Practical AI turns 100!!! 🎉

We made it to 100 episodes of Practical AI! It has been a privilege to have had so many great guests and discussions about everything from AGI to GPUs to AI for good. In this episode, we circle back to the beginning when Jerod and Adam from The Changelog https://changelog.com/podcast helped us kick off the podcast. We discuss how our perspectives have changed over time, what it has been like to host an AI podcast, and what the future of AI might look like. (GIVEAWAY! https://changelog.com/posts/practical-ai-100th-episode-giveaway) Sponsors: Linode https://linode.com/changelog – Our cloud of choice and the home of Changelog.com. Deploy a fast, efficient, native SSD cloud server for only $5/month. Get 4 months free using the code changelog2019 OR changelog2020. To learn more and get started head to linode.com/changelog https://linode.com/changelog. Changelog++ https://changelog.com/++ – You love our content and you want to take it to the next level by showing your support. We’ll take you closer to the metal with no ads, extended episodes, outtakes, bonus content, a deep discount in our merch store (soon), and more to come. Let’s do this https://changelog.com/++! Fastly https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship – Our bandwidth partner. Fastly powers fast, secure, and scalable digital experiences. Move beyond your content delivery network to their powerful edge cloud platform. Learn more at fastly.com https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship. Rollbar https://rollbar.com/changelog – We move fast and fix things because of Rollbar. Resolve errors in minutes. Deploy with confidence. Learn more at rollbar.com/changelog https://rollbar.com/changelog. Featuring: Adam Stacoviak – Website https://adamstacoviak.com/, GitHub https://github.com/adamstac, LinkedIn https://www.linkedin.com/in/adamstacoviak, Mastodon https://changelog.social/@adam, X https://x.com/adamstacJerod Santo – GitHub https://github.com/jerodsanto, LinkedIn https://www.linkedin.com/in/jerodsanto, Mastodon https://changelog.social/@jerod, X https://x.com/jerodsantoChris 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: Enter to win some awesome AI hardware bundles https://changelog.com/posts/practical-ai-100th-episode-giveaway! Sasha Rush on trends from ICLR https://twitter.com/srush_nlp/status/1289311254167924736Episodes mentioned on the show: MLOps and tracking experiments with Allegro AI https://changelog.com/practicalai/97Explaining AI explainability (Darwin AI) https://changelog.com/practicalai/91Practical AI Ethics https://changelog.com/practicalai/96Model inspection and interpretation at Seldon https://changelog.com/practicalai/48Attack of the C̶l̶o̶n̶e̶s̶ Text! https://changelog.com/practicalai/99🤗 All things transformers with Hugging Face https://changelog.com/practicalai/98Achieving provably beneficial, human-compatible AI https://changelog.com/practicalai/85AI for Good: clean water access in Africa https://changelog.com/practicalai/89Exploring the COVID-19 Open Research Dataset https://changelog.com/practicalai/86Artificial intelligence at NVIDIA https://changelog.com/practicalai/15Growing up to become a world-class AI expert https://changelog.com/practicalai/36Helping African farmers with TensorFlow https://changelog.com/practicalai/3Books “Interpretable Machine Learning” by Christoph Molnar https://www.amazon.com/dp/0244768528Upcoming 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?

  Well, I think AGI is a lot of people just think of some type of singularity and Terminator incident where like now we cease to exist because AI has crushed us into the ground and we're no more sort of thing, which I think is to me inconceivable, but whatever.

• 0:17Chris Benson?

  He's so bullish on being inconceivable, though. Like, he's dead set. He's sure of it.

• 0:22Speaker 02

  Challenge him on it, Adam. Challenge him.

• 0:24Chris Benson?

  I still know, man.

• 0:26Daniel Whitenack?

  Like, I just, yeah.

• 0:28Chris Benson?

  I don't know if I agree with that. I'm not on the inside, but this is my own opinion.

• 0:31Speaker 02

  Right.

• 0:31Daniel Whitenack?

  It's at least, I think, inconceivable that it would be sometime in our near future or my lifetime. But to me, it's inconceivable, but.

• 0:41Chris Benson?

  It depends so much, though, on how you're defining that, which I think is important when we define what AGI is, because you can define it a couple of different ways, and I'll give you completely different answers about what my opinion is.

• 0:53Speaker 02

  Being with your change log is provided by Fastly. Learn more at fastly.com. We move fast and fix things here at Change Law because of Rollbar. Check them out at rollbar.com. And we're hosted on Linode cloud servers. Head to linode.com/changelog.

• 1:11Speaker 02

  Linode makes cloud computing simple, affordable, and accessible. Whether you're working on a personal project or managing your enterprise's infrastructure, Linode has the pricing, support, and skill you need to take your ideas to the next level. We trust Linode because they keep it fast and they keep it simple. Check them out at linode.com/changelog.

• 1:36Speaker 03

  Welcome to Practical AI, a weekly podcast that makes artificial intelligence practical, productive, and accessible to everyone. We've teamed up with our friends at Packyderm for an awesome giveaway in celebration of the 100th episode. Check the link in your show notes and enter to win an Nvidia Jetson Nano, Intel's Neural Compute Stick, or Google's Coral Dev Board, plus Practical AI and Packyderm swag.

• 2:02Daniel Whitenack?

  Welcome to another episode of Practical AI. This is Daniel Whitenack. I am a data scientist with SIL International and I'm joined as always by my co-host Chris Benson, who is a Principal AI Strategist at Lockheed Martin. How are you doing, Chris?

• 2:19Chris Benson?

  I am doing very well. I think it is a lovely day to talk about AI.

• 2:22Daniel Whitenack?

  Oh, it is. It is a beautiful day to talk about AI and I was thinking, we've kind of done a lot of these Practical AI episodes, haven't we?

• 2:31Chris Benson?

  I think we have. I think we're actually hitting a milestone possibly at some point here.

• 2:37Speaker 03

  Hey!

• 2:38Chris Benson?

  Hey, who is that?

• 2:40Daniel Whitenack?

  Wait, did we not lock down our Zoom?

• 2:42Chris Benson?

  Oh.

• 2:43Daniel Whitenack?

  Who is this?

• 2:43Chris Benson?

  Who has the password?

• 2:44Speaker 03

  Surprise, guys. We're here.

• 2:46Daniel Whitenack?

  Ah!

• 2:46Speaker 02

  Hey, what's going on?

• 2:48Daniel Whitenack?

  What's going on?

• 2:49Chris Benson?

  Oh my gosh!

• 2:50Speaker 02

  Busting in on this party.

• 2:51Daniel Whitenack?

  Yeah.

• 2:52Chris Benson?

  Oh, it's Jared Santos and Adam Sokoviac of The Change Log, who are kind of our bosses, you know, Dan?

• 2:57Speaker 02

  Oh, gosh.

• 2:58Speaker 03

  Don't please don't call us that.

• 3:00Chris Benson?

  Are we in trouble?

• 3:01Daniel Whitenack?

  Now you gotta be careful what we say.

• 3:03Chris Benson?

  What's going on?

• 3:04Speaker 03

  Well, we're here to say happy 100th episode, guys.

• 3:07Speaker 02

  Yes, what a milestone.

• 3:09Chris Benson?

  It is. Can you believe we've done that many episodes?

• 3:12Daniel Whitenack?

  Yeah, 100 episodes talking about AI. Who would have thought there would be that much to talk about?

• 3:17Chris Benson?

  True.

• 3:18Speaker 02

  Well, podcasting is hard. Doing 100 episodes is harder. So, seriously, congratulations for sticking it out.

• 3:24Chris Benson?

  Yeah, absolutely.

• 3:25Speaker 02

  And not just sticking it out, but like thriving. Congratulations.

• 3:28Chris Benson?

  Thank you very much.

• 3:29Daniel Whitenack?

  Yeah, it's been a lot of fun.

• 3:31Chris Benson?

  It's been a long time. So, listeners who've been listening all the way since the beginning may recall that Adam and Jared actually interviewed Daniel and me in episode number one.

• 3:40Speaker 03

  That's right.

• 3:40Chris Benson?

  As we got going and they finally come back. They've been here all along, but now they're back.

• 3:45Daniel Whitenack?

  Yeah.

• 3:45Speaker 03

  We're back, baby.

• 3:47Chris Benson?

  So, the team is all in line together.

• 3:49Daniel Whitenack?

  We've gone full circle.

• 3:50Speaker 02

  We're suddenly behind the scenes.

• 3:52Daniel Whitenack?

  Yeah, yeah. You're always around. But what was the date of that first episode? 2018, July 2nd. So, it's been a little over a couple of years.

• 4:04Speaker 02

  What's funny, though, is the recorded date is April 20th, 2018. So, I guess we took a little out of ship down.

• 4:09Speaker 03

  First one's the hardest.

• 4:10Speaker 02

  Is that accurate, Jared, do you think?

• 4:11Speaker 03

  Yeah, probably.

• 4:12Speaker 02

  Wow, what would happen there?

• 4:13Daniel Whitenack?

  Maybe we just didn't know how to tag them correctly at that point in time.

• 4:17Speaker 03

  That could be the case as well.

• 4:19Speaker 02

  I don't know.

• 4:19Speaker 03

  Well, there was lots of stuff to figure out, like how to do the intro, what music to use, et cetera, et cetera.

• 4:24Speaker 02

  Mm-hmm.

• 4:25Daniel Whitenack?

  Yeah.

• 4:25Speaker 03

  Plus, we're also launching Other things at the time?

• 4:29Chris Benson?

  Yeah, true.

• 4:30Speaker 03

  Regardless.

• 4:31Chris Benson?

  And we had to figure out if this whole AI thing was just a fad or not, you know?

• 4:35Speaker 03

  That's true. Well, so what's the consensus? Is it a fad? Is it, is it over?

• 4:39Daniel Whitenack?

  Oh, it's definitely a fad.

• 4:41Chris Benson?

  Definitely a fad. Yeah.

• 4:42Speaker 03

  Yeah, it's going away.

• 4:44Daniel Whitenack?

  Yeah, I think it's sticking around. I think, you know, I don't know one episode that we talked about this and maybe I mentioned it a couple of times. I'm kind of thinking of it at this point, like another layer in the software stack. So just like, you know, developers don't necessarily have to be like into all the weeds of DevOps and like C I C D and that sort of thing. But if you're going to be a developer these days, like you're going to sort of interface with DevOps and C I C D at some point. So I'm kind of started viewing it that way where if you're a developer these days, if you're not an AI practitioner, you're going to sort of interface with AI related something at some point.

• 5:26Speaker 03

  What are like the most bucketized or verticalized aspects of AI that are like in products today? I know there's computer vision, there's natural language processing, sentiment analysis. Like, what are the ones that people are using where it's like, yeah, this is kind of a little niche inside AI? This little niche or a product that people are actually putting into play?

• 5:46Chris Benson?

  Yeah, well, those, those are certainly the biggest two right there. And in terms of commercialization, you know, showing up all over the place.

• 5:53Daniel Whitenack?

  There's definitely a lot of areas even of computer vision. So I would say like certain things like facial recognition or object recognition are fairly commoditized in terms of like major, major categories of detection and that sort of thing. So my wife and I just bought a building for her business and it, there's no internet there yet. So we bought these like little deer cam things, like trail cams.

• 6:16Speaker 03

  Oh, yeah.

• 6:16Daniel Whitenack?

  That like take, snap a picture if something walks by. That's my security system right now. And they have an option. It's like AI integrated. Like if a deer walks through the like zone of the camera, it'll identify that it's a deer and like ping your phone versus like a turkey or I don't know, other things that people shoot, I guess.

• 6:35Chris Benson?

  They're great. I use those to catch lost dogs around here because you know, we do all this little animal stuff.

• 6:39Speaker 03

  Ah, interesting.

• 6:40Daniel Whitenack?

  That's a better use, maybe.

• 6:41Speaker 03

  Yeah, I hate to burst your bubble, Daniel, but deer are not dangerous, man. You don't need to be protecting yourself from deer.

• 6:47Daniel Whitenack?

  Yeah, I've even so far caught other things that I wish, wish wouldn't have been in our building, but that's another story.

• 6:55Chris Benson?

  I'm just envisioning the deer coming in from the skylight in the Tom Cruise fashion, you know?

• 6:59Daniel Whitenack?

  Yeah.

• 7:00Chris Benson?

  Dun, dun, dun.

• 7:01Daniel Whitenack?

  Maybe so.

• 7:01Chris Benson?

  Yeah.

• 7:01Speaker 03

  Oh, yeah.

• 7:03Speaker 02

  I should mention my side project then.

• 7:05Daniel Whitenack?

  Yeah, go for it.

• 7:06Speaker 03

  Please do.

• 7:07Speaker 02

  So I live in a cul-de-sac and I've got some cameras outside my house. So I've fed all my neighbors faces into a little system and then I can detect whenever there are people in my cameras that are not those people.

• 7:19Daniel Whitenack?

  Yeah, exactly.

• 7:20Chris Benson?

  Interesting.

• 7:21Speaker 02

  I'm kidding. I want that. I can't make that, but that would be cool.

• 7:24Daniel Whitenack?

  Interesting. Oh, you could totally make that.

• 7:25Speaker 03

  Do your neighbors know you fed their faces into a system?

• 7:28Speaker 02

  Jared's like, dang dude, you're awesome.

• 7:29Daniel Whitenack?

  I thought you were saying you had done it and I was like, I'm not telling them. Keep me quiet. I'm on a cul-de-sac too, but you know, I don't think any of them are listening to this, so.

• 7:37Speaker 02

  Wouldn't that be so cool, though?

• 7:39Chris Benson?

  That would be pretty cool.

• 7:40Speaker 02

  That would be cool.

• 7:40Speaker 03

  Would you have to get your neighbors' permission?

• 7:42Daniel Whitenack?

  I mean, it's not, it's not that hard to, to build. Yeah, you should go for it.

• 7:46Chris Benson?

  It doesn't take that many photos to do it well.

• 7:48Daniel Whitenack?

  So I'm just looking at, for example, like, I mean, Google Cloud, AWS, like all these platforms have different things. So they have like categories, like in Google Cloud, they have site, language, conversation, structured data and cloud AutoML. So site, like these object detection sort of things. Language, of course, there's translation is, you know, very much used these days, but also like transcription, speech recognition, and then there's like conversation type stuff like, you know, chatbots and, like you were saying, sentiment analysis.

• 8:21Speaker 03

  Right.

• 8:22Daniel Whitenack?

  So those are all definitely pre-built solutions for all of those things out there, for sure.

• 8:28Chris Benson?

  Can you tell that Daniel is a natural language processing expert? You know, he just whipped those things right off the top of his head like it wasn't even, he didn't have to think about it.

• 8:38Speaker 03

  Let me flip this question on its head then. What are some untapped areas or some emergent areas where people aren't quite putting this into play yet, but it's going to be big once it reaches developer, the masses?

• 8:51Chris Benson?

  Yeah. First thing that comes to me is adversarial networks, you know, which I know we recently had a, an episode on, but there are so many uses for adversarial situations and, you know, GANs, generative adversarial networks.

• 9:04Speaker 03

  GANs. So what are some uses of GANs that we could look forward to?

• 9:07Daniel Whitenack?

  Deep fakes.

• 9:08Speaker 03

  Deep fakes.

• 9:08Chris Benson?

  That's the most famous.

• 9:10Speaker 03

  Aren't those here, though? Or they're just not that easy to do?

• 9:12Daniel Whitenack?

  I think what Chris is meaning is good applications of GANs, probably.

• 9:16Chris Benson?

  I was going for good applications, but you went right to it, Jared. You went right to the, you know, the nefarious thing there. So and we're certainly, we're certainly seeing that.

• 9:26Daniel Whitenack?

  There's some cool things, though. There's like, people are using these generative networks to like, generate, almost in a data augmentation sort of way. So particularly like, for example, in healthcare, if you need to generate a bunch of example, like, medical imagery of tumors to help you train other models that discriminate between cancer and non-cancer, it's hard to get your hands on that data because it's all, you know, very private, there's simple laws, all those things. So if you have a small amount of it, and then use one of these networks to sort of augment your creation of other data, that can really be a benefit. I'll also say, you know, just because I'm always the, the language-related person, you know, the, even though

• 10:08Daniel Whitenack?

  language and conversation is on this sort of list on the major platforms, you know, almost, like, 98% of languages of the world have no support. So I think also, you know, as the benefits of AI extend to emerging markets in the developing world, there's going to be like, new applications or new challenges in language, but also there's going to be other new challenges that are probably hard to anticipate because most of AI development has been fairly Anglo-centric. So when you're solving problems in other contexts, you know, I think there's cool stuff going on in like agriculture, like even though like we're,

• 10:52Daniel Whitenack?

  a lot of times when we think of AI in the US, we might think of like driverless cars or some like cool tech thing like that. But like, for a lot of the world, like agriculture is, is a big deal. So like, cool new stuff in agriculture or other areas as AI kind of gets, hopefully, democratized and the benefits of it extend to more than just, you know, the US and Europe.

• 11:14Speaker 02

  Is there anything like a Hippocratic oath for those in AI? Because I'm thinking like a gang can be both good and evil, right? So is there, and you think moral compass? When you think, okay, I'm going to be a practitioner or an expert or, you know, somebody doing these things, using these for good. Like, I can see the tumor detection, totally a good reason to manufacture, for lack of better terms, data to support medical, you know, research and whatnot, given the privacy of medical records and individuals and things like that. But like, is there anything out there like a Hippocratic oath for AI?

• 11:48Chris Benson?

  There is a rapidly developing entire professional field that is addressing that, which is commonly called AI ethics or, you know, responsible AI is another term for it. There are several terms that lend themselves to it and they're tied into other terms like explainable AI. And so, yeah, there is definitely the recognition, going back, Adam, to your determination to capture all your neighbor's images and enter them into the.

• 12:15Speaker 02

  I would do that with their knowledge. I will literally go to them with my iPhone and say, this is for my neural network to detect you and not the, in quotes, bad guys or bad people that come into our call to second. Like, so some realness behind that. Our neighbor down the street, like, we live in a decent neighborhood and our neighbor down the street got their tires stolen two days ago.

• 12:34Chris Benson?

  Oh, wow.

• 12:35Speaker 02

  Legit, like, he went out and his truck was on blocks.

• 12:39Chris Benson?

  That's disturbing.

• 12:40Speaker 02

  Yeah. You know what I mean? Like, we have these things happening and I think it's just a result of the economy, honestly. I don't think that people are generally that bad. I think that people get desperate and these are things that happen in desperate times and the economy is definitely suffering. So you have people that are willing to do things that maybe they're not more willing to do otherwise, but I'm not saying thieves don't exist any other day, but I think it might be a normal occurrence or a more common occurrence. And I live in Houston, so it's a pretty well-populated city.

• 13:09Chris Benson?

  Yeah, you know, if you kind of take that condition and you put it in the context of AI, we don't have safeguards. So we have this field that's developing called AI ethics. But, you know, across the world, we really have very little law or regulation that people are required to follow. So for the most part, it is, it is other laws that are not specific to this that have applicability. You know, GDPR is, is one thing that comes to mind in Europe, and that is really the only, you know, national level law that I can think of off the top of my hand. Am I forgetting something, Daniel? I know California has some laws.

• 13:48Daniel Whitenack?

  Yeah, there, there's some other regional things. And I think probably there's some others that exist now. But I think in general, people are still reliant on developing good, like, principles statements and how that trickles down to the actual development workflow, like Google and Microsoft, IBM, all these companies have developed their sort of AI principles. And that may or may not trickle down to the actual development work. Also, I think there's still a perception that like, any sort of, like, when developers hear governance, or like, you know, that sort of thing, it's just an assumption that it's going to slow down all work until, like, we can't do anything. And so there's still like, that feeling, I think, exists there,

• 14:31Daniel Whitenack?

  although people would acknowledge that it's like, there's some important problems to address.

• 14:36Chris Benson?

  You know, what's funny, I was on a call where I was having a conversation with some folks from the, the World Economic Forum about a week ago. And we were talking about this a little bit. And there's so many companies putting out so many principles that there that it's becoming cluttered and stuff. And a comment was made that, you know, there are enough really good examples out there of AI ethics. And maybe, maybe this is a moment where before you go create your own, you do a little soul searching in your organization and go pick one of the 100 great examples that are already out there, just to, to make it a little bit easier for, for people to keep track and stuff. So, I mean, it's definitely a growing field, but well, it has a long way to go to mature.

• 15:17Speaker 03

  It's funny, Daniel, I was listening to your guys' AI ethics episode, which was just recently. And I very much identified with you when you said that, because Chris, you light up on this topic, like, I can tell you're into this. This is like, where you play, thinking at this level. And Daniel feels, as a listener and a producer of the show, I think more like boots on the ground, like, putting The slapping the code against the data kind of a person. And he's like, when I hear governance or I hear these things like, he kind of resonated with what he just said. And I was with him. I'm like, my eyes rolled, my glaze over. I'm just like, oh, here we go again.

• 15:50Daniel Whitenack?

  Yeah. And I think it's what people don't realize though, is, um, and it, I don't think it's immediately obvious that in O'Reilly, um, I forget, I forget who wrote the post. Maybe I can find it, uh, and put it in the show notes, but they wrote a, uh, article about like doing good data science, quote unquote, like ethical data science actually like allows you to do better data science. And I, I think that that like is generally true in the sense that, you know, part of the whole, uh, like governance aspect is like understanding like what data produce what AI model that produce which predictions at what time, right?

• 16:32Daniel Whitenack?

  And if you actually know those things, your development workflow can be sort of supercharged because you're not duplicating your effort. You know, you, you did in the past, you can do some, you know, analysis to figure out like, you know, what parameter space you've explored and what like issue popped up when you did what. And so there is like a lot of benefit there. And I think that's why certain of these practical tools plug in, like we had an episode with, um, Allegro AI recently, and they have open source project called Trains, which helps you track a lot of these things. And there's other projects like Pacaderm, which is an open source project, helps track some of these things. So I think there is developing some tooling around it.

• 17:11Daniel Whitenack?

  It's still not like super streamlined. So there is like, still when you, when you're, you know, on the ground, doing the development, like it's not just like, there's no road bump when you try to integrate these things.

• 17:26Speaker 03

  Right.

• 17:26Chris Benson?

  Everyone has their own, their own approach to it and their own solutions. There's no standardization yet. We're, we're way before that. You know, it's in TensorFlow, but once again, they have, you know, that team has done, um, and I haven't used it, but presumably a really good job, uh, but it's their thing.

• 17:41Speaker 03

  Right.

• 17:41Chris Benson?

  That their community is for. So it'll be interesting to see how this rolls out in the years to come.

• 17:46Speaker 03

  It's kind of like a code of conduct. Similar concept, right? Like you need one, you should have one. If you don't enforce it, it's useless. Right? If you don't have the tooling to help you enforce this, then it's enforce it, then it's difficult to enforce. They're hard to write, so you usually start with somebody else's. But if you don't internalize that somehow, then you just cargo culted somebody else's values and they're not yours. And so there's lots of things there. I'm curious how the state of explainability has moved since you guys started the podcast, because that's one big aspect of this. Right? Like the scary, hard thing about machine learning and whatnot is the fact that you're basically putting data into a black box, training a model, then out comes your response. Whereas there's no, there's not necessarily a policy that created that response. It was a bag of bias or not bias, or whatever you put in.

• 18:31Speaker 03

  With algorithms, we can write an algorithm and you can go back and say, who wrote this? And you're like, I wrote that. Why'd you write that? Well, because my boss told me. Why'd he do that? Oh, here's the policy. This is a bad policy, right? And it had these bad effects with these things. It's non deterministic. You're like, I don't know how we got this result. It's just happened. But as you talked about, there's, there's progress on explainability, which is great. Both as an end user, why are you showing me this ad?

• 18:54Chris Benson?

  Daniel, you, you interviewed Darwin AI recently.

• 18:57Daniel Whitenack?

  Yeah.

• 18:57Chris Benson?

  You know, what, what was your takeaway?

• 18:59Daniel Whitenack?

  Yeah, that's definitely a good episode to, to listen to. There's some, some things happening there in terms of explainability. But I think that there has been some progress. I'm hesitant to say that, like, I feel like there's been a huge amount of progress. There's still a lot of open challenges and open questions. There is more like organized information now, though, I think. I'm thinking of, there's a book which you can actually read online. It's called Interpretable Machine Learning by Christoph Molnar. And you can just read through the entire book and he talks about like all sorts of things from, you know,

• 19:37Daniel Whitenack?

  counterfactual explanations and adversarial examples. And, you know, we just had a previous show on adversarial examples. And so there is like gradually more tooling and organized information out there. But I don't think there's a consensus on this subject either, like how to approach it. But there's just like a series of tools in the toolbox that you can use. And maybe not a, you know, people are still developing those and still actively researching them. And it depends too, a little bit on the like type of data and model and such. So when certain things happen, like, you know, there's like a model that enhances an image and you put in like, you know, Barack Obama's image and then the enhanced image turns out

• 20:21Daniel Whitenack?

  like he's a white guy and there's like, this like blows up on Twitter and stuff. That motivates a lot of like in computer vision, I think they've been struggling with this for quite some time. In other areas like natural language processing, it's probably a little bit newer.

• 20:55Speaker 03

  What up nerds? Jared Santo here, your humble producer. I'd like to tell you about something new. We're beta testing around Practical AI. It's a membership program, which we think could be really valuable for the whole community. We call it Changelog++ and it's the best way to directly support Practical AI and all of the podcasts, videos, and other stuff we create here at Changelog. We have big plans and ambitions for this, but we are experimenting for now to make sure there's interest. That means when you sign up today, you get Practical AI and whatever Changelog shows you listen to now, except no ads. I guess that means this part you're listening to right now, it'll be gone. We also have some extended episodes planned, bonus content, merch store discounts, and a lot of ideas. But since it's such early days, we are offering memberships

• 21:39Speaker 03

  at a 40% discount for early adopters. That disappears at the end of August, so head to changelog.com/plusplus to join today. Lock in that discount, get closer to the metal, and make the ads disappear. Once again, that's changelog.com/plusplus. We'd love to have you supporting us as a member.

• 22:04Speaker 03

  So let's talk about what has changed since that 2018 April to July timeframe, episode one through episode 100. I've been listening, I've been producing alongside. So I do know one thing, I'll cue you up, Chris, because I'd love for you to say this again, or say it to me. One thing you've seen and you've said, which I'm not quite sure I know what you mean by it, is that you think that we're moving into a post-deep learning era, or not post, but beyond deep learning era of AI. Can you say what you mean by that? And explain it to me?

• 22:34Chris Benson?

  Yeah, and it's actually on my mind a lot these days. It's been on my mind today before we started talking about this, you know, before we did the show, because I'm having conversations with other people about the same topic. And to set it up, if you look at these last couple of years, and when we started the show, we were still in kind of the ramp up stage where deep learning, I mean, there were stuff, there was stuff just coming out every day, because everybody was finally setting their focus to it, it was being funded, and we were getting some pretty amazing stuff coming out on a day to day or weekly basis. And, and some of those early shows, you know, Daniel and I, we would struggle to figure out which news items to include in the show and which ones

• 23:16Chris Benson?

  to just not mention, because there just wasn't the time. So it was very exciting, very Wild West, you never know what's going to happen next kind of moment. We've matured since then. And so it's definitely, we kind of went through the, kind of the rise of computer vision and all its various things. And there was, for a while, there was a new algorithm every, every week coming out in that space. And then we transitioned into the NLP period, where we had tremendous progress there. And Daniel has been right in the middle of that. I've learned so much from him. But we're also, we've matured quite a long way in a fairly short amount of time. And one of the things I'm noticing is we're still having things like, I mean, GPT-3 just came out recently, and that was a big deal. But

• 23:56Chris Benson?

  I don't see the, the like, every week things happening. And so because we're seeing kind of an evolutionary progress, there's a lot of people in this space that are starting to say, have we mined this for? All the the big new things or or maybe most of them. I don't think we're there yet. But are we getting to a point where we've had some maturity? And we're having lots of kind of just evolutionary improvements in the in the current space of deep learning. And because of that, a lot of people are turning back to kind of AGI. You know, which is artificial general intelligence, the idea of the the AI that can kind of do multiple complex things instead of this simple AI that we have, the narrow AI

• 24:39Chris Benson?

  that is very good at doing one particular task. And so there's talk of something that if you had said this to me two years ago when we started, I would have laughed. I would have, you know, like, there's no way. And that is the idea of an AI winter at some point down the road. I don't think it's quite that because it may be that the research being in a more in a slightly more mature area is focusing on lots of evolutionary things, but we probably have quite a ways to go before AGI. So I think we might have a lot of commercialization over the next few years about what's already out here. Because there's tons of industries that can use what's already been created and discovered. But how far are we from the next major step up revolutionary wise? That's not just an evolution of the existing, but a new thing coming out.

• 25:24Daniel Whitenack?

  I always, and Chris knows, I kind of cringe a little bit when we talk about AGI and like general intelligence because I think what a lot of people mean by that doesn't like it doesn't mean what they think it means.

• 25:39Speaker 03

  What does it mean and what do they think it means?

• 25:42Daniel Whitenack?

  Well, I think like AGI is like the a lot of people just think of like some type of singularity and like Terminator incident where like now we cease to exist because AI has like crushed us into the ground and we're no more sort of thing, which I think is to me inconceivable, but you know, whatever.

• 26:01Speaker 02

  He's so bullish on being inconceivable, though. Like, he's dead set. He's he's sure of it. Challenge him on it, Adam. Challenge him. I still know, man.

• 26:11Chris Benson?

  Like I just, yeah.

• 26:12Daniel Whitenack?

  I don't know if I agree with that. I'm not on the inside, but this is my own opinion.

• 26:15Chris Benson?

  Right.

• 26:16Daniel Whitenack?

  It's at least, um, I think inconceivable that it would be sometime in our near future or my lifetime, but I, to me, it's inconceivable, but.

• 26:26Chris Benson?

  It depends so much though, on how you're defining that, which I think is important when we define what AGI is, because you can define it a couple of different ways. And I'll give you completely different answers about what my opinion is on those.

• 26:38Daniel Whitenack?

  And I think this comes back, we had a conversation about the, uh, NERIPS, which is a big AI research conference and in the keynote, they were talking about AGI. And they were talking about it in much more terms, like I could grasp like certain things, like the architectures that we've been seeing in, in natural language processing that involve attention and self-attention and models that actually go beyond sort of fitting parameters, but actually paying attention to certain pieces of data more than other pieces of data. And they tied that into saying, okay, well that's a, like that's an advancement to generalization. And I can like latch on to that. That seems like some interesting steps there. And so, yeah, I think it does depend on how you define it.

• 27:19Daniel Whitenack?

  And to Chris's point as well, I think that, you know, people are going to start trying new sorts of things. And I just looked up Sasha Rush, we had him on the podcast from Hugging Face, um, a couple of weeks ago. And he was, uh, one of the organizers at ICLer, which is another one of the huge AI research conferences. And he posted a graph, like he graphed over the last two years. So while we've had the podcast, the keyword growth for certain topics in research over that 2018 to 2020. And I was actually somewhat shocked. So the, the top growth was a thing called graph neural networks, which is a new way of, um, constructing neural networks to work on graph structured data. And so

• 28:04Daniel Whitenack?

  I think that that's an indication that people are like, they pushed. The sort of architectures that people are used to. Pretty far and they're searching kind of for something new and exploring new areas. And new like types of structured data, new types of data, maybe multimodal types of data where multiple different types of data are linked. So that's really what I think of when I think of people like pushing the boundaries of it.

• 28:31Chris Benson?

  Yeah, but it's deep learning specifically. If you think of AI as as a broader collection of technologies that are advanced, you know, these are more deep learning things. And I think some of the other luminaries in this field talking more about what AGI means to them and, you know, what they think about that and stuff. And then you kind of have the whole commercialization thing going. So there's a little bit of divergence in some areas between the research community and the commercial interests, the commercial community, you know, community, because, you know, we're seeing deep learning models and architectures deployed in many different areas. But that is separate from, you know, what what leading researchers at Google Brain or OpenAI or something like that would be focusing on. So you're seeing people kind of cash in on where

• 29:15Chris Benson?

  we're at, which is good. And I think we'll see it everywhere. I mean, just pervasively. But that's different from where things are going in the future. And, you know, it also, going back to the AGI point, one of the things I think that I've learned in this two year period is I think I was unclear on the idea that if you were to look into this somewhere down the road future about, you know, AGI comes about and, you know, obviously we're a long way from that. But if at some point, I think people confuse whether or not consciousness is part of AGI or not. And so I think we'll get to tremendous levels of intelligence that has no sense of self or consciousness many years before.

• 29:55Daniel Whitenack?

  Right.

• 29:55Chris Benson?

  Any breakthroughs on on the ladder on consciousness. So, you know, we're already seeing models that in very specific areas can outperform any human in a particular task. And so, you know, if you extend out that computation capability, then if you're defining intelligence as a form of computation. Being able to solve a problem, then we're kind of, we're much farther down that pure computational road than we are. I don't, you know, you can have an amazing model that can outperform any human, but that doesn't mean that it's thinking, wow, I wish I had a cup of coffee right now.

• 30:29Speaker 03

  For sure.

• 30:30Chris Benson?

  That's a big leap right there. Self-awareness is a key ingredient missing. I'm not sure that's clear in most people's minds because if I have five different conversations on this topic with five different people, their idea of what they're starting from, all five can be very, very different.

• 30:43Daniel Whitenack?

  Yeah. And maybe a good way to put it in, one of the reasons why it's inconceivable to me is because like the way that we're approaching this path to advancement in AI, like we're going to destroy the planet way before an AI singularity like happens. There was a study, it takes like, we release into the atmosphere as much carbon as five cars during their entire lifetime when we train one of these state of the art NLP models just once.

• 31:14Speaker 03

  Wow.

• 31:14Daniel Whitenack?

  So I think if we were to say, oh, let's just like, keep rolling with this, like, we're going to have to be living on Mars by the times that the, you know, AI singularity happens. So that's like a whole nother strain of things where it's like, can we actually solve some type of the computational issues that are happening with AI? And we're going to have to like, we can't just keep doing things status quo. We have to figure out like more efficient and like creative ways of doing what we're wanting to do, I think.

• 31:41Chris Benson?

  You know, to your point there, when we do get to that point where the singularity does come about and it's self-aware, it's going to start off depressed as hell. You know, we'll have destroyed the planet. It'll be like, why did I bother?

• 31:53Speaker 03

  It'll turn itself off.

• 31:54Chris Benson?

  Why did I wake up to this?

• 31:55Speaker 03

  It'll just turn itself off. I'm way too, I'm using way too much energy here.

• 31:58Chris Benson?

  And has anyone noticed we have a princess bride thing going here? We keep saying the word inconceivable.

• 32:05Speaker 03

  That's true.

• 32:06Chris Benson?

  And, and we were saying, I don't think that thing means what you think that thing means.

• 32:10Daniel Whitenack?

  I was hoping that someone would do some other reference after I said that, but I wasn't.

• 32:15Chris Benson?

  There is a princess bride theme going in this episode.

• 32:17Speaker 03

  Do you want a peanut? No, that was too non sequitur. Let's talk about GPT-3 for a minute, just to completely, it's not actually a hard shift, because I think it plays into this concept that you're referring to, where it's like a plateauing. And a question that I have, because what happens with us who aren't deep in the field like you all are, is every once in a while, we just get impressed with the results of some sort of new technique or model. And it has been a while, so I've definitely seen the slowing for me, where it's like, I feel like I've seen a lot of the things, and it's like, yep, I've seen that before, that's cool, like different applications. I think the color transformation stuff is really cool, right, with the style transfer kind of things.

• 32:53Daniel Whitenack?

  Yeah, super cool, like de-oldify.

• 32:54Speaker 03

  Yeah, exactly. So you're just seeing like, well, let's take that and apply it to X, Y, and Z. And then every once in a while, you get super impressed. And GPT-3 just made a splash this summer on kind of like the tech Twitter and the VC Twitter and like in our space by generating some pretty realistic and even tricky sentences and phrases and blog posts even. And I don't know what GPT-1 or GPT-2, or I don't know what GPT-3 is, I know it's OpenAI, I know it's the first time I've signed up for an AI-related beta. I'm like, I want to play with this, you know, it's, it's that, but is that like, is GPT-3 kind of like the last or the next phase, is it an evolution of what they've been doing?

• 33:37Speaker 03

  Is it a brand new thing? Tell us your guys' take on that thing.

• 33:41Daniel Whitenack?

  Daniel's our NLP expert, so I'm gonna throw this to him. So I think it's an amazing achievement and like incredible results. And also like people just being really creative in their use of it. I don't think it's like a fundamental paradigm shift in terms of like how we're doing modeling necessarily. And one of the reasons that they listed in not releasing the model publicly was the fact that like it requires an extreme amount of like compute to make it run like efficiently and good, which people just don't have access to, like. Normal people, maybe like us on the call. I don't know if I want to call us normal, but.

• 34:19Speaker 02

  Speak for yourself.

• 34:20Daniel Whitenack?

  And OpenAI people as like other than.

• 34:23Speaker 02

  Everyday people.

• 34:24Daniel Whitenack?

  Yeah, it is an amazing achievement. I think it's an amazing achievement, both computationally and scale wise. And you know what they had to do to achieve that scale and all of that. But I think it is in a sort of paradigm shift in the way we're going about things, I guess. It's still, it fits into, I think, this, this model of bigger and bigger sort of general language models that people are training on huge amounts of data using sort of task, multitasks or arbitrary tasks to help the model learn about language like filling in words or you know, question answering or sentence completion or character completion or predicting next sentence.

• 35:08Daniel Whitenack?

  And like these sort of tasks that people don't really care about that much, but are used to train these like large models that are then able to be fine tuned for very many tasks that may be unrelated like, you know, translating code from one programming language to another or, you know, generating like understanding queries to generate, you know, front end components and that sort of thing. I know those are a couple of things that were shared in the changelog slack. So, yeah.

• 35:41Speaker 02

  Could you maybe hypothesize about what's after this first magic trick? So.

• 35:47Daniel Whitenack?

  You mean for GPT three?

• 35:49Speaker 02

  Exactly. Like, you know, so that I say that's the first magic trick to sort of like introduce this new, you know, generation of language prediction model and what it can do. Like, it's very impressive that that that blog post was generated. And then I think like on some ethical level, like, do I trust that blog post less because it wasn't actually generated by a human? So why do I just trust humans so much more than, let's say, anything else? That's an aside, but I think like, okay, this is the initial magic trick. What are the actual applications of it? What's, what's beyond this blog post for GPT three?

• 36:22Daniel Whitenack?

  Yeah, for me, it'll, I'm mostly interested in.

• 36:27Daniel Whitenack?

  How people go about like interacting with this model because the standard has kind of been in the past that, you know, people train a model and then at some point like a serialized version of that model is released. And you can load it into your own code and do like nifty things with it. That's, you know, not going to happen in this case for reasons that were on purpose that they are releasing this via API. They're running it internally. They can shut it off when they want to. There's interaction patterns that are governed by that API, which aren't just kind of, you can do whatever you want. And so I think it'll be interesting to see like, you know, how that influences how people are using it.

• 37:11Daniel Whitenack?

  Obviously, as you've already seen, there's a lot of creative uses already. But I think that I know why, you know, these reasons why they did this. And, you know, that's within OpenAI's set of AI principles, right? But it also somewhat constrains or puts some boundaries, I think, on, you know, how you're going to to use it. And so I'm mostly actually interested not so much in the application because I, I think it'll be things similar to what we've seen in the past, maybe just leveled up a notch, but more how people figure out ways to use the API or, you know, develop workarounds or creative

• 37:53Daniel Whitenack?

  uses and that sort of thing. So the workflow wise, it's very different.

• 37:57Speaker 02

  Back to the inconceivable part, though, this shows some level of fear.

• 38:02Daniel Whitenack?

  Yeah.

• 38:02Speaker 02

  From humans to non-human in this case, like the machine, so to speak. There is some sort of fear. Maybe it's a fear of how other humans will use this.

• 38:10Daniel Whitenack?

  Yeah.

• 38:11Speaker 02

  That's it. But the governance and the API and the restrictions definitely shows, you know, either, you know, responsibility or fear. I'm not sure. Maybe it's both.

• 38:19Speaker 03

  Well, it's a tool, right? And any tool can be used for evil or good.

• 38:23Daniel Whitenack?

  Yeah.

• 38:23Chris Benson?

  And we've had this conversation a bunch of times on the show in that, you know, like every technology, people always take technologies and most people use them for for good ends. And there's always a handful of actors that use it for bad ends. And the worry here, obviously, is especially if you combine this, you know, like we're talking about GPT-3. And first, it's been done. So, you know, people know, even if they don't have the details, they now have seen it. It will eventually get out there, whether it comes from its point of origination or whether it comes from some other group that says, oh, I think I could do that as well. I've seen the results. I have a pretty good sense. That's possible down the road. And we have seen that with previous models coming out and then other people kind of piling on.

• 39:07Chris Benson?

  But if you combine that with obviously other aspects like generative adversarial networks, and we're talking about deep fakes and stuff like that, then there is room for bad actors to do stuff. And so that is a concern. So, you know, when you're detecting that, I think it's less about, you know, it's not that we're worried that we're about to hit the singularity and, you know, what do we do about these machines? I think it's more a question about what do we do when the criminal mind out there or something decides that it's commoditized to the level they can use it effectively and they they decide to deploy it. And we have to, we have, it's part of the process. We got to sort that out.

• 39:38Daniel Whitenack?

  I mean, that fear element, to me, it comes down to a couple of different pieces. It's one that like this is trained on so much data in a sort of programmatic way that like when you scrape 80 million websites, like you don't always know like what's going to happen if you train a model on that and what your output's going to be. So there is this element of like, how do you probe all of the unexpected things that could be output from this model? That's a hard piece. The other thing is just like, I think, you know, yeah, any tool can be used for good or bad. But like, if you think of different tools, like, Like, we can all go get a hammer from the hardware store, right?

• 40:20Daniel Whitenack?

  And I can choose to like, you know, hang a picture on my wall with that hammer. I can choose to hit someone over the head with a hammer. But everyone can go get a hammer.

• 40:28Chris Benson?

  Don't do that.

• 40:29Daniel Whitenack?

  Like, that's, that's like democratized. I can go get a hammer. I can go get something else to hit you with, right? But with, with these sort of AI technologies, if like, for example, China is using these sort of computer vision technologies to detect a person's ethnicity from a street camera based on their walking gait to determine if, you know, they're part of a minority population called Uyghurs and so that they can track them and put them in concentration camps. And those Uyghurs, they can't go out to some store and get that same technology. Like, there's such an imbalance with this technology and who has access to it? And who has the compute? Who has the money to run the compute?

• 41:13Daniel Whitenack?

  Who has the facility to get the data? That the imbalance of power is really emphasized in, for these technologies in particular, I think.

• 41:22Chris Benson?

  And there's a good point there when you talk about scale, is that, you know, to your point there, it's a powerful tool that is powerful enough to where nation states are very, very interested in this and they put a lot of money and effort into this. And that just amplifies how things can go off the rails. You know, it's, it's one thing if I get a, you know, a mean streak in me and I go out and use a tool to go do something, but when people do it at that level, it's one of the challenges of our time. So, I mean, I think it's not unique to AI. It's any tooling, but it's going really fast in general. If you look at the fact that over just a decade, we've made profound advancements in this area that are usable.

• 42:06Chris Benson?

  So this is part of society. This is part of culture. We have to get solutions to this and it's not going to stop anytime soon. So rather than be afraid of it, we just need to be focused on, on answering it effectively. We, we need great minds to put their attention on it.

• 42:40Speaker 03

  So y'all have done 100 episodes and I've been curious to ask you this for a while, which is that when we set out to do this show, we called it Practical AI. And that first word has really been kind of a primary focus of the show and maybe a guiding light to a certain degree. I remember hearing Daniel oftentimes saying, "It is Practical AI." And he'll use that as a way of kind of turning the conversation into the practical aspects of deploying and using and etc. these technologies. But I'm wondering if you felt like that's limited you or made the show go in directions that you haven't wanted to, or if there's any sort of maybe an inkling of like a regret of being pigeonholed into the Practical AI podcast?

• 43:26Daniel Whitenack?

  No.

• 43:26Speaker 03

  Okay.

• 43:29Speaker 03

  Dang it.

• 43:29Daniel Whitenack?

  I don't know. I guess you could interpret Practical as being like practitioner, like, you know, in the sense of like tutorials and implementation, tooling sort of things. But we have gone into things like ethics, or like the use of AI for good, or telling stories, more so than just like highlighting methods, I guess. And to me, that still fits within like the practicalities because it's, you know, the same reason why you want to see like case studies or whatever, when you're looking at a particular tool or product, you know, something like that, where like you want to get a sense of how people are

• 44:12Daniel Whitenack?

  using a thing. You want to get a sense of how people are thinking about a thing that maybe you haven't thought about as much. So to me, I don't necessarily feel pigeonholed. I think we have gone into certain of those things, but you know, it's kept us away from... To me, it's like brought a bit of focus so that we're not always, you know, talking about Terminator and those things.

• 44:36Chris Benson?

  And I'll actually, to be honest, and since you asked, I hadn't thought of it that way, but I'm going to actually say occasionally, just to give a different perspective. I would imagine that anyone who's listened to us for a while has heard me. Between the two of us, I tend to be the one that gets out there into the speculative realm more often.

• 44:54Daniel Whitenack?

  Yeah. And I'm the downer that always is like, "I don't know how to implement that from my terminal." So.

• 45:01Chris Benson?

  I'm out. I'm running step by step directions. Daniel's like, if I can't deploy it to Kubernetes, it doesn't count. Yeah, exactly. You know, I definitely have an interest in, in looking out there into what is essentially speculative. Yeah. You know, even philosophical realms. Part of that is probably not because of my job, uh, because that's not what we do at my company. But, but just as a person myself, uh, in the defense world. And if I go give talks, you know, Daniel and I have talked about all these talks we've given over the last few years. And I commonly am asked about these speculative things because people make the assumption incorrectly that being in defense and being, um, doing AI, that I must be thinking about terminators all the time. So it's, it's something I do think about that idea.

• 45:46Chris Benson?

  It's not real life. It's not real life in the defense industry or at the DOD. But I will confess that I can't help wondering about a future like that and putting some thought into it. So it occasionally creeps out into the podcast and I think Daniel does the right thing as he quickly. He goes, well, it is practical AI.

• 46:02Daniel Whitenack?

  It's good to have the balance.

• 46:04Chris Benson?

  Exactly. And he pulls me back to reality on that. Mm hmm.

• 46:07Daniel Whitenack?

  As you can tell, probably tell from a hundred episodes that, uh, I'm not, uh, maybe not the speculative philosophical type.

• 46:15Chris Benson?

  Right. He whips me into shape immediately.

• 46:18Daniel Whitenack?

  Yes. I do struggle in those like planning, strategy meetings at certain points in like organizations where it's like, well, how does this fit with our vision? How does this like, you know, what are our transformational things that are going to happen in the world based on the stuff? And yeah, so it's good for me to be pushed into those areas. It's definitely good for me to put, be pushed into areas where I'm looking like beyond like my, my VEM window. But, um, you know, I also always like to, you know, go back there.

• 46:54Speaker 02

  I kind of see the name as more of a, less of a restriction and more of a, a north star. Rather than like, this is the direction you're going. You can go into the fringes to, because it's always good to entertain possibility. But maybe a better way to define it might be what you think practical actually means. Like, I think practical means possible and useful.

• 47:16Chris Benson?

  I do too.

• 47:16Speaker 02

  What's the lore to the, the word practical for you both?

• 47:20Chris Benson?

  I think that we keep it for the most part, and even me, despite my, my confession a moment ago. I think this is real life. So when we were starting, AI was still very young and cool. And most organizations didn't have that as a capability in-house. And that's still developing, but right now it's, it's no longer the thing where it's the strictly aspirational intent. It's now something that a lot of organizations are incorporating and they have really practical problems to deal with. Like, okay, well, we now know how to produce, to address an architecture and produce a model that can be deployed. And how do I get that into our deployment, you know, methodology?

• 48:05Chris Benson?

  How do we get it out to our customers? How do we do that? And I think that's where the bulk, that's where 99% of AI is and should be, I think, and for most people, you know, if you're not strictly a researcher. And so I think my sense is that we've given a platform for that, for people to talk about it and to come on the show and to help others. And I think one of the things that might have set us apart is that we're always thinking about, is this going to be meaningful for, for listeners who are trying to do it themselves out there? And I think that's where, where Daniel's, but this is practical AI drawing us back on the line is really important, because that's, it's fine to dream a little, it's fine to speculate a bit. But for the most part, people are trying to figure it out and get work done.

• 48:50Chris Benson?

  And I think that we actually help them get there.

• 48:52Daniel Whitenack?

  Yeah, I think useful and meaningful are two words that have been mentioned in the past a couple of minutes that are, to me, important. So there are useful and meaningful things that are not code or implemented on GitHub, right?

• 49:09Daniel Whitenack?

  You know, there's things that influence our workflow and the problems that we solve that are, you know, still useful and meaningful, but not like talking about a framework or specific open source project or something like that. So yeah, I agree with that.

• 49:26Speaker 03

  So with the course of a hundred episodes, surely there's been highlights and lowlights. There's been successes and struggles. We didn't prep you guys like bring your favorite episode or anything like that. But if you could just think back and would have been some of the struggles of the podcast and what have been some of the successes for y'all.

• 49:44Chris Benson?

  You want to go first, Daniel?

• 49:45Daniel Whitenack?

  Yeah, sure. Put me on the spot.

• 49:47Chris Benson?

  I did. See how I did that?

• 49:49Daniel Whitenack?

  I think one of the successes for me is the sort of, um, and we're always like, I think we can always do better in getting a sort of variety of guests and we're always like trying to cover different things that we haven't covered. But I do think that we've had guests on sort of a range of like from students, just like getting into AI and giving their perspective all the way up to like, you know, people like, like Stuart Russell and others that have been like luminaries in the field. So I think it's really cool that we've got that diversity of perspectives and you know, a whole lot of different topics covered.

• 50:29Daniel Whitenack?

  So looking back, I think that's one of the things that I'm, I'm pleased with is that kind of variety of, of perspectives that we've brought. All kind of bringing their unique spin on, on whatever they're, they're talking about. Also, I, I don't know if this was specifically a thing that we set out to do when we started. I don't think it was, but even from the very start, I forget, it was one of the first episodes where we talked about a TensorFlow project that was helping African farmers with this app where they took pictures of their cassava plants and you know, it helped them.

• 51:08Speaker 03

  That was really early.

• 51:09Daniel Whitenack?

  Identify disease. That was like one of the first episodes, I forget. Like ever since that point, it seems like we've always kind of had in the back of our mind, like high, specifically highlighting AI for good episodes. So I think those are really the ones that stand out for me. There's a one with, um, you know, uh, representing a, a project from Data Robot where they were working, trying to detect like waterline issues also in Africa.

• 51:34Daniel Whitenack?

  There were recent ones about like COVID-related data and question answering systems related to that, or like getting scientific knowledge out there to scientific researchers about COVID. So those AI for good ones are really some of the ones that stand out to me.

• 51:50Chris Benson?

  Yeah, I think I would agree. And I think going back to what you were saying before, some of these folks that have come on are just luminaries in the field. You know, you mentioned Stuart Russell a minute ago.

• 52:00Daniel Whitenack?

  Wojciech Czerimba.

• 52:02Chris Benson?

  Oh, yeah.

• 52:03Daniel Whitenack?

  A lot of different people.

• 52:04Chris Benson?

  Bill Dali from Nvidia. You know, we had Anima and Anand Kumar. I've probably just messed up her name and I apologize for that on the show from Nvidia. And just, I mean, just these people that have really wowed me and to some degree been heroes for us. But at the same time, then we bring on people that have never been out there at all, but they have great ideas and they have great insight and they're hungry and they're doing really amazing things. And it's been a platform for them to come out and share that. And so I think it's been, from my standpoint, a really good balance that we've struck in terms of not just going with superstar people or just going, you know, whatever, but being able to get all those different perspectives into the show. I think that's good for the AI community.

• 52:46Chris Benson?

  I wish, I wish that there were many channels within the AI community to do that. There's, going back to actually, one of the things that I admire from another change log show, which is GoTime. For those who don't know, that's for the Go programming language, is the Go community has really rallied around that podcast. And I love that sense of community. And I think that's been one of our aspirations here is to give the AI community a place to rally, to be real, to have this chat, like what we're doing today and what we've been doing all these episodes. I think I really would like to see us be more successful, you know, in the next 100 episodes for trying to bring people on board and make it a conversation and recognize that AI is just now part of life, like so many other things, and that

• 53:31Chris Benson?

  there's room for every voice involved.

• 53:33Speaker 02

  We've seen great success there, especially with JS Party and GoTime in terms of community representation, diverse voices, diverse perspectives, polite disagreements. We love that. We think that's a welcome place for good dialogue like that and conversation. And those two shows are great examples of us iterating towards that, in quotes, greatness. I don't want to call us great, Jared, but we're doing a great job there. Those shows are very representative of two diverse spectrums in software, JavaScript, Go, of course, and you know, all the ways those two languages and communities tend to go out to do what they do. And I'm happy that both of you are inspired by that, impressed by that, and desire that, too.

• 54:17Daniel Whitenack?

  Yeah, definitely. And I think one of the cool things that I've seen is there are people that are active on our Slack channel, on LinkedIn, on Twitter, and are interacting with both Chris and I, you know, a good bit. And like even today, I was getting, you know, show suggestions or like, hey, I was looking for a show about this, but it doesn't look like you covered that yet. And so all of that's really cool. I think there's been a lot of things throughout the episodes where people have suggested something, we've tried it, or we've covered a gap that people wanted to talk about. And so one of the things that I maybe didn't quite expect was that sort of ongoing continual, like throughout the

• 55:00Daniel Whitenack?

  week, outside of the podcast, conversation with a lot of people out there that aren't guests on the show, but are involved in the community. And that's really cool to see, like, people, you know, having discussions that are useful for them, even outside of just listening to the podcast on these different, you know, channels.

• 55:18Chris Benson?

  The reach, I think, you know, Daniel and I came into this not having ever done this before. And, you know, Adam, you and Jared have been doing this for years, and you're, and you're real pros at it. I think it was very surprising to me when, you know, it wasn't a big deal kind of thing, but like, when I joined my current employer and I'd meet some people and they'd say, oh, I really enjoy your podcast, by the way, just wanted to let you know that as we're meeting. It was a quiet thing and no big deal, but it was like, wow, the reach is really, you know, really there and the conversation that that creates is really there. And I've ended up away from the podcast, having a lot of conversations about topics that we address because you'll

• 56:00Chris Benson?

  meet people out there and they just want it. They're like, by the way, you did this episode on such and on this topic and I'm, it just really struck a chord with me and you'll end up spending, you know, 15 or 20 minutes just kind of redressing it. It might've been an episode for many months ago. And that's had a fairly profound impact on my awareness that that sense of community, it's not always completely apparent, but we have developed that and AI is getting there. It's following other technical communities in that way and, and the need was recognized and people are, are choosing to opt in and that's something that's truly gratifying from my standpoint.

• 56:34Speaker 03

  So my interview skills failed me once again. I asked two questions, allowing you to slip the second question, but my persistence pays off. So I'll ask that one again. Struggles. Doing podcasts is hard. Getting to a hundred is hard. Most people fade and there's a lot of work involved. What has been some struggles along the way? We know we've had life problems, guests problems, scheduling problems, you know, it's not all roses and rainbows. So what have been some of the struggles and along the way?

• 57:02Daniel Whitenack?

  I felt like for quite a while when we were starting, when we were reaching out to guests, I mean, I knew certain people in the community and certain of them were easy to kind of make connections. But when you're starting out a podcast and you don't have a long history of things, it's like, I feel pretty much like a creeper reaching out to these people on like Twitter or whatever, like sending you the random message, like you have no idea who I am. Thankfully people are mostly gracious, but there was a lot of like, no, like no response. Like there's that time period where you feel like, is this really gonna like pick up? I know some people are listening, but like, you know, every, every time I reach out to people, it's like explaining this thing we're trying

• 57:47Daniel Whitenack?

  to do and like, I'm kind of coming out of left field. So I think that was like a bit of a hard, hard period for me. And of course, anytime I've learned over time that anytime I listen to myself, which I try to do not, you know, like 100% of everything I say, but I try to listen to what I'm saying. I think it was like, Various things that you do that like make you cringe and then you replace those things like you say, all right, I'm gonna like work on that. And then there's various other things that you just replace that thing with that also make you cringe. And so it's like, yeah, there's a little bit of that too. And, you know, life happens. Chris and I, I think, have both had times where I tell Chris, I'm like, I need you to like, push this forward for a couple of weeks

• 58:32Daniel Whitenack?

  because I am like, I'm not going to be able to do it. Thankfully, I don't think either one of us had had one of those times where we both needed it at the same time. But there have definitely been times like that where we flip, flip-flop back and forth.

• 58:47Chris Benson?

  Yeah, I would agree. And, and we rely on each other. You know, Daniel said that about me, but, um, you know, without going into all the specifics, uh, listeners who've listened know, I've, we've had some challenges this year that have certainly affected my family and Daniel stepped right in and just took care of things. And that was good. And you really have to, you know, all of us that are doing this, uh, we have day jobs. So we, we need our, uh, the people that we work with to make allowances and accommodate to some degree. We have families who are supportive. We're all doing these podcasts out of our houses. And, you know, families having to, to, to recognize that you need a, a little time, a little quiet space to get some stuff done. And, um, I, I know for ours, when, you know, we have a, a house full of dogs and a young daughter,

• 59:32Chris Benson?

  and that can be pretty challenging. And then frankly, there's a burden that I didn't appreciate back when I was listening to the change log and go time and JS party, uh, before we started this, I don't think I understood the burden of trying to provide a great hour or 45 minutes of content, you know, week after week after week. I, I've learned to really appreciate that when I listen to other, is I listen to other podcasts, that it is hard work to do that. And so I've developed as a, as still an amateur in this compared to you guys who are the real pros. I've really developed an appreciation for the amount of, of effort it goes in to at the bottom, uh, what, what, what it comes down to is serving your audience and trying to help people get what they need out of it. And, you know, they're, they're choosing to listen to you for a little while, and they should rightly expect something from.

• 1:00:18Chris Benson?

  From that. And so I think that's what I would finish with is just I there's always the burden for Daniel and me to make sure that we have content that is good for people to listen to. And so for those of you out there listening, please keep coming at us with social media, all the places that we like to engage you. We are listening. We are welcoming your ideas. You are giving us our ideas for our future shows and we want to serve you.

• 1:00:40Daniel Whitenack?

  And finally, after like a year of doing the podcast, I got to meet Chris in real life and he's a real person. So.

• 1:00:49Chris Benson?

  Sort of.

• 1:00:49Speaker 03

  We'll have to take your word for it.

• 1:00:51Daniel Whitenack?

  We hadn't actually met in person before we started the podcast and that happened way later, which was kind of funny.

• 1:00:56Chris Benson?

  It was actually, it was about a year. And it was only in the beginning of January, right before the COVID stuff set out. So most of these episodes of these hundred episodes were done. Daniel and I had never been in the same place at the same time before we became friends through the podcast.

• 1:01:12Daniel Whitenack?

  We just started our COVID life early.

• 1:01:15Chris Benson?

  Yeah, we did. But it's worth mentioning, it's remarkable that it is through these episodes that everyone listens to, that Daniel and I have become very, very dear friends. And we had never met until January of this year. So in person.

• 1:01:29Speaker 03

  Amazing.

• 1:01:30Speaker 02

  Well, these struggles illustrate why we're here today, which is to celebrate.

• 1:01:34Chris Benson?

  Yes.

• 1:01:35Speaker 02

  One, Jared and I are both proud of you guys. You guys have done a great job with this show. It takes, as you've said, a lot of work to produce on all sides. So one, congrats on 100 episodes. Two, we're proud of you guys. You guys are doing a great job. Keep doing it. Don't stop.

• 1:01:51Daniel Whitenack?

  Yeah. Thanks for believing in us and keeping us on track. It's been awesome. Truly.

• 1:01:57Chris Benson?

  And yeah, and you guys have been right there, even though you you're not there from the listener's perspective, you're invisible to them in most episodes. You're actually there every episode. And I got, I cannot count the number of times that Daniel and I have thought, oh, at least we have a great post-production team that can clean up whatever mess we just created during that. And so that's really important.

• 1:02:20Daniel Whitenack?

  Yeah. And do cool promotions on everywhere. And yeah, keep the wheels moving and it's awesome.

• 1:02:26Chris Benson?

  Yeah, it's a team effort. Totally.

• 1:02:29Speaker 03

  Well, I'm not just a producer. I'm a happy listener and I fall into the category of the AI curious, which we know that practically AI listeners kind of fall into a couple of categories. You have the practitioners, then you have the curious. And I'm very much in the curious. I've never used any of these things in a useful context. What was the other word? Useful and valuable?

• 1:02:49Daniel Whitenack?

  Meaningful?

• 1:02:49Speaker 03

  Meaningful. Yeah. I've never used it in a useful or a meaningful.

• 1:02:52Daniel Whitenack?

  Oh, possible. Possible is a good one, too.

• 1:02:54Chris Benson?

  Inconceivable.

• 1:02:55Speaker 03

  Yeah, I've definitely used it in some inconceivable fashions, but I know more about this stuff than I ever thought I would just by osmosis of listening along and producing. And I'm very much conversational AI enthusiast at this point. I can talk to anybody about it and trick them into thinking that I know what I'm talking about. That's a valuable skill in life.

• 1:03:17Daniel Whitenack?

  Yeah. Whenever you need a large salary, you can go out and really, really score high on that interview, I think.

• 1:03:23Speaker 03

  I probably could. So what's next for the podcast? What can people expect soon? Where's it headed?

• 1:03:30Daniel Whitenack?

  Yeah.

• 1:03:31Chris Benson?

  You know, for me, I was thinking about this as we were coming into this. And one of the things that I really would like to hear more about is, is how people are incorporating these AI technologies into their lives and into their work. We hear that to some degree and we know the breadth that this stuff is, it just could be pervasive in every industry. And I'm really curious to learn more. I think partly that may be because I spend, you know, I'm, I'm, I'm in my own job and I spend so much time completely absorbed in trying to get my own work done, but people are using this stuff for really cool, cool applications. And I think I would like to go out there and understand where it's being applied and what are some of the things that I never would have thought about? So I'm pretty excited about that.

• 1:04:17Daniel Whitenack?

  Yeah. I mean, there's still a lot of gaps that we, that we haven't covered. Even like today, I was talking to a listener who saying, hey, let's cover something about like drug discovery and pharma with AI sort of thing. Like there's just tons of areas that we haven't covered yet. So I think like getting some people on from, from there. And I think also, you know, developing more good relationships with the kind of leaders in the field. And, you know, hearing their perspective over time is something that you can continue to expect along with like.

• 1:04:49Chris Benson?

  Those up and comers and I think there's really cool stuff as well that's happening all around the world outside of the US and Europe with AI, both in Southeast Asia and Africa and Latin America. And one of the things that I really have a passion for and would love to see on the show is us continuing to try to, try to get those guests who are really innovating in those areas and bringing the benefits of AI to the whole world, really. So.

• 1:05:16Speaker 02

  One of the best ways to circle back is to invite the listeners. We've mentioned community a lot. So one easy way, if you're, I'm sure if you listen to this, you may get the outro and occasionally we'll say it in the outro, which is, hey, there's a community. So change all the com slash community, but there's a Slack, there's lots of people in there. If you want to share your story, these stories, the ways you're tinkering, the ways you're practicing, et cetera, in AI, you can do that there. So the invitation is there, change all the com slash community, the practically I channel, hit that up. Daniel, Chris, they're always hanging there. We're in there, all that good stuff. And there's lots of other people there, too. Come and hang out.

• 1:05:53Chris Benson?

  You know, another, one of the things that I wanted to mention was I think we're at a point now where it's been commoditized enough to where it's not all about business. There's a lot, you know, going back, you know, in the spirit, Adam, of you talking about, you know, your cul-de-sac, you know, camera identification. There's so many places. I believe you. I believe you. There's so many places where we can bring kids into this and do weekend projects. We've talked about some of that at the beginnings of some of our episodes, where you can bring people in, you can take it into your schools as schools are, you know, are opening up as COVID gets under control. There's just so many opportunities to do this in with, you don't have to be a top scientist. You don't have to be a data scientist professionally.

• 1:06:37Chris Benson?

  You can do it at your house. You can, I've heard so many little home projects that people have gotten into over the last year or two. Actually, I would like to highlight some of those. So if you're a listener and you're doing something cool over the weekend, especially if your kids are involved in it, please let us know about that. I'd really like to share those ideas. I have an eight-year-old daughter and, you know, it's a perfect time for kids to get into this and it's cheap enough and you can get the equipment cheaply these days. So, so yeah, let's bring, bring your family endeavors. It's not always just about work.

• 1:07:07Speaker 02

  It's like you said, though, it's real life. It's not going to go anywhere. So it's not so much like, hey, let's teach our eight-year-olds because we have to, but more like because it's interesting and it's curious.

• 1:07:15Chris Benson?

  It's fun. It's really fun.

• 1:07:17Speaker 02

  Who'd have thunk it to actually make? AI practical to everyday users. I mean, that's the thing, right? Like, I literally want to learn about AI to the point where I can do something to help defend my cul-de-sac from nefarious people who are trying to steal our wheels. Like, come on. There's gotta be an easier way than that, and I have no idea.

• 1:07:36Speaker 03

  And there's tools out there.

• 1:07:37Speaker 02

  Raspberry Pi and stuff like that, so maybe I can arm myself with an R-Pi and a camera, of course, and different detection, and maybe my neighbors will let me take their picture and put them into my model, who knows? But that's shooting for the moon there, but it could be fun, and I have no idea where to start, really.

• 1:07:53Speaker 03

  If you're protecting everyone's tires, why not?

• 1:07:55Speaker 02

  Right, exactly. That's what I'm thinking.

• 1:07:57Speaker 03

  It's like Neighborhood Watch, only with no people. Exactly. I like it. Well, guys, thanks so much for sticking with it, having an awesome podcast, and thanks for letting us crash. Thanks for interrupting. Yeah, we're happy to join you for this one. Celebrate with you.

• 1:08:12Speaker 02

  Gosh, it was a fun crash. I love that.

• 1:08:14Chris Benson?

  If you hadn't interrupted, who knows what this episode might have been?

• 1:08:17Speaker 03

  Yeah. Wait and see, I guess.

• 1:08:19Speaker 02

  I'm curious.

• 1:08:20Speaker 03

  That's right.

• 1:08:24Speaker 03

  Don't forget, we have a giveaway going on in celebration of episode 100. Enter for your chance to win some awesome AI hardware from Nvidia, Intel, and Google, plus Practical AI and Packyderm swag. We're giving away three bundles, so you have a good shot at them. Check your show notes for details on entering. You have until the end of the month. Speaking of Packyderm, a little birdie told me they have a big announcement coming soon, and you should join their Slack channel to stay tuned. Learn more about that at Packyderm.com. Thanks to our longtime sponsors for their continued support. Shout out to Fastly, Linode, and Rollbar, and to the mysterious BrakeMaster Cylinder for these awesome beats. That's all for now. We'll talk to you again next week.

• 1:09:27Speaker 02

  Did y'all hear my kids there for a second? I guess something fell up there and my son was like, "No!" I heard it. I was just chuckling because I'm like, "I know it's, you guys gotta hear it because I can't stop it."

• 1:09:37Chris Benson?

  I'm half-blind and deaf anyway, so.

• 1:09:40Speaker 02

  I dropped the marker though, I mean, you could totally see in the timeline. Yeah, I heard it. That happens occasionally that gets edited out. Thankfully, no one gets upset, but hey, that's what works.