Licensing & automating creativity

AI is increasingly being applied in creative and artistic ways, especially with recent tools integrating models like Stable Diffusion. This is making some artists mad. How should we be thinking about these trends more generally, and how can we as practitioners release and license models anticipating human impacts? We explore this along with other topics (like AI models detecting swimming pools 😊) in this fully connected episode. Featuring: Chris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaShow Notes: Automation and creativity Goodbye, humans: Call centers ‘could save $80b’ switching to AI https://www.theregister.com/2022/09/01/call-center-ai-gartnerDALL-E can now use AI to extend images as a human artist might https://www.fastcompany.com/90783798/dall-e-image-generator-now-goes-beyond-the-frame -An AI-Generated Artwork Won First Place at a State Fair Fine Arts Competition, and Artists Are Pissed https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissedNation states and AI Undeclared pools in France uncovered by AI technology https://www.bbc.com/news/world-europe-62717599France reveals hidden swimming pools with AI, taxes them https://arstechnica.com/information-technology/2022/08/france-reveals-hidden-swimming-pools-with-ai-taxes-themNvidia says U. S. government allows A. I. chip development in China https://www.cnbc.com/2022/09/01/nvidia-says-us-government-allows-ai-chip-development-in-china.htmlForeign Affairs: Spirals of Delusion - How AI Distorts Decision-Making and Makes Dictators More Dangerous https://www.foreignaffairs.com/world/spirals-delusion-artificial-intelligence-decision-makingResources Open RAIL Licenses https://huggingface.co/blog/open_railNormConf https://normconf.com/Upcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Daniel Whitenack?

  One of the big areas that we're seeing just kind of explode is, I mean, in one sense, AI and creativity, but in a sort of broader sense, like an expansion of the things that AI is doing in a creative sphere that is paralleling, you know, a lot of what humans could do. And particularly, I think the most obvious one that we've seen recently is this DALL-E, Stable Diffusion, all of these models that are sort of text-to-image models. It brings up really interesting questions around the ability of maybe non-artistic people like myself, who I'm not a designer, I'm not a painter,

• 0:45Daniel Whitenack?

  but the things that I could do creatively with these tools is amazing and fun. From my perspective, there's this sort of really fun, positive element about it.

• 1:06Speaker 01

  Welcome to Practical AI, a weekly podcast making artificial intelligence practical, productive, and accessible to everyone. Subscribe now if you haven't already. Head to practicalai.fm for all the ways. Special thanks to our partners at Fastly for delivering our shows super fast to wherever you listen. Check them out at fastly.com. And to our friends at fly.io. We deploy our app servers close to our users and you can too. Learn more at fly.io.

• 1:41Daniel Whitenack?

  Well, welcome to another fully connected episode of Practical AI. This is where Chris and I keep you fully connected with everything that's happening in the AI community. We'll take some time to dissect the latest news in the AI world and dig into some learning resources to help you level up your machine learning game. I'm Daniel Whitenack. I'm a data scientist with S I L International, and I'm joined as always by my co-host, Chris Benson, who is a Tech strategist with Lockheed Martin. How are you doing, Chris?

• 2:12Chris Benson?

  I'm very well today, Daniel. It's been an interesting week in various types of AI news. And I've noticed some themes that will, that we can jump into along the way of some fairly substantial concerns, I'd say.

• 2:25Daniel Whitenack?

  Yeah.

• 2:25Chris Benson?

  The world evolves.

• 2:26Daniel Whitenack?

  Yes. Well, I mean, on a slightly less on topic note, but definitely newsworthy. I don't know if you're going to watch Rings of Power today or tomorrow, or you've already watched it.

• 2:40Chris Benson?

  Not yet.

• 2:40Daniel Whitenack?

  Yeah. So neither of us have watched it. So you're not going to spoil anything for me because I haven't watched it either. But tomorrow, our family, we're gonna have a bit of a marathon and watch through Peter Jackson's Lord of the Rings.

• 2:54Chris Benson?

  Oh, good.

• 2:55Daniel Whitenack?

  And then we're gonna watch Rings of Power. So it's gonna be an all day affair. And I'm sort of basically ignoring anything I'm trying to hear, you know, trying to ignore anything I'm hearing about whether Rings of Power is good or bad, reserving judgment until, until I see it. But.

• 3:13Chris Benson?

  So, yeah, that is, I don't know that we've ever talked about this because it's not directly in the AI world, but that, I grew up as a fanatical Tolkien fan.

• 3:23Daniel Whitenack?

  Right.

• 3:23Chris Benson?

  And not only Lord of the Rings and obviously the High, but the Silmarillion and all the other stories that most people don't ever want to, to read. I've, I do them all. I've read them all multiple times. So I'm, I'm one of those a little bit slightly nut job types.

• 3:38Daniel Whitenack?

  Right.

• 3:38Chris Benson?

  It's my favorite set of stories in the world. So yes, here's the conflict. My wife and my daughter aren't as enthusiastic as I am. So they, they like it, but not like I do.

• 3:50Daniel Whitenack?

  Neither is, yeah, anyone in my life, really.

• 3:54Chris Benson?

  Yeah. So, so the big thing about Rings of Power coming out was like, do I just watch them or do we do it as a family? My wife's like, yeah, we'd like to watch it together, but they don't have the urgency. So now that it was released last night, Daniel, as we, as we are recording this, I am sitting here waiting like, when are they going to be ready to do this? Part of me wants to selfishly just go watch it, but no, no.

• 4:16Daniel Whitenack?

  Right.

• 4:17Chris Benson?

  I'm going to be a good family man.

• 4:18Daniel Whitenack?

  I mean, for someone like you with that sort of background, you're getting this like visual sense of things we haven't seen in other movies, like the Dwarf Kingdoms and such, which is going to be really interesting. I mean, I hope it's really awesome, but it's different. It's definitely different. So there's a lot.

• 4:35Chris Benson?

  They've taken license.

• 4:36Daniel Whitenack?

  Oh, yeah, sure.

• 4:37Chris Benson?

  With the story.

• 4:38Daniel Whitenack?

  Which one would probably expect?

• 4:41Chris Benson?

  I agree. And Peter Jackson, though, he did a great job. He took some license with the story, but he pulled it off pretty well, in my view. They've taken more license at Amazon with this story. So I'm, I'm one of those people who has read the Silmarillion many times and I know all the detail. And I'm going to be trying so hard not to let it just irritate the crap out of me. But I'm looking forward to it because of the, the scope and level of quality that it's supposed to have.

• 5:06Daniel Whitenack?

  Yes. Well, speaking of creativity and license and all of those things. I mean, I think, and this is one of the topics you brought up to me. One of the big areas that we're seeing just kind of explode is, I would say, I mean, in one sense, AI and creativity, but in a sort of broader sense, like an expansion of the things that AI is doing in a creative sphere that is paralleling, you know, a lot of what, what humans could do. And particularly, I think the most obvious one that we've seen recently is this,

• 5:44Daniel Whitenack?

  both DALL-E, Stable Diffusion, all of these models that are sort of text to image models.

• 5:52Chris Benson?

  Yep.

• 5:52Daniel Whitenack?

  So yeah, there's this, it brings up really interesting questions around, well, I think really fun things. First of all, really like positive and fun things around like, just like the ability of maybe non-artistic people like myself, who I'm not a designer, I'm not a painter necessarily. I mean, I took art in high school, right? But the, the things that I could do creatively with these tools is amazing and fun. And like, that's a really, from my perspective, there's this sort of really fun, positive element about it. I don't know. Do you, do you see that side of things?

• 6:30Chris Benson?

  Yeah. And by the way, I'm looking forward to, you know, you know, going and touring the Louvre at some point in future years and seeing, you know, a Daniel Whitenack AI original, you know, up there. Because you know that's going to happen.

• 6:44Daniel Whitenack?

  Well, what do you think is the attribution?

• 6:48Daniel Whitenack?

  Of these works, like, how should one attribute one of these AI augmented artworks? I guess.

• 6:57Chris Benson?

  So I don't know that I have the answer. And I can see both sides as we talk about this, because I, as we are talking about some of these articles that we've seen, one of them, which was a little, maybe on vice.com, is an AI-generated artwork. One first place in the state fine arts competition, and artists are pissed. So I think that's going to happen a lot. I think the question then becomes if you want to talk about artwork, you have to be specific. I mean, I'm so biased coming from this AI community, but you know, it's a set of tools, you know, just like Adobe Photoshop, you know, was for many years, you know, in the digital space. And now AI is another set of tools.

• 7:40Chris Benson?

  And I personally see it as perfectly legitimate, but there may have to be competitions that specify, you know, and by the human hand, you know, in a non-digital format kind of thing. I don't know. It's not just coming, it's here and it's here in a big way and only more of it.

• 7:58Daniel Whitenack?

  Yeah, and I think to, I mean, we are a practical AI here. So if people are wondering how they can practically do some of these things, one of the examples is DALL-E. You can, you know, request access and use DALL-E, which now includes sort of extra features around in painting and other things. There's also, I think, it's funny, Chris, you know, we talked about DALL-E recently. It almost seems like we've moved on from DALL-E and now stable diffusion is the thing. And, you know, you can see how things, fast things move in the AI world now. At the time we're recording, this stable diffusion is, is the new new.

• 8:38Daniel Whitenack?

  And you can use, there's actually from, from stability, there's a dreamstudio.ai, which is a, you can sign up and do this sort of also AI assisted creativity work within that dream studio app. So if you're wanting to get hands-on in a practical way with these models, it's not that hard at this point. Whether you're doing that on a Hugging Face space or in Dream Studio or in the DALL-E interface, there's multiple on-ramps to this.

• 9:09Chris Benson?

  Yeah, I agree. And it's interesting. I mean, I do think that this is going to open up the spaces in a big way. Can you talk a little bit, because I know we've talked about DALL-E quite a lot over recent episodes. And so listeners probably are fairly familiar with it if they've been with us for a short while at least. Can you talk a little bit about stable diffusion at a high level? Like, what do you know of it's different? I know it may not be your specialty.

• 9:31Daniel Whitenack?

  Yeah, it isn't really my specialty. I think that there are differences with the sort of modeling technique and the data and the way they did it. I think, though, one of the interesting things is with stable diffusion being more, I guess, more of an open approach where DALL-E was sort of released in a very guarded way with the sort of waitlist type thing that we've seen from OpenAI. I mean, like, last time we talked, I think I used just used a public hugging face space to generate stable diffusion output on hugging face. There's a Stream Studio thing.

• 10:14Daniel Whitenack?

  So I think it is a kind of, it's in the same stream as DALL-E and other diffusion models that we've seen. So we're still in this sort of stream of diffusion models. But I think the power of openness and community is part of what has driven the stable diffusion model to just explode and kind of, in some senses, people move on from DALL-E to stable diffusion. So that's a really interesting point in terms of openness and how openness drives adoption of things as well, I guess.

• 10:56Chris Benson?

  Yeah. So to turn a little bit, as we've been talking about this new intrusion, if you will, into artistic output and the fact that the folks that have been doing it the traditional way aren't so keen about it at this point. It has implications in a much larger way than just whether we like the art and whether we care about it being AI-driven. It affects people's livelihoods. It affects jobs and we, this is a recurring theme that we've, that we've talked about over time that I don't think is a thing going away anytime soon. And that's, you know, how these things intrude into livelihood that were, that were previously very human centric and start affecting jobs. So we, there was another article that I ran across in the register.

• 11:41Chris Benson?

  The article's title is goodbye humans, call centers could save $80 billion by switching to AI. And I don't think that's news to anybody, quite honestly. I think that we've seen automation occurring for quite some time. It would be natural to have AI models moving into those situations. It calls out the fact that all of these things are driven by economics. Is AI is ever increasingly present in, you know, the automation of previously human jobs, it's going to be the economics driving that. I don't have anything up in front of me, but I remember in a very different industry, there was the thing we talked about briefly about McDonald's automating their, the food service a while back. So we're going to see it everywhere.

• 12:22Chris Benson?

  So as we do this, I think one of the challenges in front of us are, is recognizing it's hitting almost every industry. It hits medical doctors. You know, one of the earliest things was radiographs, but it's, it's, there's so many other areas, cancer diagnosis. We've talked about many of these. We've talked about obviously airline pilots. We've talked about fast food. This is really hitting the gamut of industry. There's really no safe place.

• 12:48Daniel Whitenack?

  I was kind of surprised. So as you know, Chris, I'm a, my wife and I are vegetarians. So when we're traveling, we have our sort of quick spots where we know there's something that is not meaty. And one of those places is White Castle, which has been interesting over time. I've noticed they've sort of adopt, been early adopters of things like the impossible meats, sort of plant or non-meat alternative meats, which was kind of surprising for me when I did that. But the last, one of the last times I was coming down from Chicago, I was going through the drive-through and now they have like in the drive-through and AI. Assistant that does some of the drive through

• 13:33Daniel Whitenack?

  interactions at certain times of the day. I think like not at all times, but at certain times of the day. And I was just totally, I shouldn't have been shocked, right? Because we talk about this every week, but I was sort of surprised at just the, yeah, I mean, it's, it's there and they're using it. And this is like in rural Indiana, right? Like AI drive through.

• 13:59Chris Benson?

  It's been happening for a while. It's just gradually trickling through every industry. And it's funny that you say that because I had a, I actually went through a Chick-fil-A. Sorry, Chick-fil-A people, because I'm about to say something not nice. But I was in the drive through and they had done this where they have the human teenagers trying to optimize the drive through to handle the traffic and they're, they're walking around.

• 14:20Daniel Whitenack?

  It's like parallel processing. Yeah.

• 14:23Chris Benson?

  Yeah. And coming from, coming from this that we do, I was sitting in the car going, yeah, you're all going to get automated away pretty soon in my head. I mean, I wasn't saying that to them. That's literally what I was thinking in the drive through. I was like, yeah, this isn't going to last as long as you think. But we're seeing that and we're seeing it everywhere. I mean, software developers, we who develop software are now have some tools that are AI driven by GitHub, for instance. That is one step toward that. So where do we go from here, Daniel? Just to, to ask a really hard, giant, global question. At some point, people are going to start to notice.

• 15:00Daniel Whitenack?

  I don't know that I, I have the answer. I think, I think we're still though, at a point where, yeah, there's certain things that fit within the distributions that models know, right? Like we talked about sort of apparent coherence in models recently. And I think there's things that fit within the distribution of what models know about. So there's, there's probably two things. One is like, what are those areas where we don't have good data sets and models don't know the distributions of, of things to, to put together? And those are still very, those will be where we shift sort of human focus. I think, but then also there even where we do have data sets and we're automating,

• 15:44Daniel Whitenack?

  you know, call centers or something there it even stresses more the biases and the openness of data and models because those biases are gonna hit really hard when they start affecting people's lives.

• 16:22Chris Benson?

  So another thing that I noticed this last week was there was a lot of interesting things having to do with AI interactions that had to do with government. And we're seeing that because as we have talked about a number of times, the political, social, cultural, and legal frameworks that we all live within, which are all we have a lot of commonality even across different countries, they still haven't kept up with the technology. And so we have another big question going forward is,

• 16:57Chris Benson?

  what is the role of AI when it comes to the different types of feedback loops that are present in between government and its citizens? And how does that affect? And one thing that was sort of humorous, I suppose, especially if it's not happening to you, was that there were some articles out about in France, they've released a model, not released, but they're using a model in the French government to identify swimming pools from satellite imagery. And apparently, I did not know this, apparently, if you're in France and have a pool, that pool is taxed. And so it affects the tax base and it affects how many. And apparently, there were tens of thousands of undeclared pools in France.

• 17:42Daniel Whitenack?

  How dare they?

• 17:43Chris Benson?

  How dare they not declare their pool?

• 17:45Daniel Whitenack?

  Dare they swim?

• 17:47Chris Benson?

  So apparently there were more than a few French citizens that were a little surprised to receive a tax notification that we're now taxing your pool that they had not previously declared. And you know, it's a little funny to me because I don't have a pool and I'm not in France, I'm in the US. But you know, it does raise the question of such tools could be used anywhere by any government and for pretty much anything that they that they can do with the data set. So that was one thing. What do you think of that? You don't have a pool, do you?

• 18:18Daniel Whitenack?

  It is interesting. I think we were just talking about how AI models are affecting people's lives, but these very practical things about how the model is built show up almost immediately. So I'm in one of the articles that we can include in our in our show notes is talking about these this pool model that at first the system confused solar panels for swimming pools with an error rate of 30%. But DGFFIP, I'm assuming that's the government agency or whatever that's that's doing this, says that it has since increased the the accuracy. So I think that's one of these cases where can a model, I mean, in

• 18:59Daniel Whitenack?

  a very practical sense, outside of the implications of it in a government rule sense, a model can do this, no doubt, but it doesn't do it perfectly. Right? So as a practitioner in the French government, if you were one of those practitioners, how do you anticipate and deal with those biases in the system and your understanding of your data? I would say that if you do some work upfront to sort of understand behavior, also do some human evaluation to understand how the model is behaving, if you analyze the biases in

• 19:37Daniel Whitenack?

  your data set, and then you release in a way that is taking into account those behaviors of your model, then people might still be angry, but they're going to be angry that the government is taxing them for their pool. They're not necessarily going to be as angry as, hey, the government is taxing me. For my pool that doesn't exist and only is thought to exist because I have solar panels on my house. And I'm trying to save the environment, you know, that's probably a different scenario, right? So I don't know the full details of what they did in terms of that side of things. But I think as a practitioner, there's a lesson here that as you release these models

• 20:21Daniel Whitenack?

  and they are influencing people's lives, the very fact that these won't behave in maybe the way that you always assume that they will, that has to be factored in upfront.

• 20:33Chris Benson?

  I agree with you and not only that, but you know, we tend to, as we've been tracking, you know, what's happening in the world over the years, we've been doing the show. And we tend to gravitate naturally toward kind of the new stuff coming out and all that. And what I was struck about this is this is, I don't know the details, but I'm, this looks to me like a, probably a fairly simple, by today's standards, old fashioned, convolutional model, I would guess, you know, that was doing this. I don't think it's necessarily the most cutting edge, uh, deep learning, uh, of 2022. But, uh, and I'm not being critical of those, if we happen to have listeners, I'm, I'm just saying you're, you're being very practical and that you're, you're using a technology that's been around for a while to do that.

• 21:15Chris Benson?

  But with, you know, that it's really mundane, but with some fairly substantial effect, it probably did not cost very much to put this model together, train it and, and get it better in the, compared to many of the things we talk about now. And yet they note that a 30 square meter swimming pool will result in about 200 euros of extra taxes per year. So every pool is another 200 euros that they can do. And apparently that that has amassed something along the order of 10 million euros in new revenue for the French government. So, you know, enough to easily pay for the time of your data scientists that are working on that, plus quite a bit more. So I would not be surprised to see a lot more.

• 21:59Daniel Whitenack?

  Well, and yeah, I wonder too, the, so if you're generating that amount, I'm thinking of, you know, the model that I'm thinking of, Chris, is, I don't know if you had this happen in your town, but the like bird scooters or these electric scooters that like are in cities, right?

• 22:17Chris Benson?

  I've used them.

• 22:18Daniel Whitenack?

  I mean, I can't comment on their whole strategy, but my impression of their strategy was we're going to show up overnight in your town and drop hundreds of scooters on the streets, right? And we'll deal with the legal implications later. So they are prioritizing adoption above the legal implications of the problems that they might see. And I think it worked, right? The companies that did that, they saw the trend coming and they received a lot of adoption upfront and they likely received a bump in revenue versus the legal implications of what they're doing. So maybe in a more pessimistic sense, like one could take the perspective, if you're the French government, well, you say your thing's 70%

• 23:01Daniel Whitenack?

  accurate, well, just release it because we'll get enough revenue that we can take care of any of the legal things that come up from us, from us doing this. So, I mean, that's a maybe a more cynical view of the of the practicalities of it, but it's, but it's a, but it's a business, that's a very much a business decision, right?

• 23:24Chris Benson?

  It is. It's funny, I think I'm usually the more cynical of the two of us in terms of what we bring, but you raised a great point. Speaking of cynical, did you see the Nvidia news over this past week? Nvidia has been having kind of an ongoing, very public conversation with the US government about whether or not they can have A100 and H100 chips in China, which is relevant because there is a little concealed kind of competition in many areas. Economically, militarily, all sorts of implications between China and the West with the US probably being kind of the center of that. And so for for national security

• 24:09Chris Benson?

  reasons, there's this ongoing whether or not they can do that. I won't jump. What do you think of that? I I'm kind of biased. I'm already in the defense industry, so I I have some thoughts on that. But but what do you think? How did that strike you when you when you found out this was going on?

• 24:25Daniel Whitenack?

  I mean, I see a variety of perspectives here, I guess. I know, like, the supply chain stuff is all messed up around the world. And so to some degree, people are just looking for what jurisdictions do I need to partner with business wise to run my business, right? But there are a lot of other implications there to around security things and that sort of thing. And I know in listening to the IRL podcast from Mozilla recently, talking about sort of balances of power that are happening with different countries around the world, that's that's something to to consider as well. I don't know. What did it trigger in your mind?

• 25:08Chris Benson?

  Well, I can see, like you, I can see multiple sides to this. You know, I'm part of this AI community. And I certainly, you know, that part of me goes, well, that's not a very sensible thing to, you know, to go do that. And that impacts a lot of livelihoods and all that. The defense side of me is very aware of that competition. And so as a strategic move that I don't think could last, if it if that were to stand, the Chinese would go and create their own capability, but that might take time. So there is a I can see all sides of the story. I don't know that this particular move, the way it's being done, is is the most effective way for the US to secure its interests in the long term. But hey, I'm not I wasn't at all surprised that that was I mean, right now, we've

• 25:52Chris Benson?

  talked about, you know, multiple shows that AI is a power lever in the world. It is for governments, it is for companies, it is for individuals, it is one of the key ways of filling power vacuums now and changing who has the power. So I think we're going to see all sorts of things all over the place with with, you know, across the way. And you you see the same thing in a non-government sense with between corporations. So and we saw with artists this week, as we talked about a little while ago.

• 26:24Daniel Whitenack?

  Yeah, it is. One of the things I think is interesting, I guess this gets to some of the power dynamics. One of the things that was surprising to me, and I don't know the stats on DALL-E, but I saw on Twitter that Stable Diffusion, the new Stable Diffusion model, they were talking about pricing in terms of the kind of market value of that training. And I'm seeing that tweet from one of the team is that they use 256 A100s in the cloud for 150K hours. So at market price, Stable Diffusion costs 600K. So 600K is by no means a small amount of money for most individual humans,

• 27:09Daniel Whitenack?

  but 600K in a government sense or in a large company sense is nothing, right? It's not even a rounding error. Right, exactly. So I do find it interesting that like this sort of because of how the infrastructure around this is being commoditized and the tooling around workflows and ML ops and distributed training and all of these things is getting better. In some degrees, I see this sort of also a positive shift, I guess, in terms of what's possible maybe for smaller teams that can still like make an impact, release a thing that like takes the industry by storm

• 27:52Daniel Whitenack?

  for a number that's not like something that's restricted to nation states, right? Like before, I think you have these sort of models that were being released where this is out of reach of anyone but big tech and a nation state, right? And I think that it's very interesting to me that something like Stable Diffusion, again, not a small amount of money for like if I'm myself and I'm an independent researcher, 600K is a lot of money. You don't have that in your checking account right now? I wish I did. But it's reasonable to think like, well, if I really wanted to, and maybe I live in the Bay

• 28:34Daniel Whitenack?

  Area, I could sell my house for 600K and train a model, right? There's like a route, there's a route to that sort of money for, or a small amount of seed funding in terms of a startup or something like that.

• 28:47Chris Benson?

  Yeah, if you really are motivated, it's attainable.

• 28:49Daniel Whitenack?

  Yeah, and I realize that also I'm, I'm kind of revealing my Western biases here. That sort of money is not, yeah, is, is not within reach for basically the majority of the world, but I, it is encouraging to me that this sort of innovation can happen from small teams and not, isn't restricted to sort of nation state actors or big tech. So that's, yeah, that's encouraging.

• 29:15Chris Benson?

  It's a great point. That would be expected over time, you know, with all technologies, they come out, they're, they're most expensive and they're most inaccessible. And then eventually, you know, they become cheaper and more accessible and, and we're seeing that happen.

• 30:15Daniel Whitenack?

  Well, Chris, I think there's a number of things that I ran across that are in some way related to the things that we've been talking about in, in one way or another. One of the interesting articles that I think everyone should take a look at is, was posted on the Hugging Face blog in August 31st of this year, 2022.

• 30:38Daniel Whitenack?

  And it's talking about the Open Rail licensing, so toward open and responsible AI licensing frameworks. And I think this is really, really relevant for people, and people should take note because this is the licensing mechanism that was used in the recent releases of the Big Science Bloom model, which is kind of the latest, greatest large language model, and Stable Diffusion, which we've been talking about the whole episode. So you can actually go in this blog post and you can see the links to the Open Rail licenses for Bloom and for Stable Diffusion.

• 31:19Daniel Whitenack?

  And what I think they argue here is, basically, machine learning models or AI models have, in one way or another, been licensed under licenses that are mostly having, have been applied in the past to either software or content. So like, software on GitHub, I might license as MIT or Apache 2, something like that. There's various implications of that. And if I release content, like, set up pictures or something, I could do that under Creative Commons, for example. That, this sort of, or like a book, if I wanted to make it more permissively licensed

• 32:02Daniel Whitenack?

  under Creative Commons, something like that. And over time, people are like, "Well, which of these do I use for models?", right? Because a model is sort of like an asset. It's files, and in some way it has relations or similarities to content. And in other ways, well, it does have a code, code associated with it, training. There's a model definition, there's config. That's sort of a similarity with code, right? So do one or both of these fit better? And I think there's additional things with models that maybe don't fit either one of those cases, right? This sort of things around biases, the use of the model, the limitations of the model.

• 32:47Daniel Whitenack?

  That's really not in common with code, right? Because in code, generally, if it's deterministic, like you have an API endpoint and it does a thing, and you can look at the code and understand what it does. Like, it does a thing. That's all it does. But with a model, this sort of like behavioral and bias things that aren't in common. I don't know, over time, if you've run across these sorts of licensing things with with models in terms of the fit of sort of data sets, code, models, and what's licensed where?

• 33:22Chris Benson?

  I have. I mean, and I think I think a lot of people in the community at large have. It's a natural question to ask. You and I are fond of saying, you know, you have to have software with your model to get stuff out there. And software is pretty well established, as you just pointed out, with the open source models, and you have some choices. And I think they're fairly well understood today. Not so, and even, and especially with Creative Commons and such as that as well. This has been a big ambiguous area for a lot of people, not understanding. So I think it's, I think it's long overdue. Glad to see it came from Hugging Face, because they always put out fantastic stuff, as we're often fond of talking about. How would you describe that, a little bit of the differences, like recognizing bias?

• 34:04Chris Benson?

  Like, how would they approach kind of doing the licensing that accommodates bias for those who haven't had a chance to to look at the article? What's what's different or what's new there they might not have seen in those other categories of open source licensing of some sort?

• 34:19Daniel Whitenack?

  Yeah, so I think that there's commonalities with other open licensing structures or mechanisms that maybe we're used to from the from the AI world. But there's, here, there's two, really, two kind of pieces of what makes this interesting. One is is open. So that's, that's like one aspect of it, open access. So that's like an access thing. But then the other side of it is responsible use. And that's really kind of where this rail component comes in. So they build off of this idea of rail or responsible AI licenses. And these, you can go to the, also the rail, there's a rail page which talks about

• 35:04Daniel Whitenack?

  the rail licensing effort, which is been talked about in an ACM paper, Behavioral Use Licensing for Responsible AI. But basically, these rail licenses say they include behavioral use clauses, which grant permissions for specific use cases and/or restrict certain use cases. So if you think of a model, again, it has similarities with open data or open code, but it also has this behavioral aspect to it. And so what the open rail license kind of does is it grants sort of permissive access and redistribution and that sort of thing, as you might expect with an open license.

• 35:49Daniel Whitenack?

  But then it has these clauses which talk about responsible use. So for, I can give an example here.

• 35:57Chris Benson?

  That'd be good. That's the one thing I'm wondering about is responsible is kind of an ambiguous.

• 36:02Daniel Whitenack?

  It is. Well, and I think that the structure, really the open rail license is, I think, a structure. And the open rail license that you would adopt for your model, maybe you could adopt sort of stable diffusion's open rail model if you're, if you have a similar model. But other models are going to have other implications, right? So you have some flexibility in the responsible use clauses. For the stable diffusion license, I'm just looking here, there's sort of attachment A all the way at the bottom and it talks about use restrictions. And this is really this kind of clause that, in my understanding, is really important here.

• 36:44Daniel Whitenack?

  And they talk about you agree to not use the model or derivatives of the model. And then they have a bunch of things. So one of those things is, in any way that violates applicable laws, that's okay. So kind of boring. But if you, if you go down a little bit, these are very interesting ones, right? I, you agree to not use the model to defame. Disparage or otherwise harass others. And there's other kind of interesting ones, maybe specific to behaviors that they anticipated use of the model that they wanted to avoid. So there's one that says, "You agreed to not use the model to provide medical advice and medical results interpretation."

• 37:27Daniel Whitenack?

  So obviously you could generate images with stable diffusion, or maybe interpret images with such a model that are medical related, right? The medical imagery is a type of imagery, right? So maybe that's a use that they anticipated and had internal discussions about. And so there's all of these use restrictions or clauses that go into the sort of rail part of the open rail license. So it's still open in terms of access and distribution type of things. But then there's these clauses around responsibility that are explicitly included.

• 38:05Chris Benson?

  That's interesting. Whereas I know you and I are very focused on kind of AI ethical topics and the responsibilities associated with that. I find the explicit call out potentially short sighted in terms of unexpected, you know, outcomes and consequences. So I'll have to go through that in depth after the show and just kind of, and see what I think about some of those ideas and see if I can come up with any that I think maybe weren't what they were thinking, just as a, as a fun exercise.

• 38:36Daniel Whitenack?

  Yeah. Yeah. And I think that that's always a, it's probably, I mean, it's a pre-existing problem with any licenses, right? You can only anticipate so much. I'm reading some of the frequently asked questions from the Bloom Open Rail License. And they have, one of the, one of the explicit frequently asked questions is, "Does the license cover every harmful use case?" And so this is their response. Maybe that's useful for this discussion. They say, "No, we recognize that the list of use-based restrictions does not conceivably represent everything one could possibly do with our work. We focus on use cases which could be feasible for the model." Model at this time.

• 39:19Daniel Whitenack?

  The license is a start by us at exploring how such rail licenses could be used to mitigate harm. And we hope that these first set of provisions can evolve into more comprehensive provisions over time with community engagement. So I think they also recognize this explicitly, right? You can never anticipate everything. I think in some of the discussions we've had on the podcast over time around responsible use of AI and ethics. I think there's this obligation on the developer's part to reasonably try to anticipate harmful uses of what they're doing with the understanding that you're not going to anticipate anything.

• 40:01Daniel Whitenack?

  But at least if you made an effort to anticipate some things, then some of those things are anticipated. And you're not just sort of chucking your thing out into the world and seeing what happens.

• 40:10Chris Benson?

  I think they, for each item, they need to have a little parentheses behind it and says, and the kinds of things that you know I'm referring to.

• 40:16Daniel Whitenack?

  Et cetera, dot, dot, dot, dot. Yeah, that sort of thing. Yeah, so I, I mean, I think this is really.

• 40:24Chris Benson?

  It's a good start.

• 40:24Daniel Whitenack?

  It's a good start. I think it opens a, opens dialogue as well, which we're, on my team, we're thinking a lot about these things in terms of data sets and models that we release. So it's really good to have some sort of open dialogue around these things, I think.

• 40:42Chris Benson?

  I agree. I think, I think if this is a start and the conversation can kind of evolve, as did open source licenses on the software side. That was an ongoing, that, you know, there was a many.

• 40:53Daniel Whitenack?

  Well, there's still not agreement.

• 40:54Chris Benson?

  And there's still no agreement, but it's much more mature than it was, say, early in my career. And so hopefully this is the start of the conversation on the model side.

• 41:03Daniel Whitenack?

  At the end here, I think, I mean, what we've been talking about is hopefully very practical, but we can provide people with a couple of practical learning resources that have, have come across our, our desks over the, over the past week. One that I think fits right into our theme of practical AI and the things that we care about is an upcoming conference. That actually started, I think, out of a, a sort of viral tweet on Twitter called NormConf. This is a completely, as far as I understand, a completely free and online event, but with an amazing list of speakers, like the really, really great list of sort of AI,

• 41:47Daniel Whitenack?

  machine learning, data science, tech speakers. And what they say is, what if there was a conference all about the mundane, behind the scenes, how the sausage is made, middlebrow, unsexy, normcore stuff in the data and ML parts of the tech world? And so that's, that's the goal, to, to sort of talk about all of those things that aren't probably, what is the latest diffusion model? But I tried to train this diffusion model and I'm having trouble with my infrastructure and data and can't make it work, sort of problems. So you can see even people that we've, we've had on the show here are represented.

• 42:33Daniel Whitenack?

  Yeah, are represented in, in the list of, of speakers.

• 42:36Chris Benson?

  And future, future guests for our show are surely coming from this list.

• 42:41Daniel Whitenack?

  I certainly hope future guests, yeah, if you're, if you're, if you're out there and you're speaking at NormConf, let us know, but um.

• 42:47Chris Benson?

  Just reach out to us.

• 42:49Daniel Whitenack?

  This, I think, is a, just a great, focused, practical thing that will be happening, and I hope it continues to happen. But I'm certainly excited to, to listen in and, and see, uh, see the set of content that they're putting together.

• 43:03Chris Benson?

  Me too. I'm all over this, so I'm gonna register right now.

• 43:07Daniel Whitenack?

  Yeah, cool. Well, Chris, I, um, I've enjoyed chatting through things as, as always with you, and I hope that you have a, a good, uh, Labor Day. It's about Labor Day here in the US, so we've got a long, long weekend. But, uh, enjoy the, enjoy the holiday and good to chat as always.

• 43:26Chris Benson?

  You too, Dan. You'll have a good weekend. Take care.

• 43:37Speaker 01

  All right, that is our show for this week. If you dig it, don't forget to subscribe. Head to practicalai.fm for all the ways. And if Practical AI has benefited your life, pay it forward by sharing the show with a friend or a colleague. Word of mouth is the number one way people find shows like ours. Thanks again to Fastly for fronting our static assets, to Fly.io for backing our dynamic requests, to Breakmaster Cylinder for the beats, and to you for listening. We appreciate you. That's all for now. We'll talk to you again on the next one.