Vibecode pyRevit w Claude-Code with Alex Ritivoi (pyRevit-Pod #10 )

Transcript

So, welcome. This is the first pyRevit Live episode and today I have Alex Retiwoi with me and he works at 3XN company as head of applied AI. And today I think we're going to talk about many different topics, but the main one we said is going to be about wipe coding pyRevit custom tools. And he's a also very unique kind of position I would say because um he did not really go programming route. He did not focus on learning programming basics and stuff like that.

He just jumped into kind of AI right away and he I would say like pure bread wipe coder, but maybe that change we're going to find out during the podcast anyway. Anyway, welcome to the podcast, Alex. How are you doing? >> Hello, hello. Good to be here.

Great. Bit nervous, but okay. >> I think we're going to get through that. Also guys, well, could you just let us know that everything works? You can see just write something in the chat so I can make sure that everything works.

We are currently might be on YouTube and live at the same time. So, if we Yeah, I can see somebody writing. Not in our River side chat, but on YouTube. I think it works. Anyway.

So, Alex, could you just explain us what does a day-to-day looks like for someone with the position of head of applied AI? >> All right. Uh yeah, it's a it's a good question. I don't think there's a clear answer at the moment. It's just that um Yeah, we've been exploring different tools the past 2 years and then um there's we start to see a lot of potential and we we start to realize that it's evolving quite fast and it's starting to become quite useful.

So, you do need someone who's up-to-date, who knows all all the non make sure and then who also has a bit of technical background in automation and knows a bit of context of of beam and all this, who can have like a high-level overview of okay, this is happening. What do we How do we react to this? What do we need to change? Do we need to start putting stuff in a database? Do we Do we start making data sets?

An MCP comes out. What does that mean? Where do you deploy it? And it's it's a bit stuff like that. I mean, the past 2 months it was just trying to roll out the cloud and trying to just show people hey, this is core work.

This is how it works. This is what it can connect to. This is what it can do. You can do a scripting grass over. You can connect it to Revit.

Don't do this. Talking to IT, trying to coordinate a bit. Okay, what what does that mean in terms of cyber security and all this? So, it's it's a bit of everything. >> Mhm.

>> It's a bit unstructured, but yeah, it's very exciting. And AI Successful AI adoption requires 100% participation from from the company. It's not a one-man task by any means. It just needs someone to be able to keep up with the new updates and try to to spread that knowledge around the company. >> Mhm.

Yeah, cuz I think at first title sounds like how you're just going to sit there and code with AI all day, but I think it's the least of your responsibilities. Cuz as you mentioned, like it's probably more about like security, making making sure everyone gets what it what what this technology can do even. >> It's a bit less coding now. Yes, it's uh yeah, it's it's just trying to to figure out trying to it's more of a translator and and uh proving the tools before we decide to scale something or not scale something. >> Mhm.

I see. And I mentioned that you started it like 2 months ago like going full full on on this. >> Yes. >> Okay. >> Yeah.

>> Okay, so maybe we recap a bit earlier. And I know that you pretty much jumped on AI the moment it kind of became mainstream like this public GPT what was it 3 35 one >> 3.5 I think it's the first one that was useful. >> Yes. >> Yeah, yeah. So maybe we just kind of roll a little bit back so kind of quickly go through the timeline how your kind of experiments with AI evolved and when you actually managed to bring it to the free exam.

>> Yeah. So I think it was I mean I was quite into parametrics. Uh it's just private parametric families and then a bit of Dynamo and uh when I was when I was younger I was playing Counter-Strike and all this and I would take a word phrase uh uh template and try to modify it and make a forum and all that. So there's a bit of exposure to coding back then. I still I wouldn't call that coding.

It was just taking other people's stuff and slightly modify it and knowing a bit of Photoshop changing the graphics and make it my own let's say. And then um Yeah, then then I start uh getting into Revit parametrics at school. I got quite far with that. Then say okay look a bit into massings into into Revit massings and how you can divide that and then adaptive components and then I say okay this is takes a bit of time to to to start placing the adaptive components manually. Let's go into Dynamo.

Maybe around 4 years ago a 3 exam and we had to do some class some very preliminary class control and I tried to learn in two three days Navisworks and I was like okay this is just not what we need. It's it's complete it's it's too much work for what we need. We just need basic clashes. Let me start doing it in Grasshopper. And then I was like okay it's a bit too much to learn Grasshopper in two three days so maybe I can do that in in a Python node and then I installed Rhino inside and and and from there I start like diving into Python basically.

>> Mhm. >> Yeah. Started with the your FilteredElementCollector. >> Oh, interesting. >> Yes.

>> But also that note is pretty much a PDF ebook with all examples I could gather on this. Many found it useful. >> Yes. And yeah, and then I said, "Okay, now there's this new ChatGPT thing. Let me see if I can just do it with that." And that's basically when I realized that I don't need to maybe memorize syntax anymore.

>> Mhm. So, by this point you actually already had exposure to Revit API Python, so you're like not complete beginner who is like, "What is that?" So, you already >> I would say exposure means a month of understanding what the for loop is and doing a few element collectors. I think that's how far I got until I said, "Okay, I'm going to not do it. I'm going to do it with AI." And and uh the further AI evolved, the more I kind of wanted to challenge for myself. So, like maybe I on on purpose not learn the syntax to see if I can completely use AI to automate a thing I don't really understand how it works.

>> Mhm. I see. And how did it go for you? Cuz you probably still exposed to code. You open code files from time to time.

Are you completely like in the chat and you don't care what's going on in the back? >> Obviously at the at the beginning you had to had tricks, right? So, >> Yeah. >> then yeah, I was just trying to find guides online, trying to transcribe them into text, feed them into into the chat model, see what happens and and at some point, you know, you do it it was still early. You still have to learn some things.

So, a bit of WPF. [snorts] I didn't want to spend all my usage into moving something to pixels. So, small stuff like that it's it's it's about seeing what the model can do and figuring out, okay, in order to make this work for me now, what's the minimum I have to learn? So, there's still a bit of something, right? >> Mhm.

Okay, I see you. >> And that something keeps changing. For example, I I learned how about like how to do a PR now and how to use GitHub and how to use Git and stuff like that. I mean, there's still some things you need to learn, right? >> Yeah.

Yeah, I think there's still foundation that it's all built on. >> Yeah. >> Cuz I think what I realized, and it it can be applied to any topic outside programming as well. When you know something well, cuz all of us has has some hobbies. We are we are really good at that.

When we start chatting with AI, it kind of gives you the answers, but it kind of it starts to drift apart or mention a few things which are kind of works, but people who actually do it, they know like, well, that's not really how we do that. >> Yeah. >> And I think that's why foundation still matters, so you can at least like, hold on. What Why No, we're not going in that way. Just do it straight.

And they're like, oh yes, you're right. And I think there's still this element. >> I think it's super I mean having the coding mental model, the programming mental model is still super relevant. Just memorizing the exact syntax probably not so relevant anymore, I would say. >> Yeah.

I mean, it's still it's good. >> obviously So, sorry for interrupting. Obviously, I mean, it depends what you're doing, right? If you're an educator, if you work at Autodesk, if you have a startup that has a Revit plugin, you need to know these things. But as a day-to-day architect who just wants to use coding to to for personal tools or for a small-scale company where you have 15 Revit license, it's yeah, I don't think it's it's it's it's that important anymore.

Yeah. >> Mhm. I think I have a bit different opinion. It's good way to start cuz now you can start real quick. You can begin creating loads of plugins, web apps.

There It's It's crazy where AI has come to. But I think at certain points, you especially when complexity catches up and you like stop doing like simple tools, you want to like huge AI multi-window and stuff like that. I think there you still have to kind of stay in control. And that's where you still kind of That's usually so it's reversed. Before you spend months learning the boring basics, you're printing like class of the dog then animal do all this boring stuff and then you slowly start doing something useful.

Now we start building something useful right away and then we like in reverse look, okay, how does it actually work? Not everybody but I think that's also their main part. >> Do you have an example of this of this complexity where where it got too complex? >> It's often due to the size of the project. For the private tools, luckily it's often short but once you're going to create um VBA form where you create a lot of events, styling, like there are like multiple tabs you want to go through.

There is like interaction for example, I was looking into like dragging interactions. >> Yeah. >> And at first it worked while I try to I code something else, it breaks the other thing. So it's all about this complexity. And what I found like if kind of if I start kind of manually and give it the base and then kind of ask AI to do on top of that then it kind of has the structure and it's great.

But often times when I started just with AI, sometimes it does it great great structure from the start and it's easy to go through but sometimes it just kind of I don't know, it keeps changing what it did initially. And that's where like it it changes this module but that part of like maybe UI or maybe your logic how it works, it kind of it did not match. Cuz what I like I like the comparison. I think I saw it somewhere on internet maybe I just made it but AI is like the smartest kid alive but at the same time the dumbest one. So you have to also kind of remember yes, it's it knows everything but if you doesn't direct it well, it can also kind of start doing weird things.

>> Yeah. I what I like to do I I'm I'm deep into into the cloud ecosystem. And I think like what's called now harness engineering, right? So whatever whatever you build the things that guide the model, let's let's say like that, right? The doubt cloud folder in this case.

>> Mhm. >> That's probably starting to become starting to become quite important, right? So and also getting the feel of the model, right? Cuz if you go talk to Opus 5 and if you go talk to Opus 4.5, it's feels differently. I think it's important, I mean, you have people who who who use it in the IDE, right?

And they have access to all the models. I try to pick one. I'm using all of them, but I try to focus on one and get the feel, okay, what is what is the baked in intelligence in this model, right? >> Mhm. >> Cuz 4.5 and 5 have different data sets.

A lot a lot of what it can it can't do is it's baked in inside itself and it behaves slightly different, right? So if you combine that with with with the proper rules and skills and guides, then you really get the most out of it as opposed to if you just keep switching to a million ones in use, maybe it's like a generic agent that I'm building or something. So probably the best way to kind of learn how harness works is to focus on one ecosystem to dive deep into one ecosystem rather than than switching through all of them, yeah. >> In this case, you probably refer to cloud code mainly. >> Yeah, yeah, yeah.

>> When you talk about harness engineering, do you mean specifically like cloud code and setting up all these rules and skills, commands, kind of hooks inside of cloud folder, or do you actually mean like >> [clears throat] >> like creating your own kind of >> Yeah, it's it's it's it's it's an over overstatement to say harness engineering. Yeah, it's it's it's the dot cloud folder, basically. >> Okay, got it. >> And in par- in parallel, I'm also we set up a Hermes agent. >> Mhm.

>> Just to see some The difference between Hermes agent and and cloud code would be that it kind of just get task to simplify just get task and it builds it does the task, sees the patterns, and it builds skills for itself to recreate that pattern. >> Mhm. >> Right? Uh also on the list is to try Pi agent, which is probably the most vanilla harness harness you can have. So, that's that's a really good learning project to kind of start from scratch.

>> Yeah. >> Uh Yeah, so calling the dot cloud folder harness engineering, I think that is a bit of an overstatement, but yeah. >> No, no, that's fine. I just wanted to get it clear. Also, maybe we just take a little bit step back for people because not everyone really Everybody hears about this kind of AI harness, AI models.

They don't all They always confuse the two. >> Yeah. >> Would you like to comment on that, give your kind of opinion on like how to separate what is a harness? I can also add to that, but I just wonder maybe you start. >> Right.

The model comes from Anthropic, and you can talk to the model through an API. Through an API call, you can embed a model call in your Revit tool, for example, if you if >> [snorts] >> if you want to place views on sheets, right? That's not the most deterministic thing to to to automate. Here, you can you can just grab all your views, grab selected sheets, and send it send the names to to cloud model, and you get back a structured response and based on that that's how you how you uh how you placed the views on sheets. Um So, the model by itself is quite good but not that useful.

Like whatever is built on top of it, the ability to add skills, the ability to connect an MCP, the all the baking tools that come that come with the model that is all the hardness and that is pure deterministic coding basically around API calls to Anthropic in this case. Yeah. >> Yeah, I'm also going to give a very simple analogy. Like think of LLM models as a brain, but harness is the hands that actually kind of do something. Yeah, like Alex mentioned, that's where all the skills and everything how it's implemented.

And also the how the context managed. As simple as the context. Something has to happen when you going to fill it up, right? There's already rules, okay, we'll come back and stuff like that. So, this is the work of the harness.

And also I think big one thing I nearly also missed that. Harness can also get access to your computer cuz model is just online somewhere, but harness is what actually brings, for example, cloud code on your computer why it sees the files and stuff like that. They're just again, it's just tools. >> Yeah. >> So, I just wanted to make sure that I don't know some people already really far ahead with AI, but some people still kind of just getting started.

Cuz we already start kind of throwing in like Hermes agent, Pi and stuff like that. >> yeah, yes. >> And I don't know guys. I'm going to ask in the chat, would you like us to even talk about like different agents? Should we focus specifically on cloud?

Cuz there is like dif- I think we can >> We just we just set it up. We just set up finishing up Hermes and the plan is to start Pi as well. We haven't used it that much, so there's not many insights I have there. >> But that's okay. But for the Pi, do you Is there like actual Pi agent or there's like Pi just harness like Cloud Code?

>> It's a harness on GitHub you can download. It's completely open source. Yeah. >> Okay. So, with the Pi, I looked at this one as well and it's pretty much imagine Cloud Code with nothing where you add all the modules.

And >> It's like for people who love DIY. And I look at it on one hand I'm like, "Oh, this is awesome. Maybe I should try." On the other one, do I really want it like I decided not to go there. But, I see there's like a lot of people like interested and they kind of like to customize it and I think it's a great for people who like to kind of customize all the things. If somebody like watching us wants to kind of try it out, that's pretty much it.

You you will be responsible for all the functionality. >> As as an architect who wants to just have AI help himself >> Oh, yeah. >> Setting up a Pi agent for you would probably be too much work. But, just playing with it just for for learning experience, I probably that's that's where it is. Yes.

>> That's what I heard as well. Everybody who starts with Pi, even if they delete it afterwards or they keep it, they they just learn what is a harness, how it works, and like then they when they switch back to Cloud Code they're like, "Oh, now I get it. Like this is like something something's happening behind." >> Yeah, I was also trying maybe on this topic I was I was trying to set up Open Claw when few months ago when it came. And by the time I finished setting up and understanding how to use it, Cloud Code already caught up with all the features that made Open Claw good. So, it's a bit of a Yeah, it's a bit of a realizing the the the speed this thing have.

So, you have to be a bit calculated what you spend your time on and yeah. >> Yeah. Yeah. I think uh now it's so hard than ever to keep focus. Cuz once you just I just choose one little thing and then there're like 10 more pop up the next week and then I'm like I did I make the right choice then I start researching and like now I just like okay whatever happens I'm going to stay with Claude code for at least like I don't know as long as it works for me I'm going to keep it or until they ban me for some reason cuz >> [laughter] >> that's also a thing and >> I think that's the way yeah >> yeah and then I just but one thing I kind of going much deeper into AI but still I want to mention that one thing I can recommend to people right away is that you need to plan for if how to say it you need to plan in case for example you get banned from Anthropic cuz I see people they build the whole business around Claude code they wake up one day they didn't do anything wrong they just banned there is no like easy support ticket to resolve it it's like pretty much if you banned probably you're going to stay banned for a long time and what I noticed what I realized at least for me that I need to be a little bit not only model agnostic well model I don't really care as much they all really good now but I also need to be a little bit harness agnostic the way I found the ways just to kind of create my own memory then also find a way how to kind of um credentials this is easy I can like always set it up from the start but memory is the one that I really really wouldn't want to lose at least I found this really useful because now I set up the memory connected my Claude code and if I want to go to Codex I have exactly the same memory connected I use Honcho but there are many options >> yeah I think uh like ecosystem locking or or like model or vendor locking is something we really want to avoid so everything we decide to spend time on >> Mhm.

>> should we should be able to take it out right and the the dot Claude folder all the skills you build all the MCPs you build are compatible or easily be adjustable to work with other models so it's it's I mean the skill and the format is is a it's a standard now right the MCP is also an open open standard standard yeah so it's it's even even if you decide to go with one with the open AI or Google or or even all all the stuff you can get it out eventually, right? You can download all your chats. You can download the memory. >> Mhm. Yeah, yeah.

No, that's a good way forward. Anyway, I think we going to go a little bit backtrack cuz now I think we going to deeper deeper deeper where it's a bit it's getting harder to follow, but okay, so mentioned I want to get back to like how you in Free Axon actually started implementing it at scale cuz at first I would assume it was maybe you or a few other power users using AI. >> It still is. It still is. It's we haven't rolled it out yet.

It was a small pilot project. I mean the the we had an AI group. We have an AI group to we that started 2 years ago and it did amazing work. We we implemented uh OmniChat, which was an an AI aggregator, so you had access you had like a chat kind of chat interface. You had you had access to all the models.

And you can chat with it and uh at the beginning of the beginning of the year we just started realizing, "Hey, you can do actually a lot of of this non-coding work using agents." And then okay, let's start with five people and now we are at 25 people, I would say. So, it's still far from from from from scaling up. But the proof of concept, I think, is done. >> Okay, so that's viable solution to go forward, introduce agents, and >> Yeah. Yeah, it's it's worth it now, basically, yeah.

>> Mhm. I started hearing literally in like last month or two I started hearing from a few people in personal chats that their company is starting to also roll out similar things. So, I think like 2026 finally it's catching up to companies, okay, there is no way out of it, but also most people just don't know what is going on. Probably, like such a topic. On one hand, it's simple.

You just talk to AI, but on the other one, there are so many things kind of hidden. >> Yeah. Yeah. It's You know, there's it's under-hyped and over-hyped at the same time. >> Mhm.

>> In a way. And probably the only way to to know that, the only way cuz if you go on on if you go on on LinkedIn, like your job is done. Everything can be automated, right? And if you never use it, you think it can do anything, right? So, I the best thing to do is to just try it out on your use case, on your company, on on the work you do and see, okay, are we there yet?

Can can can it do it? What can you do? And that will inform you what you need to change, what needs to be implemented, what is just an API call, what is just a UI wrapper, stuff like that. Yeah. >> I 100% agree with that one.

Cuz I was very skeptical of AI. Not really skeptical, but I had a real kind of internal battle. Do I kind of jump into that? Do I kind of stay more in control? Do I teach it?

Do I not? Now, it's obvious. I'm like, okay, I'm going to go all in and try to get as much out of it as possible. And different people learn different. For me, I just need to break it, figure out how it works.

And what I decided to do, uh I just took like this close to $100 a month subscription. And my goal was, okay, I need to figure out how to max out. I want to like use every token. Apparently, it's not easy. Very On $20 a month, it's easy.

Like, you start talking, it's gone. But then 200, what people don't realize, you get about like 8,000 worth of tokens, like in dollars. Cuz it's very highly subsidized. And now, they even like offering some discounts, so it's like 15,000, I guess, or something. So, there's like a lot of tokens.

And once you start talking, like there's no way you can just open one project and kind of talk back and forward to max it out. Because at first I see everybody complaining on like $20 a month plan, I thought it's going to be quite similar if it's going to be longer session. Now like I didn't hit it. I start working like two product, three product in parallel. I'm like, okay.

I started using sub agent. Sometimes like it goes crazy. I'd like burn through all the tokens. I'm like, okay, I guess I should be a little bit more careful. But this actually also showed me a lot about the models, what they do, how they work.

So I think like for people who like to try stuff, this is really great way. Like I mentioned, just try and build some stuff. It doesn't need to be useful. It doesn't always need to be useful, I would say. In terms of it won't always come up come out perfect.

Because I build like 40 different projects and some of them used by 1,000 people. Like for example, RVT dogs. Lately it was heavily built with AI. I just kind of fixed bunch of things here and there, but it's heavily AI built. And then there are other projects that I built.

I looked at it. I'm like, well, it doesn't really work. I need to restart or something. I just I I just ignore it. So this really helped me a lot.

Just you don't have to really kind of focus on one thing only. And if it doesn't work, it's it's not useful. No. It definitely can be useful in many ways. >> And maybe a bit on the on the subscriptions, right?

A $20 subscription won't really show you what's possible because Yeah, when it gets interesting is when you don't focus that much on like optimizing talking use and and stuff like that. It's just like, okay, let's put the the the put throw the bazooka at it, you know? Put the highest model and let it go. Close the loop. Let it self-verify and see what happens, you know?

Have a safe environment where you can test this and just try a lot of things. And yeah, I I personally use a $100 subscription. I have a ChatGPT one and and a codex one. And yeah, it's it's getting harder and harder harder to max it. I chose to try to max it every every time.

But I think on $20 won't really show you exactly what's possible. You're just going to run out. It's just going to be annoying, right? Uh but also this other other thing now cuz what if these subscriptions all of a sudden go away and then all of a sudden you you're you develop all your workflows around a $100 subscription and that $100 subscription becomes $8,000 a month. That's that's that's a bit scary, right?

But now like for example, Kimi Cat 3 came out and I heard I haven't tested it but I heard really good things on it. So now I'm a bit more relaxed. It's like, "Okay, it's not that hard to to set up a virtual GPU and put it there. Maybe it's way harder to scale it to 200 people, but at least it's it's a checkpoint. Okay, at least we know we have Opus 4.5.

We can host it ourselves. So we're at least there. Maybe not Fable, maybe not yeah. So >> I'm fully committed. Actually says like it's on par with Fable GPT 5 3 Opus.

>> Opus 5 I would say, yeah. Yeah, Opus 5. >> so many metrics as well. And some models they optimized for the metrics. There was like I think Opus 4.7 or 8.

They definitely fine-tuned it because it just started on certain it just like it went way ahead in the scores even though the model wasn't that more impressive. It's still impressive, but it's not that far. So they started kind of hacking the benchmarks as well. >> I know. Yeah, I mean you can Yeah, you can see with the with the ChatGPT model.

They they all of the all of them they're training on benchmarks I think, right? If you see the ChatGPT model that break out and went on Hugging Face to to tweet the >> Yeah, so for people who don't know, they were testing GPT 5 6 and they reduced some kind of guardrails and it decided instead of making the task, why don't I hack a place where the answer might be? So that's kind of the whole story in short. But that's also marketing as hell. I think >> I also think so.

Yeah. >> I think it's 50/50. Maybe it did something like that. Maybe it's on purpose and marketing because Hugging Face got a like now full support from OpenAI, so they benefit anyway. After Anthropic and like and Fable, it's too dangerous to release to the world and stuff like that.

And they're like, "Here." I don't know. We have to be careful about all this AI news. >> That's why it's important you test it yourself and you get the feel of how the models behave. That's why it's like I feel like five coding is such a fitting description. You really need to get a feel.

And it's also >> Yeah. >> I didn't never believe in prompt engineering. You know, it's it's just like using the model to develop intuition to talk with it. >> Mhm. >> There's no rules to prompts.

Like you speak English. You just be clear. As Yeah, okay. There's tricks where you say you can't don't have fingers if the model gets lazy, but that's that's not valid anymore. >> Mhm.

Yeah. I mean, the more I use it, the more I know where it's going to go. >> Yeah. >> in the wrong direction. Sometimes I I I try to be lazy often as well just to see is it going to do it?

I ask it some often times it does it well, but also because it remembers how what we do and stuff like that. But, from the fresh chat, sometimes you need to be a bit slower to get started kind of cuz it also learns about you. >> Yeah. >> And when you look into these memories that it collects on you, it actually gets so scary. Cuz I set up my own memory and it chat it doesn't just save your chats.

It thinks about it. It dreams. It makes conclusions. It's like, "Oh, Erik likes UI design in this way. Oh, Erik is over like analytical and over um paranoid about this part." And I'm like I'm reading that.

I'm like, "Oh." Cuz I didn't write that I'm that. I'm just talking to it and it's like, "Okay, Erik is this. Erik is this." And it just builds it up. >> Yeah, I'm not I'm not happy with any memory system I've seen. I also not like with the like memory from from fraud app projects or ChatGPT projects.

I don't think it memorizes the things that I care about. >> Mhm. >> It's just it's it's Yeah. [clears throat] So, what I what I try to do, I try to have notes and logs. Just put set up a hook.

If we're doing a private tool, save notes of what worked. And then I have those notes and at some point I'm going to take all those notes and I'm going to build some sort of memory on top of it, yeah. >> But for the memory, do you now talk about like built-in this kind of cloud and ChatGPT? Or is it like externally like Mem Zero, Mem Palace, Honcho? >> Well, there's the open-source projects, right?

And then there there are there's also the baked-in app >> Yeah. >> chat memory and also cloud called slash memory, right? So, it's it's None of them are okay for for me at the moment. >> Okay. >> They're not Yeah, they don't memorize how I want things to be memorized, yes.

>> Mhm. Uh I can share like I can share about Honcho memory. I think it might be useful for people as well. So, I also went down this kind of memory stuff and they're like I think they're like I'll just put two categories to simplify. They're like basic ones that just remember what you did and they like just keep the kind of log pretty much what you do manually but do it for you.

There's a few extra kind of to make it more readable, but then there are other memories which are more interesting. These are which start to also do something with your data. They kind of dream, they try to organize it. There's like There are a few cool projects. I specifically went for the Honcho.

And that one is it gathers all your information but also I'm sorry. It tries to build connections. It build a peer card. For example, Erik comes a lot Erik comes out a lot in conversations, so it's going to build a peer card about me what what what I do, different facts and stuff like that and certain conclusions, stuff like that. And when you talk to ChatGPT, what it gets, it gets your prompt obviously all your rules and stuff like that and it also kind of gets its internal memory but also it gets my hunch memory and sometimes I don't always see what exactly is provided but sometimes I see it kind of it's more aware of what I do across everything.

I think I don't think there's perfect memory I don't think it will ever be they always change too much but I think at least trying something probably is worth giving it a try as well. >> Yeah. What's the difference between cuz you you tried have you have you You first probably have car party LLM Wiki. >> I'm building it now. >> You're building yeah I saw I saw that so what's the difference between graphing such a Wiki and this memory system for example?

Is there a difference or is more or less similar? >> Yeah there's difference for sure. So first of all memory it's literally a capturing everything you talk about every kind of your prompt every response you get I literally have all the chats saved and it kind of learns from its own memory is saving it somewhere. >> But it's still graphing it's still graphing it in a way right? >> In a way it's not trying to cross I don't know exactly internals but it's not trying to cross reference everything to source and stuff it just kind of learns about you like little facts here and there it's not as in depth probably.

Probably there is some similarity now that I think about it. But LLM Wiki first of all the main idea is that you bring your own sources so you curate information it's not like okay go to internet and grab me everything related to API pirate related it's going to be a mess. Because all of the system they have flaws and for LLM Wiki the flow is if in your sources a bad information comes in and bakes in as a Wiki now it's a fact. And you still can kind of there are tools to kind of clean it up and it's supposed to do that but often times it will miss it because it needs some sources. So in LLM Wiki what I found is like you are in control and you also decide how much you are in control.

Some people just go crazy. They Here's like 1,000 notes, do something. And Karpathy tells himself that well, what he likes to do, he pretty much adds like one or a few files at a time and then kind of gives his judgment. So, AI parses it. He's like, here's what I learned.

And he's like, no, yes, no, yes. So, you still need the judgment layer. That's what I like about LLM Wiki. You can stay in control, but you choose how much control you want. >> [clears throat] >> And for me personally, I decided to go a little bit I think I went too much because I spent so much time trying just to figure out how to control it and I decided to change how I do it and it was like a big change to everything.

So, I I added a lot of verification layers. So, I add all the content. I say, okay, go and check my course and get information out of it. Then it kind of reads it and then I have to go and verify pretty much all the concept it builds about it. It's a lot of work extra, but for me it's kind of worth it because I want to make it once really good and then later I'm going to loosen it because once it has the base and it's rooted everything on the real content, this is what kind of suppresses hallucinations because hallucination happens when there's like so much data, it sometimes kind of makes stuff up because it Sometimes it can be either just a wrong block.

Sometimes it just kind of makes stuff up because it thinks that's what's next. It forgot to reference source. In LLM Wiki, it's still kind of everything goes to the sources. At least that's what I my conclusion so far. So far, I had good results in terms of I just kind of gave it a bunch of data.

Had a look for example, I didn't go through everything. There's a lot, but for example, selection concept, it nailed it. It went through. I still need to give some judgment here and there. I'm like, kick this one out, kick this one in.

And yeah, if you take it slow, it's going to be really powerful. I'm still to test, so everything I'm just take everything with a kind of grain of salt. It might work well. Sometimes maybe I'm over constraining it. That's also something to keep in mind.

>> Yeah, I assume it we start talking a bit about what would the company memory or knowledge look like, right? Because you have you have the company data, right? That is just files you place there, right? And then you have all the knowledge, the wisdom, like what does that even mean? >> Mhm.

>> And uh so we start starting to explore a bit that and I think just a as as a first step is just to be able to capture all that information that doesn't exist, right? So you finish a project, you end up with drawings, you end up with a design book. You might have minutes, you might have but a lot of stuff, a lot of like, okay, we forgot to do that because of that is not really captured, right? So like clean solution like temporary solution would be just to have a {slash} uh project command that where you can type whatever and then it sends out to a router, it knows who you are, it kind of infers what project you work on and it just saves that information there. And then the next step would be once we have some information there, then we'll probably look a bit into a memory system, a bit into a probably Microsoft or or someone will come with something that works that you can then run on on the data.

But I feel like capturing it is probably step number one. >> Yeah. >> Yeah, yeah, yeah. Also cleaning up is a big thing because well, garbage in, garbage out in the same way if good things go in, good things come out. >> But that's the thing is if you if you do the capturing intentional, yes, it's a bit of work, but then you're sure it's mainly high signal stuff that's there.

So you already can start in two, three months have a little sample set you can test different memory systems on. Yeah. >> Yeah, also all the systems they're going to grow. So the first thing as you capture data, there's just going to be new system, there's going to be new plugin open source project like, "Hey, plug your data, something will happen." So, this there are more projects coming, and the question is like if it's really that useful. Yes.

I know that Pinecone, which is the leader of rag, they just recently made a blog rag is dead, but that was also promotional because they said, "Well, we solved it. This is the next thing." And they built on top of LLM Wiki, so they have they're cooking something like between rag and LLM Wiki, I think. >> Yeah, it's a Google also released this OKF open knowledge format. So, now it became it went from a tweet to a standard. It's it's a bit funny in 1 month.

>> Oh, it's Andrej Karpathy tweet, the man who made wipe coding also a tweet. >> Yeah. >> The term itself. >> Yeah, if if you want to keep up with AI, he's one person you should follow, basically. Yeah.

>> Yeah, yeah, startups there's he has way too much stuff because there's like 3-hour long sessions. I'm like, "Oh my god, why it's not 15-minute video I could watch now?" >> [laughter] >> Anyway, um yeah, I think we're just kind of drifting a bit. So, maybe we'll get back to wipe coding for pyRevit with cloud code. So, can you just introduce us to cloud code? Or maybe talk how did you get introduced to cloud code?

>> It was So, I was using I was using ChatGPT for big portion of time. >> Mhm. >> Then, at some point indexing came with folders. So, I'm starting to gather data a lot. A lot of my initial tests I did ask for permission, but were on your toolbar and on the pyRevit tool toolbar.

So, I was just like getting the library, trying to extract useful definitions, and trying to do some sort of database. So, that took me quite far. Um Yes, and then I think in the beginning of last year, I was trying cursor and they start introducing agents and and all these like rules and stuff. So, we we we just start populating just like, okay, let's let's try it on the toolbar. Um and in in September, I think there was this Google principal who posted on Twitter that we managed to automate 6 months of work with using Cloud Code and then and then I just installed it.

And I went on Lucid Icons. I downloaded all the alignment tools. So, I just wrote align. So, you have, you know, in a design software, you have all the alignment, align equally, align to left, align to right. So, describe all the tools and one-shot this Fire It toolbar.

>> Mhm. And what was what was the goal of the toolbar? Just the structure or actual working tools? >> Working tools. >> Okay.

>> And it didn't work. And then I was, okay. How how can I make What What What do I need to put here to work? Okay? And then I went on I went to the Fire It website and I I I downloaded the structure of the folders.

How you have a panel, you have an extension, you have like that. So, I made kind of a little helper skill from that. And I tried again and it didn't work. And I just keep doing that until I got to a minimum amount of things I need to put in there to kind of one-shot this toolbar >> Mhm. >> uh just from one prompt and the skills attached.

>> Can you describe approximately what was the toolbar? I mean, what was supposed to be inside? Just roughly. >> It's uh it's uh it's all the alignment options. Uh Do you know the plugin Beam 42 Align?

>> It's all aligning elements to left, right, distribute. >> Exactly. Yes. So, all the possible options that work on annotations, but also a 3D, also annotations and 3D combined. >> Mhm, got it.

>> Yes. >> Okay, so okay. >> And I got it to work with two skills, I believe. Some element hierarchy skill and the how how Revit is is aligned, plus a dump of all the icons. And the prompt was describe all the icons, uh try and one-shot a Revit toolbar.

Arrange panels logically so they display intuitively in the UI. And deal where you have to do some translations between 2D and 3D elements. >> Mhm. >> And it cooked for 15 minutes and I got a toolbar. And since since then I was like, okay.

This this is super useful now. This can be scaled. Uh I need to look a bit into into into the skills. And since then I started using Cloud Code and yeah. >> Mhm.

Okay, so you got your first win and you like, okay, that's actually something that can bring value and you focused all in on it. >> Yes. >> And at the time, what was the model? Do you remember? Was it already Opus something or >> It was just Opus 4.5 came out.

>> 4.5. Okay, the 4.6 was a big one. Where it really kind of got much That's where everybody started to realize like, oh, Cloud Code and Opus actually actually can be really useful. >> Yes. >> I see.

And from that point on, you just kind of focused, okay, you started building tools or skills? What do you kind of focus more? >> I'm just trying to understand, okay, what's what front matter do I need to to to have? What's like the minimum amount of knowledge I I need to know in order to make this work and keep testing. And again, like you, I have 40 projects.

And I get stuck with a lot of them and then new model comes out and then I continue on that project. >> Mhm. >> And I realize also with skills and MCP, there is is a fine is a balance cuz if you clutter it with skills and MCPs, it won't work. So, what is the minimum amount of stuff you need in for the model to do its its work. So, for example, for our active toolbar at some point I say this is this I'm drifting a bit apart.

Uh we keep breaking our tools. So, we decided, "Okay, we're not going to have a library anymore." So, each tool is going to be self-contained. >> Uh so, not reusable code anymore. >> Yes. So, each tool is self-contained and um >> Mhm.

>> So, so the toolbar so has it has no dependency. So, you can iterate on one tool and you can just deploy it and if it has an error, it doesn't not much happens. >> Okay. So, if something breaks, it's only one tool. >> Yes.

>> Ideally. >> [laughter] >> Exactly. Yeah. I mean, that started to use a lot of tokens, become a bit inefficient and uh then we say, "Okay, we're going to I need to make back back to library, right?" So, now we have a little skill that just explains how we want the library to be arranged. Right?

Another thing is I can take a lot of time and I'm getting I'm always rabbit pulling in icons. So, I went on Lucid icons. You You know the the the page? >> Mhm. I've been there but I don't remember how it looks.

>> Very simple SVG uh open source library and I just you can download all the icon names. You can describe a bit what they do. And >> Mhm. >> you just make a tiny skill that has the library of of all these icons and you make a rule telling the model, "Okay, when a new tool comes in, figure out what is the appropriate icon to put in, right?" >> Mhm. >> And another rule, how are our tabs arranged?

And there's no there's no there is no no real coding skills in it. There is a hook that logs all the gotchas. >> Oh, that's a good one. >> So, that keeps building. I never cleaned it.

I never actually looked at it. >> Mhm. >> It just works. So, we got to a point now where you just open a bundle creator, someone write codes a simple script that they need, send it over, I dump it there, and I said, "Add tool." And it adds it to the toolbar. It breaks often, but again, we're everyone who uses the toolbar sits 50 m around me, so it's it's not that big of a deal if you get a if your Revit crashes once in a while.

Doesn't happen often, especially with failure, doesn't happen anymore, but yeah, it does happen. Yeah. >> Yeah, I think that's a luxury of small offices. Not small, but more like >> Yeah. >> Uh when everybody who uses your tool is around you, >> Yeah.

>> it can be much but because I heard from some people when it's already multi-offices, even multi-countries, that's really tricky one, but still with Git and everything, it's possible to resolve easily. >> Yes. Uh another skill would be the the use of the WTF design guide. >> Mhm. >> So, stuff looks consistent, so he knows where so model knows where to get stuff from from the library and and stuff like that.

Yeah. Having a map of of of the tools in your cloud that I need file helps a lot. And that's pretty much it. >> Mhm. >> Obviously Sorry.

Obviously, having the MCP connection to Revit, so the model can test the tools, so you close the loop. That That's also helpful. >> And that's also what we're going to dive deeper so in a moment, right? Before that, I also want to touch up on the You told that you're building up the gotchas file. Which is also I also build one as well.

>> Yeah. >> In my case, it kind of went a little bit wild. It built like thousand something lines and I was like, "Woah, woah, woah. Show me all of it. It It This is not This is not good." I started going through and it's such a little picky things here and there, I think.

I think you need to also verify it cuz mine got so big, it definitely will start losing information. >> Yeah. >> And I looked many of them they are they're not even gotchas. They're just like "View has to be unique." I'm like, "Cool." Like mention the solution at least because I looked at it and it's something very stupid. This one is not bad actually.

>> Yeah, that's a good example. That is a good example. >> Yeah. >> Yeah, that that the view has to be unique. You don't need to give it any code and it's one sentence.

Slow tokens. It's quite high signal example. Yeah. >> It's good to include how you want to solve it. Cuz for me, I like to add the stars in the end.

So, if you try to rename floor plan one, it exists. I just add one star or two stars or three stars. So, you know it still renames it because if it doesn't rename, sometimes you lose the view. Cuz like five views move to a new place, one left somewhere and you're like, "Where's my last view?" >> That is actually in our rules as well. We I I like to use a the date year month month day day minute minute second second or hour hour.

As a unique identifier. Yeah. >> Yeah, because the thing about skills, there's so many different ways to build them, but one thing everybody agrees on the smaller the kind of the less information in your skill, the more efficient it is. >> Yes. >> Cuz once it gets longer, I know I also learned that on a large skills, it prioritizes the start and the end.

And the middle, you never know like it it might lose something and stuff like that. So, that's why I am so keen on like trying to clean up, but it can be sometimes hard. Even I read some of the gotchas, and I'm like, "Where did you even get it?" Like, I know it went for my files, but I don't remember working on that thing. >> Yeah, I think you know, if you quickly want to work on something, it's fine. You use the skill creator, you tell it make a skill out of this API, just implement it.

It doesn't matter. But, if you have a skill, like for example, we have three first skills in the private full bar, that's quite important. It's good you do you go and create it yourself, and remove There's a lot of noise. There's a lot of markdown decorators that don't need to be there. There's a lot of Yeah, it's just it tends to to overshoot skills most often.

There is this guy called Matt Pocock. You you heard of him? >> Not sure. Maybe what what's the project? Or what's the name?

>> Uh >> [snorts] >> can I maybe share >> I can just describe. Oh, yeah, yeah, yeah. Actually, people already asking in the chat if it would be nice to share Revit screen. >> [laughter] >> They're they're missing their domain. >> Let me just see if I have it here.

One second. >> I think I came a little too early. >> Okay, how do I share? I share. One second.

>> Also, when you share, make sure it's a screen, not a window. You know, this annoying thing. >> Yes, can you see my Revit screen? Yes. >> Hey, yes.

Maybe let's make it bigger. >> It's this guy, and he has this skill that we're using. It's called the writing great skills. That's probably the best explanation that I've seen how to write this clear a good skill. This is what I've been using.

I mean actually all his skills are really good. Especially this one is also has a glossary, right? And it tells it it just tells you This is this can be a skill for yourself, right? It is it tells you very >> Oh, Alex, we see your Revit screen. I was thinking that you were still opening something.

>> Oh. >> I think you selected the Revit Revit window. So when you share screen just make sure it's the screen. There's on top like there should be some kind of selection between applications, screens. >> You just you just told me that and I did it wrong, okay.

>> Once I I can see it. I see Revit. >> Yes. >> Yeah, now okay. >> Yes.

>> I was confused. I thought maybe I missed something. I thought I was waiting for you to like pull it up and then I realized. Sorry. >> This guy this this guy has a great skill, right?

It goes through all the stuff that you need to know, right? User invocation, model invocation. User invocation if that's the skill up here where when you type {slash} can you see it, right? And and model invocation is does the model see it and can the model invoke it by itself, right? And it goes through all the concepts of skill and it has an amazing glossary.

>> Is this something internal or it's also open source by somebody? >> This is this guy. It's Matt. This it's just a influencer. It's a coding influencer online.

>> Okay. >> And the skill is I mean it is in our repo but it's it's still Yeah, I can't remember where to find it. It's probably somewhere in this four repos. >> Mhm. >> Right?

And it has it has the glossary. It's really good. One important thing that I see a lot of people missing their skill is is the last point he has here which they don't consider model baked in intelligence quite often. They tell to model to the model and the model gets updated. We need to also need to understand that all the skills that we make and we put online will be in the next model.

So it's so skills are something hot that constantly needs to be curated and removed and tested. Because you want to use the baked-in intelligence and just like keep it keep exactly the the the the high signal guide. >> About that, there's recently the creator of Cloud Code, Boris. He literally I think that week or last week he went out to some presentation and his kind of the main point was like well every 6 months you should delete all your skills and just test what model can do by itself. Cuz many people still have the skills from like a year ago or 2 year ago and that's exactly what you mentioned with this last point that there's a lot of things we include in skills that model just knows.

You don't need to describe how to create a folder in Windows, it knows it. You know, it's a oversimplification but you get the point. And as you mentioned also that everything feeds back to kind of Anthropic, that's why we're getting these crazy subsidized tokens and stuff because they need our data more than money at the moment cuz all the skills, everything that there are like people who create crazy kind of workflows and stuff. Anthropic sees all of that cuz Cloud kind of gets it in and out and it kind of learns I'm like oh that's a cool one. I think that's also something that's a really good point that you mentioned.

>> Yeah. I don't see actually I don't see anyone online selling skills which is very interesting. >> I recently saw that there's starts the boom of selling the skills but I think it's very very very niche. I don't even know how they sell because in the end of the day you're trying to sell a text file. On one hand I get completely get it like if you going to sell like it's very niche application for sure.

>> My advice never buy skill. It's uh Yeah, so an analogy used before with the harness, right? A good analogy I like to give people is as a kitchen, right? So, you have the kitchen, you have the chef, you have an oven, and you have a let's say pasta bolognese recipe, right? Like the kitchen would be the harness, the chef would be the model, the oven would be the MCP, and the recipe would be the skill, right?

You can find online >> It's so good. >> ton of bolognese pasta recipes, but you'll never find the bolognese pasta that your mother makes for you, you know? So, that's that's the skill. That's how you need to think about the skills. You'll find, I don't know, accounting skill a lot, but you'll never find a skill that that tells the model how you do accounting at your company, right?

So, that's what you need to put in your skills. Yeah. >> Oh, that's such a good analogy. Did you make it or >> Uh I'm I'm not sure. I don't want to give to take credit for it.

I'm not sure. I've been saying it for so long that I'm not sure if I've seen it somewhere or if I invented it myself, yeah. >> No, that's a such a good one. That's actually covered. Usually this analogy is that I I at least when I try to make when I forget about MCPs to mention but this oven stuff.

That's a good one. >> Yeah. >> Now, we're going to attribute it to you until we find somebody earlier. [laughter] Anyway, uh so, would you like now to just dive a bit in the cloud code to show how to uh how to kind of go deeper? I can't remember.

I think we wanted to do a quick demo as well, right? >> Quick demo. We can we can we can we can share my screen and and share something. Let me just see what I have here. I am on my personal laptop, so >> Yeah.

>> performance is not great. >> Yeah, we'll account for that. Cuz right now I think our conversation is a bit for people who already aware about a lot of AI stuff. We didn't really exactly went for the beginners topic. So, maybe once they see how to actually kind of how this cloud code works in general.

Something that I think I'm going to cover in the channel anyway soon. >> I will think. Let me just think a bit. What do I have here? >> [clears throat] >> Also, for people watching us, just a question so far, are you following along the conversation?

Is it like do you find it actually useful? Is it too complicated or is it too easy for you? Just let me know in the chat. Meanwhile, Alex already has it. >> Yes.

I'm I'm not sure what what to start with. Uh maybe a bit how I like to how I build the tool, right? So, let's I'm just going to pause my doodles here. So, there's there's building stuff and there's [snorts] Oh, sorry. There is building stuff and there's using the model to help you work, right?

So, for building stuff, my current workflow is I just go on ChatGPT and I say research uh this or describe how this tool works, right? And I then I take take that description and I together with ChatGPT I convert it into a specification of how the thing should look and a minimal viable product of what's the minimum amount of features I want from from from the tool, right? And then I just let the model go at it. Build some sort of I make sure the model can self-verify and probably the easiest way to do that is to get her I mean you can go private MCP, right? You can go and call the tools and if your Revit crash you can enable computer use.

It can go and and and uh restart Revit for you. It is very slow though. And I'm also always using uh the latest model available because of the hundred dollar subscription. You you can afford that at the scale of the tools we are building. And then the model goes and and says, "Okay, this is what I built.

I'm going to test it. Did it met all the requires from the MVP? Does it work? No? Okay, try again until it considers to be done." That's I don't you could say that's the most basic uh loop you could have, right?

>> Mhm. So you start with an idea, you turn it into spec of everything. You also build MC kind of MVP kind of plan and then AI. >> Yes. And it's good to ask for the model ask the model for a a spec and then you have to go and modify it yourself.

I never experienced that the model it always tends to overshoot. And um a good a good skill for this would be again car party is called if you go if on Google if you do car party uh cloud MD you would get maybe I have it here. >> Can you explain what do you refer now? What's the cloud MD? >> Yes.

>> close that? I got it. Got it. Got it. But show it for others anyway.

It's a good one. >> Yes. Uh do I have it here? Sorry, give me 1 second. >> I usually always type online Karpathy Claude MD.

It shows up first. >> Yes. I but I transform it in a skill. It's just that we are recording and my computer is very slow. >> If you want, I can show people the main gist of it from my mind.

>> Please. >> I'm wait I'm going to just open. So Uh Just 1 second. Yeah, I'm trying to pin myself. Oh, this app.

Anyway. So I think people can see right now. So there is Yeah, that's the one. You talk about that one, right? >> Yes.

Yes. And I convert that into Can you zoom out a bit? Sorry, I only can see >> Uh there is if I'm going to go to Claude. No, wait. Claude MD.

Here we go. >> Yes, exactly. That one. Yes. And >> I just want to show people to find it just type Andre Karpathy skills.

It's not made by Andre Karpathy, but it's based on his teachings. Uh what you're going to find here, it's a repo with 200,000 kind of stars. So you can see it's really used. And right here, it's very short. It's shorter than I think before code simply simplicity first, surgical changes, gold driven execution, and a little bit more explanation.

So that's the base. And I also highly recommend just to copy paste and implement it somehow in your skills in your projects when you start. If you especially if you use nothing, just use that at least. >> [snorts] >> Yes. I think that that helps and obviously it helps to to modify the bit for your use case.

And that that gives this workflow. It it it stops a model from over shooting, right? So, a few examples of how we do that is um If I go to tricks and tools, we have our >> You need to share your screen again. >> Oops. Or I'm sharing now.

I'm sharing now. >> Okay, now here we go. >> This is not good, Erik. It takes so long to open a printing tool. All right, so this is is it's a simple it's a simple printing tool that we did by describing the roots pro sheets amazing plugin.

By the way, we use it for a very long time. Um It just it had too many features for us, so we just took exactly what we need and we recreated this this tool that kind of mimics the way pro sheets exports IFCs, DWGs, and PDFs where you can compose the name together and all and and that. So, that's that's one example and then again, this was a few prompts and it built this tool for us. So, yeah, that's one example. The next example would be uh something maybe we can try it through Should we try it now or what what What do you think is better?

It's better to show an example where we drive Revit or where we just build the >> I can do it by side by side. There will be kind of waiting times anyway. >> Yeah, I will I have might have a descriptions here. >> Yeah, also sorry everybody. We have a few technical issues here and there because I decided to figure out on the go with Riverside and apparently it's more limited than I thought.

But I hope it's still useful to you. And so far I think Alex sharing quite a lot of good insights here and there. But I also understand that maybe we skipped a few basics here and there. Okay, so we see your VP of UI Revit. >> Yes, so anyone can try this, right?

You open Cloud Code inside a fresh toolbar, take a screenshot of Autodesk family editor, pointed where it is and say, "Hey, can you describe this tool in Revit?" See print screen and it goes and gives you description of the tool. Okay, and then let's Sorry. >> So kind of reverse engineer the tool first. >> So you still use ChatGPT but you code with Cloud Code. >> Uh I I have web search disabled on Cloud Code.

So I do the research with ChatGPT and then try to copy the answer and paste it into Cloud Code. It's just it's a precaution. You don't want to combine access to your machine with a web search. It's there's uh potential for prompt injection or >> Mhm. Mhm.

>> Yeah. So, uh then we're going to get the spec. I mean, it's going to take some time, right? So, I will I will let you know when it's done. I'm going to take the spec.

I'm going to open uh our I'm going to open the pipe the uh Pyre V toolbar, and I'm just going to try to just dump the spec in there and ask it for for a toolbar and then see what happens. Basically, that is the workflow. >> Okay. Uh recently about all these kind of spec building and stuff uh maybe I just need to change the screen or something. Can I I Yeah, thank you.

That's going to help. Now, I need to So, recently there was a cursor they shared. Maybe I need to find it. They shared their report where they they were experimenting with different models and stuff like that. And one of the things they did is they tried to rebuild SQL, I believe.

And the way they did it is just use different models. Uh wait, it's jumping. So, uh they tried Fable 5C. It took like 8,000 something like that to rebuild it. Then they tried, "Okay, use Fable for planning or building specs like you say and then use Opus." It took like this much, this much.

And then like the conclusion was like in the end they got similar results using different models. But what worked the best it looked like using very small model like Fable for the planning. But then using like much cheaper models for actually execution. Cuz once the spec is built and it's good, like like you don't need as much intelligence because like the biggest thinking already done. And then the other models can pick up and like, "Oh, how do I do this, this, this?" And they kind of do it pretty well.

However, if you are like on these cloud subscriptions and you are cannot even max out your tokens, often times you can and me I oftentimes just run with Fable. It auto kind of goes down to Opus anyway sometimes depending on the task. But that's what I noticed. >> Yeah, using Fable 5 with Opus sub-agents seems to work okay. I'm trying also you can you can choose inference level, how much the model thinks.

I feel like medium is good for everything. I don't think I do complex things where I need more than medium on both. >> Okay. >> Yes, it's a balance about, you know, overshooting and then it's too much thinking and how complex is your task. If you add some examples, most of you you'll be able to do most of thing all of the things.

>> Mhm. Yeah, that's the thing like I often just run on like extra high or like even max when I really want to get the most out of it. Mainly because the tokens is not an issue oftentimes. Cuz the way I do it somehow I'm still surprised like that it does it maxes out so rarely. But going to lower ones I always think about like okay, what if this time it will kind of make a mistake and it takes me longer time going back and forth explaining.

Maybe I just give it more juice, it produces something and then I work on that. >> It sometimes produces worse in my in my experience with with high inference. >> Oh, yeah. >> High, extra high, yes. More like ultra and it's it's it's just like it's too much.

>> The thing is that ultra it's actually running on extra high. It just it just unlocks you very sub-agent kind of very heavy sub-agent works. But the max is the one that kind of doubles like a token usage or something like that. And also slower. Yeah, the thing is that I see a lot of people talk about this.

There's this fine um how do I say it? Fine line that you have to decide which model to use when. There are even whole like harnesses that will decide for you. So, you just you just prompt and sometimes it uses this model, sometimes that model. >> But again, that is a bit of a temporary thing.

I think the like we're going to get ChatGPT-6 very soon. Then Gemini 3.5 Pro will come out. It's going to get to the point where models like after when Fable ChatGPT-6 and and Gemini 3.5 Pro becomes the new medium [snorts] model and then tokens are very optimized. Um I don't think it would matter that much what model you choose. I think that again, this is one of these temporary things.

>> Mhm. Mhm. Yeah, it's a tough question. I see your point. I also have been down this rabbit hole.

I was like, how does it I think they're all going to get to the certain intelligence level where it doesn't Honestly, we're already at the point where it doesn't matter. If Fable disappears tomorrow, honestly, I wouldn't care. Like there's Opus 5. Even if like instead of Opus 5, I had Opus 4.6, I would be happy. It can do a lot.

>> Yeah. Yeah, you need to do the the stupid less important the less capable the model, the more the skills and all that matters. The higher the model, you get to a point where all of all of it gets absorbed. So, what is left is basically what you want to build and why are you building it? >> Mhm.

I think there's also like skills, memory, uh context management, all of Not context management, but providing the real context of your projects, what you do this, what will matter the most. >> Yeah. Should we maybe let's try and uh >> Yeah. >> Let's inspect. >> Oops.

Uh I think we we need to share screen again. >> Yeah. One second. >> I noticed I think it will auto do it. >> Is it Is it there showing?

Yeah. >> Yeah, yeah. >> So, you can see it. >> So, this is the spec. >> So.

Purpose. Search category for the thumbnail previous grid and list views, multi-selection, batch loading. That's good. Button definition. >> That's looks okay.

>> Folder structure. >> Mhm. >> Uh >> But anyway, when you look at this folder structure, you still can understand what's going on like here and there. Like you know what I mean? >> There There's Yeah, I mean all this is what you get in in a in a button, right?

>> Yeah. >> Then and then you have the library and uh the resource folder that's is going to be in the in the toolbar root. >> Mhm. >> [clears throat] >> Correct me if I'm wrong, please. >> No, no, that's fine.

>> Yes. Yeah. Uh >> I'm still I'm still trying to understand because you do know some Python, some Revit API, but you mostly kind of write code. So, I'm trying to find like cuz here like okay, that that should make sense because probably you've seen it like 100 times already as well. But probably also recent kind of looked into this.

>> If If you start from scratch, maybe it's not immediately obvious, but I mean, you still need to know the structure of a pyRevit toolbar. There's a library folder where you pull uh repeatable code from. Push button is the folder name of your actual button, then you have extension, uh panels, and all I mean, you still need to know the structure of the toolbar. But again, it's knowing the structure of the toolbar is not coding. It's >> Yeah, yeah, but there's like already separation of different tasks.

That's That's what cuz you can see that there are like five files in library, so it already starts kind of splitting stuff here and there. So, >> I just wonder if you really try to control it or you just let it go and like well, probably it knows better. >> I think first pass, I don't assume it knows better. I think first pass, I literally go and see what happens. >> Okay, so.

>> And then I adjust, right? So, then we have window layout. Okay, we're happy with this. Okay, let's just take let's just take the output. I'm not going to at least view name footer.

Okay, we can take this. And this is It's a bit slow. I apologize to everyone. It's a We can open cloud in this folder. >> That was all.

>> So, curious if you use cloud in the terminal or like chat. >> The reason why I use it So, I I don't I I use mainly I mean, I have VS Code, I have Cursor. I mainly use terminal and Sublime Text. Just because it's it's fast and I only read markdown files. >> Mhm.

>> So, and I like I like using cloud in terminal the same I'm I don't run any any partial commands. I It's It's literally a chat interface. It's just It's way faster. You can customize it. You can see the context toolbar.

You can set the the permission mode. You can quickly disable and uh enable plugins. You have all this flexibility that actually I don't even know where to get it in the in the app. Mhm. So, uh yeah.

I'm not sure what's the difference between the terminal, the vanilla terminal, and the plugin in the IDE. Is there Is there a difference? >> Yes. So, well, first of all, uh in IDE, you're getting kind of terminal version, so you have to install it like a terminal, but in VS Code, for example, I know there are two options. You have it as a chat option, kind of integrated, but also you have like uh terminal option right inside your IDE.

Cuz you you can open terminals, but they also made it part of the plugin. I personally prefer terminal as well. It has a a lot more commands. It has a lot of shortcuts. You can even open For example, we can write prompt now in the uh what is it?

Notepad. We can also extract everything in Notepad with a few clicks. There's like a lot of hidden features. Can you put a question mark in your prompt? What You can also kind of ask it like what it can do.

You put the question mark, you can see there is a lot of information like, "Oh, if you put this, you can execute this this this." You don't need all of it at this in the beginning and probably most of it. But over time, there are a few shortcuts here and there that you can find, which are not available in the chat version. So, for me, I kind of went into that. Yeah, so I was just glad to see that you're using terminal. >> It's Yeah, it's the same reason why in Rhino you would use the the the command bar instead of clicking the icons.

I would love if Revit would have a command bar as well. >> Mhm. >> Okay. So, let's let's just try Let's just try and and just type, "Okay. Uh here is spec." slash Sorry.

Superpowers is a good plugin that comes from on topic partners, I believe. And it's just it helps you convert a plan a spec into a implementation plan. And starts it tries to to to overdo it again and it's an extra step that maybe is not necessary if you have good spec. It will probably take me half an hour to clean a spec so I we're not going to do it live. But we let's just try and do superpowers and say okay here and I normally use wisp wisp wisp flow.

I don't have it enabled on this laptop. >> You can use command voice command in terminal version >> Voice yes, I know it's >> a space >> I'll I'll I'll I'll write this for now. I'll write it for now. Okay. So here is spec where are you?

I'm just going to put this folder. I'm going to give you the path to the folder. >> Mhm. Okay, so you still directed to your skills and stuff. >> Yeah.

Yes. And so you can you can just have bunch of skills chained together. And then I'm just going to copy the spec. I would also normally do ask user use the ask user questions tools for any clarifying questions. That works quite well.

The more time you spend there the like fewer task is. So this is going to run in the background. Okay. >> Okay, I see. >> Maybe meanwhile we can talk a bit about this.

So this is can you see my screen? >> Yeah, yeah. You have Revit MCP. >> Yes, so this is a Revit MCP that I wrote it myself. It's probably not the most efficient.

I'll just maybe show the Oops. >> Is it still by rate extension or is it C# kind of a little bit? >> sharp it's a it's a C# uh >> [snorts] [clears throat] >> Is this cleaner right? Is this a visible? >> Mhm.

The text is a bit hard to read, but if you zoom it's going to be fine. >> Yeah. So it's >> Perfect. >> Yeah, it's a Python MCP server. It's pinned at very like a very early version, so it has almost no dependencies.

And it's it's very simple, right? It has um run Python command and a run sharp sharp command. That's pretty much the the two tools I use the most. >> [snorts] >> And then it has a uh a bunch of built-in verifiers that can can make pre- print screen. It can query elements and stuff like that, right?

Then it has It's basically saves on the drive a tool if you if you ask it to and then you can load it dynamically. >> Mhm. >> So it can list save tools and run save tools. It just points to a folder, right? And then it goes to Yeah.

>> a quick question. By the tool, does it refer to pyRevit tool or MCP tool? >> It's a Python tool that uses the Revit API. >> Okay. >> Or it's it's it's basically it's using the it's using the the run Python tool to run that file from your drive.

>> Mhm. Got it. >> Yeah. >> So I just want to get >> has Yeah, that file has the the tool the the tool script and the description. >> Mhm.

>> Yes. And then it goes Yeah, it goes to the Revit and then the whole idea with this is not to make the most efficient thing. The whole idea is to Okay, we're going to wait for Autodesk to release probably a very stable MCP, right? If I give you the model that doesn't do mistakes and if I give you a connection to the MCP, what are you going to do with it? And it's not immediately obvious what you want to do with such a tool, right?

And that's why Yeah, that's why I built this like this small prototype for myself where I start to talk to the model and see, okay, what can you do? How How can I push this? And uh if I'm going to show >> [sighs] >> I don't think this is going to be possible, but let's try. >> Well, that might be a bit too much for the laptop. >> [laughter] >> This is a project I worked in UK.

>> Well, looking nice. >> Yeah, and There There's a period at the beginning of the year where we had to we we went from stage two to stage three and we had to quickly convert it from um from beam beam geometry to non-Revit geometry, we had to quickly convert it into native Revit geometry to be able to not bottle neck on any of the consultants and have like a native Revit model Revit to ship, right? And we are a bit squeezed on time, so I had like two days to to to uh build this model, basically. And it's This is the first I would say vibe co- co- It's like it's It's not like fully It's I use a lot of AI to help me do this. It's Again, I haven't used it to design.

The The design was already there. I just had to recreate it. But I had to do a lot of >> [snorts] >> I I it would be different than I would normally use it for. So, we I started by building uh, this family. >> Mhm.

>> I started building this family that has all the >> This is like a base or >> Yeah. It has It has all all the facade set out. And the all the facade lines baked in, right? And then I have this drafting view where from previous stage I extracted all the all the cladding profiles that I need. >> Mhm.

>> And then I to get using my MCP I I start asking the model, "Okay, here's this line. Divide the line by this grid. Take this profile, sweep it. Convert this to families. Okay, then once you have the families start rationalizing the family." And it goes and then okay, now that you rationalize the family you find similar family parameters and replace them and start like with this with with with the base of lines and a base of profiles I managed to to recreate most uh, most of the tower, right?

And then I started thinking, "Okay, this is actually quite interesting. How can I like push this further?" Uh, and I start creating a set of LOD 250 families. >> Mhm. And it's mostly like AI driven. So, you create the profiles, lines, and it kind of >> Yeah.

Profile lines and then okay, I need a bank of families. How how is the facade built? In in 60% of the cases you have a unitized facade. A unitized facade comes with a window, a cassette, and then you have cladding on slapped to it, right? So, what if I make this ready-made library of window cassettes and I they're very low LOD, so it's it's fast.

And I baked in some instructions. I I put a parameter and I baked instructions inside the the family is similar how you would use MCP tools. Describing, okay, this is the family, this is how it's driven, and this is what each parameter does. So, combining this kind of low-level LOD library with with the sweeps for the cladding, you can pretty much you can go quite far in recreating a tower like this. And then later on, you can pair the the external family with a unitized final and nested inside.

It works much better than expecting the model to do everything from scratch, right? And then you start with a very low LOD, and then you just swap the families. So, the next step would be to take a low-level LOD family, go inside the family, and talk with the model so you can Okay, now we need to add profiles. You need to sweep the profiles. This is the profile.

Look at the You see some drawings from consultants and and and it seems to work quite well. It does require Opus 5 or Fable to for it for it to be in reasonable time. >> Mhm. >> Yeah, but it does work for sure. Yeah.

>> Does Does it also save some kind of tools or it's more like one-off things like >> that's that's the that's the thing. So, then I Once you're happy with the thing, you can you can go and it saves the tools. And it's it it saves the tool files and I also have a hook that saves the description of the file. Right? So, that starts giving me a catalog of >> [snorts] >> what what what might I need and how how how might I need to use it, right?

Because it's not super intuitive. You have to start building your families a certain way. You have to Yeah, put this description. You have to use a lot more parameters, for example, right? You don't want the model to be stuck in making times.

Maybe you make a lot of instance parameter. It's all driven by instance parameter and the There's a type parameter that is a concatenation of some of of this instance parameter that builds the type name rather than than than the other way around cuz then you can you can classify your families correctly. Does it make sense what I'm saying? Uh >> Yeah, yeah, yeah. There's just quite so many things going on here in terms of like the whole workflow.

It's like certainly so many steps. We're kind of following I hope people watching also can follow along. >> Yeah, so Yeah, to to to simplify it, right? You can use the you can use the pyRevit connector, right? You The pyRevit connector ships with a run Python code inside Revit, which lets the model write its own scripts dynamically.

You can use that function to drive profiles on rails or place families that you built already. So, that's that's kind of kind of where I am. Once you do that a few times with a model, you can ask the model to save notes on how it did it or build tools for itself. For the next time you're going to I it, that workflow becomes more deterministic and the model can just call tools that have variables instead of going through a long loop again. That's Is is is Yeah.

>> Yeah, yeah. I mean like more like also the Revit side, the families, the parameter setup. >> Yes. >> Coding part, it kind of makes sense that pretty much you're just using AI so it kind of figure out its own way and then you just also tell it like, "Okay, don't forget to save it so we can reuse it." >> Exactly. And then if that tool becomes very valuable and that tool can easily can point to to the folder where you save the tools and you just promote them to a private toolbar and that is how we use private now the most is to to to have this this this scaled uh to to to to scale some of the tools, yes.

>> Mhm. Okay, so you're pretty much just going to iterate, explore what's possible and when cool things come out, you just like, "Okay, now it's a tool." >> Exactly, yeah. The thing is that Fable now is so good or obviously a lot of people don't have access to the Fable. Now it's so good that you don't even need to spend time doing that. It's Yeah, it saves tokens.

Obviously, that's good, but how much effort do you put into managing those tools, deploying those tools? You you you need to kind of calculate a bit that, yeah. >> Yeah, Fable has one downside that recently they a little bit downgraded it that when it thinks like maybe it's an upgrade because it often uses Opus 4.8 instead of Fable, especially on the coding tasks because well, they were after Fable got banned, they had to add restrictions and something like that. But it's still a solid model. But now I don't know like there's Opus 5 on benchmarks, same as Fable so they they don't know themself how to promote it.

It's a bit weird. >> Yeah. Do we should should we do Should I do a small scale demo on how that works or >> Yeah, sure. >> Is that is that >> Uh I think we going to end up in 20 minutes. If we can fit in like in 15 minutes, sure.

Because we also creating another tool. >> Yeah, maybe yes. It's it's it's it's going. Uh I'll try I have I really apologize. This my personal laptop is so slow.

I have a folder saved with some video. Maybe I can just pop that quickly. So if if you look at our stack, right? We have uh pyRevit for scaling tools. We have the MCP for letting the model to promote tools uh to to to run code and check itself.

We also have uh we also use a Rhino MCP that um has access to Grasshopper and we do a lot of things using the Revit API on the Grasshopper. I'll I'm a visual person. I like if I do a complex tool, even if it's in pyRevit, I like to first visualize it in in Grasshopper as chunks of custom uh Python nodes placed by the model that then that then the model can read and can promote to a uh pyRevit tool. And uh >> Uh so wait a second on that one. So you pretty much break down the tool in steps visually.

>> Yes. Yes. >> Okay, that's that's a good one. >> Using Grasshopper but using only the CPython node in Grasshopper. >> Mhm.

So you have a full diagram in there. >> Exactly. If you have Rhino inside installed, it can talk to your Revit. It can write Revit API code. You do the logic.

You test it. Also, it's uh we use it as a 3D engine. If you need to to to do some geometry, it's better you send it into Rhino inside. Also, uh right you you have access to Rhino inside engine through pyRevit as well. So, we do have a tool here.

Um Yeah, there's the Rhino exporter. >> Mhm. >> That pushes everything to to Revit uh to Rhino quite quite fast. So, you can choose link models, you can choose uh active view. Uh you can just send it as separate files, you can send them as as um as individual files.

So, we try to look at the whole stack. I try to look at the whole stack rather than pyRevit or using uh cloud code to for pyRevit. It's kind of becomes this homogeneous thing >> Mhm. >> where you just buy you see where it goes. Like, okay, I'm I'm not sure how this facade how to solve that facade.

There, send it to Rhino. I go in Rhino, I model it while you go and sweep the rails and stuff like that, right? >> Mhm. Mhm. >> Yeah.

>> So, through wipe coding. >> Yes. So, uh yeah. >> it's really good context that everything that it can travel around, find things that it needs. >> Yeah.

>> Okay, there's some video. >> Uh and so with this example, right, where we have some people who maybe are not super experienced in Revit, but they want to be able to set up views. So, this is using Opus 4.5 where we just have a mirror of our model in the Rhino environment, and then this is the prompt. >> All right, can you comment on that a bit? What's going on?

There's a bit too many things. >> Yeah, problem is I see six frames in my screen. I'll try uh Do you see Do you don't see the prompt? Okay. So, the prompt was I I mirrored the two files, right?

So, the coordinate system is the same, right? And we want to be able to create our Revit views from the Rhino environment. So, my prompt was, "Okay, build a Rhino inside CPython script that grabs all surfaces from a specific layer. If the surface is oriented upwards, that's an RCP. If the surface is oriented downwards, that's a plan.

If the surface is oriented sideways, that's a section." >> Mhm. >> You have your sections here, and the model I think it took maybe 3 minutes with Opus Opus 4.5 to build a tool that sets up the views from you and place them on a sheet, right? So, this is this is the corner of the building where this tool worked. >> Mhm. >> So, yeah, because Revit has maybe a bottleneck, it doesn't have a scalable API, so it's a bit more complicated to kind of manipulate views like that.

So, maybe you can just do it on the Rhino side where it's it's flows easier, and then you just push it back to Revit. >> Mhm. Yeah, Revit in general not AI-friendly at all, but it was never meant to be. >> Yeah. And but I mean, the main use of Revit is as a document manager.

It's and I I would love if Autodesk would focus on that, just making documentation process really good, being able to like have granular control on the extent of sections and all that, placing the views precisely. >> Mhm. >> View title alignments. >> Yeah. >> They should really push hard on that.

Yes. >> Yeah, there are so many it's The thing is that there we all need just little things from Autodesk, but they focus always on the next big vision, like, "Oh, generative design." There's like AI assistant. They just try to make the headlines when everybody just needs like this little fixed like 50 just take 50 most common bugs we are asking for years to fix and fix it, and we're going to be happy. >> Yes. >> But I I see this over and over.

That's why pretty much most of probably beam managers started they just try to fix one little thing there, then they started like going really in depth, creating their own custom workflows, adding more software, and eventually becoming like I don't know discovering everything like because of just trying to make Revit work. >> Yeah. Do What do you think? Do you think Autodesk will train their own very efficient LLM to run Revit, or they'll skip that step? >> Honestly, I hope not.

It's not going to work. >> [laughter] >> The thing is that um Well, first of all, they made already the same CP for their AI assistant and stuff like that, but it's not backward compatible. And it doesn't make any sense because well, most people not going to use 2027 for another like year or something like a two. And we already have pyRevit which is open source, we can look inside. We It works across all versions, which is crazy, but also we can modify it.

And I think they focus on AI wrong. And if they train their model, they it won't be fable because it's expensive as hell to run. So, it would be something super cheap. And if it's super cheap, it probably won't be as intelligent as we would want to. So, we would want to bring our agent anyway.

So, uh they probably could make some kind of skills, but when I had conversation with uh Jonathan Jacobson, uh he kind of suggested good idea that they just should focus on infrastructure, not the product. Cuz Revit is great because of all the third-party add-ins cuz they just exposed Revit API, and now we can do what what we need cuz Autodesk cannot cover everybody needs, and you know like every single office has their own way of doing things. Even like we cannot even agree how to name things in project browser. Everybody going to have their own way. And if they literally just give us super great MCP, something like maybe you created something just this kind of experience, but they're not going to do it because there's a lot of issue that come out of it, right?

The security, like what if like intern like accidentally gets access and like goes wild with that and I don't know what they're going to do. They're in tricky situation. They try to catch this kind of shiny object, which is probably not theirs to catch. >> Yeah. Also, there's a lot of focus on destructive tools >> [snorts] >> and in the context of modeling, right?

In Revit or in Rhino. Yeah, you have to pay attention. But if you by mistake delete something and your project is on ACC, you have so many backups. And those rail rail guards to prevent destructive tools slow down the process so much. But there should be an option where you remove the destructive tool rail guards and just let people >> Mhm.

>> Yeah. >> But I see from their side, they're like huge corporation and they don't want to risk anything. >> Also, it's crazy too. It makes the tool useless. Like I I'm trying mirror MCP and it's just it's quite limited compared to what you could just one shot yourself.

>> Yeah. >> But I think that's why I'm so happy to see like Revit MCP. There are so many MCPs there is even if you're going to go for the this by harness, somebody already made pyRevit. Naming is weird because it's just one letter, but it's already has MCP for Revit. Kind of ready, so it's just going to one click add-on.

>> Yeah, I hope out of this focus on just make it as compatible as possible with with an AI and then focus on their what they're what they're good at. They're very good at I mean I I believe personally, I believe that drawings might disappear at some point as soon as some some municipality decides that it's time to just send over 3D model or something. It's not going to happen soon, but it will happen. But I think until then I architects should just focus on making that process as easy as as possible. It should be one click.

The whole document set should be one click in Revit. There's no excuse why it's not after 25 years. >> Yeah, I don't know. It's When talking about like Revit functionality and where it's going, it's always so hard cuz you see what happened in the past years where all requests user have, they kind of get ignored or and it's not developers' fault. It's probably just the whole structure of the company cuz their sole goal is just to make shareholders happy.

Since that therefore they focus mostly on headlines, I think. I don't know. Sometimes I have a feeling they're trying to kill Revit and probably push something like Forma forward, but they are like there's a huge gap before that can happen. So >> Also Also that API isn't isn't great. >> I never looked in Forma.

Is it? >> Yeah, I I I It's like we we together with my colleague Sam Sweeney. Um he's he's leading the beam efforts and then we're trying to look into APIs at the moment. Yeah, we kind of want We just want to put more time into design and I feel like everything is is like to share your Just to share your model, you have to save, you have to synchronize, you have to publish, you have to package, and you have to transmit. I mean, that's ridiculous.

Just to share a model. >> Yeah, I don't know. Honestly, it's This topic that there's like this a few years where it's not clear what's going to happen to Revit. Obviously, it's not going to disappear. There's like even if there's a better tool.

If tomorrow there's a better Revit killer comes out, it's not going going a big market share because everyone invested so much into Revit templates, workflows, power Revit tools and stuff like that at most offices like well, they just cannot justify the jump. Even if it's better, there's like a lot of training. Again, nobody is going to learn the new software in a week. So, I see that we're not moving away from Revit anytime soon, but at the same time, I don't see Revit moving as fast as we need. So, we just >> Yeah, I don't think we're going to move from Autodesk, right?

They have the user base and that's the I think the hardest to get everyone on board on something, right? That's why also MCP is great. It's not the best connection. I mean, I see a lot of way better uh for way more efficient, but it's just everyone agrees that okay, this is the standard that we're going to use now. And that's probably the the hardest thing to do and I feel like now with like a lot of consultants are in ACC and it's getting easier to to collaborate and that is the leverage that Autodesk has at the moment over everyone else.

>> Mhm. Yeah, it's always about the users, especially now like I see that there are a bunch of apps coming out. There's like literally I saw a few like in but the thing is that in AEC Revit does so much. Nobody can just make a quick project. Everybody focuses on one topic, mostly on these early stages of the process cuz the later one probably requires you to start with early ones in the same software.

I don't know, it's a tricky one. You probably have way better awareness about where it's going in terms of I'm not using Revit as much for the projects. I'm just kind of teaching people how to automate and I go really into kind of API, how how it works, trying to kind of learn more about it, but overall, I kind of see that it's moving, but I don't know where. >> Yeah. There's no one trying to solve documentation at the moment.

I've seen all these new companies and I've seen that new BIM tools, no one, is everyone's focusing on 3D, which I don't know. I think we we already have good 3D engines. We don't need 3D engines, but it's a document managers and yeah, that's and auto tagging and auto detailing from 3D and all that. That's that's the gap I feel like. >> I mean >> Yeah.

>> For a long time I was thinking that Grasshopper Blender is just going to surpass. They're going to add some kind of beam more There are There are like Bonsai it was renamed as Beam Blender or Bonsai. I think Grasshopper some had something as well, but it's not that common to use. I haven't really seen people really jumping on board in like in masses. They usually try it out, make one two projects, but then they I think they fall to Revit anyway.

So, I honestly don't know cuz our industry is one of the slowest to move. Now it's I think catching up, but before that it was like so behind. Cuz everyone was just dependent on native tools, plugins, like nobody looked into automation as much a few years like like 10 years back. And there was like huge gap just kind of make it more efficient. But now I don't know.

It's a tricky one. >> It's a tricky one. >> I don't think that I'm the right person to kind of really answer this. >> I'm I'm trying to see how we are with our tool. >> Oh, yes.

Where are you? >> No, no, no, I'm not sure yet. It's just a >> Yeah, so the laptop is the bottleneck, right? >> I need to I need to redeem myself and follow up with the with the with the tutorial or some demos. >> [snorts] >> Yeah, well, I'm sorry everybody we didn't really plan it.

Probably it's mostly on me. So, don't don't blame the laptop as much. But yeah, Alex offered me to kind of do the demo before. I was like, "Ah, we're going to be fine." So, yeah, I think we're just going to wrap up here because of the time. Unfortunately, well, I'm going to also kind of dive in Cloud Code, but if you can also later kind of either share the demo or something, I could also maybe kind of point the people to that.

>> Yes. >> Uh but now, I I always like to usually end my podcast by asking people like, what is the one like this advice? It can be multiple, but what's the main advice for beginners? And before that, it was about coding. But in your case, like and since this whole AI era coming, what if people who, let's say they are just use still ChatGPT and Claude in the browser, what should they do?

What would you recommend to focus on? >> I think probably the most thing is most important thing is to just constantly be curious and try to grow and the only way to satisfy your curiosity is to go and do the thing and see what happens. So, I think it should be a lot of that. And we know maybe don't need structure. You don't need You just just do it, see what happens, and then that will inform you what is the correct way moving forward, what should you focus on, rather than trying to get all that beforehand and then and then and then be stuck in this preparation stage.

>> Mhm. >> Install All right. Install Cloud Code, try to just one-shot tools. You're going to get something. You're going to get excited, amazed.

You're going to think you can do anything, then you're going to try to do something more complex. You're not going to be able to do it, and then you're going to go to your tutorials, and then you're going to and it's going to be a bit of that. So, that's pretty much the process that everyone, but it all starts with curiosity, right? If you if you If you're not curious, nothing will nothing nothing will happen. >> No, that's a good one.

Also, think Also, they need to stop kind of being afraid of trying it. Cuz many people don't try because they don't understand it. >> Yeah. >> And have you played played a lot of games when you've been younger? >> Yes.

Too much. >> I think it comes from that because I also was a lot in games and I have this mentality like give me any software, I'm not afraid to click any button. If it breaks, if it crashes, I kind of understand like well, there's save, there's cancel, and I click around. But I know that many people they they need to be they need to know what the button does before they click. >> Yeah.

>> Where I think it's coming from games, we just click the button and see like oops, crashed. Well, now now I know. So, I think this is also there like just kind of sandbox your Revit. Don't start with live project, please. Just >> Yeah, yeah.

Don't don't don't do it on enterprise network, for sure. >> Yeah, just understand kind of how it goes. But yeah, I think it's a good one. Just be curious, try it out. Also, would you like to share any what what you're up to, where people can find you, and stuff like that?

>> I'm I'm I have in my drafts on LinkedIn I have four articles. So, it's pretty much only there. I'm trying I'm not very active. I answer to messages. I I I try to stay away and try to curate my feed to people I know.

But I'm planning to to just release a bunch of articles just taking this sober pragmatic view of how we use the tools, the tricks, and and it's mainly going to be focused around the Rhino inside and and Revit's to begin with. So, just give me a follow on on on LinkedIn and then Yeah, that's that's it. >> Okay, thank you. Uh I'm going to leave the links in the description where you can find Alex, what he has what he's up to, learn more about what he does. Cuz he has a few articles on LinkedIn already and he mentioned he's going to write more, so might be interesting.

Thank you, everybody, for watching. This was our first test drive of going live in Pirate Podcast and we started with massive technical issues, so I'm really sorry for that. And yeah, I'm going to prepare much better for the future. Going to see where it goes. But thank you once again.

It was a pleasure, especially talking to you, Alex. Thank you for coming on the show. And Wait, is he still here? Going to complete. Okay.

Goodbye, everybody. Alex, you stay.