1029: How AI Brought a Podcast Back From the Dead, with Linear Digressions’ Katie Malone

20:17Katie Malone:But they tend to not, it's just not as nimble to get 10 ,000 people kind of going in a particular direction. I think I do kind of wonder, some of this is just reflecting where we are as a society right now. This might look very different in five years as people have had a chance to acclimate a little bit to some of the AI tools. People are maybe a little more fluent with it. Some of the norms that I think we're figuring out now might have settled in a little bit. Like it might be, you know, you feel like it's not okay to send AI slop to your coworkers or something right now. I hope it doesn't.

20:55Jon Krohn:Please stop. If you're that one guy, just stop.

20:59Katie Malone:It's not one guy though. Yeah, that's the thing. It's like, you know, my AI slop is talking to your ASLOP. I talked to Tom Davenport a few weeks ago for my podcast. He's wonderful. For anyone who doesn't know Tom, he's been writing, especially in enterprise data science and analytics for decades. And he has coined the term process SLOP. And I think it's an idea whose time is rapidly approaching of like, there's a whole process. I'm a job applicant and you are the hiring manager on the other side. And it's just like our AI slop going back and forth. I have AI generate my resume. You have your AI that reads it, that automatically sends me some kind of reply, like whatever.

21:44Katie Malone:Anyway, so I think that those are challenges for us in general and in particular in large organizations where you might not have direct personal relationships with the folks that you work with. You're kind of relying on the machinery of the organization and some of the processes to kind of get things to where they need to go, then injecting AI into that all of a sudden is like not necessarily, you know, fitting in exactly with how these things are working. And so there's also, I think, a really important part of the, what you might call change management or something, like just how do you get people, how do you turn that aircraft carrier?

22:20Katie Malone:And I don't know how much I have to say here that's deeply insightful or specific and insightful besides it's just really hard work. And I think it's interesting to see in some ways as there's new companies that are popping up, they're obviously approaching how to build businesses in sometimes very fundamentally different ways. established companies are retrofitting their operations and their technologies to varying degrees of success or maturity at this point. So it's a, it's an interesting, uh, I guess, period of high flux for us all to be in.

23:01Jon Krohn:Tom Davenport is great. We had him on the show in episode six, four, seven, he writes about a book a year.

23:06Katie Malone:Did you just have six, four, seven in your head or did you like pull that? Okay. I have a spreadsheet. That is impressive. Okay.

23:13Jon Krohn:I did a command F of the word Davenport and it brought me to the right row. All right.

23:18Katie Malone:I feel better now.

23:19Jon Krohn:Uh, there are a few that I have memorized for various reasons. For example, I was the guest with Kirill Aromenko hosting it. Like, so my first experience with the super day designs podcast, I was a guest in episode three, six, five. And it's so easy to remember because there's that many days in a year. Um, and then there's another one. There's episode seven, 777 is my number one that I recommend to people who don't come from a technical background, uh, because it's with, uh, another big author, Bernard Marr, if you know him, a big author of business books and in recent years, AI books, and he's got millions of followers online, really great introductory episode for non-technical people on what generative AI can do.

23:59Jon Krohn:Anyway, I, I think those are the only two. Um, but, uh, Tom in his episode, there's a funny line where he quips that somebody else, because he's a professor as well, and some other professor quipped at him that Tom Davenport has never had a thought that he hasn't published.

24:22Katie Malone:He is prolific. He is prolific. I'm proud to say, have you published with Tom Davenport? It's like an Erdox number or something. I'm proud to say that my Tom Davenport number is one. We have a Harvard Data Science Review co-authored article. Yes. about model deployment back in the day. Yeah, arguing that that was a thing that people needed to worry about because it was not as apparent at the time as we thought it should be.

24:48Jon Krohn:Nice. Well, let's digress less. And I don't know if we are digressing linearly or non-linearly, but I will bring us back to a little bit of the track that I felt I had us on, which is that when you are doing things with agentic AI, so you are managing things, I don't want you to talk about things that are proprietary to anywhere you've ever worked, but are there particular tools or platforms that you prefer, say personally, that you think our listeners should be using?

25:23Katie Malone:Well, I don't know if anybody's heard of Claude Code. Highly recommended.

25:31Jon Krohn:They're our biggest sponsor this year. Oh, really? Our audience here is Claude Code. Literally every single episode.

25:38Katie Malone:Nice job, Anthropik. Well, good taste all around. Yeah, I mean, I think the thing that I would talk about, and it's the thing that I have my hands on the most, is probably the podcast setup. Because I have some agents kind of, to borrow a phrase, I think, I first heard this from Dan Shipper over at Every, but he's got this phrase, a cloud coat and a trench coat. So the idea is like the core of the agent, you know, the main capabilities to do tool calling and reasoning and planning and things like that is all cloud code. But you kid it out with some specialized MCP and instructions. And then all of a sudden it's a podcast producer instead of just a coding agent.

26:20Katie Malone:And so anyway, so that's been one of the tactical pieces for me of that, you know, burnout recovery, like relaunch the podcast, but without it becoming a huge time sink. And I so beyond that, I don't I tend not to be like super fancy with my setup. One thing I've been having a lot of fun with lately, though, is adding some telemetry to that agent. I found it to be kind of fun to wire it up to, in my case, Arise that is like an observability platform. Sure. Yeah. Maybe familiar to many. Yeah.

26:56Jon Krohn:From them on the show. Yeah.

26:57Katie Malone:And they have, among other things, an open source version of their tool. So I had some fun a few weeks ago wiring that up. And it's interesting because then you get to kind of look, you know, under the covers, as it were, watch a lot of what the agent is actually doing in servicing the request. And for a nerd like me, like that's that's actually kind of fun. That's some of the the learning that I get. So it's it's both a experimental subject for me and the thing that actually kind of gets my gets my work done on 745 on a Sunday night. and like we need to post the episode. Like let's fire up the agent where there's some work to do.

27:38Jon Krohn:So I might be teasing hairs here in a way because both of these tools are offered within the same product. But it's interesting to me that for podcast production, you're using Claude Code when it kind of, it seems to me intuitively like Claude Cowork might be the more, I don't know, like it's interesting that you reach for Cloud Cowork. I mean, maybe that's just because of familiarity and you being so technical, but it seems like Cloud Cowork is set up to be this, because what aspects of podcast production involve code? You know what I mean?

28:15Katie Malone:Yeah, that is a great question. And it's actually something that I've toyed with a little bit. So the boring but true answer is I started, I set up the agent before Cowork existed. And so it's part of it is just like, you know how I said a second ago, like change management is a whole thing and getting people to like change what they do, like guilty. And I have explored a little bit, poked around at getting, you know, what would it look like for cowork to do the same stuff? I think probably some of this, at least for me is reflecting that I'm technical by background. Like you called out some of my, you know, early experience, part of what I want to do with some of these AI tools is feel a little more technical again.

29:01Katie Malone:I'm not like writing code, but I don't know. It makes me feel kind of cool to see like a terminal window going. And so anyway, I did experiment a bit or consider a bit porting it over to cowork when that launched. It wasn't clear that the benefits were, you know, so overwhelming that it would be, that it would be worth the effort. And so again, kind of the boring, but, but true answer is I was like, okay, well, I think I'd rather spend two hours, like building the telemetry software into it then, then during the port over. But that's a good, that's a good catch. And I think that it's a, it would be a totally reasonable thing for someone else who's listening to this and wants to do something similar, but it's just a little bit inverted for them.

29:43Katie Malone:Like I bet it would work quite nicely with co-work. Yeah.

29:46Jon Krohn:If you're starting today, but yeah, if it wasn't around and it's already working, if it ain't broke, don't fix it. It's interesting how you mentioned how like just having a terminal window or whatever makes you open makes you feel cool. Uh, cause I also, I, you know, I, I use them wherever I can, like just to change directories, make a directory, like grep something. But actually I think as I was thinking about that, I was thinking, Oh, do I just like doing it because it's cool? Or I like, I don't want to lose that skill that used to be something I needed to be doing all the time. But actually, uh, as we're speaking, another thing came into my head, which is, I think that that experience of being in the terminal window, it's a lot easier to be in a state of flow than when you're dragging and dropping all over your machine for some reason.

30:32Katie Malone:Oh, I think that, oh, I think you're onto something there. I had never really thought about it before. Um, and I mentioned it before a little bit, like I love flow as much as the next gal. Um, and, uh, I think a lot of people do, uh, yeah, there's just something a little bit special about you. Like you, you put on your headphones, you turn on your ambient techno or whatever, you fire up your terminal window. And I feel a little bit less like a whatever, like pencil-headed manager for a few hours. And yeah, I do like having some amount of understanding how the internals are working. And so I think if I were fully in co-work, it's probably not impossible to have those internal views, but like, I'm going to be pushing this code to GitHub.

31:17Katie Malone:Like I want to have a poll, you know, um, pull requests that I can open up and see what the diff looks like. Um, I have a little bit more of a technical understanding now of certain concepts. Cause I've asked my, my agent to explain itself a few times when I'm wondering how it's doing something. So I find that to be a, a bit of a virtuous cycle for me.

31:39Jon Krohn:Regular listeners will already be aware that I'm obsessed with Anthropic's Fable 5 model, and it has taken over my working life. I'm writing a technical book that includes LaTeX files, mathematical notation, Python code examples, and Fable 5 and Claude Code handles requests I make across whole chapters with accompanying Jupyter notebooks end-to-end, work that a few short months ago would have been dozens of separate requests with way more manual fiddling required. With Fable 5, it just works, essentially like magic first time. Claude is the AI for problem solvers. It's the collaborator that understands your entire workflow and thinks with you, not for you.

32:15Jon Krohn:Whether you're debugging code at midnight, building a financial model, or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. For problems worth solving, get started with Claude at claude.ai slash super data. That's claude.ai slash super data. And check out Claude Pro, which includes access to all of the features mentioned in today's episode, claude.ai slash

32:38Katie Malone:superdata.

32:41Jon Krohn:Makes a lot of sense. One thing that, so I was yesterday scripting and recording an episode about OpenAI's GPT-6 Astra, which is not yet publicly available, and I'm not cool enough or working at a cybersecurity firm enough to have access to it yet at the time of us recording this episode. and one of the things that I'm super excited about I so I actually like I haven't it's been probably a year or so since I've had an open AI subscription I have been you know all in on Claude for some time now but I would be really curious to try or hopefully it won't be too long before Claude code does this one of the new functionalities about it that I think would be so awesome and help us like kind of stay in the state of flow and just kind of collaborate better with a machine is that apparently gpt6 astra when you're using it in codex by open ai their coding environment it will if it needs to ask a clarifying question of you that will happen you know the chat box will open back up and you can provide a response but it will continue to have agents running in the background on tasks that don't depend on your answer.

33:55Jon Krohn:And it's such an obvious thing to be where we're going with this technology, but I haven't personally experienced it yet.

34:02Katie Malone:Yeah, that does sound interesting. I'll have to poke around at it. I'm a little bit curious about it too. I myself am also pretty heavily on Claude, again, mostly out of inertia. It sounds like the open AI stuff is neck and neck and depending on the week, there's cool new stuff that's on both platforms right now. Yeah, I think the flow state is a particularly interesting one. Again, I think about this sometimes in the context of being a manager versus being an IC. I really miss the flow state. That's much harder to get into when you're... That was my experience anyway when I started managing people.

34:41Katie Malone:You're bouncing from meeting to meeting. You've got people pinging you in a Um, and, and you, you can't really like turn it off in the same way and be doing your job effectively. Like kind of what, what you're, what you're being paid to do is to keep all of that stuff moving, like just run around touching all the plates. And there can be like days when you feel like you're in the matrix or something where you're like, you've got all the bullets flying around and you're redirecting them all. And I guess there's like, you know, kinds of satisfaction that one can get in that. Also, if you're, you know, an extrovert, that might sound like it would bring you a lot of energy.

35:21Katie Malone:But I know there's a lot of folks that they're like, that doesn't sound like my idea of a good day coding. And so, like, what do we do in a world? And how do we help? How do we get, you know, the really valuable and unique contributions that those folks have and not make them just, you know, the little drinking bird from The Simpsons that sits there and hits like, accept, accept, accept. except for all of the requests from Cloud Code, but that they're still engaged in a way. I think it's a real challenge. And so some of these UX, UI experiments around how to allow the system to go and not get blocked on you while also keeping the correct around of oversight.

36:07Katie Malone:I'm very interested in seeing what form factors we experiment with in that space. For sure.

36:14Jon Krohn:Great points all around. I'm going to bring us back now a ways. We're going to digress less. We're going to undigress. We're going to follow that line back to the main thread a little bit more. We were talking about AI slop, and then you brought up Tom Davenport and process slop, which wasn't a term that I heard before, but it makes so much sense. And so one thing that we pulled up in our research on you is that as far back as 2016, you complained on how Twitter was more and more becoming a platform for bots to talk to bots. It's a problem that you just - I don't remember that, but that sounds like it.

36:49Katie Malone:It was 10 years ago. I don't know why you would remember.

36:51Jon Krohn:But it's a problem you described as bot contamination and as bad for Twitter analytics. Now we have AI slop contaminating human accounts. I see it on LinkedIn all the time. particularly it seems like if I make a LinkedIn post, hashtag AI agents or hashtag agentic AI or something like that, I will get dozens of just, it's getting hard for me to say, okay, this is definitely AI slop, but it's, but I'll get like 20 different random people that I don't know. And they don't usually comment making the same point in different ways.

37:27Katie Malone:And I'm like,

37:28Jon Krohn:Like this is surely, and there used to be a time when I could definitely tell it was AI Slop and I would just delete it. But now it's kind of like, well, what if two of these 20 people really thought this? I don't want to delete theirs and I can't tell which ones they are. So what the heck can we do about this? Especially at a time, you know, like you and I both creating a podcast. And as we talked about on the Harvard Data Science Review podcast last week, well, in the episode that was recorded last week. and came out, well, today, yeah, so it came out at the beginning of September. Or I think it was the last day of August.

38:04Jon Krohn:Because they were like, oh, it's the August episode. We have to get it out. So it was like Monday, August 31st is when that episode of the HDSR, Harvard Data Science Review podcast came out. And yeah, in that episode, you and I both talked about different ways that we're using AI, you know, for us at the podcast. We use it to create summaries for the superdatascience.com slash podcast page because it does it perfectly. It's just so good at it. And I don't feel like that's a place where people, you know, this obviously is taking the thoughts and opinions of myself as the host and the guest, our real opinions, and it's just condensing it down into a paragraph that is perfect, easy to read, technically correct.

38:49Jon Krohn:And so it's like, that seems to me like a perfect use of, um, of AI. And, uh, but yeah, you talked about an example where, so you create a podcast newsletter where the underlying episode, the intellectual framing, the curation of the topic, the subsequent editing, you do all of that for the episode, but then you ask a model, probably like Fable 5.1 or something today to take that and turn it into a newsletter for your podcast. And then when you run that through a tool, like, so there's a sub stack tool to try to identify what content is AI generated or not. And you said that that shows up your newsletter shows up as 90 % AI generated, but that's because you just used AI at the final step.

39:35Whereas a lot of the heavy lifting,

39:37Jon Krohn:almost all of the work was done on the front end by you.

39:40Katie Malone:Yeah. It's an interesting spot to be, And I'm not too precious about it. Like that 90 % doesn't keep me up at night. But it did open for me, you know, kind of this interesting vein of exploration. And I'm trying to do a little research on the side to see if I can like understand this methodologically a little bit more because I'm just a nerd. And that's kind of my reaction to these things. But I think it's an interesting... They do have the concept in Pangram, as I understand it, of like AI assisted. So they're not quite so naive. believe, and I'm not saying that they are, that they don't see any gray area in between, but how do you interpret that 90 %?

40:19Katie Malone:What does that mean? And I think, yeah, at least for me, I'm perfectly happy to say 90 % of the words here were written by AI. That is factually correct. But in terms of the content that it's running on top of, yeah, I see a big difference between that and if I were to just do something like, hey, Claude, write an episode script in the style of linear digressions on a topic of your choice. I'm going to sit here and read it. We pass it through the newsletter. That's substantively different from the curation and presentation choices that I'm making upstream of that summarization task. So I think it's an interesting thing that we haven't quite figured out how to tease out yet either.

41:09Katie Malone:Something I wonder about too, and I didn't get a chance to ask Zhao Li about this, but I'm going to try to wrangle him into doing linear digression sometime and get his thoughts on it. But as it happens, my husband is an academic. He's working right now as a editor on the board of one of the journals in his field. So he's looking at, I don't even know, like dozens of papers per month, probably more. And, you know, not surprisingly, there's maybe some distribution of those that seem to have the obvious tells of being AI generated, others much less so. And their journal has some policies around, you know, in general saying that you shouldn't be using AI, but I'm not sure exactly how far that goes.

42:02Katie Malone:And I think it's an interesting challenge for folks like him and for other people in this field that are acting as like curators of content, whether it's academic material or if you're in the publishing field, maybe traditionally performing this role of kind of selecting from content that's been human generated in the past, you might be calibrated to think of this material as having been thought through by a person. And I'm going to approach it in that way versus something that's more just wholesale AI generated. And they make different kinds of mistakes. And I think it's I feel like I'm being a little bit muddled in my thoughts here.

42:42Katie Malone:And maybe that's because it's just a muddly topic inherently of when you have a human and an AI and they're collaborating with each other. You know, how much how do we think about the contributions of each of those? And I think the final product is not a tidy sum of like the two individual pieces. It's a messier combination. And we don't have all of the machinery right now, I think, to try to tease apart how that might be different depending on exactly what role each of those contributors plays.

43:12Jon Krohn:Right. Yeah. We just, yeah, lots of talking from both you and me, lots of questions, not really any answers, but, you know, maybe something for listeners to think about. But something that you do have answers on is you have a lot of content on your podcast about like these, you know, like these laws or these principles or like these kinds of like named concepts. And I wanted to dig into basically what I'm planning to do for the rest of this episode is to just talk about a bunch of these because I think people will love them. So the first one that I want to talk about is what's called Bainbridge's Ironies of Automation.

43:48Jon Krohn:and so good choice in an episode of your podcast called the impact of gen of generative ai on critical thinking you discussed bainbridge's ironies of automation which is when automation gradually erodes the human expertise needed to catch its own edge case failures especially when humans are supposed to be there to co-pilot automated decision making but end up just rubber stamping everything that's simpsons bird hitting the approve button that you were describing And yeah, I think the big problem here is that that could theoretically mean a time. And I feel like there's clever enough people at Anthropic and OpenAI that are going to, I don't know, figure something out.

44:28Jon Krohn:But maybe it's the biggest risk within an organization where you have kind of this agentic system. You have these processes that are self-correcting, that are recursive in some way. And yeah, you just end up over time, the process slop, the AI slop, it ends up corroding the whole process.

44:46Katie Malone:Yeah, it's a really interesting challenge. The idea of as the automation gets better and better, the human naturally has a tendency to trust it more. they become less experienced in dealing with these cases themselves. They lose some of that expertise. And then in particular, I think there's an interesting failure mode that introduces itself out at the tail ends where there's the most complex cases. Let's imagine the case of like a physician maybe, where if you have AI that's handling maybe some of the most straightforward cases, which it's probably very qualified to do for simple stuff, then that means that you as a physician are not handling those cases.

45:33Katie Malone:The only stuff that gets escalated to you is the trickiest stuff, the corner cases, the things that are not easy for an AI to handle, but you're out of practice at that point. You're kind of rusty. You haven't been doing kind of the reps day in, day out of the simpler stuff to be sharp and catch the stuff that's out of distribution a little bit. And so I think, yeah, that is a real challenge. And it's something that I think about sometimes in my work. I'm curious if this is the same for you, where even when I know from experience or I strongly suspect that the AI is going to do a pretty good job, I will sometimes kind of force myself to go in and to change something for the better.

46:18Katie Malone:It's not just like go in and make a change for the sake of making a change, but like think about where I want to make a change. I think the newsletter is actually a good example of this. I don't think I've ever put out a newsletter that didn't have some edit from me. And sometimes these are minor because it does a good job out of the bat. At this point, I have a prompt that's pretty good and it gets my voice for most cases, But I always read it and I'm always looking for something to change because I think that act of engaging cognitively with material keeps me sharp in a way that I would not feel the same way if I were just kind of piping it straight in.

46:59Katie Malone:But I think it's a hard stance to take. And to some extent, this was something that Tom and I actually were talking about a lot. if you're in a business setting, some of the advantages, the efficiency gains that you get from AI, they start to disappear if there always has to be a human in the loop overseeing and approving and managing. So it's a little bit of a double bind, I think. And I don't know how optimistic I am that we're going to come out on top on this one, to be honest with you, but here we are. But it's got a fairly catchy name, which is nice. Bainbridge is, yeah, a law of automation.

47:44Jon Krohn:If an AI agent caused an incident at your company tomorrow, what evidence could you produce? Most teams have the prompt and the final answer. Everything in between, the commands it ran, the files it changed, the credentials it picked up, are gone the moment the terminal closes. Origin closes that gap. A sensor on the endpoint records the agent's work as a trace. Who started the session? What was asked? What the agent reached? And what changed on one timeline? And because the sensor sits on the machine, you don't rely on the agent's own account of itself. Origin lives on the endpoint because that's where the work happens.

48:19Jon Krohn:Coding agents in a terminal, local agents, agents calling MCP servers on a laptop, none of that passes through a cloud gateway and Origin sees it anyway. When something unexpected happens, you read the session in order from prompt to outcome. Origin is endpoint AI observability. See what a trace looks like for yourself at originhq.com slash SDS. Yeah, it's cool. Let's move on to other ones. We can just kind of try to rifle through them here. One that we hear about a lot, but I always need to jog my memory on, is Simpsons Paradox. oh this is the one simpsons this is the one about um like how the overall trend can be running in

49:06Katie Malone:the a different direction than like the the group wise trend this is a deep yeah exactly so it would be like if you look at the whole i feel like i'm getting a pop quiz right now benford's lock go

49:18Jon Krohn:good heart's lock go it is it's so unfair i get to do whatever i want and i can make them successively more challenging as well. No, I won't do that. It's just a fun, I like, no, I literally had to, I had to look it up just now to remind myself, but it is exactly what you were saying. I was like, I think there was like a few neurons firing with this image of where you, if you look at something across the whole population, like let's say you have data on humans. And so there's men and women in there. And when you look at it over, like if you fit a trend line, you fit a line to all of your data points.

49:57Jon Krohn:It's like going, it looks like there's an increasing trend, but then if you break it down so that you're looking at both genders separately and now you fit, so you fit two lines to the data, you end up with both lines decreasing instead of increasing when you look across both groups. And, you know, it could be caused by something like, you know, men on average are taller or heavier or something. And so it's like, yeah, you get this weird overall effect when you look at both groups together. But when you subset by these important distinctions, these important categories, the important groupings that we have in the data, you end up seeing the trend going the opposite direction.

50:38Jon Krohn:And probably you can talk about it for a bit and I'm going to look up some real world examples.

50:44Katie Malone:Well, yeah. And what I'm wondering, the one that I was thinking of right off the top of my head. So something I've found pretty interesting to try to follow is this debate about whether what AI is doing to the job market right now. And I think it's a pretty complex picture where it's having heterogeneous effects at different, you know, junior roles might be impacted very differently from, you know, mid-level to senior, for example. And so teasing out exactly what the, what's happening with the job market overall and where you see changes, what's attributable to AI versus any other macroscopic cause might.

51:24Katie Malone:It's tricky as far as I've gotten. Many people who are much more well-versed in this are working very hard on trying to answer that question. But I wonder, that could be an example maybe. I wouldn't be surprised. I'm not claiming to have seen this, but just by way of illustration, that you might have some sort of result that says, you know, overall salaries, let's say, might be going up for the population. But within, or let's say they're going down, that might make a little bit more sense just intuitively. Let's say they're going, you know, down for the population, but they're going up for the highest earners and they're going up for, you know, the middle and lowest earners.

52:04Katie Malone:So I think that would be Simpson's paradox. So how is that happening? Well, it might be like the people are migrating between those groups. And so even if the tippy top of the distribution is doing better, there's fewer people in that part of the band relative, you know, people are kind of falling down the ladder a little bit or something like that. So anyway, yeah, it's one of those things. It's called a paradox for a good reason, because you're like, how can it be going up and down at the same time? And yet, there it is. So I'm sure there's, you know, economists are thinking about this as they're analyzing that data.

52:36Katie Malone:But those are the kinds of things that But yeah, well, I think as much as, you know, AI is very good at data science at this point, I think you'd probably agree for a lot of use cases. But thinking through and understanding, you know, some of these little trickeroos is still part of what, you know, what we still have to do. Not to say that an AI couldn't think about Simpsons Paradox.

52:57Jon Krohn:Yeah, I don't know. I guess like I honestly have no idea. I haven't tried with like Fable 5. to do this kind of thing, you know, to kind of just give it some raw data and say, find the trends and try to trip it up on Substance Paradox. Like, I wonder to what extent, like maybe it would be better looking for all these different kinds of paradoxes. Maybe you'd have to prompt for that. I don't know. I don't know. But I feel like at some point, it's going to surpass us in a lot of these ways. I don't know. But I think the key thing is the real world context. Like, I think that there is usually somehow there ends up being key pieces of real world context that haven't made it into the context that you provide to your agents working on it.

53:45Jon Krohn:And maybe that's like a big problem to solve. It's like a data engineering problem. Like I need to have like a recorder on me at all times, making sure that everything I say, and then like an agent takes all of that stuff and routes it into the context windows of all the relevant different projects, the different clients I have.

54:02Katie Malone:um so you say you have that set up or no this is like so okay i'm trying to like say interesting

54:08Jon Krohn:yeah because what i find because the a big part of my life these days is spending time thinking about how do i get all of the relevant context into the context window for an ai agent to be able to do this task for me yeah and that is a that is still a place where a human can provide a lot of value. Uh, hopefully I'll be able to provide value there for a while. Cause I would say that that has kind of become, that might be the number one, like single task that I do in a typical week. Uh, it's like I spend hours every day just for different tasks. Some of them are legal things. Um, you know, like reviewing documents, you know, doing a first pass.

54:55Jon Krohn:Do I, you know, is this something easy? Is everything easy about this form or, you know, like what changes has, you know, this new client's lawyers, what changes have they made to this 50 page contract? Please just spit it out for me. But then you're like, okay, but what other context is it going to need? What are the relevant emails or what are the relevant conversations that I need to kind of type in manually? Um, anyway, so we'll see what happens. But I, so I have, I have a real example of Simpsons paradox and apparently it's, uh, it's one of the best known ones, and it's really easy to understand.

55:27Jon Krohn:So it was a study of gender bias among graduate school admissions at UC Berkeley. And the admission figures for the fall of 1973 showed that men applying were more likely than women to be admitted. And the difference was so large that it was unlikely to be due to chance. So it was a statistically significant finding. But when they took into account information about the different departments being applied to. So you break it down based on departments. It showed that the different rejection percentages reveal the different difficulty of getting into the department. And it showed that women tended to apply to more competitive departments with lower rates of admission.

56:08Jon Krohn:So yeah, men were applying to less competitive ones. And so it looked like they were getting, yeah, it looked like there was this gender bias.

56:17Katie Malone:That's a good example.

56:19Jon Krohn:Anyway, so there we go. Let's quiz you on the next one. You know, you've done stuff like A-B testing, which I don't think we really need to get into that much, but you go into that stuff into detail. Network effects, shrinkage, Stein's paradox, Zipf's law. I don't know if there's any of those that stood out to you as another one you'd like to talk about.

56:38Katie Malone:Is Zipf's law the one about the first digit in the numbers?

56:43Jon Krohn:We did actually have that come up in our research. That is something called Benford's Law. And it isn't one that I looked into more. And it isn't one that I'd heard of before. Can you describe Benford's Law to our audience, Katie?

56:53Katie Malone:Yeah, this is a weird one. And the gist of it, again, this is a deep cut. But if you look at, imagine you have a bunch of numbers that are taken from measurements of some system. And the thing that's wild is you can kind of pick a lot of different things that you want to put in here. It can be prices of stocks in the stock market. It can be measurements of distances that people are traveling. It can be vote counts in election returns. And if you look at the first digit of those numbers, they tend to, it is not evenly distributed. So you may naively think you have a bunch of these numbers, like let's take vote counts, for example.

57:35Katie Malone:I want the vote counts for every county in Ohio for governor. And you look at the first digit, you would think naively that it's peanut butter spread across all the digits, zero to nine. And that is not the case empirically. In fact, there's a tendency to have many more ones and then it kind of starts to fall off and less less common to have digits near the end. Of course, you still get those digits sometimes, but it is very much not something that you would think by naively thinking about it. Anyway, what I like about this is number one is just weird and fun. And number two, it escapes me at the at the moment.

58:17Katie Malone:But there's you know, it's not it's not a total mystery. Like people have noticed this for a while and there are some explanations of how that can even be the case. And I don't know. I'd have to go back and listen to my old at the same to see if I can refresh my own memory. Yeah. But it's interesting. Yeah. Weird, right? Right. And it's news to me because I definitely, if you told me that, you know, you have some large sample of data, whatever it is, I would assume that the integers in those data would be evenly distributed.

58:49Jon Krohn:Like you said, peanut butter, peanut butter spread, you know, uniform distribution across those different buckets, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9. We should have roughly even numbers. It's going to be a tiny little bit of variation, but they should be roughly the same proportions. and as we have more and more samples, I'd expect them to be increasingly uniform. But Benford's law is that you get ones more often. And so it's something that's used in fraud detection algorithms. Because when people simulate fake data, they of course have even numbers of all the integers.

59:21Katie Malone:Exactly. And I think one of the things that I, you've picked up on something interesting that no one else has ever picked up on before. So I'm duly impressed by your research process here. but I do at a certain point in making this podcast there's this you probably know this this pressure too you're like all right I need to come up with something like what are we gonna do I've talked about a lot of stuff already so what's what's something else that we could go go go mine and anyway I I think that there's this this turned out to be a a rich little vein of content that I stumbled upon at one point, like it came up with some Wikipedia list or something of interesting little numerical properties or laws that have are named after people or whatever.

1:00:09Katie Malone:And most of them end up being sort of interesting. So I think I've done a few more at this point. But if we kept going, I think I would not maintain my my score in this pop quiz here. So

1:00:21Jon Krohn:leave it at that. Yeah, no, we will stop here, especially because we're running out of time. But I will tell the audience because I looked it up just now. So Benford's paradox refers to the counterintuitive phenomenon where the number one appears as the leading digit in naturally occurring data sets 30 % of the time.

1:00:38Katie Malone:Yeah, it's crazy. It's like not even close. Yeah.

1:00:40Jon Krohn:Yeah, you'd expect 10 % or I would have expected. And the number nine appears least frequently it's less than 5 % of the time. And oh, our human intuition would assume each digit. So there's nine digits, not 10. So you'd expect an 11.1 % chance, not 10%. But yeah, in fact, it is quite often, apparently it's something to do with, although our standard counting system is linear, natural processes often behave logarithmically.

1:01:08Katie Malone:Exactly. Yeah. And the thing that's fun about that is it also, it's, so you might be like, oh, well, what if it's not a base 10 system? Like what if I did like a base two system or something? And it, it, it's independent of the base system. So yeah, it's something in the generation process. So if you were thinking that, that was one of my first reactions was like, oh, well, what if we just did base two or something like that? Still there.

1:01:33Jon Krohn:Hey, hey, this is your host, John Crone. In addition to hosting this podcast, did you know that I run an AI consulting firm called Y Carrot? Yes, that's the letter Y and the deliciously crunchy veggie. At Y Carrot, our team pairs decades of ML and software engineering experience with recognized expertise in the latest AI techniques, such as all the key generative and agentic approaches. From problem scoping and proof of concept, through to high volume production deployments, we've got you covered at every stage of the AI project lifecycle. To learn more, head to ycarat.com. From there, you can click partner with us to give us some context on how we can help.

1:02:12Jon Krohn:We're looking forward to hearing from you. Again, that's ycarat, Y-C-A-R-R-O-T dot com.

1:02:22Jon Krohn:And there's a pretty good explanation here, which is that to go from the number one to the number two, that requires a 100 % increase in whatever you're counting. Whereas to get from the number nine to 10, that only requires an 11 % increase. And so there's more like resistance to getting to two than there is to getting to 10. And so you end up stuck at one more often than nine.

1:02:47Katie Malone:There you go. today we learned yeah that is weird well i did promise that we would talk about cern we didn't

1:02:53Jon Krohn:really get into that so well particle physics will have to wait for another day unless reason to have a next time yeah yeah exactly um but in maybe we could have an episode dedicated to it because i don't think we ever have had a particle physics episode so i'll make a note of that that

1:03:12Katie Malone:sounds really fun. And, uh, uh, yeah, in the meantime, I'll, uh, I bet they're doing some crazy stuff in physics right now. So, uh, yeah, that sounds, that sounds like a blast. Yeah.

1:03:23Jon Krohn:Fable and GBT six Astra are doing crazy things in physics right now. Um, I have no doubt. And so as we wrap up this episode, I ask my guests always the same two questions. One of them is for a book recommendation. Katie, what have you got for us?

1:03:40Katie Malone:I've got a book that I am reading right now that I'm really enjoying is called Music, The Brain and Ecstasy. I'm only part of the way through it, but I think folks who like this content may be similarly charmed by it as I am right now. And it's all about, so we have this experience as humans where we really like music. Some people like are obsessed with it and have, of course, like incredible skills. I think everybody likes it to some degree. and so it's going through the auditory and neural processing systems of like what we experience when we experience music and why we like it so much um and so it's got a lot of like music theory to it and it's talking about how like acoustics work and how your inner ear works and how your brain works and how you know signals reach different parts of your brain at slightly different times and we just like really love that for some reason i haven't gotten all the way through it but music the brain and ecstasy.

1:04:36Katie Malone:The author's name is Robert, uh, Jordan, uh, J O U R D A I N. And, uh, yeah, I've been enjoying it a lot.

1:04:45Jon Krohn:And does this, this has nothing to do with the music being even better if you're on MDMA.

1:04:50Katie Malone:I don't know. There's a lot that I haven't read. I would not be surprised if it's in there, honestly, you know, this, this is your brain. This is your brain on drugs and like, boy, Oh boy, does it light up? I don't know. Good question. This is your brain on ecstasy.

1:05:01Jon Krohn:This is ecstasy on ecstasy. Oh, that's true, yeah.

1:05:04Katie Malone:I think he meant it.

1:05:06Jon Krohn:I know.

1:05:07Katie Malone:I did not. I didn't get the pun at the...

1:05:10Jon Krohn:Yeah, I was like, oh, I don't know. She understood, yeah.

1:05:14Katie Malone:Right over my head.

1:05:15Jon Krohn:Yeah, yeah, yeah, yeah. But I guess it wouldn't be surprising either way, even if he hadn't used the word ecstasy, but something else. Cool. Well, the final question that I ask all of my guests is how can they follow you after this episode? Obviously, we know about linear digressions. What else?

1:05:31Katie Malone:Yeah, Linear Digressions is the big one. I'm not really on Twitter these days for my own mental hygiene, basically. But Linear Digressions has a few ways that you can join along. So the podcast is the main one, get it wherever fine podcasts are sold. and then the Substack newsletter, which is my fable laundry. It's Opus, technically. My Opus laundered transcripts, if that's your jam, and a little bit of content that's kind of my own original takes in any given week, unique from what we have in the podcast. So yeah, I would love to, if you've liked this, that might be something that you get a kick out of, so come on over.

1:06:17Jon Krohn:Yeah, I think there's a chance that if you like the super data science podcast, the podcast that you might like most in the world other than this is linear digressions. In a vector space, I suspect we occupy the same location no matter how many dimensions.

1:06:31Katie Malone:We're on that first principal component together.

1:06:33Jon Krohn:That's right.

1:06:33Katie Malone:That's right. And I just, as an aside, like it's been really, I've really enjoyed this. Thank you for this opportunity. This has been really fun. I love what you're doing there too. Like I think at some point I'm going to try to get you on my show and fangirl a little bit back in the other direction. So just thank you. I think I'm in awe of your productivity. I think you are over 1 ,000 episodes at this point. And someday, maybe I'll be there. But I have a lot of respect for what you've been doing here for the community year in, year out, week in, week out. I know what it takes. And it's just an honor to be here and to chat with you.

1:07:15Well, thanks, Katie.

1:07:17Jon Krohn:The honor is all mine, I assure you. And I look forward to being on Linear Digressions. It would be a dream come true. Thank you so much for taking the time out of your very busy day. And I look forward to providing our audience with a particle physics episode in the future.

1:07:34Katie Malone:Thank you, John.

1:07:35Jon Krohn:I love that episode today. In it, Katie Malone detailed how she relaunched linear digressions after a five and a half year hiatus because AI cut the production burden that had burned her out. She talked about why the skills of a people manager transfer to managing AI agents, Tom Davenport's concept of process slop, Bainbridge's ironies of automation, where as automation gets better, humans trust it more, lose their practice on the easy cases, and end up rusty exactly when the hardest edge cases get escalated to them. And we talked about Simpson's Paradox, where a trend across a whole population can reverse when you split the data into groups.

1:08:16Jon Krohn:And Benford's Law, where the leading digit of naturally occurring data is a 1 about 30 % of the time, and a 9 less than 5 % of the time.

1:08:26Katie Malone:Sounds weird, but true.

1:08:29Jon Krohn:As always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for Katie's social media profiles, as well as my own at superdatascience.com slash 1029. Yes, for episode 1029. Thanks, of course, everyone on the Super Data Science podcast team, our podcast manager Sonja Brejevic, media editor Mario Pombo, partnerships manager Natalie Zajski, researcher Serge Macisse, and our founder Kirill Arimenko. Thanks to all of them for producing another super episode for us today for enabling that super team to create this free podcast for you.

1:09:06Jon Krohn:We are deeply grateful to our sponsors. And if you'd ever like to sponsor an episode yourself, you can get the details on how to do that by making your way to johnkrone.com slash podcast. Otherwise, help us out by sharing the podcast with someone else who would love to hear an episode like today's review the podcast on whatever podcasting app you use if you write reviews on apple podcasts that's particularly helpful for us and i'll read them on air subscribe obviously if you're not already a subscriber but most importantly just keep on tuning in i'm so grateful to have you listening and i hope i can continue to make episodes you love for years and years to come till next time keep on rocking it out there and i'm looking forward to enjoying another round of this super data science podcast with you very soon.