a16z Podcast - What Happens to Design After AI?

Episode Date: June 24, 2026

Anish Acharya speaks with Microsoft VP of Design John Maeda and Impeccable founder and CEO Paul Bakaus about how AI is changing the practice of design. The conversation explores the relationship between design and technology, the rise of AI-powered creative tools, and whether automation raises the floor, the ceiling, or both. Maeda and Bakaus discuss software craftsmanship, taste, creative judgment, and why some aspects of design may become increasingly automated while others become more valuable. They also examine agentic workflows, the future of user experience, the role of designers in an AI-native world, and how new tools may reshape the relationship between designers, engineers, and software itself.   Resources: Follow Anish Acharya on X: https://x.com/illscience Follow John Maeda on X: https://x.com/johnmaeda Follow Paul Bakaus on X: https://x.com/pbakaus Get the GitHub Copilot app: gh.io/app Stay Updated:Find a16z on YouTube: YouTubeFind a16z on XFind a16z on LinkedInListen to the a16z Show on SpotifyListen to the a16z Show on Apple PodcastsFollow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Designers when using Claude, as opposed to engineers using Claude, would consistently get better results. And it's because of the language that they use. We have to remember that design in the European sense came from royalty and the desire to be distinctive because they were working with scarce materials. What's interesting about this era is that this idea of taste doesn't fit when all the materials are available to everyone.

Right now, everybody's trying to solve whether LLMs have taste. These models have millions of definitions of taste. LMs have been trained on the output of humanity, not on the input. So what led to a design decision is not something that the LLMs know. Maybe advice for our design engineers in the room. How have you in the past effectively communicated the value of an instinct versus a deadline or even another instinct, which may be less important? Yeah, that's a tough one because AI is making it easier than ever to create software.

But what happens to design when anyone can generate an interface, a website, or an application with a prompt. Some argue that AI will commoditize design. Others believe it will make great design even more valuable by automating routine work and freeing people to focus on higher order creative decisions. In this conversation, A16Z general partner Anishaara sits down with Microsoft VP of Design, John Mehta, and impeccable founder and CEO Paul Backus, to discuss design, software, creativity, and what happens when AI becomes part of the

creative process. I want to actually talk a little bit about perhaps to begin with the relationship between design and technology. I think there's some people who may view design as sort of more spiritually close to art and others who see it perhaps in a more utilitarian fashion, you know, the design of everyday things. So perhaps, John, I can sort of ask you to kick us off and talk to us about how you see the interplay between the two. Well, first of all, glad to be here. I used to think about this a lot when I was at Clarenda Perkins and thinking like, why is that the design is important in 2014-15? Like, why was it important? It was because this weird company called Airbnb was unusually successful. And if you connect to why design became important, it was because of mobile.

Before mobile, desktop experiences could be crappy and it was okay because he didn't use them very often. But mobile had high usage and therefore it was bad all the time. It would be painful. So that's when it sort of started to happen. In terms of the relationship now, however, I'm so excited to be on this with Paul because I've been a fan of this moment when we'd be able to auto design. And that's since the 90s when I was at MIT.

We thought it was going to be possible one day, and now it's very possible. And Paul, I think, as someone who's really at the cutting edge of all that. Incredible. So elaborate a bit for us. What do you mean by auto design? And is the way that it works today, Does that match the sort of idea that you had back in MIT?

Well, I mean, if you go back to the 1980s, there was a woman named Muriel Cooper, who was unusual because she was a trained print graphic designer. She designed those book fans, no MIT Press. There's a beautiful logo at MIT Press. She designed that among other pretty major things at MIT's history publication. She was one who imagined a world where people wouldn't want to use the terminal

and might want to use this thing called Helvetica to look at things on the screen. So she predicted the whole desktop publishing revolution, electronic publishing revolution, and at the time she also had a bunch of people from the MIT's AI lab. And MIT AI lab was asking, how do we create things by machines

that humans are good at? And so her lab focused on how to automatically design things with quote-unquote AI up the time. Incredible. Hall, give us your view on the relationship and interplay between design and technology. Has it turned out the way you expected? I think we're still working out, to be honest.

But I do see a lot of it converging in interesting ways. I remember seeing one of John's talks about algorithmic design a long time ago at some conference. And I always thought that I was always an oddball that was both designer and engineer. And I never wanted to be boxed in into any of those areas. But I think the reality is that it's no much easier to build yourself tools. that help you with that, that help you algorithmic, create design and supercharge your design with technology. And I think that's really exciting.

So to me, they're one and the same. They've always have been. I mean, John calls it auto design. I think we're basically, like, to use John's words as well, we're kind of like raising the floor in areas that can be mechanical. And I think that's really exciting because so much of design is spent with the things that we can automate. as opposed to doing like the really high level of thinking

of what actually needs to be done, what needs to be built. And that's something that I'm really excited about. Paul, do you think there's a tension between craft and automation and design? Yeah, I think so. I think there's ultimately, everybody has to think for themselves

how much they want to automate away. There is definitely still the intent and the judgment that is needed from a human side, I would say, on figuring out the what and what not to do. I actually really greatly respect people like Steve Jobs because Steve was one of the greatest editors in the world, right? He said no to so many things, as opposed to saying yes to so many things. And I think having a viewpoint is still important.

So I think ultimately the machine can work for you, but the same way you take a photo on a good camera, ultimately you figure out what you want to take a photo off. There's a lot of post-processing in the camera, right? There's so much, I mean, if you look at an iPhone camera today, there's so much happening on. device, that photo that you took is literally like not reality anymore, right? But there's still some sort of human viewpoint. John, maybe to bring it to you, what do you think are the limitations

of the sort of the cloud code codex and all the agents today when it comes to design expression? Oh, wow. Well, I mean, everyone's saying it's limitless and it's that, not the, whenever the Cambrian period, blah, blah, blah. There's so many adjutants that are like, oh, my gosh, amazing. I think the reality is that this has been a long time. coming. In the same way, we saw writing being automated before LLMs, and we saw marketing assets being automated as well. So we just spread that to the kind of, we went upstream on this idea of designing things, which, to Paul's point, it requires restraint, it requires all kind of things, which is the opposite of engineering. Speaking as a engineer as well, if I can make it, I'm

going to build it. Oh, I'll do this. I'll keep doing this and this and this and this. And so restraint isn't part of the equation, unless you're an architect. Designers always think like architects of the human experience. Now, with the harness and the models coming together, we have a nice combo platter, plus the e-vowels. We have the feedback loop. That's actually really brand new. And that's why I'm so excited.

We've been experimenting with Paul. We have this something called the GitHub co-pilot app. It's a new app. And the challenge with that has been to keep the human taste and craft and all the high-tech goodies. And to Paul's point about raising the floor, humans should be able to raise the ceiling of what they can do. So we have human animators making the craft

and the final decisions around some of the super-duper high extreme polish, but we want to have the coding agents to the part that have the drudgery. Yeah, that makes perfect sense. I want to pull the thread about the new Git product, but Paul, why don't we actually take a step back and you can talk to us about Impeccable. What was the intention? What does it look like today?

And then let's talk a bit about how you and John are working together. Yeah, of course. So Impeccable started as an operations project that I needed for myself. So I used to contract for a couple of my friends in the last year to figure out how far can I push AI engineering and how much can I 10x or 5x or whatever myself. And I know many are going through a journey right now. In the process, I had to build a whole lot of tooling for myself. and one of those tools was impeccable.

And ultimately, I noticed that whenever I would tell Claude or Codex make this better, it would do a really terrible job because they didn't know what angle of better, what does better mean in the design context, right? And I also noticed that designers when using Claude, as opposed to engineers using Claude, would consistently get better results. So why is that, right? The model is literally the same. The website is the same.

The design system might be the same. but the designers still get better output, even though they're prompting a machine. And it's because of the language that they use. It's because oftentimes engineers don't use the words. Things like vertical rhythm or negative space or make this bolder or quieter. They don't have the same vocabulary as a designer

who's been in the game for a long time. So the first instinct that I had was to bring that vocabulary to the actual harness, to the agent harness. And that was the first iterative. of Impeccable and that already made a huge difference. And interestingly, I thought that it would alienate designers and it would mainly appeal to engineers, but that was not the case.

I think a lot of designers are being asked right now to move closer to code and I think that makes a lot of sense and maybe they're waking up themselves to that idea because everybody's moving closer to code. Code is the substrate that the agents use, of course. And Impeccable feels like a way to get there much more quickly and with the language that they use. So I would say it's actually like, you know, if I look at the target audience right now, it's half engineers, half designers, a whole bunch of PMs as well, founders, entrepreneurs, but that's how impeccable came together. It is an agent skill that's completely open source.

It has lots of subcommands to steer. And then I kind of built a visual iteration mode on top of it and a quality layer that removes slop. And of course we can talk about that too. But it does stop the overfitting of the models. So not everything is cloud-based. as I call it, or, you know, instruments, the rive italics, or purple gradients. But, yeah, it's ultimately these three pieces. So vocabulary, it's

a quality layer and it's a visual iteration mouse that runs in your own co-base. Knowing the vocabulary, it's sort of like a set of magic spells, you know? Either you know them or you don't, and I love how impeccable makes them more broadly available. John, talk to us a bit about the analog

between what Paul's doing with Impeccable and Kai's power tools. Well, in the 1980s, late 1980s, there was this application that was called Photoshop that was amazing because it suddenly lets you draw pictures and play with photographs, or whatever. It was popular among photographers and people like that. But when something called Kai's Power Tools came out, it blew up the TAM for Photoshop because people could do weird algorithmic things with their faces or all kinds of

filters because Adobe Photoshop opened up this plug-in architecture. So people could write code, could put things in that plug-in format. And so when I saw Paul's work, it reminded me of Kai's power tools, but for design, and less weird, but very focused, very functional. When I saw impeccable, it reminded me of just giving away my era of when PostScript came out. Most people don't realize that PostScript was a miracle because the Adobe guys, the Warnock, the manual for PostScript is beautiful.

It's incredibly well designed. Those engineers understood graphic design better than most graphic designers. So PostScript properly encoded the perfect set of primitives to implement visual graphic design of the day. And so it required people who were technical and also design-minded. Same with Donald Knuzh, before that with tech.

It was an incredible typographer, a mathematician. So tech was a miracle in the combination of function and form. So when I saw impeccable, not to embarrass Paul, I was like, whoa, this is like the postscript moment. Someone actually gets that it's not just filters, it's also about subtraction. And it was all the different dimensions a designer might take. And so in that I saw a really skilled engineer designer, and that got me excited.

That's tremendous. That's so cool. Maybe, John, talk a bit about how you guys are planning to work together. Oh, well, I'm super excited because I recently picked up all of GitHub design. And great timing because we were launching this GitHub co-pilot app that's already in GA. And it is different than things we built in the past in that, as you know, the craft bar is very high now for applications. So we wanted to set the bar with this app.

And when they had a product Mario said, hey, do you know this impeccable thing? And I was like, oh my gosh, it's one of my favorite skills. And so I reached out to Paul, and Paul said, hey, oh my gosh, you know, we kind of made me, we know each other. And so we had this wonderful conversation, by techs. And since then, a few weeks later,

we're going to, we're going to, Paul's going to be the first one to be in our integrations of this powerful design skill because we believe that design is important in the digital, what do you call it, harness age. I think Paul has the best implementation of it. And Paul, how do you sort of see the product showing up within the new get, you know, impactful showing up

within the new get product? Well, first of all, I'm super excited to get impeccable into the hands of more builders and solopreneurs and individual enterprises. And I think GitHub is obviously a great place to do that. Everybody's using GitHub today, and it's a great way to roll out impeccable to more. I think the building in the skill as a built-in tool

is one thing. I think there's a lot more different directions that we can take from there, and that's what I'm really excited about. The GitHub co-pilot app has a, pick and polish mode, it allows you to pick and change something and quickly edit something.

And of course, Impeccable has its own implementation of that. And right now, you need to opt in and it's a bit clunky. And I think through first-party harness integrations like the collaboration that I'm doing with GitHub, we can actually

improve that significantly and make it so that it's a lot smoother. There's nothing there to be announced yet just now, but I'm eager to work with John and Mario on actually improving the impeccable and GitHub integration. So, okay, so follow-up question, Paul.

So obviously, Git is one of the most important platforms in the world, especially at this moment in time. So let's wind a clock forward in an imagine a world in which many people are using the new GitHub product and impeccable to have, you know, these sort of designs that are much improved from what you get default from the models. as those designs become in distribution for the next gen of models,

does the next version of impeccable have to steer in a different direction altogether? Or are these sort of evergreen best practices that live regardless of what's in or out of distribution? Yeah, that's a great question. You know, that's already happening. That's the really interesting thing. So in 2022, 2020, everybody thought AI slop was purple gradients. And to be honest, the models aren't really doing that anymore. The models aren't producing purple gradients anymore.

I mean, maybe like, you know, cheap models. But overall, they don't. I think the world has moved on to beige backgrounds, tinned backgrounds, instrument serif, eyebrow text. You know, there are the typical AI tells that you look for now. If you use clod design, oftentimes you run into a design like that, for example, but it's not unique to clot design. So how did this happen, right?

We basically said, okay, no purple gradients, and then the model steers into different part of the latent space and this picks the next best thing. Right. And so ultimately, slop is a moving target, and every time you move to another part of the latent space and everybody adopts it, you guess kind of like algorithmic Uniclo or IKEA or something, right?

And now that's not necessarily bad. It doesn't mean the design is bad. It just means the design is not unique. And you want uniqueness if you're building something like a landing page that attracts people's attention. So what I've tried to do with impeccable, and I'm constantly updating those rules, but what I try to do with impeccable is to build these types of anti-attractors. For example, instead of just going to a different font to the overused fonts, it actually tries to create a random seat that picks something based,

on a small user interview with you and maybe even some scripts that run that create randomness to actually steer it into completely different space of the latent space. I do have colors, I do have fonts, and we have a bunch of different things to create some sort of uniqueness. And there'll be more there that I'm going to do, but I really do want to make sure that not everything ends up looking the same. I ended up doing that before, by the way, like the reason why purple gradients, do you know, the reason why purple gradients are purple gradients tell me uh tailwind so tailwind's uh default theme

has purple i see because of the default theme uh the purple gradients were introduced everywhere that's at least the common saying right now uh the common theory and ironically i was at fault of coloring the web orange before yeah and the reason why is because when i started jQuery ui we've created a theme framework, a theme roller kind of thing, and the default theme was orange. I don't know why orange. I'm not sure why I picked orange at the time,

but I called up the whole web orange overnight, and I learned my lesson. The power of defaults, huh? Yes, we're sure. John, do you think uniqueness is an important attribute in design and maybe broadly talk about your view on uniqueness and slot? Well, as Paul described to me, I think we're in this era, because you can see all the new model companies

that are building designer models or visual design models, whatever. We're at this era where design will be further commoditized and automated so that a new kind of design emerges. I think that's what is the next way that will happen because of this. And the design and tech report I give every year for Southwest, I've been pushing, moving from UX to AX, agentic experience, An agenic experience is non-visual. It is the world of robots.comptus.coms.

It's command line dash-help. It's a world where we're designing for agentic affordances instead of just visual affordances. That force multiplication will occur once we can retire more of UX, I believe. But it'll require this sort of surge of automated design solutions

that cover 80% of those cases. So 20% can be spent by real humans, but the humans will move on to AX, I believe. Interesting. Can you say a bit more about it? I mean, what would be the role of human taste and judgment in AX? Oh, well, the thing that every designer knows eventually is that, and actually there was a famous typographer, Eric Speakerman, who designed a very popular soundsterifist called Meta.

This is before Meta became a company name. became a company name. Okay. And his company in San Francisco was called Meta Design. Okay. At the time, one of the top design firms in the world. Meta design speaks to the fact that design can either be the application of craft or the abstraction of craft.

And so I think of the abstraction of craft is, at the moment, what is the best way to design things? There was a time where being a design craftsman meant you could take, a picture of Helvetica, cut it with Xacto knife, put rubber cement on the back, and you were amazing. We moved to different higher-level primitives of the page, et cetera, with desktop publishing. Now, I think we're going to move even more meta, and so it means that designers are closer to programmers,

but programmers of interaction. And I think Paul probably he subconsciously is working in this way, when we talk about e-vows, Paul. evals live in latent space construction as literal meta-design. I think that's true. I think there is a whole area of design here. I also think you're probably right with the evils.

And the interesting thing with A-axis that oftentimes here you're designing for a different audience and that's the agent, right, as opposed to the human. I think there's, I guess my viewpoint is that that area of design will be increasingly important. I also think there's still the area of human-centric design as well. And I think, I mean, to the point that we met at the very beginning, that is shifting towards higher level work, right?

Sort of the last 20% of what makes something unique. And it's interesting. I mean, we've seen so many AI startups that are trying to enable creative work, right, whether it's music or video or whatever. but we haven't really seen the next Hayama Miyazaki from runway or some other video startup, right? We haven't really seen,

I think it's very fair to say. I don't think we have seen our music, AI startups next, I don't know, BDI or a co-player. And why is that? I think a lot of focus right now is raising the floor. And going back to that,

I think that raising the sea, I think there's too little time spent on that specifically, and I'm really excited about it. So I want to build tools that also cater to that, so kind of like both ends of the spectrum. But I do think that the world is shifting towards higher-end experiences. The audience is demanding more too. That's interesting, right?

If everything looks relatively polished and clean, if we're raising the floor and you have a clodd design website, that it looks cloth design, but it's a legit website, right? It's a good enough website to get the job done. Then the only differentiator will be whatever we craft on top of that. And I always thought that was very interesting. I spent a lot of time, like many of my design engineer friends, we spend a lot of time crafting something really unique and particular

and sort of like, I don't know, like a slider that is beautifully animated, right? But oftentimes that work would drown. That work would not be seen by higher ups who have completely different business goals. I think it's the moment for these people right now to shine. That's one of my theories. I think the craftsman can kind of grasp the last 10%, 20%, and really put a unique spin on something, I think that's also really interesting.

So I do subscribe to AX, but I think there's also that next level human experience that I'm very excited about. Yeah, they're both interesting. You know, on the agentic topic where my mind goes is, will agents value aesthetics? Will they have their own definition? Because you could argue

that something that is aesthetic or appears to be high quality. There's information in that that we as humans value. We make decisions, at least in part, accordingly. What will be the analog for agents, if anything at all? I don't know.

John, what do you think? Speed, clarity, things that, and verbosity, and all the edge cases. So the best example I like is the one of, you have to have a really good dash-dash help. You have to have great error messages. But Paul is correct in that the ceiling raising can occur once we can wick away all the generic good quality design.

And that's going to be extremely exciting. It echoes the era of the arts and crafts movement of the late 1800s. When machines took over, humans developed this next level of craft. And so that will occur. And I think tools like how Paul is thinking of can enable that. At the same time, from a functional perspective, there's so much runway on the AX side that I believe that primarily visual design thinkers will be able to crack that faster

because there's so many gnarly problems in it that an average engineering mind will not be the right-shaped mind, I believe. I think that's actually a good point, if I can respond to that out real quick. I think a great example would be API design and also like CLI design to your point, right? Do you mention dash-help? But I'm hearing this over and over,

and I agree with it. Agents themselves are not really good at designing those interfaces yet. So they're not really good at designing API interfaces, designing CLIs. And that's kind of ironic because they're the user, for them now. But I do think that's right. I do think information architecture

is another great example, right? It's a certain type of designer that's very good at crafting a good navigation. And they have to think differently. Yeah, absolutely. Paul, the design of impeccable is example to me of great sort of computational craft

for people who understand how to use manipulations of agents and code So that's why another thing I admired about your work, it's very computational. Really interesting, but do you guys, is it an instinct problem then in terms of raising the ceiling or is it a tooling problem or is it both? Because if you said, you know, John, you said

that the brain isn't quite shaped the right way, but we also have a tooling gap. So how much of it is instinct, how much of it is tools? I would say it's probably a combination of the tool. I think there's definitely a lot you can do with manual labor and manual craft that you still can't do with AI today or with current tooling, I would say.

I think there's some CSS that I still manually write because the agents aren't good enough for edit yet. But that's usually things like motion. So right now there's no good feedback loop for motion. I've built this shader library in the past. And I had this one shader that was animating rain upwards after working with the model for a bit. And it's not the models for it.

The models just take screenshots. And it's like, well, I guess that looks like rain. But it doesn't understand temporal resolution. Now Gemini does a little bit more than that. It samples frames on a video with sort of like one frame per a second. But even that doesn't really give you the full picture. So it's still really, really limited. So there are areas of design like motion that are still super limited to what we can do today.

But I also think the bigger problem is that it's not that we can't build tools to accelerate those things. I think we totally can. It's just that we haven't yet. I think there's a lot of tools that can be built, can be explored here in the space that really target craftsmen that want to go beyond,

that really want to raise the ceiling. So I think right now it's a pretty barren landscape for tools that really push the envelope. I would say Suno is actually a great example. are something that started for, you know, beginners and sort of vibe sound designers, right, that wanted to just one shot a quick song. But now it expanded very much into real music production.

But there are producers using it for all sorts of things now. They've built a whole studio around it, and they are really trying to push sort of the last 20% now, which is interesting. So they went through a big shift there. So I think that would happen more and more. I think right now it's still quite manual. Really interesting.

John, do you view instinct versus tooling? Well, as Paul was saying that, and thanks for all these questions, I don't think about these meta questions a lot in my job. Well, first off, the vision models don't really see the way we see. And there's not enough training data to help a vision model think like a designer can see. Because there's so many dimensions in the latent space of a visual mind still, create a visual mind versus looking for license plates

and more functional things. I think that also the number of people who will raise the ceiling will not be a high number and it won't be that valuable in the mass market. So I think it'll never be automated.

And that's good news for people who want to stay in the craft of this kind of work. I feel like right now, we're just making it easier for more people to produce average and above average work, which is the lesson of task-based automation. But those people who are dying for the creativity to last,

it'll always remain there. Only problem, the customer base will be smaller for people who are in that. It's like, do we know letter press, letter press printing? Oh my gosh, I love it. But not many people will pay for it. So it's that market mechanics. Yeah, it's really, really interesting.

And one of the things I believe, you know, Paul and I have spoken about this, John, is that we're going to have a lot more products with small but still compelling markets. You know, the million-dollar TAM products, the $100,000 TAM products, because they were economically infeasible before. So we can sort of imagine a world where it's more of a digital main street that can or a digital homesteading that can support a product that may have smaller scale, but be more opinionated and be that last 10%. If I may add to that, the thing that I feel over and over is that the reason that we're going to be a little bit of the reason. why someone's willing to pay that much more for that bespoke better thing is they're paying for human trust and accountability. So

for brands and products where people demand that level of like, can I trust this? That human ability, that human smell could become even more valuable than ever before. Yeah, I love that. Actually, Don, this is a good moment, too. I want to read a quote, this is you, Forbes 2010,

which is, the software industry is poised to embrace its craft heritage. A 2020 software will return to a cottage industry with bespoke apps made by many. We will discover the value of authorship. I mean, it seems like you might have been a few years early in your prediction, but it seems like we're getting there now. I mean, is that how you see the world playing out,

many more small apps that are all sort of more unique and opinionated? Absolutely. I was five years incorrect on that one. It was a hell of a prediction. But, no, I mean, that's something I learned by working with non-computer based designers and artists when I led Radan School Design. That was after a career at MIT, where it was all engineering, and then back to the handmade.

And I was struck by how much human emotion, human pride lives in that community. And I felt it would have to surface through in the software industry. But as a longtime proponent of computer programming for artists and designers, whether through processing or scratch, it was too hard. for those people to engage. But now it's that that floor of getting in has been lowered. Everyone's coming in. It's a really great time right now.

It's also very confusing. So that's why Paul's approach is unique in that it's high quality and high thought-based, which gives hope for more people to not choose the wrong solution. Paul, do you think there's an opportunity to start to shape the design mind through something like impeccable?

because it feels like it touches on a part of the solution that no other design tool does. That's a great question. I think for sure it can. And I think it can shape ironically or interestingly, both the design and the engineering mind. So I've had people walk up to me and say, well, we're now using Impeccable as like a default thing

that we installed on all of our engineers and designers' computers. So why? Interesting. And it's because they now communicate much better after using it. And I thought I was really interesting. Engineers are using design language more. Designers use engineering language more because the tools by using this careful vocabulary

and approach actually teach you what works, what doesn't, give you the language to express what you want and what you don't want. And so I do think there's a teaching moment here. And I learned something by working with models all time. I've always prompted that way, to be honest. I always wanted to get more out of models by having it test me and challenge me

and sort of like teach me new concepts. And there's some cool skills out there, like Matt's grill with dog skill, for instance. But impeccable does that in some ways too. So I do think there's a lot of opportunity there, and that excites me. I think that really touches on a topic that is dear to my heart

and was part of a lot of discussions that I was in recently, and that's cognitive surrender versus cognitive delegation. I don't know if you heard those terms, but really cognitive delegation is, you know, I'm using Google Maps, right? And Google Max tells me

where I can go as quickly as possible, right? I mean, I'm going to root my way to something. But what if I let Google Maps decide where I want to go to. Now I kind of like cognitive to surrender to the application. And I think that's very true for LMs if you're not careful.

Like with LLMs, oftentimes, LMs, for instance, plan mode, right? You sounds like a good idea. You prompt something and then it creates a beautiful plan and then it's eight pages long. You're not going to read through that plan. Right? I mean, you're going to scroll through that plan

and skim it and it's like, I guess the model knows what it's doing. and then just click, okay. And so now you kind of like surrendered yourself to the process. And there was a great conversation with a professor from Harvard and a bunch of others in a recent conference that I was at where they touched on, you know, how students are now working differently and how that is becoming a problem. So anyway, I think cognitive delegation is great.

I think delegating to the model is great. but preserving that point of view, making sure that you're still the one driving is also really important. And I want you working with impeccable to be the sort of game that back and forth where you're both active collaborators. So Paul, question for you on TACE.

Right now, automating things away, goals and loops, multi-Asian orchestration, that's what everybody is talking about. But the question that nobody seems to be asking is, where does it all lead? Are we sort of, you know, as you said, surrendering human intent? And what is the role of human intent and taste more broadly in these products we're creating?

Yeah, I think it's an interesting question because right now everybody's trying to solve or figure out whether models, whether LLMs have taste, right, as well as diffusion models. And I think the answer is a little bit more complicated in saying yes or no. I think these models have millions of definitions of taste. They are trained on so many different interpretations of taste. But one thing that is very useful to know and think about is that LLMs have been trained on the output of humanity, not on the input. So what led to a design decision is not something that the LLMs know. So it's an approximation of taste, I would say, for a certain type of audience at a certain time.

So you can get pretty far if you know what you're asking for. But then the question is whose taste is it? Oftentimes, taste is rare and human. And I think the way we talk about taste when we talk to humans, it's not something that can be easily recreated or spread to lots of different instances. And so I think can the models be tasteful? Yeah, they can design tasteful things. But I also still think that something that has a human viewpoint goes far beyond

And so I think amplifying human taste is an interesting problem, a more interesting problem than creating a replication of taste at the model layer. Yeah, it's really interesting. If you sort of substitute the word taste for humanity, it sort of carries the same meaning, perhaps even a more powerful way. John, do you think taste is innate or is it something that's cultivated? I think taste is always cultural. and different cultures have, quote-unquote, higher taste,

specifically because they've usually been around longer, they've had the material around longer, and therefore they've had maturity. For instance, if you go to Denmark and you sit in a furniture, any type of furniture, like chair or sofa or table, you're like, wow, this is nice stuff. When we walk around the U.S., it's like, I mean, I'm sorry, but not that great sometimes, because the U.S., we invented the styrofoam plate, for instance. You go to Japan, you know, centuries evolved approaches to raw materials,

and not only that, scarcity of material. When raw material is scarce, we tend to make it more precious and design things well, taste emerges through scarcity. The last point on this is we have to remember that design in the European sense came from royalty and the desire to be distinctive. I've got fur, I've got gold, I've got emerald, I'm important. You're less importance because they were working with scarce materials. I think what's interesting about this era is that this idea of taste doesn't fit

when all the materials are available to everyone. That's a really great point. Yeah, that's absolutely fascinating. Here's a comment from your recent blog, Jana, that I think I'll read. So the hard truth is that in rooms where decisions actually happen, taste as a whisper and velocity as a megaphone. It's almost impossible to defend a design nuance or instinct against a looming deadline. Engineering isn't the villain here, incentives are.

We build organizations that reward the measurable over the meaningful. Agree, disagree? How do you see that show up in the rooms at your end? I mean, it's all about the leader, right? We mentioned Steve Jobs, a unique combination of design, business, technical sense. That's very rare. And I want to go back to Paul's stuff and not to embarrass them again.

So what's interesting about impeccable is the design is very interesting from the API perspective. And Bill Atkinson, the engineer who recently passed away, who worked on the Macintosh, specifically people may know this thing called Quickdraw, which was Apple's graphics library.

Photoshop happened because of the design of Quickdraw's API. It was impossible for there to be a Windows version of Photoshop because it didn't have the same API shape in DirectX. So I would argue that Paul design engineers are able to design the shape of the APIs really well. So they can work with agents as well as they can with humans. It's completely fascinating.

I had no idea. What was the sort of unlock in the API surface that made Photoshop possible? Oh, my gosh. So the Macintosh, Pictra, had this routine built upon this idea of regions. Everyone had polygons. The Mac had regions based on pixel shapes, which you could like do logical set or and. And so things like Floodfill were super-duper optimized for speed.

So again, great APIs are not just shape but performance. That's super interesting. That's absolutely fascinating. Paul, how do you think, you know, maybe advice? for our design engineers in the room, how have you in the past effectively kind of communicated the value of an instinct versus a deadline or even another instinct,

which may be less entombed? Yeah, that's a tough one because certainly I've failed many, many times on this one. And I've developed on strategies, but oftentimes it's about getting a leader into the right mindset to make a decision that I would make, right?

What I mean by that is, and John is right, I mean, if you have a lead who's mostly focused on economic outcomes, for example, doing that week, right? Or mostly focused on performance. Ultimately, I think it's your job to present to them in a way that makes them dream the way you're dreaming

about the future. And that's something that I learned the hard way over and over and over. you have to bring them along to where you want to go and get them on the same emotional journey. Maybe that's the right way to frame it. And I think once they see what you see, once they see the type of future that you want to create,

then you're on the same page. Now that's way easier set than done. It's I think an art form to start with something that they need. and then subtly shift in that conversation towards something that you both might want in the future. I know that's maybe a bit abstract. So on that topic, I was having dinner like last week. You know, Jay Perrick is my boss, formerly the meta.

He's used this word that I hadn't heard in a while, which is conviction. And I think it's such a great word because someone, a leader with quote-unquote taste, has conviction. But it's not just like, you know, designing conviction, it's not just businessy conviction, it's not just engineering conviction, it's that combination in a leader when they have conviction. It is a bet, but it's a bet aimed at the global maximum,

not the local maximum. And so I would say leaders that are doing that global maximum conviction make a huge difference. And that's why product leaders, again, to embarrass Paul, I think Paul shot for a global maximum for design automation. And he hasn't given up yet. And that's a beautiful thing. I love it. You know, Ben always says that the job of the CEO, the job of the leader is not to have the idea, but to recognize the idea when they see it and bet the team on it. Yeah.

there's also you know another Steve Jobs sort of anecdote where he's very well known or he was very well known for shutting down

people just to test them right so the conviction is both on the leader and on employee interestingly because very often somebody would walk

into the room with Steve and Steve would just test him and say like you know I think that's a stupid idea even though he might think

that's a really great idea But just to test them, right? And you see, like, no, well, dude, will they just cave and walk out of the room? Or would they actually double down? Will they like, no, no, no, Steve? I think you're wrong, right? And once he got that signal, once he got that signal of conviction, he often would say, yes, let's do it.

And I think that's fascinating. So I think what John said is really interesting. I think it goes both ways. Oh, my gosh, that's an perfect example of LLM as judge. We have to, on both sides, on a general. of conviction and the feedback of conviction. You need both, the anti-matter matter matter. Love that.

I love that. Well, you know, we'll record this pod again in a year and we'll see if all of the software that's been created through the new Git product with assistance by impeccable turns out to be more thoughtful and, you know, in that upper 10%, and that would be a fun outcome to see. John, accept it. Incredible.

John, any closing thoughts for us? I, after picking up GitHub design, I realized how so many people grew up at GitHub design who were computational designers slash engineers. There's Macs at Notion who used to lead design at GitHub. There's all the Bursale people. They're all over the place. And I was asked to somehow bring back the franchise. And so I'm excited because Paul's work represents that kind of seminal,

design history-infused approach that can really make it easier for more engineers to design reliably and also more designers to be able to reach for the skies. I love it. Paul, any closing words? Well, I'm just super, super excited

to see what will happen once we've raised the floor and what insane, crazy experiences will build afterwards and design afterwards. That's the kind of future that I want, as opposed to the bleak future where everything looks same and people feel like they've surrendered themselves.

And so I really am excited about that new era of craft. And that's something I'm working towards. Extraordinary. Well, cheers to a new era of craft. Thanks to the good work from both of you. Thank you guys both for being here. Excited to see the product out in the wild and play with it ourselves. Thanks so much for having us. Thanks for listening to this episode of the A16Z podcast.

If you like this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our substack at A16Z.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. taken as legal business, tax, or investment advice, or be used to evaluate any investment or security and is not directed at any investors or potential investors in any A16Z fund.

Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures.