Rebroadcast - Uncertainty and the Limits of AI with Chris Butler

Speaker A: Hi folks. 2023 has undoubtedly been the year of artificial intelligence. So that's why at the end of this year, as we wrap it up and look forward to 2024, I wanted to rebroadcast an amazing conversation I had last summer with my friend Chris Butler from Google, where we talked about the promises and perils of AI.

Speaker B: Welcome to the Thinking Leader brought to you by Red Team Thinking. Bad leaders react, good leaders plan, and great leaders think. You'll get new ideas and insights from business executives, military experts and innovative thought leaders to help you lead more effectively and um, better navigate your complex world.

Speaker A: Hello and welcome back to our show folks. We have a very exciting guest this week. Chris Butler is a lead project manager for Google's Core Learning Group and even more exciting to me, a very deep thinker on Uncertainty who I had the pleasure of meeting through his excellent contributions to the Uncertainty project, which if you follow our work you will know is very near and dear to what we are doing at uh, Red Team Thinking and uh, what we like to talk about on this show. Marcus is off this week, so it's just Chris and I. But you are going to, I think, really enjoy this conversation. I know, I'm looking forward to it. Chris, welcome to the show.

Speaker C: Thank you, uh, so much for inviting me. I'm really excited. I've been a big fan of uh, your work for a very long time as well. So excited um, to be here. Thank you.

Speaker A: Thank you. Maybe to start things off, can you tell people a little bit about your work at Google?

Speaker C: Yeah, absolutely. So I'm a lead product manager within the Google Core Machine Learning group. And what that means is uh, Google is a very well known company that does a lot of different things in

Speaker A: particular, I've heard of it.

Speaker C: Yeah, maybe you have used a product or two of ours. I mean we only have I think 8, 1 billion user products or something like that. Um, uh, but you know, so the core machine learning group is a group that's specifically created to help enable machine learning within Google as well as outside of Google. And so um, you know, that includes helping uh, product teams within Google to land uh, you know, machine learning features. Uh, it means that the frameworks and services that you might use um, to help build those things. So things like TensorFlow, Keras, um, those are all projects that come out of the core machine learning group and are actually open sourced as well. Then finally we think a lot about the infrastructure within uh, Google. And that means things like uh, servers, uh, accelerator chips, even things like chip co design, uh, for the Tensor Processing Units, which is an accelerator that we use inside of Google. Um, and so we kind of COVID all of the stack from very applied to very foundational kind of infrastructure. Um, and that's, you know, it's a group that I think is kind of at the center point for a lot of the Google innovation that's going on right now, especially around large language models. And so anyways, I would say like it's a very interesting group to work for. Um, my role within the group is really to focus on uh, three things. So strategy. So how do we kind of wrangle or uh, steward the strategy for this group that does a lot of very different things. Um, the second thing is really PMing the PM experience. Um, sometimes it's referred to as product operations. But I end up thinking a lot about the dynamics of the team. How is it that we either adopt processes to help us do a better job or actually remove processes to allow for more emergent behavior and then finally just like special projects. So if um, my VP needs something thought about, um, it's something that I end up helping out with. And um, so anyways, it's a really exciting role and it's a very interesting team to be working at inside of Google at this time.

Speaker A: I love something that you just said there, which is sometimes removing processes, um, which is something that almost never happens in most organizations.

Speaker C: Yeah, it's very hard to do.

Speaker A: Right.

Speaker C: Like I think, um, one of the things that I've thought a lot about, I guess as a way to frame the type of job that I do, um, and so people are familiar with Dave Snowden's work, which is around like Cynefin and the idea of the domains of Cynefin, right. We have uh, clear, complicated, uh, complex, chaotic and then kind of disordered in the middle. And um, now we just had Dave

Speaker A: on for the second time a few weeks ago.

Speaker C: Oh, that's awesome. Um, I love, I love his work and I think he does.

Speaker A: We do too.

Speaker C: Like, yeah, really interesting stuff to help me kind of frame the way to think about problems. And so I think my role in particular is really existing within the complex space, usually around organizational dynamics. How does this PM team work together and then taking the way that they end up working together and seeing if there's something that should be kind of turned into a systemic process or procedure and move that into the complicated domain. Um, but the problem is, is that you can be building things inside the complicated M domain, um, where it is no longer helpful. Right. And that you need to go back into the complex domain and start to think about like where what are we doing here? Um, the biggest version of that inside of Cynefin is of course like the drop off from like clear to chaotic. Right. Um, because you think you've gotten it all figured out but then you really don't at all. And so that's, that's my, that's my job is that liminality kind of between complicated and complex in that movement is, is really kind of what my job is meant to do. Um, inside this company, this group.

Speaker A: That's so interesting. And you know, one of the things that we talked with Dave about when he was on um, a few weeks ago was uh, his, his new, new focus on what he calls Esterine mapping.

Speaker C: Yes.

Speaker A: And I think that that gets into that in a lot of ways too because you know, I think when I think about things I've seen in organizations that I've either worked with in my current capacity or that I used to cover as a journalist, um, you have legacy processes in place that go unquestioned long after their efficacy and even their reason for existing has gone away. I mean Ford Motor Company was a great example of that. There were, there were things that Ford did that were, were legacies of the way Automobiles were produced 75 years before and made if you were starting over would make absolutely no sense. And that's why one of the things that we, huh, advocate with our clients a lot. And one thing we, that we practice ourselves rigorously is know Marcus and I, every quarter we have a conversation where we make a list of everything that we're doing and we ask ourselves if we were starting over today, would we still do this?

Speaker C: Yeah, that's right.

Speaker A: And then you have to have the, the, the discipline and the courage to say if the answer is no, then I'm going to stop doing that.

Speaker C: Yeah, that's right. And, and if people kind of can't remember why you're doing something, that's usually a good reason to stop doing it. Um, but I think it gets to also kind of like what are your assumptions about the way the world is working? And so, and I think this, this, we talk about strategy or decision making a lot of the time. We have to have a working set of assumptions. And um, if you forget about those working set of assumptions or you don't write them down or you don't constantly kind of monitor them for signals that are showing that those, those things are changing, I think you get into like a really bad place where you're just Doing things because you think they're the right thing to do. And that's where you end up getting into a space of, I think, you know, talking about domains. You're in the disordered space, actually, because you don't know what you're doing anymore. And so I think that's really, it's really important to understand.

Speaker A: Well, I like another thing that you, that you touched on too, which is that you're doing this to clear the way for new, new processes too, or new ways of doing things. And if you don't make that space, then. Then you're not taking advantage of, of the innovative thinking that's always going on around us about, about new ways, you know, of doing things. And, and, you know, every. It seems like right now, particularly thanks in large part to the work that you and your team and, uh, others like you around the world are doing, there's an opportunity to rethink just about everything right now. And I really, I really believe. I mean, we've said this on the, on the show and you know, I mean, there's, there's very few, few facets of human endeavor that I can think of that are not or could not be touched in some way by machine learning, by large language models, by AI, uh, at this point. And it's all gone from being something that's speculative to something that's actual and happening right now. And you know, it's, It's. It's creating, you know, and I'm always amazed when I see leaders who get that and excited to see the ways that they take advantage of it. You know, I was talking with, with. With a leader that we work with who, who is, um, a fire chief and of a large department. And he said, you know what? I, I was, I was listening to your show about, about AI and the opportunities and I was on. I was making a call to one of our, to our, to our, uh, our, uh, employment attorney, who I realized I have to consult many times a month because, you know, our department rules require that anytime I write up somebody or give the, you know, put someone on leave or discipline someone or let them go, I have to write a letter and it has to go through the employment attorney and get vetted and all this. And I just decided to ask, sorry, what are your chat. GPT to, uh, to write the letter I was supposed to, to send. And then I sent it to a friend who's an attorney just to say, hey, is this kosher? Is this right? And he's like, yeah, I couldn't have written it any better. And he said, so now he said I'm saving my department, you know, more than $10,000 a month in legal fees. But I mean this is something that is so it's not even core to his job, you know, in a certain sense. But it's like people are finding these ways to leverage this stuff every day.

Speaker C: Well, that's the part that I guess I want to always add a disclaimer about these types of technologies right now.

Speaker A: Right.

Speaker C: Is that it's not so much kind of the common case that's the issue.

Speaker A: Right.

Speaker C: Because what you're talking about is there's probably well known regulations, uh, there may be kind of private documents that refer to policy that may not be part of the training set that that particular large language model is included. And so I think that's the only thing is it's not so much like maybe the 10 or 100 times where it does the right job. It's the one out of 100 times where it basically comes up with some type of fact. Uh, that is not a fact. Right. Because it's not using actual references in the world. And so I think this is where there's still a lot to do when this type of thing. And I think also, you know, something I've been thinking about a lot is just how do we decide what is the appropriate times to use different types of technologies. Right. Maybe getting back to your main point about how machine learning and these types of things, non deterministic systems, the way that these, these systems are kind of trained and then utilized will change. I think almost everything about the way that software development is done. Right. Kind of today you end up building something that's very heuristic based. Right. And things still go wrong by the way, like there's always bugs because I think we're still not great at actually creating uh, or disambiguating things that we want to do and understanding the goals or the outcomes that we're really trying to go for. And that's the reason why like software is so hard to do perfectly. And I think it's going to be helped in a lot of ways through these non deterministic systems like machine learning. To be able to say like we don't know what the rules are right now, but we have lots of data around this and we have a way to understand or track whether we're actually getting the right outcome for people. And so then these will become part of almost everything uh, that I think a software developer ends up doing. Now there's Always going to be specialization, right? There's always going to be new state of the art models and there's a big discussion right now about uh, in particular model architectures are constantly evolving. Um, and this is more of a model centric view inside of the industry which is that we can keep on kind of building better models, right? And we just shovel data into those different and better models. We'll get better and better results. There's another side of the world which is more data centric, which is that actually highly curated data sets are potentially very valuable and that those are the types of things that we should be focusing on because it ends up kind of shortcutting all of the energy that we end up using to be able to train these humongous models. Right. And so I think there's this constant tug and back and forth between this idea of like tons and tons of like huge models and then very, very specific curated models that are based on specific data sets. Um, I think there's like some interesting papers that are starting to come out about how there's this kind of collapse that happens for large language models when large language model output is fed back into large language models. And so I think it's, it's almost going to be like there was like the data sets pre, large language models could be incredibly valuable rather than the ones that are post large language models that are just like non human synthesized data sets.

Speaker A: Oh, isn't that fascinating?

Speaker C: Yeah, I mean it's kind of like I, I'd heard a story about how like there's a certain type of steel before a certain era that was like really valuable because it didn't have all the, in it.

Speaker A: So yeah, it's actually, so what's happening right now, it's actually, it's actually a, uh, there's a, there's an international incident going on between Malaysia and China, which is probably what you're referring to. Um, and the British, because the Chinese have started uh, salvaging um, the wrecks of, of the British Asian Asiatic Fleet that was sunk uh, in the battle of, of I think the Battle of Java Sea in World War II, which are protected under international law as, as military graveyards. Um, but the Chinese are, are pulling them up because the steel used in those battleships was forged before the first atom bomb test.

Speaker C: That's it.

Speaker A: And apparently if you want to build scientific instruments that measure things like quarks and things like this that are used in things like, like the Large Hadron Collider and stuff like this, you have to use Steel that was forged before the first atom bomb test. Because every, and this is, this is such a freaky thing. Every piece of steel that was forged since then has traces of radioactive elements in it. And by the way, so do we. This is something that. This was a little bit less shocking to me because I was an anthropology major and I was shocked about the same thing back in 1989 when our professor introduced us to the concept that you didn't have to go to the trouble of radiocarbon dating human bones if you're just trying to determine whether this was a murder scene or an archaeological scene. Because there's a quick test you can do. Anyone who was born after 1945 has a different, has different bone chemistry than anyone who was born before 1945. Um, uh, because we all have little, little shiny bits of, uh, atom bombs in us.

Speaker C: That's amazing. Yeah. And I think this, like, this issue that we're talking about, right. Is like, it's going to be, it's going to be interesting because there's another paper I just read recently. It was talking about kind of like the limits of this data as well. Right. So they had postulated that the limits of high quality language data will be exhausted by 2027. Um, now there's a huge margin of error that they have in these graphs. You should go read the paper. Um, and they think about the amount of kind of video and imagery data as well being exhausted in the next couple decades. Um, this doesn't take into account, say, personal data, right. Which includes a bunch of other kind of like, issues around privacy. And then it brings into another topic which is kind of called federated learning, where you end up having learning at the device that you're actually using. But, um, there's a lot of interesting kind of topics that come out of this idea of like, how does synthetic or synthesized data buy these, like, large language models to then train smaller models? Um, how does that either help us or hurt us? And does it amplify biases that we don't want? Like, biases that cause us to have worse accuracy for, like, the true outcomes we want. So there's a lot of, like, really interesting subjects around this stuff. But I, I think like, in the end what it comes back to is really like, what are the human outcomes that we're trying to do? Right. Like, I. There's no such thing as a machine learning algorithm that does not interact with human beings. Right. And if it did, it wouldn't matter. It would, it would be the box in the corner that Just like has a bleeping, a beeping light every now and then that doesn't matter. It's the one that actually interacts with human beings. And that, that idea of like, we're always just, we're always in some way intermediating human beings with technology. And so that's, I think, something that I try to bring back to this. And the way I think about, you know, I think a lot about kind of organizational dynamics. And um, I've been really interested in this, this uh, new kind of practice called, uh, design fiction, uh, which is something that is, you know, maybe not super new. I mean, it's been around for maybe a decade or something like that. But the practitioners are mostly discursive designers that think a lot about that are kind of in alignment with kind of speculative future people, um, and thinking about kind of like, how do we ask provocational questions a lot of the time about like, how does, how should we be doing things today based on these far future kind of speculations. And I bring this back to organizational stuff because I keep on wondering, you know, inside of most organizations we just keep doing the same thing over and over again. We're just like, this is good enough. Or like, I'm smart, I'm smart. I got hired by a smart company. So I'm just going to do the thing that I think is smart when the reality is like, all this stuff should be in some way questioned and we should have provocations about like, are we doing the right thing the right way? And this gets back to your point about the process.

Speaker A: Yeah, I really like that, that methodology because, you know, I'm so, you know, I. One of my favorite authors is William Gibson. And you know, I, I, uh, read an interview with Gibson, um, probably like 20 years ago, probably maybe even, maybe even a little bit longer ago than that. Um, and, and the, the interviewer was asking him, like, it's uncanny how many things that you have described in your books are in the process of coming to pass. And, and you know, they were asking in this, you know, reverential tone of like, you're, you're this amazing visionary who sees. And he said, look, he said, I'll tell you, all I do is I look at where things are at today and I extrapolate it out to the, to the next, you know, the next logical or illogical, um, step. And, and that's what you're describing there. Um, but I mean, it is amazing if you go back and, and read books like Neuromancer Account 0 and Stuff and I mean how much stuff that he described, I mean it's almost to a T. How things have, have played out. Um, and most interestingly to me, not on the technological side, but on the social side, um, on the societal side. He, you know, his books, his books, uh, that he wrote, you know, back in the, the early 90s and stuff have a lot about the, the, the extreme, uh, division between wealth and poverty that occurs in the United States in the first half of the 20th, 21st century, for instance, and the expansion of homelessness in the Bay Area. I mean even literally. I mean there's a whole trope in his books about homeless camps taking over the San Francisco Bay Area. And these are books that were written, you know, 20, 30 years ago, you know, and, but that's what you know, but, but you know, he was looking around and seeing like this is getting worse every day. So what if, what if we turn it all up to 11? What does that look like? You know, and, and so I think there's a, I think, and I think that that's a different approach than most innovators taken. And so I think it really is a valuable. And I want to come back to something you said though. Um, it touches on something you and I have talked about offline. This idea that of the real value being in, in more specialized large language models. And, and I, this is something that I've been fascinated with ever since I, I did a little bit of work with Paul Allen back in, I want to say like 20, uh, 15, I think 20, um, 16. And at the time one of the things that I was most interested in that he, that he was working on was, was funding the development of I'm m. Using a lot of stuff in AI. But one of the things that his team was focused on was creating a, an AI that was entirely focused on one task which was reading the medical literature and being able to synthesize peer reviewed papers and then with a simple prompt, tell scientists. Tell a research team that's working on a particular problem. Here's the current state of thinking on X with citations so they could go back and check it themselves. And you know, I mean the, the, the potential for that was amazing because you know, uh, as his team was describing me, you know, we can, we can give a team of doctors that's working on, you know, a particular gene therapy a summary of the, of the, what their peers have, have been doing on this topic over the past 10 years in 30 seconds. And this would have taken, you know, a team of graduate students, you know, six weeks to do, you know, 10 years ago. And I, and the reason I thought it was fascinating for, for just the value in and of itself in that area. But I was like, I want that too. I would love to be able to have to, to not have to read 10 websites to find out what I want to know. I would really love, you know, to have a machine synthesize that information and present it to me in an executive summary. And I started to think as a writer and stuff, you know, what a force multiplier that would be for my own work. And that's why, you know, I've told you is I'm, I'm so excited actually um, about what you guys are doing with search right now. And I think that, that, I think that has the potential to be like the killer app of this current stage of large language models. Because it, while it's cool to be able to have you know, an A.I. write your letters, you know, your, your employment letters for you and I think that's totally game changing. I think the ability to rapidly synthesize and information in a lot of different domains, um, is going to be an enabler for a lot of innovation from people who are going to take that and use that in ways that you all aren't even thinking about right now.

Speaker C: Yeah, I mean the, the possibility of like summary summarization and like even just this idea of like translate anything to anything.

Speaker A: Right.

Speaker C: So you, you're not a legal expert. Translating something from that, like legal speak to something you can understand is really, really powerful. Um, I think there's something interesting about what you talk about when it comes to like scientific ah, discovery as well. Um, that's something that probably Gary Klein, uh, who is someone that you had I think on your second episode and I'm a big fan of his work as well around decision making. Um, and so I think what's really interesting about this idea of like tacit knowledge and so they've, they found um, that this idea of like innovation within organizations, it's very hard to actually transfer um, the specifics of a particular, like the discoverability, sorry ah, repeatability uh, of experiments ends up getting sometimes difficult between different labs because there's not the tacit knowledge of how to run the experiment correctly. Um, and so it's not just the idea of like embodied tacit knowledge. Like I'm able to like when I'm dipping this thing in a test tube in a certain way, it actually helps increase the ability for that result. But even the idea of like Cultural tacit knowledge, uh, which is something that like when we talk about like the Toyota production system, right, like it was very hard for a lot of people to replicate that. Not just because like you could get the book and you could read it, but if you didn't embody the knowledge, you didn't have the culture that was actually pushing for that type of thing now. And that, that goes both ways, right? Like there are, there are cultures that end up being very heavy handed and end up like perpetuating themselves in some way. And so I think like, what I'm very interested to see over time is for these types of like summarization capabilities or this idea of I want to just understand kind of like what's going on with regards to this like industry. I think that is, you know, in some ways if you look at the output of these types of models today, if you're an expert in a particular field, it will some, it will usually be more middling kind of uh, understanding of it, right? If you're a true expert in that industry, you'll look at some of the output of this, you'd be like, uh, you know. Yeah, that's kind of true, but it's not the most interesting part of this, this, right? So I think there's like still a lot of work to do. Um, and this is maybe where I think going back to like provocation of people I think is actually something that this can potentially do very, in a very interesting way. So like one of the techniques I end up using a lot are like different card decks. You know, I have like oblique strategies and um, a couple other things like that where I use them to like basically create a little bit of confusion inside of a process that I'm working on. But it's not a lot of confusion. Like oblique strategies is very kind of, um, not opaque but like, it's very conceptual. And sometimes people that are incredibly literal, they're like, what is this? This is, this is bs. Like I don't understand what's going on right here. But adding a little bit of confusion actually causes people to reframe something that they're thinking about. And so I think there's some interesting directions to go with this where if you have enough of an understanding of an industry or of a context, like my personal context, what would be enough confusion to add to my current work right now that would push me in a little bit further direction than I would have gone otherwise? And I think that is something that large language models can potentially start to innovate around. Is this idea of how do we have these relationships that we're working together? And it's that the machine is there to not only assist me in say, summarization of something or better understanding of something or starting to understand something something or teaching me something. But it's also then for me to go back and critique that in some way or cause uh, cause me to then think about things in a slightly different way. And I think going back to the uncertainty project stuff, I think the biggest problems that humanities have is we do not do good decision making. And I don't mean in a logical way. I don't mean in a first principle way. I mean just in general we don't have good discourse. We don't have a good way of actually figuring out things a lot of the time because we don't do a good job of also navigating the emotion is like the emotions that are involved in decision making and the way that people want things or don't want things. And so I think there's like a lot to do and I think one of those tools will be this, these types of systems in some way longer term.

Speaker A: I think it's interesting that um, I mean this is such a, such a low level example of what you're talking about. But one of, one of the most entertaining things I've done with Bard is, is as, As a former full. Because I was a.

Speaker C: Thank you for using Bard. I appreciate it.

Speaker A: I, I use all of them, I will say because I like seeing the different results. But um, but, but I, but I like. It's easier I have found to have conversations with Bard, um, because the, the guard rails are a little bit less in um, um. Talking about kind of classic philosophical problems. It's interesting to see, to like challenge it and, and try to get into turning its logic back on itself and in it. And it, and it. But it yields sometimes some very interesting responses, um, that I have found thought, thought provoking in myself too. And I, and I, I think that, I mean we're like, this is. I feel like we're, it's, you know, we're at like the, the, the pre kindergarten level of this right now. But you could see where it's going.

Speaker C: Yeah, well, I would, I would argue that like, you know, a lot of mysticism practices are really just trying to help you reframe what you're thinking of. Right. Like, I, I'm not a very, like, I, um. I don't believe a lot of mysticism or spiritualism or anything like that, but I Think like I have a pack of tarot cards that I use and that actually using those as like I've done product tarot readings for people. Um, and you have to be a little bit of a charlatan to read and interpret the cards. But like, once you do that, it's about kind of asking questions about is this. So I just, I just interpreted something for you. Is this the right way to interpret? And this is not very different than me writing, you know, like a one pager and then showing it to an engineer and saying, do you understand what this means?

Speaker A: Right. And uh, this is something that's really core. You know, you get into something here, this disruptive element. I mean this is very core to, to kind of the principles of, of formal red teaming. The. When the army started setting up red teaming, red teams, they made a rule that you could not be on a red team more than. It varied from 18 to 24 months. And uh, ah, they're very emphatic that red teams could not become tenure positions. And the reason was, is because even if your literal job is being a contrarian and challenging people's thinking, if you do it long enough, you will stop challenging your own thinking. And even just using the same tools over and over again to analyze problems will, will start to dig an intellectual rut where you'll start to, to get sloppy, lazy and say like, oh, this looks like this. You know, so it must be that. And so, you know, they put in place mechanisms like requiring teams to use a, uh, different set of tools, you know, on problems and things like that, just to shake things up because it's. And I do think it gets to laziness, you know, I mean this is, this is one of the key takeaways that we get from, from Kahneman and Tversky is, is, is that we, our brains are lazy and we. And therefore, if you give our brain, if you're not careful, doesn't matter how smart we are, it doesn't matter how well educated we are, it doesn't matter how aware we are even of uh, this stuff. If we're not careful, we will lapse into lazy thinking. And so you have to throw a wrench into things is the best way to disrupt that. Yeah.

Speaker C: And I would add two things to this. I think like one of the key characteristics of human beings and the reason why we will, I think, constantly outthink in some ways. Right. Like again, not at chess, not at things like that anymore. Right. But we will, I think we'll out think when it comes to Human values, um, is because we are incredibly lazy at our core and that our job is to constantly, like, satisfy, ah, a certain situation and for us to find the right type of trade off between like, really good decision making and effort versus the idea of kind of like, uh, efficient use of resources in some way. And so I think, like, when we talk about laziness, right, like, And I hold 100% agree with you, the reason why those biases. Right. Exist is that they've evolved over millions of years because of the way that we interacted with our environment. And so they're not bad. They're shortcuts. They. The problem is, yeah, they're heuristics. Right. And, and even like Giga Zehner, um, who talks a lot about like fast and furious, like heuristics and things like that, he believes a lot of these things, but he sees them as benefits in some way. And so I think, like. And I'm sure like, Kahneman, uh, would also say the same thing, right?

Speaker A: Oh, yeah. I mean, my classic example I talk to people about is Kahneman calls optimism bias, for instance, the engine of capitalism.

Speaker C: Yeah, exactly.

Speaker A: If we didn't suffer from optimism bias.

Speaker C: That's right.

Speaker A: I mean, why. If you were Orville and Wilbur Wright and you didn't have optimism bias, why would you, why would you jump into this. This wooden, wooden cloth kite and jump off of a cliff when the last 12 people who did that died?

Speaker C: Well, they were much safer, I think, than the other people that were doing aviation. But, But I totally agree.

Speaker A: You know, you know, it's like, uh, you know, no one would ever start a new restaurant in America.

Speaker C: That's right. Yeah. You know, you're 80% out of business, right, in like 18 months or something like that.

Speaker A: Yeah. And so it's like, it's just, you know, so they. These things all obviously have purposes. I mean, you know, if, you know, if we didn't have loss aversion, you know, we'd all have traded our cows for a bag of magic beans a long time ago.

Speaker C: Well, and to your point, like, I think that's why like, group decision making is so interesting to me as like a dynamic is because, um, there are a lot of. There was a, um, particular study that was talking about. There's a. It's a very specific kind of bias where you're given four cards and on one it either has like a letter or number on it and then you're supposed to guess. Uh, on one side there's a letter, on one side there's A number, um, of a card, and there's four of them. And then you're supposed to guess how many cards do you need to turn over to be able to understand whether these. This particular rule actually takes place. If you're. As an individual, one out of five people will get it right. So it's an incredibly low kind of, like, react. Like, people end up saying, like, too many cards, they turn over too little. It's like these two specific cards you're supposed to turn over. Essentially, when you do it as a group, though, and you have a discourse, the likelihood of you getting it right goes up to 80%. And so I think there's, like, interesting ways that we talk about, like, and there's a whole school of thought around decision making that's really around discourse or argumentative decision, decision making, which is that we need to be able to talk to other people to be able to make better decisions. And this is kind of like how rubber ducking works, if you're familiar with the term, which is. I'm not like, yeah, so. So rubber ducking comes from. I actually can't remember the exact origin. I just remember a story about the origin, which is like, uh, there was an engineer that went into their kind of manager's office and was like, I'm having a problem. I can't figure out what's going on with this system right now. And he says, okay, well, I'm busy right now, but here's a rubber ducky. I want you to just ask. Talk to the rubber ducky about it for a few minutes. And the guy, he's, uh, kind of like this kind of weird. And so he starts talking to the rubber ducky. But because there's like, an otherness that is now created in some way because I'm externalizing it, I'm verbalizing, I'm having discourse about it, even with an inanimate object. I've now taken the problem outside of my head with all of my understanding and now turn it into something where I can look at it as an external being in some way. And so they then figured out the problem, basically. Um, and so this is the issue of, like, just by. And this is something that I think almost every engineer, uh, they'll have that moment where they're, like, talking to someone. Like, I just can't figure this out. Here's the problem. And they're like, oh, I got it. I realize what the problem is.

Speaker A: Just, I do that all the time. I'm famous or infamous, depending on who you ask. Ask for Answering my own questions. Because like you say, just the, just the act of articulating it in getting it outside of your head. But you know, this goes. We'll take a break and come back to this. But what I'd like to talk about when we come back is, and this is something you and I have talked about before, but and it's something we've talked about on the show a lot is this is why I firmly believe that the real opportunity with machine learning, with large language models, with AI, whatever you want to describe it as, is not as a replacement for human decision making, but as an augmentation to human decision making. Let's take a break and when we come back, we'll dive into that. Hey folks, Bryce here. If you're listening to this and you're liking what you're hearing and you're wondering, am, um, I a Red Team Thinker? We have an easy way for you to find out. Just go to the Show Notes, click on the link there to our free assessment to find out if you are a Red Team Thinker and what you can do to think more effectively, to lead more effectively, and to make better decisions faster in your complex world. Like I said, the link is in the Show Notes or you can simply go to our website, redteamthinking.com. check it out. I can't wait to see how you score. Welcome back. Well, this has been such, uh, a fascinating conversation, Chris. And before the break we were talking about AI, machine learning, large language models as an augmentation to human decision making rather than as a replacement for human decision making. And I just love to get your thoughts on that because I think AIs can play the role that you're talking. They can be a much better rubber duck, um, than an inanimate object perhaps.

Speaker C: Yeah, I think you're right about that. And I've written a couple posts recently. I did maybe as a way to think about this. Uh, I did a workshop at south by Southwest a year ago that was called how does the Roomba Really Feel about Dog? And it was kind of this thinking about the idea of uh, animism, which is kind of like a spiritual stance that every object, every object, ah, whether it's animate or inanimate, actually has a spirit inside of it. Right. But taking that to maybe tenet of Shintoism. Yeah, exactly. Shintoism, First Nation, like, ah, a bunch of different groups believe in this type of idea. Pagans. Right. Like it's, it's all, it's all related to that type of, of thing. Um, and so I think what's interesting about it is once you start thinking about this from the standpoint of like for the devices in your home, right. How should a lamp act is the question. Right. And it probably should be able to hear your commands, but it probably shouldn't talk back to you. Right. It probably shouldn't be like Beauty and the Beast where everything's like singing at you when you walk home. But it should be able to like blink its light probably. It should be able to dim, it should maybe change its color. Like those are things that you expect a lamp to do, right? And so my only question about this is that when we start to now combine things from like large language models, we can start to get into this realm of you can do. And I've done some experiments with this a little bit. Like, um, you, you as an agent are going to act like a lamp now, right. And what I want you to do is what would make a lamp happy? Theoretically, right. And a lamp would feel happy when it could be on when someone's in the room, right. But it's also dimmed when there's a baby there that's trying to sleep or something like that. And so you start to create these like really interesting circumstances that every device in your home could in some way start to be like an agent inside of that home. Um, and so I wrote another follow up piece uh, a couple of months ago that is like, once we, now that we have these large language models, we can start to say that all of these different agents could start to communicate with each other in a way that starts to allow for. Basically the big problem with smart home automation is that routines, um, end up being very brittle, right? So like the interfaces, you get to be able to turn off light end up being like if someone enters the room or not, or like this day of week, this hour of day, or like sunset and sunrise, right? Those are kind of the main ways that you end up doing this. Problem is though is that like if you just do it based off someone walking in the room, if there was a baby asleep in a crib there, that's a bad thing to just automatically turn on the lights, right? Um, you have to special case all of these different routines. And so I don't know, I think there's something interesting about what if there was like a household chat where every device was talking to each other in some way, maybe you did it by room or something like that. But you can then also understand a

Speaker A: web of awareness too.

Speaker C: That's Right, that's right. Yeah.

Speaker A: By communicating with each other.

Speaker C: That's right. So the doorbell, the video doorbell should be able to say to like a device inside the home that someone's here or that a package got dropped off or something like that. Right. And the thing that then gets very interesting about large language models is that that kind of translation from anything to anything aspect is that I can then go into this like household chat and I can troubleshoot what happened. And then I can say like, so why did you turn on the light? The baby was asleep in that room. Right. And it would say, well, because someone walked in the room and that's what I'm supposed to do. And uh, then I can say, well, no, if somebody's asleep in that room, do not turn on the light when someone walks in.

Speaker A: And I just create from that.

Speaker C: That's right. Well, and learning is also, you know, I mean, personification of these things. I, I always, it's. We don't have the right language to talk about this yet because the way that these models work today is that they can, it's not so much learning, but it's that you can now add to this, this text that is kind of the prompt engineering addition. But yeah, like longer term reinforcement agents and reinforcement learning, for example, they will learn actively to be able to do a better, better job. There's a lot of work to still do around, like how does that get federated? How does learning work on devices that are very power constrained? Um, but I think there's some really interesting concepts that start to come out that are like, how do we deal in an edge and a way? And so to your point, going all the way back to your point about decision making, augmentation of decision making, why don't you have a series of agents that are in the same chat as you that are maybe helping? So the idea of trying to make a decision about something. Why wouldn't you be able to have an agent that could bring up past decisions that were like this one? Right. Or another agent that could double check the math that you're talking about and estimations. Right. Um, why couldn't you have an agent that goes out and retrieves some information that might be interesting about this? Right.

Speaker A: Well, you know, you, you, you and I had had a really interesting lunch a couple months ago in San Francisco. Yeah. And we started talking about this concept and I, I wrote a piece about this, this afterwards based on something that you got me thinking about with this, which is that, and we talked about it then is what if you applied this to the military decision making process, which. A key tenet of MDMP is that you have to come up with three different courses of action. And what if you programmed agents with different personalities, one incredibly risk averse, one incredibly aggressive and not constrained by risk, and another with another personality. And instead of having human planners come up with those three courses of action, those agents could come up with three course of action. But then you as the human decision maker still are the arbiter that judges which of these three is the, is the, is the one I want to take or what is the fourth course of action that I as a human being am going to synthesize from these? I think that could be incredibly valuable. Um, and I think it could be credibly valuable not just in a military setting, but in business settings as well and a lot of other settings. And that's, that's it. This is the stuff that I find so exciting because, you know, it's going to change what's possible in terms of the time horizon for thoughtful decisions, I think. Yeah.

Speaker C: And I mean, there's going to be some interesting kind of trade offs with this as well because like a technique that I use whenever talking about strategy with a team is after they've kind of formulated what they believe their strategy or decision making model is, right, Um, I then ask them to pick kind of five random people throughout their entire organization and then try to think about how that person would contextualize that decision making model. Right. And it could be like, in the case of like a technology world, right? It could be like a customer support person, an engineer, it could be a vp, it could be um, someone that does like procurement, right. What would that mean for all these people to do that? And I think we get into the realm of kind of what is referred to as digital twin, and there's a lot of things around like the idea of an agent that ends up simulating what you would want in some way. Um, there's a lot of, kind of, there's a lot of questions that start to come up about whether you should be consenting into that type of process. Right? Like if you do it yourself and you own it, and you actually own that agent in some way, you could then decide how you want to use it and when to use it. And that's a very important aspect of automation. Um, but like in your case about the idea of the coa, you know, would it be interesting to then pull digital twins of every person that's inside your company to understand, like, how would that impact their particular part of the plan. And what would they say? Um, now again, there's a trade off here, right? Because it's not going to be the true thing that that person would have said, but it's going to be a simulation of what they could have said.

Speaker A: Well, that's where you get into another important bias though. Which is, which is automation biases.

Speaker B: That.

Speaker A: One of the things that I think we're going to have to increasingly guard against as we deploy systems like this is people will have the tendency to forget that. People, because we're lazy, will start to forget that. It's not, it's not the truth, it's a truth about that. And, and that's the danger.

Speaker C: Yeah, well, and, and this is why I think when we talk about organizational dynamics, when we include things like agents inside these organizational dynamics, even though there is like potentially a digital twin for you, that could speak for you in some cases, right. In certain cases you would definitely not want it to. And so I think what we end up doing is we then start thinking about like who should be part of a decision making process is a really important question, right. I call it like deciding how to decide. But a lot of the time inside of companies, they end up creating a very muddled process which is the discourse and the decision maker and the communication, they're all just like jammed into one process which ends up turning into like a thousand meetings things. And so unless you end up start splitting those things out, and this is something I'm writing about for the Uncertainty Project right now, which is like just the discourse part of this. How do we separate out the discourse of a decision? Um, I think that's where to your point, like who should be included and is it which humans should be included and which agents should be included and which data is necessary, right, like to be able to make a better decision. And in some cases it just becomes like the thing again, going back to the inherent laziness of human beings is that there are going to be times that we also want to experiment with things and there's, there's like domains that we don't have really good knowledge about. And so how do we use some type of like abductive thinking which is this idea of I have some evidence of something, but I make an imaginative leap somewhere based on my experience and my background. And so, um, that I think is like, that's where again, the human is much better at this type of abductive thinking right now.

Speaker A: But it's still fascinating to get these inputs, you know, as Long as you're aware. As long as you recognize them for what they are. You know, it's interesting. I was just there was there. I was reading a piece a couple weeks ago in Vanity Fair. It was their turn to do the piece that everyone outside California feels compelled to write now about, you know, oh, how. How is California going to survive its current death spiral? Um, but it was a bit more thoughtful than most of them. And one of the things that the author did, um, as part of the piece was have, uh, one of your competitors, ah, AIs, write a piece on the state of California today, um, as though they were Joan Didion. And it's uncannily. It sounds every bit like Joan Didion wrote it, but it's also very insightful because it's all about California today versus the California of my youth and stuff like. Like this. And, you know, and as long as you recognize it. Yeah, that's not actually Joan Diddy. Unless, you know, stipulate that it's still. I could see, like, you know, it's interesting to think about what would. What would she say about this, you know, right now? You know, and then what. What does that make? What. What prompt does that provide for me to think about?

Speaker C: That's right. Well, and I mean, again, another strategic kind of concept that I end up using is, um, if you put yourself in someone else's shoes, again, you kind of get to see the otherness of the thing that you're working on. And so I've asked, like, what would it. What would a famous person think about your plan?

Speaker A: Right.

Speaker C: And, um, so I think there's something interesting about that. Like, I. If I think about kind of the people that I try to think about, like, what would they say about this? And there's a little bit of, like a parrot. There's definitely, like, parasocial relationship issues about this type of thing. But, like, you know, what would Dave Snowden think about this plan?

Speaker A: You just gave me an amazing idea here. There's a tool that we teach called Four Ways of Seeing, and it's designed to look at problems from the point of view of other stakeholders.

Speaker C: Yeah.

Speaker A: And one of the challenges of it is how do you, you know, the subjectivity of it? You know, like, how do you really know what they. What the other thinks of the situation.

Speaker C: That's right.

Speaker A: But I. I just wrote a note to myself that. That using. Using large language module model, uh, uh, agent to play the Persona of the other.

Speaker C: That's right.

Speaker A: And ask them what they think of it. I think it would be such a fascinating thing.

Speaker C: Yeah, absolutely. Like, what would John Boyd. What would John Boyd think about this? Right? Like, um, what would, what would, what would these people that are both like, living and dead and even just like other famous people, you know, that be.

Speaker A: That brings up a question that, that, that, you know, I'm obviously a layperson in this arena, um, though a very interested one. What leads AIs to lie?

Speaker C: Yeah. So again, I'm not, uh, as technical as the people that are on my team. There are engineers that really understand. And if I were to I guess, uh, say what one of the engineers that I know would probably say is that these things are really just taking an input and then they're running it through a function that then creates an output. Right? That's really all these things do. And so the problem goes back to what I was saying about we, we don't really have good language for this right now because what we do is we say that this thing is lying or this thing is hallucinating or this thing is imagining or dreaming or something like that. Right. Like that was the original one, was

Speaker A: that none of which are really accurate statements of what it's doing. That's right.

Speaker C: Because it's, it's not. It doesn't include a human that is doing these processes that is constantly also evolving and moving. Because, like, I think that's the other thing is that we are existing or perceiving time in a certain way, that we're constantly changing as people as well as we actually like move the world. And there's probably like some things about like free energy principle and all these other kind of interesting theories that come into that. But.

Speaker A: Yeah, well, I'll give you an example. You know, I asked, I asked a large language model. I said, I said, you know, give me 500 words on this topic. Include at least two quotes from relevant experts in the field. And it did. But then when I checked those quotes, they were not actual quotes. The experts that it chose were actual experts in the field. I could find no evidence that they said any of the things that it quoted them as saying.

Speaker C: So the reason why is because a lot of the time, um, these things are meant to be. They're probabilistic in such a way that they are trying to guess at what the next word. And again, guess is a very inappropriate word to use in this case. But it's, it's saying like the, the there's a probability, distribute like distribution of different words that could come up next after this body of knowledge and There was actually a really great post by someone where they were talking about like, what would be the actual embodied experience of these things being. And what it would actually be is that you would be loading. You would have a word that would be in front of you. You'd be slapped. You're Rip Van Winkle. Okay, you've been asleep for three years, but you know everything previous to 2020. Um, you're slapped awake and you're shown one word and then you go back to sleep and you say a word. Sorry, you get slapped awake. You see one word, you say the next word. Or you have a distribution of words that are of different levels of temperature or something like that, and then you go back to sleep. And then if you have another word that was before that, you're now slapped awake again. And you see two words, the one that you wrote, but you don't realize you wrote it and you're now writing another one. And you do that for however long, you end up generating this content. So the thing is that it's because there's lack of data, right? It's because there's. You're asking a prompt that has maybe lower confidence. And then also these models are not built to actually reference real things today. It ends up being kind of like a big jumble of just like words that are then trained now the sequencing of those words is what matters. So that's why like a lot of time people call it kind of like a, ah, really fancy autocomplete is really what it is.

Speaker A: And that's why I would imagine it's, it's. These large language models can be very effective at saying, write a piece in the voice of Joan Didion, because it doesn't actually need to know anything specific that she said. It just needs to analyze all of her extant works and get her voice and understand her perspectives on things, but it doesn't actually have to reference them.

Speaker C: Yeah, and so that's what I would say is, I think it's, that's why it's such an interesting technology, is that it probably shouldn't be used for a lot of the cases that people are using it for because of the failure cases. Um, but again, the fact that it can in some way synthesize human language or synthesize kind of human communication I think is going to be a powerful tool long term. It's just that when and how we use these things and how we put constraints on them. So again, going back to the agentive kind of group of agents type of thing, like the way A lot of these systems work today is that there's a bunch of different things that end up filtering input into the model and there's a bunch of things that filter output from the model. And that's usually to remove things that the policy does not want to deal with. Right. So like pornography is a good example of that. Right. Or other things.

Speaker A: Here's an example of how I use this that I think is, I may be wrong, but I feel is like an appropriate use, uh, uh, given the current state. As a, as someone who, who, who among other things is a writer, I'll, I'll be writing something about a topic and I will say give me 10 bullet points of things that's, that you would want to think about if you were, if you were contemplating doing xa. And then I'll, I'll look and I'll check off. Yeah, I already, I talked about that. I talked about that. I talked about that. Oh, that's interesting. I didn't think about that one. But then I'll go and I'll write, I'll, I'll write what I would say about that. But it's, it goes back to what we were talking about a few minutes ago. It's using that to, to, as the rubber duck to kind of like disrupt your thinking and think about things you might have missed. And that again is what I see as the ultimate opportunity with this.

Speaker C: Yeah, no, I think you're, you're right about that. And that's why, you know, one of the things I've done with my writing to use this is there've been a couple of times where I wanted to write a post about a particular topic, like in product management or decision making. And so I've like posed it to the uh, system and said, please write me an article about this. Right. In this kind of tone. And so I would then read the article and then I would basically write about anything else that it didn't say. And the reason why is that like, whatever it has inside of like all of its training set of data, um, is kind of just like the average of the Internet. And yeah, you can make it do spicier stuff, but I think there's like, like one, I think differentiating is going to be more interesting in the future. And then I think to your point, that kind of adding a little bit of confusion to reframe what you're thinking about, I think there are some like, really interesting aspects of this. But getting back to your point about like when you should or should not trust these systems, is still not like a perfect. I guess this is more of a human computer interface type of problem. And it's a lot about like interpretability of the state of the machine. But I think there's like, there's, there's a great paper by Paris, ah, I think in like 96. And at that point most automation, and this was, it was all about trusted automation. And at that point it was mostly like aviation systems, military and academic. Right. So that's where they were pulling from when it came to this stuff. And the problem is, is that you trust either too much or too little the system. And so that dialing in of what is the appropriate level of trust not only for the system as a whole and my experience with the system, but also in that exact moment for those exact circumstances becomes incredibly difficult. And so I think there's a huge amount of work for us to do within the industry around that because like, I need to be able to interpret when this system is probably out of its realm of actually doing a good job. Right.

Speaker A: Like, Right. And I'm familiar with that paper and I actually wrote a bit about this in my last book, uh, Red Teaming. I mean it's very troubling. You know, uh, more than 50% of pilots in simulators will make errors one way or the, or the other based on either trusting the, the automation too much or not trusting it. And, and it's, and these are all, all life and death errors. And it's really terrifying how regularly that happens. And if you unpack most of the aviation, major aviation incidents that we know about, like we don't know what happened to uh, to Malaysia Air for instance, you know, but you know, particularly like I think about the Air France uh, incident off of Brazil, that was completely a case of very experienced pilots trusting the, the automation rather than their senses and doing things that, that, that, that a cadet in flight school and a Cessna would have known is wrong. Based on just listening to what was happening on the cockpit voice recorder. You could hear everything shaking in the cockpit because the plane was stalling, but the automation was telling them to continue to pull up. And yet, I mean like, uh, you know, anyone who's, who knows anything about flying knows that that shaking. And you could hear like one of the flight manuals fall off a shelf and stuff like this. How did they not recognize that they're stalling the plane, but they stalled the plane.

Speaker C: Well, and maybe adding into this as well is that we also do the same thing when dealing with humans. Right. And we, we have different ways of Dealing. So like the, the hero of the, uh, the hero of the Hudson.

Speaker A: Right.

Speaker C: Um, he was just, he basically ignored all of the, the advice from the air traffic controller, uh, which was to return back to the uh, the airport. And he just knew that he was not going to be able to make it. So he's like, you know what? I'm just gonna, gonna land here.

Speaker A: Well, you know, it's fascinating. I did this is, I actually did an episode with, with Gary Klein on this topic because Gary, Gary sat with him, Sully Sullenberg, and, and Gary spent several days unpacking this with him. And what he did is even more interesting than that. He kept coming up with hypotheses and then rejecting his hypotheses as new data emerged in a very systematic way. So, you know, yes, they told the initial thing, they told him to come back and he, it took him, uh, you know, 10 seconds to say that's not possible. So he asked for a vector to Teterboro and as soon as he got on the vector to Teterboro he, he, he assessed the situation and realized there was no way he was going to make Teterboro. And, and it went to, and you know, he did not, he did not want to land in that. No, but he went through these options very rapidly but consciously and ultimately there was no other option left.

Speaker C: Yeah, the, um, the idea of decision making and the way we end up making decisions, um, I think ends up coming down to usually one individual making a decision but assessing the information. And I think that is another aspect of this whole process that I think gets very muddied inside of technology companies like uh, the idea of consensus driven decision making is something that's very common inside the technology world. And I think that's a mistake to think that when the reality is like you want to have kind of one person assessing and rejecting or accepting options but making a decision in the end. And um, that should depend on their expertise, that should depend on their decision making authority. There's a bunch of things that go into that, but in this case that's why you have a captain right on an airplane.

Speaker A: And that's why, you know, I firmly believe that applied critical thinking skills are going to become the most sought after trait in leadership as we continue down this path of having increasing opportunities to leverage machine learning, leverage large language models and such like that, because that's what is going to be the thing that differentiates is having a human decision maker at the end of the line who has really good critical thinking skills to be able to assess all that information, all those options, and then make the final decision. There's something I want to talk about. I know we're getting short of time here, but, you know, you talked about, um, kind of throwing a wrench into things, um, as a way of disrupting thinking and, and leading to innovative insights. Talk to me a little bit about randomized constraints, which is kind of gets into that area as well.

Speaker C: Yeah, I mean the, the thing that um, I've, I found very interesting is, is this concept of just that we usually, you know, as, as kind of Kinnaman would say, or in any. Is this idea of like optimism. We're very optimistic about how things will work out, especially for decisions that we end up making ourselves. And so, um, I think, uh, Gary Klein with the pre mortem is a great example of trying to push on like, what, how will we fail? Or how could we fail? Um, but one of the things that I've used as a technique is, um, to really start to apply random constraints to a particular, uh. Like when we do a bunch of ideation, we do a bunch of what's referred to as like divergent thinking.

Speaker B: Right.

Speaker C: We're trying to get as much variance and ideas as we can and then we start talking about like convergent thinking. We're trying to now say, based on kind of how we value things. And that's usually your strategy or like prioritization or something like that. Right? Like okrs could even be another way that you do that.

Speaker A: This is, this is Red Teaming 101 is moving from divergent thinking to convergent thinking.

Speaker C: Exactly. And so that idea of randomized constraints I think works really well in the convergence because if something like X would happen, what would that mean for the solution that we were, the solutions that we were considering? And so that could be, you know, in the case of say, technology is like, what if we only had two days to build this? Right? What would you do then? Like, and that, that, that is a very much like a constraint in the way of like, what is most important about this actual thing? Or what do we need to learn? Or what do we need to do? But the other side of this is like, what if I gave you like an unlimited budget to do this? Right? What would that mean? And so I'm playing with constraints in a bunch of different ways that I think like, uh, you know, Dave Snowden would talk about them more as like governing constraints. Um, I think there's like, again, I still am trying to internalize the way that like emerge emerging constraints and governing uh, constraints actually interoperate with each other. But I think like, this idea of playing with like, what are the rules of the system end up helping you actually understand what's valuable. And, and I've even taken this, I've even taken this to the idea of like, I, uh, want you to come up with a bad idea right now. And the idea of a bad idea is that there's a very complicated way that you evaluate what is good inside of your head and you use probably many, many dimensions. And so saying what is bad is. You're trying to take the opposite of that. But, but it's not always true that that thing is a bad thing, because if you relax one of the constraints that you would have had otherwise, that could be a good idea. And so the example did like a session that was like, around adversarial product management and at the beginning of that, trying to bring home the point about bad ideas is I asked everybody what would be some bad ideas, uh, with regards to a meetup, right? And so someone said, don't advertise it at all. And I said, well, that's a bad, that's usually a bad idea. But what if it was like, actually a very exclusive thing, right? I only wanted word of mouth to actually like people to hear about this. That would be the way I would do that then, right? Or what if you like, you only, you only want to be like it. You want to be good, you want to be nice to people, you don't want to be rude. And I was like, well, I've been to like many restaurants. Like, there's a restaurant called Dick's Last Resort, where basically their job, the waiter's job, is just to like, insult you the entire time. And people seek out this restaurant to do that, right? Where you go to like, stand up comedy, because people are.

Speaker A: I don't get that, by the way, but I know it exists.

Speaker C: And so I guess, like, the point is, is that we're taking things that, that are potentially norms or assumptions about the way the world works, and we're flipping them on their head in some way because the constraints are actually really key to the way we end up making decisions.

Speaker A: I love that, I love that idea of playing with constraints. This is something that we, we talk a lot about in our work. Um, and it's also something I practice both professionally and in my personal life. I mean, you know, you know, one, one of, one of the, the. You, you, you just, you just used one of the, the, the prompts that I used in my, with My wife often when we're talking about whether we, certain decisions that we're making, you know, and like I say if, if we had unlimited cash, what would we choose here? You know, and then it's helpful because you know, it's like, are you, are you making a decision based on a constraint and is that constraint a useful constraint? Um, or if, if, if, if we had to make this decision in the next 30 minutes, what would we do? You know, um, and playing around with those things, like you say that it's incredibly useful I think, um, to do that. And it's, it's, I, I think, you know, again all these things are about kind of shaking us out of the lethargy that exists within our mental processes that we're not often aware of. And anything that does that can yield, yield really valuable insights, you know, and you know, this idea of what's, tell me a bad idea. There's a similar technique that we do about, we call the enemy within and it gets into looking at, you know, if you were, if you were insurgents that were sent by a competitor, you know, you know, and you know, like, let's say, you know, like, you know, if I was doing this with Google, I'd say, you know, imagine that you are a team from Microsoft that, that cunningly 15, 20 years ago, Bill Gates, you know, convinced you all to get jobs at Google and you've worked your way up the Google ladder, quietly biding your time and now you're in senior positions in Google and you've just, the balloon has gone up and you've been activated. What would you do? What non destructive things would you do? There's the caveat, you know, can't poison the, poison the water supply or things like that to bring Google down. And then after you do that analysis. All right, now what are, what are those things is Google doing right now? And it's always, it's always stunning to organizations how many of those things they're actually doing right now.

Speaker C: I love that CIA, uh, M manual that was from I think the six.

Speaker A: That's what inspired this.

Speaker C: Totally.

Speaker B: Yeah.

Speaker A: It's the OSS manual from World War, from World War II.

Speaker C: That's right. Yeah. And so like that's something that, going back to design fiction, I've been kind of exploring, um, been using this deck of cards called the Work Kit for Design Fiction by the Near Future Laboratories and combining that with uh, a kit from the ready, which is all about organizational design. And so I've been looking at tensions and then what would be a Future way for us to think about how organizations might change to address this. So one of them was too many meetings. And so, um, I ended up writing a piece that would be a press release about how everybody got a neural link and that we installed some new software that if you were attending a meeting that you shouldn't, you would actually feel physically bad. Um, and so what would that mean for that type of thing? But, um, let's take it like a step further. Like, let's talk about like, that's just like technology, right, in some way. And, and again, very, very bad technology. Like, again, this is not meant to be like a positive use case of this type of thing. But like, what if we start talking about the way that we actually have expectations about the way we do work? So if we have too many meetings. Um, you know, one way that Shopify has started to deal with this is they think of meetings as a bug, which I don't agree with. I think meetings are very valuable in some ways, um, or neutral. And they can be good or bad. Right. Um, but like, rather than just say canceling all meetings, what if you said that you could only have if you had a meeting, it had to be your entire group and that was the only way you could ever have meetings. What would that change about the way you actually do? Information flows in some way? Or what if you can never have a meeting above one on one?

Speaker A: Well, here's a, here's a great example. An artificial constraint that one of my, one of my clients uses really effectively. Um, I was working with a private equity firm down in Texas and they love pre mortem, but they also have a tendency to go down rabbit holes, as I think all of us deep thinkers do. So they came up with something they call a hall brawl. They conduct their pre mortem standing outside of one of the director's offices. And that's the constraint because after about 20 or 30 minutes, people don't want to be standing in the hall and they, and they do it, they don't do it in a private hall, they do it in a hall where, you know, admins and stuff are walking through and stuff like that. And the whole point of doing it is to make them, make it, uh, as erudite as possible still to have that rigorous discussion, but to do it in a way that, that it's going to be uncomfortable to kind of have an endless discussion. And, and the other constraint is whenever, when, when they decide it's over, a decision has to be made.

Speaker C: Yeah, that's cool. That's Great. Um, and I, I mean it's very much in the line with uh, like standups or something that are used a lot in the technology world to kind of like get it, get an understanding of what's going on and build kind of understanding between each other of what's, what everybody's doing, but they can like drag on and on. And so one of the things I've seen some people do is like, uh, everybody is actually, um, while you're doing your stand up, you have to be planking at the same time.

Speaker A: I was just thinking about that. Yeah, Yeah. I just shared on LinkedIn, uh, somebody wrote something about that. I shared it like two or three days ago. Yeah. And you know, this is not, you know, we think this stuff is also new, but you know, Alfred Sloan, who was, who was the legendary CEO of General Motors back in the, the 50s, um, Alfred Sloan and even the 30s, I think I can't, I will have to fact check that.

Speaker C: Um, but don't ask, uh, a large language model.

Speaker A: Okay, we won't ask. Um, but Alfred Sloan had a rule that I. It's something I've advised a lot of CEOs to adopt, is very simple rule. If something, if somebody poses a question in a meeting, should we do X? What about this? You know, we have a problem with. Yeah, a decision has to be made within 24 hours. And he was religious about it and he was famous for, you know, somebody would come up in a meeting, he'd say, what, you know, you, you. And you think, let's figure out what, what we're going to do about this. And he would, he would stop people in the hall the next day and be like, you got 12 more minutes to make a decision. Now the decision could be, we don't have enough information. We need to set up a task force to study this. But that was, you know, he pushed back on that. But you, you know, he wouldn't, it wasn't, it wasn't so extreme that he wouldn't allow that. But the point was, is a decision has to be made in 24 hours and think about how many great ideas we have that just get lost because we can't make a decision. And then people forget about them.

Speaker C: Yeah.

Speaker A: And, and if he set up a task force, his rule was it has to have a stop date. So like, okay, we don't have enough information to make decisions. So in seven days from today, the task force will report back and then we will make a decision. And I think that discipline is incredibly helpful.

Speaker C: The time boxing. I think is really important too for this type of thing because decisions can drag on for forever.

Speaker A: Right.

Speaker C: Um, I mean we do a lot of work where we have like a Google Doc that then has a lot of commenting on it. And so one of the things I've wanted to experiment with is uh, that you know, one, not all comments are equal. Right? Like I, I, it's not true that like every comment is of equal value. And then second of all, like you shouldn't have to answer every comment. Right? Like it's, it's, you should definitely try to hear everybody that wants to provide some information. And then finally like there should be a time by which you actually clear these comments or, or at least like archive them in some way so that if you wanted to go back you could understand like what was the decision or how did this like comment actually change something. But I almost imagine that there should be a suite of tools that are available that are like, there's like a countdown timer on the top of the uh doc and that at that point all the comments are just resolved automatically. And so how would that change the dynamic of the way that we end up having a discourse that is say, asynchronous rather than synchronous. And that combination of asynchronous and synchronous then I think also plays well together. And then you start adding in agents that start to, if there's a question it, there's an agent that tries to answer it. And um, there's been some interesting work too about like how to use these types of agents in uh, education. And so it's not so much about the idea of like taking these to now they're just going to do everybody's homework. But it's like you are now the teacher as an individual and you need to critique how this large language model is actually outputting and go and research like did it get these things right or wrong? Right. And that type of teaching is very interesting.

Speaker A: Yeah, it creates opportunities, it creates challenges too. I was just thinking, as you're describing this, um, I, I, I never, I, I bounce around through different email uh, uh, clients all the time. And I spent a uh, uh, a long time using Spark and Sparks Sparks Party Trick because it's, it's driven by the Inbox zero principle, is that once you, once you see a message you have to act on it and if you don't, it's archived. And in practice it was exactly what I wanted, you know, because it's force, it's creating that artificial constraint that forces Action. But what it ended up doing was stressing the hell out of me, like really seriously stressing me out every day when I was doing my email. Because it's like literally when I say you have to act on it, if you go 60 seconds on the screen, it will archive. And so if I got a phone call when I was in the middle of reading an email, I'd be like, shit, you know, there. And you could, you know, obviously you get it back. But the point is, and it's interesting, they then ended up making that an option rather than a non toggle. But I mean that, but then, but then what's then is then it's no different than any other email client, you know, type of thing. So it's like I, I, I think that it's interesting to play with these things. Asher, there's so much more we can talk about. I, as I've said to you before, I could talk to you all day, we're going to have to have you back on the show. We've gone over our normal time here, but it's been really enjoyable. Chris, thanks so much for taking the time to talk with us today.

Speaker C: Yeah, thank you. And yeah, and I'd love to, love to come back as well. And uh, yeah, thanks for the time. I mean, I think there's a lot of really interesting stuff that's going on around not only large language models, but I think we're getting to the point that decision making, um, is something that we should take more care with. Um, and we shouldn't just like, like sleepwalk through it. So anyways, I definitely want to plug like the Uncertainty Project again. I think it's like a really interesting place to talk about that. Um, but yeah, it was really great to talk about all these things.

Speaker A: We'll put a link to that in the, in the show notes. Absolutely. Yeah, it's wonderful. Thanks a lot.

Speaker B: Thank you for tuning in to the thinking leader. Uh, check the show notes for more information about the topics covered in this episode. There you'll also find a link to our free assessments. Click on it right now to find out if you are a red Team thinker, um, with a red Team culture.