This article is a transcription of the premiere episode of the second season of IDology - the industrial design podcast by Mindsailors. You can watch the entire episode on YouTube or listen to the audio version on Spotify, Apple Podcasts or Google Podcasts.
In the premiere episode of the second season of IDology, our company’s COO, Voytek Holysz, sat down with Piotr Dalewski, a mechanical design engineer at Mindsailors, and industrial design company, and the co-founder of Mindsailors and senior designer himself, Rafal Pilat, to talk about Generative Artificial Intelligence.
We can all admit, with all its prons and cons, that Generative AI is a great tool for product designers and alike. Some would even say that it helps to skip some designing steps... But can it really revolutionize your product design workflow? Should you really rely on AI? Let's find out!
Voytek Holysz: Okay, I think we can continue this conversation because we have had it before. And I remember that you, Piotr, mentioned that before we talk about AI in design, and why AI isn’t necessarily a ‘solve it all tool’ for industrial designers, we need to firstly answer the question, why are our products designed and how?
Piotr Dalewski: Yes, I think the understanding of the project background is crucial for the success of the product. At the beginning of our design process, we try to deeply understand the vision of our customer, that we know that this specific device should fulfill some specific needs. Those products are very often innovative, they are not on the market yet and we have no other products to compare them with. So actually, we are just limited by our imagination but this imagination isn't that easy to mold into the prompt, we can put in some GPT or AI browser.
Voytek: Yeah. You said that, if I understand correctly, this is a case of understanding ideas behind products in order to make the products good and AI is something that is very literal, like you mentioned prompts, they need to be very, very literal for the AI to generate something because you can put an idea into a prompt the AI won’t know what to give you.
Piotr: Yeah, actually, our customers don’t want to have generic solutions, generic problems. Our ideas come from sometimes many hours of discussions, many emails, some kinds of workshops, when we analyze those needs of the users, of the manufacturing companies, and even the vision of the investors or even all stakeholders. So, to understand those needs, and mold them into the brief, it takes a lot of time and also a lot of empathy, like we need to understand the whole vision and whole logic process.
I think what is also important, is that many customers have different visions, different products are directed into different customers, but they also have different needs. For example, when some kind of company releases their first product on the market, they are focused on low volume manufacturing, and in design, especially in the documentation or the further development, we focus on the solutions that are fit to them. So, for example, we avoid the complicated or sophisticated assembly ways, because those would enlarge the cost of introducing into the market. And if we have a large company, they have already released a few products, and they know their market, and they know that they are able to invest large amounts of money, we can focus on more advanced technologies.
So, before we design, we have to have a vision of this product. If we want advanced design, we can suggest double injection molding or other overmolding, or when we are designing for the mass volume production, we think more about the assembly: so we are using some kind of snap click-fits, something that snap fits make assembly very easy, and they limit the labor. So all those needs influence our design works so it would be very, very difficult to write the prompt that includes all those items.
Voytek: Rafał, how many years of experience do you have as an industrial designer?
Rafał Piłat: About 18.
Voytek: So that’s 18 years. It’s very important what you said - you’ve mentioned empathy, that’s 18 years of negotiating, extracting information and extracting intention from other people: from clients, or co-workers, or manufacturers like you have mentioned and that’s the kind of work that just based on data or information wouldn’t be basically possible without empathy, without understanding how another person thinks, what they think of when they say something, because that’s not always the same. So do you agree that this empathy can be a key to it all? That lack of empathy simply says “No, this won’t be possible”?
Rafał: I would say, one of the biggest issues when working with clients is, like you said, extracting information from them. Very often we meet with clients who have hard times verbalizing their ideas, describing to us what they actually need, they are not even aware of certain problems that their product should solve. So, that’s why we not only have to interpret the wishes of our clients, but we also need to be able to imagine, to put ourselves in the role of people that will actually be using those products.
So, I guess, this is one of the most important things for a designer, to be able to get into someone else’s shoes and try to imagine how the product will be used and what potential issues and problems can arise from using such a product. Current AI models have a very limited knowledge in terms of interpreting our ideas and, let’s say, finding solutions to problems that we have. They are trained specifically either on text or pictures; what they lack is expertise on, let’s say, problem solving, which is related, for instance, to movement, construction, manufacturing process, etc. So, in this regard, I think that the current gen, maybe the next generation AI or some kind of expert AI models will be able to work out certain problems or address them in some way but the current gen is very limited in terms of creative process and creative approach to solving issues or problems, which I mentioned.
Voytek: Yeah, it’s like I remember we were, I think you once brought up a meme about how graphic designers are safe, because if AI was to replace them, the clients would firstly need to say what they actually want. With every creative work, it’s a process of mutual understanding before we can get to a result.
Rafał: Yeah, I am sure that not only industrial designers, but all kinds of people who are dealing with clients and are working in a creative field will agree that working with certain clients can be a guesswork, rather than trying to interpret their intentions and their, let’s say, putting a product brief into materializing into something feasible. So certainly, we are safe in this regard.
Piotr: I would like to add some words about this data extraction for customers. When we start the product, we usually have a brief. And we might say, it's kind of a prompt for us. But actually, we are not only extracting this data, but we also validate it; we try to understand, as more experienced ones, the purpose of this device. We try to validate all those needs that customers see, but also we try to define the problems that might occur and maybe add some additional functions that will improve the product. So it’s not about generating the design, but we also try to develop or refine the input data to achieve the best results.
Voytek: That’s also interesting because you have compared the project brief to a prompt for designers, which is a brilliant comparison. And still, when you give a prompt to AI, at least for now, it just gets to work. And like you said, the first thing you do is look at the brief with scrutiny, like with critical confrontation, with facts to validate if the brief even makes sense from the standpoint of the project. This is something that, at least I don’t know for now, I am not aware of, that AI can do something like that.
Piotr: Well, it’s also very important that we are not doing it ourselves. Our team of designers is focused on that, we have got a workshop or brainstorming with a customer. So, I would say, that there are many, many “computers” that think about the best solution. And what is also quite funny during these workshops and meetings is that I have noticed, it’s not that easy to communicate with each other because the words are the same, but actually, we think something different behind them.
For example, the client says, “I want something futuristic,” and for someone, it seems like maybe something starwars-like, another person will think maybe of Interstellar or cyberpunk, so the same word has a few meanings. That’s why during the early stages of the design process, we communicate not only with words, but we also prepare mood boards. Sometimes when it comes to aesthetics, we communicate more with the pictures than with words. When we try to explain something to each other, we also move our hands or try to communicate not only with words, and when you are working with an AI, it’s not impossible. That’s a huge advantage when it comes to working with a design team.
Voytek: But also, what you have mentioned, “design me something futuristic”: AI would just take the cyberpunk examples, the interstellar examples, etc., mash them up and give you “something futuristic”. That’s not the way a designer works basically. I mean, as a designer, you use mood boards and inspiration, but not that literally like AI does.
Rafał: Right, I guess it’s always a matter of taste, but aesthetics of a project or of a product is just part of the design because, again, the key element in each design is its functionality and construction. We cannot design, or I mean, we try not to design only things that are just looking nice - they need to serve a purpose. We are not creating an image, we are creating a visualization of something which will eventually turn into a product.
So, the principles are completely different when you are thinking about building something from blocks, and when you are just trying to paint a painting. Those are two completely different processes and they require completely different approaches as well.
Voytek: Okay. I also remember we have talked about how AI handles reverse engineering, sort of, because we are talking about creating something brand new, new ideas, new products, new solutions, which doesn’t necessarily mean new shapes or inventing new colors or whatever. But still, it requires a certain amount of originality, which is a problem for AI that basically works on a set of images. Also to understand like you said the functionality, because we know it’s meant to be a device, not just a nice picture. You’ve mentioned you had tried reverse engineering for your designs with AI. How did that go?
Rafał: It didn’t go very well because when I asked, I think it was Midjourney, to interpret or describe an image what it actually does see, it had a hard time trying to identify not only what the device was, but also what was the function, so the guess was completely off. But then again, I shouldn’t be surprised, because in the end, even if we show our design sometimes to people, they also have hard times trying to guess what the device or what the object is, what is the functionality of this, what the main purpose of such a device is.
So, I guess, we shouldn’t be expecting some kind of revolutionary skills from an AI model, which we should remember is trained on something, which is already there existing: so if it does come to a problem of analyzing an image, which does not resemble anything existing in the world, it will have problems interpreting it. And it works also the other way around: when you expect a very creative shape or, let’s say, aesthetics or a construction, which should be original, which should not be a copy of something else, sometimes, let’s say the looks don’t correspond to the functionality of an object. AI won’t be able to work out on such a specific design.
So, I guess, those are currently the biggest limitations, in my opinion, when using the AI engines to generate even initial ideas for objects.
Voytek: Can you give me an example of what you are describing?
Rafał: Yeah, sure, we had a project which was a table for hairdressers, it was a hair dyeing piece of furniture.
Voytek: Like a workstation?
Rafał: Workstation for a hairdresser with a purpose of making the hair dyeing process easier. We had an internal context, whenever someone came to our office, whether it was a client, or a friend, for some time, we had a prototype in our office, and we always asked the people, “Okay, can you guess what this device does?”. I don’t think there was one person who was even close to guessing what the functionality was, I think that the best guess was that it may serve some kind of makeup, like station for applying makeup or something like this, but no one had guessed that it was actually a hair dyeing station or a hairdresser device.
Voytek: Is it to a point, at this stage at least, that we are talking about AI tools in product design, that they are trained on data that is very universal, so AI can tell you that it’s looking at a medical device. It would answer like a regular person would: “what’s this?” - “this is some sort of medical device”. But if you showed that same device to a professional doctor, or someone who works with similar devices, they would probably be able to say, “This is a portable ECG or whatever.”
Rafał: Yeah, I guess this is partially related to the actual teaching process of the AI because I tried to train AI with my own images with a set of images that I have prepared. You can use, let’s say, the engine to actually try to guess what it sees, but you can expect the best results when you will actually make the description by yourself. And this description should be as accurate as possible. For instance, if I am showing an image of a person who is six-years-old, and the next image is of a person who is seven-years-old, then I can teach the AI that the face in the first picture was a different age than on the second. So, this is related to teaching the AI how it should interpret what it does see.
Obviously, the better the description, the better the end result will be. But still, like I said, even humans with generic knowledge sometimes will have hard times describing the purpose of the device, which they have not used or seen before. So, I guess even doctors could have problems when we would show them, for instance, my spiral, which is a portable spirometer, it does not look like any other spirometer. So I guess it might be problematic, even to a doctor, to try to guess what this device might be.
Voytek: Okay, so, for now - it is, to some extent, also a matter of technical expertise or experience on the side of AI. Because like you said, it has general knowledge, general concept but lack of technical expertise sort of makes it still unaware or unable to interpret what exactly is that it should design?
Piotr: Actually, we don’t really know what the basic knowledge of some kind of AI is, we know these are trained models. The crucial thing is that we need to know, which was the input data. So, right now, we can generate some text, some blog articles, and create the pictures. But looking forward, when it comes to designing whole products - I mean, just creating whole CAD models and tooling from just the prompt.
I think I see one big issue that AI models are trained on some kind of data, and this data should be validated. I mean, the proven knowledge, proven designs, including all details like in design: draft angles or bending radiuses, or just the technical data. If we want to train some, for example, such an AI algorithm that creates this data, we should feed it, we should teach from the proven designs. And actually, these designs are not available anywhere on the Internet, well… There are some kinds of databases like GrabCAD, where users can share the RDS but they are not validated, they are not proven by toolmakers, they are just random data, let’s say and many of them can have some flaws, errors and things like this.
So, when it comes to the investment, because the product development is a large investment, we have that in mind, we should focus on validated data. It would be very, very hard to get this data to train some kind of AI model to give us good results. Because this is secure data: we are not sharing our products and projects online and even large companies are not showing their products, you can see just the picture or some kind of render, but you don’t have a whole formula on how to manufacture it, you can reverse engineer it.
Voytek: We can guess to some extent.
Piotr: Yes, or you can buy, disassemble and measure everything which is a very laborious, costly process. But actually, I think, training such a model would be possible in some closed environments, large companies to some extent for some kind of configured products, maybe some kind of roller conveyors, stuff like this.
However, it costs us lots of time and lots of knowledge, which isn’t that easy to possess. And the sources of this knowledge are very limited, usually to be trained in some kind of design area. You have to be hired in some kind of manufacturing company and then you will be learning step by step. So, I think this is quite complex.
On the other hand, as we are looking into the past, when some kinds of technology developed, its cost increased on the global market. I mean, when new art softwares, that allows you to design, appeared actually, it caused the increase of product release on the market. And I think that development of AI will definitely change the market, it will change the way designers work. But actually, I think it will be more like a tool, or I would say, an ally.
So, I think that people will be more creative, more inspired to develop new products and more products will appear. But there’s always going to be a need for somebody to prove it, to double check all the facts, and actually, to fulfill those products to bring them on the market, we will need somebody to guide the process.
Rafał: Yeah, the way I see it, those tools, those AI enhanced tools, will develop into some kind of assistance during actual design or modeling. I imagine that maybe we will have some kind of creators that will speed up our work, for instance, you will select a few parts and try to merge them and then a machine will suggest the best way to merge them, or even will select, based on the simulated weight of the parts, what would be the best nuts and bolts to use for such assembly.
So, maybe, we will see some kind of wizards that will enhance our work, that will simply make our work more efficient. But I guess, I do not expect a model that will do the full work for me. Because at the end, this is a discussion of knowledge versus experience.
I think the machine will always have some kind of theoretical knowledge but without actual practical verification of someone else, or practical validation of a prototype and experience coming from working on different kinds of projects or products, it will never be as efficient as a human who actually solves specific problems, or has a vast general knowledge on different subject matter related to product design, in this case.
Piotr: We need to be aware that design costs are just a fracture of the whole manufacturing or whole design process, depending on the project, it can be like about 10 or 15% and to develop this project, you need competent experienced designers or decision makers who will take responsibility for all those choices.
Rafał: And not only designers because this is such a complex process to develop a new product that apart from designers, you need engineers, you need specialists in other fields as well.
Voytek: You need logistics, you need marketing experts, business experts, financial taxes, safety. So you need to understand all those stakeholders or perspectives in a product.
Rafał: That’s right. And each of those persons has some kind of role and influence in every project. So I guess, it would be wishful thinking to expect that AI will simply do all the work for us.
Piotr: These are just our assumptions. Maybe in 10 years these episodes will be very outdated.
Rafał: And we will laugh.
Voytek: Yeah. Because there is a subconscious human need to tell ourselves that we are so irreplaceable and so unique that we think “no way, I will ever do it” but who knows? Maybe in time there will be separate AI for each expert, they will be working in tandem in AI industrial design offices.
Rafał: That’s true. Who knows? You cannot tell for certain how the models will work in 10 years, how the AI research will advance, maybe the machines will do a better job than we do.
Voytek: Well, probably until there’s a huge leap in computing power, it’s not going to change.
Rafał: Yes. We still need to wait for actual AGI models, which have not yet appeared. And I guess, until we do have something like this, we cannot expect actual creative work and problem solving tools.
Voytek: Cool. Great talk, guys. Thank you.
Rafał: Thank you.
Piotr Dalewski, a mechanical design engineer at Mindsailors, places emphasis on manufacturability, design for assembly, and product optimization within the value engineering approach throughout the R&D process.
Rafał Piłat is an entrepreneur and co-founder of Mindsailors, an awarded industrial design company, with over 18 years of experience as a designer himself.
Voytek Holysz is the COO of Mindsailors with 16 years of experience in running a business in creative B2B services, marketing, sales and video production.
IDology #13 - When using AI in product development is not a good idea
Schedule an initial talk and get to know us better! You already have a basic brief? Send it over so we can have a more productive first meeting!