What AI Developers Want Artists to Know about AI

Summary

In a fascinating discussion between Proko and an AI industry insider, the podcast explores the disconnect between the world of AI and the art sector. Highlighting the ethical concerns surrounding data collection and copyright issues, the conversation addresses the transformative power of AI in reshaping industries, including art. While AI holds the potential for massive economic shifts and even utopian potential, the transitional period poses significant challenges, with potential job displacement for artists. As the dialogue unfolds, it reveals both hope and caution, urging artists to engage with AI developments and push the technology's direction towards constructive symbiosis.

Highlights

• The conversation highlights the importance of open dialogues between AI developers and artists to foster understanding. 🗨️🎭
• There is a small, influential group of people in AI, predominantly working in San Francisco, who are driving significant change. 🌐💡
• The concern about AI automating artists reflects broader societal fears about automation and job replacement. 🤖🚫
• Current AI models can commoditize tasks, which may lead to a dramatic reduction in costs across all industries, including art. 💰🔄
• Artists are encouraged to engage actively in AI discussions to influence the paths of this technology. 🎨➡️💬
• Establishing ethical norms for AI is challenging but crucial as the technology can dramatically alter creative industries. ⚖️🧩

Key Takeaways

• AI and art worlds need to bridge their communication gap to address mutual concerns and opportunities. 🎨✨
• The AI industry recognizes the ethical issues around copyright and data use, yet struggles with effective solutions. 📚🔍
• AI's rapid growth may lead to a societal shift as significant as electricity, promising both utopian possibilities and serious challenges. ⚡🤖
• The concept of pushing 'the boulder' emphasizes the need to guide AI progress in beneficial directions. 🚀🎯
• Traditional and human-led art still hold irreplaceable value amidst AI advancements, but artists must adapt to collaborate with technology. 🖌️🤝

Overview

The dialogue between Proko and an AI expert brings to light the communication gap between AI developers and artists. The discussion uncovers a mutual lack of understanding about the operational spheres of AI and art, emphasizing the urgent need for more synergetic cooperation between these fields to address the concerns and leverage the opportunities presented by AI technologies.

Throughout the conversation, the narrative dives deep into the ethical considerations of AI, particularly around copyright and data use. It highlights the industry's self-awareness regarding these issues but also its struggles to find effective solutions. The potential for AI to drastically change various industries is acknowledged, with emphasis on the profound transformative impacts akin to revolutionary technologies like electricity.

In reflecting on the potential societal impact, the discussion frames AI's progression as a 'boulder' rolling downhill, increasingly unstoppable, yet steerable. The dialogue positions artists not as powerless bystanders but as capable participants who can and should influence AI's trajectory, turning potential conflicts into collaborative progress for mutual benefit.

Chapters

• 00:00 - 01:30: Introduction and Initial Thoughts The chapter titled "Introduction and Initial Thoughts" begins with a discussion on whether to start immediately or take some time to settle from all the excitement and activities, like restaurant bookings and general hustle. It quickly transitions to a significant observation that there seems to be a disconnect between the AI community and our world. The speaker acknowledges this lack of communication as a major issue and emphasizes the need to bring someone in to bridge the gap. Furthermore, it's pointed out that there is a small subset of individuals, around 200 to 300 people within the AI community, who have the capability to engage in meaningful conversations that could potentially connect the two worlds. The chapter encapsulates the initial thoughts and plans on addressing this communication gap.
• 01:30 - 03:30: Understanding the AI World The chapter 'Understanding the AI World' explores the small and concentrated nature of the AI community, particularly focusing on AGI (Artificial General Intelligence) Labs. The transcript describes how this world is comprised of a small number of people, estimated around 300, with many not directly involved in the work. The AGI Labs are limited to a few geographical blocks in San Francisco, emphasizing the need for open conversation within this concentrated community.
• 03:30 - 06:30: Impact of AI on Artists and Developers The chapter discusses the impact of AI on artists, referencing a video by Steven Zapata. It focuses on the ethical concerns related to data collection in the AI industry. The discussion also provides an overview of the current AI landscape as a prelude to diving into specific ethical issues.
• 06:30 - 09:30: The Potential of AGI The chapter opens with a discussion on artificial general intelligence (AGI) and its implications for the world. The narrator emphasizes understanding the broader scope of AI rather than immediately diving into specifics. They express that the current generation has not yet experienced a transformative technological shift, hinting at the significance of impending advancements in AGI.
• 09:30 - 13:30: AI and Artistic Creativity The chapter discusses the unexpected shift in technology, particularly in AI, and its impact on artistic creativity. It highlights that initially, automating artistic fields wasn't a targeted goal of AI development. Many in the AI world did not foresee AI encroaching into the realm of artists, as creativity was assumed to be beyond the scope of automation. The narrative conveys a sentiment that such advancements were not anticipated even a few years ago, with a particular focus on the complex, intrinsic elements of art that make it challenging for computers to replicate.
• 13:30 - 17:00: The Ethics of AI Data Usage The chapter discusses the underlying technologies crucial for AI advancements, particularly Stable Diffusion, emphasizing their origins in self-driving cars and robotics. It highlights the importance of AI's capability to understand and interpret the world, facilitating tasks in medical imaging and robotics, such as instructing a robot in object recognition.
• 17:00 - 21:00: Economic Implications of AI In the chapter focused on 'Economic Implications of AI,' the discussion centers around the initial efforts in teaching AI to understand concepts, using apples as an example. The process highlighted involves downloading numerous images from the internet and associating these with text, which appears to have been done without a deep understanding of art. The narrative suggests a broader attempt to make AI grasp vast aspects of the world by feeding it vast amounts of data from the internet.
• 21:00 - 28:30: Future Projections and Concerns The chapter titled 'Future Projections and Concerns' explores the philosophical and practical aspects of art creation in the context of AI advancements. It discusses the distinction between traditional art and modern creative practices influenced by artificial intelligence. The focus is on how AI is changing the way artists and designers conceptualize their work, with an emphasis on optimizing new creative processes. The conversation hints at the broader acceptance and integration of AI within the art community, reflecting on both predictive and innovative shifts in artistic practices.
• 28:30 - 37:00: Collaboration and Community Building The chapter titled 'Collaboration and Community Building' touches on ethical concerns in the art world, specifically about using AI to replicate and produce more of an artist's work without their consent. It highlights a general consensus that such practices are unethical. The conversation suggests that people working in AI understand the negative implications of this and recognize it as something to be discouraged.
• 37:00 - 41:00: Concluding Thoughts and Advice for Artists The chapter discusses the perspectives of those involved in AI development, highlighting their sympathy towards artists amidst the AI revolution. It describes the collective effort in the AI field akin to a 'boulder rolling down the hill,' symbolizing a massive change. The AI community aims to guide this change positively, acknowledging that such a transformation will bring both good and negative outcomes.

What AI Developers Want Artists to Know about AI Transcription

• 00:00 - 00:30 do you want to just jump in or do you you want to do you want like to settle down a little bit from all the all the hype restaurants book it over here like literally just like yeah as far as I can tell um nobody from AI world uh has really talked to our world um and right that's a big problem this is why I want to bring someone in yeah and there are basically like 200 300 people in AI world uh that are like maybe capable of sort of like coming and talking about stuff
• 00:30 - 01:00 um like like the world it's just like really small it's like really really small like um yeah like opening eyes maybe like 300ish people um and like not everybody sort of is like directly working on stuff uh well I mean this they're pretty lean I guess um and then sort of there's only like a few other sort of AGI Labs um and almost like like a very significant portion of them are like literally just in like this few Square block radius that I'm in right now in San Francisco um and like basically I think just like somebody needs to like open a conversation uh because yeah so this is
• 01:00 - 01:30 great thank you for coming on um yeah I'm gonna be talking a lot about Steven Zapata's video that came out you watched it um and a lot of artists have that on their mind yeah and you know the ethical side mostly about the data collection side of it right yeah before we get to that um I want to give you a chance to talk a little bit about the AI world because I know in in the emails we've exchanged
• 01:30 - 02:00 um you you discussed AGI um you know artificial general intelligence and just generally what this whole AI shift means for the world because it's important to understand the bigger scope of it before we narrow in on the art AI part of the story right so let's do that let's jump into your first point that you made to me um is that this is uh the people alive today uh have never really experienced a huge
• 02:00 - 02:30 shift in Tech like what's about to happen I don't think anybody really um in sort of AI World um sort of came to work that day and was like let's automate all the way all the artists like but she's like like it's just like don't think like anybody was attempting to do this at all even just like a few years ago um and you asked him like what was like the last thing that would get automated away from AI it's being like Oh like probably artists um like probably there's like intrinsic stuff in there that's like really hard to um you know have like a computer do um and
• 02:30 - 03:00 um like I think it was just assumed that this would be really hard and a lot of the stuff that people were working on um like the the core technologies that sort of led to stable diffusion um are like mostly for like self-driving cars um or other computer vision problems um of just like it's a useful thing that if you're making a self-driving car or you're making like robotics um or you're like making Medical Imaging um that like you should just like know what is in the world like if you want to tell the robot that it should like go pick up the Apple over there um that like the way that you tell the robot to go pick up the Apple over there it's like you teach it what a bunch of
• 03:00 - 03:30 apples are um and then you like teach it that it should like have some concept of like what that apple is and then just like Bradley if you want to do this um like you you just kind of like need to teach it about like everything in the world um and like the best way to do that is just to like download every image on the internet um and then like um just connect the text to every image on the internet um somehow that I think was like mostly done with like no understanding of art at all like like if you talk to many of the folks who are building this um and you ask them like um like what their conception of like
• 03:30 - 04:00 what an artist does they're definitely thinking in the like capital a art um in that like oh like they're much more talking about like a philosophical concept of Art and like they don't really have an idea in their head of like what sort of I don't know like a character designer does uh maybe yeah I don't think it was like a thing in which people were trying to do there's definitely like now that this is totally the case like now like folks are definitely like um sort of optimizing on this um all the way down to things that I think are like just I think sort of agreed upon in um in the AI world and
• 04:00 - 04:30 also in the art world that are just like relatively stunning um like like taking individual artists and then like making the machine that like generates more of their um their art like without asking them and without like that is like obviously um like I don't think anybody would like disagree with you and what's the general like consensus in the AI World about that like are are you guys internally doing anything to kind of discourage each other from doing that stuff oh I think everybody understands that that's an thing to do but yeah no I think um I I think almost everybody who like works on AI is
• 04:30 - 05:00 like relatively sympathetic to the artists like otherwise I don't think I would be like coming on um okay um I think the thing to note is that like from everybody who works on AI um it it's like a game of many many people who are doing this um and so like the boulder is like rolling down the hill um and it is a large change that it's about to happen um and everybody I think in AI is like trying to push the boulder in a direction that is like good um and like positive um and I think like lots of people understand that like if you make a large change like a lot of stuff um is going to be really really good and a lot of stuff is gonna be really really bad um
• 05:00 - 05:30 and to sort of get back to your point before is like AI is not like a like a like the iPhone it's like electricity it's like going to be a much much larger technology um than I think most people in the current world like kind of get at this point right if the news was about like real things and not I don't know whatever is currently I mean I guess the election is real I I would say that's real which is currently happening but like average news on a slow do you say is just like not uh maybe about things that like are sort of core important
• 05:30 - 06:00 things but like if the news was sort of optimized to be tell you what is actually happening like the core thing that it should be telling you that's happening is like oh AI is happening um like that this is probably the largest thing that will like occur in a long time um and the reason that we know that um is because it is a technology that like probably can create Utopia like the sort of definition of AGI um at least my definition of ngi that I've heard sort of repeated and like people are very vague about what this actually means um it is like at some point the AI can do everything that the human can do at like but cheaper um and like more effective
• 06:00 - 06:30 um and that like literally just like means everything like absolutely everything you possibly of that that's the thing that I can do um and if that were to happen it would mean the labor costs of everything would go to zero um and if the labor costs of everything goes to zero all prices of like everything go to zero like like um it becomes you are paying for electricity price um so like consider this microphone right um like somewhere there's like the materials that were gathered for the microphone um they're somewhere like the transportation that was gathered to the microphone the design of the microphone
• 06:30 - 07:00 um and then like all the way to when it is like here um from when it was like some material somewhere um like almost all of the prices paying somebody to do that um and you might say like oh that's like um you know that's important that we like pay people to do that um but like the AGI case is like no no actually it's like bad um because the alternative is everything is free the alternative is like your rent is pennies the alternative is that like you never have poverty um and like when you talk to AI people it's like they have kind of this mentality of two things one is that um
• 07:00 - 07:30 if you are like in it um and you like know what's happening you sort of recognize that this is like the largest thing that will happen in our lifetimes um and on the other hand you also recognize that like it's a really big thing and like electricity it could be like like it could cause problems um it's probably going to be more good than that depending on who us there's all sorts of people who tell you that it's bad or bad than good um but like if you roll the boulder correctly then like it's actually really important that you make it more good than bad like like the the point is to run aside the Boulder and like shift it into the direction
• 07:30 - 08:00 that's like Utopia and not the direction that's like really like shitty and bad when people hear that AI is going to completely change everything it could be Utopia but it could be bad and it's like all right most people will say but but it's so dangerous what if it goes bad we shouldn't let this happen it's like that's impossible right like yeah that's kind of what you're saying is that we don't have that choice anymore we all the only choice we have is to run alongside the Boulder and try to push it
• 08:00 - 08:30 towards the good yeah say you like banned it in the US um like China Will Roll ahead like like China like if you're like wondering why the U.S um just like sort of like you know kneecaps China's chip production it's because like this is this is not like a small sort of tiny problem um like like it's like a nation-state problem um I don't I don't think there's a way to stop the boulder at this point and there's lots of people who are like attempting to stop the Builder and I like I think even they would agree that
• 08:30 - 09:00 they're just like isn't the way to realistically do that the only thing you can realistically do is sort of push it in the direction that you want and the more you sort of like demonize um or that's the wrong term because it like everybody is on your side um like like nobody is nobody nobody in AI is not on your side uh like they are like they genuinely like are trying to they're good people who are trying to do good stuff um and like on what's part of our side though right like and you're talking about when you say your side you mean like artists that or just Humanity
• 09:00 - 09:30 okay I mean like um like artists should probably be other people who are rolling like running aside the boulder um all the folks in AI are also just running aside the boulder like I think there's a perception that you're sort of like pushing the Builder and that's like not exactly it kind of is what's happening but like the way you the way you like push the boulder in the direction that you want to go is you like push the boulder uh that you push it in some direction um but like the boulder has been rolling for a while
• 09:30 - 10:00 um if you see it as like here's the Boulder and like there's like the AI folks and they're like pushing this way and you're like trying to stop the boulder like that is not at all what it's current but it's like AI folks and they're like trying to push it in the direction and like yeah um they're also like willing to listen to you on like which direction you want the boulders to push in um but like if you're trying to like stop the boulder um like the AI folks yeah I think you'll just get run over by the boulder in the same way so the other thing that's to be clear um is like all the programmers that work on this have the same problem like
• 10:00 - 10:30 almost certainly within a few years you're going to get like models that take um like text and spit out the whole codex yeah they're going to replace you yeah like like probably before artists too because like artists go can like fall back on like traditional media um which like requires Robotics and all sorts of other stuff to automaker or something but like before the art gets automated the programmers will totally get automated um okay so so here's the one the group of people that won't be automated that will probably profit the most from this it's the people funding the stuff right and I think this is what
• 10:30 - 11:00 a lot of artists are going to gravitate to here is that like you have people who are going to make a lot of money pushing the boulder towards this and yeah they're listening to what artists want but are like are they really on our side if yeah if it's better for them to just push it forward as fast as possible so that they can make their returns faster you know what I mean so
• 11:00 - 11:30 like some problems of this is I'm not confident that the ATI Labs um at this point will make a lot of money um for the same reason that like electricity companies don't actually make that much money because they all get commoditized away that AGI may not have a thing that prevents it from being extremely competed with other API Labs um so like if you're making AGI somebody else can just make AGI and then they've spent a whole bunch of money and you've spent a whole bunch of money and like there was a lot of money involved um but
• 11:30 - 12:00 because you're competing against each other your margins are really really tiny um and actually where the money flows is the people who are using it um which is currently the case so like for openai right now the customers of openai the people who are using it to do stuff are making way more money um than the a than opening itself and the same with stability and that yeah sure I get it but then you know you I'm not talking about like the employees of the companies I mean like yes the owner right like if if stability or stability what it's like a billion dollar company
• 12:00 - 12:30 now right okay so maybe this is the owner would right I don't know okay so like yeah let's clear it up what does that mean um when a company goes and raises money for the first time or like early on um like the valuation is like mostly fake but then it gets sold it gets sold to Google or whoever is a bigger player and then the owners do get a billion dollars like sort of um unless they get sold for like under the amount right stability is a particularly odd case that I think almost everybody who is in AI is looking
• 12:30 - 13:00 at stability and very confused um so the the question that I think a lot of AI has at the moment is like it is stability a company um like right is it for profit or not it's definitely a for-profit company and they raise us for a profit company um uh are they making profit right now kind of um profit are they making money um uh kind of they have a app on top of stable diffusion that seems to be charging money for it uh whether or not that's like you know making more money than they're spending is the whole question the pitch I guess of stable
• 13:00 - 13:30 diffusion is that they're probably not going to sell to uh consumers probably they're probably gonna sell it to of businesses right and the thing they won't be selling is the AI um the thing that they will be selling is like other software that uh like is helps you use the AI um itself will basically be uh so what is that yeah I mean what does that mean why why is that an important Point um it's me trying to figure out how they're actually going to make money um and if they're going to be a company that like succeeds well they could help
• 13:30 - 14:00 major employers of artists yes and eventually any industry but let's just say artists to replace artists to be much quicker and cheaper and then if you do that that's that's a lot of money if you provide a two of them that is that is a good point um the thing I'm saying though is that like um I don't think stability is the only people who will be able to do that I think like lots of people uh based on the way that they're doing scarier I know I know so like yeah I'm not trying
• 14:00 - 14:30 to like um anything um uh I'm just saying that like um if you're trying to figure out like oh like who's gonna make money here um yeah I'm not convinced that it is stability and I'm also not convinced that it is open AI um I'm not convinced that the people making the AI are actually the folks that make money I'm more convinced that people who using the AI for whatever reason um are the folks that make money um whether that's like the companies on top of it whether it's the users on top of it um like that that seems the companies on top of it aren't yes those
• 14:30 - 15:00 gonna be fun founded and funded by the same people uh no uh they're founded and funded by anybody um like it's somebody in their like dorm room like the folks on top of it it's like anybody can go and use it so like sort of available for anybody to go do uh okay I see what you're saying so so you're saying these models they're made open source so every anybody could just build on top of it but okay I guess one point that uh
• 15:00 - 15:30 Steven Zapata made was that the people making the models have power because they're the ones updating the models so if there's a lot of money in having the next better model sure then they can charge let's say you have a startup that builds on top of something sure do you keep that open source forever or do you see oh we can just stop making the new ones open source and just have an API available from now on right
• 15:30 - 16:00 so um in my opinion for stable diffusion instability right now it would be a very bad business decision for them to do that um because the waste um stability is like most likely to make money is to sell tools that let you use the open source model and as long as the open source model is like really really good um then your tools that plug into that model are going to work and like that's that's like your mode that's how you retain customers is that you have this free open thing um that like hooks into the tools that you're using but if somebody else comes along and they build like a better version um then like your tools don't plug into
• 16:00 - 16:30 their version um and so like now the tools where you're making money from um like that that business dies um and this is like a common Trope So like um like red hat did this with Linux um and like GitHub does this would get um and then like for sale does this with something called xjs um and like there's a bunch of companies that like play this exact strategy and like they have historically never abandoned their open source project because if they did their whole company with that um so I would not suspect that however Steven is totally right um that like other models that like don't have this
• 16:30 - 17:00 business strategy yeah they're like 100 they will they will like almost certainly just like have the better model that's updated in um in their own world the problem that I think they're going to run into is that they're all like effectively equivalent like if you have a perfect version of this one that you just like type in the image or like that you want to generate and it generates the absolute perfect one of anything you could possibly imagine um like those are all interchangeable with each other's which means that anybody who like comes along and does and just builds their own version of it can like come along and like Drive the price down
• 17:00 - 17:30 um and so the the pitch of stability if you were to like make the bolt case for stability is that you like come in um and you give away something totally for free um because you recognize that it's all going to commoditize down to zero like everybody's gonna beat each other to push it down so why not just give it away for free and then make your money somewhere else uh right okay so you said that they're the same people aren't building the tools but then stability is building the tools that's the way they make money so they're building the tools tools for other people who are building tools so this is maybe a problem um yeah yes I see this kind of being similar to
• 17:30 - 18:00 like what Google did right they made Android they made it for free anyone can build a phone and use Android yeah but then eventually they made the pixel phone and now it's like becoming like one of the best and eventually probably they'll just dominate that it's like they're just kind of using that they're allowing other companies to help them improve their thing and then they build their own tool on it and because they are so connected to both they're able to just build the best tool yes so I would I
• 18:00 - 18:30 would agree with that like I think that's correct um I think like you basically nailed that like yes you build this like open source thing and then you build this other thing and it's like tightly integrated and then you just like um your open sourcing like lets you win this other Market yeah I guess what I'm saying is like the um uh I think Steven gets a lot of things right on a lot of things I just don't think this particular one is right um I think like um you don't think that they're gonna make I don't think they will make um stability closed Source at some point I think other people will totally make their own model that will be closed Source but I don't think stability will make their sources yeah he's correct in
• 18:30 - 19:00 that the the companies that work on this will totally have control over their own model um the thing he I think is not thinking about is that lots and lots of companies are going to be working on this um and they are all going to compete against each other um and that's going to like sort of Drive costs down and make more like way more options for it um to the point where stable diffusion is sort of forced to push their thing for free forever um okay that doesn't necessarily actually help anybody though like this is not this is not like the like maybe the good news that uh folks want to hear but like I think maybe um the benefit of
• 19:00 - 19:30 having me on is that I can at least try to tell you real things and not fake things absolutely I just want to ask the question I think everybody's gonna want answered here is the more ethical part of collecting the data um so people are concerned that copyrighted images or used to train the data that now are being used for for-profit tools and the artists are not being properly or are not being compensated at all um
• 19:30 - 20:00 what is your I guess your opinion on that um because I know the law hasn't really settled in any way yet so we don't really know but what's your opinion on this uh I think you're basically right like um like copyrighted images trained to use this and then you can sort of generate other images that are um like uh maybe close and latent spaces is to sort of like correct AI word to use um uh or just you can generate images that are almost identical and like probably should like like if you were to create that image and post it on the
• 20:00 - 20:30 internet like then that would be considered sort of like a copyright violation um okay I think you're basically right like like basically like um I have no idea what the lava stands but like I agree with you ethically that like um training on copywriting images is sort of bad um and like Steven makes this like wonderful point that I think um like like if if I were if I were sort of trying to push this a lot um I would basically say like oh like why why is um why are folks trying so hard to keep uh copyrighted music um out of the training set um and they're not trying very hard to keep copyright images outside of the
• 20:30 - 21:00 train set which I think actually there are a lot of people who are trying to do copyrighted images outside of train set um so I basically like agree with this the thing I would like heavily caution folks on is that like I am sort of here to tell you like like real things and not fake things um and then like I'm like so what's the reality here like the reality is that it doesn't matter like the reality is that like I don't think I don't think the copywriting like you can take out the copyrighted images entirely and I think you can still produce a model that would um generate images um like and so like copyright is
• 21:00 - 21:30 important but the reason that you have copyright is that like you don't like you don't want somebody to generate images that look like yours um like that's that's the like original reason right um and so you can take the copy right away and you can like you can make the law stronger or whatever you want to do um but like I don't think it will succeed um I think you like the reason it can generate a face is because it's been trained on a lot of um like non-copyrighted images that were like um just like photos like just like random random photos dash cam footage like yeah I don't think anyone's concerned about photos though but like a
• 21:30 - 22:00 specific artist's Styles which you've already said you agree that's kind of an thing to do yeah I would say is it possible to get specific Styles without using art that has that style yeah I I would say so it would just happen how is that possible uh my bet is that it would look like a person describing the style um to the point where like they just work with the model um like you would have you had sort of generic art model and then you have like somebody who's like trying to create something that looks like this
• 22:00 - 22:30 um and then they sort of like describe the style okay I mean I guess as an artist who I would not see how it's possible to describe such subtleties of a visual Style with words but not with words oh how would you describe it okay so this is actually the thing that I would like like to try and get uh like the reason why I think like and I would just like throw this out here um like if you want to meet a bunch of AI folks in San Francisco and you happen to be in San Francisco and you're like like I
• 22:30 - 23:00 absolutely but like like sort of open to the world um like I think there should be like a stronger dialogue here um yeah and if you want to but like my sort of selfish interest is that um right now the tools that we're building for programmers are like very much designed by programmers to make programmers more powerful um but like we're just like like I have seen all of the startups who are doing this and there's just like not that many startups that are doing the tools that make artists more powerful you see lots of folks who are just like generate image and make folks who are not artists more powerful but very few startups that are actually like doing that and I think
• 23:00 - 23:30 part of that is that like if you're like there's just like not a lot of like sort of overlap between art land and AI land um and like totally should be so if you are somebody who like wants to do this um like you're trying to build tools like this like please come talk to me like I would love to talk to you because I think this is like actually like a better strategy but that's my suspicion is that you can build that you can build tools that uh like would just like be like appropriate and you can like sort of um give um you can describe things both in words and um in um like and draw I not exactly sure how to do it but from like an information Theory perspective
• 23:30 - 24:00 like the information is there um and you could probably figure out how to train a model to do that um like basically you describe and you sort of tweak it and you tweak it and tweak it um and like maybe you um you can both draw and you can also subscribe um and eventually you can get there okay so the the I guess the answer that I'm hearing here is that the way to get a style without trying it on a bunch of data of that style is to work with artists to train it on that style so work with artists who are cooperating
• 24:00 - 24:30 um yes uh like I think that is a way to do it like there is a world in which AI um goes down that path um and so when I'm saying like oh you should like change the boulder like you should like um push the boulder in a direction I would say artists should be like pushing the boulder in that direction yeah this is my you know what's probably gonna happen where I I agree with you that it's probably a waste of time to try to stop these companies from using copyrighted data although I do also agree with Stephen that it's dangerous to not do anything
• 24:30 - 25:00 because it sets a precedent of like no one's going to do anything so I think it's worth it just to kind of have that pushback but it's in the end it's probably just gonna not do anything um and what's gonna happen is there will be a small group of artists that sell their their data and that is enough data for these models to be trained and that will replace the artists anyway and now what you have just this really small group of people that are getting paid
• 25:00 - 25:30 and you know 99.99 of artists don't get paid because they're not the ones cooperating with we're working with the the AI companies right um that's okay here's here's another point in which I'm going to be more real and less um great great one way to think about AI um and in general sort of automation overall is that you can kind of take like a GDP of the Art Market um and say like okay like here's all the money that is flowing into artists at
• 25:30 - 26:00 the moment um and there's a thought that like Steven mentioned in this video and you sort of just mentioned now um which is that like you're taking um you're like having sort of artists that um are like somehow compensated um in some direction in the product or via like a royalty or something um the issue with that direction is that what automation does is it makes the total value smaller um like say it's like you know like a billion dollars it's probably not a billion dollars probably way more than a billion dollars um but it's a nice whole
• 26:00 - 26:30 number um like the actual value of like in order to generate stuff that would like money being flowed into um into art is smaller um like much much smaller like possibly like in the you know thousands of dollars instead of like um uh well not probably a thousand like probably more like Millions it's like but it's like it's like an order of magnitude um smaller than it was before and so like you can get um so getting an artist a piece of the pie um is not going to be as successful as you might think because it might be much more like when people say oh um you know
• 26:30 - 27:00 the the companies are using my data to sell ads to me and all this other stuff um and I'm part of the product um and so the strategy is that like we will give um Everybody money um in uh like exchange for their data that we're collecting and then you like actually go look at like how much money it's um per person and it's like under a penny um and that is the problem that I think you're gonna like run into is that the the actual value um that uh it exists for every sort of incremental artist um in the training set
• 27:00 - 27:30 um is like not very high um and that is the problem of Automation in general but it's also like the good thing of Automation in the grander scheme it's a bad thing in automation for like individual artists but it's the good thing when applied to like all of the economy um because what it means is that all of the economy become like every product you can possibly imagine becomes like wildly cheaper like becomes to the point where you start to get like houses that are pennies um because like the whole process has been automated away and like that is that is sort of like Utopia Vision that is like counter
• 27:30 - 28:00 intuitive um but it's also like the history of modern society in that like it used to be the case um that it would cost like like a Year's wages or something like that to produce light uh and because like you would need to get this like candles um and you would like uh like the effort to provide a candle was just like a lot um and so there was these luxury goods and then like electricity comes along and they have light bulbs um and like um Energy prices fall and you start to get light for free um like totally free um and like if you had gone back to
• 28:00 - 28:30 somebody in the candles and said like oh this is gonna like automate away um like most of the work of doing Candlestick making like that's like really really like it's like obviously bad uh for uh for folks but like you go today and you say like okay um should we go back to the point in which nobody actually has light um that's like bad um and like in the art particular it's like obviously like yeah there's a boulder it's rolling down the hill yeah I got it okay so you you talk about this
• 28:30 - 29:00 future where everything is pennies where it it could potentially be better yeah um the thing I think that worries me and and a lot of people thinking about this is the transitional period before it becomes this Utopia however that happens yeah there's gonna be people who get incredibly wealthy
• 29:00 - 29:30 because they're providing all of this and a lot of people pretty you know going into poverty because they're being replaced and it's not just artists it's like just industry is getting obliterated are there conversations right now about how to go through this transitional period of everything's becoming slowly cheaper but people's wages are going down much faster than prices are going down because the wages are going down like instantly like you don't have a job anymore right
• 29:30 - 30:00 like what I'm just I don't know do you have kind of an answer to that or sort of um yeah I don't have I I have answers that won't sound real but I think are generally considered real which is I think it's most people assume fast takeoff what do you mean fast takeoff means um like the AI gets better at improving the AI um and that you don't need a research power you just need compute power
• 30:00 - 30:30 um and so you get everything within 10 20 years uh oh really good yeah yeah like like some people have shorter timelines like five years uh like like three to five how is that possible I don't have three to five years but uh because you don't yeah like no how is how could that possibly happen in order to have an industry that's building houses for pennies so much needs to change in the physical world
• 30:30 - 31:00 it's not just about software becoming better it like what how this is what I meant by like when the news is talking about anything that is not AI the news is like okay you remember like it's been there's been points in the past um in which like the phrase like trust the scientists um like you know we trust science has been sort of like rolled out or like you trust experts on particular Banks the experts in AI right now um everybody who is like working on it right now sees the progress understands it the most clearly
• 31:00 - 31:30 um has really short timelines um like there are certainly people who you'll talk to and they'll be like oh yeah like 100 years or something like that um but opening eyes seems to have pretty short timelines from everybody that for everyone I seem to think and I think there's like literally these like charts on the internet like um like uh different different groups and organizations and like when they think it's going to occur and whether or not they think it's going to be good or bad um and like the sort of average case is like 20 35. um 20 35 20 35. within 10 to 15 years there's a few reasons that are pushing this um one is that AI doesn't take off
• 31:30 - 32:00 like a normal technology normal Technologies um need to happen really quickly because they need research to happen um and they need like people to go and like sort of randomly like roam around in the forest of knowledge and like accidentally discover things um well not totally accidentally but you know what I mean like it's like it's sort of random um AI is not that case AI is you you get more compute you make more chimps more chips and like Nvidia is producing like uh sort of 10 to 20 000 a100s a quarter um and like then you just get ADI and like like you
• 32:00 - 32:30 literally just like train it and it is currently training and like most of the models exist already like in the labs like they they're just like unreleased and exist um and models of AGI no no no no no the next better model like the 10x better model like it's just like it exists in the lab and it's unreleased because like you know there's versions and they like they they do a big training run and then they like wait and then they release that one and then they do a big training round and then they wait um and so like yeah um this doesn't actually solve your like initial problem but like
• 32:30 - 33:00 um when we're talking about like like a future um like a future Crazy Town world we mean like like now like like like within the next sort of 10 to 15 years uh and to be fair I I don't actually know so okay um the things that would prove me wrong like the things that would make me skeptical about this is is one that there's like um it's just like really hard to get chips um and like you're compute bounded so like you're you're your speed to AGI is like um based on like how much computer you can power you can get um and right now we just like don't have
• 33:00 - 33:30 enough a100s which are the chips that you need or the sort of comparative chips um there's other ones that you can use but like it's primarily a100 switching these like big chips made by Nvidia um and there's just like not enough of them then when you need chips like in order to build more chips you need like Fabs and like resources and power and just like physical stuff in the world um and so that needs to happen but like if you've been noticing like like like the whole US Government right now has been like you know sort of pushing on like how do we get Fabs in the country and like how do we get uh like like more
• 33:30 - 34:00 um like there's a whole big tips act um and so like that is maybe the thing that would push you to be slower but like really you just need compute um to go back to what you were saying before though um like how do you how do you solve the problem at the moment um of like folks uh my best answer is that I don't know um because again I'm trying to be like more realistic I think a lot of AI people will come in and they will be like oh like oh like Capital art will never die and like I think the best thing that artists have at the moment is traditional art um because you're
• 34:00 - 34:30 basically right um that like the the thing that is harder to do is like physical stuff in the world um that is changing because robotics is getting good um that like all the same advancements for AI and this was actually these this was like the original purpose um like when people were creating these computer vision models they were making it for like um like uh self-driving and self-driving is literally just for self-driving cars but it's also for just like robots in a factory um that sort of world and that's all self yeah self walking uh so that's all getting yeah good um like really good
• 34:30 - 35:00 um and and some of the reasons why this is getting good is it used to be the case that if you wanted to build like Factory um you needed like this robot and you'd like install it and then like it would cost you like ton of money and then you have to like program it for this very specific task um but now you can get like sort of cheaper one that you're just like literally like a human like moves the arm and says like do this and then like do that and then like that's it um and it's just like so much easier to set up and there's like folks working on like sort of humanoid robots um that are just seem to be like far better um than than what you could reasonably do
• 35:00 - 35:30 um but the the thing that most sort of uh to go back to I guess what I was saying before um is the most likely outcome for fast takeoff uh is that you get a model that is good at making another model um that's when I was saying before that's uh the like you know sort of within one or two years you get right text and then it writes all the code um it's basically that like the speed at creating new models um might be increasing um and there's kind of this like Loop
• 35:30 - 36:00 that you can maybe have um which is that you make a model that's really good at making models um and like very quickly improves itself um to unbelievable levels um and like maybe you get like a gpd5 or whatever um that you like go and you describe in gpd5 and you like write so the way TV3 and gp4 and gp5 works these language models um is that like you write the first bit of text and then it writes the rest of it um and so like maybe in like a gb5 256 you write like this is an academic paper describing the cure for cancer
• 36:00 - 36:30 um and it just like generates the cure for cancer because it's like been trained on the internet knows all the things about the entire world and it's just like Drive The Cure from cancer um for like the whole paper um and so you start to get like there is basically what I'm saying here is that there's a few different things that could like maybe cause a fact to take off one is that you can do the same thing for like like uh for new AI models you say like you describe like um uh here is a new revolutionary new AI model uh that is like 100 times faster than what we currently have um and like uh way better at like real world stuff and like write the whole paper um and
• 36:30 - 37:00 then it's just like write simple paper um or you get one that's just like sort of like looped into itself and what it's just um it's just like self-learning um and it gets better at the learning process itself um and then another one is that you get uh like stuff we can't really predict but like might occur um which is where you use it on science um and you start to do the like here's the cure for cancer by the cure for cancer thing um and you just like apply that to as many fields as you can um and you start to get maybe new technology that we didn't know that existed
• 37:00 - 37:30 um and that one's like definitely that's more like the risky sort of thing like who knows if that will actually occur but like that's totally happened in the past like in the past we've just like accidentally had these like large technological spikes um and the thing about AI is that like um once you have sort of the tbd4 tb5 db6 um like you get like infinite scientists just as many computers as you can have all of those are now lean your scientists um and so this is this is when people are sort of saying hey we're about to get Utopia 15-ish years like maybe
• 37:30 - 38:00 longer who knows but like like really really this is like the most important like like Century Of The World um that's why people are saying that and it's because like you don't need do research really like you do you'll totally get new research but you just need compute and we just have never been in a situation like that in humanity the one thing that I think I don't have a I don't have very strong confidence that will happen in a decade though is like there's a lot of legal structure that would need to change for a Utopia and like law just moves so slow
• 38:00 - 38:30 it does um it moves really fast when National Security is up on the table though okay um and uh the thing that maybe should shift your timelines um is that China is also trying to do this yeah um and there's a lot of bad stuff that you could do with sort of an infinite like a bot that can just sort of you know invent new technologies on the Fly um a way a way to think about AI is that you are adding like a trillion new people that don't need resources uh like
• 38:30 - 39:00 it's the it's the Manpower brain power of of like like as many people as you can make copies of um like these sort of you know um people's making the wrong word but like the power of of many many more people um and that's just like such a large thing um that could be used in all sorts of different ways um that there is definitely sort of like like a nation-state race on a lot of this how confident are people in in this Vision though because I mean it
• 39:00 - 39:30 never turns out the way people I mean there's a few people that predict exactly what will happen but then the majority of people are wrong right I think most people aren't sure what will happen I think they're just like the only things people are confident is that the models will get better and really quickly because they like to remember and the numbers like going up and the more computer power like computing power you put in like you can like draw a line and just like you can you just like draw the line and be like oh if we add you know X more computers we get this performance and like that's the only thing people are confident and so like when people are saying like Fast takeoff um they're just
• 39:30 - 40:00 like literally saying like oh like we need X more computers we've bought X4 computers they will be here in like over the next year as they are being built and then like we just look at our graph and we just say like when we have export computers then we will have this model that can do this crazy stuff because and then like what will actually happen is way more speculative um yeah one thing you can do is you can go look on like metaculous um attackers no I don't I it's like the
• 40:00 - 40:30 forecasting website um it's like n-e-t-a-c-u-l-u-s uh I really think it's called metaculous but I've never actually said it out loud um and somebody's corrected me so that's how I would say it I guess what data does it use to predict stuff oh it's like a game for forecasters so um you can compete on it and then the better uh forecaster you are um you sort of get ranked more and so it's like the world's greatest forecasters go on meticulous and they just like predict events and all sorts of stuff so but you can go look at like
• 40:30 - 41:00 the AI stuff and that's like one way to sort of predict things um and like and is it generally that as a group these forecasters are mostly right when you kind of average out everyone's opinions is that kind of what that's the theory but I mean it looks like they have so many forecasts going on right now that close in December it's like they must already know yeah that like we're generally right if you average out our opinions or not or is it just like like you said it's a game that there's like do a few people
• 41:00 - 41:30 win or something they have you know information that's like under NDA and stuff um but like you know when they tell you short timelines like and other people don't and they're the ones driving the boat you kind of just go with the short timelines um that's it though yeah there's lots of stuff that's like pushing it back um you're correct that like there's a lot of legal stuff um honestly my suspicion is that um like construction
• 41:30 - 42:00 of new Fabs and the resources and all the stuff that you would need to build it's just like a lot of it's slowed down by like just generic humans permitting processes in like local governments and stuff like you want to build like a new Fab or something or like um uh it's the same with like Bridges or buildings or anything like uh yeah I mean you were talking about how it's all about compute but you know all of that will require physical computers right and somebody in the world is going to control the land
• 42:00 - 42:30 that has these resources and that has value right that has power not if you do it in space yeah like when you start to get to this like um yeah is that 10 years from now though uh okay so the problem of putting like chips in spaces that uh uh it doesn't conduct like there's nothing for it to uh dissipate that the heat on the chips um so like uh in sort of Earth there's like air and air cannot can can get the
• 42:30 - 43:00 in space that doesn't exist and that's a problem um there's maybe solutions to that um but uh the point is more that it's like an accident like it's like a limit um like you're like approaching some curve of like costs go to zero um and you're not necessarily saying that you're gonna like get to zero um but you're just saying that like whatever costs of everything that exists right now that should just like Follow by like orders of magnitude and we've seen this happen like all sorts of other times like we've seen this happen with like electricity or light I guess um like the
• 43:00 - 43:30 cost of light um we saw this happen with like um just computers in general like computers in general just like went millions and millions and millions of times cheaper um because of other computers um and the thought is that we basically will just have that for everything um because AGI maybe um yeah unclear yeah got it uh one question I forgot to ask you and this is when you were talking about people in the industry typically you know don't want to screw over artists they they want to
• 43:30 - 44:00 be fair and everything so you are leading um the AI Grant right and so you're gonna you're working with a lot of these startups who are going to be using uh data sets that potentially were not collected ethically um are you guys pushing them to collect their own data sets more ethically or if they are using data sets that were collected just for research to not use
• 44:00 - 44:30 those data sets for commercial purposes so I don't know if there are that many startups in our um like in our badge um that are collecting large data sets on the internet um most people like there are there are a few um but most people are um uh using sort of existing models that have already been sort of trained and these existing models were trained on data that was probably collected unethically ah true uh well um yes maybe
• 44:30 - 45:00 um so are you gonna push them to not use those models probably not yeah so that's what I mean is like when you when you said earlier that you know we're on your side it's like sure generally but when it comes to your goals then you will not be on our side right you're gonna just say ah well you know this this startup has their own goal like what I'm saying is that like the side that I would recommend folks like
• 45:00 - 45:30 sort of again like science are really a shitty metaphor um I know I know it is but when it comes to for-profit companies it's not really a bad it's like there is a side it's like my the profit of my company versus the people getting hurt that it is kind of like those are pretty clear sides I guess yeah like what I'm saying though is like if you want artists um to
• 45:30 - 46:00 like continue to like sort of go into the future in the same direction like like probably better tools to make are the ones that like use the existing stuff and then move in the direction um of like art artists and AI work together maybe uh I'm not totally sure if that actually will work um but like that is that is the side to go in um like if you're gonna try and push the boulder in some direction um I'd rather be like real um than like sort of yeah because I actually don't even know if I believe that um I believe believe what well artists
• 46:00 - 46:30 can push the boulder a little bit or no artists can definitely push the Border that's um that's the case okay I'm just not sure um like there's a part of my brain that's just like trying to figure out what products people will use um like you need to like shift the economic incentives um so that um like there's an artist somewhere in the middle like if you want to continue to be in the middle then you need to have an artist in the middle somewhere right and you need like an economic decided for them to be there and the best economics incentives I can think of um are like maybe um the artist can be in the middle somewhere and like they're you know it's
• 46:30 - 47:00 like a procreate app um and it like does a lot of the rendering for you um right and like you go and you draw something and it sort of like um fixes little mistakes here and there like maybe like cleans up your line work um it like makes you a better draftsman um and uh maybe that's the like direction that it goes in um the other um way that it could go in is like what is the art being used for like maybe um the art is like an input to um to like a video game and and actually what happens is that the artist doesn't get
• 47:00 - 47:30 automated but all the programmers do and like all the rest of the video game kids are um automated and then artists can just like make the video game by drawing stuff um and they can draw in like the whole rest of the video game gets made so that's a maybe though I guess whatever that's a 100 no that's that's total I've seen a demo of that like Athletics no I know but but I mean there will also like why do that when when you if if you could just generate the art cheaper and faster without the person well you're also going to generate the rest of the game as well like yeah I know I know I understand that yeah I'm
• 47:30 - 48:00 more focused on the transition period I I'm sorry I'm pretty confident that the future is that like most art is generated by machines and not people um uh I'm not 100 confident about that the numbers I mean like oh sure machines can make so much art so fast it's like impossible for humans to keep up sure my point here is though like in order for artists and AI Engineers yeah and I had
• 48:00 - 48:30 companies to actually work together there has to be uh signs from both sides that we are actually interested in your in working with you sure and one thing that you can do you personally Evan is push your the startups that you work with to treat artists current in in with respect in that
• 48:30 - 49:00 transitional period And if if you don't say if if they are using these models that are trained on on a you know unethically you you know where I'm headed with this right like if you don't know where you're headed with this but the problem like well screw you I can't work with you right like the problem is that like as far as I can tell at the moment um there's not like a lot of companies um that are like are in our batch that are like like their MO is like we're automating with artists I don't think we've like funded anybody who's like we're like automated by the artists I think we're funded folks we're doing all sorts of different things um in lots of
• 49:00 - 49:30 different places and some of them are going to use Sable diffusion um as like a thing but and like some of them up some of the weights in stable diffusion are like four uh we're like totally generated on on like copyrighted images um but like are they using stable diffusion to like automate away like um the character designer like no like no like I don't think we've like funded anybody who's like automated away with artists uh We've totally funded people who are doing stuff close to automating uh like parts of the like redundant jobs uh for some types of um folks in Creative Works um but we have not like uh like redundant stuff with like the um
• 49:30 - 50:00 much more like the co-pilot strategy where like somebody is um it's like a tool rather than like a replacement like as far as I can tell we have like we have not funded anybody who's like automated like that's the thing that isn't to say that we won't actually um like I just want to be like super real about that yeah the problem is at the moment I cannot tell you um who is an air grant um because uh they have to announce themselves um but there are there are tools in here that I think you will see um that are like positive this one in particular that I can think of that I think um almost everybody will see it's like a positive
• 50:00 - 50:30 development the thing I can do from my position is encourage folks to build startups um that push the boulder in the direction that they want to go in like you just need people to do stuff like like you need folks to come along and like build the tool that like um lets the artist and the um the AI work to together like otherwise you're just like not going to get that and it's like not the case that if somebody doesn't build it just like oh somebody else will like um right so they're they're artists just need to create the company that they
• 50:30 - 51:00 want yeah I I would say so like somebody somebody who's like gonna watch uh one of your videos is gonna do that um right right like somebody presumably should um but or I guess what I'm saying is like the wrong thing to do is sort of to dig your heels in the right thing to do is like to pick um like to go and like do stuff um and like make push the boulder in the direction that you want it to go in um make stuff um and like you're much more likely to have an impact on the Future No I get it there's a lot of subtlety in that in pushing the Boulder and it's artists are
• 51:00 - 51:30 hoping that ethical issues aren't ignored because there may be some legal loopholes right that could be taken advantage of which I don't know is that even like should we count on that like when billions of dollars are involved okay so I agree with ethical issues on like copyrighted images like probably shouldn't be in the training side the problem that I'm like would like artists to sort of like pay attention to is that like even if you were to get rid of
• 51:30 - 52:00 those copyright images they don't solve your problem great I know it doesn't solve your problem in the long run but it could it could uh pad the fall so that it's not so hard because if all if we say you know what we cannot use these models that we're trained on ethically we have to recreate it ethically now what that does is it means all that funding that you got now you have to use a portion of that to pay artists to pay people whose jobs are
• 52:00 - 52:30 being replaced in order to create data you can't just use it on computer no I don't think that will actually happen I'm just like genuinely under the suspicion that like you can you can remove all the um like non-copyrighted or sorry you can remove all the copyrighted images um and you can still create a model that generates art well not new Styles though like new style shorts yeah new Styles you still have to work with artists to create those new Styles I'm not totally sure I don't know if that's actually the case oh okay got it see that's why I'm not saying that there will be ways to
• 52:30 - 53:00 work around it but yeah probably like I that's my suspicion um is that you you can get rid of sort of the copyright claim and you still like and a good sign of this is actually there's a lot of folks who are like literally trying to do this and it's like on one hand you can be like oh it's like Optics but on the other hand it's like it's possible that the people training on those data sets like barely even knew that there was copyrighted data in there um often because it was like framed it's like a non-profit or something like the the sets are so big
• 53:00 - 53:30 um that they just sort of assumed that it was reasonable data to work with um and they're just like compiled by this like nonprofit that like you're like oh like it's just a thing and like you know I don't know if it's gonna solve the problem the thing I would like more recommend folks do um is go the other directions but um one thing I want to bring back to though is that when like right now where art is being used um as like inputs to stuff so things like designing um like games or movies um the rest of the movie and the rest of
• 53:30 - 54:00 the game will be generated as well and oh yeah probably the person who's going to be doing that generation is actually the artists like okay say you automate away all the things um and you generate always all the things but the person who is best at doing it is the person with taste and that's the like thing trained artists actually have um that is like the most useful once you generate all the things is that they just have right maybe maybe that's a maybe too because I mean Steven said in his video too that is like we are currently training the I mean actually
• 54:00 - 54:30 that's a question I had is is this true when I use mid Journey yeah and I am putting my prompt in yeah and then I select which prompt I think is best am I training them to pretty much replace the prompter to write good prompts and then choose the best outputs from each prompt are they like using that data to train a new model yeah so like if you were using some application and they're like oh is this good or bad um and then you're like it's got like a little thumbs up and a thumbs down um and you like click the thumbs up like you're almost certainly training um like to to be closer to a
• 54:30 - 55:00 taste well no but more of like mid Journey gives you four options and then you choose which ones you want to upscale which basically you're telling it which of these four is the best one yeah no I would I would guess that is probably the case that you are like training it to make uh produce better images okay so then you're kind of giving it take you're training it on taste the thing you're training it on is to be more accurate to what you like said and that's not the same thing as tasting um kind of when it comes to Art though I mean maybe yeah more accurate but also what prompts are what prompt you put in
• 55:00 - 55:30 is kind of taste there's a little bit of overlap there there's overlap there's overlap yeah like um maybe uh so like right now if you like went on stable diffusion and you were like you know such and such and then you were like 4K art station or like AKA art station that you're like putting these other like sort of spell words and you like summon the name of like some artists who like had no idea that they would be in this training set um yeah and like for some reason that's like just you know what I mean like when you have all that extra stuff at the end yeah is probably going to go in like that that's just like not gonna occur in the future I think like that stuff is
• 55:30 - 56:00 what you're training away You're training away the like the rest of it all the sex for like spell words right and sure that's not exactly the same thing as taste no that that part of it is not that's just kind of like little tail you know decorations that you put on to improve the the output but there is still a major like um the main body part of the prompt that is the creative part that the person is putting in yeah you know that that could be like taste that you know you're using
• 56:00 - 56:30 creative descriptive words that get good results and then the community votes on which ones are the best as well that's taste when Community chooses which ones are the best sure and that's happening on my journey as well I think more when I'm at taste I meant like my bet is that you'll be able to like generate a game and you will like be able to like Define the art style um by like you individually like draw a bunch of images um and then like you create like or like um you do all the sort of initial parts of it and then you let go and you like
• 56:30 - 57:00 describe what you want the game to do um and then it like creates it generates the game for you and then you're like actually I want this part to be different um and then you like describe this a little bit more um and then like oh actually I want this character and I want to have like bigger horns um and then you like show off the horns and you're like I want this aesthetic style to be like this and then you like draw out the aesthetic style um and then like that's what I'm saying is like um that combination of stuff that that's what I'm assuming you will be able to do within like two years three years maybe why would we need some an artist to do that like if if an AI could know what the general public would
• 57:00 - 57:30 want in a game yeah so this is this is the question of AI um the hard to predict thing about AI is that you you keep running into this problem of like um why don't you you just have the AI do that yeah and then you just keep doing that forever um and then you run into like oh wow labor costs are just going to go to zero uh that uh you you just keep going back until everything is just like oh like you just can keep replacing every part um the the real maybe counter to though why don't you just have everything the AI do that is because uh like you want
• 57:30 - 58:00 to sure but that then that's that's a hobby that becomes a hobby yeah and I think what you were saying when you said that artists will be the ones making games I thought I think you meant that they will be paid to make those games I don't know if game industry will um like like I think gaming games will be like generated in like like dollars like the dollar is worth their compute time um right so so basically it eventually everybody will just generate their own game that they want to play yes but the problem with entertainment is like
• 58:00 - 58:30 like if you write a movie and you like tell the movie twist at the end like people actually do enjoy it and it's also like social like you want to know like stuff that other people like so like I don't think actually everybody will generate their own games and stuff like more realistically what's gonna happen it's gonna be a lot more like YouTube um in that like way more people will be able to do stuff that previously only folks could do before but like one totally real thing that will happen is like you will get something that used to be like a Hollywood movie and costs like 100 million dollars or whatever to produce and you're gonna generate it in like 20 seconds of compute time
• 58:30 - 59:00 um yeah and that's almost certainly going to happen within like a few years maybe is it a few years or 10 or 15 years or is that the same thing 10 10 to 15 years is like the AGI thing that's that's people's AGI timelines um the like 100 million dollar movie generated um in uh in like you know 10 seconds or 20 seconds or however long ago time uh like that's that's like like almost there now um
• 59:00 - 59:30 so like you kind of like that's being trained at the moment yeah like by who uh probably a combination of meta they have like video stuff um I know uh or like my my sort of hunch is at um the various folks who are using stable diffusion will sort of like take it and do a hacky version of it so you might have something that like converts each still frame to something um and then probably like uh like a stable diffusion or stability will let go work on video probably like um so there's a bunch of other sort of
• 59:30 - 60:00 labs out there um maybe landscape of AI is there sort of like four important figures maybe depending on who you ask um uh but like there's opening eye which is probably the leader in all the things um openai doesn't really care about art that much they made Dolly but I think it was almost like a a side project for them um like it was a side project spun out of like it's really useful to have uh like something that knows about the world yeah but then they didn't stumble on it and then they leaned into it so I don't it like doesn't matter if it was intentional or not like they at some
• 60:00 - 60:30 point they realized what they were doing and then they went into it like from the Investor's point of view and me hanging out in San Francisco and just like the general atmosphere is no they are not the general atmosphere is they lean out of it well how they named it like did a lot of stuff um like the general vibe that I got was that open AI is working on AGI um and Dolly is a thing yeah they built but like what openai is going to be known for in like you know 20 years is that they built ATM yeah yeah of course this will be uh so oh yeah it's one of them
• 60:30 - 61:00 um the other one is deepmind um I'm not convinced that deep Minds makes stuff that is released to the public ever um I've heard rumors that maybe one day they will um but deepmind has uh so deepmind is like sort of a subsidiary Google um so you can basically when you hear deep bind I guess you can also sort of hear Google um uh they have done a whole bunch of stuff uh and sort of Never released it to the public they just work with governments like what's why do they work on it if they don't they publish papers um and they've released stuff oh that's
• 61:00 - 61:30 not true they actually have released after the public um they just haven't shipped it as a product that you can like buy yeah I see a lot of AI stuff and it's like it usually seems to me like they're just it's this product that they're using uh they're trying to get people to train some other thing it's like check out our AI thing but really we're just training a new thing that dude like the Casper stuff and they like get you to train do you mind is a really particular org within Google okay and why aren't they trying to are they just not trying to make money like well what's the point of it if they don't
• 61:30 - 62:00 release products if you legitimately believe that you are going to make AGI and you're going to make Utopia um like that's infinite value and everything else is just a distraction from infinite value um like everything else is distraction for making ATI um and so for a while lots and lots of the AGI Labs did not release anything because why would you ever release anything that would slow you down from making AGI which you think is about to happen I see um and oh this is going to totally
• 62:00 - 62:30 offend a bunch of people um in that I'm saying that like whatever um but in my opinion this is my vibe My Vibe check on understanding um the sort of folks in uh in AI land is that the sort of core companies that seem to be important um are uh like open Ai and the deepmind um those are maybe like the two leaders in the space um and then uh like stability seems to have sort of this importance because they have a lot of popularity um around sample diffusion um and it's just like I think it's the first thing that let people know the
• 62:30 - 63:00 sort of Open Secret of Silicon Valley that like the AI is here um I think maybe a description is like um right now I would say we are in like fast tech takeoff of like media Generation Um so when we're talking about like Fast takeoff before of um like what fast takeoff might feel like for AGI and the speed at which things are coming out like the just how quickly stuff is advancing and like every you know month or so you see this thing and you think like oh well could never do that and
• 63:00 - 63:30 then like a month later he does it um that is what that is what like Fast takeoff of AI will look like um and it appears that we're probably just gonna have that for like the next 10 years like I think you're just gonna have that so stability is maybe another one and the other one is anthropic um which is like uh the folks that um trying to make the AI do good things and not bad things um uh so they do um they are all of these things are like pseudo research Labs too so like Stephen has this thing in his video um that I think maybe misunderstands
• 63:30 - 64:00 and what all of these things are a lot of these projects are not made by like maybe standard moneyed interests they're made by people who legitimately think they are building Utopia um and I'm relatively confident that they are building um at least something very large that might lead to Utopia um like they are quite confident that they are building ATI um and that's probably true um and when you are that confidence that you are building the most important technology in the world it's just like not your motivator anymore um like when you're really like when
• 64:00 - 64:30 you're really really confident that like the thing that's gonna happen is that you're gonna make all prices go to zero um and you're gonna end up in this world in which like everything is like free you like look over at your kids um and you're like oh my God like if I don't do this like how shitty would it be for me to like not do this for my like future children um it's just like make a world in which they like never have poverty like they are just like like it's just all free uh and they like never have they like don't have cancer because like we cured cancer with it um it's just like you you look at that and then you look at like the sort of bad sides of it and
• 64:30 - 65:00 you say like oh what if this went like really badly um and like what if actually we made this really powerful thing um and it like went really really terribly I and then you look at it and you're like well like how much do I care about like um like making sure it goes right um that it's like really really important to make sure that it goes right um and you look over your kids and like I don't want them to like you know get sort of like destroyed by the like evil API or destroyed by the like um like the sort of foreign State that's like using the ATI in some horrible way um or just like somehow leads to some dystopia and then just if you actually
• 65:00 - 65:30 believe that the money doesn't really matter um and so most of the AGI Labs like all of them uh um though I don't know if stability is really counts themselves as an AGI lab um or just what's the AI labs are research projects they like started in research they like were founded as research projects um they like don't they don't really hire people whose like job is is to like make money I'm I'm the person who like um job it is like to make money I guess that's like the thing that I normally do um it's like to try and go and like figure out how this
• 65:30 - 66:00 stuff makes money um but like from my perspective they don't seem to hire those folks they seem to hire folks that genuinely think they're trying to create Utopia um or trying to prevent some really bad thing from happening didn't they hire you if your job is to figure out I I don't work for any of these Labs um I work right um I am um I am a venture capitalist got it okay yeah yeah yeah these your your Grant is for startups building on top of all this stuff yeah like I I know at the industry well um but I uh uh not I don't work for
• 66:00 - 66:30 any of them but so what I'm what I'm saying is for like Steven's video um is he goes through and he sees that there's like a bunch of non-profits um and I think he has this like assumption that like the reason that they're non-profits is like for this like Eagle reason um and the actual answer is that no they're just like non-profits um like they like straight up boom or existed as nonprofits because like the people making them uh first one of them is like a lot of them are already really wealthy um like a lot of people who just like started this um are like pre-existing debuffing so why do they need to have a hybrid of non-profit and
• 66:30 - 67:00 for-profit if it's legitimately like just a non-profit because they started as a nonprofit um they started as a non-profit so why do they have to add the profit part of it ah because you can't get the capital by the GPS um so it used to be the case um that you didn't think that you needed like lots of money to do this um and then what in 2017 Transformers um like turns out you can just like scale these things forever and then it became the case that this became a capital intensive thing and so people needed to figure out how to create an incentive in which they could like um you know get compute money into it
• 67:00 - 67:30 and so like the reason opening AI so uh one thing about opening eye setup is and to sort of explain open AI setup um is that openai is set themselves up as a venture company um in which the uh the return that you could get as an investor was kept and any money beyond that went to um the um like went to sort of this nonprofit that hopefully does something good with it um and the reason that is is because they legitimately think they're creating Utopia and and I'm not entirely I I don't disbelieve them um and that's not like a like a
• 67:30 - 68:00 malicious thing like if you if you look at this and you work in it it's very hard not to see that you think that this is going to be a very large thing um and you are basically right in that there is a reasonable chance that you generate lots of returns from that they are literally set up so that like you direct the lots of return to somewhere that's actually good the other thing to mention is that he like mentioned this concept of like oh like um the investors can only make so much money and then you like look how much money it is um and it's like uh the gist is like when you invest in a company it like grows and you make some multiple on that
• 68:00 - 68:30 um and uh a common thing for a successful Venture company um is like thousands of times multiples so for example um uh stripe is worth something like 70 billion or and then they like fell down another like 40 billion or something like that um but their original investor uh y combinator uh invested somewhere like around the range of like 2 million um and so you just like run that multiple around and it's like 40 000 multiples or something like that and so like when open AI says that they're gonna only cap you at 100 times multiple well that means that if you put
• 68:30 - 69:00 in your money you can only make 100 times your money that is actually small in Venture World um that is like a small amount of money because the way Venture World works is that um uh you invest in a very very tiny chance that something is going to succeed and you're going to take a really big mac um so you're betting on like space space technology um and open AI is like we're gonna make the biggest spaceship technology we possibly think of we're gonna build the craziest thing um and it's going to be really really valuable but like it's basic technology and it may not work there's my work though yeah it kind of
• 69:00 - 69:30 seems like it's working they put money into a lot of companies most of them fail and then one of them will be 50 000 X that's what pays for all the other money that they lost but if if their investment is capped at 100x it won't pay for all the failures is what you're saying right yes but the folks that invested in opening eye yeah I think they are genuine um actually one one thing about Silicon Valley in general is that people are genuine uh like the East Coast other money on the other side like in Finn like Financial world like who
• 69:30 - 70:00 knows like I don't know I don't really like are going on oh yeah oh yeah this is very strong like East Coast versus West Coast um okay or maybe like uh East Coast you might consider to be sort of like lawful neutral um and then like West Coast Silicon Valley is very like chaotic good um and uh the East Coast looks at the West Coast and thinks um why aren't you lawful and then the West Coast looks at the east coast and thinks why aren't too good Silicon Valley versus New York maybe the reason is that like nobody goes into building HEI um for the money
• 70:00 - 70:30 like you could just go be a Quant uh like you you're doing this because you want to do the largest impact you can possibly have um and if you legitimately think that this is the um the biggest technology that will ever happen um and it's probably the case that it is it's just not the thing that you're concerned about one thing to note is that like I don't think there's actually any benefits to being a non-profit really I think it's like um like I don't think you'd be like that um like I don't know copyright
• 70:30 - 71:00 protection or something like stars as far as I know I mean what Stephen said I don't know if it's true but they're exempt from certain you know being able to collect or not being able to collect some data and use it to train models like if they were not a non-profit they wouldn't be able to collect that data I don't know if that's true um I I would totally be able to be proven wrong but as far as I knew that was um uh if I were running a non-profit and or maybe it's not the non-profit part of it it's the the research lab the
• 71:00 - 71:30 research part of it that yes the the data is being collected for research and I think the example he mentioned was a UK company was it lion lion is is like not a company lion is like an open source project yeah I guess the uh the people that made it it was done by a non-profit research lab right yeah yeah uh many non-profit research Labs that kind of were collectively also funded by um like hugging face and stability and some other folks yeah and were they able to get access to some data because they
• 71:30 - 72:00 were researched nonprofit no I and I don't think so um uh interesting as far as I know that isn't the case but I could be wrong and I um like I could also be wrong about all of this but I'm just gonna throw that out there that I am I'm like trying to say true things and not fake things um as far as I can know because I don't think it's like useful to sort of like yeah I don't know there's a lot of like like AI folks that will kind of come and they'll talk to they'll say like
• 72:00 - 72:30 publicly um these like pressy stuff I guess optixie stuff um of like they don't say what they know or I don't know and I just want the bad press to go away instead of having like a conversation about what's going on I just want to try and steer the direction in a good direction like steer the boulder in a direction um and I don't think that you like accomplished that by sort of like pretending as if the problems don't exist um and like sort of only saying like oh it's like only Utopia until forever um and like it's only good
• 72:30 - 73:00 um I think you have to like you have to like work with folks in in order to make something better and that's kind of why I make this like open claim to like folks in San Francisco like if you are in San Francisco you would like just like like to meet a bunch of AI folks um is it like mostly willing to like sort of meet with you and talk to you and they probably like even like um would like want to meet with you um because they probably aren't like we even reach out to like who wants to talk to us like uh probably literally
• 73:00 - 73:30 the researchers at the companies um it's like like literally just like the individual people and how do we reach out to them if you're in San Francisco um and you would like to contact out I can just like give an email welcome to uh like I am obviously not like um uh infinitely available um but like yeah yeah like you and Steven uh we could like uh like you want to come out San Francisco uh we'd like bring some other folks um uh yeah that'd be pretty cool
• 73:30 - 74:00 there's not that many people actually working on AI uh it's like actually really small um uh and I I don't think um the sentiment in AI is not that like um it's not anti-artist yeah yeah no I get it I think this is a great idea a Meetup makes it it makes everybody human to each other correct right um versus a name on a LinkedIn you know page
• 74:00 - 74:30 CEO of um I think this is a great idea I think we should do it yeah yeah okay cool I'm done I'm down for this I I have a place in mind people to invite cool and we can try this yeah we'll we'll keep talking through emails and I'll get it set up yeah okay I have one final question yeah your your advice to artists was to help push the boulder down the hill in the direction that they want um and the way to do that is to go and make the apps that they want and help to to do all this most artists like
• 74:30 - 75:00 if we're talking about the millions of artists out there they're not going to be able to do that right um they're not going to work on AI and all that stuff they just they want to just draw and paint right what's your advice to those people like how can you pretty much not how can you survive through this transitional period as an artist yeah yeah what's your advice for that like what what can they do anything okay so there's two things
• 75:00 - 75:30 that I would advise them um one is traditional traditional art is probably um like very hard to um to automate away for a long time uh actually maybe a few things the other one is just keep doing what you're doing um yeah I probably should have been more clear about this before um I'm just like not sure yet whether or not uh artists will get automated or artists will automate away everybody else um and this is kind of what I was trying to say with like the um that would be very ironic though um it makes sense if art is currently the thing that's being like AI if I yeah
• 75:30 - 76:00 it's visual like programming is also being aiified right now like like totally like at the same pace and rate as uh uh as art um and then like so this video at the same pace and rate um and so you can kind of assume that like other entertainment media yeah more likely to be generated it's more likely to me at the moment um that uh the sort of that the amounts of people who are currently artists will probably be like better at um you know using the other tools that are available for programming and um or like you know
• 76:00 - 76:30 can automate away the programmer um that can let you like accomplish larger things like create the game and create the movie um instead of like being the input to the game in the movie that you just like create the game in the movie I'm just like more willing to leave but that is the case um yeah the third thing is like um I just actually don't believe the case that um that like millions of artists can't learn how to use AI tools not that they can't use AI tools it's more that they they can't make the apps they can't go in oh yes they can what oh
• 76:30 - 77:00 wait and I guess they can but that's not what they want to do they want to draw and paint right right so you're telling them to basically go get a new job which no I I'm telling you that in like two years programming's automated like I'm telling you that all the other stuff all the other stuff that like you think you can't do at the moment it's like also being automated um yeah like those are occurring at the same rate and one like the programming is actually happening at like a relatively faster rate than art like I'm not convinced that you won't be able to create your own tool just like just in general okay
• 77:00 - 77:30 maybe a better thing to do is just to like like learn about it as much as you can and don't try like learn about it from from people who are doing it I guess if you're learning about it from this like off-hand way um I think you're gonna like no I agree you're not gonna learn like real stuff you're gonna learn like fake stuff it's probably the best option at the moment um is there a resource that you would recommend for for learning about this stuff like I would go try and play with gp3 um that's actually like the most feasible thing
• 77:30 - 78:00 um like we obviously have a thing that's like a we have a product called every prompt which is like um a playground that you can go sign up for free and then like try out the language models if you want uh I'm not here to like promote every product you want to like I mean I I want to know what you're doing um it's fine you can say it but like it's a playground that you can go try out these models um and like one thing you might want to like try it to do with it um it's just like kind of like write you
• 78:00 - 78:30 a story or something um it is like going and playing around things these things like um former Technologies they were like four really technical people um the way that you're gonna use AI in the future to build things and to play around with it is you're going to just use English and or or actually literally any language that you speak really um like you can go in um the large language models and speak in Chinese or Mandarin I guess um and then it also work um and um uh but you're just gonna like talk to it
• 78:30 - 79:00 like a normal person um and technology is going to become much more um like feasible to be used and create and build upon by like regular people um in a way that just like wasn't the case before so with every prompt are you is it just gpt3 at the moment is that the only model that you're using yeah so the gp3 is like a collection of multiple different models um oh yeah so um it's just language at the moment um but you can get it to do a lot of interesting stuff
• 79:00 - 79:30 um and playing around with it and just like seeing what it does um is a way to think about it if gbd3 does not freak you out in the like how good it is uh then the uh I've been playing with it for a few years now it's been kind of kind of crazy yeah yeah the newer models are even good also if you played with it um TV3 before um uh it's like seems to be like quietly updated um so the GB3 of today is like way better than G3 before um uh I'm not 100 sure of that but it's just
• 79:30 - 80:00 like no there's they're launching new things like yeah I jump into the playground and I'm like oh there's a new one I could select uh Da Vinci uh instruct for example it was a new one that did not exist you know last year instruct fine tuning is like maybe the thing that like really took off for um uh teaching people it the explanation of instructors it used to be the case that the way language models worked is like you type the beginning of the text and then it just like completed the rest of it and now the way language models work is that you say like write me a story like literally the words write me a
• 80:00 - 80:30 story and it does it um and uh instead of like right in the beginning so it's just like more usable yeah going and playing around with all of this stuff as early as possible is like genuinely a very useful thing to do um and then also um I've said a bunch of things that are like really firm predictions um like uh that the models will get better um and like it's like very likely that they will get her really quickly and the thing that I think you should keep in mind is uh you don't actually know
• 80:30 - 81:00 whether it's going to be good or bad yet um even for you individually um but like it's just like it's totally possible that you like look back in like five years and be like oh wow like that was just like Photoshop um and like Photoshop actually turned out to be really good for us um and like I don't want to suggest that it is or isn't but like um I can't predict and my job is predicting uh like my job is like I'm sort of trying to guess what the future will be like um and trying to learn that um and so if you think you have a really firm prediction about what's going to happen right now um other than just the models are going
• 81:00 - 81:30 to get better um keep an open mind yeah okay cool what thank you Evan yeah sure really appreciate you coming on um I'm sure people will have a lot more questions and I look forward to continuing the conversation with you and uh meeting up with you and others yeah it'll be a very interesting few years I think it's a really weird views yeah it could be a really really weird views um
• 81:30 - 82:00 or maybe not like it's also very possible all the partners still sound like like literally um that's true there are people who will tell you that just like oh like the current stuff won't scale and like we'll eventually hit some point and we'll plateau and um yeah I don't think that's like the majority opinion the majority opinion is definitely just like infinite returns to scale or just like very large returns to scale um but um yeah that's in the cards okay cool well thank you so much really appreciate you coming on and uh yeah I'll talk to you later sure talk to
• 82:00 - 82:30 you later