AI UX Building

Speaker A: Foreign.

Speaker B: I'm Rob Hirschfeld, CEO and co founder

Speaker A: of Rack N and your host for the Cloud 2030 podcast.

Speaker B: In this episode we dive into how user experience UX is going to be changing with AI as the interface. We talk about interesting concepts about stochastic and deterministic systems and how all of these things relate to what is easy and effective to build with AI, especially by coding AI and what is difficult and challenging and accelerates AI without actually

Speaker A: being built by AI.

Speaker B: Really fundamental challenges and discussion topics that we've been hinting at and in this episode dive straight into. I know you'll enjoy it.

Speaker A: Are we gonna do away with traditional form based ux? Like uh, you know, I can see a time in the not too distant future because I'm already talking to my phone a lot more than I used to just for text and transcription. It's, it's convenient, fast and accurate. Yeah, right. I can I convey more than you know, you know BRB that you know, I could see a time in the not too distant future when I'm not filling out a form, even if a form's being presented right where I'm literally saying, you know, hey, put my information in this form and you know, log, log me in, put my information in this form and go do something. That's what you know, you know, a year ago people were thinking agents would be helping you with M and, and that's, you know, that I, I, that would mean pretty fast that we're going to stop building forms. Like my whole, the whole idea of me uh, putting together a UX with a form or something that I'm going to click through or anything like that. There's two sides that are problematic to that. One is the user has to take information, they have to understand the context. But also when you're building that form, you don't have an opportunity, you have to ask all the questions up front. Right. You could conceivably turn that into, you know, a dialogue or a wizard where you're submitting the form into an AI and the AI can say I don't actually have enough information, but based on this I think you want to do this and I need this additional information for you. Can, you can already provide. It's already happening.

Speaker C: It's absolutely already happening. I mean there are a couple of

Speaker A: um,

Speaker C: frameworks that I've already seen with for agents that are doing um, SQL queries by, and, and you, you will go and say you don't know you're doing an SQL Square query. Actually, you know you're going to say, I want the, I want to report on the following, or give me a list of X. And, uh, you know, the agent will come back and say, well, you know, I, I don't quite have enough information about, you know, some aspect of what you're asking. Where should I be getting this? You know, should it be from, um, my list of states, you know, is it. Does it have to be from Georgia? Does it have to be, you know, something else? So, yeah, those, these are already, this is already happening. And there are some ways in which they, they get somewhat standardized, but you don't think of them as forms. You don't think it's, it's just not the notion of a schema.

Speaker A: Oh, yeah, you don't. Yeah, yeah. So we're, we're. I, um, this to me is st. That's the stochastic system. You're saying you're, you're putting the AI in front and you're saying, you know, ultimately I'm going to drive something behind the scenes. It's a deterministic process, you know, and, but all that interface in front of that. We're going to let the AI, you know, in term, determine your intent, ask you additional questions, come back.

Speaker C: Yeah, yeah.

Speaker A: And, and so a whole U product, UX almost. It's weird though, because at some point you're like, all right, where am I adding value? What am I delivering for that? I mean, I guess if I want to buy an airplane ticket, who's the we here?

Speaker C: When you say where am I?

Speaker A: I'm just looking at me personally. Um, um. But I would then extrapolate that to the, um, consumers or people in general, right? I'm buying an airplane ticket. My conversation, you know, I don't go to the, the airline app. I would go. Or, you know, maybe I may. I actually, I wouldn't. Here's where things get weird, right? I'm going to go to my phone, right? And then say, I want to travel, you know, to this city, um, you know, on these dates.

Speaker C: And, you know, you know, my habits, you know, my likes, dislikes, you know,

Speaker A: my credit card information.

Speaker C: You have my cred card information is, you know, what, what do you need to know? What do you need to know that you don't already hear to set this up for me? Um, yeah, yeah, no, that's, you're, you're, you're asking the, um, you're asking the agent to act as a concierge. It's not just an interlocutor. It's, it's literally, it is a concierge that, you know, hopefully has your best interests at heart, and that's what it's about.

Speaker A: Yes, I, I heard an interesting discussion, uh, on, um, Hard Fork, where they were talking about how potentially sensitive data would be leaked across these agents. Oh, yeah?

Speaker C: What was the upshot? What was the upshot of the conversation?

Speaker A: So they were talking about new. They had gone to the OpenAI, um, Developer Summit, Dev days. Dev Summit, and they were, you know, OpenAI is announcing all these integrations with all these partners. But, you know, the, the challenge with these integrations is that they have to share information. And it's not clearer because talking about an agent, I keep sitting in front of Windows. Sorry. Um, it's okay. I.

Speaker D: The.

Speaker A: Because you're talking about integrations with an agent where I can say, well, I need this information, I need that information. Be very easy for Cambridge Analytica, they made the comparisons, um, to, to say, I need, you know, additional information or I want to know everybody in the social graph or whatever, to start pulling that data back as part of the, um. Yeah, yeah.

Speaker C: I mean, at what point do you, you know, question, uh, why do you need that information? Um, what's the, you know, what's the purpose?

Speaker A: This is, this is, this is the interesting thing, right. The nice thing about a UX in a form is that it says, I need this information to proceed. You can decide to, you can, you can decide with, try transparency that you're going to get it or you're going to give it. Um, and it should be, you know, at that point, you're syntactically complete in the request with, with this. And there's, there's good and bad. I could see a great thing where, you know, you could be, you know, booking an airline ticket and it could say, you know, oh, there's, you know, this flight. I know you prefer directs, but this flight's in your preferred time zone, time travel thing. Or it lands here, or, you know, this is your favorite. You know, uh, right. There's, there's, there's layers in this that become very attractive. Rather than me just booking a flight, knowing what I need to book, um,

Speaker C: well, you know, you've just also laid out something and that is you're making a request that is being built sequentially. It's basically, it's an interrogation. It's like, let me help you fill out the form here. What's your middle name? What's your phone number? And you don't have any gestalt of, oh, geez, you know, I'm going to have to provide all of this information in order to make this request. Uh, I'm going to think a little better. I'm going to think better of that. Um, so there's a linearity to the way that's being done in many of these cases. So I mean my, my approach to that is if I give permission to an agent to act on my behalf, somehow I go through a, uh, kind of a, uh, a kind of a ritual training. It's what it comes down to. It's like I want you to write a standard email response or. Um. One of the things I was, I was looking at, uh, I had to make a bunch of introductions and I. When I make introductions to. From people, my premise is it would be good to kind of fill in the little blanks. You know, this is, um, you know, this is Rob and um, um, he's the, you know, he's the CEO and, and co founder of Rack and. And blah, blah. And here's my. Roughly, here's my history with. With Rob. And you know, he's a great guy and he really, really understands X, which is why I'm putting you together. Hey, Rob, this is, you know, Joe Dokes and so forth.

Speaker A: Well, yeah, the matchmaking.

Speaker C: It's a little bit of matchmaking, but it's like a, um, you know, it's the one paragraph. Why I'm introducing how I know you.

Speaker A: Yeah.

Speaker C: Why.

Speaker A: I think it's a template, but it's, it's right. It has to be a template in

Speaker C: a way that it.

Speaker A: Yeah.

Speaker C: And that you. I can, I can, I can pretty much train, uh, an AI, you know, an agent to do that.

Speaker A: Right.

Speaker C: Correctly and not. And not reveal, you know, um, something that I know that you've shared with me in confidence, for example, or, you know, something like that.

Speaker A: Makes sense.

Speaker C: But there are so many of these queries that you. You've just been talking about, you know, planning out a trip, planning out a. Um. That by itself. Um, just making all the travel arrangements, coordinating them around a kind of a schedule of production, thinking about contingencies. Do I want to, um. You know, are part of the decision. Is part of the decision making. Making it as inexpensive as possible. Is there, Are there other factors involved here? You know, I really absolutely, positively have to be in Las Vegas on this date at this time because I have a, you know, an obligation to be on a panel or speak and you know, so forth.

Speaker D: So on.

Speaker C: Those are, those are not. Those are not easy things to do linearly. Serially I should say.

Speaker A: Right. Well, uh, AI has been good at travel planning. I've met people who are like, okay, I asked AI to tell me what, create an itinerary for this trip with these, and that's it. They've had very positive experiences with that.

Speaker B: Um,

Speaker A: yeah, it's an interesting dilemma because I think that we're seeing very productive outcomes we're not evaluating yet, or it's hard to anticipate, impossible to anticipate just how much of our data is going to get shared or how much, um, potentially systems are going to be manipulated behind the scenes based on our feedback. Right. You know, I, uh, just heard, ah, this was on npr. They were talking about the big retailers using AI to target sales for people. And they, they think they're going to be able to juice juice revenues a little bit by having better, um, targeting better ad.

Speaker C: Are we talking about like, I'm, I'm, I've got kind of micro sales. In other words, I've got a sale on, and it's specifically for Rob, and I've got a different sale on that's specifically for Rich.

Speaker A: Correct. And then also targeted down to, you know, what product I think you would be likely to buy or. Right. Oh, no, but, but, but you get, you know, that's the way they're positioning it is, oh, it's a discount that I'm being presented for you, but it's going to juice revenue because you're going to buy something that you wanted, blah, blah, blah. But I'm giving you, you know, but the same is true on the other side of increasing prices. Um, right. Or, you know, sending you to the resort where, you know, you, it's more expensive, but it fits your, your profile rate. There's, there's matching and there's beneficial matching, but it's, it's a funny.

Speaker C: And then there's predatory matching.

Speaker A: There's predatory matching and there's slippery, you know, it's a very slippery slope between the two.

Speaker C: Ah. So getting back to the deterministic versus the stochastic, that seems to be a, you know, a focal point of what you're thinking about or concerned about.

Speaker A: Well, it's, it's in that, um, you know, the AIs, it's pretty clear to me that the AIs are going, are already there or going to a path of the stochastic interface. Everything we've been describing, I would describe, I would say is a stochastic interface. Um, but those have to be supported by, um, some foundational abstraction. Um, and that's A deterministic system. And so if you want to speed the thing that. And this is. I think part of the question is how. Where's that line drawn? Does, um, does it help to have a determining system that gets up to a certain point, does that then drive better AI integrations? Because I, I have seen the, the more you have AIs, try and build a process that should be repeatable. The, The. You get a lot of slop. You get a lot of. You know, it's just like trying to. You know, I was. I was working on a AI picture for, um, a presentation, and I had part of it that I really liked and part of it that I really liked. And then I. My prompt was like, you got the left side and the right side. Now give me an integrated image. And it built a whole new thing. Totally new.

Speaker C: Yeah. No, I mean, look, it's not.

Speaker A: It's not good at doing that. That part of the.

Speaker C: Well, here's the question. Is it because it's stochastic or is it because, um. What do I want to say here?

Speaker A: There are.

Speaker C: There. You can make a case that even with kind of very kind of locked in abstractions, you can get yourself into trouble. And in, in this particular case, it's a. You know, it's beyond the agent's ability to understand or your, Your ability or willingness to describe specifically what you mean by putting them together. You can see the picture. You can see the whole. You can see how it would all fit together. And, you know, it would be, you know, do you have to go.

Speaker D: Do.

Speaker C: Does the agent have to go into a. A, uh. An aesthetic conversation about, uh, composition and, you know, placements and, you know,

Speaker B: perspective,

Speaker C: uh, of the picture and, you know, all of these. All of these types of things. It's making some assumptions. Yes, it is stochastic. It's going to go, all right, I'm looking at the. The world of pictures and uh, yeah, I'm going to slam it together like this. It's either that or there's. There's a. The other issue is we've come up. What you're coming up to also here is a sense of. At what point do you assert that an agent understands aspects of taste has, you know, which, you know, taste is taste. By that I mean, you know, you do things aesthetically and that can be. You can do. You can, you know, aesthetics have as much to do with the way you code, the way you build an interface, other types of things. Um, that's, you know, it's. It's kind of how do you put those? How do you codify or how do you represent that knowledge? You know, there's a. There's kind of, uh. You get to a point of, of abstraction. Layers of abstraction where, you know, that's what I think. I, that's what I think Cory Doctorow is all about. About shidification and shitification. It's like building abstractions on top of abstractions on top of abst. Actions. And yeah, you made it easier for, for someone to build code with first with compilers and then with, um, various kinds of layers of.

Speaker A: That's funny because I, I don't. I don't think the insidification. I don't think that's where he goes with insidification. I think his, his concept is strictly a, um, commercial transaction perspective. Where the insidification is, is that you are intentionally decreasing your level of service because you have a cap. You have a, you have a captured,

Speaker C: uh, You've captured the audience.

Speaker A: Yeah.

Speaker D: Yeah.

Speaker A: No, I, I don't. I don't think it's a quality statement. I think. I mean it is a quality statement, but I, I think it's. It's a, um.

Speaker C: Maybe I'm talking about something else.

Speaker A: It's. It's a good. His. That's a good way to describe the intentional breaking of the. The implicit contract when you created, when you created the service relationship. Um.

Speaker C: Okay, that.

Speaker A: That's me. I, I think. I think we, we do see, um, a, uh. A degradation of the user experience. Um, for I, I think as. Especially as the user experience gets broader, uh. I. It's funny because I thought you were going someplace different before you went to the aesthetics piece.

Speaker C: Well, um, Taste is what I was getting involved with. Yeah.

Speaker A: I, I think t. And taste is important and I think taste is, um. One of the things that has people concerned with the AI Music and AI slop is that you end up. It's not slop. You end up with bespoke, um, creations for the, you know, for the individual. And you're like, oh, I like this type of music and this type of thing and this type of thing. And you get something that's exactly suited to your pace. I always. That always bothers me. I'm off on a little bit of a tangent because to me, each human ends up with a set of preferences, but those preferences are formed somewhere and there's a significant amount of, of enjoyment that you get from shared experience. Uh, and so it's. Or.

Speaker C: Or getting, you know, part of. Part of. Part of being human is Being surprised or being pleased by something, you know, that you haven't heard before, I haven't seen before. All of those very, uh, buy into that.

Speaker A: It's sort of, it's sort of. I, yeah, it makes me nervous that we're losing something on the other side. I, I think I, I, I look at, but I look at the other side because, right, we're, we keep going back to AI Creating bespoke experiences. And that's the stochastic piece. It's interacting with the specific stimuli that it has to. To make decisions and solve that one, that one unique problem. Right. AI is this is, uh, and, and this is where I've, I've been trying to get. It's like you're, you're, you. We have this new emerging ability to treat each request as a unique experience, which is fantastic. It, it's, it could be incredibly powerful. On the back end of that, there's a, there's fulfillment components, the deterministic parts where you actually feed that into an industrialized, uh, process. And it needs to be a set of processes are consistent, repeatable, predictable. And what I, what I, what I'm not as clear on is how deep the personalization goes into that, into, into that.

Speaker C: How, how, how deep you want it, how may maybe how deep you want it to go without, you know, finding yourself losing, uh, some, some capacity or some capability to satisfy. I mean, I, I really, um. There's a, there's a certain point that says, all right, let's go back to the music idea. Yeah, I'm listening to an artist, you know, um, that has created something versus, I've said, you know, I like stuff with this kind of backbeat and I like these kinds of instruments and so forth. I'm actually composing something for myself, but I might not be all that talented a composer, quite frankly. And so, you know, what I'm going to get is, yes, arguably what I've described. But, you know, it, it could be pretty, could be pretty anodyne. You know, it could be pretty. It could be pretty m. You know, it could be a nothing burger. All right. As opposed to, you know, here's an artist that, you know, has real talents at this, understands music, has a particular feel for it. And part of being human as opposed to being an agent, is the capacity to be surprised and be, to please be pleased by the encounter of something that hasn't happened before. That's one, one aspect.

Speaker A: Yeah, it's an interesting, It's a. We're, we're. There's a whole Other topic for us. Look at that. That, that is, um, about some of what you just described. Allows people to be creators without building mechanical skills on the back end, which, which in some ways is similar to the same. You, uh, know, learning how to play guitar, piano or tune or sing tune. You know, uh, those are, those are all deterministic skills in that you have to build them over time. There's, you know, you, you. There's tremendous discipline involved. There's a lot of things that require that type of discipline. I was thinking, you know, my example didn't go towards the music as much as, you know, if I have a, uh, if I'm manufacturing cars, you know, we're getting to a point. I'm thinking about some of the, like the, you know, the robotics on the new. New car. Vehicle stuff where they have robots building, building the car. Right. Very. Um, you know, it's, it's the, the neat. The process is programmed. Not it's deterministic, but it's programmed where, you know, does AI reach back into that process? You could see somebody ordering a car, having a, uh, an A.I. do all of the work to say, okay, I'm gonna, you know, this is the feature set you need. This is the, you know, color the trim. Right. There's parts of that manufacturing process. If they're 3D printing the dashboard, they could, you know, adjust the sight lines to your height when they manufacture that car for you. Um, now does the change. Let's, let's. All right, let's take that example because I think that's a, That's a fun.

Speaker C: That's a good one.

Speaker A: Practical example. So the ability to pre program the, the dash, the, the dashboard configuration to pre print it differently. I'm assuming there is somebody, there's something, there's somebody that made a decision, oh, we could change how this is printed and it becomes something with parameters and features. It's still a deterministic capability of the vehicle assembly line. Okay? It's not, it's not. And there's. So there's a firm interface and there's firm guidelines. I mean, we have regulatory pieces that makes cars safe so that you can, you can do that.

Speaker C: Still has to have an airbag. It can't have sharp edges.

Speaker A: Yeah, everything that, that dashboard has to conform to a whole bunch of pieces to it. So how far into that process does the a. Does an AI interfacing with the user get to reach? Right. Do they basically say, oh, I can, you know. Right. If you give you height and I can do this, I You know. Oh, wait, you know, you like, um, you know, pumpkins. You know, we're gonna, we're gonna put pumpkins, you know, we're gonna put pumpkin symbolic symbols into it. Um, right. You know, we're gonna basically, you know, that, that's type of customization school. I, you know, it's. Well, think we're getting to a point where those things are going to get pulled more and more out of the deterministic process. But the deterministic process, actually, the robustness of that process is critical to enabling,

Speaker C: to making that, uh, work. Well, it's got to be in.

Speaker A: This is what I'm struggling with. Right. So, so there's this interesting thing of. You need that deterministic processes that can have AI, uh, flexibility. This is, this is the whole. All right. This is the whole thing I've been trying to articulate. The quality of your deterministic processes to allow AI to reach in and have variants is really important here.

Speaker C: Okay, I'm. Now this is, this is a better idea.

Speaker A: What you're struggling to, Struggling to articulate. So if, if I, uh, you know, we're building a car, we need the car to be delivered. There's places where reasonable customizations could be made. And having a assembly line that accommodates that is really valuable because now, right. I have an AI in the front that actually enables all sorts of interesting things to be done with that deterministic process within boundaries. The deterministic process also has to be able to know where it can accommodate those types of changes.

Speaker C: And so, and in some case where, uh, kind of make a decision, actually determine who's made the decision. Let me, let me give you another aspect here. If by offering me the ability to, you know, custom build, you know, have a bespoke dashboard, physically, you know, placement of seats, what I can adjust what I can see and so forth. Um, what does that do downstream with respect to servicing the vehicle, repairing it if something goes wrong? Again, those are constraints, those are guidelines, or you tell the customer, yeah, I can, I can make that happen. Uh, let me give you a little bit of advice. However, this is going to. Downstream, potentially cost you money to get it repaired or replaced.

Speaker A: Right.

Speaker C: Um, it's like, yeah, I can give you this, this kind of a paint job and this with these colors. But I can also tell you that what you've asked for and what you're doing here is, uh, quite sensitive. And if it gets scratched, if it gets. If you have a, if you have a, you know, a parking lot incident, you know, again, Your, your whole point here is you're, you. You've offered up all these degrees of freedom, right? One of the things that you could do with an AI that actually makes that more realistic and more possible is to have the implications of the things that you have ordered. MHM predicted. I can tell you from understanding the experience that, you know, happy to do this. You tell me you want it. Great. Let me also inform you of the implications of doing that.

Speaker A: That is a predictability on deterministic systems that is missing and needs to be. Now we're, now we're, this is. Now we're talking about where we're getting to the right question. It's like, all right, I can influence things more. The deterministic system needs to also be able to echo back predictive feedback because it's going to be able to respond or it's going to. So part of, part of what we need to be thinking through is how we build deterministic systems that are, that have AI, uh, interfaces. That's. This is the thing I'm struggling with as we think about what we do with Rack N It's. We have a deterministic system balancing.

Speaker C: It's balancing intent, its intent and the implications. It's, uh, in, in many of these cases, I want to stand this up in these places and I want it. And I want to be able to, you know, power down that, that rack, you know, under, under certain conditions. Right. Okay. I can, I can make that happen. Let me make sure you understand the implications. Do you know the implications of that?

Speaker A: Right. And I, uh, I think our first generation on that is just going to be. No, but right there, there's elements of, uh.

Speaker C: It's going to be like I talking to it everywhere and anywhere. The first thing you say is, no, that's too hard. It's not the way we do it. No. I want it to be configured the following way. Okay, I'll do it. But it's going to cost you m. The universe.

Speaker A: The universal AI problem right now is that, Right, AI will build whatever you ask it to build. It has no knowledge of cost or, you know, maintainability.

Speaker C: Unless, Unless you have built that into the user interface and into the, the whole process.

Speaker D: You.

Speaker C: It doesn't. I mean, it's there to satisfy your, you know, it's the genie in the bottle. Yes, master. I'll tell you what. You, you know, I'll do what you want me to do.

Speaker A: Yeah, the genie is exactly the right analogy because the genie's wishes always have unexpected consequences in them, which which is, which is why the, the determinant. This is. This, this is to me exactly where I wanted to get with this part of the conversation. Because what we're saying is, look, AI systems are going to be able to talk to you and ask you what your greatest wish is and then try and go figure out how to build it. And then the systems on the other end there. There's still something that has to be responsible for. I can't do it that way. I have to deliver this. I have to have. Right, you. You have to have the, the backing system has to have guidance to make sure that it. There is. There is.

Speaker C: Yeah.

Speaker A: Uh, a, I mean a, A realistic process for that wish for fulfillment.

Speaker C: If, if the only constraint on an AI is I uh, can't break any, I can't break any laws of physics.

Speaker A: Right, that's right.

Speaker C: And every. Anything else that you request. Yeah, I'll, I'll work it. I'll do it. Did you tell me to think through the implications? Do I know enough about that? I could actually give you the downstream implications.

Speaker A: That's, well that's, that's ends up being the. What was actually the paperclip problem. I was listening to one of the, the author of the paperclip problem. He's like, yeah, people dumb that down way too much than for what I intended it to be. But um, but yeah, the, the idea of the AI, you know, you asking for a paperclip in it, it, you know, going, going way out of bounds on that is. Is very real. The deterministic guard. You know, it's funny because I, I didn't think of determinist systems as guard rails and I still don't think they're guard rails in the AI sense. But you know, the, the AIs. AIs should operate in a constrained space from what they can deliver. I guess it's one of the reasons

Speaker C: at least a def. A defined. And if you said if it was a defined space as opposed to a constrained space.

Speaker A: Defined space. Yeah. And so, so I uh, do think there's going to be increasingly I'm uh, looking at just our space and then we're, we're just not out of time. But like when I look at our space. Cause I think it's an interesting example. Of course I think it's my space. Um, but, but the, but the, the idea is right, our competitor, our comp. Our competition is. And it's always been this do it yourself bunch that, that people build together in all sorts of bespoke ways and so an AI. This is what, you know, I think about an AI could, you could say, hey, I want you to build something out of these bits and pieces. And it's going to go build it. But there's no deterministic, no constraint. It's doing whatever it thinks is the right thing to do.

Speaker C: Uh, it's either doing what the, it may be doing what it thinks the right thing to do is, or the right thing to do is listen to the boss and do what he tells you to do.

Speaker A: Correct. Well, that's, this is the, the. And the, the challenge becomes, well, you did it this way this time. But I, now since I wrote a whole bunch, I have an AI prompt driving it. I want you to do it this way. And it changes the way it delivered that, that process. Right. What, what we're seeing is increasing value by having a higher deterministic entry point where the outcomes are much more controlled and uh, the requests are the same, but the processes are much more deterministic. That's an interesting part of, uh, tempting to rebuild everything from parts.

Speaker C: Yeah, yeah, yeah. And another aspect by the way, is going to be memory and the retention of knowledge. The next time you ask the AI to build X, does it. First of all, did the first time. Did it take good notes, appropriate notes? Second, the second time, does it know to look for those notes? Does it, does it know that it's done this before and would be a good thing to kind of look at precedence? Um, and how do you do that? Which precedent? Again, these are aspects of setting and maintaining or recreating context. It's anticipating downstream implications of the action you're about to take or that you're being asked to take, told to take. No, uh, these are, uh, this why people are going to get, get themselves into a lot of very weird situations and putting constraints or definitions on agents in an ad hoc fashion whenever it doesn't do what you wanted it to do this time.

Speaker A: Uh, right. All the experiential learning is problematic. Yeah, it's such an unconstrained area.

Speaker C: Um, it really is unconstrained. It's undefined. It's not even a matter of constraints. I'm not even sure we know enough to put proper constraints on many of these things. Yeah, you can overly constrain. You know, it's just, it's, it's a, uh, it's a, it's Wild West. So you. This is where learning comes in. This is where memory and history comes in. And it's also where forecasting and thinking and incorporating Downstream implications of the action you're about to take or if you're an agent comes into.

Speaker A: Um. And I think that's going to be the distinguishing factor for agents. It's a really.

Speaker C: Absolutely.

Speaker A: Oh, it's funny. Uh, I'm, um, holding my comments for next week. Uh, this is good.

Speaker C: I came across some interesting new approaches to models, including one model that is

Speaker A: um,

Speaker C: trained on 7 million, not billion million parameters that is beating quite handily all of the latest, um, LLMs, um, in things like the uh, the uh. Ach. Um. What's the ACI Agh2. Anyhow, the, the models that are, you know, kind of the, the, the super hard ones where even the, the frontier models, you know, are, you know, getting rated at, you know, 8 and 9%, uh, on this, on a 100 point scale, you know, for, you know, and it's beating them by a couple percent. All right. This is tiny.

Speaker A: That's significant. Yeah, yeah.

Speaker C: For that level. And, and it's a different, it's not a transformer model.

Speaker A: It's.

Speaker C: They, they've, they've taken a different uh, approach. It came out of the uh, Samsung AI Labs just was written out.

Speaker A: Okay.

Speaker C: So I'm looking into it.

Speaker A: That would be fascinating there if there was a different.

Speaker C: Yeah. Anyhow, cool. Great to see you. Uh, good to m. Talk to you as always.

Speaker A: I, I feel like this was, this was a youthful breakthrough. So I, I appreciate it.

Speaker C: Good, good.

Speaker A: This, I, Yeah. This balance of, you know, the, the interface between deterministic systems and stochastic systems I think is, you know, understanding that um, is going to be really important because I think deterministic systems are actually going to be good. Deterministic systems are going to have, you know, have value. Maintained value.

Speaker C: Yeah, absolutely. I completely agree. I, you know, I, uh. My big thing is how do I approach incrementally a situation where I feel comfortable enough to give agency to an agent to do certain things without consulting or doing more stuff before it consults with me about, you know, next directions, next instructions, kind of where I'm um. Definitely a bunch of my time.

Speaker A: Yeah, I want to see that. Wow.

Speaker B: In some of these episodes I really get a very clear insight about how to describe the challenges that we're facing. A really fundamental shift that AI is bringing into how we think about building systems and what makes things work well and what things are going to continue to be challenges. Uh, the hard problems that are still hard problems. If you're enjoying this conversation and if you're listening to me now, you probably

Speaker A: are please let us know. We want to hear your input, get

Speaker B: your ideas, hear what's working for you.

Speaker D: Thank you for listening to the Cloud 2030 podcast. It is sponsored by by RACN, where we are really working to build a community of people who are using and thinking about infrastructure differently. Because that's what RACN does. We write software that helps put, uh, operators back in control of distributed infrastructure, really thinking about how things should be run, and building software that makes that possible. If this is interesting to you, please try out the software. We would love to get your opinion and hear how you think this could transform infrastructure more broadly. Or just keep enjoying the podcast and coming to the, uh, discussions and laying out your thoughts and how you see the future unfolding. It's all part of building a better infrastructure operations community.

Speaker A: Thank you.