In short
Podcast Summary: This Week in Startups - E1747
Episode Information
- Title: Demoing Google’s MusicLM, AssemblyAI, and other AI tools with Sunny Madra
- Host: Jason Calacanis
- Guest: Sunny Madra
- Air Date: [Insert Air Date]
- Podcast Link: [This Week in Startups](https://www.launch.co/podcast)
Episode Description In this episode, Jason Calacanis invites Sunny Madra to discuss the latest advancements in AI technologies, focusing on Google’s MusicLM, AssemblyAI's LeMUR, and other AI-powered tools. The discussion provides insights into how these tools work, their implications, and the current state of AI development.
---
Key Topics Discussed
- Launch of ChatGPT’s iOS App
- ChatGPT recently launched its iOS app.
- Offers features similar to the web experience with added functionalities, like home screen widgets.
- The app provides a more convenient way to access AI-driven information.
- Demo of Google’s MusicLM
- What is MusicLM?
- An AI model capable of generating music based on user prompts.
- Demonstrations:
- Users can input creative prompts (e.g., "drums that sound like rain and thunder").
- The output music showcases the model's ability to interpret various genres and styles.
- Training Data Concerns:
- Discussion on the training data used (possibly sourced from YouTube).
- Concerns about copyright implications and the model's reliance on copyrighted material.
- AssemblyAI’s LeMUR
- AssemblyAI's tool allows for automated transcriptions and Q&A generation based on audio content.
- Demonstrated by pulling transcripts from a previous podcast episode and generating answers to specific questions.
- Highlights the efficiency and accuracy of modern transcription tools.
- Wondercraft AI
- A tool that creates podcasts from bulleted points.
- Demonstrated by generating a podcast intro based on provided topics.
- Offers customization options including voice selection and background music.
- Stable Diffusion’s Deep Fake Technology
- Discussed the advancements in deep fake capabilities and their implications for media and entertainment.
- Examination of the ethical and legal ramifications of using AI to generate content that mimics real people.
---
Key Takeaways
- AI Tools as Creative Aids:
- Tools like MusicLM and AssemblyAI empower users without technical skills to create and manipulate content, potentially democratizing access to music and information creation.
- Navigating Copyright Issues:
- The dialogue highlights a growing concern regarding the use of copyrighted material in AI training datasets, suggesting a need for clearer regulations and guidelines.
- Market Impact:
- There is an increasing competition among AI tools, pushing companies like Google to innovate rapidly in response to emerging competitors.
- Future of Content Creation:
- The convergence of AI and human creativity presents both opportunities and challenges, with implications for how content is produced, consumed, and monetized.
---
Additional Notes
- Sponsor Messages:
- Issuu, Vanta, and .Tech domains provided sponsorships during the episode, each promoting their services tailored for startups and tech companies.
- Call to Action:
- Encouragement for listeners to explore the discussed AI tools and consider their implications for their own projects or interests in the tech space.
---
Conclusion Jason and Sunny’s engaging conversation sheds light on the current state of AI technology, demonstrating its capabilities while also addressing ethical considerations. This episode serves as a timely reminder of the rapid advancements in AI and the necessity for stakeholders to navigate the evolving landscape thoughtfully.
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Transcript
Automatic transcript. May contain errors.0:28I think it's interesting. sign up for an annual premium account and get 50 % off when you go to issue.com slash podcast and use promo code twist. That's I S S U U.com slash podcast and use promo code twist Vanta compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a sock to report fast. twist listeners can get$1 ,000 off for a limited time at vanta.com slash twist and dot tech domains are the go-to namespace to build anything in tech and home to some of the world's most innovative startups secure your dot tech domain today and lock down a one-year domain for ten dollars or a five-year domain for fifty dollars at go dot tech slash twist today all right everybody welcome to this week in startups and really it's this week in ai uh we've never seen anything like this so i made an audible i called an audible i decided to make a call in the middle of the year to have one of my great friends on sunny madra and vinnie vinnie's out today um and we just decided gosh instead of talking about all the nonsense around where we think ai is going we just thought we would show you what ai is doing today and this will serve as an incredible roadmap for you this is your ai mba this is your master's degree this is your phd tune into this week in startups i want to show you what's being built and sunny madras here he of course is the co-founder of definitive intelligence which i have a little slice of I put a little investment in there and that lets people do analysis of information on the blockchain and other data mining analysis services.
2:23I think I described it pretty good. Your own personal data analyst. Your own personal data analyst. Ooh, this is a new mission statement. Updated, updated. Updated, evolved. Okay, I like that. Your own personal data analyst. So we talked about this in chat GPT4. i don't know if it was maybe a month ago three weeks ago it felt like three years ago but i think it was literally three you could upload a csv and ask the data analyst at chat gpt4 hey tell me about this file and it would make charts and give you some analysis so you're going to build the industrial strength version of that or you're building the industrial strength version of that yeah yeah we we already have it you know you know chat gpt is good if you have 100 megabytes of data if you have terabytes or petabytes then come talk to us all right so if you're if you're screwing around you can use the code interpreter uh on chat i wouldn't say screwing around if you're just doing something with small data but you know if you're a big enterprise and you've got you know transactions for the last 20 years of your customers that's going to be you know several terabytes or even approaching petabytes for you and you need to use a different set of tools big news uh chat gpt launched their ios app yes which really does change your behavior um i have been trying to get you know i had this weird thing i realized mobile chrome does not let you set the default browser i didn't know why that's the case like i guess chrome is getting greedy google's getting greedy safari's getting greedy they want to own that real estate i like to set my home page up at what i want it to be but this is now i'm going to just stop using browsers i'm just going to go straight to the chat gpt4 app and current interpreter is now built into it so did you download the app and i had the app by the way like three months ago they put me in the test flight yeah yeah and it was single use now it seems like they're trying to get parity between the web experience and the app experience it's not perfect yet but it's getting there yeah and i mean it doesn't have all the advanced features like some of the stuff that we've been demoing but what i found interesting is they did do like a like a you know like a home screen widget so i think that's what you were looking for at one point right jaykel and so you can with the latest one you can make it a home screen widgets you can just push something on your home screen and it pops up and you can ask your question but it what we really need is i need to be able to say hey chat gpt as opposed to saying hey siri and so that's the change that we really need is can i use siri as an interface for it and why do i need to use siri for it i guess maybe on an android phone i could do that because Android lets you integrate into the operating system a little bit deeper than iOS does.
4:59All right, so we've been playing with a bunch of stuff this week. Big conversations all over the internet about regulation. We'll talk about that in a bit. But what have you been playing with this week? Well, you know, let's start with someplace fun and then we'll get into sort of the meatier topics. And I know you're a big fan of this, J.C.L., so I think this will be a fun one for us. I'm going to make sure I share this properly. Okay. Okay. So Google this week, uh, or, you know, launched this or made available to some folks, uh, what they call music LM. And what this is, is basically a, um, you know, a model that is able to generate music based on prompts.
5:40And so, um, you know, what, what we can do is we can take some of the recommended ones here. And so you can do, let's start with something simple. you click that here and it'll give it just a few seconds and it will generate this and so okay so we're playing some funky music here what was the um so that was the prompt here i'm just gonna read out to everyone it was drums that sound like rain and thunder so that one is not all that exciting um you know well i'll do another interesting one here which is uh chill out elevator music and then we'll get into some some more kind of fun ones in a second and for the folks watching here you can see some recommended prompts it comes up with so you know it's okay so this is you know it's kind of interesting feels like a bit um it you know early music generation we've gone through this before but where it gets kind of much more interesting is when you prompt it yourself so i'll do one jay cal and then you'll probably make it better for me so let's do tulum uh beach party the old tulum beach party yeah with um uh a bit of uh what do you want to put in there let's say cyberpunk yeah cyberpunk okay or blade or something let's see see if we get anywhere close so there we go
7:04maybe maybe i heard some cyberpunk in there yeah yeah yeah um you know you can do atlanta trap music with a New York 80s flair. Oh, okay. Yeah.
7:30You can see it's got a bit of... Sounds like Jay-Z's voice. No voice, it's just music. I thought I heard... It's just music. But you could hear Jay-Z's voice going to something like that. You could. You totally could. Yeah. So it's kind of getting something. So this has been trained on copyrighted music, I assume, or on maybe a library of... Okay. What is this? New York 80s rap. Yeah, this is New York 80s rap. Yeah. Back in the day, we used to on the block.
8:09Fly shoes. can't lose boom bap okay listen if you're running a sales team you got a design agency you got a media business you know the hassle of one pagers right they never format correctly on mobile you can't track them properly it's a complete mess i have worked in the media business for many years three decades in fact i'm getting old and nobody had figured this out until now you have to check out Issue. It's issuu.com. Issue is the all-in-one platform for you to create and distribute beautiful digital content. It's not just one-pagers. Those are important, of course, but you can create marketing materials, magazines, catalogs, portfolios, and so much more.
8:49Beautiful, evocative collateral to help you sell your product or service, but it also comes with amazing analytics that you can just pop up on a dashboard and you're going to track the reads, the total time spent, device breakdown, are they watching it on their iPad, their phone? It also works seamlessly with tools like Canva, Dropbox, MailChimp, and InDesign. Having a trackable magazine, if you want to level up your marketing needs, you need to use Issue. So get started for free or get 50 % off an annual premium plan at issue.com slash podcast and use the promo code twist. That's I-S-S-U-U dot com slash podcast and use that promo code twist so they know I sent you for your free starter account or 50 % off an annual premium plan.
9:34Yeah. So, you know, I think this is pretty fascinating. Like, where is this applicable? So, I don't know if you're familiar with Rick Rubin, right? Legend in the industry. Big fan of the All-In pod. He just sent me a copy of his book. Oh, awesome. Yeah. Oh, really? Okay. That's incredible. He personalized it too. So, where does this go? Let's talk about that a little bit. So, you know, he gave this pretty cool and interesting 60 minutes interview. Yes. Not that long ago, right? And one of the things that, yeah, there we go. Awesome. We got it queued up here. And one of the things that he talks about here is that he has no musical abilities.
10:06He does not know how to play instruments or use the mixer board. Maybe play, let's play the clip real quick. It just became a meme with Anderson Cooper. But exactly what he does and how is difficult to describe. Do you play instruments? Barely. Do you know how to work a soundboard? No. They have no technical ability. and I know nothing about music. You must know something. Well, I know what I like and what I don't like. And I'm decisive about what I like and what I don't like. So what are you being paid for? The confidence that I have in my taste and my ability to express what I feel has proven helpful for artists.
10:53So curating. really having a great year and curating stuff and being confident in your selection process and if you take that alongside you know a tool like this i think there's a lot of people that you know fall in a similar category that don't have the skills to you know maybe play an instrument and or run a soundboard but know the type of thing that they want to listen to and so i think tools like this will evolve to support folks that are not just songwriters but that have a great ear for something but i had no way of creating it didn't know how to pull up like any of the music tools i don't know if you've ever tried to use garage band or something that i mean this stuff is it will take you 10 000 hours uh yeah you know the proverbial 10 000 hours and in fact i just saw a clip of um the lead singer of the smashing pumpkins talking about how you know people will be able to get 50 different riffs and instead of having to spend a decade or two in their parents basement you know like he did finding interesting ones they'll be able to just do that with the software and it will still be a selection process so you're defending creativity in this process i'm wondering this is gonna this also was a copyright uh and there was a copyright uh hearing last week that kind of went unnoticed where people are just talking about who gets the rights to this so did google say how they train this did they find stock music which is available royalty free and so what that means is somebody creates a beat based upon what beats in the 80s were like but they don't um copy anybody specifically they just copy the genre and then they put it available in a thing like splice or there's a couple of other places you can buy beats that are royalty free you can put into any song you don't have to pay a royalty you don't even have to give credit so I'm wondering how Google trained this because Google is smart enough to know that they should not train this on Spotify or something.
12:49So, did they? Yeah. So, I think, and Nick is sharing the same thing here. I think it's trained on samples from YouTube. Ah. Which is a very fascinating thing because they know what's copyrighted in YouTube. Obviously, they do a ton of work around that. Jason, you've talked about that a lot. And then they have a bunch of stuff that is just done organically. organically people are creating that's what you know youtube has so much content on it now if you're a creator and you've created a beat it's been uploaded there so my understanding is that they've taken clips from youtube to got it here and i guess we didn't hey guys you didn't put where this these quotes are you put two great quotes in here where are those quotes from oh sorry um that is from google's data set on kaggle which is where they host everything i can pull it up on the screen okay so but is that how this was trained or you're guessing yes it's from their GitHub for the Music LM.
13:42Alright, and so producers just pulled up that they built this, they trained it. Audio set consists of an expanding ontology of 632 audio event classes and a collection of 2 million human-label 10-second sound clips drawn from YouTube videos. Huh? 10-second sound clips drawn from YouTube videos. It doesn't say if those are copyrighted or not. The music caps database contains 5 ,521 music examples, each of which is labeled with an English aspect list and a free text caption written by musicians. So if those 2 million human labeled clips were of copyrighted content, I think there's a copyright claim here.
14:31Maybe they're saying they're only using such a small amount of it, or maybe they're using people in the public as a way to defend themselves against copyright violation by saying those people made these clips so then therefore you have to blame them so in my scan of this jaco we were kind of looking at this when we were preparing yeah all the clips seem to be of like uh user generated content not okay not yeah so not like oh here's 10 seconds of rolling stones or something like that the reason i bring this up is because we have talked about sunny many times uh google waited to release this because google is so aware of copyright claims because they bought youtube at a time when youtube had a multi-billion dollar you know lawsuit with viacom and youtube was going to be shut down for all the copyright violations going on there and they were able to navigate that so google is incredibly savvy on navigating fair use they use robot.txt in their google search so you can opt out of being crawled they get permission from people sometimes generally they let people find music um that other people are using in their videos and then let them shut down those youtube videos or claim the monetization of that wasn't that the big unlock jay kelroy i've heard you talk about this a few times is when you can put music in but they'll basically they'll demonetize you for that video and they'll pick up the monetization so if you have a video that has you know millions of views we made our own track here and then we copyrighted it so again i i guess we could take these tracks and then we could claim copyright on them even though the ai built it i wonder how that works i wonder if that's in the terms of service so if you and i sang over one of these beats then we put it into on youtube and we use distro kit to publish it to spotify whatever we claim copyright on it now if somebody else uses it in their video we can then get an alert based started being um tagged in a database content id is the system con id id flags it tells us and we can just say automatically shut it down so the beatles might want to just shut down anybody using their video music and videos but somebody else the rolling stones might say yeah let anybody use it by default just take their money and so you may have seen all in two weeks ago we put the balanciago ai video that somebody made yeah at the end of the show and then it turned on ads on the show so then we had a choice either turn the show off re-upload it or just deal with the fact that somebody else is going to make you know five thousand dollars off that episode which i think we just decided the latter well yeah so and that that's an interesting one just okay so then it was it the music or was it the music that they use somebody owned the right set of music so they didn't make ai music they used somebody else's track that i guess was used in a balanciago stuff before i'm not gonna play it here okay crazy and here's the quote from this is a really interesting discussion here's um the lead singer of the smashing pumpkins my general uh thought is that um ai will change music forever because once a young artist figures out that they can use ai to game the system and write them a better song they're not going to spend 10 000 hours in a basement like i did they're just not but are you do you still have the ability to get real art that way well ultimately art is about discernment right so uh like somebody was telling me the other day about uh how an uh a famous rap artist would work they would bring in all these different people and they would sort of pick the beat that they were most attracted to okay now let's change it to ai hey ai give me 50 beats listen i'm not really feeling ai give me 50 beats from the 50 most famous rap songs of all time okay well i like number 37 that that inspires me yeah so he basically nails what you just said you can use this as a tool to be inspired use your curation your thoughts on it yeah i was actually gonna just put in what he said in in the tool right like should we give it a try oh yeah good uh right he said uh give me a beat uh that is uh inspired
18:52by the 37th greatest song in hip-hop in rap yeah 37th okay yeah he said just see if he goes and finds a list on rolling sound rap song okay ever right see what it does i don't think it has the 30s no no maybe no see it's not well that could be like an eminem kind of beat
19:21and i killed your cat um yeah okay maybe it's got some work to do but i think here's the thing this is going to come out with weak results because it's not trained on copyrighted music so i think this is weak and it's a 1.0 right yeah but you know what i would say is like youtube is full of very talented people that you know can produce like almost the exact same quality uh yes you know copyrighted music so you could hire people yeah but here's what would be interesting you're going to be able to take that language model run it on your laptop and take every single dire straight song or every single m &m song jay-z song biggie put it in there and then you're gonna be able to do it on the copyrighted stuff aren't you yeah yeah yeah and so that's what people are going to be doing and that's something google is not going to release but this is in fact an opportunity and i guess this goes back to what grimes is saying you know like hey just use you know what's her tool like elf something um so she's got a tool that they she partnered with somebody just put out there and she said people are making songs that couldn't possibly be me because i have a speech impediment of some type like a lisp or something and they elf.tech the enunciation and this is better than i can enunciate yeah um i think one got published right just last week um i think it did yeah and so shout out to dot tech our new sponsor here for elf.tech with grimes i wonder if grimes secured a bag for using dot tech um if you're a sas or services company that stores customer data in the cloud then you need to be uh sock to compliant you knew that from a third party and you need that third party to close big deals.
21:14And if you want to get compliant easier and faster, you need to use Vanta V A N T A Vanta makes it so easy for you to get and renew your sock to on average van customers are sock to compliant in just two to four weeks, prepare that to three to five months without Vanta and Vanta can save you hundreds of hours of manual work and up to 85 % of compliance costs. This is a total no brainer. And Vanta does more than just SOC 2 compliance. They also automate up to 90 % compliance for GDPR, HIPAA, and more. You can't afford to lose out on major customers. We all know that. Listen, it's a hard year. Last year was hard.
21:50You can't lose those major customers because you don't have your compliance dialed in. Just work with Vanta. Get your compliance automated and tight and tight is right. Lock down those big deals. Here's the best part. Vanta is going to give you$1 ,000 off. That's 10 hundies. Get$1 ,000 off at vanta.com slash twist. That's vanta.com slash twist for$1 ,000 off your SOC 2. very very interesting moment in time how would people tell i guess is the question where this came from because it's not giving you sources and this is the thing i think all these ai's need to do and i think this is where the legal issue is going to come into play again i don't want to obsess over legal stuff but either you regulate yourself or you get regulated that's nature of business of life if these people don't show their work and it would be possible to show here's what the ai here are some of the source material the ai used to make this beat it is possible to do that it is possible right like it is like taking bits of it but it may be it may be you know several bits maybe you know and it's all possible yes and i think that's where we're probably going to have to head towards because the black box nature just that's also very much on the closed side of things and i and and i you know what i anticipate is as we move to the more open models jcal what we'll see is the those ones will have the kind of references baked into them so people will want to see where it came from and that'll force the folks that are doing it in the clothes i also think the challenge is for um and this is interesting because um uh it came up earlier today in terms of i think a lot of folks are let me try to find this link but talking about like what these models have been trained on and i'm going to try to find this link as we're just talking about it here and and send it to the uh to the guys but like um there is a really interesting um share here that people talked about what the training set was and one of the things included the enron emails uh for many of the models like so there's a big open source set from you know back in the late 90s 2000 era yeah so that got dumped there's no copyright yeah it's just in the open because of the lawsuit right yeah it's a lawsuit but each of those individuals i wonder who owns the copyright to those emails it's not public domain it would be enron's ip yeah here it is there's no enron left to sue so i guess you know in order to have this is a really edge legal case is sort of interesting to me but not anybody else the interesting thing is in order to have you have to have somebody bring the case and there's nobody to bring the case for the enron emails you know when a company goes out of business it's like you know who's going to defend you know you have to find who owns that copyright there is nobody owns that copyright no there might be a shell company somewhere that does um which is super interesting but uh the grimes uh oh what oh here's a chart what is this chart so this shows uh what someone let me read this uh for the folks listening what's in the pile data set from ai euler that is being used to train a lot of the open llms right the following chart analyze the tokens in the test set there are 22 sets in pile from sources like wikipedia github uh you know books three etc and you know if we kind of zoom in here uh you see at the bottom it's enron email so there's like 600 000 enron emails in the tree yes yeah so uh hacker news yeah is in there uh i guess that must be open sourced um and there's youtube subtitles now that's interesting i don't see twitter in here yeah i do see stack exchange which didn't give their permission i'm sure so that's an issue pubmed extracts somebody probably owns that so this is really interesting we're going to have to have data sets the reason ai is going so well is the data sets right and the combined with the hardware combined with the software you take out one of those it's kind of could narrow the um the success of this correct correct yeah i mean they you know at the end of the day these are systems which use you know math to predict um the most likely kind of next character in a sequence that the human would like to see based on what has been trained on so if it's not trained on as good material you're not going to get as good predictions and so yeah i i can see that really changing and then and you know you've talked about this before with what's happening with poe and others right where they've started to close their data sets off and say come to us directly if you want to sort of use an uh use our llm if you want to experience our data on this and by the way in music like there are all kinds of crazy drive-by lawsuits where people try to you know claim that their song from 30 years ago inspired somebody and ed sheeran just got sued yeah he decided to take it to the mat usually these things settle out of court sometimes people do actually steal people's beats and they settle with them other times people are using chord progressions you know gcd cfg whatever uh g you know whatever just chord progressions that are obvious um and that exists in many songs here's ed sheeran explaining this what did you play for that jury if i was the jury yeah what did you say to them so it was um so my one is um when your legs don't work like they used to before and then there's have i told you lately that i loved you and then um um people get ready there's a train of coming um and then uh what was the looks like we made it look how far we've come my baby and oh she breaks just like a woman i mean there was there was 101 songs that what i was saying is like yes it's a chord sequence that you hear on successful songs but if you say that a song in 1973 owns this then what about all the songs that came before we found songs like from like the 1700s that had similar uh melodic stuff and so what's the url for this uh just so people who know um if people want to play with that test kitchen just type in google's music lm you should be able to find google's ai test kitchen that's the best way to get to it ai test kitchen is the name of where they're putting these experiments yeah it's a ask ai test kitchen dot with google.com so great just type in ai test kitchen you'll find it so they're just sharing all their work now because they were getting their butts kicked by at least in public perception and so google's in the game just generally speaking you feel like google's catching up the gap is closing what do you think yeah the gap is closing you know it's all hands on deck it's a different mentality it's shipped quickly you know all these things that have been in around beta for a long time or private or becoming available um you know they are realizing you know there's a war for developer mindshare As much as there is for sort of the financial side of the world wanting to hear about it, the developer mindshare is really important.
29:12So they've really started to make a big push in terms of engaging with developers and making the tools available so people can play with them and see them. So that's been a distinct shift in the last, I would say, even 30 to 90 days. yeah this hugging face uh dot co just keeps coming up over and over and over again having i guess one some amount of just getting in early and being a place where developers can find the latest stuff but what do we got next okay so we're gonna look at assembly ai i thought this one was really really interesting okay um they they do a you know very interesting job you can basically so if we just let me i'm gonna go to a different window for a quick second because this is one's in progress um so you show up and basically um you can you can drop in a url okay for for any anything with a transcript so i what i did was because it's and it's a good discussion point i took last week's all in pod okay dropped in the url it it it pulls it does a transcription itself and so here's that transcript on the left and then what it does is it lets you create like a q a around it on your own.
30:19And so what I did, I have a couple of examples queued up here. I said, you know, what was Friedberg's position on the AI hearing? And, you know, I think this is a good one. Friedberg's position is that regulating open source AI models will be impractical due to the rise of edge computing and many models already available, making it hard for regulators to track and audit all AI models. That was his position. Yeah, pretty good. Yeah. So you have signed up for this. This isn't available to everybody yet. This is in beta. yeah and uh i signed up and i thought it was really kind of what do they charge to do a transcription um i think if it's under is like 24 a month or something like that i don't quote on the prices yeah oh well i'm looking at it right now audio intelligence is 0.005 per second so i don't know what that is for an hour i'm doing the math right now yeah 4.05 times uh 60 times 60 i don't know what that equals but looks like two bucks an hour if i had to guess i don't think i did the math right there fascinating and so you could ask uh What does Sachs think of Biden?
31:36Okay. Or did Sachs mention Biden or Ukraine? Okay. Then, uh, how does the Sachs mention?
31:46I mean, this is the kind of internal joke. Like, everybody who's on the left, even some people on the right, they play a drinking game. And be careful with this drinking game. They do a shot anytime he says Ukraine or Biden, and you take two shots when he says Hunter. Yeah. if he says the media okay no he did not mention biden or the ukraine says interesting interesting well this is like i'll tell you what's interesting is very uh sometime around last year maybe 18 months ago people started contacting us saying hey we can do all of this week in startups all of all in we'll just transcribe everything and i was talking to nick about like hey i would like to do this for every all 1700 episodes of this week in startups and we had budgeted originally like five hundred dollars per episode to have human transcription then machine transcription was like 150 bucks an episode plus we wanted to clean it up put the people's names in it and thank god we didn't do it because we started this process of doing like 10 episodes and looking at the results and now i think this is all going to get so good that even doing it is a waste of time because it's going to be built into every app in other words when you're on youtube the transcript is going to be so rich and great on youtube i don't know if we should even publish one does that make sense like i feel like this is going to go from 80 to 99 in the next six months in terms of fidelity yeah so i'm just for anyone that's watching i'm basically went to assemblyai.com slash playground v2 and all you have to do go here drop the link i won't do this again and basically you're off processing it um i did another kind of quick summary from here i said does it know the people how does it know the people and separate different people because that is what like we had a service called temi t-e-m-i that we're using for a while we'd use descript and i think we have to identify the people's names is that correct producer nick yes descript is the best one at they actually can identify different voices and then you just name the voices so they identify them and you tag them basically it all it does it automatically it's pretty great i think there should be a thing at the beginning of podcasts when you're doing this where you say hi this is jason calacanis my twitter handle is twitter.com slash jason my website is calacanis.com and then it so if everybody did something like that nick at the beginning of a podcast sunny then when you run it through these it would just go find you and then it would add to the signature hey if you ever hear this voice again attribute it to these two urls it's like a new robot txt yeah it would be like just a way of defining over and over again who somebody is i have to say just in my time working here i started here in 2019 at the beginning of 2019 the transcript the auto transcription world has completely changed it's gone from like a dollar a minute to like this one was what two dollars for an hour and a half of all yeah it went literally from a hundred dollars an hour to two dollars an hour you're exactly right because it used to be rev remember rev.com was the original and they would do human transcription that was their big thing and it was a dollar a minute then temi came out but they were you know not a not perfect and now it's descript and they do every single thing for you and it's a monthly they don't even charge you on a per transcription basis it's just a monthly charge crazy and and then the use of these transcripts is something i've got to get the legality around what i want to do there's somebody who can do this for me i don't know if we can hire a developer to do this for this week in startups i would like to just take the entire this week in startups corpus and normalize it and publish it to coda notion you know publish it to the web um and then i've got a great firm that you should work with the small like ai focused law firm we should we should uh yeah well i don't want to sue anybody i just want to create my own website no but i thought you said you wanted the legality around it so they'll help you no no no i mean the legality what i want to be able to do is come up with a license kind of like grimes did where i'd let people come up with a license that is if you can do whatever you want with the archive as long as you link back to the show page and um yeah it's non-commercial yeah maybe even let people do commercial do interesting things with it let them monetize a little bit rev share just a rev share yeah maybe rev share yeah 50 50 rev share uh do whatever you want 50 50 rev share interesting uh all right let's continue continue on this one so like the next question is a summary of the congress hearing section and i think it does a really good job like going back to your point on like it's identifying what each you know of the host is talking about so sort of what sam's claim is and then you know what chamath jumps on about in terms of what did you say about chamath is it says like right i'm just going to highlight it chamath says sam went further than he initially suggested by proposing licenses for compiling and training models as a form of know your kyc know your customer verification right um sax argues that sam is seeking regulatory capture to maintain open ai's lead by creating red tape um you know friedberg we already got that accurate yeah yeah and so um i thought i thought this was one of the better summaries pulling out each person's position and you know like to your point no one had tagged themselves we've not done any work in terms of identifying uh you know who is who and it really i think it really did a good job here yeah it's fantastic yeah okay great so this is lemur l-e-m-u-r assembly.ai assembly.ai or assemblyai.com yeah so that's the company and that's their product uh very cool what else do you got in the delightful demos that we do here demos on monday demo or die demo or die we usually do them on fridays but like uh at work but now we got mondays as well well i'm just going to cue up the next one so this one was really fascinating um okay there's got to be 20 people doing transcription like stuff like this so i it i wonder if they all just wind up in the same place which is perfection okay everybody who's a transcriber is now out of a job the concept of transcribing is over i don't think it's over yet but you should really level yourself up with these tools is what i think so if you're a transcriber you use these tools and then you have to build a product or service on top of it and the product and service on top of it would be as best i can see i would want somebody to make sure that the names were attributed correctly and any uh you know pertinent links would be added to it so if somebody was talking about assembly ai lemur deep link to it to the google ai sandbox deep link to it right so but of course ai's i bet you could ask ai to go through and find the relevant links and somebody's going to make something that puts hyperlinks into documents that are correct i think so that job too all right everybody you know i've invested in 300 early stage startups and you know having a great name is critical but it's hard to find a great domain name sometimes so people will then compromise their search for a name based on what's available.
38:51And you know what, there's a new domain name I need you to check out. It's called dot tech. Yes, there are now dot tech domain names. And this is the one go to namespace to build anything in tech. A bunch of people are using it. And I even got Jason dot tech, which I'm using for a micro blogging site that I'm coming up with. It's not exactly Twitter. It's kind of like not exactly like Tumblr. It's just going to be Jason dot tech where I give updates all day long about little tiny things that I don't want to share on social but that maybe I want to share with y 'all. So go ahead and get your own dot tech domain name, and you can lock them down.
39:23Now there's tons available one year domain 10 bucks, five year domain 50 bucks. It's a great deal. Now's the time to go get a great dot tech domain name. Again, the word go dot tech slash twist is where you're going to go get these great domain names. You type geo dot T E C H slash twist to lock in a great deal right now you want to do it today, you want to get in there and lock down a great domain name. And listen, if your company's already up and running, maybe you want that for your tech team or to have it as a possible brand extension later on. So if you've got a great brand, you want to get the.tech version of it as well.
39:58Okay, so let's go to this next one. So I'm going to open this one up. This is really cool. Wondercraft.ai. Okay, Wondercraft.ai. And it basically can create a podcast from a bulleted set of points and so in this example i said let's do the twist weekly ai what is the podcast about it's uh the ai a weekly ai podcast discussing and showing the latest in ai products i took three sections from the notion that we normally use here so we talked about uh open ai app perplexity which we haven't got to yet here and then apple bands chat gpt and uh you just put some links in there too uh to the source links in there yep correct and it you know you kind of hit it off and then it's going to generate here we don't we don't have to generate it because it's already generated so here we go so i'm going to play this now make sure the audio comes out so it's generated here's the transcript of generated so all we gave it was a title topics and you know some links and i'm just going to play the little intro here and we'll get started already
41:10corny music welcome to twist weekly ai the podcast that keeps you up to date on the latest developments in ai join us every week as we discuss and showcase the newest ai products don't miss out on any of the exciting breakthroughs in this rapidly evolving field okay it's sufferable that voice is insufferable and now she's going to ask for likes and subscribes oh but but hear me out here it gives you an option to use your own voice oh when you're doing this yeah and i can retire yeah and as i was going through the process here you'll see you know one of the sections was to to choose your own voice and you can do a lot of editing here you can do a speaker you can it's got a lot a lot of choices here you can add your own voice having in julia so it has all these different voices and it's also add your own voice and add your own music which i didn't do i just kind of ran through this relatively quickly so you can go through this process and you can basically uh add your own voice now who's going to use this well one if you're a one man show or one woman show and you're trying to one person show that's yeah probably more accurate yeah everybody's one person show and you're trying to put something together and you don't have the resources that the great resources that twist does uh and you know you basically you do your research you find things that are interesting you prop it up here and you can basically get a podcast going you know quickly and you can use that as maybe not your primary so what i've seen people start to do is people that are publishing newsletters who are already doing the work or using this to create a podcast just as another media for them and you can imagine if you're an enterprise you got something interesting you want to talk about you can do that in there so i'll pause there because this one probably strikes closer to home for you guys so i'll let you react yeah i mean like many things uh you know you want the human's actual reaction we talked about um rick rubin and his reaction to stuff and his taste so what this doesn't have is taste it doesn't have comedic impact it doesn't have the surprise the banter that you and i would have now how much of people tuning into this is for the banter and the personality i would argue significant how much of is it for the news value and the information well podcasts are entertainment therefore and information right so this week in startups all in are giving you information and they're entertaining you it's going to be very hard for this to be entertaining uh authentic and informative but i could see this working really well if it was just reading you stock quotes headlines etc so for something but you know if i listened to it and i didn't know and then i found out i would feel dirty i would feel like i've been tricked so maybe that's just a short-term thing maybe an ai personality is the future and there'll be some version where this joe rogan uh and we talked about this before you know if i could have i don't know i'm trying to pick somebody who's no longer with us johnny carson and i can have johnny carson interview jesus christ or johnny carson interview i don't know somebody who's died recently who's an incredible artist but you know um how are stern live on forever right or biggie smalls you know like so now you've got johnny carson interviewing you know biggie smalls it's like well that's fascinating so there are kind of permutations here that you can't get yep um but i do wonder uh how soon it would be how soon it will be before these are witty and entertaining can an ai entertain you and do you authentically want to be entertained by an ai is i think where we're going to get to so if the ai made you laugh you know there's like a there's a level of can it provide information that's accurate and we all agree it's on the road to being accurate right but it's still delusional sometimes and hallucinations and you know whatever so but and then the next step is can it actually be entertaining and can you bond with uh the personality and i think that's what the movie her was about is can you cross that uncanny valley yep in which you could fall in love with and listen people fall in love with podcast shows they love listen to the show they use that word love could you fall in love with this artificial podcast of course you could and so will that happen of course and will we be competing against you know i don't know of the top 110 years will five or ten of them be virtual sure why not completely possible i actually don't feel i i I feel like there's an artist out there who could make, there'll be an auto GPT or what they call it, baby GPT.
45:58Yeah. Auto baby. There's a few of them now, right? So now let's make it a baby auto GPT where we, we unleash this thing on the world. And we just say, go find interesting people and then create a fake interview with them. And could that be interesting? Yeah, it could. And could you make a custom interview? So I wanted to talk to this person about the topics you're interested in. so i mean the mind kind of can go to weird places you could have you know if i really love johnny carson as an interviewer and he you know or howard sterny retired could i have howard stern having an absurdist interview with dolly you know or i don't know pick a you know an actor who's no longer with us um and that could be fascinating and sure why not this is a good segue to the last thing i just wanted to show quickly and then let's come back to the startup that's building this wondercraft uh a wondercraft.ai i think it's interesting as a podcast host i can tell you a lot of these podcast hosts suck i guarantee you their ai will be better than 50 of podcast hosts in under three years and that's not a really reflection of their ability at wondercraft ai or ai that's just a function of how bad some podcast hosts are yeah and the sheer number that are showing up so exactly yeah well this is an interesting maybe what do you think i didn't get your reaction there what do you i gave my whole analysis there do you think i'm being objective or do you think i'm being precious because i'm a podcaster i think you are but i want to tie it to this last demo then we'll close out on the discussion because precious a little bit okay great fair fair enough you know because uh and so let's kind of tie to this so this is the last really cool demo so this is a combination of stable diffusion control net eb synth and fusion where they took a couple of scenes from cowboys versus alien cowboys and aliens and they replaced current harrison ford with young harrison ford so i was going to play this video here real quickly tell me he's a fool
48:07don't you remember me boy jake she's in a better place all right so your reaction first to that i've never seen this movie before cowboys versus aliens i've never seen it cowboys and aliens i think cowboys and i said when did that i don't know when that film came out i don't know anything about it um i i think harrison ford is kind of my style icon along with daniel craig so this kind of hit both notes for me everybody can obviously see the likeness and the personality and the ruggedness that i do have and i do exude that male rugged confidence uh and style icons uh of my two style icons i like them old i like them young but so it's a bad example for me because i feel like old indiana jones old harrison ford an aging daniel craig is kind of strikes a nerve with me that they get better with age you know like sean connery got better with age i like these guys who don't get a bunch of like botox and they just are like yeah listen it's gonna i'm gonna maybe they do i don't know but yeah they kind of like yeah i'm gonna be old and weathered so i like the weather but it it crosses the uncanny valley the luke skywalker deep fake from uh mandalorian three years ago it didn't some kid fixed it they hired that kid and now i feel like this de-aging stuff is a whole awesome space and uh i would watch you know uh alec guinness is a perfect example who played obi-wan i would watch an old obi-wan series and if they could get alec guinness as a state to secure the bag for 10 million bucks disney plus please do the old adventures of obi-wan that's incredible imagine alec guinness like on adventures as a jedi knight yeah incredible and i heard he did and i you know i just saw a tweet about uh empire strikes back and yeah the deal that he made to be in empire strikes back it was like some rev share deal that netted him like tens or hundreds of millions of dollars which at the time yeah you're talking about alec guinness i think so yeah he played obi-wan in the yeah he got paid a lot he secured the bag yeah yeah um but in empire strikes back he did some weird rev share thing yeah yeah well he george lucas funded with his own money and some loans and the money from the toys he had the sequel rights he funded empire strikes back himself and so that's the other story that's crazy is george lucas just yeah it was like the entrepreneur's entrepreneur and he almost went bankrupt doing empire strikes back because they were filming in like finland or something or norway the hoth scenes and then a blizzard happened and then that i don't know if it was pinewood studios or wherever like a bunch of sets burned out i mean it was like problem after problem after problem there is an amazing tweet thread that has all those details in it it's uh if you're a star wars fan you know all this because he's talked about it over and over again yeah um yeah and i didn't realize but like luke skywalker got injured that's why some of the in the motorcycle motorcycle accident yeah and that's why the tauntaun uh scene he gets his face ripped by the tauntaun yeah and then they put him in the back the tank and they did that because he actually got yeah all right yeah so this was a deal so i think it was uh like guinness right half a day half a day and was paid uh you know quarter point of the film's gross yum yum skis yum yum skis all right okay coming back to it uh just kind of closing out on your question so i think there's a couple of different ways to think about it i think one um we always want to be able to produce more content and we're not capable going back to the sort of the music example so i think giving people the ability to create more podcasts is excellent i think where the technology has to come together is like what i really like to see is a combination of what we just saw with the harrison ford piece combined with a generative ai where someone can take the style and we've seen that in these the music ones that we've seen are like the ai drake where the underlying music and the tempo and the beat is done by someone else and then the voice is replaced.
52:16Now, if those two things can sort of come together where someone can understand the style of J Cal and then, you know, you're willing to do the revenue share and say, Hey, I would really like J Cal to be my world's greatest moderator of my podcast. And, you know, he doesn't have the time to do it, but he's willing to do it through sort of what you talked about, like through his voice, he's made it available through a platform that he has, and it understands all the inflections and it has like a way of coming up with unique and interesting questions i think that would be really great i think that would be sort of and i feel like we'll get there sooner than three years where someone will take all the nuances that you've put together in the you know hundreds of episodes it's got to be more than that it's gonna be thousands of episodes that you have 1700 episodes 1700 episodes and it can come up with sort of all the uniqueness of jay cal as a moderator love it and yeah that that's what you know we'd love to see that if you're doing something like that we'd love to demo that tell me what you think about this harrison ford clip and then me being precious uh about it like do you actually think that these things will replace podcasters i said like i could see in three years you know it would be these folks at wondercraft ai or another or you know even like a generic model uh or you can roll your own would be able to make a host that's better than 50 percent of hosts but i don't think they're going to ever replace the top 25 percent or top 10 percent i think if you're in the top 10 you're a virtuoso you're just so good at it that you're not going to be replaced but maybe there'll be an ai that could crack into the top 10 but i don't think you replace the top 10 what do you think yeah you know i sort of think of content as this like pyramid and at the top it's like you said it's the folks that have really you know tuned into the craft and put the work in and then sort of you know at the bottom of the pyramid you have everything else i think there'll be a lot more at the bottom but i do think we'll see something crack into maybe not the top top of the pyramid but just below that where we're comfortable the cost base is more affordable for someone who wants to listen to it and so i don't think you're being precious but i think i really like what you said earlier you should embrace this and you should kind of lead the charge and say hey let's create you know the world's greatest moderator as an ai for podcasting that can be there to moderate your pods if if you kind of led that as a charge i think then you know then you're not being precious about it i'll get permission right now somebody can train an ai and make me as the world's greatest moderator and you can make the jcal ai totally cool with it and as an experiment uh non-commercial experiment whatever see where it goes and people can then create me interviewing jesus christ i mean i think that's what you gotta that's the benchmark you gotta have me interviewing jesus h christ or who would be another historical figure well i mean i don't think we the jesus christ is maybe a bit more difficult but yeah you don't have video of jesus christ you doing steve jobs there's a great video of you and steve jobs yeah there's a great video of me and steve jobs interacting right yeah do me interviewing steve on the program jesus no no well that's kind of like steve jobs is tech jesus so have me interviewing this okay whoever makes the best version of this i will give a ticket a vip ticket i'll give sidi's vip ticket to all in summit if somebody can create how many minutes 10 minutes or more yeah 10 10 minutes or more whoever creates the best jcal interviewing steve jobs but it has to be now and it has to be it has to hit on current events so he has to talk about about how bad siri is yeah him freaking out about siri me telling him like me criticizing siri that'd be one of the themes um talking about um maybe he could talk about uh ai um and what he thinks the future of it is maybe he could talk about uh what would be a current affairs topic trump him talking about trump we never got to hear his take on trump uh hmm china maybe talk about china the relationship with china i don't know people don't care about his politics yeah it was not as they what's attack they don't attack i think it's talking about br and ar yes me interviewing him at the launch of the augmented reality that's okay that's the scenario okay so there's a 7 500 vip ticket at stake there's a bunch of ai people are racing to make jcal interviewing steve jobs on the launch of the ar glasses and you just throw the ar glasses in there because there's people who have them um does it have to have video two of us i think it does right has to be some video bonus point bonus points bonus points for some video yeah uh or at least images because you just have the audio have images but uh yeah we'll do creative and we'll announce the winner whatever at some point that's a big$7 ,500 prize that's a big one well i mean it's also going to be sold out so those tickets like i got 25 tickets 10 for family yeah one for this contest one for my twitter contest and then i've got for my lps and then i've got my close friends who are now telling me that i need to give them tickets because they're co-hosting this ai segment so uh i have to get you and video ticket apparently i think vinnie's what's the vibe on tickets have you have you what's the vibe in our circle on tickets are people losing their minds or what's going on what's the back channel on the oil and summit uh well you know i think uh everyone it's a it's an interesting spot because i think everyone assumes they have a ticket because i think as much has been said oh i see everybody assumes that they have it everybody in our poker circle seems to have a ticket yes because as much has been said there's 20 see here's the thing there's 25 tickets per bestie i told my wife five six for family immediate family members yeah i have to give like 10 lps my major lps i gotta invite them to the event of course i can't do any of my startup founders uh sorry uh but i can let them buy maybe the generation and then i guess i got some close friends i gotta think about hmm that's a tough one it's a tough one how do you feel about working the badge station being a volunteer, Sonny.
58:32No, the volunteer. Everybody wants to be a volunteer for this thing, too. All right, everybody. We'll see you next time on Sweet Starters. Bye-bye.
From the publisher
This Week in Startups is presented by:
Issuu is the all-in-one platform for creating and distributing beautiful digital content. Get started with Issuu today for free or sign up for an annual premium account and get 50% off when you go to http://issuu.com/podcast and use promo code twist.
Vanta. Compliance and security shouldn't be a deal-breaker for startups to win new business. Vanta makes it easy for companies to get a SOC 2 report fast. TWiST listeners can get $1,000 off for a limited time at http://vanta.com/twist
.Tech domains are the go-to namespace to build anything in tech… and home to the world’s most innovative startups. Secure your .Tech domain today and lock down a 1-year domain for $10, or a 5-year domain for $50 at http://go.tech/TWIST today!
*
Today’s show:
Sunny joins Jason to discuss the latest in AI. They chat about the launch of ChatGPT’s new iOS app and demo Google’s MusicLM (1:24). Then, Sunny tests AssemblyAI’s LeMUR and Wondercraft AI before wrapping with a look at Stable Diffusion’s deep fake capabilities (29:38).
Follow Sunny: https://twitter.com/sundeep
*
Time stamps:
(0:00) Sunny joins Jason
(1:24) ChatGPT launches its IOS app
(4:59) Sunny demos Google’s MusicLM
(8:14) Issuu - Sign up for free or get 50% off an annual premium account by using promo code twist at https://issuu.com/podcast
(9:33) Having taste, and how did Google train MusicLM
(21:02) Vanta - Get $1000 off your SOC 2 at https://vanta.com/twist
(22:09) The legal issues with training MusicLM
(29:38) Sunny demos AssemblyAI’s LeMur
(38:34) .Tech domains - lock down a 1-year domain for $10, or a 5-year domain for $50 at http://go.tech/TWIST
(39:58) Sunny demos Wondercraft AI
(47:18) Stable Diffusion’s deep fake technology
(51:36) Creating more content with AI
*
Read LAUNCH Fund 4 Deal Memo & Apply for Funding
Great recent interviews: Brian Chesky, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland, PrayingForExits, Jenny Lefcourt
Check out Jason’s suite of newsletters: https://substack.com/@calacanis
*
Follow Jason:
Twitter: https://twitter.com/jason
Instagram: https://www.instagram.com/jason
LinkedIn: https://www.linkedin.com/in/jasoncalacanis
*
Follow TWiST:
Substack: https://twistartups.substack.com
Twitter: https://twitter.com/TWiStartups
YouTube: https://www.youtube.com/thisweekin
*
Subscribe to the Founder University Podcast: https://www.founder.university/podcast




