Thoughts about this claim about Grasshopper and AI

Hello everyone.
In the last months I’ve been encountering many posts (this, this and this) about an architecture studio that claims to be bulding “artificial intelligence” scripts with grasshopper.
When I dug out a little bit more, I found that their so-called “artificial intelligence” is only galapagos, a built in functionality within GH. For what I can see from the pictures and videos availiable in their website there is no such thing as Grasshopper AI, and that they are only using the ignorance of the mainstream architecture media as leverage for god knows what. Although Galapagos is a very powerful tool, I think that calling it AI it’s a lot. Galapagos doesn’t learn anything or is trained against anything to then make its own descitions as AI systms do, and ultimately they didn’t invented anything, they just coded some IF statements and that’s it. It kinds of bothers me in some way that media portray this guys as if they have invented the wheel when they clearly didn’t. I don’t know, maybe I’m wrong and I do not understand the true scope of this Flinch3d thing and there is trully some kind of AI behind it, but it doesn’t seem like it. As a seasoned grasshopper user I feel kind of treated like a fool.

What are your thoughts?

By the way, here is a very interesting post about galapagos written by David, if you don’t know what I’m talking about.

Best regards
Sir Ernest Shackleton


I think AI today can be distinguished from automation.
People in the past called programmed toaster(…) or laundry as AI implemented product and they sort of agreed to it.
But since the emergence of successful cases in deep learning showed up, those automation is becoming to be distinguished from so-called AI now, I guess?

I saw this project from GSD’s website that used GAN to interpret plan drawings and produce new drawings from that.
Its much sophisticated than the OP’s link but still not mature enough than human designer.
However I believed this kind of approach would be used in the future to translate non-BIM data into BIM.
I’m not professional so I only know the surface of the whole picture…

Let me begin by saying,there AI kind of covers generative and data driven design. But the A.I is most directly related to some kind of dynamic machine learning that almost depicts human-like decision making capabilities. However I have not seen anything remotely on the same lines being worked through over grasshopper. But generative and data-driven design has been on full throttle for quite sometime.
(Refer link)
There have been many attempts made at single and multi objective optimisation over grasshopper using evolutionary algorithms such as Galapagos and octopus. In all the cases above there is no ‘machine learning’ all the outputs are an output of smart coding and conditional statements and final judgement if of the designer ,which option to pick for further iterate over.
Now coming to the links shared above they are just simple conditional statements ( what-if area of room is so, place the bed ‘here’) in short ’ no intelligence’.
A.I has been dead for a long time. But generative and data-driven design is being seen as a way forward, but still stuck at a nascent stage(atleast for the masses).


1 Like

As far as I know, AI stands for Awesome Images. So awesome, that people will try to show you 10 in one second. This is why they often come in the form of GIFs. :slight_smile:


Well far as I can tell this is what architects like to DO. What academic-types do, anyway. It’s part of the same thing as when architecture students get excited about the bloody “golden ratio” and design everything to look like a sea shell–well that’s what they did back in my day. It’s trying to find some kind of math to describe the “ideal” way to design something–where this originated I have no idea…well I do but it’s about politics which I’m loathe to discuss–and “AI will do it” is the logical endpoint of this obsession.

1 Like

If you break it down to what AI is, no matter if Genetic Algorithms (e.g. Galapagos) or Deep Learning (which is a different form of ‘weak’ AI) its just basically a better way of guessing/trying things automatically.

You could say that this is also how our brain works, which is partially true. But since we don’t learn to solve just one task, we are actually able to understand things. To know whats going on, really helps in not blindly trying and guessing. A trained neural network does not understand, it just knows how to categorise an input compared to other input (to filter out). Furthermore an AI is currently not really creative, because of this, since it can only categorize based on what it has processed (“learned”) yet.

A provokative person could say that in our context, using an AI is rather a sign of lacking knowledge and creativity rather then being truly innovative. I personally find it very hard to spot a real usecase of AI within a design process, which cannot be or is better solved through traditional processes. And if there is such usecase, creating an AI for this sort of work really requires expertise and time. I really doubt that most architectual graduates do know enough about this topic. I definitly do not , but for me I rather read such posts as slapstick comedy, which is especially funny because of this slight natural narcimn designer/architect but also engineers have.


Maybe why current AI boom doesn’t interest architects so much than other profession is because most work what architect do is relatively not that much complicated.
Conventional programming is automation of something what human knows how to do.
But the new approach is making computer to find the solution that even human doesn’t know how to.
Doing what human doesn’t even know is the magic but architect’s job is to deals with what they already know and we already have engineer and partners that helps us.
Maybe real estate field can exploit this skill more than us?
I can’t think other than translating old drawings to CAD or BIM. What value would this technology bring us in future? :sweat_smile:

The genetic algorithms are within the AI as a method of multiagent optimization, of many others that there are. The thing is, what seemed smart yesterday may not seem so smart today. But it’s within the domain of AI-associated tasks. There are many AI outside of machine learning (which is only a subset of the domain of tasks that have in common that you can learn from data), like agent theory or graph theory.

What they have done is AI depends on how they do it. If they have a program that generates plans using conditionals, then no, it is not AI. But if they use statistical/probabilistic models or some kind of rules in which “intelligent” behavior emerges (that is, it is not explicitly written) then yes.

It is true that many people abuse these terms but it is also true that it looks more impressive than it really is.

And no, AI is not going to replace architects, they are going to be the tools of tomorrow’s architects.

Having said that, now the issue is how to measure generativity, as an algorithm is able to explore the parametric space within the design space (which is not a physical space but the set of possible variations). Then measuring this you can say that a generative model is more “intelligent” than another for a particular task.

Machine learning isn’t magic, it’s steroid statistics. The creative applications you can make will be very impressive, but you will also lose interest in them very quickly because they depend on how you teach them. This is going to eat up everything that can be automated, but the creative, very partially. It will be something like they will win us over by making thousands of different solutions, but without the brilliance or genius that we can have. Look at the paintings that are currently making AI models, they are impressive but they don’t have a deep impact like the great artists do. One might think that at the rate we’re going, there’s not much time left to make genuinely talented generative models, but that’s not true IMO, I think it takes an AI-complete to get there, and I doubt we’ll live to see it. I think that these solutions, after the initial impact, will be on one hand a tool, very useful for the creative process, and on the other hand, solutions of the type standard/what everyone does/fast and poor.


An algorithm can spit out dozens of designs that meet some particular criteria. In the end, it still requires an eye to separate good design from bad.


… I’m sorry Ethan, I’m afraid I can’t do that


Maybe I was being too glib, because there is a field, Neuroaesthetics, which seeks to map out the pleasurable effects of art on the brain. If we reach the point that, based on a sufficiently large training sample, we know how much the brain will “light up”, for a given input, we won’t need a good eye any longer. Just plug in the design criteria, and presto!, the perfect object.

1 Like

Interesting. Some time ago I was reading a lot about Neuroesthetics and how it might affect architecture.

From what I found so far, I don’t think that Nuroesthetics could be translated into architecture.
It’s one thing to “find” the “pleasurable” aesthetics on a framed piece that can be showcased in any type of context/space, and another - to implement those same principles on a building.

Architecture is driven by the surrounding context. You will never design the same building in Berlin and New Delhi no matter how much your building sites are the same, and if you or the AI does - it would be a colossal failure.
On the other hand - the Mona Lisa, will look equally spectacular wherever you showcase it.

AI will impact architecture in a huge way, there shouldn’t be any doubt about that, but it won’t replace the human principle of design.
AI will be a tool that will help us get realtime feedback on building performance, building site requirements, budget constraints etc.

The profession will change a lot, but in my opinion, there still needs to be a trained specialist to manage all those tools.

The only downside I can see is that there will be a lot of architects left without work since these tools will automate a lot of the processes in the industry.

One thing that will need to change for sure is that architects will need to learn how to code.
Programmers, even these days, are becoming a standard part of teams in architectural practices.

As for Finch3D - if you look around their website, the creators are not mentioning AI even once as far as can see. The media was representing their tool as an AI since it was presented a year or two ago, so it seems that the creators are well aware they are just making a complex algorithm and nothing more.

1 Like

Yes, within a range of controllable aesthetic criteria. You can now train a network to learn how to categorize or rate a collection of products if you find a way to measure those aesthetic criteria. Imagine a clothing e-commerce, instead of categorizing products by type of garment, you can automate the categorization by more abstract words to describe the aesthetics, such as summer or winter, warrior or delicate, or almost anything you can imagine. The problem is that you can’t get out of your measurements, you need to retrain it (or use another NN) if you want to differentiate between hippie and classic for example. They are models dependent on their definition and this is a problem for complex problems or adapted to specific projects. If you can measure what a perfect object means, then you can optimize it.

I fully agree on the first part. It might worth to note that Neuroesthics does not create something new. Its an optimisation on given input. If you feed in Rembrands as positive example, the result will always be a Rembrandisch derivative, but nothing new or as you said nothing that fits the context.

When I see on what great failures and risks architects encounter nowadays, I can tell that the last thing architects as whole need to do is to learn to code. To know how to code is indeed benefical for some specialist team members, but seriously the greatest lack I see today is the stronger seperation of concerns. My impression is that creating a design and actually make it buildable seems to be very decoupled nowadays. While the creation of a complex designs greatly improved, the realisation becomes more and more a challenge. Especially looking at all these great failures here in Germany seem to prove it to me. My point is that applying AI on a design is nothing trivial and I still not see the real need for it.
What you are describing seems more that you are looking for a way to validate your doing and this requires analytic tools but no AI. Automatisation is not necessarly always a good thing. Especially when architects become programmers, when do they actually deal with all the rest of this extreme diverse and already very demanding job?

Software development alone is a full time job. Anyone who coded something more complex knows how long it takes to create something really useful. The idea “lets write a short GH Script with AI” is totally unrealistic. At least if you expect something useful…


Analysis and visualization are central tasks of AI today and have a huge impact also on design, especially on parametric design (and for huge product collections). If you have a customizable product, a piece of furniture, a room, a modular/adaptive building, or an industrial or design piece, based on a generative modeler, you have an explosion of permutations or possible variations that neither the designer nor the client have the facility to find good solutions easily. You can filter using your own criteria, but there are/will be better alternatives. The analysis and visualization that AI can give here will help to remap this space of solutions into a space of semantic solutions, where you map different types of results by features and allows you to compress or code a parametric space into a feature space where exploration is much more fun, so that instead of using parameters that change measurements or quantities, you have parameters that change aesthetic or functional features, and even do arithmetic operations between them (give me a sofa characterized by garden, party, coffee and modern but no plastic), bc all these attributes can be converted to vectors using ML and represents patterns in the algorithm and its parameters. This solves a real problem of product configurators, which simply offer a lousy experience for the customer if you offer many parameters, or offer very little variability if you offer few parameters, currently this is solved by using preconfigurations that offer some variation, but this is no longer really something configurable or customizable. With AI, parameters can be design features rather than design building factors.

1 Like

Yes, you are right. I took the whole “coding” thing way too much of proportion.
Let me rephrase and clarify…

Architects should learn how to build their own tools.
I think that every single architect should learn Grasshopper or Dynamo. Or, at least, learn their tool’s scripting language.
Automation in today’s industry is very valuable and gives you space to dedicate more time to design and building processes when you add the possibility to automate analytics, documentation, organization etc.
I completely agree that there is no need and that it’s very unreasonable for architects to be full-stack devs but they need to be able to build their own tools, and understand the tools that others share with them.

Also, you are “on to me” - My philosophy is that it’s not enough to sell your pretty renders and specialist jargon, but that we, as architects, should be backing up our design decisions with data.
That’s where those analytic tools come into play. If you can connect your design and building process with tools that give you realtime feedback and validate if everything is “going in the right direction”, I think that we will be able to create better spaces.

I imagine tools [data-powered systems] that can be shared from urban planners to one architect, then the next, and the next - and every other iteration of those systems would be more useful.
And here is where AI could come into play.

But, the whole building industry needs to change and implement technology into the buildings and surrounding environments themselves, so that we can collect all that needed data to make smart systems that will aid the design and building process.
If we are going to talk about AI in architecture, first we need to talk about where we are going to find all that data for the specific local contexts to power those systems and be trully valuable.

Since there is no data like that, we can only talk about automating processes and analytical systems that are as smart as the data you can provide and yes, we simply cannot refer to it as AI.

To recap:
No, architects shouldn’t be software developers, but they should learn the basic principles of coding and learn how to use more sophisticated tools in order to design and build better.

1 Like

Agreed, there are cultural and environmental factors, and as @TomTom says, everything depends on the sample used to train the AI, so you’ll never get to see anything new and better because the algorithm would have rejected it before it reached your eyes. On the other hand, we are animals shaped by evolution, and maybe our “aesthetic space” is already highly constrained, so how much “new” is really out there?

1 Like

I’m not sure if you mixing a bit parameterized cad models with AI. Such permutations can and are created without using AI at all. You can even filter, both the input and output, based on analytic methods.

I believe its the other way around. An AI can use an analytic algorithm as “Fitness function” or “Weighting function” (depending own which kind of AI you are using). I mean you are right, using a genetic algorithm is really obvious for such scenarious. But the difficulty is the fitness function. It can also be that you are missing a lot of interessting permutation if your fitness function is simply not good enough. Or sometimes, the optimal solution is simply not the best or most aesthetic one.

Of course this will also depend on the actual use case. But what I see, when looking at the links mentionend above, is that the ai-ish optimisation is not validated at all. If people say without a high amount of permutation, they wouldn’t have found an optimal solution, then this is questionable. Its the chicken-egg problem. If someone is able to create a great fitness function, doesn’t this person already know how to find a good solution without using AI. The definition of itself is already pointing to whats leading to a good result. Balkonies in the south, squarish floor plans, good proportions of each room, etc. If you know these triggers, why not creating some parametric models directly considering this?

The other problem I see with these configurators is the simplification of design which takes place. Some people will disagree here, but I believe its not doable to keep a certain level of detail if you keep everything parametric. Anyway, I’m not against the usage of AI at all. My point is just that it really has a much weaker impact on any design related industry as most people believe. Its not even something new, even in Architecture. There is work out there being almost 50 years old. Nothing really changed until then. There is simply a great mismatch between theory and real world application.


Interesting thoughts everyone, thanks for sharing.

I agree the industry is full of nonsense around AI. Let’s clear it up by learning more precise terms and connecting them to real, useful applications. I think our industry can have a more informative discussion if we focus on what AI/Machine Learning(ML) tools do, rather than what they are. Rather than throw around broad terms like AI, we should discuss particular algorithms, or even better, what these algorithms do. Most of us probably don’t even know what to google to learn more. I’ll try to fill this post with terms I find google-useful, so please forgive my obnoxious use of fancy words.

The way I see it, we have a variety of algorithms at our disposal (Genetic Algorithms, Neural Networks, Decision Trees, K-Means Clustering, etc.) that do a variety of things (Optimizers, Regressors, Classifiers, etc.). What we want to do determines which tool/algorithm to apply.

I’m a skeptic. I worry that it’s too easy to unknowingly build assumptions into the fitness functions that optimizers rely on to generate the “best” design(s), making the results not actually optimal at all. Also, optimizers take the designer out of the loop. I don’t think design involves optimizing at all, as it can be so subjective. Rather, let’s accept that we are usually exploring or refining, not optimizing. If so, I suggest we focus on Multi-Objective-Optimization (MOO) so that our results are a collection of “better” ideas given the assumptions not a false single “best.” In fact, I think our industry should focus more on Design Space Exploration (DSE) techniques that inform the designer rather than optimizers that replace the designer.

Regressors basically interpolate and they go hand in hand with DSE. If you iterate a parametric model to generate a design space, you can apply a Regressor (Neural Networks work well) on the collected data to train a surrogate model of the entire space (See Design Space Explorer, Crow, Owl plugins). This is useful to replace slow parametric simulations with real-time predictive engines, which can then be used to power other generative techniques. Once you have mapped the entire design space using a Regressor, you can explore it using optimizers to identify “best/better” options, run sensitivity analysis on each parameter (Multiple Linear Regression MLR) to guide criteria prioritization, understand covariance, etc. Dimensionality Reduction techniques like Principal Component Analysis (PCA) and Self Organizing Maps (SOMs) come into play here, as they help clarify complex data. I like DSE because it is all based on building a good underlying parametric model of the design challenge to generate the necessary data/design space. We don’t have to have or find the data, we generate it ourselves. Once you have modeled your challenge parametrically and translated that problem into a database we can apply all sorts of AI/ML tools to mine that data for insight and apply it in practice. How are others applying these techniques? Please share with the community so we can all upskill.

Am I correct in thinking that generative floorplans are fancy applications of regressor applications? They “learn” from past designs to choose a new design based on a complicated interpolation of the training data? Rather than expect our algorithms to deliver a completed floorplan we can also break the problem into parts. WeWork showed a great example where the tool predicted the expected number of workstations that could go into any shape. This seemed more useful than generating a final layout (it also did that) because the designer would likely want to tweak it anyway. I’ve heard of Reinforcement learning techniques like Generative Adversarial Networks (GANs), Deep Q Learning being used for these types of problems.

I’ve struggled to find good applications for Classifiers, or clustering algorithms. If we had the data we could classify go/no go decisions, identify problems in on-site construction images, flag documentation issues, but these all seem to require data that is harder to gather. What have others found? Gaussian Mixture Modeling (GMM) seems promising because it returns the probability that an input belongs to a particular cluster, rather than an explicit prediction. This fuzziness seems applicable to our world. GMMs are used to separate noise from background, like separating multiple instruments out of a single audio file, or a moving object from a still background. Perhaps edge detection/ object detection could be useful?

Are there other major application groups I’m missing? What do others think about focusing the conversation on what the algorithm does rather than what the algorithm is called?

I like Scikit Learn’s cheat-sheet:



I like your point about parametric modeling encouraging a certain level of simplification. I’ve been struggling with this as well. My current approach is to intentionally model generalized (non-detailed) models that deliver information about any final design.

For example, if I know something about Design A and Design B, and Design C is between them, I know something about Design C even if I didn’t explicitly model it. The trick is building generalized models that return useful insight. Once the general direction has been informed then we can manually refine.

How do you get past this dilemma?