Archie: an engineering AGI for Dyson Spheres | P-1 AI | $23 million seed round

Hey guys, in this video I want to tell you more about what I've been up to since late last summer. Uh I co-ounded an AI startup called P1 AI together with Paul and Adam. Uh Paul on the left hand side is a former chief technology officer of Airbus and United Technologies. Adam on the right hand side is a former engineering director of Airbus and together we teamed up. We raised $23 million seed round to help us build an engineering AGI that will help us design physical systems. And so briefly, if you haven't seen the launch, as I mentioned, we raised 23 million from Radical Ventures uh led our round and uh we have an amazing lineup of angels uh that supported us. People like Jeff Dean who is a legendary Google engineer and people like Peter Valinder who is a VP of product at OpenAI, people like Bob who whom we might know as the CEO of VV8 and there's also Louis Vet and many of many other amazing uh angel investors we're building. We want to really move away from designing general intelligence just for the world of electron so to speak to have impact only in the software in the digital world. We want to move it over to the actual physical world such that it has impact on the world of atoms. And we want to use this engineering AGI to help us build systems as complex as Dyson spheres or airplanes all the way to relatively speaking simpler systems such as commercial atri systems. So, heating, ventilation, air conditioning systems for data centers, stuff like uh water towers, stuff, stuff like chillers. And so, in this particular video, I want to walk you through the current the P, the demo that we have for the residential ATrex system. Um, and I'll demonstrate to you how Archie, our agent, helps me design a new a new product.

Okay. So, let me open up my Slack channel here. And I'm going to first start by inviting Archie, which is, as I said, our system. So, I'm going to invite Archie to our channel. I'm going to hit enter here. And then I'm going to tag Archie and tell it, "Hey, wake up. Time to work. " And so Archie is now in the background going to be spinning up and loading. And while that's happening, I'm going to type in the first uh command, which is going to be um hey arch um load up the oops wj for design um and visualize it and compute its um performance. Okay. And so I'm going to wait for a second. And then you can see now it's ready. So I'm just going to hit enter. And then it's going to start streaming messages here in the debug mode in the Slack. I'm just going to move over to the actual web UI which is nicer. Uh so you can see on the left hand side the first thing Archie is doing is it's trying to it's a complex system. It's an agent that interacts with stuff like file system and bash and and like Python interpreter and also various custom tools that we have and also uh LLMs that we postrain internally and uh different types of surrogate models. And so you can see here it's first trying to basically find the WGA4 system which is a system modeling language document which is a centralized representation of our design. Uh and so it loaded it up as you can see here and then on the left hand side it started evaluating the performance. So that can be actually seen in one of the tabs here. You can see the performance metrics stuff like coefficient of performance and cooling power and the cost is currently hidden. Archie then proceeded to visualize the system in 2D and in 3D. So I can also show you a schematic. So you can see here this is a representation of this particular residential latex system that we're dealing with. You can see every component has uh certain uh characteristics. And then we also have stuff like 3D CAD. So you can see here a bounding box view of our system. And you can also see a more detailed view of our system uh with like the fan, the motor, the evaporator, condenser, the purple thing uh all in one place. So then if I open up the boom tab, you can see the actual components that our system contains currently. It's a it's an SCI compressor. And so you could open it up and see compressor we are currently using and you can find various metadata modelica and cml uh about this component. And then finally the summary on the left hand side uh about the performance of the system. Now let's

actually go to the interesting part. Let me try and modify the system. So let's say that regulations kicked in and for some reason the coefficient of performance really has to meet some threshold like maybe 2. 3. So I can maybe tell our system a query like the following. So can you increase the coop to 2. 5 while keeping the power constant? Okay. So I'm going to hit enter here. And so what's going to happen now in the background is the large language models that we pro that we post train. So all these are we are not using cloud or openi or xis gro or whatever. we have to use open weight models and we have our custom pipeline and we have a whole part of the company dealing with designing synthetic uh designs and that then the team that I the AIOR is basically using as the input to post-rain large range models and surrogate models which act as basically an efficient implementation of a simulator. Okay. And then you can see Arch is currently generating all of the visualizations from scratch. And let's see. So we finally got to coefficient of performance being 247 which is not quite there but very very accurate and then I think the cooling power was initially 1. 7 tons and so it did increase the power to by 0. 2 2 tons. Even though I told it to be constant, obviously system is still uh something we we're working on. This is more of a demo. Um and let's see. So let's see what Archie wrote here on the left hand side. So to achieve the target COP of 2. 5 while maintaining a similar cooling power, several significant changes were made. Uh it we had to upgrade the compressor from SCI to some different uh module. Uh and then uh it changed the blower motor, wheel, um it changed the evaporator coil configuration and various different things. And then it basically he basically concludes by saying COP increased from 1. 55 to 2. 37 which is very close to target 2. 5 uh and then summarizes the rest. So yes, so we basically achieved the first objective here. So what happened here in the background is that our system that we called u architect basically generated uh like a cisml document of this new design given the previousl uh representation of our design and basically modified and then came back and communicated that information back to the orchestrator. Let's keep on pushing uh the system further. Let's tell it the following. So we now have

cooling power being 1. 9 tons. Let's tell it um so can you increase me the cooling power uh to 2. 5 tons uh please keep the let's say length and width constant and set the height to 1200 mm. I'm going to click enter here. And not all of these queries are necessarily going to be super representative of what uh design engineers in these companies in these big H companies might care about. This is more of a demo just to demonstrate the capability of our current system. Okay. So let's see what is going on. So again in the background now um our models are modifying based on the new instructions in an iterative fashion the existing design and then re-evaluating them and revisualizing the whole new system. So let's see. So we I think now okay so we got to 2. 4 tons. I think we were 1. 9. So and that this is very close to 2. 5 tons that we asked for. And then let's see if the system dimensions. Okay. So here you can see that uh length and width didn't change significantly. And then the height is set to close to 1200 mm. So it obeys those geometrical constraints that I specified through my input prompt. Okay. So let's see for the final summary here. Okay. So we can see that the compressor changed to Copeland compressor and there has been many other changes and then you can see here that Archie basically is aware that the target was 2. 5 but it only appre approached 2. 45 45 tons. Uh so let's see let's do the following. Let's go to the BOM tab and open up. So you can see now it's a Copeland compressor here and I here have here the previous one which was an SCI. So if I open up the Copeland compressor, you can see it looks very much different. And so with this new compressor, the system characteristics are completely different as is to be expected. Okay, let's test one

more query. Let's say that you want to swap out a Copeland compressor for SCI compressor because for whatever reason, maybe your company all of a sudden cannot deal with with Copeland anymore. They're running out of business or whatever the situation might be. So, you might tell uh Archie, hey, go and swap me Copelan for SEI. Let's do that. So, replace the compressor with and then I have to actually see an actual component number. So I'm going to use KN104 WV VAMT from SCI. So let's just go and run this command. So we should be able to successfully replace the current Copeland compressor with this other compressor and let's see how Archie handles this use case. Okay. So Archie is working in the background as we speak. It's modifying the system. It's repeating the evaluation. Okay. So you saw the the system that the compressor got changed swapped for SCI. And so now if I open this one up, you can see it's a different type of compressor. So Archie did uh modify only that component. And um you can see that didn't really work well in terms of the performance characteristics of the system. The power decreased, the COP decreased. But here I just again gave you the fact that this Archery can do this is something I wanted to focus on. Okay, let's switch gears a bit

and let's load up this different um design that we have um uh saved on the file system. So I'm going to tell Ry, hey um go and load up WJ4 broken design and visualize it for me. Okay. So with this it's going to again uh find it on the file system, load it up um rerun call various tools in the background to help uh Archie visualize uh the system and then we'll be able to see uh ideally in the cat tab uh the new design in a few moments. Okay, I think that's it. Um yep, that's the new design. You can see here that this the yellow thing is piercing through the this gray thing which we call bassband. Um and if we open up a detailed view, it should be you should be able to see it here as well. So now one once you have these types of like geometric interferences in your system, you might want to reasonably fix it, right? You want to fix any such interferences and Archie can actually support such a use case. So I can simply tell Archie

something like, "Hey, please go and fix any interference in the design. Okay, I'm going to submit that query and let's see what happens in a few moments once this is run in the background. Okay, I think it's done. So basically what it has done is it f the changes show that there was a positioning issue with one of the components. The translation coordinates were adjusted from this number to this number and then so you basically modify X Y and Z for this particular component and then rerun various analysis there and then finally mentioned what uh it did. So it adjusted the position of the component. The main changes being translation of this particular component and then if we actually now go and take a look at the bassband you can see this thing is actually fixed. So everything looks good. Okay, that was just an example of how we can uh modify a residential HVAC unit. Um, do stuff like manipulate the cooling power, the coefficient of performance, uh, swap out some components for components from a different supplier. Um, fix some geometric interferences in the design and whatnot. Uh, in a way which is fairly autonomous. With this, I'll be hiring for research engineers for

multiple roles. We're looking for people who know how to post-train large language models at scale. Um, and also for people who have experience with reinforcement learning from execution feedback in particular or I think OpenAI is using verifiable rewards as the name for this type of u training techniques. We're also looking for people who understand vision language models and people who've been generating synthetic data at scale. So if you fit into any of these roles, please do reach out. you'll see the link and on here on the overlay on the screen you'll see the roles and please go and apply. I would love to chat with you. We're also hiring for back end and envelops roles. So if you like what we are building please join us. Uh we are building a worldclass team of people who know how to build both AGI and have deep expertise in building physical products and we want to ship this and sell this to big industrial OEMs and sell it into the labor budget. Meaning we'll be selling this as a digital workforce as opposed to as a tool. Right? So that means if we demonstrate on our internal evaluation data sets that we already have built together with domain experts, if we demonstrate that our system is better than their engineers are, then we could build this at some number some percentage of the um cost uh of the salary of those design engineers and basically help them build a digital workforce. Thanks for listening. Again, if you like this idea, please do apply. It's going to be down in the description.