. And the goal of this was we wanted Replicate users to have access to all the metadata about the public models so that they weren't limited by this explore page to be able to discover things.", "start": 81.582, "end": 109.628}, {"text": " Just one thing I wanted to point out is that this is not all public models. It's all public models that meet certain criteria. So if the model is public, and it has a published version, and it has at least one example prediction, then it will show up in this list. And there's a few thousand of those.", "start": 109.763, "end": 127.718}, {"text": " Cut to the chase. Our new friend Bilal found this API recently and used it to create this amazing new website called RepliCover. And RepliCover just does a beautiful job of listing interesting models and gives you a way to view them by whether they're trending or popular or new, and gives you ways to dig down into particular types of models. Welcome Bilal. Thanks for joining us. I just wanted to take some time today to just talk about how you came to build this project.", "start": 128.157, "end": 157.79}, {"text": " Yeah, no, thanks for having me. Yeah, I think I like a lot. I think a lot of your users love replicate using replicate, but I think we all agree the discovery feature isn't the best. And I know you guys are trying your best to make it better. And I think it will eventually get a lot better. But right now, I think one of the problems I was having was.", "start": 158.701, "end": 176.96}, {"text": " I, I like find a model, like somebody would tweet about the model, Oh, this cool model. And I'm like, Oh wait, I was on replicate a bunch of time, never saw this model. And I think the problem with curated collections is right. You either have to have someone who's on top of it and just continue manually updating or things get lost and you have to do the work yourself and dig into the lists. And so I really wanted just, I just started doing something basic. I was like, I wanted basically something I could get up trending.", "start": 176.96, "end": 205.175}, {"text": " for replicate models, where I was like, I just want to see not just the most popular models, but by first by category. And, and then the second thing is I wanted to do it by date because yeah, stable diffusion is going to be the most popular models, right? It's been the most popular for a year now, but.", "start": 205.276, "end": 222.016}, {"text": " There's so many better models now, right? STX is out. So just doing a sort on popular models is just going to fill up spaces with these OGs. So I wanted some sort of a filter on dates. I'm like the most popular model, but that was the last update happened within the last week or something or 24 hours. And so I thought that would be a powerful way of doing it on the models. That's where I started. And I, then I saw, stumbled upon, like you said, the API. I was like, oh, this is cool. Maybe I'll just throw together something for myself. And this is what happens with my projects. I just start something.", "start": 222.016, "end": 251.159}, {"text": " And it's just simple and then it gets a little more complicated and then I go down a rabbit hole. And then this is where I, this was what the MVP I was going for was this place where you can come, you can see the replicant models, you can sort, sort them based on popularity or the latest, and you can have a category filter and a time range. That was my basic MVP.", "start": 251.159, "end": 270.987}, {"text": " Which I hope now the thing though, I'd walk you through a tech stack of it and how to do it. It's just an, it's an XTS and a tailwind CSS side. So pretty simple. And we, what I use for the actual data is planet scale because I didn't want to just hit the API all the time. And also the API takes a while to get all the models. There's about 1200 models and increasing. So it's unless, and there's no sort built in as far as I know. And so I.", "start": 271.544, "end": 298.966}, {"text": " I couldn't just pull all the models into, it would just be too slow. So what I do is I do a daily sync where I just pull all the models to a sync with my planar scale database. And then I, with planar scale is it's a service that's basically MySQL under the hood. So it's a relational DB. So just all the searches happen pretty quickly. And so once I have it there, I can do the sorts and the filters pretty quickly. And it's just a SQL statements under the hood. Amazing. Yeah. On that topic.", "start": 298.966, "end": 325.173}, {"text": " Actually created an NPM package with this in mind. I published this NPM, this package, which updates itself every day. So it's a thing on GitHub where it basically just has a actions workflow that runs once a day and hits the API to pull down the latest changes.", "start": 325.865, "end": 346.351}, {"text": " And then it just pushes them to GitHub and pushes them to NPM. A database is one way to accomplish that, but you could also just have something that just does an NPM install on this package every day. And you don't have to worry about. Oh, that's awesome. Where does it store the data though? It's in a, it's in a JSON object in the model or in the repository in the package. So there's two flavors. There's the one that's all the metadata. And then there's one called models light, which are both available as ESM.", "start": 346.84, "end": 372.997}, {"text": " uh, imports. So if you just want the light one and you want to do like a browser side bundle, you can do that that way. Nice. Also, just while we're, we're on the topic of you mentioned, sorry, what was the database provider you just said? PlanetScale. I, yeah, this is the other day when I do project, I'll like sometimes I use them as an excuse to check our technology. So PlanetScale has been on my list. I have used Superbase before I've used other ones, Upstash like for Redis and stuff, but I wanted to just", "start": 373.57, "end": 399.912}, {"text": " Get familiar with PlanetScale because it's one of the best solutions for a SQL database for Next.js developers like me. And it's the free tier is awesome. It's 5 billion. Like reads and stuff. So I was like, okay. So we use PlanetScale and what is it called? Prisma.js in our open source Scribble Diffusion app. So Scribble Diffusion is a little app that lets you draw a picture and then use, um,", "start": 400.621, "end": 427.351}, {"text": " this model called control net to generate an image that sort of matches the outlines of your sketch. And one of the things when you're building an app on replicate is you sometimes need to store files in the cloud somewhere, either because the replicate API is going to delete those prediction inputs or outputs after one hour. So this is actually, this is an open source app that we have connected to PlanetScale. So if anyone is watching this and wants to know how to use a Next.js app,", "start": 428.431, "end": 457.591}, {"text": " With a database ORM and a SQL database, this repository could be a good place to start. So anyway, sorry to take you off track, but yeah, definitely. No, no, that's super cool. Matches the tech stack that we've been using as well. Nice. Yeah. It's a very powerful tech stack. And I also use the app router. So you'll see like in the, this was like, for me, this was a new one. I really like this paradigm is instead of using state, if you.", "start": 458.03, "end": 482.752}, {"text": " All of one of the tabs, you actually update the parameter as in the search URL. So particularly with server can be annoying to pass back props and stuff. So it's just easier. And for me, this was the first time I've actually done it using this. And I was like, oh, this is cool. This makes life easier. And it's also, you can go back and it'll take you to the last stage because it started, oh, you can bookmark it. So if you, so, so these are small things that I really, I geek out on these kinds of small things. And I think they're super powerful. Yeah.", "start": 483.41, "end": 511.844}, {"text": " Yeah, so that's like a so it's like a single page app that behaves more like a", "start": 512.097, "end": 516.367}, {"text": " like a regular server ended up where you can actually move around between different applications. Yeah. I wouldn't really call it a spy anymore because Next.js isn't really a spy, but it's, but yeah, there's all this sprinkle, like simple commands you can use for it. But yeah, that, that covers the filtering logic. And then I know you want me to talk about the categories themselves. I thought that was fun to figure out because initially I was just thinking, I just used the collection because I saw the curator collection. I'm like, Oh, that probably means there's a category.", "start": 516.788, "end": 544.396}, {"text": " field or something, but there wasn't. And I was like, oh, okay, this is going to be, I'm not going to manually label everything. And I wanted some sort of a category. So what I ended up doing, which I think was a fun trick was I basically manually labeled a few models and it's in the prompt.", "start": 544.396, "end": 560.039}, {"text": " Itself. And I basically gave it a list of categories that I wanted it to, and I wanted it to categorize itself. And then I basically ran a loop through every model and I passed it into GPD 3.5 and said, Hey, categorize this. And that way I was able to categorize the models. I also wanted.", "start": 560.039, "end": 580.508}, {"text": " Not just categories, more like tags, because you can have multiple categories because you can, you can have a generative art category and you can, and it's a technically also text to image generation. It's text to image generation, but maybe it's QR code monster or something, which is more artsy. So I wanted the ability to have multiple tags. And right now this is still something I'm sure. I think there are too many categories right now. It can probably be less than a dozen, but, and maybe better wording, but. This is this, I thought this approach worked well. You can also.", "start": 580.508, "end": 608.96}, {"text": " Did you use GPT to come up with this list or did you come up with this initial list by yourself? I use you. I'm a very lazy guy. So I basically, yeah, of course, some of the, I clustered these together and it gave me a lot of categories. Then I initially in my initial prompts, I said, you may add more categories. And then GPT went off the rails and added like 10 different categories. I was like, okay. And instead I. Redid it. So the initial part was like a slog, but, but now it's there. So what happens every day is around new.", "start": 608.96, "end": 635.774}, {"text": " seven or noon, my time, what a process get kicked off. It goes to replicate API, it grabs all the models, and then it does a sync. It just updates the run counts and the trending counts, which I'll come talk about in a second. And then what it does is it says, if there are any new models that were added, then go", "start": 635.774, "end": 654.607}, {"text": " and run loop through Jadgpd and then add categorize them and then add append the role. And that takes a bit. And then I prune any models that were deleted as well. So there's when anything with state has becomes a little complicated, which is why I'm hoping one day you guys will just do it on your own. And then I can just wrap it in. Are you doing function calling for the output of this?", "start": 654.607, "end": 676.628}, {"text": " prompt or are you telling it to output JSON and then parse it? It's the second image. You can go and you can see the output. It's just gives me an array, which is not perfect. This is why it's not perfect too, because sometimes it'll just mix something. And then I do some sort of catch all where I'll put it in an uncategorized bucket. So again, this is how to get there in a couple of days. There's lots of holes there.", "start": 676.628, "end": 696.777}, {"text": " But, and I actually think a more button up process, if you guys did it on your end would be that you cluster maybe the a hundred or so most popular models with categories. And then you do some sort of embedding clip embedding clip type of thing where you try to find them the nearest based on their description. That's definitely on the list. That's definitely on the list. So I think, but the idea of just machine tagging, I think is pretty cool because obviously it saves a lot of time. You don't need a team to do it and you only need.", "start": 696.777, "end": 724.385}, {"text": " You let the machine do the work 90% and the 10% curation team comes in and fixes any things. Or if there's a new class of models, which obviously the model hasn't seen, you create a new category for it. Yeah. And so what I was going to add here is that if you used either OpenAI's function calling or now we have LLAMA function calling, you can just specify the JSON that you want this prompt to output. And then you won't have any problems with it outputting it in the wrong format or whatever. Yeah, that's true. Yeah. I need to look into that too, but that's a good call.", "start": 724.385, "end": 753.629}, {"text": " Hey, what was that? Yeah, so I just want to know. Yeah, the grammars thing was yes. Yeah, I've seen this, but I haven't even played with it yet. But basically, you provide a JSON schema, when you're making an API call to llama, or these are some of these other models. And then you're guaranteed to back structured data. Oh, nice. That's awesome.", "start": 754.321, "end": 779.6}, {"text": " With a lot of two models with support for grammars. Okay. So apparently all of these. Yeah. All the, all the llamas we have support for. Oh, so they have it built in. I thought it would be like something like a line chain wrapper or something using the prompt to call a function. Yes. If you, if you click on one of those, you can see a good example of. Yeah. So you specify the JSON schema there. Amazing. Okay, cool. Something to play with. I haven't even.", "start": 780.106, "end": 809.131}, {"text": " looked at this stuff yet, but it's pretty exciting. So yeah, I've noticed right when I came to replica cover, like in the first few minutes of using it, I was like, Whoa, I've never seen this model before. So even like for someone who works at replicate, you would think that we would know about all the cool stuff, but it's still actually, there's so much going on. If I set this to the last seven days, so there have been these five new", "start": 809.823, "end": 835.439}, {"text": " video generation models added to replicate in the last seven days, which is pretty remarkable. Like really cool to be able to slice and dice the collection of public models that way. Yeah. I think there's so many hidden gems. That was my motivation for this. That's why I initially, I don't know if you figured it out, but the replicate was like.", "start": 836.435, "end": 852.466}, {"text": " Replicate and discover. So I initially, I was trying to go with Repliciscover, which I was, I really wanted Repliverse, but it was taken. I was like, ah, Repliverse would have been a cool name, but ended up with Replicover. And then, oh, and then really quickly for the trending part. So what I do there, which I think that I need to change as well is right now I just take the last, I subtract the latest run count with the old run count. So it's basically a 24 hour.", "start": 853.023, "end": 878.538}, {"text": " increase in the run count, which I'm trying to do with training. Problem there now I'm noticing is obviously someone who has a million runs there, the absolute value is always going to be larger. So what I'm, I might do a minimum threshold of a thousand runs and then do a percentage on top. But I think one thing I think there's a huge gap is not just find the latest models and stuff, but models that are like really rising in the ranks and catching them. So again, the analogy to get uptrend, I don't know how they do it, but you'll find it. They update that every 24 hours. So having that kind of training phase.", "start": 878.538, "end": 908.373}, {"text": " is pretty powerful. Yeah, I was trying to do that actually. I did a little experiment with that for this NPM package where I was thinking, okay, every time you run, do a diff against what was run yesterday and see, check as like a comparison of the total number of runs for that. But that's imperfect too. So as soon as I saw a replica cover, I was like, you know what,", "start": 908.373, "end": 927.492}, {"text": " I'm not going to bother with this stats thing anymore. I'm just going to link to all the websites that people build that use this data that are going to do a better job of actually like making it easy to slice and dice it and sort it. But really excited to see that you're thinking about that. I also noticed, so I was looking at your website and it looks like, okay, Charlie and I were just hanging out last weekend and we were talking about", "start": 927.492, "end": 949.666}, {"text": " this exact thing. So before we even saw that you had this project UTLBR, I was saying this thing happens to me all the time where I'm watching a YouTube video and I just want a transcript of it and I also want to generate like a short summary of it so that I can either remember what it was for, bookmark it, or like share snippets from it with my team. And then literally this week I found out that you made it, which is so cool because now I don't have to build this thing. And then I noticed", "start": 949.784, "end": 979.4}, {"text": " that you have like a premium version and that's pretty cool. I was just wondering how this project came to be and like how you ended up figuring out how to like apply a premium model to it and whether like curious if it's like successful for you and if you're able to if you're promoting it at all or yeah.", "start": 979.805, "end": 998.3}, {"text": " Sure. First off, I mean, I'm glad you guys liked it. And yeah, UTL there is fine. It's like actually something I, one of my first projects I made was built with my brother over a weekend, three, four years ago. This was even before, and summarization was so difficult then, and we had a packed together product and it actually got a lot of upvotes on product time and it was doing fine. It was making like two, 300 bucks a month, a year, a month, sorry. And we were just letting it run on autopilot. And then what I did was.", "start": 998.384, "end": 1025.3}, {"text": " The whole LLM thing happened. I was like, wait, I need to revamp this. So I actually, cause initially I was using, the summary wasn't that good. And it was taking, I was, we had all these like different tricks to it. Then I use all of the chat GPT under the hood. And it was like, this is way better. So I revamped the whole website, put it out there and now it almost pays my rent. So it is, so it is like probably my most successful product. And the SAS version definitely makes my, we have about.", "start": 1025.3, "end": 1050.477}, {"text": " I think right now it's around $1,800. I'm not a crazy millionaire or anything based on this, but it helps a lot letting me play around and build fun stuff like RepliCover. And the SaaS, the free version of what I like, what I think one of the reasons it's so powerful is with projects is it lets people come in and immediately do something.", "start": 1050.477, "end": 1070.727}, {"text": " And you can just try it out for free. So I get, we get a shit ton of traffic because of that. And people just use it for free. And then some people like, I don't want rate limited, rate limits, and I want summaries and send the signup for the pro package. Now, I, as much as proud as I am about UTLDR, I do think there's a limited time window for this because I do think that eventually Google will just build this into YouTube. I'm surprised they haven't done it already, but I know that's on the cards because.", "start": 1070.727, "end": 1098.335}, {"text": " summary on automatic summarization of videos and stuff. It's coming, but for now it helps. I think the other thing that really sets us apart from other summarizers, because there's a bunch of other YouTube summarizers, we do it by language. So we get actually a ton of non-English speakers coming to the site. They'll do Spanish, they'll do French. And so getting the transcripts immediately in their language helps. And also then the summary is helpful as well.", "start": 1098.335, "end": 1123.006}, {"text": " So is that using like whisper under the hood for transcription or how are you doing? Are you getting it from the YouTube API?", "start": 1123.006, "end": 1129.688}, {"text": " There's a secret, there's a black hat there. No, it's not. It's that there's an open source repo that kind of uses, Google is going to take, they're not going to care, but there's actually an API that lets you grab Google's YouTube subtitles themselves, which is why, which is so funny because I, when I posted it initially, I got so many ML engineers, dude, your transcription is so much better. What model are you? I'm like, yeah, because I'm using Google. That's why it's better than yours.", "start": 1130.802, "end": 1154.697}, {"text": " And because the thing is it's instant, right? You put in any YouTube, you get instantly, even with Whisper, like fast, or even the best site in the game right now, like DeepRamp stuff, it still takes a while. So the instant, that's why what happens is like Google, what it does is it back processes all this for me. And so new one shortcoming here is like when you put in an initial URL, if it's uploaded within the last hour, you're not going to get anything because Google takes a bit time to process it.", "start": 1154.697, "end": 1180.516}, {"text": " Got it. But for the most part, people love it. It works. Hopefully Google won't shut me down, but I think I need to be make a lot more money for Google to even notice it. So I think I've been down this rabbit hole before of trying to find the secret YouTube transcripts API. I think, I think there is one, you can Google it. And honestly, I want to use YouTube's API. It's so ridiculous. Google is.", "start": 1180.516, "end": 1203.297}, {"text": " one of the worst developer products. It's so crazy. They have such good data, such good services, but their documentation is crap, and their products are crap, and their limitations are crap. So I actually want to do it the right way. I want to use YouTube's API. I want to pay them for it. Can't do it. They give you 10,000 units a day. One video subtitle costs 1,600 units, so I can do five transcripts a day.", "start": 1203.297, "end": 1226.382}, {"text": " And otherwise I have to go through the rabbit hole of getting approved and all that, which maybe cost dollars and all that. It's just crazy to me that you do same with Colab. I know you guys don't probably use Colab a lot. I would love to buy Colab Pro. It's just, they've just done a shit job of marketing it and actually making it worthwhile. And so there's so many Google products that you can just, if you ever think, Oh, Google already has a product. I shouldn't build a product. Build a product.", "start": 1227.209, "end": 1252.572}, {"text": " because you can probably do a better job of building a nice UX. What is next for you? Are you going to keep refining your existing products or do you have like a list of new ideas of things that you want to build using ML models from", "start": 1252.572, "end": 1267.608}, {"text": " from OpenAI or Replicate or I'm just curious. I'm the classic shiny. I got too much shiny objects in Drome. I get too, too discarded. This was not on the cards. I literally had a plan of, I was working on, I have another product called planescribe.com. So that's like a serious kind of a B2B almost product. So I was working on that and then I saw this and I went down this rabbit hole. So I go.", "start": 1268.08, "end": 1287.892}, {"text": " down different rabbit holes as well. Like I literally before this past week, I built this Giphy app, which is also using Replicate. So it's AIGifts.Vercel.app. And it's basically the same. It just uses the GIF model, the hot Excel model and stuff. And yeah, I love building these one-off things and stuff and using them as legions for my main products as well. But there are, what I do is I have a couple of products that I'm more serious about that. I'm like, okay, this is more of a business.", "start": 1287.892, "end": 1313.272}, {"text": " that I need to focus on. And then I build these one-off free tools, like replica cover and stuff, which is the worst case. No one cares. I learned something best case. I get to talk to cool people like you. And because of that, yeah, it's yeah, it's a nice time. I think to be a front-end engineer or just be a full-time developer who wants to build products, there's so many cool tools you can use. And this is open source. So if anyone's interested, they can just work it and use a different raw replicate model and do it as well. So amazing. Yeah.", "start": 1313.272, "end": 1342.178}, {"text": " Do you have any questions for us before we wrap this up? Or anything you want to add? Yeah, no, I think that about covered it. Hopefully people find the site useful. And for you guys most curious about what's on the card for Replicate and you guys have any, what are you most excited for in the next few months in terms of, it doesn't have to be Replicate specific. It could just be in general about what's coming. You think it's coming down the pipeline in the ML, AI industry.", "start": 1342.887, "end": 1368.554}, {"text": " I'm really excited about these music generation models. Just this past week, MusicGen fine tuning came out and also a MusicGen model where you can give it a chord progression and it plays a remix of your input song with that chord progression.", "start": 1371.018, "end": 1387.437}, {"text": " And I think that's awesome. I'm also still really excited about text-to-video. Text-to-video hasn't quite had the stable diffusion moment, but that feels like in a year from now, the text-to-video is going to be a lot faster and a lot higher quality. I agree. That would be a game-changer. I looked at animate-diff-on-replicate again, which is cool, but it only does it for a couple of seconds. You've got p-collapse.", "start": 1387.876, "end": 1409.797}, {"text": " We haven't hit the runway model because it's just right now, it just typically happens. You have players like runway, which is just wait list for the Gen 2 model. And it's so expensive, you can't really do it. So once we have an open source version where you can do it, and there's two versions, right? You got the one text to video where you put in a prompt and you get that. And I'm actually more excited about the other one where I'm like,", "start": 1410.067, "end": 1431.835}, {"text": " I take a video and I can just style transfer the video. And I'm actually surprised we haven't gone more far from it. People have hacked together these animate diff loops and stuff and just do two second at a time. But I'm surprised. Have you guys seen that rock, paper, scissors, like video? That was so cool. Like people like they use, they film themselves on green screen and they.", "start": 1431.835, "end": 1451.326}, {"text": " turned it into a 90s fantasy anime. And I'm like, that's for me, that's been one of my dreams for the last five years to build my own Netflix short anime. And I think, I think we'll very quickly get there where people can build their own like mini series and stuff. And it's going to be a creator's paradise. So yeah, I definitely agree. Video generation, audio generation, you got the music thing, but then you also have the voice changers and stuff.", "start": 1451.326, "end": 1471.627}, {"text": " 11 labs and play it on HD are doing awesome stuff. They're like more further on that. It's just on the dubbing thing with 11 labs. Super exciting for that, for that part. And yeah, there's so many every day. I'm like, Oh my God, there's so much stuff coming on. How do I do everything? Yeah. That's pretty much how I feel about it too. Like I have a list of things that I know. I haven't even checked out that like JSON grammars thing, which is like the sort of equivalent of open AI functions, which opens up the door for so many interesting applications to be built.", "start": 1471.627, "end": 1501.462}, {"text": " Yeah, so much cool stuff. And yeah, yeah. And one other project I mentioned, which I'm doing right now, I'm super excited by, I don't know if you guys are into React, but Remotion, which is, it's not like AI specific, but basically it lets you programmatically generate videos. And I think that's so powerful. By the way, I love that exclamation mark, do you think? I've never seen that before. That's pretty cool. You can search that. Oh, that's the duck, duck, go, bang. Yeah, shortcuts, nice.", "start": 1502.035, "end": 1530.318}, {"text": " That's awesome. So you can like type in bang Wikipedia or Google Maps or whatever. Yeah. Turns your address bar into.", "start": 1530.318, "end": 1541.54}, {"text": " command line. Nice. But yeah, ReMotion. So yeah, it's just basically a programmatic way of building videos. And I use React as my hammer. So that's why maybe I'm attracted to it. There's other, you can use FMPEG RAW as well if you're super hardcore, but I've been playing around with it. And I love the fact that you can", "start": 1542.097, "end": 1560.406}, {"text": " Just like have code that generates these videos and stuff. Now you couple that with some of the stuff, which I think people are doing, but you're going to see more of it. If you've seen these short automations and stuff, it's so cool. Like you can take, you can use an AI model like stable diffusion or, or even an API based model, which is more mid journey level or whatever. If they ever release an API for.", "start": 1561.283, "end": 1582.276}, {"text": " But if you get those images or videos, then you use transcriptions using Whisper or other models, then you use 11 Labs or one of these text-to-speech APIs, and you stitch these things together, you can basically generate complete media.", "start": 1582.276, "end": 1599.59}, {"text": " And I think that is so powerful. I don't think we've completed. We have it. I think very few people are actually understand the impact of it and are working on it. And, but we're going to see in the next six, 12 months, how this is going to explore in terms of TikTok and videos. And for someone, for a guy like me with an accent and who's maybe not the most, the best person in terms of explaining his ideas, that's so appealing that I can just have code and then.", "start": 1599.59, "end": 1623.603}, {"text": " create these amazing animated scenes, but then also these voices with these effects. And so I think the convergence of what we talked about, video, image, we had the moment, right? Video, hopefully we'll have some audio. I think we had it. And it's just going through the latest stage of finding when these converge together, we're just going to have this explosion, this Cambrian explosion of ideas and content creation. It's going to be amazing. Yeah. Awesome.", "start": 1623.603, "end": 1648.527}, {"text": " Thanks for meeting with us today, Bilal. It was really great talking to you and keep up the good work. Yeah, I appreciate it, man. Thank you so much. Yeah, this was awesome. Thank you. Thank you. Take care.", "start": 1649.759, "end": 1660.424}]Generated in
Want to make some of these yourself?
Run this model