How context determines AI success. SEEK AI talk

Lara (Senior Product Designer, SEEK)

Richard, our next speaker is now going to give his talk and then we'll have a quick break before we open up the panel.

Richard Simms (Principal Product Designer, SEEK)

I'm glad that worked. It's always a nervous connection. So my talk's called Inputs to Outcomes. And before I get started, I just want to also, as mentioned earlier, pay my respects. It's a privilege to meet on the Wurundjeri country of the Kulin nation. And I add my respects to traditional owners and ongoing connection with this land. My name is Richard. I'm a principal product designer here at SEEK. I spent all my career designing and experiences, product experiences through Europe and APAC.

But I'm passionate about finding clarity through continuous discovery. And I want to share today with how AI is helping me do that. And so what I mean by that is your team already uses AI, and you already fall into one of these three buckets. Each produces a different kind of output and vary in the different definitions of success.

Richard Simms (Principal Product Designer, SEEK)

So you're either vibe coding, which gives you speed, but you're letting AI own the output. You're in the middle camp, which is AI assisted. It gives you consistency, and it helps you accelerate your breadth and your depth of what you're able to do as a designer, or the third camp. Forgive the name, I don't have a good one, but more spec driven. I'll explain it in a second, but it's where you're working with a shared context, where teams and AI can use the same information, which is what helped me in my journey.

I don't wanna share. with that with you today. Again, I said that twice. So let's just back up a little bit.

Richard Simms (Principal Product Designer, SEEK)

So what is vibe coding? So it's purely, vibe coding is purely generative. You describe it builds, the metric here is speed, how fast can I get something working? It's perfect for experiments or internal tools, but the context is shallow. It can be used, use a design system to make it look right, but it doesn't understand the rules of the componentry. AI can be used great outputs at speed, but it's without the strategy, and it doesn't align to our outcomes.

Where assisted AI, it's like enhancing your traditional workflows. It's kind of where most of the teams, where you are most of the teams are today. You still own the output. AI just helps you go faster. It's writing the copy, iterating, summarizing the research reports. The metric shifts from efficiency to efficiency and quality. It's incremental, but it's not transformative because context still lives inside the individuals and it's not really shared across the teams.

So this thing that I've come up with, through the lens of discovery, it's kind of collaborative intelligence. It's similar to how engineers mob in their pair programming. But instead of doing it with the, instead of mob programming with code, you're mob programming with discovery. When the triad comes together, the PMs, the designers, the engineers, and they reason together with an LLM using the shared context.

Instead of in the middle version where you're bringing your own LLM and you're empowering yourself, this is where we're empowering the team. And so the team owns the output, it reviews the content the AI generates, and AI just becomes a way of an extra teammate. Doesn't replace the thinking, it just helps you scale it. We measure success here by decision quality, learning velocity, business impact, and it's where output turns into outcomes.

But which workflow makes sense, really depends on the product itself and the lifecycle that that you're in. When you're in an early stage, you're exploring ideas. Vibecoding makes sense. It shines. It's the thing where you can describe something in natural language and see it working in minutes. It's perfect for going broad and iterating fast. But as your product matures, complexity creeps in. There's more flows, dependencies, APIs, architecture decisions. And now AI needs more context to reason effectively.

And that's where we shift from AI as a generator to AI as a collaborator. Using this bit driven workflow, ensures that every new idea fits the system and it aligns with the strategic goals. So AI role changes from a product as the product evolves from maker to accelerator, co-pilot in that decision making process. But the challenge are tools of fragmented content islands. Mirrorboards, decks, docs, research papers, they all need to be machine readable.

If AI can't see across them when it generates its output, it's operating blind. It's kind of why most AI feels like it's missing the nuance, because it can't see it.

Richard Simms (Principal Product Designer, SEEK)

So why content, why content, why content drives output quality? Think of context as a multiplier. The more context the AI has, the business goals, the research data, the code, the better the quality is. The less it has, the more you rely on the look at the prompting. It's context, not capabilities that predicts the AI success. In my eyes, it's machine readable discovery items, like treat discovery artifacts like code.

Everything should be marked down, versioned, queryable, like PRDs, opportunity solution trees, hypothesis. When human and AI can share the same artifacts, we can both reason together. You preserve that institutional memory of what, and you make everything that you interact with in AI context aware. So I've got a couple of examples. This first one is how we've applied this thinking at SEEK. I work on something called the career feed. It's on the homepage of the logged in experience.

And we want to help candidates improve their job recommendations. We vibe our way through some early prototypes, explore how AI can interact with the candidates in different tones and formats. We used AI-assisted, like a junior designer, to generate the first ideas and stretch our thinking. We mobbed together with an LLM, bringing product design and engineering together to craft the prompts, the question logic, and real customer insights and how the AI recommend the models work.

So it had all the context that it could help us form good opinions about the output it was creating. And then this, this is my side project. This is an AI audio app called Speezy. It converts and summarizes newsletters into a personal podcast feed. It's got weekly digest, and I really just wanted to see how far AI could help me go. I started in Replert, just vibing, describing what I wanted, letting the model just build it. The goal wasn't Polish, it was learning fast.

How, like, what APIs, how they worked, where things broke and what was possible. The kind of things I figured out with this app was like how to scrape articles from URLs. How to handle text-to-speech limits. How to split long articles into multiple sections so I can then stitch the audios back together again. How to put it into an RSS feed and publish it. How to credit the author and make sure the sources were added automatically and that the content was right.

I built an automated workflow so users could get the new content pushed to them via email and it was personalized and updated based upon their preferences. And on top of that, I built a Striped Checkout flow so people could pay me because why not? I wanted some money. But those early messy prototypes were just a way of me working out what the spec was. It was all throwaway. I just understood what the system needed and it was text in, audio out. It was just basically me understanding what I needed to do.

So as I moved into cursor, I threw away the replet builds. This is where the structured version with the real content really shined through. Because I knew what I needed the product to do. I knew what I needed it to make. I had co-written all of the documentation along the way. When it came to the development time, it was quite straightforward to let the AI help me.

So, if I gave me the directions but spec driven, whatever you want to call it, gave me control and that shifted from more of the exploration to the making. And this is where I started feeling more like a real teammate. So, we build what's now becoming like a unified layer, connecting the research analytics, the semantics, and the product strategy together. It stopped becoming, AI stopped becoming this black box mirror and started becoming how well the team understood its own systems.

So there's four things to kind of take away from this. It's AI is augmented, it's not replacement. Context is the new architecture, collaboration is where judgment matters, and measure AI in real world outcomes, not by how much you can do. And we don't really have an AI problem, we have a context problem. For me, AI isn't just another design tool on the toolbox. For as long as I've been around, the tools have come and gone, whether it be Omnigraphil, Photoshop, Sketch, Figma, Framer.

They're just all ways of communicating an idea. AI is kind of the same, but it's now letting us actually design as play with the materials that we're designing for. That's me. Thank you. Thank you so much, Richard.

Lara (Senior Product Designer, SEEK)

Okay, I am going to ask everyone to come back and take a seat, please. Grab, you can bring your drinks with you, and we'll get started for the second part of this evening. I'll give people a moment to walk over. Thank you.

Lara (Senior Product Designer, SEEK)

So we have already heard from Richard and Nina, but I would love to now invite our other panelists to introduce themselves. Just Ben and Angus, you're happy to tell us your name, your role, obviously I've just said your name, your role in a quick sentence, just about like how AI shows up in your work, that would be great.

Angus Tate (Product Design Director, Zero)

Thank you. I'm Angus, I'm a product design director at Zero. I manage a team of designers. We are, well part of my role is scaling AI tools across the business, across the design team. And we use AI in pretty much everything in our work, from writing to designing to planning to everything we can think of right now. And we are grappling with the same challenges that everybody else is talking about as well.

Ben Baron (Principal Content Designer, Canva)

Thanks. I'm Ben. I'm a principal content designer at Canva. Been there for a little over a year before that I was at Google, before that I was at Netflix in Meta for a while. So done the big tech thing for a bit. Google, I worked on Gemini, so I'm happy to share a bit about that as it pertains to some of these chats. At Canva, we use AI a lot, although I don't use a whole lot in my daily work because I've spent a lifetime building an AI product for Canva. So if you use Canvra AI, you've probably seen that. But yeah, excited about this.

Lara (Senior Product Designer, SEEK)

Amazing. Oh, amazing, thank you both.

Lara (Senior Product Designer, SEEK)

So to kick us off, I'm gonna ask you something simple, but I think important to align on. What do you think we mean by AI output? And what does good look like? Because AI output can mean a lot of things, sort of a sentence, an image, a recommendation. When you think about your work, what counts as AI output to you? I will pass over to Nina maybe to start.

Panel

And feel free to jump in if you have something you want to add to the conversation. Did you not listen to my back? Yeah, yeah. Yeah, so I think pretty much anything generated by AI, whether it's text, image, voice, these days, yeah, that's that's output, right? And that's something that we need to measure to make sure it's good.

And good, well, it depends on sorry, that's that was going to which is a ongoing joke with every question that we talk about as AI is it depends so yeah but that's something that I you would need to define before you get started and it would be different for every situation whenever go rich yeah no that's the the bad joke but I think what defines the output Sorry, the output of AI, I think depends on where we're talking about the first draft of something that helps us iterate and get faster on Mooses through a design concept, a copy, something like that.

But we're also talking about the output that would, for me, when we talk about the career feed or talk about the side projects, is what the end user is interacting with and how much rigor we've put through what they're going to touch and whether it actually then moves a business metric. And we're measuring everything via the same things we do five years ago. It's by the same metrics because we have the same business and we have the same business goals.

So what we measure AI output by A depends on where it is on the flow, how close it is to the output of the product. It's measured by the same things as everything else by the product. I think it's pretty simple. I think it's an output is there are really two things that matter to an output. Is it what the user wanted and asked for? And does it solve the problem that the user wants it to solve?

And I think sometimes we can get a bit lost in, like, can it generate, like, a video, but with some more scrutiny, sometimes that video is not great. At the end of the day, AI is only going to be successful and stick around if it does things that are helpful. I think in terms of, I think less in terms of output and more in terms of like does this thing solve a problem.

But in terms of that experience that you have with it, which I think is part of the question too, I think less of it in terms of just an output and more about a conversation. A lot of content designers spread out into this, this discipline called conversation design. In my old role, that was my official title as a conversation designer. And a lot of what we were thinking about was the output, like how does the model respond, also what does a good conversation look like?

Excuse me, how do those responses adapt and change the context? How does it reflect different moments back to users? How does it remember things, how does it feel natural, but at the end of the day, how does it solve a problem? Anything to add, Angus? The one extra thing is going to add, we talk about output in very similar way. It's zero, output. Everybody's really familiar with what that is.

The next part of that for us has been more of an agent-based workflow where the AI is doing the task for you and the output is not a generative output. It's not an LLM anymore. It's an agent. The time it takes to do that task that you can do yourself could actually be longer than you doing it, but it's taken away the toil and you don't have to do it, but you may need to verify that it did it correctly. Great answers.

So we touched on this a bit, but when we think about the output and I guess what does good look like, there's often a gap between I guess what's technically impressive versus what's at the boot versus what is actually useful or human. So in your view, what makes a good AI output not just an accurate or a fast one? I think increasingly good output is cheap. I think that I'm sort of not, it sounds funny, but the cost pressures of AI are going to come down very hard on businesses.

If things are being run at a loss, there's a lot of compute being burned just on tasks that probably aren't delivering a lot of value. So at a certain point we need to be figuring out like where does that balance between cost and latency and quality come in. It's a tricky question to ask. Again, I'm not an engineer, so I can't really answer it. But that's going to be part of what good looks like.

And I think that the more we can get ahead of those questions and have frameworks to think about it, especially in eVouses and even when we're talking about it, the more we're actually going to be able to build these more durable products. But I think in terms of good and from a qualitative perspective, good looks like does it solve the user's problem. And that sounds kind of hand wavy, but it's true. It's like, what is a good UI look like? Does it do the thing that is right?

Is it in the, does it show the appropriate brand qualities? Same goes for any kind of generative output too. But because it's generative, you know, doesn't solve that problem. So big question, not a great answer, but yeah. Oh, I thought it was a great, I thought it was a great answer. Anyone else have any thoughts? I think when we think about what is good output, it matters a lot what domain the output is in.

And if it's your area of expertise and you're the master of some, if you've mastered some sort of area and design is the area that we are all in, when you think of what is good, you can see when it's not good and you can see it straight away. And I think we're all becoming more and more exposed to AI outputs just out in general. The more we get exposed to it, the more we detect it and the less satisfying it's going to become and if it's in our domain we're going to see the difference.

When it's not in your domain and we're doing this a lot now is sort of boosting your skills in other areas if you're outside of design, especially if it's let's say like product marketing or something like that where you're using an AI tool to boost your own lack of skills, you might not actually recognise whether it's good or not. It could actually be really bad, but it's better than what you've ever produced yourself, and therefore you think it's good. And that's probably the day. with where we are sometimes.

That's a really good call, I think. It's like that balance of expertise. There's a thing that I've observed a lot in working in AI products for a couple of years now is this paradox of expertise, where people think AI is amazing at everything except what they're an expert in. And you see it, like there's, does everyone know Mark Andreessen, that venture capitalist? Okay, see some nods.

Big venture capitalist, billionaire, funded Facebook, all this sort of stuff. there's this clip of him on his podcast from a couple of months ago saying that like AI is going to replace every job except venture capitalists. Like verbatim almost. And he runs head first into the point without realizing it. And the funny thing is, is he then goes on to say, because a venture capitalist has to be, you know, a therapist and a business person. Like you're just listing jobs, dude. But this is the thing. It's like you need that.

Great point because it's like you need to remember that We're not experts in everything so we need those frameworks valuations for applying that expertise Building expertise in areas that your products going to be working in What we've been talking about quite a lot when it comes to good and AI is taste and That's something that you develop as a designer or a writer through many many hours doing the work And therefore when AI produces something, you can look at it and go, okay, that's good, or that needs a bit of work, or that's just rubbish.

And it's kind of, and that's not something, and I think it's exactly what you've said, right? Like if you're not an expert in it, then you think, oh, whatever it's producing is better than what I could produce on my own, and therefore it's good, but that's because you haven't necessarily developed the taste that an expert has developed. So in this situation where we're in now, where AI can not deeply but very easily generate all sorts of things, are we then eroding our taste, right?

Or how do we kind of preserve our taste and will teach people taste so that they know what's good and what's bad? Yeah, like we're not ready to outsource our judgment that's come from the years of experience that we have. I'm sure everyone uses LinkedIn here. It's a great website.

But I think there's a lot of people on there who seem to have just discovered movies for the first time, where they see like the worst piece of AI generated video known to man, it's like an alien one with 12 fingers on each hand, and no character or plot, and they're like, Hollywood is dead. I'm like, all right, dude. We haven't seen a movie since Avengers 1, and think movies are dumb. But this is the point of taste. It's like we need these to be out there, right?

And I think that is, if we're not self-aware about AI, then we're just going to keep poisoning the well and pissing off normal people who are like, I'm sick of being told AI is great because everything I see is crap. It's like, let's have some humor. This is a general point, not about anyone here. And backing me up like, taste is really important. Yeah. I think these have all been really great points.

And I think just to add some flavor to it, I think this is where we shouldn't be shy of getting involved as designers into the prompt writing. When the output is affecting our end users, we need to bring that taste, we need to bring our expertise and collaborate with those who are actually writing the prompts and be the one that owns the experience.

If you're an experienced designer, own the output of it when it affects the end user, collaborate with the others to make sure that our taste is there and present when it comes to actually being part of the conversation. Does seek designers not write prompts? I do. Oh, you do? Good. Okay. Good. Thank God. Well, that's a fantastic so great because I was going to ask now about metrics.

Panel

So we often look at, I mean, the common ones that everyone looks at when it comes to AIs, things like accuracy and speed, but they don't necessarily show how users trust the experience or so how do you find the right balance between, I guess, the numbers and the human side, the balance between, I guess, what's technically right and then what feels right to your user? your user. I mean, Richard, I'm going to pick on you because you just talked about it. Yeah.

How do you understand if like, okay, you've got all these metrics in your framework, but how do you understand, like, is this actually a good experience? I mean, I guess it comes down to a lot of what Nina's saying. It's taste. I think a lot of it doesn't change, right? We're We are still trying to deliver products and we still have business metrics. I'm a product designer, so I still have to adhere to the metrics that shift the business. I still need to make sure that we're hitting our business goals.

Yes, AI might be helping us achieve that and the output of that might be part of it, but it still needs to be moving in the right direction. Yes, cost and latency are great constraints that we have with AI. We need to make sure the models are fast enough that they're not adding too much latency and they're cheap enough that we can run it at scale. But we need to make sure that we're building towards the goals of the company that we've set out and I think that has not changed.

I think that's still the goal that we need to do is make our businesses profitable. Yeah, we're similar. We're measuring things the way we always have with additional metrics. as well. When we first started working with AI, giving AI to customers, we did what most other companies did. We bolted on a chat box to our product, which is what a lot of products were doing. And of course, you start measuring like number of prompts and what are people prompting and all that kind of stuff.

And our tool is called JAX, which stands for Just And the number one prompt was hello, Jacks. And that was one of, that was the total number of prompts that a user would enter. And that would get the response and then they'd leave. And following on from that, when we started to investigate like what else is going on, what do people dissatisfied with, a lot of it was the prompt returned an inaccurate answer.

Most LLMs, like the models around today, one out of four answers is correct and the other three are confidently incorrect. So the LLM will either hallucinate or tell you that it thinks it's right and you have to be the expert to know that it's not right. We're in the accounting space and you can't make mistakes like that. We have compliance deal with, we have businesses who are running the whole business on our software.

So some of the other metrics and some of the other learnings from that, we obviously opened up a big wide chat bot and you can do anything with it and it's hard to control. The success is happening in the narrowing use cases. So one example is a really great designer in our team, her name is Darryl Powell, and she just presented this piece of work from a little while ago where the idea was to help businesses get paid faster. So it's a metric that we've always cared about.

Like every small business, they don't get paid fast enough and it really impacts their business. And we measure time to pay. So when you issue an invoice to when you get paid, it should be 30 days, but it's often 90 days, 100 days, whatever it is. In our research, we're seeing that a lot of our customers feel a lot of anxiety about going to that person who hasn't paid them and sending out that message. So Darrell had this AI use case where they built basically a letter generator.

It basically will give you a message to send to this person who owes you money. And we have a lot of data around what language works to get you paid faster. And for example, in the US, you have to be really direct. In Australia, you have to be a little softer. And if you get the tone right, you actually get paid faster. And we have the data to show that. So the long-winded answer of this is like the AI use case here, we're just measuring time to get paid. That's the thing that the customer cares about.

And we have this AI opportunity now that will help them do that much faster and get over that anxiety and run their business well. That is such a great example. Have you seen the Bunnings AI? No. Let's see it talk to Jack. It gives you no end of joy that Bunnings has. Bunnings AI. Tell us. I just saw it on the website, it's useless. I think it's just a GPT3 wrapper that just talks about products.

But I think, I do think it's, kind of we're talking more about the, I'm in the Gen AI group, but a large part of our role within business is to push on teams to justify the use of AI. So I think sometimes people like, I'm going to use AI because it's going to look great performance review. But a lot of the time, it's like, no, that could be done much better with just like a script. You don't need a model duct tape to your thing. You can just, there are better ways to do it.

So the framework I like to think of is that AI or an LLM really, you just need to do something either new, better, faster or cheaper. And if it can't do any of those things, then it don't use it. And if it does do one of them, then prove that the user value is going to be retained. Because at the end of the day, I think we end up competing with our own products, right?

And I think sometimes, especially you can take Canvas as a good example, it'll generate you designs and it does it increasingly well and I think that's a good thing and we should be doing that. There was a period of time we're still in it where we're competing with a template library, right? Like what's easier? Is it easier for you to go and search for template and just design? Or is it easier for them to generate something that may or may not be right? That's kind of what we, that's the paradigm we're in.

So going back to user problems that you've been talking about and just asking questions about what it actually does well.

Panel

amazing so I might open it up to the audience now does anyone have a question for our panel

Audience question

hi I'm just curious you talk about taste and obviously about you know the app that needs to suit what user needs right do you think there's a role there sure can you like I think we have to use feedback to inform the quality of the product. Go-dansel would be yes. Yeah, it's in like... It's getting like direct response or like gathering data. It's pretty much a Q&A, so how do you think the quality of something that's been generated has been well-meted to systematically improve it?

Panel

I don't, canvassily not the level where we're collecting live feedback like that. And I don't know enough about the data that we do collect to say, I don't wanna say anything except there's some canvass here. It's gonna get me fired. So, you know, the press, I know you're not the press, it's all good. Yeah, I think that like in pretty effective ways, just to like see what good conversational resolution looks like, right, comes back to say to your conversation side. does the user go and do a business metric, right?

It's like, do they do the thing that you expect them to do? As a result, publish a design, create a design, edit the design. So, sort of clumsy answer. I think there's a range of ways that you can do this, right? So I remember when, because I worked on the job description with Lara, and what she did was she scheduled user interviews, and then she found out what role that person was, and then she generated role descriptions in advance and showed them to that person. Right?

So I used Nina's excellent and Suan's excellent prompts and then I took the job ad that they had most recently hosted and then I showed them that AI generated, our AI generated version of that role to get feedback on the, what they thought of, I guess the quality of the output and then used that feedback to help improve things. So that's like your low-fi version if you can't do anything else. But I know for career feed, your data scientists have built an online thing where we could go in and evaluate, right?

But you could show that to us. Yes, we did the evals for the recommendations on the home page. We went through the ability to test evals. So we had the model produce the output that we wanted to show the candidate. And we, as a team, went through that and gave it a thumbs up, thumbs down, said yes, this is working, this is not working, gave feedback. We had multiple prompts and multiple LLMs to test that against as well.

So we could see which LLMs were producing the better responses and which ones weren't, and then we crafted the prompt through various iterations of going through different prompts to make sure that the output was great. What we've done on the output from a candidate side is also we've given just thumbs up, thumbs down. You see it on all of the LLMs as well, right? Just giving the user the ability to say, This is crap. This is good.

And then that enables us to know if there's some mistakes with the inside out models and we can gather that feedback. We've done quite a few different LN experiences where we've got a thumbs up, thumbs down, and then we're asking for free text as well to capture inputs. If there was a problem, we can then go and look at it and try and capture why it wasn't right for that person. Yeah, it's really helpful if you have good tracking as well.

So like you can see, for example, for us, We have a lot of different types of hires. So if a small business owner is sort of like recruiters or government hires, like if they're all consistently like thumbs down on a certain thing, then like we can better understand, you know, if a problem is limited to a certain segment or if it's universal. Yeah. Does that answer your question? Yeah. Yeah. Anyone else have a question?

Yeah. the saturation point of AI, like it's in every product, we get to see it as users, we get to see it as product owners or designers, when are people going to start getting sort of sick of seeing this is built by AI and it's just become sort of the day to day, like this is just the product. What's the... I think it's already shifting. So we have a big conference, it's called Zero Economy, we get all the accountants together and they go wild and it's like the best time of the year. We run it every year.

I think it was only about a year ago we were basically doing a Vibe Check on what's to do with AI, how are you feeling about it, and the response at that time was it was fear, it was like I don't want them to do this, I don't want AI to come and take over. And we just had our zero con a couple of months ago and the tide had completely changed.

The more I the more I use it, the more I'm saying to get comfortable with it, and the more I'm accepting that it's around now, it's here to stay, and it hasn't taken my job, I'm actually getting more time to do other things that are more strategic that I know that the AI won't do right now. I don't know where it'll be in a year from now, but it feels like it's going to change so quickly that it might become invisible in the not too distant future.

There's a way to think about it that's always been invisible. the first time it's actually been visible, right? I mean, even though you define it, a lot of things have been powered by LLMs or SLMs. Machine learning is a pretty broad field. It's powered a lot of stuff out there in the world. So I don't know, I feel like this is a weird period of hype and visibility, but it will just fade into the background, for sure. I could tell you when, but I do feel like it's increasingly, you know, on us.

There's definitely a trend at the early part that you know wasn't it where you had to make it shine everybody had to show it It was like front and center. It had all the sparkles and now over time It's just like de sparkling just becoming part of the juiced evaluation with how about AI I really there was a price exactly and I'm so but the though I saw some iPhone screen protectors that were AI ready. It's like I'm not joking. It's like because it was ready for Apple intelligence, but it was the same screen.

I'm still ready for that. Well, yeah, I'm still waiting for that to happen. Richard sent me a photo of a slider that they showed some of my work and features that I've worked on over the last sort of year or so. also. And there's four different releases in each one. The sparkles and the gradients become progressively less and less prominent.

I feel like the jazz designer, whatever you want to call it, the gradients that are popping up in the sparkles are starting to decrease, but there's more UI around transparency. So what is the model thinking? And what are the steps it's going to take? And where are the intervention points if you want to stop it from doing something. It's not disappeared, the sparkles are going, but it's not disappearing, there's not less UI, there's more UI now so that you actually know what's going on and you start to trust it.

Again, that might not be there in a few years, something else might be the more pressing issue than the transparency which is a big deal right now. Question on that? Hi. I basically wanted to ask a question for you. I'm going to go back to everyone. I thought your concept of taste was a really beautiful, unique taste. I've had a whole look at self-improvement and look into this. But in the world that we're leaning into, where AI comes up across the web, it's become increasingly recognizable from human-made content.

How do you select and improve that taste where in the world, currently, they are slow as no one knows what is always good? I challenge not being able to tell what's AI slow. It's recently been invading my Google feed. Like, you know, when you're on the Google homepage and then it gives you like all these articles, recommendations and there are these like random websites with very random articles and I'm like stop showing me there I stop.

So I like to think of this in terms of I guess the first industrial revolution where things stopped being made by people and started being made in factories and And we're in this position today where we can buy very cheaply poor quality things that are very bad for the planet. And I don't think just because we buy lots of cheap, poor quality things that are bad for the planet, that we don't know it's cheap and poor quality.

We know what the good stuff is, but we are making a decision, whether it's a monetary decision or other decision to buy the for equality thing. And I think it's the same with AI. So we're supposedly in the middle of the next industrial revolution. I'm not sure. about that. But we can definitely see the en-certification happening, right? And it comes back to being useful.

Like if content is not, is not out there solving someone's problem, if it's not, you are not writing about something that is that people want to know about or want to learn about, then you're just creating crap for no reason. I also think there's going to be a reverse cost pressure there. Right, like open AI is burning what 15, 20 million US dollars a day on free sorrow? Tens of billions of a year, right? That literally cannot last forever. So I think the slop will be pressed down, right?

And businesses will probably be like, you know, it was free for me to generate 800 articles in five minutes using this model, but this model is now costing 10 times as much. And I also just think that like there's people still making stuff. I think artists are still going to create. I think musicians are still going to make music. Like the whole idea that the output is the only thing that matters is a fundamental misunderstanding of the creative process. And a lot of people in this room will know that firsthand.

It's the process of creating it that's where the joy comes from. There's a really good article in New York Times from three years ago. I'll find out I'll show you the Flora Social contributor. But it was an interview with a bunch of science fiction magazine editors who had been doing it for decades. And they were talking about how since chatGPT came out, they had to close off submissions because there was so much crap just getting sent to them.

Like, they went from getting 40 submissions a month to hundreds or thousands. And they could tell that it was AI. It was crap. It was clearly written by people who hadn't written read or written before. But because they'd done this, they're like, this is good enough to publish. But they figured out how to circumvent it and they're still publishing their things and it's not AS-istive. So there's still stuff out there. I think Nina made a great point about the Industrial Revolution.

I can find a really beautiful piece of handmade furniture or an instrument or something and I can appreciate the craftsmanship in that in a way that I can't when I see a chair that's rolled up production line. So I think it's just a little bit more of our own sensibility, our own taste in knowing how to avoid it that comes into play as well, just in terms of general consumption. There's a different sort of value in it being authentic as well.

I think we're starting to see a little bit of a new trend where people are starting to acknowledge that it's human-crafted and human-created, and that's the new spiral that gets hand-made. I think that will be the next. I also think that AI is just not, LLMs are not rational thinking beings. They don't, like I've read a lot of AI generated fiction and that sort of stuff. And the interesting thing is they just get words and analogies really wrong, right?

They try to talk about how something feels, but they clearly don't know what it is to feel, right? And it shows up in the narrative, it shows up in the words they choose, it shows up in just these like ham-fisted things that they do. Like you don't have to be like a literary genius to say that doesn't make a lot of sense, right? Like that's not how it feels to be under summer rain or that's not what it feels like to be. It just sometimes it just doesn't make sense.

So, and it doesn't, so uses so many adjectives, you can pick it from a mile off. So, you know, I don't think LLMs are taking any creative jobs anytime soon. Yeah, I put a lot of user interviews sort of synthesis into an LLM a while ago and it was sort of like seven out of ten participants at this and I was like, I only interviewed eight people. Who are the other two? Where have they come from? And anyone else? Any questions? Ah, sorry. Yeah? Yeah.

I got probably the other side of this, I got from a university lecturer. In the universities, AI could be considered like AI is bad, don't use it, I don't like it. So we're trying to adapt and trying to see like, and thinking for like future thinking of UX designers and thinking like, they have to be prepared to be using AI.

Audience question

So I'm wondering for you, which are the important causes that you see or that you would like to see in the new emerging designers that they know about the AI or how they would be better prepared to be able to focus on your work. I think that's a really challenging question in universities. We have graduates, graduate designers who work in our teams who have come directly from university.

And I had a conversation with one recently where I said, if you wanted to, you could use these AI tools and do half of your job if you really wanted to in half the time. But I said, like, the harder part you could go down is you could do everything from scratch the way we used to do it. Which would you prefer? What would you like to do?

Panel

And she said, I want to do it the hard way. I want to go through and learn design as it was before it became fast and cheap. And she had good intentions, and she probably lasted about three months doing that. And then after a while she realized, oh, I can actually do this a lot faster using AI. She's going to be a great designer either way. It's a fantastic thing to see her try out all these tools.

But it's going to be difficult when people come into the job market in a couple of years from now when those tasks that we used to do by hand or in person day to day, you won't necessarily have a role model anymore to see someone do that. And you'll have to figure out another way of doing it. I don't have the answer to the question, just like I share that same concern, like things are coming up and how do we deal with it? Yeah, that's something that keeps me up at night as well, right?

Because like I said before, like we've developed taste because we've done this, we've put in the 10,000 hours to develop it. So what happens to new designers or people just leaving school, we will have access to these tools, then they don't have the chance to put in that 10,000 hours and learn the skills that they need. Yeah, I don't have an answer to that either, but yeah, I think about it a lot. I think keep it simple, just like, just use the tools that are good now and don't try and predict the future.

I'm probably less convinced than most people that AI is necessarily here to stay. I don't think it's gonna upend design in the ways that we think. I think things like cursor are going to democratize all the stuff we can do. I think LLMs will help speed up elements of UI writing, but I just don't see a world where LLMs trained on established data are going to be designing entire products, I just don't see it.

And the fact is that the value of designer brings is not spitting out some pixels, it's in thinking about a creative way to solve a problem. Right? There's a really great Emerson quote, It is a philosopher, American philosopher from the 1800s. Anyway, he had this great quote. He was a Shakespeare scholar and he said, show me the poet that taught Shakespeare how to be Shakespeare. And that's the truth of the LLMs. They cannot do anything new. They can't. That's not an equivocal debatable thing.

They cannot go beyond their training. So they're never going to solve complicated new product and UI problems. They're never going to design the iPhone. They're never going to design the next iPod. These things are like beautiful innovations. So I think that they're useful tools around the edges that can help us do things faster. Absolutely. Cursor can democratize access to code, but keep things like really focused on what's going on now and don't be like, well, it's here to stay.

So we need to think about the whole new process. We don't know. And again, I think there are a lot of things that are going to change maybe that impede the spread of AI. And that's okay. Again, models may change, but I don't think AGI is around the corner. I think there's going to be creative problem solving done by chat GBT. It's going to be a tool. But then do you think that doing the grant work, I suppose, helps to hone that ability to do the, to be creative and to solve problems creatively?

Yeah, I absolutely do. I think, I think using a tool like cursor, you still need to know what a good output looks like, right? You talked a lot about that. So I think like the tools as they are, and probably going to be for a while. Who knows? I don't know the future either. But still going to be like, you need to know what you want to get out of this thing, right?

And that requires knowing what good design looks like, what good information architecture looks like, what a good flow is going to feel like, what good UI writing is going to be for that moment in the experience. So I think we as designers do have a, and I've been there, so don't get me wrong, I'm no perfect or anything. I'm just sort of sharing from my experiences that like sometimes we can get a little hung up on what's just around the bend.

I really think it's healthy to think about it but I don't think it's healthy for us to plan for an eventuality that we can't see that may never be here. So yeah I think to all the points here you've got to do the work, you've got to know what design looks like, right? Hi, what was your question? I actually have to do the one now for you to think on. Go for it, yeah. I'll try and phrase it into the question, so I can answer it differently.

So I like to go into new designers and think about it as a failure pass, a cheat, all of that's actually your point around great problem solving and your point around using the tools. We were just talking after your talk about how using like

Audience question

in some ways, it replaces crazy aches of an exercise, that's a bunch of ideas quickly. So can you, to turn it into a question. I know. It's thought to instances where sometimes like people are trying to use, like can you measure the use of AI against those rubrics of failing past even often? An example, I have one designer that ever since Figma make came out, who just does the vibe coding, just as the big man make prototyping instead of just looking the clicking events up together, but it just takes longer.

So can you talk a bit about where you see some of the limits are in it as they tool fail fast, you know.

Panel

I think sometimes in doing the rapid craziates using an AR tool and generating an immediate prototype, you're almost always working with tailwind and shad scene or whatever it is. And like the components are there, they're already designed. And you're just recycling them over and over. It's hard to break. It's hard to get the AI to break out of that box. And that's probably where it's big downfall is at the moment in that it's very fast when it's got the patterns there to use.

But when you want novel patterns, it's pretty useless. And that's where the paper and pen is still winning, or the whiteboard or whatever it is. However, there's a really good path to production that's emerging. And we're playing around with Figma, but not Figma make. It's more like, how do we use our components, like our real components? And how do we prompt and get them to build the prototype?

Not through Figma make, but using MTP using the I think it's called co-connect and having that two-way street where you can go from being in Figma to being in a browser and then pushing it back and making changes in your area where you're comfortable. There's still like an almost tangible thing about moving components around on the page as opposed to telling the AI to increase the size of something and then you go go backwards and forwards like three times before it does it the way you want to.

So I think, but I think there's an efficiency that's about to happen with it. It's actually not quite available widely, but there's a really good connection now between your components and actual browser and going both ways. And I'm probably like reluctant to predict what that's going to mean, but it's definitely going to be a good, a good improvement for what we're doing at the moment. not going to replace the good old thinking on the whiteboard though.

I do like the analogy of like crazy eights and how you can use big remake or any of the five coding to do that. And it depends on like for me, it really does depend. But what is the what's the material that you're designing for? And if it is for a chat, see and react code base, then maybe it is a good tool to start in because it's the same as the output, right? You're starting in something that you're building for, so it's got that context.

But if it's an app or something else, or you just want to go abroad, then yeah, cool. Pen and paper and going off script does make a lot more sense. But having the context and treating AI as a tool, like to your point around, like what does it mean for junior designers and what do they need to do? They're gonna be empowered with this as their starting spot. And I think there's a lot of growth that someone can get from going broad with this is additional to there.

And so that Figma to the, to all of the other design tools we have, but we shouldn't get hooked into Figma as a design tool to solve design problems with like, we're solving problems. We're not solving how it looks in Figma. Like it's got to come back to solving problems. I love how Ben described that earlier. I think that was really good. Yeah, no, I think the, I also think that plus one It's like it's like design is in a linear path.

You like it's not like you start with this and then you go to this and then that thing that you started with turns into the thing you ship, right? Like for example, at the moment, it can do this a lot of designers doing prototyping. That's all taking place in cursor. But like the thing that is output is just a proof of concept to like say, okay, this is generally how this product needs to work. Then we'll go back to square one with all those things.

All right, we have a sense of what we need to get to now, but we need to go and design that from scratch. So I think that we need to remember that these tools are allowing us to do new things in quicker ways and that's fine. And I think your point about wiring up the prototype, if MAKE can help a designer do that quicker or in a more comfortable way, it's probably fine as long as it's not designing the prototype. I mean, even if it is, maybe that's okay.

But you see it's kind of like, and that's where I think AI is going to really find its home, is just making small things easier. And I think the like grand promises of it are like 30% true 70% maybe not I just I just made that up so don't hold me back ever But I think that the it's little things where I think LMS can be really powerful I think it is useful to just keep going with this for a second But like you're building upon diverse thinking is building upon different directions you can take it in but I'm still often when I'm doing this I'm still coming back to figma.

I'm still doing specs in figma I'm still going back and building out the different directions and the different error states and scenarios and all of the different things we always did write because we still need to help understand all of the different scenarios we go through. I don't think I'm not there yet in terms of what I'm doing in Big Mermaid or Vercel. Those are tools to get to a point where actually you can articulate all of the different area scenarios and all of the different things.

I still have to apply that thinking and that thinking for me still happens in specking it out and you're still going to have a PM popping into your slack being like hey buddy I've got some spicy feedback and then okay I'm working late tonight so it's like those are the things that like sometimes it's easier just to go into figure and do the change rather than be like alright now I'm going to re-prompt whatever and like you can't change human beings right sorry.

No no so I'm going to take this in a slightly different direction so when I first started as a young copywriter at my first internship. I had three months at this place and I had one job every day. I would go into the agency, I had one brief and the only thing that I was asked to do was come up with 50 ideas. So every day for three months to go into the office and come up with 50 ideas at the same brief. Okay.

And the reason you get asked do this in advertising is because the first 100, 200 ideas you have, all the shitty ideas that you get out, and you only get to the good ideas when you've got all the shitty ideas out. And actually, that's what I find most helpful with AI in my process these days is I'll be like, I've got a headline to write, give me 50 versions of this. So it gets all the shitty ideas out, and I'm like, okay, well, there's a good direction there, there's a good direction there.

And then I can kind of tease together like five good headlines, rather than have to go through the process of coming up with the 50 CT ones first. Yeah, so that's actually been the most helpful for me, especially because like, we're very, very small content design team at Zeke. And I'm I'm constantly context-fitting and I don't have time to really get into that focus state sometimes to just quickly pump up what's needed. Any other questions here?

Audience question

Ben, you don't think AI is a real place, does it? Yes, it is good, like you said, especially for rewriting something, having this bit of it's more succinct and that kind of thing. But we're fine and be good to, I'll just get your feedback. We're fine, it really can't be scrolls and it just can't do it. It can never be, hey, I want some feedback on this design, does it work? I can turn it around to my colleague and five or 10 minutes, wrap it back before, hey, try shifting this thing, there's nothing here.

So even as sophisticated as the AI gets, can you ever see it being as good as a pairing of the human being? No. I mean, I don't think it'll be, I mean, for context, I've worked on training Frontier

Panel

Mom. Like I end this, I have seen how these things are trained. And I'm going to like be honest with you, the people, the data that's being used to train these things is not very good. It's just like, for example, Red is not the source of truth of everything. Okay, the red is very funny, right? Like, the pizza, delicious, right? But no, it's actually worse than that. The way we used to train Gemini is you would be like, okay, we need to train Gemini how to write a limerick.

So I'm going to write a, as a conversation designer, I'm going to write a brief for that, write a bunch of sample data. Then I'm going to go to this data harvesting company, like Surge, or there's a bunch of out there. Surge is one of the ones we used. And they would then hand that brief over to people getting paid 65 cents a day in Sub-Zyron Africa to write me 5,000 limericks. And then we dump that into the model. So these things aren't getting trained by experts. They're just not.

So you're literally never going to get the Claude or ChatGPT or Gemini or Meta AI, any kind of specialized good design, literally never going to happen. It'll give you generic things, it'll propose ideas that sound correct, but it doesn't have that. It's not a human. It doesn't know how a human is going to ingest this thing on a screen. It doesn't know what it's like to have eyes, right? It doesn't fundamentally understand these things and it's been explicitly trained to it.

So I think that it can be a tool that can speed things up and help with certain things, but I don't think it's ever going to replace what a human being actually fundamentally does. And I think that shows up in labor statistics too, right?

Like if you take out, look at US labor statistics, and there's some graphs that have been making their rounds on my favorite website, LinkedIn, where it's like, stock market goes like this, jobs go like this, AI is killing things, like okay, take out the Magnificent Seven, the big tech companies, and it all tracks pretty much the same. It's like, it's just, it's a bit of mass hysteria out there.

And I think the fact that we're all here having conversations three, four years in, anthropic, open AI still still can't tell grandma what she should be doing with their products. Means this isn't the life, the tectonic shift. As I've said, cost pressures are coming. This is not in maybe, this is in the wet, this is not niff. Open AI will go public, they're going to be accountable, they're going to have to be profitable, they're going to have to jack up the price of Bunnings AI.

And I think to your point, for a lot of businesses, there's just experts there. What's better for the businesses? You two chalk. Sorry, I want to tell you. No, that's great. So before we close out, I want to ask one more question.

Panel

And that is around what role do you think designers play in making sure that these systems are ethical? Ben, and we've talked about this briefly, so I'm going to pass straight to you about reducing bias. We've talked about Gemini and all sorts before. Well, yeah. Okay. So the story there is I was I worked on Gemini early last year when there was that it was generating black George Washington's right. And that was that was a problem, right? Like it was wrong was inaccurate.

It was like a you could see there were like bias leaves out pulling and it was not working where it's intended. And there were other things in there too. It was not who would generate a Kamala Harris poster, but it wouldn't generate a Trump poster. We can can debate the ethics of that, but it should be fairly even hand. Now I think in a couple of years there's going to be some really interesting research on the extent of the digital colonialism inflicted upon the world by these companies.

These are products built and essentially trained by North Americans being poorly translated into every other language and just force in front of them, right? And I think that's okay to a certain point, but that bias starts to leak out into other conversations into the slop that is thrown out there into the world. But I think in terms of like, debiasing it for our users, it's just like we need to have good policy people who understand the technology.

You need to have people who are really ready, like everyone on this panel too, who's willing to sit down and red team it to like push it to its breaking point. You need to have people who are just spending all day thinking about what is the worst thing a human is going to do with this. Like literally that's what I was doing a lot of at Google. It was like can I get... And we've seen so much of it in the news personal and chapter 3 started up about all the... Right. Right. The assisted suicides essentially. Right.

I had to read mine calm to train the model. Right. Like it was... Because I had to think, okay, what is the worst thing somebody's going to do? And like, okay, can I make it...

Like I was typing into the awesome abhorrent stuff and I would use some guardrails free one and it would generate some of our stuff We're doing like see Sam just stuff that should not yeah, but it's in there So I think like you need people who are really aware of this who are understanding of shifting tides This is a new role that's signed as plays Understand like how much of people gonna miss it because it can destroy a brand right? Yeah So to answer your question, I think designers need to be very aware of it.

We need a red team We need to understand that these things are generative and they can be broken and The worst thing that can be done with it. Somebody's gonna do something even worse than that Don't trust people But still don't trust me I think we're all kind of biased and around in our own ways And I think diversity helps break some of that and all just keeps in checks all right just keeps it the more diverse our teams are, the more that we're working together across functional and making sure that we've got different people's opinions helps with some of this.

But as you say that you've got to keep checking and people will use it for the worst possible ways. But it's just making sure that you've got a diverse source of people checking the output. You're making a point. I was conflating bias with like policy violations and they can be different. I think you're right.

It's like one of the problems that I think going back what I was saying about digital colonialism is that what me as an Australian upper middle class person thinks is bad or wrong may not be what somebody in, you know, a working class person in Indonesia agrees. You know what I mean? And vice versa. And I think that for me to impose my, so okay, I got to give an example, right? So in an AI model I've worked on the past, not Canva, the word Negro was bad, right?

And that was because white people who just thought, this is a bad word, I feel uncomfortable saying it, they bad. But a lot of feedback came through from older black Americans being like, that is how I refer to people in my community. I am now being like, it's always racist that you have decided that I can't say that, right? So like bias is very, very contextual and very, very much about the individual user. So it comes back to like what is the policy of the brand, right?

Like what are we comfortable allowing and not allowing? But that's hyper nuance as hyper regional. It's not an easy thing. And I do think that a lot of companies that are doing and actually no one, no one really here with maybe with the exception of Canva is biting off a bit more than they can maybe chew with it. It's like I don't think you quite realize how bad this can get. But I say that having sat in a dark room at Google HQ writing poems about Hitler.

What of the most memorable, oh, one of the most memorable meetings I think I've had. Not long after the start of it, so you can, we had to come up with a list of bad words that we weren't allowed, are going to allow people to put into job ads. So sitting in a room with, you know, people from different disciplines, and I was quite at the time, thinking of all the possible swear words that you can think of and listing them out so that we could, but it was an interesting exercise.

But then, similarly, some people, there are some legitimate use cases, we have, we are in seven different countries, and in some countries, some words are more accessible than in other ones we flag it. So we do allow people to say, I think in Australia, it's a girl's school or a women's school, but I think we have a caution for other people because the amount of ads that are saying that they don't want any women applying or something like this. So it is different in different countries and it is a challenge.

And it is like Richard said, context is everything. The square word is conflicting some context. Some squares are great in some contexts. Yeah, yeah. Well, thank you so much everyone for joining us today. And I hope that you've all learned something and will take something away from this. And thank you all for taking time out on your Wednesday evening to come.

I know it's sort of late and after a long day of work, everyone wants to head home, but I hope that the food and wine and the great chat was a good incentive. Please feel free to stick around until eight o'clock. Have another drink, have some food, have a chat.

Panel

the speakers will be around if you want to chat to them more. But just yeah thank you all for