This is Making Better Decisions. I’m your host, Ryan Sullivan. Decisions are where rubber meets the road for organizations. Each week, we’ll be learning from people who are on the front lines of turning raw data into better outcomes for their organizations. This show is sponsored by Canopy Analytic, helping companies make better decisions using data.
Ryan: Hey everybody. Welcome to another episode of the Making Better Decisions podcast. Today’s guest is an analytics and AI powerhouse, started his career in analytics and BI.
Ryan: Has since done a major upscale, getting a master’s in
Ryan: artificial intelligence to add to his undergrad math, math nerds for life, he’s now moved into software.
Ryan: Engineering, data science and AI, specifically natural
Ryan: language processing on database metadata. Please welcome senior software engineer of data
Ryan: platforms, at Prion, Christian Damare Bates.
Christian: Hey everyone. Excited to [00:01:00] be here.
Ryan: Yeah. Thanks so much for coming on. I am super excited. Um, Obviously today we’re going to get into some fun stuff that I don’t always get the opportunity to talk about. Um, and I think I’m, I and the listeners will, will learn a lot from you. So without further ado,
Ryan: what’s one thing you wish more people knew about using data to make better decisions?
Christian: yeah, the, this, this question got me thinking a lot because. The fact of the matter is requirements gathering is, I think, one of the biggest things that, and biggest blockers to delivering good products, um, is really like how to communicate, um, effectively so that you can track the work that you want to do, um, and get what you wanted out of the work, right?
Christian: Um,
Christian: we spend all this time investing in, like, clean, cool
Christian: sprint boards
Christian: on JIRA. Um, And [00:02:00] acceptance criteria and what even needs to be accomplished is a really difficult problem to solve because you can make so many choices and there’s like option paralysis that comes with making choices. Um, and not only just the choices on how your app
Christian: can go, but also how to communicate
Christian: between business minds and technical minds.
Christian: Um, is a
Christian: real, I think, telephone bottleneck.
Ryan: Yeah, for sure. I mean, I think I assume this to be the case just about everywhere, but obviously my experience being mostly
Ryan: with, with data and
Ryan: building reports or databases or whatever
Ryan: the case may be, you know, you really don’t. ever get a fully clear picture because of what you’ve talked about. I mean, there’s the other problem of even if I assume I could capture perfect requirements, I’m kind of making the assumption that they’re going to stay the same long enough [00:03:00] for me to be able to complete a solution to them. Um, you know, but then there’s, you know, obviously business is hard. Otherwise everybody would just be rich and not working anymore. I don’t even know what would happen, but you know, like, The alternative to that, which I’ve seen a lot of people lean into is just like, we’re agile, man. Like we just, you know, we’re agile.
Ryan: And I’m kind of like, okay, so like you don’t plan and everything becomes the highest priority as soon as it gets put up. Right. And they’re like, yeah,
Ryan: so I think there’s like
Christian: One of the most stressful
Christian: environments to be in,
Ryan: Yeah. Hmm.
Christian: as an engineer, and especially like, I think of myself as a
Christian: creative
Christian: mind. Um, I sell
Christian: myself as a creative mind, right? Um, where
Christian: I think like software engineering in general is almost a
Christian: pure creation. It’s a creation
Christian: from nothing. You’re writing logic to create a vision of a world.
Christian: Um, [00:04:00] and like, you have to think
Christian: at different levels of abstraction to really
Christian: represent problems, but like you’re taking this picture of the world
Christian: and writing it in code.
Ryan: Yeah. Yeah.
Christian: as an engineer, um, or as a
Christian: data set or what have
Christian: you to try to boil down.
Christian: I mean, the universe is a hard problem, like we couldn’t compute it, but like, what can we compute? What’s the level at which we need to compute to be competitive, to address needs, to, um, maybe you want to explore a little bit and it’s not necessarily a need, but if we unlock some secret, Um, that the data is going to, tell us, um, maybe we have competitive advantage.
Ryan: Yeah. One of the um, the things that’s kind of emerged as a pattern
Ryan: after getting the opportunity to talk to so
Ryan: many smart, experienced people on the podcast is.
Ryan: They’re the reason why we can point at every one of [00:05:00] these systems or, you know, like frameworks, whatever you want to call them and point out ways in which they don’t work is because almost any approach has to be really bespoke to the people that you have in your organization.
Ryan: You know, if you have people who are building things that also are the types of people that like understand the business really well, well, then you can just go to them with a business problem and be like, Solve this problem, please. Here’s your budget. You know, on the other hand, you may have an organizational structure where the skills lie in different areas.
Ryan: So one of the things that you and I talked about a little bit was what is generalizable about that? So like, what are some of the concepts as we think about like building a new thing, whether it’s data AI, anything like that, that is you know, Applicable to everybody. What’s the one takeaway that everybody can
Ryan: get out of this?
Christian: Um, [00:06:00] I was thinking about this? earlier today actually, um, especially the, this is like AI Masters 101, um, Information Theory 101, um, is
Christian: the concept of entropy. Um,
Christian: I don’t know.
Ryan: we’re going crazy. I love it. Dig in. Go for it. Go
Ryan: for
Christian: I love entropy, but the fact, Wikipedia has a really pretty chart on it. Um, insert chart here on YouTube. Um, where it’s the, um, it’s a chart of the probability of an event versus the entropy of the event.
Christian: Um,
Christian: And basic, basic level what entropy communicates is how much information there is in a decision that you’re about to make. Um, or rather
information that you would gain from making a decision under the circumstances, like the probabilistic circumstances. And almost the, the maximum way to gain the most information is a 50 50 coin
Christian: [00:07:00] toss. Um, is you are going to deal with true random. You’re going to flip a coin and say, okay,
Christian: heads or tails, I trust fate. Um, because any. If you already are at either a zero probability of something or a one probability of
Christian: something, you’re not gaining any information there by making a decision. Um, and so I think that’s one rule that I wish
Christian: people could, could come out of this decision making
Christian: process with, is if you want to gain the most information about your field, and this is probably something you want to do early stages in startup, you want
Christian: to try to effectively
Christian: behave randomly to start.
Christian: Um, because any target you shoot at is going to gain you more information.
Ryan: So I knew talking to you that we were going to get
Ryan: way nerdier than any other podcast. And I, I’m, I’m, I
Ryan: think I’m just going to lean into it. So I have, I have a fun story
Ryan: that I think you’ll get a kick out of based on one of the research articles that I think was most impactful. So like, you know, the big takeaway that I got out of what you just said is you know, [00:08:00] obviously from a math perspective, right?
Ryan: I know what you mean when you say random, but I don’t think, you know, you’re
Ryan: like, Hey, we’re a software SAS company. We’re doing hot dogs today. It’s like, okay. Okay. you know, like, you know, so like I, I get what you’re
saying, but it’s like. The takeaway I think is to try a lot of things and gather information quickly.
Ryan: Right? Like the Silicon Valley paradigm of fail fast, right? The math
Ryan: actually backs that up.
Christian: Well, and that backs up Agile, which we were just talking about, right? Like, the agility of being able to quickly pivot on a user’s requirements. This is hugely beneficial early,
Christian: um, but also this is, this is in a world where like the coin flip 50 50 shot is basically we have no, no idea what’s going to work and what’s not.
Christian: But as we gain more information, we’re going to lean either further towards the, this is like a 0 percent chance of working or this has a 100 percent chance of working because we know our users want this.
Ryan: yeah. [00:09:00] So there was a, there was a research article
Ryan: that I read. This was probably
Ryan: like 10 or more years ago now, it was in this, um, field of
Ryan: kind of computer science as surrogate for larger research. Um, and it’s called multi armed bandits, right. So you’re probably familiar with this. Yeah. I see, I see you getting excited already. So they basically, you know, When you go into a casino and you look at the slot machine, some people will, you know, rather smugly call them one armed bandits because you pull the little thing and it takes your money most of the
Ryan: time. So, what was essentially engineered was, okay, well, instead of just pulling one, what if I put a large number of these, you know, slot, digital slot machine, You know, analogs in a row. and then I actually make different payouts for each one of them. So like for one of the ones in
Ryan: the string, right? You, every time you pull the lever, you get [00:10:00] a million dollars, right? It’s like a really broken slot machine, just forks over the cash. Right. and
Ryan: then in another one, every time you pull the lever, you lose a million dollars.
Ryan: And then like everything in between, and it’s like all randomly distributed. And then they made this competition where they said, all right, programming nerds. We want to have a competition where you write a program to try and win overall. so you get 10, 000 opportunities to go in and like, make some play, like some levers, you pull some levers, you don’t pull, sometimes you watch other people, all of that stuff the programs that came out on top were extremely focused on observing other people.
Ryan: And even though they couldn’t get any payout in the turns that they spent observing the other programs that were running, like you just mentioned, gathering information is the most valuable thing that you can do.
Christian: Yeah, I mean, and that’s, um,
Christian: you just [00:11:00] brought reinforcement learning into this conversation.
Ryan: Let’s send it, man. Let’s go for it.
Christian: yeah, so I, like some
Christian: vocabulary, uh, I might avoid that, um, but
Christian: when you start that, like all of those slot machines and you think of them, they’re all equal. They’re all starting at this, okay, like if I pull the arm, I’ve got like a 50 percent shot that I’m going to win and a 50 percent shot I’m going to lose because I don’t know.
Ryan: Yeah,
Christian: Um, and so this, is kind of going back to that entropy conversation where like if I pull the arm, on this one slot machine, I lost.
Christian: Okay. Like my experience tells me that I’m going to lose on this machine. I don’t know if I’m going to win on the next one, but my experience tells me I’m going to lose on that machine. So maybe I’ll try the next machine and the next machine, until I win. And eventually each of those slot machines comes out with this probability of winning. Um, and that’s kind of the goal of reinforcement learning is to learn that behavior [00:12:00] out at infinity, right? Like I imagine I could pull those slot machine arms an infinite number of times. I can’t in the real world, but like, that’s what the computer is trying to get at is how many times do I need to pull these arms, each of them, to figure out which one’s the best. Um, because then I can just, I know which one’s the best.
Christian: I can just hack that later. Like I’m going to hack the victory. Um, and so that’s the reinforcement learning term for. The, like, list of probabilities for all of those arms and, like, mapping it to the arms is called a policy function. Um, you might, you might hear about optimal policy. um,
Christian: if you ever hear about reinforcement learning, and there’s a ton of reinforcement learning in the language modeling
Christian: space right now with the
Christian: GPTs. Um, reinforcement learning
Christian: from human feedback
Christian: is the algorithm that
Christian: was
Christian: effectively used
Christian: to train the CHAP approaches. Um, which is
Christian: I, when I pull a lever, I’m gonna get some reward if I won.
Christian: [00:13:00] Positive reward
Christian: if I won, or negative reward, pain, whatever you want to call it, if I lose.
Ryan: Well, now I feel bad. Now I, feel like I’m
Ryan: causing chat GPT pain when I give it the thumbs
Ryan: down button for a bad
Ryan: answer. Yeah.
Christian: don’t know. There’s the
Christian: way that
Christian: the way that I think about it, too, in that modeling
Christian: space
Christian: is
Christian: really
Christian: like
Christian: it’s a distress. It’s a it’s a signal. And I think that’s one of the most
Christian: awesome things about the AI space in general, is it really is software engineering for the human mind.
Christian: Um, because we’re trying to go to AGI, we’re trying to go to general intelligence, um, we’re trying to come up with something that can truly behave human in a lot of ways, and a lot of ways we’re trying to like transcend humanity, which is a little scary. Um, but a true human machine or thinker would be maximizing joy and [00:14:00] minimizing pain. Um, that’s reinforcement learning.
Ryan: Yeah. Yeah. So
Ryan: I know that I’m going to want to swing back and
Ryan: talk a little bit more about not just the, the cool things that are out there.
Ryan: Um, With large language models and you know,
Ryan: what I think pop culture talks about as AI, but also, you know, Hey, there are all these other cool things, right? Like random forest and, you know, like other machine learning algorithms that are like really great for data.
Ryan: But one thing that I, I want to make sure that I touch on before we get too deep is, You know, when we first met, you were doing business intelligence, data analytics, same stuff as I’m still doing. So what is the going from kind of you know, the most complicated math typically that I do on a regular basis, you know, is like, division, right?
Ryan: It’s mostly about like aggregating, making sure that, I have the right sets, making sure that well, [00:15:00] did you subtract out this product line? It’s very much like simple arithmetic, but complicated strings of it to map to business
logic. Whereas some of the stuff that you get to do You know, is now like kind of going back to our math roots, right.
Ryan: All sorts of you know, linear algebra and, you know, you know, multivariable calculus and, you know, vector spaces and all this? fun stuff. You know, what is that, what has that transition been like? What’s that grow up story been like? And, you know, tell me what you’ve enjoyed the most about it.
Christian: humble brag over here. I had like taken linear algebra in high school.
Ryan: Oh,
Christian: like returning to linear algebra was a little bit of a challenge, but not, not too bad. I mean, there’s a lot of great content on YouTube, right?
Christian: Um, I love 3Blue1Brown, shout out to him and his work
Christian: on,
Ryan: how I learned how to code my own neural
Ryan: nets was from his videos.
Christian: yeah, and his Linear Algebra series is incredible
Christian: for people trying to understand what’s happening in the math of it. [00:16:00] Um, because it.
Christian: really is like spaces shifting
Christian: depending on how you multiply things. one of the things that’s kind of eye opening about the,
Christian: especially the AI textbook world, um, is how much linear algebra starts as for loops. it’s really like, if you think about a vector product, it’s a for loop over a range, applying two vectors, multiply it together. um, I don’t want to, I mean, that’s in
Ryan: I I get you. I
Christian: it’s for loops.
Um, and like, whereas in the database world with like, you can kind of think of those things as joins even. Yeah. 100%.
Christian: and so I, yeah, and then I think one of the other things that, like, the upskill that really takes some getting used to is the thinking of everything as a function. Um, think of every AI model that’s out there [00:17:00] as some black box that, has some math behind it. Um, but, like, early, like, algebra will say, okay, f of x equals x squared plus, I don’t know, 4x plus 4. And there’s
Christian: some factorization, like, there’s all this
Christian: simple algebra
Christian: that you probably learned forever ago that
Christian: maybe you don’t use.
Christian: But like, we don’t really care what that actual function is. That’s what the neural net’s learning.
Christian: We care about what is that f of x that we’re trying to model.
Ryan: Yeah.
Christian: Um, and even more importantly, especially in deep learning, there’s a theorem out there that basically says it can learn anything. It’s basically the Universal Approximation theorem or something like that I can’t remember fully what it
Christian: means, what it is, but it basically
Christian: means that neural networks,
Ryan: I don’t know. Yeah, I
Ryan: guess.
Christian: but, but that, and that’s why deep learning algorithms are so hungry is, because they can learn exactly how to spit out the values that. you put it in the training set.
Christian: It’s why they are prone to memorize if you don’t [00:18:00] train them right. Um, it’s because they will learn. exactly what you tell them to learn. Um, and so there’s a trade off of generalization there versus, do you want things memorized?
Ryan: yeah. So, I want to frame up a little bit. You know, the next chunk of the conversation. So, obviously, Neural nets have been around for a long time. Most people don’t really know what a neural net is. It is essentially just computer programming that uses the human brain as kind of a structure that you then try to mimic on a computer.
Ryan: Cause we, you know, we looked around and we said, well, hey, that, that’s That thing that we all have in our heads is pretty good at thinking. Let’s try to make a computer version of that. And they’ve gotten increasingly better and better and better, and they can do increasingly cooler and cooler tasks. And we have different ways of implementing them. but I think this came to the forefront. of pop culture [00:19:00] with the release of chat GPT, you know, it just like took the
Ryan: world by storm, everybody’s using it. It is almost synonymous with artificial intelligence and kind of, You, know, non technical conversations, you
Ryan: know, you’ll have an AI conversation. the
Ryan: whole thing’s just about like, how do you do prompt engineering for chat GPT?
Ryan: And. So we’re kind of at a place where this like unbelievably cool, but also very, very deep And complicated technology is being talked about in a public space. And so there’s lots of people out there talking about lots of stuff. And you know, transparently, I think that most of them. Either don’t know what they’re talking about, or they do know what they’re talking about, but they don’t know how to talk about it. So one of the things that I wanted to do is as someone who is just familiar enough to be dangerous and mess a bunch of stuff,
Ryan: up [00:20:00] from someone that actually went out And got, you know, a graduate degree in this stuff, start off by telling me, you know, a little bit about like, what is this technology? Not so much from a technical standpoint, but kind of like, how should I think about a large language
Ryan: model?
Christian: I think one of the things that’s so challenging to describe this problem is it’s so easy to describe the functions of the human brain and how it maps to deep learning, right? Um, we think of a neural network as a bunch of neurons. Um,
Christian: and the way the neurons in the brain work is they connect.
Christian: Uh, they have some amount of or absence of electricity in them,
Christian: and they send it along, and they send it over their connections, and eventually I say
Christian: a
Christian: word, um,
Christian: and the machinery that does that is so large,
Christian: and the way that the brain works, I, [00:21:00] ha, insert number of neurons here. A
Ryan: number of neurons here.
Christian: Large number of neurons
Christian: but we can, when we look at the GPT models, we’re looking at billions of parameters.
Christian: And whenever they say like four billion parameter, they mean four billion neurons and synapses all together. They might lump that number together, or they might just describe it as neurons. But the parameters are the things that are changing as it learns.
Christian: And that happens in your brain too, um, as you get neural pathways that form and that changes as you experience the world.
Christian: the deep learning
Christian: to the, the brain
Christian: process makes a ton of
Christian: sense,
Christian: but when you start talking
Christian: about the GPT series, um, you kind of have to start all the
Christian: way back
Christian: at recurrent neural networks.
Ryan: Mm hmm.
Christian: Um, and what they, what they
Christian: were
Christian: trying to do,
Christian: um, and basically, one of the things that’s really difficult about a [00:22:00] basic vanilla, uh, multi layer perceptron, or you can just, it’s a linear model, basically, um, simplest neural net you could build, think about that, that’s not like a, just, a neuron, like the simplest neural net you can build that’s not just a single neuron, really can’t take into account time.
Christian: which is, which is wild. and the reason it can’t take into account time is because when you think of a neural net and its inputs its inputs represent just like a frozen moment. It represents a single entity. It’s features about a single entity. and you can start to add time features to the dataset.
Christian:
Ryan: Yeah. I think one of the things that is most interesting about that concept. So like when I type something into.
Ryan: you know, the chat window and chat GPT is web portal, you know, it could be anything. from, you know, give me some good trip ideas for Austin, [00:23:00] Texas, or it could be something like, you know. here’s eight pages of prompting information and background information context.
Ryan: And then I like ask you to do some sort of formatted task, like whatever the case might be very, very simple to very, very complicated. It always functions the same way, So, When it it like does a
Ryan: little bit.
Ryan: of thinking and then you’ll see it. It kind of like just starts like vomiting text chunk by chunk out onto my screen and somebody that doesn’t really understand like the technology of what’s actually going on there may not Know that it’s, it’s actually just like constantly doing that process that You talked about of like generating this next word or chunk of words token, right?
Ryan: Just like constantly generating this information as it goes. And so one of the things that I think is, really interesting about this topic is, okay, cool. So, [00:24:00] obviously, You have like the, the group of business people who are just like, cool, you know, tell me how to make money off of this. And then you have other people, you know, who are some of the types who are, who are building some of this cool stuff.
Ryan: And they’re like, we’re building Skynet,
Ryan: you know,
Ryan: but like the, the thing that is most interesting to me is how can we learn about our own thinking from seeing Like what’s happened as we built a mockery of our own brain, you know, like when I think of sentences, I tend to think of them as the unit, but you’re absolutely right.
Ryan: Like information comes as we’re working through a sentence.
Ryan:
Ryan: I want to I want to change up gears a little bit and talk about really your, your kind of new specialty.
Ryan: One of the things that I have found as a, you know, a hurdle working with a lot of [00:25:00] businesses is that they have, you know, lots of information in lots of different places. The things that Logically connect to one another may sometimes be difficult to find.
Ryan: You may have naming issues, data typing issues. Once you have it all together, you end up executing, you know, slightly different variations of very similar logic over and over again. A lot of these pain points. You know, that obviously I think of in my head is like fairly technical. Like most people feel as just like, you know, these are business problems and like, they’re hard to solve.
Ryan: And like the technology just can’t really do it. One of the things that’s actually becoming really, really exciting. is the idea of having, you know, better information architecture. So as you have started to focus [00:26:00] on database metadata and then on doing natural language processing on that metadata, you’re kind of working in that space to solve it.
Ryan: So for someone who’s not familiar with. You know, I think most people know what a database is, but explain a little bit. What is database metadata? What is natural language processing and how are you using NLP on the database? Metadata?
Ryan: Yeah.
Christian: full circle back to the beginning of our conversation. Um, but database metadata if you a quick Google of information schema if you’re new to SQL and you want to Query information schema go look it up Um, at the high level business level, you, you, can think of, uh, an Excel worksheet containing data about customers as some concrete data about your customers.
Christian: maybe you’ll build charts on it and all that. but each row, maybe it represents an accounting transaction, maybe it represents a
Christian: [00:27:00] customer.
Christian: in database metadata tables, Each row represents a table name, the size of the table, the, when it was created, when it was last modified, all these attributes about the table as if it were the most important central concept.
Christian: And then there’s a table for columns, there’s a table for, uh, foreign keys, or like the, the joins that you are able to do in the database. what uniqueness needs to happen in different tables, like maybe a row needs these three specific fields to always be unique for each row. but, so that’s what database metadata is, is it’s information about the concept, the contents.
Christian: at the level of a table or a
Christian: column
Christian: or a join.
Christian: and then,
Christian: the thing that’s really interesting about
Christian: databases and like, you can track it over
Christian: time, uh,
Christian: with software engineering
Christian: practices
Christian: is, especially in recent [00:28:00] years where
Christian: compute has become less
Christian: and less of an
Christian: issue, like
Christian: memory and CPU has become less and less of an issue, Um,
Christian: column names have
Christian: gone from,
Christian: super
Christian: short,
Christian: like AID. What’s an AID.
Christian: Um,
Christian: to maybe we’re in a books like Amazon
Christian: world, and A actually stands for author. and so you’ve got this
Christian: concept of author ID. Well, author is a natural
Christian: language word that we would use to describe the entity
Christian: represented by that, that
Christian: ID. Um, And so as we get
Christian: to a world where we’re able to be more and
Christian: more verbose,
Christian: more specific about, the language we use to describe variables and programs or fields and tables or, we gain the ability to communicate the meaning so much more clearly to ourselves, our peers. And even machines. And as large language models start to learn, [00:29:00] Oh, I have an actor ID over here, Or an author ID over here, and an author ID over here, Maybe that’s joinable. They’re the same thing, and that’s like a really clear, equivalence, but, One thing that we might be able to do is say, AID exists, and author ID exists. Well maybe we’re able to do some character level encoding with some language model that takes author ID down to AID. And all of a sudden we can infer that relationship and we gain information about how that would join to older data sets.
Christian: and so there’s a lot of opportunity to use language language modeling, um, specifically even on columns and, and inference processes to take old data sets and make them more relevant to newer ones. but that’s just scratching the surface on a relatively
Christian: simple information retrieval problem, a like scoring problem, like how similar is AID to [00:30:00] AuthorID?
Christian: Close enough, I guess. Um, but what becomes really interesting, and this goes back to the requirements gathering process, and defining good requirements and good, acceptance criteria, is When people talk about the business problem they’re trying to solve, they write documentation about how the system needs to work. they send emails, they write Word docs, they write JIRA tickets. All of these things are language. the way that my mind works when I go to model a database is I look at it and I say, cool, what are The nouns and what are the verbs? The nouns are author. Okay, cool. I probably need a table for an author. because I want to capture all possible authors that we’re we’ve, we’re tracking for maybe our book sales. Oh, books just popped up. So I probably have a table for a bunch of books. And authors write books. Write is the verb. So maybe we have a table in the middle that says author wrote book. And it links their IDs.[00:31:00]
Christian: And so you start forming this language
Christian: that’s really
Christian: explicit about the problem we’re modeling. And those are things that you can start to detect in documentation, in tickets.
Christian: You can start to summarize down to things. And once you start to do that you can really automate the
Christian: database modeling process.
Ryan: Yeah. That, that was a, like a, a perfect, like such an easy to follow journey from, um, Hey, we need to build some sort of report to now it’s
Ryan: like, Hey, well, like the documents that you already have And
Ryan: the stuff that already exists in your database
Ryan: how can we use some of that to make that process.
Ryan: much, much easier?
Ryan: So obviously I think one of the followup questions to that is let’s imagine an organization that. You know, magically doesn’t have like a ton of, you know, data infrastructure. They don’t have a ton of, uh, you know, any of the stuff that we’ve talked about, whether it’s like BI and [00:32:00] reporting or databases or any of that stuff, right.
Ryan: Like outside of their operational stuff, like they’re just kind of, you know, maybe they’re a business where it was like, Hey, we, you know, the old CEO just liked everything on paper. Right. And, you know, new CEO comes in. What’s kind of your recommendation for what’s some of the best stuff to focus on? Is it kind of like, Hey, get some basic reporting up and gradually improve, improve, improve, or is it just like, we’re going straight to the moon?
Ryan: Like, what’s your recommendation for what people should start thinking about and implementing first?
Christian: I think it goes back to the entropy question, right? Um, so going back
Christian: to
Christian: entropy, there’s no wrong
Christian: choice
Christian: first on
Christian: your
Christian: data model because you don’t know what information is going to be most relevant
Christian: to you until you start reporting on it. and a lot of ways that people will start is they’ll start with this one big table mentality, um, and shout out to Google BigTable, BigQuery. But like, if you start out with this one big [00:33:00] table, where you are like, Okay, I have a bunch of books, so I’m gonna track the book title, the book author, and the book price, maybe the book publish date. And that’s all you really care about right now. as data starts growing horizontally, uh, i.
Christian: e. you start adding columns, maybe you start caring about the author’s birthday because you’re into astrology. Um, i.
Christian: don’t know, um, but like, but like, you start adding these extra properties about the data horizontally, well eventually you’re gonna hit this critical mass where the table is so wide I can’t interpret it. um, or I can’t maintain it. Um, and that’s where you start thinking about like normalized forms and data engineering. really that means just how do I make sure I
Christian: don’t have too many duplicates so that it’s unwieldy for me to maintain manually. and those duplicates being like the author wrote 10 books and I changed the author’s name on one book but not the others, like I screwed up.
Christian: Now I lost reporting at the level of the [00:34:00] author there. so maybe that’s when I start to split author information out because I know that I have too much information about an author to model in that one big table approach. it’s really difficult for me to maintain and what would be easier for me to maintain is just a link to an ID.
Ryan: this is, for what it’s worth, probably one of the most clear
Ryan: statings.
Ryan: Of what I think the right
r to this
Ryan: question is, which is
Ryan: like, sometimes agile just means like whimsical and chaotic. And I don’t really think that’s actually like a really good idea. Cause then You end up just like, it’s bad for everybody.
Ryan: But the idea of. We’re going to just try some stuff. And as soon as we get information about what works or what doesn’t work, we’re going to just change the weighting. So like, if we find out that something works, we’ll invest more in that thing until returns diminish. If we find out that something is bad or doesn’t work, we’ll pull back on investment or even stop doing it.
Ryan: And I think that you [00:35:00] really. hit the nail on the head. Like, how do you know if, you know a
Ryan: given report, data set, whatever the case may be, is, meaningfully going to provide you with good reporting or be the back end of a model? Like, do any of these things until you have it to give it a try, you know? And there’s this kind of like fairly iterative, like, let’s start Small, fast, and light.
Ryan: Get some feedback about What worked. Learn from that feedback and then take intelligent next steps and then just continue that process. and you know, especially, you know, here I am talking to this, you know, AI expert and they’re saying, Hey, let’s, let’s start simple. I think that There’s also like kind of a flip side to that coin though, right?
Ryan: So like, there are all people like you
Ryan: who are out there who love thinking about AI, who just by preference want to be out in those complicated spaces. [00:36:00] All of the companies that you, you start, you work for, you’re a part of, you know, whatever the case may be for an individual AI person, you are building something that is ideally a good or a service.
Ryan: To another person. So I think that when I think about the balance between, you know, where do I start, what do I focus on when I’m focusing internally, it’s like, well, do I have AI experts on my payroll? Well, I, I don’t, so I’m not going to focus on AI, but I’m going to make sure that I’m constantly learning about what the smart AI people who are out in the marketplace have built.
Ryan: And then if I find a place where I can pull that in as a service and use it, that’s a great idea. So. That’s kind of my take. I’m curious, as somebody who’s out building cool AI stuff, like, how do you view that blend between build and use, simple or like really cool cutting edge?
Christian: well, to be clear, I’m super excited to [00:37:00] be building all this cool AI stuff, um, but I’m not on the research side, I’m very much on the, um, the platforms engineering, which I support the researchers to do the really cool, like, learn how we can push the limits here. Um, but my focus is on how do we turn those things into services?
Christian: How do we make it as easy as possible to turn those things into services quickly? So we can pivot. So we can serve a new model while serving the old one. Um, so we can monitor that these models are doing what they say they’re, they’re doing. Um, and that they’re doing them well. Um, and that the data that we have about them isn’t. Like drifting off in a direction that doesn’t match what we assumed it to be doing.
Christian: I think monitoring is key. especially monitoring, you have to have a focus on humanity.
Christian: Um, I think that’s kind of the, what you’re saying there. but, and I’m, I’m very grateful that my master’s program emphasized [00:38:00] a human computer interaction course with every AI masters.
Christian: because You learn a lot about what kind of information you need to gain from people
Christian: to keep your AI product functioning effectively, but also how to hide AI in places that
Christian: create real user value. And so right, like chat GPT isn’t really hidden, which I think is one of the things that’s so incredible about it is like, you know, you’re chatting with an AI.
Christian: but just the interface and the humanity of like speaking to something is, I think what brings people back, the ability to have a conversation
Christian: instead of. Like this one sided, I’ll call Google a, one sided experience
Christian: because it kind of is.
Christian: Where you type your question into
Christian: a search bar and it spews out
Christian: a hundred search results.
Christian: I mean millions of search results in a quick amount of time, but you’re gonna look at the top ten.
Christian: It feels so less personal. And I think that’s what really [00:39:00] resonates with people in interacting with
Christian: these kinds of
Christian: products. And so a lot of the question is how can you make
Christian: the data gathering
Christian: process As personal as possible, as personalizable as possible.
Christian: so that you can,
Christian: make sure that your, your product is focused on the human interaction.
Ryan: so that kind of extends
Ryan: back into the question that I was talking about, you know, like, okay, so there’s this balance
Ryan: between. What skills do we have in house? We lean into those skill sets. What skills do we not have in house? Those we outsource, you know, either to, you know, a contractor, you know, good or service provider, something along those lines.
Ryan: So thinking about, okay, you’re a business person. You have this strong feeling That, everybody else is doing AI. So you have to say that you’re doing something AI related as well, or you feel like you’re going to get passed. And, you know, while I, in general, kind of like agree with that sentiment, right? like like, Hey, be a lifelong learner, figure out [00:40:00] like what cool stuff is out there and always be trying to
Ryan: improve and implement, like, yeah, sure. Of course. But you know, what I, I heard from, you is really, really about that. interface between the realities of your organization. I, I, I love talking about companies and other organizations.
Ryan: Like there’s a reason we call them that. It’s an organization of a group of humans and systems and computers and programs and all of that stuff. it’s all of this, you know, network stuff together that
Ryan: achieves some goal, whether that’s, that’s, you know, profit and a for profit business or some sort of mission goal and a nonprofit.
Ryan: And, you know, You know, what I heard from you, which was really, really cool is it’s like, focus on where do you have gaps as an organization and what tools are out there that.
Ryan: can specifically fill those gaps. And You talked about making, like if you [00:41:00] have a clear picture of, what it is, like, what is the gap and what are all of the
Ryan: things that surround that gap, then and only then can you really get.
Ryan: A great fit with a tool that will fit that that then is Like kind of invisible and really functions very well. Otherwise, it’s just kind of, like, you know, you’re another pop up window. And, a website that’s linking me to an LLM that’s just going to, you know, keep asking me how my day Is or like something else bad like that.
Ryan: You know, like, what can I help you with? and it’s like, I know you’re an AI that’s trying to sell me something. I’m all good. You know? So for somebody that’s not an AI expert, how can they think about. That specificity, how can they think about finding the right tools for the right jobs that are actually going to be tied to improved business outcomes for them?
Ryan: As opposed to just like, Hey, I plugged chat GPT into my website’s chat bot.
Ryan: it is less about understanding what [00:42:00] tool to use and more about what is my core value proposition
Christian: in the sense of if my core value proposition is to, think about a customer service org for a second. If my core value proposition
Christian: is to serve customers as effectively and as quickly as possible with the correct answer to, their question, then you’re almost always going to choose a language model.
Christian: Right? Like, the fact of the matter is, the, the interface that the human is, is, is connecting to you on is human language. whereas on the flip side, if you’re thinking about, image content violations as my core product, I’m going to choose some image model. and potentially, I might choose some multimodal model, where, It does some image to text, and then now that I have both, I can do some extra learning on both mediums of the language and the, and the image, looking at sensitive content, as an [00:43:00] example, and maybe content moderation as a piece of your software, say, meta is, is super invested in making sure that graphic images don’t make it to children on their app. you’re almost always going to choose that multimodal, or pure image, or some combination of expert models that do all of it. and so I don’t want to necessarily pigeonhole like, Oh, the next new tool and how do you evaluate it? I think in a lot of ways we have come up with some really interesting approaches to modeling these, these integral parts of our interaction with the
Christian: internet. and There are places to plug them everywhere. It’s a matter of what, what is the core mission? What is the core requirement of the system you’re trying to, you’re trying to use? because there are, there are thousands of language models out there. They’re trained on specific, knowledge domains like medicine or law [00:44:00] or, um, but under the hood they’re functionally a GPT that they then made better at a specific thing.
Christian: so in a lot of ways, your, your first entry point is what is the core, the core mission of the feature I’m adding? What is the core value add? What is the requirement? and then you can start thinking about the
Christian: medium that the AI
Christian: needs to know about. And I’m
Christian: sure there’s a,
Christian: there’s a, um, a model out there.
Christian: Yeah. Um, and I think most companies aren’t thinking about
Christian: like, do I pick this architecture of a neural net or this architecture of a neural net
Christian: because the fact of the matter is you can download pretty much any architecture from Hugging Face, like.
Ryan: You want to know what’s actually, as you were talking, I had this thought to myself, cause you were describing these different kind of, Either structures or implementations of, you know, different artificial intelligence. And I think that the best way, [00:45:00] like the best simple analogy that I can think of that you’re, what you were just saying, kind of jogged me into is saying, you know, how are you using AI is a lot like asking someone, how are you using humans? How are you using computers? You know, it’s just kind of like, Oh, I don’t, I don’t know, We kind of like, do a lot of stuff with them. Like, What do you want to, know, Right. And So I think this, like this drive that people are like, well, how are you using AI. It’s just like, I don’t know, I guess I’m like, whatever I want that makes sense. you know, like if I’m trying to categorize something well, something with learning that categorizes things, or if I’m trying to like create an automated chat, well, I’ll use a large language model, if it’s a. you know, automated chat on a specific topic. Well, then it’s gotta be that you know, Like, that’s, um, thinking about AI.
Ryan: Or even [00:46:00] just large language models, like, even a you know, small piece Of the overall AI puzzle, maybe, maybe not small by importance, But you know, by count for sure, you know, is like kinda not thinking about the whole picture and not thinking about how you can use it. I absolutely love how you tied everything back to kind of core value proposition.
Ryan: Like what is your objective? Whether it’s a business objective or, you know, company, personal, nonprofit, whatever the case may be. I absolutely love that. and I think that’s probably like the best takeaway to think about with all this AI stuff. So, I also want to give everybody the opportunity to get to know Christian a little bit, instead of just being the smart guy that knows a lot about AI.
Ryan: Tell us a little bit about, uh, your background hobbies. Like, what do you like to do for fun or outside of
Ryan: work?
Ryan: Tell us about you.
Christian: Yeah, I mean I’m a musician. Um, I haven’t released [00:47:00] my solo stuff yet. but I have been working on an album for a while. I think like I love creativity. Like, like, I love a good like paint and sip. I, love a good, Um, I love writing poetry. I
Christian: love writing. Um, and I think one of my, I, mean, one of my biggest fears about the AI world is that we’re trying to have AI replace artists.
Christian: Like, that’s mind boggling when like art is one of the most fun things you can do. Why wouldn’t you use it to replace the mundane stuff?
Ryan: That’s
Ryan: that’s actually really simply put.
Ryan: Yeah.
Christian: because the mundane stuff? is annoying. I’d much rather be focused on the creative thinking and like, I think one of the biggest, we, we talked a little bit there about, um, like AI powered organizations.
Christian: I think one of the best ways organizations can become AI powered is empower their workforce to use AI and to empower their workforce to [00:48:00] partner effectively with AI. Um, and I think, like, ugh, I could go down the rights for robots rabbit hole, ugh, um, but I, I love my, I love my philosophy, I love my, um, futurism, I think a lot of our big, like, dystopian fears about some of these sci fi topics, um, Like, I think
Christian: I think the biggest dystopianism is arguably capitalism’s application of AI.
Christian: but, AI fundamentally is, is built in our image. Like, it learns from the things we do, it learns from our biases, it learns from the data that we generate. Um, and
Christian: it learns from our behavior. And if our
Christian: behavior.
Christian: is to replace art, then AI’s incentive is going to be that art is irrelevant. and I don’t like
Christian: that world.
Christian: I think creatives, creatives make some of the best stuff happen and make some [00:49:00] of the most interesting
Ryan: that is such an interesting point, right? Like tying the dystopian stuff back to how it’s tied in
Ryan: our own
Ryan: image. Right. And it’s just kind of like, Hey, this thing that we built modeled on ourself, that’s just like way faster and has more
Ryan: resources. If you really don’t like that thing, like maybe we as humanity kind of have to look in a mirror cause it’s like, yo, this is just like you, but really fast and with a
Ryan: good memory.
Christian: Yeah, and, and better at, I mean, again, I said better at memorizing. what it needs to do in a scenario than we are. Um, and, but anyway, that’s not about me. That’s about my stance on the AI stuff about me. I love, I love the creative stuff. I would love someday to live in a world where AI handles the mundane stuff and I could be an artist all day.
Christian: I could write my music and people will listen to it or they won’t, but my
Christian: wellbeing
Christian: doesn’t depend on it. But also I, I think the AI music side of
Christian: things is fascinating and you bet, [00:50:00] you bet. I’m using at least randomization
Christian: as a
Christian: tool in my music tool toolkit,
Ryan: I love that. That is the intersection of AI empowering creativity and artistry is just so interesting. Like, you know, like you hear from so many, you know, artists and creatives of all types that, you know, like you have this idea, whether it’s, you know, You know, just like pops into your head or it’s something that gets refined over like trying different things and something catches your eye.
Ryan: And then it’s like, okay, well now I have to spend, you know, 20 hours making this look like it just took me 10 minutes. And that process of refinement. Can get empowered. It’s like, hopefully that makes everything more prolific and there’s more cool, interesting stuff, you know, to make me question my existence.
Ryan: Cause
Christian: Well, and that’s that’s the, [00:51:00] that’s the difference in the,
Christian: reinforcement learning. Vocab 1 0 2, I guess, uh, that’s the difference between exploration and exploitation, right. Um, and exploitation is a really interesting choice of word there. but exploration being exploring possibilities and exploring multiple arms on that bandit or on those bandits to find the one to exploit.
Christian: and I think especially with creatives, like AI isn’t inventing anything new. Um, it takes humanity to invent. at least at this point.
Ryan: Yeah.
Christian: And so relying on humanity for that invention process, giving them the tools
Christian: to
Christian: explore what that invention could possibly be, and then allowing the human
Christian: to exploit that exploration to really land on something they can be proud of, I think is a wonderful creative process. Uh, for
Christian: creatives out there.
Ryan: I love that. I’m excited. Hopefully, hopefully it goes that way. Um, [00:52:00] now.
Christian: not looking so good so far.
Ryan: so if anybody, uh, thinks that you’re smart or wants to have a conversation with you, what are, what are the best ways to reach out and
Ryan: get in contact with you?
Christian: Yeah, I mean, LinkedIn is a big one. Um, and then, I, this isn’t exactly the email that I, I monitor the most. but, so I, would definitely recommend LinkedIn, but my my email for my GitHub, is FriendlyTime at Gmail. com
Christian: spelled with minimal vowels. So, I think the only, the only vowel is the sometimes
Christian: Y in there.
Christian: Thank you. So,
Ryan: I
Ryan: love it.
Christian: I, you can always
Christian: post that in the description.
Ryan: Christian, this has been a conversation that was number one, kind of So,
Ryan: full of of
Ryan: dense knowledge, but there was always that moment where you would check back and make it more approachable and
Ryan: connect it back to the last thing that
Ryan: we were talking about. I am super excited [00:53:00] about some of the things that I learned.
Ryan: And I think also the
Ryan: opportunity for some of the listeners to kind of understand a little bit more on a deeper level, what. You know these
Ryan: crazy technologies that
Ryan: are really rapidly changing, um, the
and the business landscape and all that stuff are. So thank you so much for taking some of your time to come on and talk to me and teach everybody about some of this stuff.
Ryan: It’s really
Ryan: appreciated.
Christian: Yeah, I, I’m always excited to talk about AI. I’m always excited to talk about, uh, NLP. and I’m just excited for that next step in my career, uh, to really, like, jump off on that NLP train and start helping businesses gain as much information from their knowledge bases as possible. And that’s what Pryon’s doing.
Ryan: I love it. I love it. And I also want to make sure to thank all of the listeners. If you laughed at anything, you learned anything today,
Ryan: please make sure to go out and tell a friend, give us a rating, subscribe. Hopefully it’s a good rating. Um, Christian, thank you [00:54:00] again so much for being on. And this has been another exciting episode of the Making Better Decisions podcast.
Christian: That’s a wrap for today’s episode of making better decisions for show notes and more visit, making better decisions dot live a special thank you to our sponsor canopy analytic canopy. Analytic is a boutique consultancy focused on business intelligence and data engineering. They help companies make better decisions using data for more information, visit canopy analytic.
com. There’s a better way. Let’s find it together and make better decisions. Thank you so much for listening. We’ll catch you next week.
Sign up for our newsletter
Stay up to date with the roadmap progress, announcements and exclusive discounts feel free to sign up with your email.