Subscribe
Share
Share
Embed
Conversations with designers, founders, and builders behind some of the best work
00:00:00:00 - 00:00:09:17
Unknown
I think a lot of developers are mourning right now, the flow state that you get when you're like a master of your tools and you know the syntax really well
00:00:09:17 - 00:00:13:20
Unknown
you feel like your expertise is, like, driving it forward,
00:00:13:20 - 00:00:21:21
Unknown
And, like, I think it's just really hard right now for developers where we're trying to find the new rhythm, where we're like, I've lost my flow state because,
00:00:21:21 - 00:00:48:21
Unknown
Amelia, Wanting Berger started as a front end developer, spent eight years building dashboards and data viz for startups. She's the kind of person who does not care what the material is code pixels, Figma, whatever. She just wants to wrestle with an idea and make it become real. She's been at GitHub. She authors these custom crafted blog posts that wrestle with philosophical ideas of human computer interaction.
00:00:48:21 - 00:01:11:04
Unknown
And now she's building tent, a desktop app that's trying to answer the question if code syntax is disappearing, what does the developer's workspace start to look like? And she thinks in layers, moves up and down the abstraction stack the way most of us switch between browser tabs. And she's building for a moment that most of us haven't been able to articulate.
00:01:11:04 - 00:01:41:20
Unknown
The old flow state is disappearing, but the new one hasn't arrived. Not yet. We talk about the eras of AI coding tools. Why she thinks the spec is becoming the new source of truth, and what she means when she says the work is shifting from implementation to intention. This is state of play. Let's get into it.
00:01:41:22 - 00:02:00:06
Unknown
The thing that I'm really trying to do right now, whether it's with our podcast or any of the other media that we're putting out, is just explore what I think is one of the wildest times in humanity, frankly. And there's a lot of people in a way that's not like hype ridden is. There's a lot of like anxiety out there.
00:02:00:08 - 00:02:16:23
Unknown
You know, there's a lot of people who aren't sure what's going to happen. I'm concerned about the short term impact on the labor markets. I'm hopeful that long term it's going to be worth it. And the way that we're kind of just looking into that is talking to people who are building in this and who are thinking about it.
00:02:16:23 - 00:02:38:02
Unknown
And it's clear through the stories that I saw on your websites and the designs that you're putting out. And obviously the places that you're working, that you are thinking very deeply about this. What's interesting is your career kind of feels like as I was digging into what you've been working on, it seems like you've moved up and down this ladder of of abstraction, right?
00:02:38:02 - 00:03:14:17
Unknown
Like from design to kind of absorbing engineering, to really trying to, like, pay attention to market through VC. Now working with intense or the product intense, does it feel like you're juggling a lot of things, or does it feel like it's all kind of coalescing into one big effort? Yeah. Yeah. My career is interesting. I started as, like a front end developer, and I worked for like, eight years as a front end developer at SAS Dashboard Analytics start ups, but, like, pretty small.
00:03:14:17 - 00:03:42:22
Unknown
So at one point, I was like, only developer, only designer. So doing like a lot of things like kind of like not super nice, right? Like data viz, design development. But like I kind of had this trend from like an engineer over to like recently I've had more and more jobs where the job title is designer. I feel like for me, writing code is just a means to build things.
00:03:43:00 - 00:04:02:23
Unknown
And like, I don't care what material it is that I'm building in. I just like to conceive of things and like, work hard and then like something comes out, like something pops out and it like is a thing that started in your head. And like, for me, that's like the main motivation for why I do what I do.
00:04:03:00 - 00:04:26:14
Unknown
And I think over time it's gone from like like the craft of the code and like, what does the code look like? And learning new syntax to like, more like I'm not like a super like pixely like visual designer. I'm more focused on like, what is the material that we're working with and how can we build things that change how we live our lives and like, what behavior can they afford?
00:04:26:15 - 00:04:56:16
Unknown
Code used to be like separation of concerns, right? There is like a QML files and CSS files and JavaScript files. Right. And it was like these were all separated because they were different languages. And one was the style, one was the structure, one was the behavior. And as things got more and more dynamic, it was hard, right where you're like, I'm working on a button, it's buttons, CSS button, JS and button that HTML and and said, like the way I think about it is like it's a button and I want to work on the button.
00:04:56:18 - 00:05:23:05
Unknown
So like with modern JavaScript frameworks, they're more component based. So instead of like slicing vertically, we can slice horizontally, right? Based on like what it is that we're working on. What is it that can we can build across like different mediums. And I think like with AI, right, there's more and more designers who are writing code in service of building their designs or like building a product.
00:05:23:06 - 00:05:42:03
Unknown
And it feels like a similar like we were slicing vertically and now we're slicing horizontally. It seems like, you know, a lot of this, what's happened with kind of the invention of AI is we have these layers. When you think of these like abstractions, you start with kind of the math. And that's very specific to, researchers and people who are really building that out.
00:05:42:03 - 00:05:59:22
Unknown
And then on top of that, we you could start to say maybe the next layer and you might put a couple things in between here. But then we had chat, GPT three come out, and that really started to open this up to where we're going towards critical mass. And outside of that, then you like go a layer above that.
00:05:59:22 - 00:06:16:17
Unknown
Maybe you'd call it agents. How do you think about the abstraction layers from the math up, and where do you think a lot of the industry is going to start? Where do you think most of the industry is going to capture themselves in those layers? So there's there's like a few different layers, right? There's like level of abstraction.
00:06:16:17 - 00:06:38:23
Unknown
So with programing language just right, it was like punch cards and like you were like and then like assembly. Right. And you're like dealing with like the like raw like computer code. Right. And then like it's like very hard and like it takes a long time to build anything. But you're like very in touch with, like the material itself.
00:06:39:02 - 00:07:12:07
Unknown
And then you go up another level of abstraction. It's like higher level and higher and higher level programing languages where like, like with typescripts, you're or like with one of these, like modern JavaScript frameworks, like a little bit of code goes a long way because like, like compiles out to this like bigger thing. And it seems like the next level of that is like, now that we can talk to agents that can compile it into like the, like raw code, we get this, like really high leverage so that we're like less in touch with the materials themselves.
00:07:12:08 - 00:07:35:22
Unknown
Like, I could not tell you like how a lot of the things I build are actually like running from like a hard and but I get to do more things right and like I'm building more and I'm building more and more complex things and like, it's at a cost, right. But like, we like continuously like as a society and as we build tooling, we're going up and up in, in the abstraction.
00:07:35:22 - 00:07:58:23
Unknown
And I think there's the other levels, which I think you're getting at, which is like where like we have agents and then we want like Asian orchestration libraries, and then we're like, okay, we have agent orchestration, but like, let's zoom out another one and like, let's orchestrate the orchestrators and then like, like there's like this never ending, like, wait, you just throw an agent at it and then like, you're up another level abstraction.
00:07:58:23 - 00:08:21:18
Unknown
And like seemingly to no end until, like, you get to the very purpose of why we do anything right. Like I would say, like agents aren't going to eat ice cream for me because there would be no point in that, right? Like the point of me eating ice cream is like, because it's delicious and I enjoy it and you can't automate that.
00:08:21:20 - 00:08:40:15
Unknown
I love the way that you frame that, because it does seem like if you're kind of in the tech bubble the way maybe you and I are, that sometimes. Are we just building this, like, Rube Goldberg thing where we're building all the plumbing just because we can? And look how cool this project car is. But is it actually useful?
00:08:40:17 - 00:09:04:07
Unknown
And when you think about what you're building with intent and the things that you've some of the things that you've kind of called out, where does the abstractions start to stop for us in the industry, whether that's designers, developers, like, what do you think in the next? And I hate putting timeframes to it, but over the course of the maybe the next two years, where do you think a lot of us are going to find the most value out of moving up that layer?
00:09:04:09 - 00:09:26:06
Unknown
Do you think processed code gobbling up of like, the SDKs? Is that just the future of all code and we're not coding anymore. How do you think software is going to be built moving forward? It's funny because like like I learned to code back in the days when like Dreamweaver was a big thing and like with Dreamweaver, right.
00:09:26:06 - 00:09:47:10
Unknown
Like it's images, right? Like you draw like you have these images and they get sliced into image maps, right. And like the code falls out of the end. And like if you look at the code, it's horrible, right? Which is like there are a lot of reasons it didn't stick around. But like a derivative, a div, a div, a million divs, which is, you know, a lot of websites or so like that.
00:09:47:12 - 00:10:09:19
Unknown
But then like screen sizes got more complicated and everything became dynamic and like that didn't work. So what I think for the way developers work is we have these loops, right? Like where you start with a plan and then you implement. So you write the code and then you like look at the thing that was implemented and like that's like the higher level architecture for any given thing we work on.
00:10:09:19 - 00:10:29:03
Unknown
But it's also like the micro like as we build, we're going through these loops over and over again. Like, like, what are we gonna do next? Implement the thing, look at the thing, and then plan what you're going to do next. And you're doing it over and over again. And like most of the time, is spent with the implementation because code just takes a long time to write.
00:10:29:05 - 00:10:30:13
Unknown
And a lot of like brainpower,
00:10:30:15 - 00:10:56:07
Unknown
I think a lot of developers are mourning right now, the flow state that you get when you're like a master of your tools and you know the syntax really well and like, the pacing is really nice. So like, you're like pretty drawn into it and like, engaged, for the whole time, you feel like your expertise is, like, driving it forward, like you're like, I like, train for a long time.
00:10:56:07 - 00:11:06:08
Unknown
Look at me. I'm at the end of the montage. And, like, I think it's just really hard right now for developers where we're trying to find the new rhythm, where we're like, I've lost my flow state because,
00:11:06:08 - 00:11:13:07
Unknown
like, I ask an agent to do something and it takes like 20 minutes and like, what am I supposed to do?
00:11:13:07 - 00:11:46:16
Unknown
Then maybe I'll, like, start this other thing, but then, like, it's kind of this, like, herky jerky thing. The journey for developers is going to be building tools and figuring out workflows that bring the flow back. But the work itself is going to look different, right? It's going to be less heavy on that implementation phase and more heavy on the like planning, thinking, brainstorming, being intentional phase where I think, the medium isn't necessarily going to be code.
00:11:46:18 - 00:12:05:07
Unknown
I think like as you increase the resolution of the thing that you're building, right. Like usually when I'm like, I'm going to add this feature, I'm going to add dark mode to my website. It starts as like literally I'm going to add dark mode to our website because, you know, I'm supposed to. And then like as you work through it, you're like, what does that mean?
00:12:05:07 - 00:12:30:05
Unknown
Is it for like all of the pages? Like, what do I do with the images? And like, you like kind of work through the problem at a pretty high level. And as you work through it, you get higher and higher fidelity for like, the plan of what you're going to do. And I think like oftentimes it goes from like the medium of natural language.
00:12:30:07 - 00:12:50:09
Unknown
And then like maybe there's a spec that's written up and you read it through and it's like words, but it'd be hard to have like diagraming tools and like, kind of like design tools exists for a reason because they're like, you, right. Level of fidelity in the right physics for like that middle phase where you're like, I don't know, maybe this thing goes over here, right?
00:12:50:09 - 00:13:05:14
Unknown
Like that's pretty hard to do in code, but it's pretty easy to do with and like Figma and then like maybe the medium shifts in its code and you're like, you have a prototype and you like, work with it and you iterate on it. And then at some point it's like the code itself and you're like, this behavior isn't right.
00:13:05:14 - 00:13:30:11
Unknown
Or like this needs to the transition needs to feel a little bit more like this. And it like, I think there is a new flow state that, is more focused on, just like crafting and bringing something in, like into the world that's just like less code heavy. But it's still flow state and it's like, still engaging, and it's still like you bring something into existence.
00:13:30:11 - 00:13:52:16
Unknown
And, I think we just need the tools that make that possible. How do you begin to research something like this? You have a background where you've done a lot of research into a lot of these different paradigms, and it seems like there's a lot more to consider because for a while there, you know, I was starting to subscribe to the idea that across the modalities that we're building software for, we're starting to maybe reach a little bit of a ceiling.
00:13:52:21 - 00:14:11:21
Unknown
There's only so many ways I can reinvent a table or some progressive disclosure patterns. How do you start to think about it when there might be and I don't know if there's like, science behind this, but for me personally, I'm very I Pilt I am using agents for all sorts of things, and I find myself having to do things like if I'm coding.
00:14:11:23 - 00:14:28:02
Unknown
Yeah, of course I have to hold variables in my head and I have to remember a lot of things, and it feels like that is amped up by a factor of like 3 to 5 when I'm using agents because I've got seven different, things happening. I'm orchestrating. I'm having to zoom in and out very quickly. How do you start to research?
00:14:28:02 - 00:14:53:18
Unknown
What are some of the optimal ways for someone like me to work with an orchestrate across AI? Does that start to change the interface patterns you think about? Yeah, I think we're very much just like still figuring things out. And like, I used to, design and build tools for, you know, like marketers and publishers, and that was a lot harder because you'd have to, like, talk to someone whose job it is today.
00:14:53:20 - 00:15:13:10
Unknown
And I've been very, it like a blessing and a curse for the past, I don't know, five years to focus on tools for developers. So it's like I have my own instincts of, like, build a thing, try it out. This is shit. Or like, this is great, and I use it and I'm going to keep working on it.
00:15:13:10 - 00:15:36:00
Unknown
A lot of it's just like trying new things and having intuition until like, oh, this feels new. Or like we could push this in this direction. Like figuring out what is this new medium and what are the new things that we can build around it. And like starting with something very fuzzy and fleshing it out and trying a million different things.
00:15:36:00 - 00:16:02:14
Unknown
Like I like to live between design and development, where, the questions to be answered are in the behaviors and in the dynamic dynamism of like the tool, as opposed to like I've talked to people who spent a year designing the undo button on Twitter. Right. Like that does not bring me joy. What brings me joy is like what could be and like, how would that feel experientially?
00:16:02:16 - 00:16:26:17
Unknown
So for me, the like prototyping design process is all just like build the thing, try it out. Is it bad or is it good? And then push on the things that are working and discard the things that aren't. So if we see this kind of new way of work starting to take shape right now, it's almost a running joke on on like social media that you have a lot of people working with agents and it's like, oh, what did you build?
00:16:26:17 - 00:16:45:04
Unknown
Well, I built the thing to make my agent better, and it's kind of the installation phase that maybe we're in. But some of the things that people are building, it seems like it's just a recurring pattern is a way to visualize what their agents are actually doing, and then to maybe like inject control. If an agent starts severe.
00:16:45:06 - 00:17:04:11
Unknown
And I've seen so many different ways I've built my own where it's like, I want to see a node diagram, and I want to see all the things that I can do to to the extreme where people are building, like full game interfaces, and I can send my little character over to the tasks they need to perform. Are there any emerging patterns that you're starting to notice as you're kind of building out?
00:17:04:12 - 00:17:28:11
Unknown
Yeah, it is fun to see all of the experimentation. Like people, people get crazy with it. And like it's also hard to know from Twitter, right? There's so much noise and you're like, man, they got like 3000 likes and like it is cool, but like, no one's going to be using that next week. So like there's like things that look cool, but then you try it and you're like, I can't, I can't get anyone any work done like this.
00:17:28:11 - 00:17:55:10
Unknown
I think, like, spiritually, like things should be as visible and transparent as possible. I see the like AI coding tools progression as a series of like, arrows, I suppose. Right. Like there was like GitHub Copilot, which was like, we're in the editor and we're writing code, but it's like fancy autocomplete, right? Like it's just kind of moving sidewalk as I walk along chat in the IDE.
00:17:55:12 - 00:18:16:06
Unknown
Right. Like these agents are pretty good at like answering my question. So let's have them answer more questions. And then it's like, well, now they're quite good at writing code. So like maybe I can just ask them to write the code for me. So like the agent like chat in the ad phase. And then I feel like people just like, felt like we need to tear things down.
00:18:16:07 - 00:18:40:05
Unknown
So then we switch to like CLI tools and like I use tmux and there's like ten terminals going at once. Because you can kind of rethink your workflow and you're not in this like very designed, like maximally real estate used like ID right. Like I think of VS code and it's like a power tool designed around me reading and writing code.
00:18:40:07 - 00:19:12:15
Unknown
But maybe that's not what I'm doing anymore. So we need to, like, tear it down to rethink it. And then we're in the CLI and we're like, wait. But like I have no visibility into what's happening, right? Like I've got agents going. But like in order to figure out what any of them is doing, I need to like read a small title or like scroll up in this chat log and like, they're all working on different things and it's really hard to see like anything, like all the things that are happening and like it's just like all the information and also none of it in kind of a weird way.
00:19:12:17 - 00:19:41:12
Unknown
And now it feels like we're going into a new era where it's like, wait a minute, interfaces are actually quite nice sometimes. So people are like building the app, which is what intent is, right? It's like going to a dump to rebuild our developer tooling around our new workflow. And I feel like we're now bringing back in like these visibility observability transparency tools where like, we can see all the agents that are running, we can see all the code changes.
00:19:41:12 - 00:20:07:06
Unknown
It's really easy to go through them. Maybe we're adding tooling to make it easier to do common things like code review and like making a PR, things that are usually in interfaces, right? Like this is a common thing that happens. Let's add a button for it. But like the the trick with it is like there's so much going on that like you can't just show all of it at once because if you show everything, you can't focus on anything.
00:20:07:08 - 00:20:32:19
Unknown
And so it's been like an interesting exercise to figure out how do you focus on what's important at any given moment and like allow the developer to, like, open the hood and like see what's happening with the engine if they need to. But like by default, the hood is closed. And like, like make it this like recursive interface where you can go up and down, right.
00:20:32:19 - 00:20:57:18
Unknown
You're like up here working at the level in ten and you're like, wait a minute, what is that agent doing? And you can like dive in and be like, what is that agent doing? So like I don't know if there's any like specific interfaces that are magic bullets right now. The other thing that this makes me think of is like with data viz, the whole point is for it to be a lossy compression of the data, right?
00:20:57:18 - 00:21:18:19
Unknown
Like the data itself is like massive. You could get insights out of it, but it'll take like a year to like read through like every row of, of data in a data in like a huge data set. And what you're doing when you're visualizing data is to, like, get rid of most of this signal so that you can focus on a specific perspective.
00:21:18:21 - 00:21:40:17
Unknown
Right? Like if you put something in a bar chart, the the map is not the territory. You're losing a lot of the like, like texture of the data itself, but you get to focus on a specific thing. And I think those same principles apply to like an inner piece over like a massive amount of information is like we need to have that higher level perspective.
00:21:40:19 - 00:22:03:12
Unknown
So, so that you can like get stuff done and have a perspective, but still make it possible to go down to the like nitty gritty level. And so in a, in a, in a workflow like that. And now I'm thinking, you know, the way that I work in our team is, is five people. Like we're very small and we're definitely we have the luxury of of experimenting a lot.
00:22:03:14 - 00:22:20:07
Unknown
And I'm aware that, like the way we work today is probably going to be to change a lot. But as we talk to a lot of people, like the question is, okay, but what is practical for a team? And especially a team at scale? And when I think about that question, I start to ask myself, where is the source of truth?
00:22:20:07 - 00:22:43:00
Unknown
When you have humans and AI working on the same sorts of projects? And I know that maybe you've talked about like the Living Spec, and for me it was like, oh, quad project spaces was wonderful because I got to like, upload new context. And then my, you know, the outputs were really great. And then now it's like, oh, actually, I can tell the agents just, hey, update the document as you make progress.
00:22:43:02 - 00:23:11:18
Unknown
Is it that the spec starts to become the team's source of truth? How do you think about that moving forward? So one of the new primitives that we've introduced in the intent app, which is desktop app, trying to replace it for developers, is the concept of a note, which is just like a place for collaboration and communication and like it's like a rich markdown editor, basically, that has a few, like new affordances, right?
00:23:11:18 - 00:23:33:00
Unknown
Like subtasks and things like that. And like, it just feels like there needs to be a place for that source of truth. Right? Like as you talk to the agents, there's so much information and like, it's kind of like this progress log of all the things that have happened, if you want to know, like the current state of things, you need to like, scroll up and like figure out from like this log.
00:23:33:00 - 00:23:53:17
Unknown
It's like figuring out what a code base is from the git commits. We need this place where it's like the source of truth that the agents are working from and the humans are working from. And there's like some really interesting explorations around what notes are like helpful for agents to all work from. So they have like a shared understanding of the work that's being done on any given task.
00:23:53:17 - 00:24:18:08
Unknown
And like that I'm still fleshing out. And then there's the other use, which is like, what is what is useful for humans to understand. For like you, you throw like ten agents at something like, what is the information that is important for you to like, take out and have like a report, right? Just like if you had a team of ten people, you don't want to know exactly all the things that they're doing, but you want like the report at the end of the week of like, this is what we did.
00:24:18:08 - 00:24:38:22
Unknown
And I'm like, this is like the high level takeaway. I, I, I'm curious what you're doing. Right. Like, you know, I find myself and I feel like I saw someone share an interesting post. Maybe it was Karpathy. I'm not sure who, but it was a screenshot of I think maybe Codex is sidebar, and it was just like a like 20 different threads of agents and it's like, this has to get solved.
00:24:39:00 - 00:25:00:00
Unknown
And it was interesting to me because it is at a point where there's, there's a lot of cognitively I'm paying close attention to what's to how I'm thinking about things. What are the shortest feedback loops I can have with any of these automations where I still feel like I'm driving it and not being just pulled towards whatever's happening, I don't have the right answer.
00:25:00:06 - 00:25:22:23
Unknown
For me, it's pretty primitive. It's like, oh, I've got obsidian and I can pull up a doc and get an idea. And hopefully the enforcement layers I've set up with my agents are actually being enforced. And therefore, when I look at these documents, they are accurate. That's not always the case yet, but that's really where I live is I look at the documents, I look at the progress.
00:25:22:23 - 00:25:45:03
Unknown
So I have like intent. Funny. And then I have the, progress. And then I diff like is the progress matching the intent and if not whoa whoa whoa slow down. We got a, we got a course correct. And that's kind of like this rough loop I have that doesn't feel perfect. It feels like a sputtering project car, but it also kind of works sometimes.
00:25:45:03 - 00:26:13:23
Unknown
And then when it works, it works very well. And that's where, like, I'm living right now. But I again, yeah, I'm like one operator. My team is five people, but we're not touching the same code base or project every day. My, I am so curious how teams are going to think about that, because I've heard people say, oh, you know, in the future it's going to be, you know, you don't have agent orchestration over here and you can be managing these layers of agents, and then you have human orchestration over here where you're just continuing to manage the humans and both are going to live.
00:26:13:23 - 00:26:52:03
Unknown
And how do you do that? I mean, how do you think teams are going to operate around this? What do you predict might happen? And I know predictions are tough and but what do you hope might happen. Yeah. Rich train of thought. There. I think I think the thing we keep circling, right is like, the problem is we're still trying to work at the level of abstraction we're used to working at with, like, the nitty gritty or used to day to day writing code and, like, we've turned the speed up and, like, there's more people working on things because some of them are agents.
00:26:52:03 - 00:27:18:16
Unknown
And it's easy to have agents like work on more and more things. And we're still trying to work through our existing tools and the like, existing medium of like, code and like texts that we've been, like, trying to work at. You can't read books faster, right? Like what you need like SparkNotes or something, right? Like you need that compressed version of like the raw signal in order to properly manage.
00:27:18:16 - 00:27:46:02
Unknown
Which means you need like proper like compression of like give me the gist. You need trust in like, you ask an agent to do something. It's going to do exactly that in like in the way that you expect. And you also need trust that like that compressed version is in sync with the actual work itself, which is like something we saw with people have built tools around inspect urban development.
00:27:46:02 - 00:28:08:05
Unknown
But like once you have a plan, the work starts and then it diverges, right? Because like naturally when you start doing the work, the real world isn't as clean as it looks on paper. And like things need to change or, you know, some of the work takes a lot longer than you speculated. And you need that plan to stay like syncs, right?
00:28:08:05 - 00:28:28:15
Unknown
You need the bi directional sync between, like, the compressed version and like the expanded version. And like we've tried to build that into in time, where we call it like a living spec and like everyone works from the spec. So it's not just like this document you created that you then delegate. It's like the, how did you put it?
00:28:28:15 - 00:28:56:18
Unknown
I liked that, like the source of truth for what everyone is working on. So that like when you delegate agents, you see what who is working on what, like through the spec, and the agents will update the spec based on what they're working on and what their findings are. And like I find that's like, just like a really nice working surface where you have this like, like short, concise document that captures like at a higher level but not too high a level.
00:28:56:20 - 00:29:13:04
Unknown
What is the work to be done? And from there you get to see what the what is the work being done and what is the work that was done. We need systems by which we can compress and then, like keep that compressed version in sync with the actual work being done, which these agents are quite good at.
00:29:13:04 - 00:29:38:09
Unknown
So, like, we're in luck that like, we have the technology to be able to build these systems. It's just it seems to be a matter of like having them as a nation to like, think of how things could be different and then like just putting agents in like the right harness. So I'm going to take a step back because there's a lot of people who are listening to this, and there are certainly going to be a portion of them who love everything.
00:29:38:13 - 00:29:55:19
Unknown
We're nerding out about here. And then there's a lot of people who are trying to understand what's happening, but they just don't have the time. And a lot of what we just said, they're like, what did I just listen to? I don't understand these layers yet. And I want to I want to grounded a little bit. And so two things I'd like to do.
00:29:55:19 - 00:30:15:09
Unknown
One is I kind of want to hear from you. How are like, do you have an example of how you're actually working around these concepts that you prefer right now, even though that might change? And then after that, I actually want to pull up some of these side projects that you build these stories, these narrative websites. And I kind of want to dig into how you put those together.
00:30:15:12 - 00:30:36:09
Unknown
Can we do that? Yeah, yeah, that sounds great. This is the app I've been working on. And I bring it up because this is like how I do, my day to day life. Right now, the core unit of work within intent is not an agent chat, which is something we need to make more clear, actually, because it's quite different from the way other apps work.
00:30:36:11 - 00:30:59:05
Unknown
The core unit is a workspace, which is just supposed to be a bundle around any given task. So you can see here, like for my website, I've got like I was trying to play with rebuilding one of these leaderboard charts. I've got like, a few blog posts I'm working on, maybe show my blog posts in a grid.
00:30:59:06 - 00:31:25:10
Unknown
Here's another blog post I'm working on for, like, the app itself. There's like bugs reviewing paths. There's some, like, really cool features that I'm working on. I show this because, like, I am juggling so many different things. And like, the thing I like about this workflow is they're completely isolated, and they're. Which makes them really easy to, like, pick up and put down.
00:31:25:16 - 00:31:55:08
Unknown
So if I look at any one of these, like, if I open up, or something that was working on like AI native canvas, right. Like this is like I was playing with, what does it look like if, you have think jam as, like a, brainstorming tool within your, you're like, coding environment and like, when you pick it up, you have the workspace as you were working on it, other than my dev server is down.
00:31:55:10 - 00:32:21:23
Unknown
So this is like a preview of, I have a sandbox that I usually use to like go through all the edge cases for like different interfaces. So like this is, this is like my working environment. Like if I had like 20 desks to work from, and one desk per thing I was working on, and I had, like, all my papers scattered on the desk, like, this would be like my desk as I left it the last time I was working on this.
00:32:21:23 - 00:32:46:16
Unknown
So I had like one agent over here where I was like, going back and forth on like, hey, this one looks really funky. Can you clean that one up? And over here is like my view into like all these different edge cases for this like whiteboarding tool. Right. Like and then if I look at another one right, like what does it look like working on this, this blogpost like this is, this is the desk as I left it.
00:32:46:18 - 00:33:10:01
Unknown
So my flow looks a lot like all right, maybe I want to add this feature. Let's like, like kick off a new workspace. Right. Like let's go ahead and add dark mode to my website and I'll like kick off the new workspace and like we will do like two with different models. And what they'll do is they'll come back with a spec like a, like fleshed out plan of like I did a bunch of research.
00:33:10:03 - 00:33:29:05
Unknown
Here's how that would fit into this code base. Here's my proposal for how that would work. And you can, like, go back and forth on it, tweak the spec itself, and then like at some point you're like, go ahead and implement it. And it'll like delegate to some agents to, to to work on the plan. And at some point the plan is done.
00:33:29:06 - 00:33:52:14
Unknown
And like because these things are like so isolated, I just like I what I do every day is I jump in between, like all of these and like, whether it's something really tangible, like reviewing a PR or whether it's something really speculative, like, what if we had a whiteboard in our like, on our developer tools? I just, like, kick things off and then like, try to juggle them, right?
00:33:52:14 - 00:34:12:18
Unknown
Like I pick one up, I'm like, what does this look like? Where was I with that? What what I'm seeing here is this, this progressive disclosure of, like, I want to drill into I want to see all of my workspaces. I want to drill into one of those items I can open up. Like then it starts to show maybe a linear like, here's the research, here's like even further research or some of the progress being made.
00:34:12:20 - 00:34:40:20
Unknown
And and then I wonder at this point, because for me, when I view interfaces like this and it's this is very analogous to like the workflows that I'm experiencing, I when I'm drilling into things, then I want to open up one element, one node into like a Zen mode, you know, so I can again, I can go backwards if I want to, but when I get in here, I want to open up that so I can just be absorbed into the one node for a second and then zoom back out and go back to actually let's see if the research is still matching up.
00:34:40:22 - 00:35:01:10
Unknown
Is that is that one way that you're thinking about it? Is that how you work recursively diving in until you're at the, the like level that you're focused on and that needs to be pushed forward. And then like zooming back out and then like zooming back in on like another part. Right? Like if it's a map, you're like, all right, I want to like, go over here.
00:35:01:10 - 00:35:28:19
Unknown
So zoom out and then zoom in, like to like the city. And then you're like, wait, let me check on, like, this other part. Zoom back out and you zoom in until you're, like, in Ikea and, like, looking at, like, the different shelves in Ikea, right? Like, some of what we're trying to build into this is like keeping you oriented as you mix between like, personal blog post and like really big feature at work and like adding a landing page to something, right.
00:35:28:19 - 00:35:52:05
Unknown
Like it's really hard to context switch like so much. And we're doing more and more of it as like AI accelerates our work. So just like keeping people oriented between these switches feels so important. So I can show I'm looking at one of your blog posts here. Our interfaces have lost their senses and you've got some really wonderful not just designed.
00:35:52:05 - 00:36:16:06
Unknown
Right. Yes. This right here and so I'm curious. Two questions. One, first, I'd love to for you to walk us through how you kind of built this and your reason for making it exist, but two, what I'd also love to know is that process you just showed us was that was this a product of that process you just showed us?
00:36:16:07 - 00:36:39:22
Unknown
Yeah, yeah. So this they wrote, I think like a year ago, eight months. I realized that I haven't written a blog post in a while. It's like a few things, right? Like I'm like always collecting things that I feel like push my understanding of something forward as well as, like I saw, like I saw some like image of like felt people.
00:36:39:22 - 00:36:58:14
Unknown
And I was like, oh man, that's really cool. How could we push on that? Like what that feel like on the web? I like how all these threads of things that come up for me as I'm like exploring and as I'm trying things out, other people build trying stuff out that I've built and like at some point they go here into like a larger train of thought.
00:36:58:16 - 00:37:23:23
Unknown
And if it's not something that I've seen out into the world, in the world, then like I try to put it into a shape that is generative for other people. A lot of these ideas kind of start from this rough place of just a collection of thoughts and things. Humans that are felt, some like higher level research maybe, or some concepts that you're trying to articulate before you get to the kind of build phase for something like this.
00:37:23:23 - 00:37:46:10
Unknown
Is there. Where does all of that? Where do all of those pieces live for you? It's changed. Like I think when I wrote this one, I was really pushing hard on like, what does it look like if ChatGPT is a thought partner? But so like going on long walks and just like rambling and having it respond back to you, right?
00:37:46:10 - 00:38:05:11
Unknown
Like in the way that you use it, you use like a pencil and paper to like put your thoughts down and then look at them and react to them like, can we externalize our thoughts and have them repeated back to us with like these limbs, right. And can they, like, turn the thought and can you see it from different perspectives, so that you can evolve it?
00:38:05:11 - 00:38:31:03
Unknown
So I think some of the, the like evolution of like, this is just a bunch of like random thoughts. Can we pull it together into one type thread? I think some of that came from just like using CBT as a sounding board. And then like so there's one thought, I'm trying to think one like thread. I'm trying to cohere today and like that has been conversations with people.
00:38:31:05 - 00:39:02:09
Unknown
It's been sketches on paper. It's been like, sketching out in Figma, like, like, this is what I mean when I'm like, I talk about like, can we make our coding tools, brainstorming tools for thought that where you can change the medium through which you're thinking things. Because different tools have different physics, and they let us explore different parts of it with different levels of fidelity.
00:39:02:11 - 00:39:22:06
Unknown
So I'm constantly just changing, like, okay, what is what does it mean to explore this in this and in this medium or this other medium, like this is the level of the thought as I try to work through it. So this is like, all right, so here's the timeline for development. We have planning, we have implementation, we have review.
00:39:22:08 - 00:39:56:18
Unknown
But really in the in the implementation phase, like it is mini versions of that. Right. It's planning implementation review. Like you put something down, you look at it, you make a new plan. And now the agents are doing this part. Not only is it shorter, but we also aren't doing those like mini planning steps, right? Like we don't are our understanding of the problem does not evolve as the implementation happens, and we're also doing less planning up front because it's easier to just be like and dark mode because the agents will figure it out, right?
00:39:56:20 - 00:40:18:08
Unknown
And then they'll put something together like you could put like three words into a prompt box and have like a full feature. Like you do not have to think very hard upfront, but like it doesn't necessarily go to a place that you wanted. And also like you feel less ownership over the end result. Right? It isn't really something that you built.
00:40:18:08 - 00:40:38:10
Unknown
It was just something that you're like, I don't know what add this whatever. And then there's this tangent of like, Elon's are really good at expanding text into more detailed text. And the reverse like what we've talked about before of like compressing a lot of text into a smaller amount, the bare minimum of like a dark mode.
00:40:38:15 - 00:41:16:16
Unknown
Right. That's about this. As like small as you could, like point it, a direction and then like on the other end is like the full code, right? This is like all of the logic that goes behind, like, what is dark mode on my website and like, every single edge case is, is, necessarily in this like fully expanded version and then like somewhere in the middle is like a spec where, like you, you say like a dark mode, the agents come back with like some expanded version of it that's not fully expanded because that's too much information and it's hard to steer from there, but it's like it teases out some of those assumptions
00:41:16:16 - 00:41:40:11
Unknown
that are in this, like, totally compressed version. As agents do more of the work. It seems like most of the work now allows, like if we trust that agents can properly execute code like this stuff, the implementation and the review falls out of a sufficient spec. Like if you've detailed enough of the edge cases to put your intent in there, they can just build it.
00:41:40:12 - 00:41:57:21
Unknown
Right? So most of the work is now squeezed into this front part. Like the plan. What I'm finding is I, as I talk through this, is like, people do not like the word plan. They're like, oh, well, we're all like PMS and we're going to work and like Jira now. And it's like, that's not really what I mean when I say plan.
00:41:57:21 - 00:42:19:08
Unknown
So like that part needs a little bit of like fleshing out. We still get these loops of like think, implement and then review. I just think the medium is going to change. Right. Like this starts with like natural language and more natural language to spec and then like diagrams and like sigma level and then code and then like you're iterating on the code, right.
00:42:19:08 - 00:42:46:06
Unknown
Like I think it's, it's the same loop. It's just very like intention heavy. Something that's clear about you from this conversation is you are very you are a very intentional pathfinder. And I really appreciate you walking us through how you do that and whether it's like the products you're working on from GitHub to intent to the websites you're putting together to tell these stories, there's a lot of deliberate pathfinding you're doing.
00:42:46:06 - 00:43:17:17
Unknown
And I and I love that process. What would you say to people, the engineers, designers, PMS who are frankly too employed to spend this kind of time pathfinding in this new kind of way of work? Where would you point them in terms of what signals to pay attention to and what signals they can safely avoid? For now, the only way that I learn things is to have, like an end in mind and then look around and figure out, like, how do I get there?
00:43:17:19 - 00:43:40:08
Unknown
I think we like, I think everyone should be playing with these new tools because, like, the physics of them is different than the tool, than the mediums and the tools that we're used to working in. And I think, like if you sit and think of like, what are the things that you need to get done and like, is one of them like, can you try to get there with while exploring one of these new tools?
00:43:40:10 - 00:44:11:12
Unknown
And just like pushing on that and like, like, hey, I need to do this thing. I usually do it with this one route, but like, I saw this, this thing on Twitter and like, maybe I'll try it out and it might, like, crash and burn horribly. It might cost you, like, $10, right? I think we owe it to ourselves in this moment in time to, like, figure out what is this new technology and like, as you use it, extrapolate to what it could be because like, things are changing, like the models are getting better every like month.
00:44:11:14 - 00:44:33:18
Unknown
I think a lot of people try it and like I hear like I tried using an agent like a year ago and it was shit. And it's like it like they're completely different today than they were like two months ago. So just like continuously trying to work things into your workflow and like, keeping an eye out for, things that work and things that don't and, like, just keep trying.
00:44:33:21 - 00:44:58:15
Unknown
Amelia said the agents weren't going to eat ice cream for her because the point of eating ice cream was for the experience, and you can't automate that. And I think that's the cleanest way I've ever heard someone describe where the abstraction stops. You go up and up from punch cards to assembly to JavaScript to agents and then orchestrators and at some point you hit purpose, the whole reason you're doing the thing in the first place.
00:44:58:16 - 00:45:21:23
Unknown
And that's the layer where Amelia thinks the humans stay. And I also liked her point about compression, where the most expanded version is the code in the most compressed version is at dark mode, and somewhere in between. That is a document that teases out the assumptions without drowning you in all the implementation details. She thinks that middle layer is where most of the work is going to live now.
00:45:22:00 - 00:45:31:07
Unknown
Personally, I don't know what the next answer will be, but I do think the question she's chasing is the right one. That's the episode. I'll see you next time.