Skip to content
What the #%!@ is Vibe Coding?   episode artwork

Apr 24, 2025 1 hr 12 min

What the #%!@ is Vibe Coding?

from Invisible Machines podcast by UX Magazine Hosted by Invisible Machines

Product design sage Tim Wood returns to the podcast for a conversation about the deepening integrations between experience design and AI, including a dissection of the trending term “vibe coding.” Currently a design lead at Meta, Tim has hands-on experience leading AI-first design and this episode explores what that means on a tactical level. Having been the Principal Product Designer for AI Platforms at AWS, and Head of Design for Amazon Q Developer, Tim shares hot takes with Robb and Josh on what automated mainframe migration means for SaaS providers as well as the broader business community, which leads to the real-time birth of another new term: “placebo swipes.” ---------- Support our show by supporting our sponsors! This episode is brought to you in partnership with OneReach.ai — offering turnkey private architecture for building and orchestrating your team’s AI agents with guardrails. Use any AI models Trusted enterprise-grade security No vendor lock-in Discover more about the power of orchestrating agentic AI: https://bit.ly/4jBfVRt --------------------- Our new book is out NOW! The revised and significantly updated second edition of our bestselling book about how teams can successfully orchestrate AI agents, Age of Invisible Machines, is now available everywhere! https://a.co/d/im5rb8F #InvisibleMachines #Podcast #TechPodcast #AI #AgenticAI #AIPodcast #UXDesign #ExperienceDesign #AIDesign #UserExperience #AIAntegration #SaaS #DigitalTransformation #Coding #VibeCoding #TechTrends #EnterpriseAI #aiagents #aibooks #bestsellingbooks #aws #meta

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

What the #%!@ is Vibe Coding?

0:00 1:12:57
of matches

Transcript Auto-generated, may contain errors

What the fuck is vibe coding? Yes. Rob, you showed me a bit the other day. Today, we're going to talk about vibe coding and a lot of other things with our friend Tim Wood, who's a returning guest. But just to set the stage, Rob, tell me, describe vibe coding. You kind of showed me, but it's sort of this unique hybrid almost of design and development. Yeah, it's the last thing I would have called it, but I get it. I get that people are vibing out. I think it's a developer term. I think designers call it designing. And I think the more we do it.

The more I'm doing it and the better the models get, the less coding there actually is and the more designing there is. And I don't think it's more than 12 months before it's not vibe coding, it's just designing, which is, you know, it's designer's time now. It's all those tools that we use. I design real in real time and it just manifests. It's crazy how it improves my design ability. It's, you know, I know that we made that migration to Figma and that was a big deal because it was more interactive, but this is like a whole nother level.

Yeah, I watched it and I'm sad to report that it was a bit of a vibe, but I do think it was more designing for sure. And Tim Wood is a great person to have in to talk about this stuff. He was principal designer for AI and AI platforms at AWS and then eventually became head of design for next-gen developer experiences. And a lot of that work was working on Amazon Queues. I think it's the developer tool, which can migrate mainframes, which is really, really huge. Yeah, I think people haven't really let that sink in. That just seems to be like not talked about enough. And I think maybe a purpose by Amazon, but if you can migrate...

you know, from a mainframe in two weeks, which took years, you can migrate anything, meaning all of your SaaS software, click of a button. I can just see Amazon and other cloud providers just cranking through these migration scripts, being able to migrate from, yeah, from licensed software to open source, just bam, bam, bam. Yeah, I don't know if the death of SaaS bell has ever rung louder, you could say. Yeah, death of SaaS and almost in a way the birth of open source because, you know, because now, you know, we're just going to produce software and it's all going to be open source and yeah, crazy, crazy world. Like open source is on a roll now.

Well, and Tim, too, has a cool perspective because he's able to also talk about designing the experience of migrating the mainframe. And that's really fascinating, too, because that's this kind of AI first product design where the humans are really kind of collaborating with the AI and making decisions. It's like a very forward thinking approach. Yeah, the UI for that, I think, represents a page out of the AI first book for instead of these bolt on. components that people are putting into their traditional UIs that are AI, these little magic wands or whatever. We talk about a real first, AI first approach to design, which is huge. All right. Yeah, it was cool because we also got the chance to talk with Tim about one of your favorite things, Rob. Ambient intelligence, yes. Ambient intelligence, yeah. Specifically, even when like the non-electronic objects in your life now have the ability to tell you stories and communicate information.

Yeah, Providence, they call that, right? Like, it's when a painting has a whole story of who painted it. And, like, we spend a lot of money on stories. Providence is, like, the number one driver of value in these paintings. So now everything has Providence. You can walk into a thrift store, like, get lost just finding the backstory of everything. I once wanted to, like, I still want to do this. I want to attach a RFID to a dollar bill and connect it into the cloud and have an AI agent attached to it so people could talk to the dollar bill, find out where it's been, who it's talked to as it passes through from hand to hand. I think it'd be...

It would be fun. So yeah, everything will have provenance now and you just hold your phone near it. And what a cool concept. Like you said, we'll never get out of our house. Yeah, we'll never leave home again and just talk to all our stuff. And a couple of quick notes on this episode too. As the title suggests, there's a little bit of salty language. There's an appearance or two from our producer, Elias Parker, and somewhat randomly, but I guess for good reason, this conversation opens with a discussion about blackberries and bear shit. So enjoy that as well.

Hey, exciting announcements. Age of Invisible Machines is out now in paperback in this newly revised and updated second edition from Wiley to reflect all the things that have changed and to reinforce, I guess, all the things that haven't changed in the world of conversational AI or agentic AI as we're now seeming to call it. So the book contains the same evergreen information that you can find in the first edition, but there are also three new chapters about generative AI and architecting agentic solutions. Also a bunch of pearls of wisdom throughout the book gleaned from amazing conversations with all sorts of luminaries right here on this very podcast. Age of Invisible Machines, second edition, out now in paperback from Wiley, wherever you get books. And if you feel so inclined to leave a review of the book, that helps us out greatly, especially on Amazon with all their crazy algorithms. So yeah, buy the book, read the book. And also on the way, the QR codes in this book lead to an intelligent digital worker that we have built, an agentic solution basically that lets you talk to the book, interact with the information in the different chapters of the book, and have one heck of a good time talking to a book. Hey, you ever talk to a book? You're about to talk to a book. Grab your copy today.

Maybe you want to get a copy for a friend. Maybe everyone on your team needs a copy. Maybe everyone you work with or everyone who works for you needs a copy. Maybe everyone you've ever met should get a copy of Age of Invisible Machines out now in paperback, a newly revised and updated second edition from Wiley. All sorts of fun stuff in here. Go forth, grab a copy and enjoy.

Please support our show by visiting our sponsors. This episode of Invisible Machines is brought to you by onereach.ai. Building good AI is hard. Onereach.ai gives you turnkey private architecture and the tools your team needs for building AI agents that can do real work. Their powerful GSX platform lets you avoid vendor lock-in, use any LLM in the marketplace, build secure solutions with reliable guardrails, incorporate legacy software, and create production-ready automation with enterprise-grade safeguards. For anyone who is serious about automating their organization, the GSX platform from onereach.ai has everything you need.

to get started. Learn more about orchestrating AI agents for your team by visiting onereach.ai. We've got this land, a couple of acres, and it's all designated wildland. And there's a bunch of non-native species in there. In Washington state, blackberries are a huge problem. These Himalayan blackberries have taken over the whole state. So we cut the blackberries back, and then you've got to plant native species to take control of the land. And so we're in the process of working with the county and the state to help return this land back to its natural state and habitats for wildlife and stuff like that. So we've got to go out and have to cut willows. Do they want you to replace the non-native invasive species with native invasive species?

With just native species. Yeah. And they take over. But is it supposed to be like aggressive stuff? Like that's going to. Yes. Like somewhat. Are they tasty blackberries at least? Oh my God. Yeah, they are. And everybody loves them. In fact, what happens, Josh, is that, you know, I spend a lot of time mountain biking up in the mountains over here and you will find giant fucking piles of purple bear shit like all in the fall. It's crazy. Like we're talking like a foot tall, just pure purple seeds and everything because everybody's gorging on these blackberries and spreading the seeds everywhere. And they're delicious. They're huge. When you say everybody, does that mean it's not all bear shit? Some of it's like you're not sure.

Not entirely sure, yeah. But at least it's not yours, right? That's not mine, but you know how berries are. But that's like accelerating the already invasive species, right? Because then they're like getting consumed and fertilized into a new place. Let me tell you guys another story. I have apple trees out here too that, you know, in the fall, they're pretty productive. In fact, we spent a lot of time pruning them and getting them good to go. What happened was we had about six apple trees and pear trees full of fruit. In two days, all of that fruit was gone, 100% gone. And what happened was one family of bears, like just a mama bear and three cubs or two cubs, came in and ate probably about 200 pounds of apples in two days.

Wow. Right? And what they do is- That's like my kids. They eat it and then they barf it back up in the grass. They can't possibly consume it because they just eat everything and then they just barf it back up. Then they come back the next day, barf's gone. Okay. Well, at least they clean up their barf. It's all really interesting stuff here. Most party guests don't. Rob, there was another kind of pile of shit that we were talking about earlier. Well, we were talking about vibe coding, which is- Well, Rob is not a huge fan of the term vibe coding. I hate it.

Although he gave me the opportunity to watch him vibe code, although he would call it designing. But he was indeed designing a web app using a conversational interface that was, he could use different LLMs to write different pieces of code and test the code. And he was just going over and over again. And apparently some people are calling this vibe coding. Yeah, I can't stand this term, but the idea is like you're now coding with natural language. So you're just telling the thing what you want and just changing it. Yeah, I know some people want to hear about this, this vibe coding term. I know people who want to know this. This term that it seems to be picking up, which is terrible.

Because to me, it's like, I don't know why the word designing needed to be renamed. I thought designing was pretty good. But like, yeah, people are like, oh, we're vibe coding. And I'm thinking, well, that's just because you weren't a designer and you don't actually know what designing feels like. Because if you were a designer and you started vibe coding, you would be like, oh, I'm designing. Because that's what you're doing. They're not coding at all. There's no coding. The AI is coding.

and now you're designing. But it's interesting. It just feels like it's developers that want to be designers but realize if they call it designing, then they're no longer qualified. Well, yeah. Gosh, do you remember, Rob, when the whole designer-developer thing blew up and there was the whole diviner? Diviner. Remember that term? Oh, yeah. Okay, I use it all the time. I was wondering if there was ever developers.

They're vibing now. Yeah, whatever. But yeah, it's interesting that so fast we've gone from like not coding to designing the pure design. We were just kind of jamming on that a bit and how fast that just happened. What's interesting is how much developers are doing more designing with AI than designers.

So there's this weird paradigm and that's where we want to stack AI first, right? There's this weird paradigm where the designers are still designing the apps that developers don't want to build anymore because they're vibe coding. It's just weird, right? Like their developers are doing the designing and the designers are still designing for the developers. of yesterday. What the hell is going on? Like, why is it that designers are so late to this? Well, you know, I think it has to do with the nature of the models themselves, right? That's a really good question. And I don't think designers are late necessarily because a lot of the folks kind of like in the UX space obviously want to experiment with this technology and are playing with it. But the nature of the models, like these are fundamentally, well, at the beginning for a lot of these things, they were sort of linguistic transformers, right? And the code is just another language. And so the ability to go from natural language to code.

seems to be a natural leap for these types of things. Then you have image generation and things like that that are going on. But the image generation that's happening that maybe designers are interested and want to play with doesn't really give you the outputs and materials that you need as a design professional to assemble things or to build and design things. You can go into... you know, mid-journey or one of these sort of image prompt tools and put in a prompt and you get like a thing back, right? But if I'm a designer, like I'm working in something like Figma or Illustrator or, you know, something in Creative Cloud. And I need like that, I don't need just a flat image. I need that image exploded into all its layers. Like I need the actual like human API with that file to be able to like manipulate it and use it in the context that I want to. So it's just an example. Yeah, it's that fine tuning.

Yeah, and I got to do stuff. I got to do stuff that maybe the AI doesn't know how to do. I need to deconstruct some things or I want to tweak some things. And just doing that through a prompt continuously isn't necessarily the most efficient way to do that. So I was actually using... Yeah, I'd be like an architect designing a building through natural language. Yeah, you could start there. Make the door shorter. Yeah, you could start there. Why not just grab a scale tool and drop down? No, that's way too little. So I went through this process actually yesterday. I had a...

engineer asked me to do something for them. And I, you know, it was like a little bit of a side task. So I'm like, I wonder if AI can help me out with this. So I was, I jumped into Adobe Illustrator, like the latest version, and they've got their Firefly, like vector drawing beta, you know, generative UI sort of, or generative experience in there. It's all data. And so I told it, I gave it a pretty detailed prompt about what I wanted. And I got back a pretty good vector rendering of this icon that I was thinking about, kind of based on my description. And it did some things that I wasn't really all that excited about, but some other ideas were good. But then when I went to inspect it more closely,

Like, you know, it was this very sort of geometric line art type of thing with circles around it and, you know, basic geometry. And I went and like sort of inspected the objects and the objects were all kind of weird. Like circles weren't perfect circles. The way it was organized was sort of weird. Some of the paths were intersecting in strange ways. And it wasn't, I couldn't use the file. Like it wasn't useful to me, but it was useful enough in terms of what it inspired. It's like, okay, I can take this idea that it created. And I went and actually.

hand reproduced about half of what the AI did. And I did it 10 minutes, but the AI wasn't able to generate that thing for me at a production level quality, right? So I just- It's like cardotide, yeah, in a sense. Yeah, it was almost like, it was like maybe the sketch stage for me. Yeah, that's a good point, yeah. Verbal idea in my head. I told it to somebody else and they sketched something else and I'm taking it now, rendering it out. But I had to go in and like precisely, you know, assemble my vector art, you know, in order for it to do the things that I wanted it to do. But it was weird. I wasn't sure what the AI was going to do, but then seeing the actual result, it's kind of really interesting.

yeah i was that's when coding i was realizing like so if it's if it's not vibe coding what is it like if i i gotta give somebody something better and it's more like peer repair programming that's what it feels more like um and i think there's probably a similar term to just hair designing yeah co-design you know it's yeah it's like a co-design concept that's happening and you're riffing off each other um and that that is really cool and really fun um and that's interesting that's an exciting Yeah. And it's interesting because like that, where the technology is at today from a design perspective is it can, it probably can be an effective collaborator with you. And I think my experience yesterday was like an example of that. My expectations were a little bit higher. Right. And it kind of came in a little bit lower, which seems to be the case for like a lot of this AI stuff. Right. But what's really, I think going to be more interesting for designers on like maybe the UX side is, you know, if you start, if you take the visual design tasks away, like some of the more UX oriented tasks around like.

how are you going to card sort an IA, right? Or organize a domain of knowledge into a sort of coherent structure. Like I'm more excited about having an AI help me with those tasks because those are the... You know, it requires a lot of analysis and processing and, you know, sort of like in many cases can be sort of like a multi-dimensional problem. And you need to explore like maybe many variations very quickly and then test those variations. Like I think that's where AI and UX, the intersection starts to get interesting in terms of the tooling, right? Because then you start to accelerate all these weird things that are very tedious for UX designers or UX researchers to do. And that's the thing I've been noticing a lot is...

The tedium, like, so to abstract this out, when you talk about AI agents, it's, you know, if you really want to be sophisticated about it, you've got to think in hierarchies. You've got to understand that task hierarchy. So if this is intelligent agents getting tasks done, then you have to understand that tasks come in hierarchies, right? There's, and we talked a little bit about this, Josh and I, you know, it's, I want to, you know, plan a vacation. Well, then there's the flight. That's.

one level below the vacation then there's how you're gonna get to the airport that's a level below then there's you know packing your bags um and what time you're gonna leave and the things that you need to do before the trip things you need to do after the trip like all of these are like branches and hierarchies um and in some cases ontologies around tasks and we're so used to in the UX world being like lower on the hierarchy. So, you know, in the day we're designing FedEx's package shipping, right? Well, we're way down, like ship a package is so far down the hierarchy. And that hierarchy is like, well, what are you shipping? An organ. Oh, well, okay. The UX designer, that's like a checkbox. Is it an organ? I remember doing this with United Airlines where we were like starting to plan shipping and realized like,

Well, shipping organs is actually one of the things that they ship on airplanes commonly. So getting a heart or a lung from point A to point B is a thing. And you think like, we're designing so low in the hierarchy that now we can start designing the full experience of shipping an organ. We don't have to do just shipping anything. Now we can...

broaden the scope of how we think about design at a higher level and start building experiences that are more suited to broader workflows and tailor those workflows to shipping an organ versus shipping something. And I think that's what that represents is a lot more work for designers because there's a lot more variations of things when you start sequencing these smaller components into and orchestrating them into higher level workflows. It's just like a mathematical algorithm. There's more numbers. which means more variations and more variability. And so a lot more experiences to design. But what does the world want? You know, the top 10 experiences for shipping an organ that works, right? With proven capabilities, not some human who figures it out on the fly.

So anyway, whether you're planning a vacation, this is all higher level design stuff that's usually out of our scope and now is in our scope. Yeah, I think you hit on an interesting point, Rob, when you started talking about the hierarchy of tasks, right? So this has been my experience trying to create like my own workflows with AIs, right, for various reasons. One example of this is I've been trying to recruit and search for candidates for some roles that I have open on my team. And, you know, obviously I've got a big recruiting team kind of looking at all that stuff, but I like to have my hands in recruiting to get the right candidates myself. But I wanted to do like a large scale search. And I was like, well, I can just jump into LinkedIn and start using their tools to sort of figure out things. But I'm like, well, is there a better way to do this? And so I've been using perplexity like all the time for like everything. I love using perplexity. I think it's just a very well done tool. And I pretty much have sworn off Google entirely for perplexity. So I was like, I can search for candidates, but why don't I just start asking perplexity? Hey, I've got, I'm looking for a candidate. Number one, this is the hierarchy. I'm looking for a candidate. I'm looking for a candidate with.

this educational background and these qualifications. I'm looking for a candidate with this education qualifications and professional background, right? So I'm starting to stack that up, those messages up. And then what I say is like, I'm not just interested in getting that information, right? I want to know like out of, you know, 10 of these people that you can find, Perplexity, the agent, like that you go out and find, can you rank them for me? And can you explain for each one of these candidates, provide me their LinkedIn profile and then explain like why you think they're the best and better than the person below them, right? And so I started stacking up like these sort of instruction sets in my prompt to sort of get the feedback.

You know, like the first initial tries that I had, you know, I didn't have that hierarchy in there. But as I started to sort of stack up the tasks and make the sort of decision processes clear for the model, like then my outputs that I got were the quality started to skyrocket. Right. And now like was the agent able to like, you know, fully able to dial in, you know, what I wanted? No, it made mistakes and did some weird stuff like they do. But, you know, there was probably out of, let's say, 10 people, probably two like bullseyes. Right. Which is actually a pretty good result when you think about like the recruiting process. Right. You know, like it can be really hard.

So, yeah, it was only a minute away from reaching out to them itself and setting up a schedule, right? Like, I mean, really, technically speaking, it was, you know, that's just 15% more. And what was really interesting, I did all this, I actually turned on pro mode in perplexity, in which you get to see the self-reasoning unfold. And I would just watch that as I was putting this stuff in. It was just following those instructions and talking to itself or talking between the multiple agents that drive it. And it was just, yeah, it was pretty amazing that it got that far with that level of quality for such a weird abstract concept. Yeah, so that's what...

That kind of brings me to two things I want to chat with you about. I want to focus on the first one, which is this AI-first design concept we were discussing, I think, several times. But as you're saying, there's this temporal concept of events. You're watching Chain of Thought. Now, what is that? It's really an event log, right? It's things happening. You could be inserted into that event log. And then I asked him this, and then he said this, and then we went through another. There's a...

There's a longer version of that, right? That seems a bunch of the hierarchy together and a bunch of little human in the loop components that would be involved in it doing more work. But one of the things I think a lot of people kind of lose sight of is the new paradigm, how to be AI first in thinking. And I was thinking about what you did at Amazon where it was migrating the... you know, the mainframe workloads over to AWS. And it was going from like years, or was it years to weeks, right? It was essentially like, you know, yeah, man years to basically weeks worth of effort. Yeah. And when you, as a technologist, when you really sit back and look at that, you think, well, that's one of the hardest migrations. That's one of the hardest things to do. Like pick the hardest thing, go for mainframes. Why do I say that? I think I can support that on the fact that if it was easy, we wouldn't have mainframes anymore. There'd be nothing to migrate, right? Clearly it's hard. And here we are, like it's happening.

And as people think about like these SaaS programs and our old software, and they think, oh, well, you know, SaaS can be around for a while. It takes a while to move them off. I think people are misunderstanding or misseeing the fact that if Amazon can hand you a script to migrate from your mainframe, they can give you a script to migrate from Salesforce or ServiceNow or like piece of cake, right? That's like nothing. And so what we're probably going to see, and who knows if they're even planning it, but I think it's going to happen. We're going to see that cloud providers are just going to start handing you migrations.

from your SaaS software to their core services. Essentially with AWS, it's moved from some paid for program to open source and core AWS services. And then someone might say, well, the software is better when you pay for it. And the answer is, until now we have vibe coding. The open source software is going to be significantly better than it is now and good enough. And so we're going to...

we're going to see just this massive, massive shift of shifting off of those expensive business models of SaaS software to just press a button, migrate off of Salesforce and pay for the electricity basically that you're using. Yeah, we were kind of likening it to what Klarn is doing, but on steroids, I suppose. Rob, I think you're spot on. I mean, AWS sees the future of all this, right? Obviously, it kind of hits their flywheel pretty nicely, right? You can automate the migration of the service, then they're burning compute and whatever else, storage up in their cloud, and the money's rolling in from that perspective. But the interesting thing is, just kind of thinking about what you just said, if AWS can do it...

probably anybody else could do it too. So is there a pending sort of battle that's going to happen? It's almost just like it's PG&E. At some point, you're just going to get it from your electric company because at the core, you're just going to pay for electricity. I mean, look at OpenAI. They're investing in electricity. Well, fundamentally, everything just comes down to compute, right? You got to flip a bit, and flip a bit takes an electron. Yeah, we're going to have Grok on, and they're working with the Middle East, and it's this interesting concept. They're going to get over a billion dollars in chips that they're going to buy. And essentially, it's the idea of, like, we don't need to ship our electricity from the Middle East to the U.S.

we can sell it to them in the form of inference, an inference, right? Yeah. It's like inference is the new wires and electricity grid. And they're now selling their, so they're just going to invest in massive data centers that does steep inference. And we're going to get the electricity through this text-based, you know, essence optical wires. Right. Like what a crazy idea that the new electric grid is data. Well, I also think it's really interesting that...

Even in the industry, when you start talking about compute and inference and all that type of stuff, they don't refer to it in terms of like petaflops or exaflops and compute power anymore. Like it doesn't matter. It's just, just megawatts. And it just blows my mind. You know, I hear you hear all the time, like, oh, we're not going to invest in a certain scale of compute. We're investing in a certain scale of just energy. Yeah. Right.

Because the nature of the compute can change and can be modified, but the energy consumption is the number one limiting factor at this point, which has, I don't know, I'm scared by that, right? Like if you're building data centers the size of Manhattan, that's a ton of energy. But if you've got to produce energy, there's this other thing that has to be produced as well, right? And that's carbon, which is not optimal, right? So it's kind of a strange time. It is. I think...

I think it's tough because the question is, is it a zero-sum game on energy when it comes to compute? In other words, are we just going to do less compute locally? Is it going to move or is it just more compute overall? There's a lot of compute going on with inefficient software. There's a whole efficiency component to... I think it's at all the levels in the stack. You have how these...

how these data centers are configured, right? How you're transmitting the power to the data center, like just the physics of the data center alone. And then all the, you know, how are you storing your data? Like, how is it being piped back and forth? Everybody is looking for all these very incremental optimizations because those incremental optimizations scale up to massive impact. You know, when we're talking about the scale required to do some of these things, it's unbelievable. Imagine the cost savings of just moving your mainframe compute over to AWS and the energy savings. Yeah. Right? Like, what does it cost to run those mainframes? Like, geez. And what...

So it's hard. There's like an efficiency gain. There's always a loss of moving energy from one side of the country to the other side of the country. So if most of our compute is now an inference, and it can be generated in the Middle East or at the source of energy, I mean, it's almost like we're going to put these little tiny data centers right over a... oil rig at some point. You're just going to stick a straw out of a server into the ground and be like, okay, suck the oil in closer to the source. Turn energy into compute directly. I just think this concept, I mean, it's AWS releasing Q and its ability to migrate from mainframes. It's just massive. The implications are huge and I think people are just missing it. It's just not being talked about.

But what's equally as interesting to me is how you designed it and what this looked like. Because I feel like it's really an example of what interfaces are going to look like in the future. This concept of a temporal flow of events around a project or a workflow. And then people jumping in and jumping out to make decisions along the way. And at some point, I'm imagining as you guys were going through this, the speed of which you could migrate was no longer about the system's ability to rewrite code, but just the people's speed in responding to these decisions. Yeah, yeah. I mean, yeah, let's dive into that. So I think first and foremost, as we started getting into understanding what that transformation product was going to look like, I was working very closely with my UX architect, Ashani, on this. And one of the things that we were talking about is like, okay, this is really an experience.

Like this is truly an automated experience, right? And so there's a lot of different processes happening and a lot of things happening concurrently. And the way that we were thinking about the interface and how to approach it is, well, what is the substance of this experience? And what is the sort of key information that the user needs to understand? And so it's very different from a more traditional kind of user experience problem where you have to present this functionality and you have different calls to action. You're trying to move people through an architecture of UX where they understand the relationship between concepts and the actions that they want to take and how that's going to create like specific outcomes that their tool is going to help them produce, right? When we were thinking about this tool, it's like, well, it's kind of...

doing all the things for you, and it's doing many things at once. And what are the sorts of things that, like, what are you actually doing? What is the theory of human in the loop model here? Like, are you, do you have to understand all the details? Do you have to understand sort of critical messages? Like, what's the situation? And really what we ended up doing, Rob, was just kind of leaning back on classic UX heuristics, right? Like, okay, we have this system. This system is doing something.

the most important thing that it can do, especially given the nature of this work, right? These are large scale systems moving from point A to point B. There's a code transformation happening at the same time. It's, you know, very complex and even, it was so complex that even like the sysadmins maybe on the client side, right? Don't even know the full nature of their code and how it operates, right? That's kind of how complex it is. So, you know, really what it came down to is like, it's all about status communication, which is like a very basic old sort of Jacob Nielsen type of heuristic, you know, communicating system status to the user and kind of keeping them up to speed on all that stuff. And so,

After we were kind of experimenting with these different sort of visualizations and thinking about, okay, how do we show progress or how do we communicate progress? It really just came down to, well, it's just, we're talking about a hierarchy of tasks or a log of events, right? Where we understand all the different sort of critical events that we need to know and understand and just flagging the things that we know will be important to the user if some threshold isn't exceeded or if some threshold is exceeded, right? And just kind of building the experience around all of that. And it's really just all about...

trying to drive as much communication to the user as possible, just so that they're aware of what's happening. And it's really not about control, right? It's very feedback. It's not like your experience using perplexity too, where you want to crack it open and see what it's thinking. So you can know that it's making the right decisions, that it's following the hierarchy that you set out. So there's another one with that too, right? Where you're kind of monitoring it. It's just very feedback. We go to this push versus this pull in UX, where we're pushing what's happening next. We're taking the user through what they're doing versus asking them to drive. We're driving and just letting them observe and then pulling them in when needed.

And it really shifts from a dashboard to a feed. And now you just talk about, okay, that's if you think single user. But in the concept of the mainframe migration, you've got now multiple users on a single workflow or threaded workflow that rolls up to a single hierarchical workflow. And then you're pulling in the right person at the right time. And so you have this thread of feeds and this decision of who needs to decide what when. Yeah. Which is just, it's not new. I mean, you know.

Most humans on Earth right now spend a majority of their compute time in a feed of some kind. But as UXers, the dashboard seems to be where we're focused versus the feed. What's interesting to me is thinking and looking at innovations around UX innovations around the feed versus the dashboard. Yeah, that's a really interesting idea, Rob. And I think, yeah, it's like, what is a feed? I think the execution log model is just this idea of a list, right? But this thing could probably take lots of different shapes and forms. And what I think is kind of the way that you're talking about this is it suggests that there's maybe this future where our experiences are just...

a series of status communications, either in serial status communications or parallel status communications, in which we have to make sort of critical decisions that something else goes off and does, right? We all become these sort of like project managers in a way, right? It's a, you know, if we think about this sort of agentic future that we're sort of sprinting towards very quickly. If you think about the current context of like... how organizations operate today. And I work, I've been working in, you know, sort of large organizations that produce like a lot of software. And there's always a series of decisions that have to be made. And one way to think about this is what is the, like the frequency or velocity of the decision-making process? Maybe velocity is the best term for this. So when you have a bunch of humans working together in a large organization that's socially connected in different ways and different levels, the decision-making process can be somewhat slow, right? And you have organizations like Amazon, which have like an optimized culture to facilitate and speed up decisions because they see the speed of decision-making as being directly correlating to their revenue, right? So.

So does this sort of future, this agentic future, suggest that we can now accelerate the decision-making velocity across an organization in a really significant way? If I have multiple feeds coming in, that's definitely accelerating the process. And if I'm highlighting all the critical decisions for each individual, let's say there's this orchestrator in the background kind of working through all those things, can it be prompting the right people at the right time? Does that get you like two to three orders of magnitude increase in the velocity of decision-making within your organization? And that's just based on, okay, if we're moving more towards this model of...

okay, I'm always being communicated status in very dynamic ways about all the different things that are going on, then I can make, and I can go deep on those sort of inputs that I'm getting from the system, then I can make those decisions very, very quickly and very, very accurately. Yeah, I think it's, you know, as humans, we, you know, spaces, one way to call it, projects, you know, I like spaces as a concept because it's more general, but we all have these things, like health, learning, work, our house, our relationships. our kids, our dogs. These are our projects, right? And everything we do every day rolls out to one of these spaces. We brush our teeth, it rolls up to health. We walk the dog, it rolls up to our pets. We sweep the floors, it rolls up to our house. We take out the trash, like everything rolls up to these spaces. And usually, if you're like me, in almost every single one of those, you have an X objective. And I'm always tracking them as threads.

these spaces. You know, I'm trying to make progress in each one of those. And there are events happening in each one of those that I'm trying to sort of track in my mind. And there's always a longer list of ones I'm not getting to and objectives I'm not moving forward than there are the ones I am moving forward. And I sort of imagine that with AI now and an army of AI agents, I can make progress on all of these things now because they can be off making progress. And all I got to do is make decisions between them. So I'm jumping between these spaces, which are essentially just things in my life I'm moving forward that matter.

You can look at the temporal logs and see what's happening in space. And which one needs my attention. And we end up, as designers, being more like cognitive load misers, right? Just taking a 2% cognitive load off of a decision will add up over the course of a month to this many more decisions being made. And we're almost like, I was thinking about Barack Obama's gray suit, blue suit, like wake up in the morning and choosing what to wear. Oh man, you just wasted a decision. Like that's, you know, that's, that's like now as designers, we're like, how do we help people make decisions with the lowest cognitive load possible, which is a lot about giving them like the data.

that they need. You know, when you have to make decisions with less data, that's more load, right? It's filling in the gaps yourself, more prediction that you have to add to the equation. The more data that the system can give you, the faster decisions can be made. And the more certainty you get too. Right. Yeah. So it's almost this like, just, I was, in my mind, it's almost this like swiping left and right all day long, like just trying to get decisions down to a simple yes, no. Yeah. And like just sitting there and I was, and I can't help but keep going back to what...

to like where you guys were trying to get with the mainframe migration is like the closer you could get to that, just people swiping, yes, no, yes, no, yes, no, the faster that migration happens. It was always a question, too. How much do you expose to users at any given point in time? How much choice? How much choice do you give them? And this is like Ashani Mittal, who's the UX architect and main designer on this work. This is the things that she was balancing and really going hard from a UX research perspective on. It's like, how much do you need to know? And how much do you want to pollute that execution log with? If you're a human being, then basic kind of usability principles come back in. Or information design principles come back in. How do you manage this?

execution log and how do you present it in such a way that people can make those decisions quickly and can maybe just do a swipe left or swipe right to make a choice, right? Yeah. And it's complicated because every user has a different optimal level of information that they need on a project like that, especially. Exactly. Exactly. So I want to shift from, because making a lot of decisions all day long doesn't sound very fun.

So let's shift to like probably the other side to that, which is fun, which is storytelling, right? You know, if it's temporal in essence, you're really telling a story to them of the migration, right? And so the other side of this is how do you make migrating, which sounds like a super big challenge. How do you make migrating from a mainframe to AWS a fun story, right? And it kind of ends at like, if you want to make this fun, it's almost like choose your own adventure. That's the...

Like that's what the design becomes. Like this is like a story and now you get to choose your own adventure and you can decide which way this goes. So in essence... if you really lean into the feed concept, you really understand that the feed's just trying to tell a story and a coherent story. And you're trying to figure out the story points to surface and the right places to give people agency so that they are more engaged and feel connected to the story. Yeah. You're making choices at these story breaks that move it forward in the right direction. It's really interesting, Robert, you're bringing up this idea of like narrative as a way to sort of, you know, first of all, narrative is probably one of the most human things in the world, right? Stories are so...

so critical to the human experience like throughout from day one right it's so critical it's how we got here this is why we're talking about computers right and ai you couldn't do that without having stories right this is how everything begins um but what's really funny is yesterday i had a really fascinating conversation with a former colleague of mine jay castile he's a brilliant designer at aws he's working in the ai space too great guy and he actually went recently went on a family trip to uh disney i don't remember if it's disney world or disneyland i think it was disney world and he was sort of like you know i don't know if i really you know want to go on this trip etc and i was like jay like It's such a UX extravaganza when you go to Disney, like, because everything's so dialed, right? And it's like, you know, as a design professional, as you go there, like Disney, like, everybody's a creative or a designer there. And there's so many brilliant artists and brilliant people that are working there. And it's an evolution. Like, yeah, it's evolved design. Like, it's evolved, right? Yeah, it's college. It's dialed. Like, you're not seeing something on first turn. You're seeing something.

Highly evolved, yeah. Highly evolved. And so we were talking about, he was sort of reluctant, but I was talking to him yesterday. He was like, Tim, you're right. There's so much to be learned just walking around this place. And one of the things that we were talking about, Rob, was the fact that he was getting on a new ride. It's one of these new Star Wars rides. And to be honest with you, I don't know all the details of which one it was. But basically what Disney did is, what's the worst part about going to Disney, right? Standing in line and waiting with your kids or waiting in the heat, sweating, you're thirsty, whatever, kids are crying. It's this whole trial. And they've created all these different tools to sort of bypass the line. But in this new sort of mindset that Jay was talking about, they turned the line into the ride.

right like you're in line but you're actually being taken prisoner by the emperor and you're getting pushed into here and you're getting moved over there and it was actually like you know he was like the experience was amazing like getting in line and kind of going through that and then all of a sudden you're in the story right and it's that like you know these these imagineers and designers at disney are thinking about how can we get these these people engaged in the story before the story even starts or where does the story start the story doesn't start in the ride the story starts in the line and i think about the line as being like a metaphor for like the execution log right it's like here's this thing i've got to watch but okay how and where do i want to engage the user and how do i make it interesting how do i make them feel like they're a part of this yes right and where does that person have the most value yeah and this kind of gets to like storytelling being in the ux not rounded corners and saying that like um we talked about this uh imbuing intelligence onto everything in our lives so having this like digital twin concept and having like our chair, for example, be able to talk to us? And I was thinking about imagination from that standpoint. Like you're like, most people would say, why would I want to talk to my chair? And the answer is, because the only reason you would say that is because you don't have much imagination. Because that chair could tell a story. It could tell a story of the tree that it came from. It could tell a story about the designer that designed it. It could tell a story about the craftsman that built it. It could tell a story about other people that have bought it. Chairs like it. It might be able to tell you a story about your herniated disc in L4, L5. Yeah, yeah. And just being able to like,

We have this like digital twin of our lives and our homes and the things in our homes and that those things talking to each other. And the idea that your chair talks to you, most people think of as, you know, oh, why would we add a chip to your chair? Well, you don't need to add a chip to your chair. Yeah, yeah. The idea is that your chair is in a digital twin. You're talking to it as a, you know, as a sort of metaphor, right? In a sense of itself. But it still inherently has all this logic. And it comes a point where almost everything is going to be beautiful.

looking. It's not going to be enough to just be aesthetically pleasing. We're going to need a story behind everything and that UXers now become the storytellers. If you don't put a story behind the chair, there's always going to be a hundred chairs that are going to look good that we can buy. We're going to want the chair that looks good and has the story behind it that we want to tell and we want to be told. A lot of kilowatts. Indeed, Josh. This is really a...

this is really about rob what you're describing as a transformation of material culture in the world yeah right and you know i i've i've been a part of the industrial design world in my life right and i know people who design and make chairs and you're right like first of all everything is designed right if it's manufactured it's designed and and there's a lot of beautiful things out there and plenty of things that obviously are not that great but like so much stuff is amazing right now and um we have the technology i mean these We have vision models, we have complex LLMs that can recognize objects in a space and can understand things about them and can synthesize information and contextualize information for you. So yeah, why can't these things have stories and ways to interact with them that they don't have today, right? And yeah, is there some sort of virtual representation of them in this, let's say, the sort of AI that runs your home? Maybe that's kind of like the center of all that stuff, or maybe it's the AI that you work with every day is now kind of observing your environment and can imbue that intelligence. But then that opens up so many different possibilities, right?

I don't know. I'm looking at my table now. I love my little oak table that I use here in my dining room. I don't know what types of things I might ask it or what it might be able to tell me. But, you know, having some sort of like there's different scenarios, right, that probably make a lot of sense. There's these sort of unmet needs or undefined sort of jobs to be done and use cases that I think are right for the picking. And I wonder, like, you know, when we were having our previous conversation about, you know, like.

the nature of these agentic experiences and the communication status, like a lot of these discussions are around these sort of virtual environments, computational environments, and sort of liminal spaces in which we're thinking about are the different projects and things or spaces that we work in. But what about the intersection of physical space, right? And that's what you're talking about. And so there's this imbuing of intelligence into the sort of inanimate object, right? That's one piece. But then we will have, like there is the intersection of kind of the agentic experience and robotics as well. Like that's coming. We know that that's coming. And what does that mean? Does the intelligence get imbued into the robotic experience or is it more...

of an ambient type of thing. What are the future possibilities of all of that? I think robotics is just very anthropomorphized at the moment. And it makes sense. The form factor of humans to take over human jobs makes sense because human jobs are designed for the human form. But I still think it's scomorphic in a sense. It's going to be a passing thing because...

Using a robot's senses for your AI agent doesn't make sense because it's singular. When it's in one room, it can't be in another room, right? And we don't want an AI that's in one room. We want the AI that's omnipresent. That's everywhere. If you think about sensory and sensory information, we want it to be aware of everything. So limiting the AI by putting it in the physical space is just more anthropomorphic than it is practical. Yeah.

The question is, who are you talking to? This concept of a chair being smart, we're riffing on this the other day, which is the plant talking to the chair and then mentioning to the cleaning lady, like, hey, can you move the chair away from the window? The plant asking for the chair to be moved away from the window because it needs more light. I think that was something you brought up in that idea. that these things are talking to each other and they're talking to people and they're taking care of themselves. And, you know, it makes a lot of sense that, that, that, that these things happen, but where are this, where's the sensory data? Well, it just needs to be a camera in the room. Right. And you don't need every single one of these things to have sensory inputs. You just need, you know, you need a sensual sensory. Yes. Right. Yeah. Just like this, like my sensual sensory device right here.

It's interesting. You're bringing about this idea of the anthropomorphic robots or androids. You see Tesla is really running fast after that direction. Boston Dynamics has some of these things as well. To me, I look at that and it's almost like that's a bridge. You need a human-shaped robot to fit a human-shaped sea. It's like, well, what if you just take away the sea? You don't need that thing. What is a human good for? Humans are primates. We are really good in trees and grabbing things. We're optimized for living on the African savanna, more or less, or wandering the frozen wastelands of Northern Europe or whatever the situation is.

We did really well with that. Are humanoid robots the answer to the robotic question? And the answer to that is obviously no, because the most ubiquitous robots, let's say the vacuum sweepers and the floor mopper things, aren't little people, right? They're these little disks or whatever that kind of move around and optimize for their space. But I'm thinking back to, there was a really famous roboticist. That's a funny thing you just said. Why aren't they little people? Why are they making it ours? I never thought I wanted a humanoid robot in my house, but now that you said little, would I want a little totemian, maybe?

It's a little army of little robots. Yeah. Little tiny men. Cartwheeling around. But yeah, there's this roboticist, I think from Carnegie Mellon, his name's Hans Moravec, or Moravec, I'm not sure. But he had a sort of vision of, he's a very visionary sort of robotics guy. And he's written a number of really interesting books. But he had this vision for robots that weren't humanoid. They were basically kind of like fractal base. Like there would be a central body and then an appendage. And then on that appendage would be another appendage. And on that appendage would be yet another appendage. And those appendages would go all the way down to like the nanoscale where they could manipulate matter.

Right. And so it's basically you could just think of them as fuzzy balls. Right. That could move around in space and manipulate matter to do whatever you want. Like that obviously is very far out in the future. But that's that's kind of like a logical end state. Right. You have something that needs to be able to move through space, manipulate matter, do whatever. There's there's kind of an answer to that. And it's if you think about the.

tiny human versus, you know, disc shaped robot that moves across your floor. Well, the robot's disc shaped and flat because it has to move under things and has to adapt to its environment. So you want to create these machines and machine intelligence that actively adapt to whatever context they're working in. And humans are highly adaptable, but we're biological and our adaptation is driven by our biology, right? Like I'm, you know, Rob, you know, I'm an athlete. I like to do things and my body adapts as I train it to the different physical stresses that I put it in. Like when you bake a robot into a certain form, it doesn't have those biological processes or nano level machinery to adapt itself.

It's a fixed form and as such is already obsolete by the time it rolls off the assembly line. Like any automobile is. You're so right. We used to envision them as what we call mechanicals where there's these things that have all these attachments and all these attachments are designed to do specific tasks because the hand isn't necessarily the best interface for an attachment. It's great for, like you said, trees and swinging on trees but not necessarily for hammering something.

And then you imagine like a robot with all these attachments, but then you realize, well, what's the robot then? Like, why isn't it just a bunch of attachments running around? You don't need the robot part. And so, yeah, then all of a sudden you've got attachments running around your house, which are little tiny robots doing little things. And the question then comes back to our AI first. What does the interface look like?

for you to interact with these things well it's natural language first of all but it's also this like feed right it's just all these attachments are gonna have questions that you're gonna be asked and decisions for you to make because at the end of the day they're designed to make your life better and provide you with happiness and happiness equals agency so you're going to want to these decisions are required for you to be happy um whether they're necessary or not so here you are like being bombarded by you know decisions with all of these little agents that could be physical could be virtual running around making choices I think we can make a parallel. You're just swiping left and right? Yes. They're just feeding you placebo swipes eventually because your inputs don't really need any placebo swipes. That's like the next UX thing, the placebo swipe. It gives you the sense of agency. My wife does that all the time. I think I get this placebo swipe often. I do too. It feels like a choice, but you know it's not a choice. Josh, you'd be wise to trademark that term ASAP. But Rob, I think this relates to some of our previous conversations that we've had where we've talked about...

you know, the AI of the agentic system is like almost no UI, right? And the parallel that I see is like, we've been talking about how UIs can be drawn dynamically now, right? In real time. And so that theory like maps directly to the robotics model as well. Whereas like these are attachments, you know, we've got these agentic user interfaces, in which case, in many cases might have self-assembling UIs or drawing UIs dynamically as you need them. And then this maps to the physical space where you've got a robotic agent, right? A physical agent that self-assembles, you know, whatever it needs.

given the context of its use or how you're directing it, right? So the robot itself isn't a thing. It's a collection of things that self-assembles based on the tasking context, et cetera, right? And so that's a direct parallel virtual to the physical environment. Yeah, exactly. I think when people think about agents, they tend to think about a single thing and it's just not, it's going to be a ton of these little nano agents.

all with very specific tasks, all orchestrated together, talking to each other to get things done. Because nano agents make better decisions than these macro decision-making agents. This idea of an agent making a hundred different, you know, selecting from a thousand different choices is overfit. Well, yeah, it's overfit. And big models, right? Big models are highly non-deterministic. So, you know, you can get a different answer every time you ask, right? And so that's why you need the specialized agents. This is what everybody's doing now is you have this sort of specialized tooling.

and agents and, you know, sort of rag situations where you're able to make a more sort of more constrained sort of problem space for the system to operate in, right? So I think you're spot on. There was a good story that fits right into this conversation from, you guys remember we were doing some research and we ended up talking with some people at Lego for the show and they had an interesting moment of realizing the anthropomorphization problem where they were trying to use AI and robotics to assist with picking the right Lego pieces. It's an interesting and sort of obvious problem that you're probably not thinking of unless you're at Lego in operations and logistics there. They have to select the Lego pieces that go with the kit that they're packaging and shipping out. And that's actually a really interesting problem. There's like tons of different Lego pieces, different sizes, shapes, colors, blah, blah, blah. And they have to find the exact one and then put it in the kit that's going on the shelf. And so they set down the path initially envisioning like robotic arms that were like human arms and hands. And then at some point they like stops the record. Like, what the hell are we doing? And I completely rethought, like, why are we doing anything human form? Like, we just need to problem solve. We're like constraining ourselves to human form. And they ended up going with like air pressure devices.

Like basically just like super specific, like high pressure air that would like triangulate on a piece and like relocate it. But it was a really interesting, very, very powerful example of... how much better the solution was the second that they like completely dropped, realized that they were constraining themselves way too much by human forms. Yeah, it's just that anthrocentric bias, right? You can't help but have it. If you live, you know, Elias, if you live in a human body, you can't help but have that kind of bias, right? It's almost, it's very difficult to shed that, right? It's gotta be like a hand that reaches out. And everyone's obsessed with replacing jobs, you know, instead of just inventing new ones. Like, why do we have to replace jobs? Let's just invent new ones. Screw replacing jobs. That's a waste of time. Let's just replace new jobs. Let's get new jobs, not replace jobs. Let's just invent new jobs and...

And then let our old jobs become inventing new jobs. I remember what I was going to say. It's totally off topic though, but I was thinking about how the right way sometimes is the routine way, not the right way. Especially in UX and humans, like a lot of times we just want, we want to do what's familiar, not necessarily reinvent the right way. So imagine like every day you go to tie your shoes and each time you're trying to optimize for whatever, the shoe that you're putting on, the laces, the...

humidity in the air. I don't know what you're going to do that day. And you're like, and you have like some sort of card on the wall of a hundred different knots, you know, nobody, we're just, oh my God. So, so AI optimizing what we do isn't necessarily about always doing things the right way. Sometimes the right way is just the routine way for us. But interesting enough for AI, that's not true. It doesn't mean it has no real cognitive load in that regard. And when, when AI is doing, like you said, you know, it's just going to conjure up a UI.

How beautiful it is doesn't matter because we're not going to see it long enough, hopefully, if it's designed well. And it's this idea that for us we need routine, but it doesn't. It can always do the optimal thing at any given time, which is an interesting way to think about when we anthropomorphize the job that it does, we're going to box it into routines. We're going to ask it to do things in the same way, the way we do, without even thinking that it should just decide the best way each time, which essentially that's... So what you're telling me, Rob, is that we shouldn't be anthropomorphizing the machines. The machines should be anthropomorphizing decisions for us. Right. Because we're the ones who need it. We're the ones who need the anthropomorphization, not the machines. This was great, guys. All right. Well, thanks for hanging out with us again here on Invisible Machines. Don't forget to like and subscribe wherever you get your podcasts. Always remember that you can watch this podcast, too, on our Invisible Machines YouTube channel. Head over to uxmag.com. We have a podcast page there where you can explore our back catalog.

and you can look at episodes centered around popular topics. Many, many thanks to our incredible production team here in the U.S. and in Ukraine, Elias Parker, Michael Lipinov, Allison Harshberger, Ola Slessarenko, Vera Pakhodkov, and Dorena Moskovchuk. Till next week, Iwaki.

Frequently asked questions

How long is this episode of Invisible Machines podcast by UX Magazine?

This episode is 1 hour and 12 minutes long.

When was this Invisible Machines podcast by UX Magazine episode published?

This episode was published on April 24, 2025.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this Invisible Machines podcast by UX Magazine episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.