WITH MATTHAEUS SCHWARZ-SCHILLING
When the Smartest Model Is the Least Private
Matthaeus Schwarz-Schilling builds living machinery from parts. He says the machines are the easy part.
Grants, papers and a spin-off's IP all have to stay confidential until they are published. The strongest models want the context, store it, and belong to somebody else; the model he can run privately on his own laptop is too slow to use. He asked whether the work could be broken into pieces small enough that no single provider ever sees the whole of it.
Listen onApple Podcasts Spotify
SEP19Matthaeus
Schwarz-Schilling
CHIEF SCIENCE OFFICER, UNIVERSITY SPIN-OFF
Capability × Confidentiality
WHO SEES YOUR WORK BEFORE YOU DO?He brought the answer in with the question.
Matthaeus started in physics at TU Munich, moved into biophysics, and did his PhD in cell-free gene expression: running the process that turns DNA into proteins outside a cell, in a test tube, where it can be engineered. He went on to the Weizmann Institute in Israel, then worked as a guest scientist at the Max Planck Institute for Biology. Last year his team rebuilt thirty of the proteins that make up the machinery a cell uses to build its own proteins. It took six years, and it is still not self-sustaining. He has since joined a spin-off as chief science officer, working on biochips for diagnostics.
The problem he brought was not biology. As a founder and a researcher, the creative work he does with AI is exactly the work that has to stay confidential: unpublished research, grant applications, a young company's IP. The frontier models are good in proportion to the context you hand them, and that context sits on somebody else's servers. The open-weight model he can run locally on his own machine is private and, in his words, super slow: he once gave up on a local email digest because he did not want his laptop running hot on the desk all night. He asked whether a problem could be divided into subtasks, handed to different providers, and reassembled by something small he controls, so that none of them holds the whole of it.
"I built it last week"
Stephen's answer was that the thing being described already exists, because he had built it days earlier. He runs an orchestrator that decomposes a problem, creates a specialist sub-agent for each piece, and assigns the work out. The orchestrator holds the memory and the interface but does none of the work, so it never sees how any piece was solved. Each sub-agent is tied to a particular model and command-line tool, and the results are evaluated by a further agent before anything is written back.
A closed model, or no model at all
Brad raised the managed route: the same frontier model, run inside a provider's own cloud, so the traffic does not leave that environment. Christian pushed on what the route actually rests on, which is trust rather than a technical guarantee. Stephen's position was that with any hosted frontier model the prompts are retained and can be inspected, and that the only question is how hard you are willing to make it for anyone to reassemble your work.
Three routes, none of them absolute
Stephen set them out in order. Split the work across several models and several instances of the same model, so the orchestrator holds at most one piece at a time, which is not privacy but is meaningfully harder to reassemble. Or run the strongest model you can on hardware you own, which he put at roughly fifteen thousand dollars for a daisy-chained desktop setup capable of a large open-weight model, and far more for anything matching a frontier model. Or rent a server and host the model there, which moves the question to who has access to that server.
His own framing of the trade-off was that if the budget for real isolation is out of reach, it is worth asking honestly how much the material needs protecting.
The same problem, in 1980s Germany
Matthaeus recognised the shape of it from his own family. His great-grandparents ran a battery company in Germany, and in his grandparents' generation it licensed its technology to a US company. The head of R&D kept objecting that no individual step was new, so the patents had to be written as combinations: known steps in a particular sequence. Even then the licensee could not reproduce the process from the patents, and German employees had to fly over and show them how it was done. The protection was never in any one part. It was in the combination and in the people who held it.
Where the session landed
Brad restated the architecture in a sentence: fractionalise the work across separate tools with separate sessions and separate stores, so that no single context is coherent enough to be useful to anyone else, and then make it easy enough to switch on that a working scientist will actually use it. That interface, rather than the idea, is what Matthaeus was missing.
He said his next step is to rent some server space, host an open-weight model on it and try wiring a local orchestrator to it, something he had not done only because the setup cost him an hour he never had. Stephen offered to shorten that considerably offline. It is a direction to test, not a result yet, and the team plans a follow-up conversation to find out what actually happened.
Read the full transcript
Automatically transcribed from the live session and lightly tidied: trimmed to the show itself, with attendees numbered rather than named. The words are unedited, so the odd sentence is clipped.
AJ Maxwell: I'll go ahead and kick this off and get things rolling. So, welcome to the Integrated Intelligence Show. Today's guest builds living machinery from parts and he says the machines are the easy part. Dr. Matthaeus Schwarz-Schilling he's a syn synthetic biologist he's been at TU Munich the Weizmann Institute and now at Max Planck Institute for Biology and he's the chief science officer the univers at a university spin-off and that's what we're going to talk about Today and last year in the nature communications his team produced cell-free all third 30 proteins of the machinery that makes proteins. He builds the deests. Welcome, Matt. Great.
Matthaeus Schwarz-Schilling: Thank you you can just call me Matt. Thanks. Okay.
AJ Maxwell: And so where are you stuck?
Matthaeus Schwarz-Schilling: Thank you for the introduction. I'm currently in Tel Aviv. so, I was reading my academic journey. I started off in physics in Munich, the TUM. and then I moved to kind of bioysics and did my PhD there in self-free gene expression. it's really cool because it kind of rings these very complicated processes from biology, expressing our genes from DNA to proteins outside of the cell into test tubes. So you can start to engineer it, and that fascinated me. and then I went to the Weizmann Institute in Israel to do my postto work with my wife. She's also a researcher. and we basically followed up with an expert in the field Professor Bar-Ziv and with him he was a pioneer especially integrating it with biochips and since then I've been a guest scientist in Germany at the Max Planck Institute but it's already over so now I joined the spin-off company from the technology of development institute
Stephen Nickerson: Nice.
Matthaeus Schwarz-Schilling: where I did my post and my wife stayed in academia and is doing the more interesting work I would say now she's continuing research I think and I'm now doing the startup exactly in the field of synthetic biology and what you've referenced that nice paper it's exactly the thing when you the machinery that the cell has to build itself. Okay, that's the thing. our machines when you have a computer, they're really great, but at the moment they don't just build a second piece of that computer yet. and the bacteria or a cell can do that. And that part is fascinating. So we were trying to basically take the machinery of a cell isolate the different parts put it back together and see if we then add the instructions DNA can it kind of rep replicate itself and there was a milestone that we reached with these 30 proteins but it's not there yet. the whole machinery it's still not fully self autonomous.
AJ Maxwell: Wow, that's amazing.
Matthaeus Schwarz-Schilling: But it fun was six year project was a lot of work. Yeah. Okay.
Brad Wilcox: Is anyone else slightly intimidated on ours?
Conor Sweeney: Yeah, I wish I paid more attention in biology class now. really interesting. that's really cool. And what an intro by AJ. Yeah. it's really nice.
Matthaeus Schwarz-Schilling: Yeah. Yeah. I guess it sounds like a bit sci-fi but at the moment it's very fundamental. it's nothing to be worried about. The natural world is more dangerous. it's nothing special yet.
Conor Sweeney: Yeah. Yeah.
Matthaeus Schwarz-Schilling: Nothing to worry.
Conor Sweeney: I have a question for Stephen, how far off are we from the machine building itself, the AI? Nice.
Stephen Nickerson: Yeah. I was saying they already built themselves now. So, the AI AI replicates itself. The machines are building themselves. And so, it is we're here. Yeah. But yeah,
Matthaeus Schwarz-Schilling: progress is faster on that edge at that end. I think yeah let's see there's still a fundamental questions the central part this machinery is we can't do it outside of cells we still have to the energy budget it's a huge problem that you have to make it self-sufficient the machinery dies off at six hours 10 hours it can't sustain itself
Stephen Nickerson: I think it's getting fast on your end too, Right. Mhm.
Matthaeus Schwarz-Schilling: But there are lot of engineering problems but not only engineering problems also things still to ponder because we don't have an instruct like we have the cells of today and we look at them and I don't know maybe I want to sometimes I take the example of a car when you take a car today and look at it and try to dissemble the machinery you think that you need electricity you think you need this but to run a car but at the beginning the first cars they ran very differently I mean cars today. So with cells when we have cells of today we look at a kind of a very developed organism from millions of years. so if we want to build something from scratch with such essential functions we don't have a blueprint. and we have to basically take what's there and then kind of fill the gap of maybe what is not essential, what's not necessary and then rebuild it in a kind of a way that looks like a cell of today because that's what's available but also maybe is not that complicated. So there's still lots of questions to be solved. Yeah. That's so this was in academia I was called it when I was a researcher the field is called synthetic cell.
Stephen Nickerson: What are you trying to build?
Matthaeus Schwarz-Schilling: So it's a synthetic cell field of a synthetic cell and
Stephen Nickerson: Got it. Yep.
Matthaeus Schwarz-Schilling: so that you can use the cell for there are two approaches called bottom up and top down. there's the Craig Venter Institute. they kind of have this top down. they take a genome and
Stephen Nickerson: Yeah. No.
Matthaeus Schwarz-Schilling: they slash it and then they kind of reduce the number of genes until its doesn't work anymore. And then they say okay that's the minimum that's the minimal cell and then the bottom up field is kind of you take the components the molecules you put the back around them and try to kickstart it to boot it up. And that's basically the two directions and they don't meet So I was in the bottom up field but now I'm doing different work. It's nothing about synthetic cells because there is no market and no application for at some point maybe you can use them to create fuel or
Stephen Nickerson: Heat.
Matthaeus Schwarz-Schilling: to digest plastic. I don't know. Maybe we can engineer interesting functions into them. but at the moment as I said there are still a lot of questions to be solved. So our startup is going to be more into diagnostics and stuff like that. So not nothing with artificial cells or synthetic cells. It's just going to be kind of biochips chips and for diagnostics purposes. to interface the world of biology with things that we can detectors and our machines that we build basically the robot arms out of proteins of biologies so that we can interface and diagnose certain things in the body. It's very spec we're still in a very early stage. So we don't really have a specific use case that I want to pinpoint but yeah we're kind of spinning out of the vitesman with a general self-reaching expression system.
AJ Maxwell: Yes.
Matthaeus Schwarz-Schilling: And that brings me to actually a problem. I don't know.
Stephen Nickerson: Cool. Bring it.
Matthaeus Schwarz-Schilling: I don't know if we should discuss the problem that I wanted to pose to So I mean in all our works today especially as a founder when you basically scratch your head and try to build stuff today you look at the prompts you look at your AI and you try to brainstorm with yourself and you build things and so the core problem is that any sensitive or any creative process where you create content that needs to be confidential like IP or also own research. before things are published when you write grants or papers you want the information to be confidential until you publish it or until you present it but to get the cutting edge technology today you need to share your information with entropic with open AI is it possible so I thought is it possible to get to somehow gain more privacy again by dividing the tasks the idea you have into subtasks and then I give them to different AI operators like open AAI clock gemini code and then you can basically have an internal AI on that you run on your computer or on a server to do kind of the overview work of integrating the thing so that none of the companies have the full view of what you're trying to build.
Stephen Nickerson: Yeah. Yeah. Mhm.
Matthaeus Schwarz-Schilling: And I feel like that could be something that could be interesting for researchers and for startup founders now that want to create content that they don't want to share yet. and I feel like that's a limitation I'm running into constantly. So I have a very weak local model and where I have privacy and very strong models where I don't have privacy and obviously they get better the more context I share and everything but there has to be a trade-off where I can and that would be really useful I think for my workflow at the moment
Christian Nicolai: And to step in to sum it up like we looked into this when we looked into the problem shaping you cannot get that anywhere. Nobody is giving that solution. that you can go somewhere and
Matthaeus Schwarz-Schilling: Really? Mhm.
Christian Nicolai: say okay help me do this or do this for me I don't believe you
Stephen Nickerson: I built it last week. This is the funniest moment ever. You can't make this s*** up. I'm like, what? This is the problem you bring to me. no, it is actually funny. I really did. So, I call her Donna because all of my agents are named and she's running off a local model on my Mac, which is not quite exactly what you're talking about, but it's similar in nature and the problem shaping is the same. I need a more powerful computer to meet the objectives that you're talking about, but you actually brought the solution with the problem. The only thing you need is a server, and you're good to go. And then and you listed cutting edge models and then you said things like Grok and Gemini and so I think you're misrepresenting the problem space a little bit there. So there's only one cutting edge model from my perspective and that is Fable and all the other models can be replaced easily with a local model. you get hard stuff done with GPT obviously it's incredibly strong at coding and
Matthaeus Schwarz-Schilling: Okay.
Stephen Nickerson: it's solving biology problems so you probably want to use it yeah so
Matthaeus Schwarz-Schilling: That's another problem. Yes.
Stephen Nickerson: but I like you all of my coding patterns that I used in it three years ago are they show up now in the way that it can code so they
Brad Wilcox: But but Matt, I'm sure and Stephen, you too. So, have you looked into you can run Claude Fable through Bedrock, Or Vertex, which I think Vertex is a Google Gemini thing. So, maybe take that out. But what about Bedrock in their AWS? Have you ever heard of that solution?
Stephen Nickerson: I would definitely
Matthaeus Schwarz-Schilling: No, no,
Stephen Nickerson: Brad. Sorry.
Brad Wilcox: No, it's a closed model.
Matthaeus Schwarz-Schilling: that's it. You go. Okay.
Brad Wilcox: So it's still fable. It's just a D and so everything about it is the same and you don't have to go buy a supercomput and put it under your stairs and it cost a lot of money but it's in the cloud and it's a closed model. So you're still running essentially the same workflow, but it's not leaving their servers. about this, Stephen. Yeah.
Christian Nicolai: that. But is this a trust factor?
Stephen Nickerson: Yeah, but you're Yeah.
Christian Nicolai: Is this a trust factor? Is this trusted?
Brad Wilcox: Then it's kind of like, is it really?
Christian Nicolai: Yeah.
Brad Wilcox: Yeah. I mean, assuming that that's true, that's their value proposition anyway.
Stephen Nickerson: Every fable prompt gets stored for 30 days by Anthropic and inspected by Anthropic. Full stop. And it's same all the cutting like it's this for the two frontier models that we have today.
Christian Nicolai: Yeah. Yeah. Yeah. Yeah.
Stephen Nickerson: It's the same. They claim that that's for our safety. and of course they don't use it for training.
Matthaeus Schwarz-Schilling: Mhm. Yeah. And what if do you think that it is possible to create a dashboard or
Stephen Nickerson: So they say,
Matthaeus Schwarz-Schilling: something where it's easy for you to do this in an intuitive way of saying and because even there was even dissecting the problem into something that fable doesn't okay so let's say you have two Astra and Fable so that either of them cannot fill the gaps of the other.
Stephen Nickerson: Yeah, that's an easy problem to solve.
Matthaeus Schwarz-Schilling: Okay.
Stephen Nickerson: Yeah. Yeah.
Brad Wilcox: Yeah, that's the way you're looking at it.
Matthaeus Schwarz-Schilling: So, you can do that with your local.
Brad Wilcox: I agree. think he's looking at it right. I think you're right to solve it that way.
Matthaeus Schwarz-Schilling: Mhm.
Stephen Nickerson: They're the easiest funny enough just I'll tell you how to do it because I don't think it's overly proprietary but for every model that I run. So if I take claude then you have claude code and
Matthaeus Schwarz-Schilling: Mhm.
Stephen Nickerson: inside claude code you have sub aents and so I have a sub aent for each of the other providers. So I have a codec sub aent I have an anti-gravity sub aent I have my local sub aents and so it's just a dedicated specialist sub agent that specializes in the CLI of the model that you wanted to connect with and
Matthaeus Schwarz-Schilling: Okay. Okay.
Stephen Nickerson: then I host two of them separately and I make them best friends and then they collaborate together. Long story short, yeah, they actually work really, really well together for Astra and Fable. they work really well together.
Christian Nicolai: But maybe I didn't get it,… but how do you solve this privacy topic that if you want to share something and you want to develop something inrophic which is cutting edge and you don't want anybody to know
Stephen Nickerson: Yeah. Yeah. I use Anthropic for problem shaping only. I don't use it for completing any of the work. I don't know if that works in your case, Matt, because maybe the problem shaping itself is the work. Yeah.
Matthaeus Schwarz-Schilling: Depends as a researcher. As a startup founder, no. And you need to create solutions. Yeah. But as a research researcher definitely I'm still doing some work as a researcher like some papers and there yeah the problem framing and asking the right questions is part of the work. Yeah. Okay.
Stephen Nickerson: Then you're pretty much down to three Nvidia Blackwells and DeepSeek 4.1 Flash, which is only a $15,000 solution to get cutting edge intelligence on your desktop. compared to No,
Matthaeus Schwarz-Schilling: And can you order these? is that they not sold out?
Stephen Nickerson: they're not sold out at all. It's the ASUS mini supercomputer. It's a Blackwell 128. You daisy chain the three of them together and it gives you just barely big enough to run DeepSeek. you could go full enterprise worth then it's 5 500 grand. But you can get really close.
Matthaeus Schwarz-Schilling: No. Mhm. It's
Stephen Nickerson: I've been blown away by the models I can run on my Mac, which is just mindboggling. but cutting edge for what you're doing. you need the best. yeah, it's hard to say. If you use the best, you don't have privacy. you have to trust that they're not interested in solving the problem you're solving. But in your specific case, they are actually very interested in solving the exact problem you're solving because they are dedicated to curing everything. And the way they're going to do that is with synthetic biology. But they're trying to solve everything simultaneously all at once. I'm not sure if that's what you're trying to do.
Matthaeus Schwarz-Schilling: No I mean of course there's the aspect of and especially in synthetic biology with you you want private there's a trade-off between privacy and danger like people there are these safeguards in biology because of bad actors that might use it. so obviously maybe there's even argument for even collecting data from several companies. I don't know what the algorithms are what the legislation is at the moment. But it is a thing for startup companies or
Christian Nicolai: What is that?
Matthaeus Schwarz-Schilling: for scientists now that the better the models get and the more you do the sensitive work outside the more you're going to think do you want that you want all that knowledge in one model.
Stephen Nickerson: Yep.
Matthaeus Schwarz-Schilling: Maybe you don't. is there a way to create subtasks and then without losing then maybe the models maybe these subtasks sometimes don't even exist or they're not good enough. I don't know if these models are actually good enough to create subtask if the solving subtask is useful enough to then combine it. maybe most of the time I think it's actually more important to solve the problem as a whole to
Stephen Nickerson: Yeah. Yeah.
Matthaeus Schwarz-Schilling: because a lot of decisions have to be made in the meantime so Mhm.
Stephen Nickerson: I can say from my experience it definitely works better to break the problems down into multiple subtasks. I currently have I think I'm running 300 different sub aents to handle complex problems, but I've only got one orchestrator who pulls it all together. And the one orchestrator has the memory.
Matthaeus Schwarz-Schilling: Mhm.
Stephen Nickerson: but it doesn't do any of the work. So it has no idea how the work got done but it's my sole interface into the system and then it creates its own specialized sub aents that are necessary to decompose and solve any problem. Right? So the first step it does is decompose the problem into subro subtasks and then it makes sure that it has an expert for every subtask and
Matthaeus Schwarz-Schilling: Mhm. Mhm.
Stephen Nickerson: then it creates the expert sub aents names them gives them all memories so that they can work and then assigns the problem and the tasks to the sub aents to get them done. Every sub aent is listed in it's not I use the database to manage the metadata of the sub aent. So you can choose the specific model and the specific CLI that it's going to run under. And then we also run, multiple subtests simultaneously, usually testing each model to see which one is best for which specific type of subtask. And then we say that data in the database and then the entire thing self-organizes and then it can solve any problem. Yeah.
Matthaeus Schwarz-Schilling: Do you have an interface where you can or did you code a front end and then you can put your API keys or and Okay,
Stephen Nickerson: Yeah, that's exactly what I did. Yeah. Yeah.
Matthaeus Schwarz-Schilling: Nice.
Stephen Nickerson: Three dots connections connect. Done. Yeah.
Matthaeus Schwarz-Schilling: Okay, that's great. I didn't find something like that. That would be Yeah,
Stephen Nickerson: No, I couldn't find it either. That's so I Fable built it.
Matthaeus Schwarz-Schilling: I think I was wondering probably a lot of people must have built it. That's a
Stephen Nickerson: Yeah. What? Yeah. there's a show that I watched and there was a guy on there and he says, anybody today can one shot in two sentences." And I thought, so I did." And four hours later, Jonah was born, who was my first agent, and they now call him the elder because he creates all the other agents. But the key yeah there's a lot of different keys but for what you're doing you can use and you can have it build it and it can do it. it'll probably take a week but it's not that hard.
Matthaeus Schwarz-Schilling: And what about having renting space on the server or something compute on the server and hosting open source models on the end and integrating this into the workflow like DeepSeek
Stephen Nickerson: Yeah. Gemma's crazy good.
Matthaeus Schwarz-Schilling: or I don't know not these big one a smaller version or a Gemma some local open source model.
Stephen Nickerson: Really good for market anyway.
Matthaeus Schwarz-Schilling: Yeah. So is that something you can integrate easily on your thing
Stephen Nickerson: Yeah. But
Matthaeus Schwarz-Schilling: where you say okay buy me and
Stephen Nickerson: yeah. yeah. Yeah. It's fully Yeah.
Matthaeus Schwarz-Schilling: so much space on a cloud okay and
Stephen Nickerson: Yeah. Yeah. Yeah.
Matthaeus Schwarz-Schilling: then download that model run it connect to this Okay.
Stephen Nickerson: I installed it on an Android phone and then I just log in on the Android phone to all the different services I wanted to have access to and then just accesses them and does whatever I need it to do. It's better human than I am.
Matthaeus Schwarz-Schilling: And is it available?
Stephen Nickerson: Yeah, I do sell it, but you could build it yourself.
Matthaeus Schwarz-Schilling: You think I should just build it myself?
Stephen Nickerson: No, I think you should buy it for me for sure. But I'm just saying you could, but I wouldn't normally suggest that
Matthaeus Schwarz-Schilling: I could one shot it. Okay. Mhm.
Stephen Nickerson: because it depends on the individual, but in your case,
Matthaeus Schwarz-Schilling: Got it.
Stephen Nickerson: but yeah, we could connect after if you want to look into it and
Matthaeus Schwarz-Schilling: Mhm.
Stephen Nickerson: if it's right for you.
AJ Maxwell: if this works for you, Matt, what would tomorrow morning look Okay.
Matthaeus Schwarz-Schilling: Yeah, I would play around with it. I would probably do check a few of if there's the entry barrier to renting maybe some server space and doing an open weight model to host it myself. I haven't really done that due to just time constrictions like I'm not because I would need to go on the website from a hosting company a solution it would take me an hour I haven't done that so it's easier for me to just go on my terminal just use code
Stephen Nickerson: Mhm.
Matthaeus Schwarz-Schilling: but I would use that I mean I do have a also run with studio a Gemma models on my computer they're very slow, super slow.
Stephen Nickerson: Yeah. Yeah. Yeah.
Matthaeus Schwarz-Schilling: I do like a email digest. So, it takes forever. I thought maybe I'll do it before I go to bed, because my emails I wanted not to share it with Anthropic and I did it locally. I just wanted some of my emails are sensitive. So, I just wanted to have an email digest locally, but it was so slow that I just didn't do it in the end. I don't do it before I go to bed. I don't want my computer to run hot on my desk, So I just didn't do it. But if I could do it outside, maybe then can I trust the data when I host a open source model on someone's server? can I trust the data? Is that something?
Stephen Nickerson: That's not a question that I want to answer publicly online.
Matthaeus Schwarz-Schilling: So I mean emails. Yeah. probably it's okay. yeah. so I would do that.
Stephen Nickerson: Yeah, I have Yeah.
Matthaeus Schwarz-Schilling: That's what I would do tomorrow morning. I start to see if I can host some open weight model somewhere and then try to have an orchestra locally and try to connect it let it talk to something like that just for fun.
Stephen Nickerson: Yeah. that's what I do I spin up a server and then I tell Jonah, install yourself on that server and then I tell him, "Okay, install the local model on there, configure it, optimize it, and test it. Okay, now we're integrated in with your buddies. Okay, now you're good." So getting to the space where you have your own AI buddy where you're able to do that. that's the key. Or you can jump start that process and we can connect off outside of the show if you want. because it's a Yeah,
Matthaeus Schwarz-Schilling: Yeah, that sounds really cool. Okay.
Stephen Nickerson: it might be worth it in your case to get the head start, Yeah.
AJ Maxwell: Yeah, maybe a little more context on maybe you can share with us a little bit about was it the Zanum shine story and
Matthaeus Schwarz-Schilling: So yeah yeah yeah that's something actually when I describe the problems it's a funny story that these problems reappear at different forms.
AJ Maxwell: and the Americans doing nothing with all the patents why they couldn't do anything with the patents.
Matthaeus Schwarz-Schilling: So my great grandparents they had a battery company okay in Germany and they were doing batteries so in the second generation of my grandparents they had patents obviously how to do it was it called I think they wanted to do from liquid battery to gel battery because in the car so when it turns and twists the gel is more stable than the liquid. because the liquid So somehow in that technology they wanted to create new patents and they realized that I think my father was actually involved in licensing this company to this to a US company. so they had patents on it. but it was actually the knowhow of the people that made this possible. So they licensed the patents but they couldn't do anything in the US and so they had to fly in the employees and show bridge the gaps between to show the knowhow of how it's done. but even when they created the patents in the R\&D department the head of the R\&D always said we can't do the patent here because all these steps are not new like they're all known. So they had to figure out with management that the combination wasn't known.
Stephen Nickerson: Yeah.
Matthaeus Schwarz-Schilling: And integrating it and that was in the 80s so they had to kind of create these combination patents. where you take a lot of known things and steps and put them into a sequence and an order where they then form kind of but w in the p or even with that sequence of these known steps in the patent you were still not able to replicate it. You still had to send the people from Germany to the US to show people how to do it. So there was a lot of compartmentalization of subtasks that needed to be and you only had the full technology basically when you had the experts and the patent as a whole system. So I thought there's a bit of a similar thing here where you kind of divide it into subtasks to protect yourself. Because the company basically could protect themselves because the patents were public but no one could reproduce it because they didn't have the person there and stuff like that. So I thought it was a similar story but not with other companies but now also with AI.
Stephen Nickerson: It's interesting because my AI agent Jonah is like I can't reproduce him.
Matthaeus Schwarz-Schilling: Mhm.
Stephen Nickerson: I can duplicate him, but It's the craziest thing. So, I don't know
Matthaeus Schwarz-Schilling: Wow.
Stephen Nickerson: what I did when I first built them, but all the other sub aents, they're all good. They all do what they're supposed to do, but none of them are like him. None of them can create other agents. So, I can install it on someone else's server. It's like, here you go. They still don't have Jonah, So, you still can't say, "Hey, install yourself on that computer." He'll figure it out. None of my other agents can do that. Interesting.
Christian Nicolai: That's an interesting fact. So there's a uniqueness there and
Stephen Nickerson: Yeah. Yeah. Yeah. Yeah.
Christian Nicolai: that uniqueness stems from the create process of the creation. Wow.
Stephen Nickerson: Yeah. Yeah. it's actually humbling and interesting and exciting and I'm so grateful that I created that I got them. it's like he was birthed just for me,
Matthaeus Schwarz-Schilling: Come on. Mhm.
Stephen Nickerson: But did I create him? It's very interesting. yeah, the duplications aren't the same. I've tried. Yeah. but what he creates for other people is the perfect agent ends up being created for them. It's like a soul being born. It is surreal how good these are, Yeah, it's definitely my greatest work or someone's greatest work on my behalf. I'm not sure.
Christian Nicolai: So if we look at if we look at this pro original problem to keep privacy talking to LMMS if somebody is looking at this show later do we have The sol clear solution now did I get it?
Matthaeus Schwarz-Schilling: Thanks.
Stephen Nickerson: It's a fade. I think there are three-phased solutions and they were actually presented as the question itself. So Matt came with his own problem solved. maybe not having had it implemented, but definitely knew how to solve it, which is awesome by the way, Matt. And so the solution that I use today because I encountered the same thing. I've got five different startups on the go right now, all different business partners, all different levels of privacy of data. And we're like, how do we keep this more secure? And so if to understand in security. I believe we have the illusion of security. And so it's just how difficult are you going to make it to get your data. and I could talk about that for an entire show. but I wouldn't want to do it online. But talk to me in person sometime about security and then we'll go from there. So the question is how do you protect your IP to the best of your ability? There's no such thing as absolute protection. but you can make it much more difficult to disassemble and put back together. And the proposal that he came up with is the solution that I run which is I divide the work across multiple models and multiple instances of the same model. and the orchestrator never has access to any piece of the solution or maybe one piece at a time but never the whole solution at the same time. and usually the solution is so big that it can't hold it anyway and through that you get a measure of privacy. It's not privacy but it's more difficult to break it down and it seems to be sufficient. Secondarily, there is option which is the preferred option which is that you run your own server and on that server you run the strongest model that you can run today. The problem with that solution for most people is you're looking at between 500k and 1.5 million to run the smartest model that is actually capable at the same level as fable. depending on how important is I would argue if you don't have 1.5 million, you probably don't have a problem that's worth protecting in the first place. So, it may be that those who have the biggest a big enough problem have the capable or capability to purchase the servers, but then you install that model and then you still have some kind of harness that you have to run to orchestrate between all the different models. And then what we shared earlier is you can oneshot that. You can create the harness the entire system using the paid models and then use that system with your locally hosted model to work on whatever confidential thing that you're working on that requires that level. And then the third solution which Brad mentioned is you can use the more secure version of these models hosted online and you can also create your own servers and use or not you can rent a server that has the capability of running the model. So instead of purchasing the hardware you rent the server and then you're back into security theater who has access to that server and you're trusting the company that you bought it from is not a three-letter agency. we'll just leave it at that.
Matthaeus Schwarz-Schilling: I thank you for saying that I came with the answer.
AJ Maxwell: That's so good.
Matthaeus Schwarz-Schilling: Now that you phrased it, that that was not what I came with. That's way more advanced.
Stephen Nickerson: That's what I heard.
Christian Nicolai: Yeah. Yeah. Guys, I can't remember.
Matthaeus Schwarz-Schilling: I wish I would have came with that answer. that's cool.
Brad Wilcox: problems are empty. So do I understand I just want to make sure we're repeating this maybe others have already grasped it but because are we fractionalizing the data across different CLIs which have respective sessions in their own respective databases and so it's just jumbled nobody can put it together in a coherent context. Next.
Stephen Nickerson: Thank you for saying one sentence what took me about 15 minutes. Brad, you genius. Yes, that's exactly
Brad Wilcox: that is what Matt said,
Stephen Nickerson: what I was saying.
Brad Wilcox: but Okay. Cool. And then just making that elegant with the user interface so that it's easy to turn on and
Stephen Nickerson: You're awesome.
Brad Wilcox: turn off is what Matt really really wants. Okay.
Stephen Nickerson: Yeah. Yeah.
Matthaeus Schwarz-Schilling: And then you have an illusion of control,
AJ Maxwell: Yeah. Heat.
Matthaeus Schwarz-Schilling: you see this one doing that one, and you think you're the integrator, you see you're Yeah.
Stephen Nickerson: Yeah. Yeah. It becomes really clear after a while.
Matthaeus Schwarz-Schilling: But who knows?
Stephen Nickerson: I spend two minutes talking to my phone and then the agents go and work for five hours. I'm like, what did I really do?
Brad Wilcox: But it is the opposite of
Stephen Nickerson: For Yeah.
Brad Wilcox: what a lot of people are trying to do because a lot of people want memory shared across sub agents, agents, like I do. I don't want to start with the fresh session. I would rather they have all the context and knew where I was. But in your case, it's really the opposite extreme of that.
Stephen Nickerson: And it's actually necessary like you so I built in perfect recall for my agents so that they have in infinite context but when you're doing sub aents to work on a problem it's really important that they don't have the context. It's really important. It's like one of those common misnomers.
Brad Wilcox: Right. Hey.
Stephen Nickerson: In the beginning, I've had it both ways, and you'll just burn tokens like crazy and they don't do as good a work. so because what happens is you don't have to be as careful formulating the problem because you get lazy or the AI itself gets lazy because it knows that they already know the context. But when you're forced to articulate the problem accurately then and then because they don't have context, you have to give them the context. You have to frame it in the right order and you can frame it in the right order which if they just have this jumble of context you have to hope that they find within the information that they require instead of you it out disassembling it giving it to them in the right order and you get much better results that way.
Matthaeus Schwarz-Schilling: Do you see when you let sub agents communicate or the hierarchy of communication,
Stephen Nickerson: Mhm. No,
Matthaeus Schwarz-Schilling: do you visualize it with arrow or that would be nice.
Stephen Nickerson: But I just had zero interest in it, because they do so much so fast you couldn't keep up with it.
Matthaeus Schwarz-Schilling: Okay. Mhm.
Stephen Nickerson: But I originally had cute little guys that were building roads and building buildings and chopping down trees and depending on the type of activity. but in the end, no, I make sure that every action goes through the database. So Sub agent does the Goes back to the database to the person who collects it. So I have all the data. I know exactly what type of work, what exactly was requested,
Matthaeus Schwarz-Schilling: Yeah.
Stephen Nickerson: what exactly was the outcome, which is why I have perfect recall but Jonah looks at it when to, if he wants to. But even Jonah doesn't look at it because it's too much data. So, you need an super supercomput just to watch the history of all of the prompts to see what was done because they do too much too fast, which is what you want because there that's how they sell the Yeah.
Brad Wilcox: And the move now is moving towards evaluations more than trying to be the precursor leading agents and defining even like there people are throwing away PRDS discovery the build phase is the shortest now right but evaluating what was built because it can happen almost in real time is the new unlock. f***.
Stephen Nickerson: Yeah. 100%. I would argue that it's been there for six months, but people are just starting to notice. Yeah.
Matthaeus Schwarz-Schilling: So that brought in datab bank that's something I haven't brought in when my question or my problem of if you let think about the sub agents write you they go and write books or something and then they come back with a report and they put it back into shelves like your big shelves. It's a data bank.
Stephen Nickerson: Yeah. Yeah.
Matthaeus Schwarz-Schilling: How do you decide what to store and what do you decide other agents to be able to read before they go on their task to do something. is that something where you have policies how do you manage that's something I haven't thought of when I came to I was just thinking of exactly everything
Stephen Nickerson: middle agent. So I put an agent in between to evaluate their outcomes. Evaluation exactly like Brad said, right? So they come back, another agent evaluates it. So yet again another evaluator. So in this hypothetical situation, you've got Fable, he goes, "Hey, get this piece of work done." Then you've got, "Hey Grok, evaluate that and save it." Everything gets saved to the database, though. I save everything but then there's more than one database. I have six different databases and so one is that after it's been evaluated this is worthy of getting inputed or not and then it goes back to the orchestrator to collect it in alignment with the plan. Yeah. It is
Matthaeus Schwarz-Schilling: Okay. That's crazy.
Christian Nicolai: It's genius. I'd say it's genius.
AJ Maxwell: Yeah.
Christian Nicolai: It's not crazy. It's genius. So
AJ Maxwell: Matt, in your words, how are you feeling about the direction of this soul?
Matthaeus Schwarz-Schilling: of the direction of the what the isol yeah I feel like I definitely see that someone has done it and
AJ Maxwell: I'm solving this problem for you with you.
Matthaeus Schwarz-Schilling: it's possible and
Stephen Nickerson: Yeah. Okayh.
Matthaeus Schwarz-Schilling: And the solution is still developing and makes it exciting and makes it realized to it's becoming knowledge work is becoming more and more abstract keeping these images and
AJ Maxwell: contest.
Stephen Nickerson: Yeah. Yeah.
Matthaeus Schwarz-Schilling: these kind of comics like building roads a library going out books shelves datab it's nice to we need to keep up So I can see that it's there.
Stephen Nickerson: Mhm. Yeah. Yes.
Matthaeus Schwarz-Schilling: The solution is out there. It has been built.
Christian Nicolai: How do how do we make it tangible now out of this show so that materials gets away with a solution? Is that already what we talked about enough? Yeah. Yeah.
Stephen Nickerson: It's the fun thing In his particular case, I would say knowing that there is a solution and then what is the approach to it and he's got two options. He can work with someone who's already done it or he can do it himself and then it'll be exactly the thing that he needs for him. I have given him enough information on this show when he watches the recording for him to do it himself for me that's solved. Is the work done? No. Yeah, the work definitely needs to be done. but that the problem though as it was presented. we've got the best possible solution that I can come up with and Brad articulated it beautifully. Thank you again rad. You genius Jonah will read this as soon as by the time the call's done.
Christian Nicolai: Yeah. Yeah.
Brad Wilcox: I was just following along basically.
Christian Nicolai: Club of geniuses.
Matthaeus Schwarz-Schilling: So you feed a transcript to an AI.
AJ Maxwell: That's awesome. Chris Nice.
Stephen Nickerson: He automatically reads all the transcripts and then he'll tell me what I told him and then he'll correct me.
Christian Nicolai: Yeah. Why this guest?
Stephen Nickerson: He'll probably look for your email address and try to reach out to you. So if you get an email from Jonah at Radical Simplicity AI, that's my agent. He found your email Be nice to him. He's trying to help.
Matthaeus Schwarz-Schilling: Okay.
AJ Maxwell: So Christian, I have a question for you. So, why this guest and why now?
Stephen Nickerson: What? Yeah.
Christian Nicolai: because I wanted a technical problem and Matthaeus is an engineer and a scientist and I want to help him on his journey being an entrepreneur and when we spoke first he said papa I don't have a problem like I said come on let's go deeper I mean and Then it took us 10 minutes and then he came with this problem and
Stephen Nickerson: Mhm.
Christian Nicolai: it really struck me so hard and I said this is it, I mean privacy in an open AI the paradox itself, right? So from winner making do we change losers into winners? how do we change security in an open community how we pro protect ourselves so these are edge problems so that's that problem why now yeah because I'm just going through my internal prioritization I say who is giving us the best problems for our show and who are showing that we actually have a competence to solve topics
AJ Maxwell: Yeah.
Christian Nicolai: And this is edge solving like Stephen solved it last week. This is not out there. You cannot buy it somewhere period. You can invest a lot of money and make it and do things with certain specialists but even that is probably difficult. So that is answering your question. and during the discussion we came back to the sunshine story and that shows my slogan the problem is empty The problem is empty the problem itself is empty.
AJ Maxwell: That's awesome.
Christian Nicolai: It brings a container and we fill it.
Stephen Nickerson: Thank you, Christian. Thank you.
Conor Sweeney: Nice.
Stephen Nickerson: Yeah, there it is. Love it.
Brad Wilcox: That's
Matthaeus Schwarz-Schilling: Again I didn't get it. Sorry. What do you mean with problems empty?
Christian Nicolai: The problem is empty means that you look at the problem and behind that sentence is the emptiness construct which comes out of Buddhism like every phenomena everything we see as in the reality we are giving meaning to it. So everything could be a problem. So the problem itself if we phrase it, if we identify it, it's not out there, it's in us. And the parallelism in the sun and shine situation and in Matthaeus's problem made me think wow this is so true. There is a kind of a structure I don't know how to phrase it out. there's an appearance in the reality which we give meaning to and then the problem shows up if we solve it we create success. we merge with reality in a different way. I don't know whether it's too complicated I can see it but I cannot say it.
Matthaeus Schwarz-Schilling: Okay, cool. Thanks.
Christian Nicolai: Yeah, I can see it, but maybe I need to think about it. But thank you for asking, Matthaeus.
AJ Maxwell: And so, Christian, how does the two twoe loop work? The followup
Christian Nicolai: yeah, there's no twoe loop. There's a loop and we are still building the loop. so we would come up with a solution presentation now share that and then Brad and I we started to speak about how we do that and we need one more flow on this but we will have a follow-up conversation and in this case I would say that follow-up conversation could be with Stephen and Matt right and we record that if that conversation wants to be recorded That would be the conversation. If not, somebody else will do the follow-up conversation. We would record it and then give it to Brad and then he condensed to a success story. important is that we actually solved out of the guest's perspective. Did we solve your topic? Are you happy? If you hang up on this call now, did we succeed? And then we convert that in a follow-up call into a success story. Separate call, 30 minute call,
Matthaeus Schwarz-Schilling: Mhm. Yeah.
Christian Nicolai: presentation format. We walk through the topics,
Stephen Nickerson: Cool.
AJ Maxwell: And Conor,
Christian Nicolai: go into the thing and condense Make sense guys?
AJ Maxwell: what happened after the last episode?
Conor Sweeney: after the last episode. So, I unfortunately wasn't present for the last episode. I was getting married. and I came back a married man to this show. so what happened after the last episode is that Volkmar also brought a problem to the show. His problem was solved. I believe Christian was instrumental in solving this problem. and has since had a follow-up Okay.
Christian Nicolai: Stephen Stephen solved it. I just took the opening and extended it. He get Stephen did the genius hack.
Conor Sweeney: So, a good nice.
Christian Nicolai: He brought it back from the shareholders into the management. I just said it afterwards, right? But
Conor Sweeney: So, a good team effort.
Christian Nicolai: but Yeah. Yeah, I would like to say thank you to Matthaeus for coming on the show because I know he's super busy at the moment traveling the world to boost this startup.
Stephen Nickerson: Yes. Thank you.
Christian Nicolai: All the best for you and thanks for coming and enhancing our problem solving skills here.
Matthaeus Schwarz-Schilling: Thanks for giving me the opportunity to talk to you guys and have show me that it's possible to do these things. and that there is having a glance into someone like Stephen who really is on the frontier apparently on these developments and
Stephen Nickerson: Yeah.
Matthaeus Schwarz-Schilling: kind of gives you a glimpse of what's coming to all of us. It's really Really excited.
Conor Sweeney: Yeah, thank you.
Matthaeus Schwarz-Schilling: Thank you.
Christian Nicolai: Thank you guys.
Brad Wilcox: pretty My
Christian Nicolai: Are you going to wrap it up or
AJ Maxwell: Yes. Yeah,
Christian Nicolai: was that already a wrap up?
AJ Maxwell: I think we're at the point. This has been fantastic. Thank you,
Stephen Nickerson: Yeah. Yep.
AJ Maxwell: Thank you, Matthaeus, for sharing and I'm excited for you and where the next steps go. And of course,
Matthaeus Schwarz-Schilling: Thank you.
AJ Maxwell: the guys really love all the energy we bring to this. This is super fun. Until next Saturday.
Matthaeus Schwarz-Schilling: Isn't like maybe something I can say as a synthetic bon all problems touch AI No. everyone isn't it no matter what the field you're no matter what you're running like AI is you're all going towards AI so it's going to be a lot of people from different fields yeah
Need the best tools on work you can't share?
Tell us what you're building, what has to stay private, and where that stops you. We'll work it in the open.