Building Explainable AI with Beth Rudden - CEO at Bast AI

Speaker A: Foreign. Welcome to trending and Education. Mike Palmer here. Delighted to have one of our trending and ed All Stars, a friend of the show, someone I've, uh, had on a panel at south by Southwest back in the day. And I'm delighted to have Beth Rudden back on the show. Beth does a lot of things, uh, probably first and foremost, CEO of bast, AI. But, uh, let's dispense with the amenities here. Beth, who are you? In case folks don't know. And, uh, it's, it's amazing to have you back on.

Speaker B: Ah. Uh, amazing to see you. And if nothing, I have some new ways to explain what we do to test. So I, I, I brought something. People ask me all the time. They're like, what does BAS do? And, you know, I typically am like, we're middleware. We're a layer. We verify, we make AI safe. We're into, we're into explainability, we're into engineering, we're into practical application. But, um, this gentleman, and I'm going to tell this as a story because it's always best as a story. And this gentleman had such an amazing entrepreneurial journey. He and his mom went to Paris to learn how to make, uh, drinking chocolate. And if nobody has ever had drinking chocolate, I highly recommend Paris, Italy. You know, drinking chocolate is the bomb dot com. And so he came back to the United States and he started with an R and D lab and quality control group to start making chocolates for some of the biggest brands. So think about a Starbucks mocha or Tim Hortons or Pete's or, um, you know, all of, all of the things, you know, Einstein's and, you know, Panda Express and, um, you know, all those things. So he built an Empire using 12 people to do R and D quality control and operations, to serve over 6,000 brands with all of the manufacturing. Because he's like, here's your chocolate in a powder. Here's your chocolate as a syrup. Here's your chocolate as a syrup in a pop bottle. Here's your chocolate that goes with this coffee, or here's your white chocolate, or here's your chocolate caramel ribbon or da, da, da, da. Uh, and I'm like, oh, my God. That's what we do. We put explainability in AI that's exactly the same. Like, and we're like a small, mighty team. And we, we, we're AI, we're LLM agnostic. So we're like, bring your own model if you want to use chatgpt, um, if you want to use llama if you want to use Minstrel, if you want to use your own little SLM that you created, if you want to use database, you can do almost anything. But what we do is we add the explainability. Just like you add chocolate to coffee to make a mocha, we add explainability to AI to make explainable AI. So that's what we do.

Speaker A: Love it. Yeah, explainable AI. When I think explainable AI, I think Beth Rudden and uh, Bast. So uh, you've been doing it and you've been doing it for a while and folks may not have, you know, we'll include links to previous episodes in the show notes. But you got a pretty interesting background prior to founding Bast. And can you tell that like a lightning round version of that story?

Speaker B: At the time I was an archaeologist and I was digging in the dirt in Italy. And um, then I came and did my graduate work in Colorado and got to do some really cool work in Ludlow. Um, people should look up the Ludlow tent colony massacre in 1914. And then I completely and totally 100% sold out and got a job as a programmer. And I was. Do you remember back in the day where like you were put on a plane with like, there uh, was no stack overflow, no Google. There were books like you had books that you were put on a plane and you're like, you're debugging Java tomorrow. So here's the book that you read on the plane because you know, people back then didn't have smartphones in the um, what do we call that, the late 1900s, 20th century. And then I made my way to IBM and that was my home and I grew up there and I got to become the chief data officer of the managed services division. 34 billion dollar division ran around the world. Um, I was a, ah, technical executive, a global technical executive. And I did a lot of work in Central and Eastern Europe and had teams in Belarus and Kiev and Brno. And then I also ran 2P and LS and then I graduated IBM. I learned everything that I could possibly learn. And then I opened VAST in October of 22. And we are, you know, we've spent about three years really nailing down and building our product because we wanted to be a team of 12 people that does R and D and quality control with our product because we know that you can build an agent or a custom GPT model and you tell the custom GPT model to use this file as a knowledge base, for instance. So you can do the same thing now in our software it's so cool. You need to check it out.

Speaker A: Imagine check it out.

Speaker B: And so chatgpt, if you say, well, how did you get that answer? It makes up this beautiful narrative on how it got its answer. Yeah. And it's totally possible and it makes complete sense.

Speaker A: Yeah. It reminds me of me in college when I hadn't done the reading. I could come up with really plausible answers. They would sound good, but they weren't necessarily based on fact.

Speaker B: What we do is, um, we take that knowledge base and we put it into a database. And so when you get the answer, you say it was pulled from this paragraph on this page of the knowledge base. And we know that because we pulled it from the database that we put it in. So we're not making up, uh, the answers. And it's still super hard. I, uh, mean, I am regularly catfished by AI regularly. Where I find myself, I'm like, damn it, why didn't you answer me correctly? Why are you lying to me? And it does not have a notion of what. It does not understand the meaning of your words. And so I wrote this really great article. 63,000 views, man. Like, it's a, it's a, it's a pretty good one. And um, it's. A friend of mine made this picture that I had to use and it's all the logos of like OpenAI anthropic everything on catbuts.

Speaker A: I saw that one. Yes, yes. I love that one. Yes. The graphic. The graphic is beautiful.

Speaker B: Article is like seven ways that you know that or that you think that the AI understands you. Because there's so many people who are writing prompts to tell the AI not to left m. And it's, it's just, It's a pathetic fallacy, Michael. Like we can't. How do you explain that to people that really believe that this is math and science and objective and you know, what do you think about this? Have you been catfished by AI?

Speaker A: I certainly have. And the thing that's interesting to me is it has changed even the big models. Um, I mostly use Gemini now, uh, just because I'm already sunk cost into Google workplace. And of course they're always trying to upsell me to like there's always better versions with more like quote unquote, truth inside. I have noticed more and as I've done more research and I've talked to more people that there's been a lot more of the rag application, so the retrieval augmented generation. So like a lot of checking of the AI's tendency to confabulate against a more definitive source. But even within that, you know, just this morning I was asking for, you, uh, know, I use it to draft show notes frequently, so I'll drop a transcript in. And I was asking specifically for links because I needed to get the links right. It was very clear. The researcher who was citing her research was delineating five specific, you know, research studies and applications. And I think four out of five were just, you know, the links did not exist. So I asked for the link, and it just made up something that looked plausible but needed to be checked. So I would say two or three years ago, it would have. It would have been five for five. Complete horseshit. So now it's. It's kind of cleaning it up. I feel like they're building the checks in, but it's. It seems really inefficient after the fact.

Speaker B: It's sort of like sprinkling chocolate on your coffee instead of putting the chocolate, uh, in with your coffee with milk, in a cup, you know, like, I mean, so it's like, yeah, uh, do you want a mocha or do you want some chocolate sprinkled on your coffee? Yeah, and. And then it's not. I don't know, it's like some shitty powdered version of chocolate that is not made well because it. It doesn't. You know, and. And I've said this for a very long time, that the way that these models should be used is to enrich information. And so one of our, um, one of the ones that I like to talk about, just because it's. It's just. It blows my mind how much just practical application we can do with a small language model that just takes an output and translates it into a different language or into an analogy or into eighth grade. And so we've gone in and done, um, some work with Craig Hospital, and they're a premier neurospine rehabilitation center. And so, you know, the tragedy of somebody becoming a paraplegic, you know, they've been handed binders of paper and handing a paraplegic, anything is wrong. So we are working on a really hard problem of making, um, the quality of life better for patients by giving their caregivers and themselves information in a way that they can understand the information. So if you are, you know, trying to figure out it's, uh. 70% of people do not understand outpatient procedures. So take your pill three times a day. Does that mean, do you get up in the middle? How do you do that? Like, you know, what. How does this work? And Especially if you are, you know, um, your life has totally changed, and you need wheelchair ramps and, you know, the PT and the ot and it's really complicated. So Craig Hospital, it's a research institution, and they make all of this information available, but what we do is we allow people to interoperate with that information in Vietnamese or in Spanish or in 8th grade or in analogy or in, you know, something that they can auditorily understand. And to me, this opens up such a whole world of, uh, figuring out, look, we don't need the AI to give me the prediction of what those outpatient procedures are. We need those translated or enriched so that the person can understand that information. And for that, you don't need powerful models at all. You just need something that can handle what it was built to do, which is to generate transformations, to, you know, generate the enriched data in a way that makes it comprehensible. And I don't know if I did this to you, but, um, I've been thinking about it for a while, and I'm like, you know, there's a joke here that everybody in the world has been thinking about how they can push out their signals and, like, you know, engage people in this way. And I'm like, has anybody ever asked how people want to receive information? It's like, come on. I, uh. I don't know if you and I have talked about that before, but I'm like, this world that we live in is not. It's so hard. Why do we have to make it even harder by making people guess whether the answer that the AI system gave you is correct or not? That's. That hurts my brain all the time.

Speaker A: Yeah. I mean, it reminds me of, uh, the idea of epistemic security. I thought you might like that. Dwayne Matthews is the guy. He's. He's worth following. He. He's got to turn me onto this idea. But it's pretty much what you're saying, like, we don't know what knowledge is anymore. We don't know what truth is anymore. And sort of the underpinnings of our sort of understanding of what truth is, uh, are being shook. And then, if I'm hearing you right, it goes back to the old adage, garbage in, garbage out, right? So, like, the garbage was this massive, somewhat unlawful scraping of all of the data in the world to the point that there's too much in there and it requires additional processing and all the junk that we're talking about. But then also the outputs were so messy. And then, uh, to Patch on top of that, all these frontier models are now sort of making it seem like the initial output was clean by doing all these additional checks. Whereas if I'm hearing you right, you're saying if you know your source truth, start there and build your model there.

Speaker B: And we give you that ability to take. So we invert the whole concept. So instead of letting the. We call them cats, the conversational assistive technology, instead of saying here's what you can't say, here's your guardrails, we say here's what you can talk about your knowledge base. M so asking a question about baseball, and that's not in the traumatic brain injury handbook, then we ask the admin if the chatbot can go to the web to get the answer about baseball. Oh, I have another term for you that I want to put out there. It's similar to kind of catfished by AI. There's a lot of executives that might be working for their AI assistants. And I think that you have talked about that with some other people and a lot of people are like, well, we know the fix, we'll put a human in the loop. And the idea of putting a human in the loop is liability laundering. I like it because you're, you're, you're pushing that liability to that poor human who now needs to be what, accountable for all of um, the AI output. How realistic is that?

Speaker A: Mhm. Yeah, it's human in the loop washing right. I think it's big. You're cleaning things up by saying there's, it's kind of like Homer Simpson pushing the button. Is the human in some of these cases where you know the, the scope of what they're responsible for and the level to which they're able to check the AI is questionable to begin with. And then how are they trained? By the way, I wanted to ask you because I need to look back into it. I've been stealing regularly. AI is great if you're already wise and I learned it from you. But didn't you, you attribute that to someone else?

Speaker B: Nabel, um, she is one of the world's best programmers on earth and she is often known as the queen of microservices. M which is a software architecture that I have totally co opted. And so a lot of people, you know, agentic is the new AI, you know, the new black. Right. You know, it's the sexiest thing in the world. And I'm like, uh, and if you actually go and you say, hey, I want to go write an agent, it's going to send you to Zapier. Do you know what Zapier does?

Speaker A: It's like API.

Speaker B: Well, it's automation. It's an automation agent there. And it has been around, you know, since robotic process automation. And anyway, I think that's hilarious. But what I catch, and I really want to talk to you about this because it's fascinating, I use Claude code, which is an agentic harness. So it's a harness that supposedly, you know, um, gives the agentic instructions that you can then use those agentic instructions to do things like write code. And one of my favorite examples is, and this wasn't me, this is from the cat butt thing. A guy's like, oh my gosh, you deleted the database. And the AI says, the AI agent says, oh no, here, let me fix that, let me restore it. And it generated 4,000 records.

Speaker A: Oh no.

Speaker B: Like you're, you're in, you're in deep psychosis when you're like in an agentic harness writing code.

Speaker A: Yeah.

Speaker B: In natural language. Because it takes your language and interprets it in a way that could be very wrong. My CPO describes it as, it's like a seven year old who has all of the power in the world. And you're like, go brush your teeth. And the seven year old goes, well, I put an addition on your house. Isn't the addition on your house better than brushing my teeth? Yeah, I didn't put an addition on my house. So. So like these agentic harnesses are running multiple agents underneath. Multiple. Like, so it'll be like here, uh, and I even have this in my substack for posterity. Claude code. The Claude code was like, well, it was the agent's fault. Oh my goodness. I was like, first you lied to me, then you tell me it's the agent's fault. And so back, back to Aaron Schnabel on, um, microservices. And microservices. A job to be. It's a discrete task. And so when you write agents as microservices or a job to be done, that is self healing, that it knows its inputs and outputs, it knows what kind of feedback that it needs in order to be able to. Like for instance, if you write an agent to go get the weather in New York City or in Denver, you're going to make New York City and Denver a, uh, variable so that it can just go get the weather. But if that weather station changed where it gets the weather from, then you need to have it understand where the new weather station is. So that would be like A self healing agent that you would write the code to say, uh, you know, use, you know, WGYN or use whatever. That's the kind of idea that you should be able to think through. And so in our work we use agentic orchestration that looks a lot like a reasoning trace or a reasoning strategy. So a doctor could say, hey, I need to give a really complex, you know, diagnosis to somebody and I am only going to get 15 minutes with them. I want to figure out some way to tell them that I went to this website, I got this information, and then I went to this knowledge base that I got this information and then I went to my friend and I got this information and then I put it all together and then this is why I waited the information for my friend better than the information from the website. That's a reasoning strategy. Right. And the doctor looks at me and he's like, can you do that for me? Into Vietnamese. And I'm like, yep, that's, that's what we do. Uh, because that's, that's the power of doing this and using the AI in the right way where you control all of the inputs for that discrete job to be done and all the outputs of that discrete job to be done orchestrated together as a reasoning strategy in order to be able to deliver the information in the way somebody needs to receive it.

Speaker A: If I'm hearing you right, you know, there's a few keys in there. Like one is the idea that, remember I got this also from you, there's labor involved. Like there is human intellectual. Like I think you were saying, understanding is labor. Right. So like there is this idea. Yeah. And it's, and it's not, it's not an output either. You know, where I feel like frequently people just want stuff to be fast and easy and they don't want to have to think, you know, and if I'm hearing you right, like we need folks to be able to tap into the engineering breakthroughs that are real, but then to apply some human, like higher level, higher order thinking to then do the orchestration. And then the orchestration includes human actors, uh, in the design as well.

Speaker B: And the reason is, is like, I mean, this is all the way back to Conrad Wolfram, who has a great TED Talk about mathematics and about why don't we let children use calculators? And why do we. Because if we, if we let children use calculators, we could ostensibly teach calculus to third graders. You know, and so that's, that's a whole different similar train on. I am um, yep. I would not call me a techno optimist. I would call me a techno realist that I'm like, I know what the technology does. And, and so that's where I'm like, let the kids use some calculators because I, I need some more kids with linear algebra placed, you know, and there's no hand waving in math. So what Conrad Wolfram says, and this is just perfect for what we should be doing with AI is you start in the real world with a problem, a question, then you go to the mathematical world and you perform a calculation and then you go back into the real world and see if that calculation works to solve your problem. Three steps, right?

Speaker A: Yeah. It's funny you mentioned linear algebra because I've had a few folks raise that to me. I also had uh, Ted Dintersmith on a little while back and he was talking about the future of math education and he was talking about the importance of statistics and estimation and decision support, all of which make sense. And those all seem more like kind of foundational. Linear algebra is actually kind of hard. And I'm hashtag team calculus. I am always saying that it's good for some uh, of us and it's great if you're part of the tribe who understand calculus and then linear algebra is kind of built on some of those same foundations. What do we need? Like what are the, what are the more technical skills that you see being kind of critical for future humans? Speaking partly as the parent of a seven year old child.

Speaker B: Yeah, I think diffieqs are much better than linear algebra personally because your, your job is to create a function, not an answer. And that, that to me is a, a uh, much better tool.

Speaker A: Diff Diffy QS for the less mathematically

Speaker B: informed are differential equations otherwise known as meditation. I'm sorry but, but like seriously it is, it's a process and going through the process is about, you know, learning, learning how to take things in the real world and then in the mathematical world how you would solve them and then going back into the real world to see if it works. And so it's just, can you apply the function? You know? Another one that I think is super beautiful is the diversity prediction algorithm. So um, every seven year old should have a teacher that has a jar of jelly beans on the table. And that jar of jelly beans should be something that at the end of the year they reveal how many jelly beans are in the jar through the guesses of all of the different people who walk into that classroom. And taking the average of all those guesses, which is the closest to the truth. And that diversity prediction theorem, um, to me is like, that's the most beautiful thing on earth. Like that, that tells you how to solve anything.

Speaker A: And is that like the wisdom of the crowds? Is that the same idea?

Speaker B: Well, but it's, it's like what we don't know is what is the right diversity that we need in order to solve the problem. And if you're looking at, you know, if you're looking at like the Bell curve. Right. What most people don't get is that diversity prediction theorem, um, is why, you know, for sure you need X amount of representation in order to be able to have that much in your data set. Which is why anybody doing any math needs to understand what their variance and standard deviation is of the data set that they're using. And so basic, get a two tailed distribution for your confidence interval. Like that's what that means.

Speaker A: Yeah. Which is, which is getting back to stats. Yeah, yeah. And like, and some of that, you know, again, I have an n of 1 in my 7 year old, but I am testing him, like just having conversations. And it does feel like there's some like nascent capabilities that are just left alone in early education. Not even early, like, you know, now he's in second grade. But like beyond the fact that he just needs to know his times tables and they should be teaching that sooner, I'm going to put that one aside. But like, separate from that. I know, I know you do more with a calculator. He can learn his calculator and also learn his times table. Our working memory is precious, Beth. But like the practicality of math, what you were talking about before, like the ability to mathematize something so that you can play with it, uh, explore it, figure some stuff out and then extract from that the genuine insights that are actionable. That's always the way. Uh, for whatever reason, I've always thought of math as that, maybe because it came to me, but I feel like everyone should be raised to kind of operate that way. And if you need to use a calculator, great, as long as you're able to produce these insights and outputs.

Speaker B: So in the world of AI, really reduced simplification here is AI is super good at pattern recognition because it looks for the mathematical patterns. And generative AI specifically in language models is like your next token, like your next token prediction. It's far more complicated than that. But just bear with me for a second. Yeah. What I have found as a human being is the Patterns that it chooses. It's the choosing that is intelligence. And, um, incredible book, hard book, but I'm going to give it to you anyway because it, it, the, um, the call to action is the most beautiful call to action I've ever heard in my life. It's called Carbon by Paul Haakon.

Speaker A: Okay.

Speaker B: And, and it is all about how carbon is the most incredible element, substance that we are all made of. Yeah, yeah. We could go to carbon 14 dating and like all. But I think it's fascinating that, that we, we have all of this understanding of science, but it's not really teaching us the things that we really need to know as human beings because it's not teaching us that to look around us and to see the choices that the tree is making or the choices that, you know that and, and once you start seeing it, you're like, holy shit. Intelligence is all around us. It is literally everywhere because it's in the ability to choose and to choose wisely, to choose based on a, a use, um, you know, fit form to function, use only what you need, you know, reciprocity. Choose things that, that grow the ecosystem, that, that enable sustainable growth, that enable, you know, good high quality healthy living organisms or complex adaptive systems. And what I find is an AI, it chooses badly. Like, it will never get the choice right in my estimation, because it's not intelligent. Like, you know, some of the work that we do to try to force the choice. So for everybody, you can do this yourself. Take a picture of your bookshelf or your record collection or your Netflix recommendations. Just take a picture and feed that to the AI system and ask the AI what one's your favorite. It will never get it right. Or it will get it right only if you tell it that this pattern matters. So AI, great pattern detector, you know, all of the, you know, all of the mathematics to be able to say this is a pattern because it matches this, blah, blah, blah, chooses the wrong pattern every single time. Because that choice is what I would define as intelligent is to make the choice and, and to have that good choice be something that you agree with, that you as a human being and your taste and your experience and your intelligent intuition all makes sense. That's why I'm like, it'll just, it, it doesn't choose right, at least for me.

Speaker A: Yeah, that's interesting. Yeah. Because when, when you mentioned carbon, I immediately went one, uh, row down to silicon. Is that right throughout a table?

Speaker B: That's right, yeah.

Speaker A: Yeah, it's, they're, they're right there. And they're, they are similar. And there is this notion that we've got an emergent intelligence here that would be different. And then, and then, and then at the same time, you know, the more hardcore old school AI folks talk about neurosymbolic models as what was a dead end. And it's almost like people got tired of trying to solve the problem that way. So then they shifted over to this transformer model that we're kind of over engineering at this point. What are your thoughts?

Speaker B: So I, um, I'm kind of a, I, I have a really simplified view of, of neurosymbolic systems. And um, I was recently talking about this with my friend because as much as I would like to have everybody jump on my bandwagon and redefine the word agent as a microservice and a job to be done and a discrete task, it's not going to happen because that horse is out of the barn. And back in the IBM days, I learned this lesson vicariously. They tried to rename AI to cognitive computing M and just, it didn't work. Right. Like, who's heard of cognitive computing?

Speaker A: Well, yeah, yeah.

Speaker B: By definition, anthropologists define my terms. Cognitive systems are a combination. It's a binding of philosophy, psychology and computer systems. So when you're in the realm of like neurosymbolic systems, you still have to have a referential understanding of how you know what you know. And for me that's an ontology. Like, and, and I do it crudely for your hip bone connects your thigh bone, your thighs.

Speaker A: Yeah.

Speaker B: Like, I do it for like physiology and things that, that are known. Right?

Speaker A: Yeah.

Speaker B: And so if you start with an ontology of how do you know your hip bone connects?

Speaker A: Right.

Speaker B: Because you can prove it in the real world is is the answer. So you, you have an ontology of like science of real world answers. And that ontology is also known as a graph model or a knowledge representation of the entities and their relationships with, with the entities. And then psychology is the one of like why do you do what you do? I think that this is the one that is the black art that people aren't willing to really delve into because it requires self analysis, self examination, dare we say actualization, you know, like uh, it requires you to know what are your biases, what are your subjectivity. And something I read recently that I think is really, really fascinating is, you know, originally hard science, mathematics, you could never have qualitative things because qualitative things are soft science and subjective and you know, uh, Fraught with like you know, all of those feeling things that humans.

Speaker A: Psychology major. I barely graduated because I wasn't, I wasn't too into the stats side of things.

Speaker B: Yeah, and I was too. But I happen to have the stats because of archeology, which is the only reason that I am who I am is the only reason they let me in the club is because I could do stats and.

Speaker A: Yeah, and I love stats. I just don't like the idea. I don't like being forced into. I almost prefer being able to be a little more theoretical about psychology even like in the kind of post Freudian kind of.

Speaker B: Now this is why cognitive science, right? Because you can do an end mind. This one subjective understanding of you and the AI doing your complex adaptive system. People are doing science again in this, in this subjective way. And uh, I mean you, you got it with psychology, I got it with, with anthropology. Before you study something you gotta know what you're doing. Like before I took an ethnography I'd be like journaling. I am hungover today. This is going to affect the way that I ask the questions, right? Yes, I'm, I'm sure I wasn't really hungover. I use that example a lot. But um, it was just, you know, it's like that mathematicians, they were never taught that. Statisticians, they were never taught that. They never knew that their intent impacted. Which is hilarious because of course it impacts like every data scientist that says, you know, this, this data doesn't really fit. This data doesn't really fit.

Speaker A: Right, right.

Speaker B: Let's, let's get to the objective measure of um, truth.

Speaker A: Right? Yeah. But so much more is about subjectivity. Right?

Speaker B: And so much more.

Speaker A: That's the thing that we're all kind of struggling with.

Speaker B: That's right.

Speaker A: At the same time, the role of the human as a subject like the, the agentic aspects of humanity are where I get concerned around the conversation around, you know, agentic AI is that frequently we're taking decision making authority away from humans, at least when misapplied. And that the cognitive offloading can be of something that's too high up the judgment sharing.

Speaker B: Are you familiar with the Cynefin framework or complex adaptive systems? Dave Snowden. God, it's dreamy. He's the only human on earth that has scaled ethnography. He taught children to ask the elders the question, to be able to understand what's going on in a locality. It's brilliant. And he, he does all kinds of work for all kinds of places. Three letter agencies and Stuff like that. Love him. And it's spelled C Y N E F I N M, it's a Welsh word. And he is a wonderful, lovely human being. And so he and um, Dr. Delia McCab, who is a neuroscientist, and I were in Vancouver in January doing a, doing a very, very interesting, cool workshop. And Delia just tells me, she goes, beth, people aren't lazy. It's your brain's job to conserve energy. And I'm like, oh. It's like. And then she, she blows my mind again. She's like, beth, why people don't understand something that their salary depends on them not understanding is that human beings, we only have the capability of holding two paradoxes in our head at any given time. So if one of those is not actually true, you have to actually dismantle that to get another one to come up. So let's say you believe that all AI needs huge amounts of data and huge amounts of compute. That's what you know in the world. And you're probably not paying a lot of attention because that's what everybody knows and that's just common knowledge. And that's why we have to build so many data centers and why we have to harvest so much data from children and why we have to do all of these things that, that we're doing. Because that's, that's what we have to do to get AI. Right.

Speaker A: Yeah.

Speaker B: Uh, so in order for something else to be true, like maybe you just need five or six narratives in order to train models, which is what Andrew Ng says, by the way. Um, you just need five or six good, high quality data about the domain.

Speaker A: Stanford professor, founder of Coursera. Right?

Speaker B: Yep, yep.

Speaker A: Just, just keeping, try to keep up.

Speaker B: Yeah, yeah, yeah, just getting there. And, and I'm like, for that to be true, then why are we building on the data centers? You know, why do we need all this data? And you know, why do we need all of this data to build machines that are probabilistic, which means that they are probably right some of the time. And, and I'm like this stupid. So anyway, like I, I love that, like, Delia, just like every time I have, like, I'm like, why won't people understand this? She's like, oh, let me tell you about brain chemistry. Uh, or let me tell you the neurophysiology.

Speaker A: Yeah, I really like that because I've heard, I've heard the. What's is it five plus or minus seven, plus or minus two. Like working memory.

Speaker B: Yes.

Speaker A: Can only hold so many.

Speaker B: That's right.

Speaker A: Sort of atomic chunks. But if I'm hearing you right, there's only like so many kind of open problems. Yeah, only two. So very. Which. Yeah, that does make sense. That almost explains to me some of the existential angst that many of us have been going through lately, where I feel like you're confront.

Speaker B: We need to say that, that, that word angst, um, in German, the German accent.

Speaker A: Like what happens when you have four or five of these things that are kind of constantly bombarding you? I feel like some of us just. I've noticed it myself too. I just kind of shut down and I start avoiding complicated things because I can only. Maybe it's a good practice, come to think now thanks to this rule of two. But it is like I can only. And then it's hard to switch out like that if it's still incomplete. This is the Zeigarnic effect, which I talk about all the time. Our brains want to work on open ended things until they're closed. And it almost like we have to be able to leave open things alone to focus on other ones.

Speaker B: Yeah, I read this incredible. Like uh, a doctor used a scribe for 90 days.

Speaker A: Uh, an AI scribe, not a temp or something.

Speaker B: Right, right, sorry, an AI scribe. And he's like, I can't close down anything. He's like, I'm so burnout. I'm so fried. It's like every patient has 15 like red herrings that I gotta go trace down because the AI chooses badly. And you know, when he goes back and he reads the notes, he's like, this is not the notes that I would take. So now he has to figure out why the. And it was just. It's causing so much more. And I see that in programmers that are using quadcode or Codex, you know, for. Yeah. Uh, oh my gosh, eight to 10 hours a day. 12. I mean it's just like, they're like, oh yeah, there's just not. It's not good for your brain. And um. Pj, right. Like he said that it's disequilibrium. Isn't that the word that you feel when your, your mental model can't absorb the new information? You're in a period of disequilibrium.

Speaker A: P probably. Yeah. Like when you're, you're. You don't have a schema for something. Right?

Speaker B: Yeah, yeah, yeah. So maybe that's. I don't know. I'll go Astelia.

Speaker A: I love the fact that you focus, not exclusively, but deeply on healthcare. And those ontologies and those use cases because, you know, the AI scribes are becoming much more standard now. The underlying problem is that the humans were, are overworked and they're, they're, they're plowing through so many conversations, and at least if they're not taking notes, they're. They're focusing on you.

Speaker B: It's, um, it's complex and it's like, uh, I wrote another article about this

Speaker A: recently, and this is all on substack. Beth, if we want to push people to your sub stack.

Speaker B: Okay, Absolutely. And, um, please buy me coffee. Please, please subscribe. I, I reached out to. I have four or five advisors in healthcare because I'm one of those people that I'm like, I definitely know what I don't know. And, you know, a lot of the reactions, um, that I got because I, I went to all those advisors and I was like, hey, should patients, like, try to say no scribe in. In the doctor's world? And because it's, it's in. I'm always like, I want to take like, a granola or I want to take my own too, because, like, I want to have my own understanding of something that I can interoperate with. And this is just me personally, and I think that the world that we're living in, it's very difficult to trust a lot of those scribes. So your answer is a lot of human and a little AI or, you know, AI and. And human, but depending if you know it and you use it and you are the one making those choices. Yeah. And. And one of my favorite things is a granola, because, like, when I take a granola, I have a recording. And then especially if I'm like, what is granola? Granola is. It's like a transcript tool that you can run. And I was just like, I was like, I didn't do that for this podcast, but we can take the transcript and do it later. But it allows you to also put notes into it so I can say, you know, um, like, I have. When I'm listening to clients, I was trained, when your client is talking or your customer is talking, you're writing, otherwise you're fired. Like, that was my, that was my training as. So I'm always trying to take notes because those notes influence, you know, how I want to remember that part of the conversation.

Speaker A: Yeah.

Speaker B: Again, I'm putting my choice into the AI and then using the AI to detect the pattern of the words that, that are being said and then, like, write them down, which is a really hard Impossible thing if you've ever tried to like actually take dictation. Yeah, uh, it's very difficult. That's m. That's my opinion. Uh, and it's just, it's a really complex topic because of like a patient and a doctor. The doctor needs help because of the system that they're working in. The patient needs help because of the system that they're working in. And so it's like, how do you create, you know, how do you create the right environment and safety and trust? And as we all know, it's really difficult to be vulnerable and safe and you know, say things. And there is patient, doctor, patient confidentiality. It's just, I feel like we just have made it hard on everyone and, and I want people to see AI differently so that we can make it easier on the human beings because this, this. I, I don't know, this race to create godlike images of. Of men. Why. Why do we need this?

Speaker A: Yeah, non augmented Sam Altman and Elon Musk is more than enough. We don't need godlike versions of them. You know, we're maxed out as we are.

Speaker B: You know, they are more than enough just as they are. Right.

Speaker A: Like, yes, but it does speak to. And you mentioned trust and I, uh, you know, would love to have you back on. We could clearly go for hours and we should explore formats that allow us to go longer. But you mentioned trust and maybe as we're wrapping up this conversation, I've been thinking a lot more about trust and you know, trust architecture and how m. Some platforms, you know, you mentioned, Claude, like they seem to be. As a big platform, they're kind of winning on the trust architecture side of things. It's debatable, but at least certainly it's not the most advanced competition. But you know, it does feel like there is this need for people to trust what they're working with. And there's a lot of backlash against AI now. It almost feels like you're. You're almost designed with human trust in mind. I'd love to get more of your thought on that.

Speaker B: I mean I'm a. Ah. Um, so I'm married to a maintenance test pilot and you know, I'm like, we have marriage maintenance. Everything is built for maintenance. And Stuart Brand has a beautiful part one of his maintenance era book out. And Stuart Brown, Clock of the Lawn now like, you know, really believes in thinking about things in a way that we can build them to maintain. And this goes way back. But I'm always like, how something is made is Indistinguishable from what it means. And so if I'm going to trust something, I want to know how it's made.

Speaker A: Yeah.

Speaker B: And. And that, I don't know, Media guy,

Speaker A: I immediately go to Marshall McLuhan, say that the medium is the message, like how it's made informs what it is.

Speaker B: That's right. And so I don't know how to teach people that, but I think we all inherently know. Right. And so how something is made is indistinguishable from what it means. So my system, how it's made is it's made to explain how it works.

Speaker A: Yeah.

Speaker B: And it's made to. We had this as like some of the output from our like A B testing and stuff. It's like all of the other systems are made to respond or made to engage you or the purpose is to generate.

Speaker A: Yeah.

Speaker B: And to extend tokens or expend tokens.

Speaker A: Yeah. To keep you coming back too.

Speaker B: Yes. To. To. I. I call that inappropriate engagement, by the way. M and so the purpose of what I create is for people to have an explanation of how it got the answer or how it was made. Because I think that that is the only way that I could call it trusted is by letting people inspect and decide for themselves on whether they want to trust it or not. And so I'm like, let me just show, uh, my work and whether people trust or not, that is a very personal, very personal thing. And I think if you were an anthropologist a hundred years from now and you were just viewing our behaviors, we would all be behaving in a way that we are trusting a system, but we don't actively trust it. And I think that that's part of the cognitive dissonance that we all feel all the time, is that we have to use Amazon because we desperately need that conditioner for the night. Because tonight is the only time that I can take a shower and wash my hair. I'm like, do we trust it? No. But we trust it. Huh. You know, so it's like we're in this, like another one of my articles. You will love this one. Teleological suspension of ethics.

Speaker A: Mhm.

Speaker B: We're in. We've had a cultural devastation. Our entire culture has been devastated. And what used to be taboo, having a developer check their own code, now we have AI write the code, write the test case, and check itself is now commonplace.

Speaker A: Mhm.

Speaker B: What used to be non scientific and of one is now commonplace. So we have a teleological suspension of ethics. So telos is your purpose so the purpose of an acorn is to become an oak tree. The purpose of humans. We have a suspension of what used to be ethical or what used to be, you know, commonplace. And this all comes from a beautiful book. It's, uh, by Jonathan Lear. And the book is just phenomenal about the very last pro warrior. And we have very little information from cultures that have been devastated, but this is one. And he said after the buffalo were gone, nothing happened because his culture was devastated, that he didn't have any of the culture to stick. What making dinner was because they made dinner before they went to hunt buffalo, but the buffalo were gone, so now they just made dinner. Isn't that interesting?

Speaker A: That is. Clearly, I need to have you, uh, on more, but as we're wrapping up, if you want some parting shots, some takeaways, we'll include a link to your substack and get folks to subscribe. They should. We've included several books you've written over the years, and we'll reference those as well. Closing thoughts, final shots this time around.

Speaker B: Be human, Be kind. We are humankind.

Speaker A: Nice. I dig it. Beth Rudden, CEO of, uh, Bast AI. Ah, check that out. Check out what they got going on. Beth's someone you should be tracking, if you aren't already. Beth, always a pleasure to have you on the podcast.

Speaker B: Thank you.

Speaker A: And for our listeners, subscribe, tell your friends, do all the good things. We'll be back again soon. This is trending in education.