In such case AI allows you to get pseudo expertise. For example if need to pick specifications glass (transmittance, reflectivity…) and I barely know anything about it. LLMS can provide useful rundown that may allow better understanding of needs and communication between contractor and designer . Which of course could be double-edged sword.
LLMs are particularly bad at describing the finer points of materials and architecture. but regardless, as you said “pseudo expertise”
“pseudo expertise” will land you and your insurance provider on the paying end of a fat liability lawsuit and potential jail time. maybe an ai attorney could help…
No, because I never claimed any of those things.
In my view yes.
Biggest thing holding back AI or the mythical AGI the silicon valley people so desire is the raw cost of it all. When a human being can do work at a fraction of a fraction of the cost of AI then there is plenty of room for us to continue what we do.
With each new version of AI the training needed to improve it, purchase the GPUs, and the power/water needed to keep it running the price increases dramatically. The price of using these machines ranges from a few cents per prompt for something stupidly simple all the way to thousands of dollars for power users. All of this is happening currently in a market that is heavily subsidized by venture capital.
I could be wrong, but the math doesn’t make since unless you put yourself in the headspace that the people who push the hardest for AI aren’t making it to actually work, but are making it to attract as much investor money as possible before the wheels fall off this thing.
When it costs as much as a small island nation to build something that is a coinflip on accuracy, has the memory of a goldfish, plagiarizes anything and everything it comes across, lacks the ability to think contextually, has no clear sense of morals or ethics, all while feeding all of your input into a massive database where all of your information has a chance to crop up in someone else’s work.
This fancy madlib software will exist into the future and have its place as a google search summarizer or propaganda image maker, but there are fundamental flaws with it that hold it back from completing accurate, trustworthy, ethical work.
To quote IBM “A computer can never be held accountable, therefore a computer must never make a management decision.”
I consider high quality design and drafting work to be a “management decision.”
What it could be used for, and what it should be used for are two very different things…
Raw cost issue ? In the early CAD days I used desk top computers worth far more than my house using software that had annual per seat license costs that ran over $20,000. Not to mention all of the servers, wiring, power upgrades and training costs and software plugin costs.
Big companies will eat these costs if they feel it is justified and then push their lower tier suppliers/partners to do the same thing.
How many old school (non CAD) mom and pop machine shops exist today ?
I find this thread genuinely fascinating.
Many of us would benefit from reflecting more deeply on how our own minds actually operate. We’ve all been shaped by massive amounts of “training data”, experiences, books, conversations, media that we either repeat, remix, or use to draw inferences. The core process AI employs isn’t as fundamentally different from human cognition as most people prefer to believe. The main distinction is that AI executes it at vastly greater scale and speed.
We still hold an edge in certain domains, particularly those requiring deep intuition, physical embodiment, emotional nuance, or truly novel creative leaps, but it feels increasingly naive to assume those advantages are permanent. The gap is closing faster than most expect.
Absent deliberate barriers, it’s a very safe prediction that AI will progressively absorb the most time-consuming, repetitive, and rule-based portions of CAD work first: geometry generation, constraint solving, drawing production, basic validation, compliance, variant studies, and even early-stage concept iteration. The transition will be gradual and uneven, faster in mechanical design and architecture, slower in highly specialized or heavily regulated fields like aerospace and medical devices, but the direction seems clear.
If there is one thing I have learned about AI lately is that you must not think of AI as an LLMs. Why? Because all LLMs are AIs, but not all AIs are LLMs.
(ChatGPT is great at remembering and guessing interpolated between what it remembers, but it’s not “smart” and it can not use tools. Like calculus, it is bad at it. And intelligent sequences it messes up (1,3,6,10… 14?) It can reason that it should be five digits up, but still ends up with 14 because it doesn’t use math, it imagines math (yet)) In many ways it is a one trick pony: a language model)
To figure out more about this I have tested out a Sweedish web building AI lately and that is a totally different ballgame than asking ChatGPT to do mundane python scripting and failing with confidence… But with this tool I suddenly have an engineer that codes, discusses, tests and the stuff works. And I thought I should test it to it’s limits (beyond what I considered possible for it to be able to do, but that’s how you test something)
That started a journey on making a small, webbased RAW editor that works locally on images.
(It’s purpose is to let you quickly edit a raw file, while on the go, on a friends laptop, without logging in or installing anything, nor uploading your stuff, and then quickly save out a jpeg.)
A simple one that can adjust Exposure, Contrast, Highlights, Shadows, Whites, Blacks, Temperature, Tint and Saturation.
I drafted out the scope and started instructing and testing, instructing and testing and so on and so forth… and quickly saw that this could actually work and I needed a sett of tweaking sliders for an admin user too, since it started to became robust and could be useful. The tweaking was hard since I had to learn so much about how raw data works (like how different crushing pixel data towards 0 for a x% of the darkest pixels (Blacks) with a parabolic falloff is a different ballgame than blowing out pixels against infinity (Whites)) and how a pipeline works so I can instruct the correct order of manipulators. Sometimes something breaks because I ask for something stupid, and other times because it does, but every time it tries to fix the errors and can even bugtrack the problem and find the solution by itself, taking screenshots, evaluating them and figuring it out all alone. This stuff is nothing less than a gift to a designer, an engineer that never get’s tired or fed up because I go back and forth testing things and reverting things, discarding things and re implement things, over and over, never satisfied until it suddenly “feels right”.
Then I thought why not add support for EXR so I can tweak Rhino Renderings too.
And this is the result of the collaboration (so far):
My point is: No one thought this possible two years ago, and here we suddenly are, and if this can be done in 2D, then tomorrow it can be done in 3D. Will it replace web developers and engineers? No, but it will eventually replace some cad monkeys that just builds what they are told. So why not learn to be the one who tells?
Depends on availability of training material.
Re: toolpathing…some of the toolpathing programmes now have copilot built in. I have no doubt they’re scraping data to train AI.
Re: renderings…there is a mountain of images and training data available online already. Just look at the RunChat functionality available already.
More land speed car rendering experiments with runchat - Gallery - McNeel Forum
I can show you an example of Claude Opus 4.6, built me a relatively complex wpf xaml that left me spinning.
That would be great, I don’t know what wpf’s or xaml (html on stereoids?) are even, so please do!
This is one of the reasons I bought your great software. Happy to support passionate businesses who aren’t ruining their product for the sake of their shareholders…
Something to read through (yourself, not having an LLM parse it and summarize).An AI Agent Published a Hit Piece on Me – The Shamblog . Be sure to also read the follow-up posts and the referenced PR(s).
It could very well pose a risk to humans in the design/CAD industry once agentic tools become more utilized here as well. In particular the incorrect data being generated and the negative effects perculating into the future.
To begin with, I would like an AI-empowered MatchSrf tool.
I actually started using AI recently and before i started my drone project in rhino today i typed in some text to help me with my modelling
Because since my eye surgery I have been lacking alot and my confidence is low so this is my workflow to improve my Rhino skills:
And then I started with my drone project with confidence in Rhino 8
AI is just an AMAZING tool in 2026 
When an LLM-based tool works for you that is great 
I use the following base prompt when using an LLM:
When answering:
* no emotive language
* no emojis
* no reference to self (me, I)
* no yes-manning
* succinct and clear language
* mention sources referenced
* provide links where possible. If no links can be given then references to primary and secondary literature
I find the results much more palatable that way.
How can AI “judge” a rendering if it does not know about the context and purpose of the image in question? It’s a good idea to first “teach” it with a stack of appropriate photos, so that you can then ask with meaningful prompts. Here’s the first thing I discuss with participants in product rendering trainings for beginners:
Purpose
Before you open your laptop, the first thing you want to ask yourself is: What is the purpose of the renderings I’m about to do? Who is my target audience? What emotional response do I want to evoke? What story do I want to tell with my image? Basically, there are two general image categories:
-
Analytic: formal aesthetics (evaluation), CMF (colour, material and finish discussion), explanatory (assembly, functionality)
-
Evocative: design competition, client pitch, product launch, fair or exhibition, catalogue, online communication
I should feed it more images…
I don’t know, I’m just happy i have some time for some rhino modelling again and that post was ment as a joke 