The First Programmable Turing Complete Chemical Computer | Lee Cronin, University of Glasgow

1:18PM Jun 21, 2021

Speakers:

Creon

Keywords:

robot

chemistry

molecules

reaction

people

machine

chemists

computer

lab

chemical

problem

molecular

material

bit

code

space

bond

state

reagents

important

Good evening, good afternoon, and good morning to wherever you are in the world. So welcome to the sixth, I think it is now presentation on the molecular machines group series that we've had running this year. If you're watching afterwards on YouTube, also welcome to tier as well, this evening, we have just one speaker, which is Professor Lee Cronin from the University of Glasgow. What we'll experiment with today is we're going to have the question and answer session off the record, whereas usually we end up recording it and posting it online. So if people want to ask questions that are a little bit more provocative, or they want to stimulate more discussion, but wouldn't be willing to have it shared in a public forum, please feel free to do that. That's how we'll run it today. And then we can see how people respond to that after the fact. Lee, if you'd like to share your screen, I tend not to do too formal introductions, the floor basically be yours. And you can kick us off whenever you're ready. Absolutely. Okay. Thanks for the invitation. Guys. It's great to be here. I'm long known for sight and taking part in various activities, years gone by what I've what I'm going to do today is talk to you about how we might rethink chemistry does can be used to to make machines, actually, because a lot of what the chemists tell you about the way they make molecules is they lie to you. They say that they're doing molecular programming and engineering. And actually lying is probably a strong term, they are kind of talking about their artists and approaches to doing chemistry, which is not reproducible in general and relying on their skills in the bio tree. So what if we could think about it a bit differently, and use the skills of the organic chemist? Absolutely the molecular precision, they aren't able to do undoubtedly, but somehow encoded in a different way. What would that give us if we can think more systematically about building technologies. So one of the things I want to purposely do in the next 14 minutes, I will, I will go slow on some things faster than others. I won't apologize for that. I will just say keep up. And I do apologize slightly. But I want to go on the slow bit for you guys to ask the deep questions on a bit. So just self fulfilling, I won't go over all of them at the same speed. Otherwise, we won't have the same amount of time for discussion. And I think discussion and you know, is more important almost in the talk. So the talk is just an excuse for us to say, Well, what is the cheering, complete chemical constructor. So the mission of the lab, we want to make a life form. We want to digitize chemistry, we want to build chemical computers, I want to look at the transition to information, all of those things, just checking, you can see the movies moving. Great, okay. Part of the idea here is to they all unified by one important thing that we've got some degree of kind of understanding what information is the prominent will we have in, in most of science, particularly physics is we don't really know what we mean by information. Because when we have information, we talked about information, we already have an encoding and decoding system where we have some kind of engineering, that's already been done. Anyway, what I want to talk to you about is how to construct molecules and using a Turing complete approach to building molecules in reality, and this is about the same thing as running a computer program on a digital computer. It's about the physical construction paradigm that is Turing complete. That means something quite interesting when you're talking about constructing objects and producing languages, which can, which they can express themselves to formalize the ability to make molecules on demand without human interaction using state machines digitally. This is important if we're going to go to molecular manufacturing actually, in what we mean by molecular manufacturing rather than some chemist in the lab putting stuff in around one flask maybe spending some acid on the floor. sure they've got molecular control of as determined by the NMR but it hasn't really right it's not molecular control is something else. I'm making fun here if you are organic chemistry you like making stuff don't be offended. I also drop stuff on the floor and super my robots. The robots do it programming a programmer Lee program Weebly and it's quite funny. So chemistry is though programmable from the point of view of abstractions from retro synthetic glasses, it is possible to take a molecule and chop it up in your mind and think about how to make it in the oratory. And that is a triumph of modern organic chemistry, to be able to take complex objects, break into parts, work out what the reactive intermediates need to be, and then go and join them together and make that objects

in the laboratory. The problem though, is there's no physical instantiation in a physical synthesis system of the outside of human being. That means every human will do it slightly differently. Every human will have a slightly different part bit of ambiguity. It's not like running a piece of code and executing that code on our Any qualified hardware, it doesn't work. So if we want to be able to make objects at the nanoscale and make them reliably, we need to understand how we can qualify the assembly of these objects in architecture. So how do we get to Universal chemistry why I am a synthetic chemist, I like making stuff, because I realized we wanted to rather we could make any kind of, you know, we could choose any paradigm to try and make it complete computation that complete, but people aren't going to adopt it unless you exploit 200 years of literature. So you can bring that in, you enable the chemists on earth today that the chemists at the you know, at the heart end that doing it really hard designing and building and making and validating want to make it modular and reproducible. Wouldn't it be great if one of you guys one of your students could make something in your lab, a lab that maybe you took a couple of years to work out, and then you could then send the code to me and I can make it my lab and one day, that would be an amazing accomplishment? Sadly, in chemistry, that's not the case. And one of the things that we need to be able to do with chemistry is going to make contribution to new technologies is to make that the case. So I realized a few years ago, the idea is we have to automate the Rambam flask. Now, the round bottom flask is really the most essential component if you like, like the read or write head in the Turing machine. Because the the phragmoplast wish to view as a process state, so you can pour stuff into the vomfass sure, chemistry itself, the actual reactions are stochastic, don't worry about that, for the moment, that's a very good argument, we can have the act of putting something into a flask. And due to changes in the state of that flask, temperature, pressure, whatever add a catalyst is really like a state machine. And you can then start as a function of time to do transformations. So I figured a few years ago, that we might want to build a chemical state machine that would like basically take everything in the form that the chemists would use, and hook it up to a router, give it an IP address, and control the data going in and out and build up a a language that would express express all the unit operations of chemistry. And that's what I'm going to show you basically all of that in the next, say 30 minutes, and just a whole bunch of examples. So okay, so let's start with the chemical state machine. So the idea of a chemical sensor state machine is really simple. Right now lavora tree, the way that you do preparation is you write write procedure, you then you interact with that procedure, so I could write down a procedure give it to my grad student, and my grad students are all smarter than me without exception, okay? And they'll interpret micro segments, it doesn't know what it's talking about, or interpret that and do that bit differently, I'll do that differently. I'll do that with me. And before you know, each really smart grad students will their own recipe based upon my recipe, they do the experiment, get a result and update that proof. But you don't really know what tacit knowledge hasn't been captured. Okay, there was ambiguities in that process. Whereas if you digitized it, you have a code a graph, you compile that code to graph, you get a result, adaptive improve, and you can have a procedure that create a version control system. And you basically would treat probe, the chemistry synthesis like pro writing piece of code, and you would version that code and run it. Okay, that's the dream. This is not new.

There are lots of single class computing machines out there, which mirror what's going on in chemistry right now, like a DNA synthesizer, peptide synthesizer, carbohydrate synthesizer, but these are not universal. These are just almost like finite machines that can make one type of bond, okay, which is great. It shows that it is possible to use a digital code to instantiate a program that would make you know, in fact, if it wasn't for RNA synthesis, we probably wouldn't be hopefully, fingers crossed, or whatever it is getting out of this awful pandemic, right now. So, you know, that's a real that's a real marvel of being able to do that kind of bio nanotechnology actually. So, what why is it universal? What are we trying to make, are we trying to make a chemical synthesis state machine, and what this does is a chemical synthesis state machine, you put in the code, let's say to make that with the the kind of the abstract code to run your reaction, it would have a graph of the resources required to instantiate the reaction. So we have some understanding of what reflux would look like. So basically heating up something allowing evaporation and then condensation, adding something dropwise forming a solid, all these unit operations which a chemical engineer would think about the macroscale and maybe in a in a big pilot plant are brought to the fume hood to normal chemistry fume hoods. And then you then have a state machine which is then executed as a function of the input and out Let me show you look, and you have an interface with the actual physical world. And really, the abstraction of the chemical assembly make really allows us to think about state machine that can make any molecule material. This is where I kind of analogy, it was just an analogy to start with a Turing machine and then became much more concrete. So what is the Turing machine cheering machine is a, a theoretical entity that can basically simulate itself. Okay. Now, if very computer scientists here want to do more formally, you can think about what you would need for recursive enumeration and unbounded memory. But basically, if a Turing machine should be able to run any digital program, put very crudely. So now let's do the same analogy. Let's imagine a chemical reaction kicks in a field, they could do any reaction and make any molecule. Is that possible? Well, reactions aren't that hard, what you do is you throw the stuff together, or you mix this stuff together in flast. So you'd start a reaction, you would then work up the reaction at the end, you isolate and purify four steps for abstractions, go in a loop make anything? What is hard, of course, is the knowledge to choose what solvents, what workup, what catalyst, when is the reaction finished? Well, how to minimize impurities. There's lots of magic or tacit knowledge, I would rather call it the magic that is not trapped, or captured unambiguously. But our dream is to go through a thing called computation. And computation is the process of running a chemical code on any compatible hardware, and getting the same result. What is computation, computation is like taking a code, run it on some hardware get the same result on any hardware with little error. That's the dream. So that's what we're going to take, I'm going to take you through in the next 25 minutes. So to do this, we basically built a language, and I'm gonna describe how this works over the next while. And this, this language looks very simple. It's just a it's just a markup language, actually, but it's so much more than markup language, because now we've made it fully able to express itself, it's actually Turing complete. And that that has some real power. That is that I will explain as it becomes physically instantiated, so it's not Turing complete running a simulation of itself is Turing, complete with respect to running the computation that is running those reactions in your reactor in your few good for real, not a simulation. Okay.

So first of all, let's go back to the beginning where my I went my group and said, right, I'm bored of no one being out, reproduce each other's work. Let's make some robots. So everyone went, Okay, well, actually, that's not the way it happened. I went to the lab, and I'm bored of you not following my instructions. Here is the code. Here is what you do, right? It's a bit like, you know, you must obey these rules. I won't make any jokes about mask wearing and the pandemic and aliens and all the conspiracy we're facing right now. But suffice us to say, when you ask any intelligent and free minded person to follow your rules, they'll probably tell you go away. So I thought that was a fail, they weren't going to buy my robot rules. So I went to Latin expensive to build some cool robots that would, would, you know, save your time and labor and take all the drudgery out? That's better. We'll talk to you now. So we designed a whole bunch of robots. And obviously baked into those robots was a programming language I built. So I had my group went and built these kinds of open source stuff at the beginning. And pumps and valves that we built ourselves. The reason why we built our own microcontrollers. My design electronics wasn't because I didn't have enough money, or I was too much of a geek, although I don't have enough money on Amazon, but was because I wanted to control the data. I didn't want to innovate anyone bullshitting me about my data going into my object and vendor locking it, okay? I don't like data bullshit. And there's a lot of it out there in the early 2000s, where people were trying to sell hardware and make it really expensive and inaccessible. We had to try and eliminate that. Anyway, we were a few years went by and we basically built our first computer. Hopefully you can see this this is just the assembly of the robot in the lab. This looks very familiar to a bench chemist, right rotary evaporator has some glassware. There's some funny pumps and valves in the backbone. Here's the firt is the first generation the thing cleans itself at the beginnings self cleaning robot, which I thought was quite good. Now you might see now the reagents have been mixed as reaction has started in that round bottom flask. And other reactions ended now it's doing a workup of separator with a liquid liquid extraction over here. Then after this extraction, you've got this auto evaporate to basically produce crude material. Then you then add the material you need to dissolve up your crude and then you move it to the crystallizer To produce the final material in this case, I should say at this point, this is a different lecture. The reason I, the main reason for design or the computer using this language, it actually uses a lambda calculus is actually because I wanted to brute force the origin of life. Originally, all this technology was built to brute force, the origin of life, look for Molecular complexity, try and find the first replicators screening chemical space. But when I went to the funders, I said, Hey, guys, give me $10 million, because I'm going to search for the origin of life. They said, You kidding us? No, you don't know he chemistry is random guy thinks they can play with some robotics, what do you want, he says, Hey, guys, you want to give me some money to make drugs. So they went, that's better. And then that's how I built the technology on the drug side. And then obviously, in my lab, I'm using it to basically be clueless, as one of the people watching would say, wishes could be clueless and explore chemical space. But anyway, I've got quite a lot of modules. For the computer, I finally chemical modules, redox, modules, small modules that we're going to chromatography modules, this little module here, first of all, is able to make the the Pfizer vaccine all in one little robot, okay, and it's about this big, and you don't need a few birds. And this paper, this paper has just been finished right now. This is our atmosphere computer for doing reactive radioactive chemistry. So there's lots of kinds of, you know, challenging, air sensitive, and potentially very dangerous stuff, you wouldn't want a human being there. By having them separate. Maybe you can control it by computer, the thing blows up is more such a catastrophic, someone standing next to it, not excuse to blow it up. But it just gives you another layer of protection. And of course, when we were locked down, it was possible to socially distance because we just simply went into the bariatric setup, set up the robots. So she distance went away, did the chemistry from home went back in the workup. So it was quite, I wouldn't say it was we designed the pandemic for our robots, because who wants to have to be sat at home all the time, but it was a little silver lining. So they came to you system to improve the versions and we wanted to publish the first paper.

I was fairly shameless and say, Look, don't worry too much about the programming language, we just need to we just need to publish a paper that shows we can make molecules what sells well, sex sells. So we make viagara. And then we made a couple of drugs in a night on the cinema and we made them on scales, that human being would make typically useful scales, gram scale, and we were able to make them exactly the same way. And then it worked pretty nicely. But the key thing, one robots, three different drugs, same abstractions, we just had to reprogram if you think about it, now the state machine there, but the you have the reactor, the separator, the evaporator, and the crystallizer, all working together as a state machine. So you have your reagents and your input buffer, they go in, they are transformed, then when that we have purified to take your pill material from step one, put it in a holding Bay, then start reaction to take the reagent, you may take the product from your previous reaction in the next reaction and go forward in loops. So my team were very happy about this. They're pretty excited about using it. But there was one downside that they didn't really appreciate the programming language to start with, because it looked like a lot of coding. So and quite quite understandably, my team said, Well, we love organic chemistry, we love learning to to use our skills correctly. And all you've done is you've forced us now to go from doing chemistry in the lab that we love, Pat by hand to sitting in front of the computer and typing in code, and not always working. And what I tried to do when we were going through this process is say, well look, what we try to do is not just make you coders, but we want to understand the logic the hierarchy of abstractions and chemistry. And really for this, obviously, Noam Chomsky and a few other people have really, really understood what may enable us to understand the relationship between Automator and grammars in the world. And we know that chemistry is a is a language and we know that also can be treated like an automation. So what we want to do is to make a make a a kind of level four if you like Turing machine qualified chemical programming language that would constructing molecules. So what we did is we realize that chemists have gone from the stone tablet, and alchemy to calligraphy. So we all know we will Use the same alphabet. But the problem is we all don't use typewriters. What I mean by that is we don't we don't. When you write down your, your procedure, you don't capture all the information I said earlier, and that digitization doesn't occur. So that means there's always information lost. And that leaves the chemist guessing when they read the procedure. Okay, read at least lots of heart ache, because that information, some of that information, but it couldn't be because you've got a trainee who doesn't know the how to set up that particular type of reaction. It could be that the laboratory has different culture to another lab, play dry, the solvents are different way. Or they use different types of glass and around bottom flasks, or they add a to b rather than B to A, when they write down at a to b. Right. They might mean that when you write that down, they mean Oh, I betway depends culturally on whether you were you read, how do you read left to right? How do you interpret that information? Okay, how do you nest everything in the loop.

So we wanted to make a Hello World became a street. So this is the hello world and a punch card. This is hello world in assembly. I don't know if anyone here is I guess some people Creole is probably written in assembly. And here's how they welded pipes. And so, you know, her work in Python is pretty easy, right? So what we want to do is do chemistry, that very high level where we can express what we want, and it gets compiled down to the hardware. In fact, we don't care what the hardware is, the system just knows and it makes it work. I want hello world on my screen, make it happen. I don't care how that's the type of thing I wanted with the chemistry. To do that, then we had to make a programming approach to it to make any molecule we had have a reader, a chemical process or virtual machine and the way to compile it with a graph. And I won't give too much here because I want to stop in 15 minutes for questions. I'll give you the general idea. And it's I'm very happy to dig in very happy to give you links to papers very happy to give you links to code. So there's not hiding anything, it's just a time thing. And, and actually, I could be a comic nerd and tell you about how it all works, but you may not care. So what we do know though, is let me tell you how the programming actually the basically works. It works on actions that a chemist is like a cook, they would chop stuff up, they would add stuff, they would heat stuff, they would dry stuff they would radiate. So these are the kinds of operations at the top level you want to do. Okay. And because you have these universal actions, they are interpretable by any robot, and then if we write it in the right way, it's human or machine readable. And then you can verify how it works, because you can run it on any robot. We determined just by having say, just if you had the fluid handling the heater, chiller filtration, phase separation, evaporation column chromatography, only with these unit operations available, you could do 95% of chemistry, actually, it's 60 60% strictly because actually low temperature chemistry required a bit more and we now have low temperature temperature that takes a bit higher. And scientists handling textured epi and vacuum distillation takes you all the way up to 95%. Okay, so these are just extra costs an extra engineering difficulties, I would say, That's not bad 60% for a robot that costs less than $50,000. Right, it's cheaper than a peptide synthesizer, which is kind of hilarious. And this is where you in the beginning, we use natural language processing to help us populate the fields for our programming language. But it is not a machine learning programming language. It is a formal language, okay, that has a formal syntax and a set of rules that are expressible. Within that language, if I can say it's written like a lambda calculus. So there's the ADL gets, goes to their compiler, which I like rather the compiler that takes the code takes the graph, compiles it together, and makes it work. I'll give you some of the features, it's modularize, double composable transformations, they're explicit dependencies, you have the ability to express recursion in it. So bit like tree like and convergent experiments keeping, I needed this recursion for my origin of life Y Combinator ideas, but then I realized it's really good for iterative chemistry, because company chemistry, peptide synthesis, RNA synthesis, DNA synthesis, all these things fit into this paradigm very nicely. There's a formulism that we are, we describe everything as transformations, set of inputs, it goes through results, I have a sequence of instructions that we call deviation, the derivations, okay. And that's given here how these derivations work. And we're using an operational semantics sorry, this denotational semantics to label things. As you can see here, but this is just if you like formal methods, and you want to formally check that your computer is going to do what you want. Let me just show you it working right So let's just say you're a chemist and you say, I don't care about this formal stuff. Here is my here's a general organic synthesis. Here's my preparation, can I take literature converted to programming language. And this is just a little app we built that basically, the chemists would go in and take the take the procedure that was written in the literature, correct it like you're in an idle, and then basically, check everything works, then we would load in a graph of our robot. First of all, we give it project title,

generate the graph. This is just a picture of your robot. Okay, check the robot guess the picture of your robot is right. Then you take your code, the graph, checker compiles. So the resources are enough, you've got enough valves, enough pumps, things like that, simulate it, check that it works, then you're waiting and running. This is how easy it is now to take any molecule in the literature, if the literature is not BS. And the sad thing is most of the literature is bs will not be Yes, incomplete. That's the polite way to put it incomplete. But the good news is we know how to make a complete. And also even better news is that although I'm saying some of the literature is bs or incomplete, it's not that hard to make a robot that can search around the parameters and guess what was wrong and fix it. Because most people are not meaning to defraud each other. It's just loss of knowledge. So all we have to do is go rebuild the clinical literature, which is a cool, cool thing. Okay, so what we did to prove this work for lots of different molecules, we did a cross coupling reactions, we made peptides, we've made the azulene, which are really unstable molecules. In fact, during the lockdown, we started a project called Kenny phi 100, where we took the 100 top molecules that everyone wanted to digitize. And sorry, it wasn't methamphetamine or various thcs. They were just like molecules or chemistry, well actually identify I'm sorry, I am. molecules that chemists need, and are really important kind of knows in synthetic space. And we took most of the organic chemists toolbox, toolbox and we digitized it. So from cost couplings, deep protections, reductions, oxidation to enclosures, we did it. And all these molecules you can see on the screen now, we're done in our country, you know, chemical processing unit done for real and purified automatically from the literature. So we've we've done about 50, there's about another 180 million years ago. So you know, but you've got to start somewhere. And just a matter. And the reason why hopefully, you guys kind of sympathetic as you can probably remember, or imagine what it was like when the first transistors were being built and Moore's law had to kick in, were a similar kind of idea here. But if we can get the reliability, then why not just make all the molecules and literature, formalize the code, and then we have access to chemical space in a way we never dragged off before.

Okay, now I want to go on for the last 10, nine minutes, the platforms and algorithms, so I'm gonna slow down a little bit, I'm just going to show you what it can do. So I want to talk about discovery nanomaterials, I've also got a computer that's making a molecular machine that makes a molecular machine. So think about it, you've got a state machine at the macroscale that basically is going to design a molecular machine at the nanoscale, that has a state that then can act on itself to make another amount of nano machine and the outcome of that nano machine controls the macroscopic computer to do a different reaction. We call this downward causation does it cool, you start at the top human being turns it on, gets it to make them automate a molecular machine, that molecular machine makes another molecular machine or stochastic, we've added some decision making algorithm with a sensor, UV maybe, and then the outcome of that reprograms, the macroscopic machine to make a different type of machine. I don't know why that would be useful. But it seemed really cool. To make it about, you know, how do we go from a macroscopic machine to a molecular machine without human interaction, and come You know, think of ways are, you know, tailoring the way we search chemical space with that approach. But let's start simple and next to some, some platforms were built in the lab and some algorithms. So the first one I'm really cool, really, really proud of is we made a robot to basically search nano material discovery space using a physics engine that was simulate the formation of nano materials. And then and then look at how the crystal faces would default deform, and then simulate the plasma. Okay. And then what we do is the idea would be Can we have a robot where you mix stuff together gold precursor, capping ligand reducing agent, and could we then look at the plasma that we simulate, and look for the site simulated plasma only one, look for that, and then And then try and make them up the particle or the facet of material in the machine. And the way we do it like we start the seeds have an have an ideal path, or we go down this tree here, each of these levels reflects different operations in the robot. And the way we do it is we have a virtual space that we then explore. Okay, there's a fitness function linked to the UV data. And there's a chemical space and we sample it. And of course, we will then get TMS at the end, just to check we're not completely off track. And we go around and make these materials. This is what the robot looks like. So it's just a little wheel here with all the sample vials dispensing. If you look, and this is kind of the the platform here, this platform is actually open source, I released it about a year ago. So if you fancy making one, it's like like literally a couple $100 and a bit of 3d printing. And here are all the input pumps. The spectroscopy is on the left hand side here reagents from here. And the way it works is basically you have a dispensing pH control. And you basically have a UV vis characterization. And you also have a point where you actually obviously wash the reactors and reactions out and start again and keep going out of the loop. You endlessly go around the loop to find the perfect plasma on. And then when you found the perfect candidates, they stopped there. And then that is the stuff you keep for later. And you then basically crystallize that, or precipitate that and then you look at the electron micrographs. This is kind of the space of UVs. This is how we search in our class index space, which is a glorified GA, the elite the macro leaps, the way we fit them from a random to our, our you know i fitness is shown here. And then you've got this kind of this set of UV vis here and some updates. But let's just show the results. This is what we get. these are these are not selected. micrographs. It's one of my pet peeves in nanoscience at the moment is that people just basically select their micrographs. Okay, you may not cut and paste them in Photoshop, but you just keep doing the experiment until you get the ones you want, which I think is kind of, you know, kind of kind of annoying. But so these are representative, the batches and little codes on here to show you the kind of where they come out of space. These are all entirely reproducible. If you were to build a robot or have a really simple, we're not even our robot, and another robot would use a kdl. You can reproduce this very easily. And we're trying to facet them. And you can see at the bottom of the row here where we've got these different pointy bits that is harder to control, admittedly, but we're getting there. So that's pretty exciting. Whoops, I went right to the end there, that I had it back, but we'll go back. So we also made the system to look at energy materials.

This is a using the wheel now to make simple capacitors to them, put them onto a thin film to look at metal oxide space. So we integrate the discovery. And the assay is really important in the closed loop. So you're not the human isn't having to put in information. So you really have a hard link between the discovery and the operation to the device. And then we do electrochemistry and we've been making some supercaps. Many years ago, we made an evolutionary engine. And the idea was to make to again to look at the the kind of cell first origin of life. And the idea is that we have a robot that will look for colony formation and complex cell formation with very simple components. And we proved that not only we could produce these things in the lab, using fairly good random slide points, we could get these things to evolve and have some fitness. And we could do quite complex things with them. So these are the fitness landscapes but sorry for skipping that I want to get to the conclusion of the talk. We've also got some design I go back to my roots as an inorganic nano chemist, I love doing nano design with clusters. And the real I suppose reality is your crystal farmer. So you set up hundreds if not 1000s of crystallization that look for crystals. And then when you're lucky you get a nice crystal structure. This is a molybdenum 120 serum three elliptical nano Well, you can see it's three nanometers by 1.7 nanometers. And I had some fun using an out my student was competing with the robot. And basically the student got beaten by the robot and I published this paper called humans be robots and the referees went apeshit was hilarious. Now was the student in theory to the robot. No What the robot did here was kind of cheating. The robots workflow was I'm going to pick a pH at random. So for me, every pH is the same, I have to go through the same procedure, wash out my sample vial, add my acid, add my face, check with the agents. While I want the new reaction. When the student did it, students like I'll just put my sample files in a row and modify the pH and go up increment by increment in dilution, because there's less unit operations. So the student minimized the unit operations where the robot did the same number. So actually, the reason why that the robot was able to explore a great crystallization spaces just because of the fact that a student was not random. So you want to do your students to be random where possible, but but auditable, auditable random is the answer to everything. But anyway, that was a cool thing. We're also using closed loop reactivity, doing organic chemistry with no targets in mind just mixing stuff randomly in the lab, which is very scary to the safety officer. So of course, we don't do it randomly we tell the safety officer, but we don't tell the student so we don't have any bias, biases everywhere, including in organic chemistry. So if we remove the bias, we remove the knowledge of what we're making, what we're using, we can look for new reactivity. This is causing great fun with my organic colleagues, because they think I'm some kind of heretic but it doesn't matter. These are robots we're building to search chemical space, it's good fun, because we're going to use the same discovery systems that use neural networks, we're looking at rate reactivity, to basically also make drug discovery robots, which we've got going on the lab right now to make inhibitors for various kinases, which I'm going to skip over here, just because I'll show you this part, because I think it's quite good fun. So one of the things we're going to do in the computer right now is basically coming up with this is very deadly. JOHN Murray lane, Savalas type chemistry. And start out, of course, we shouldn't hit leaves Fraser out, where the idea is that we're going to try and take a series of molecularly interconnected molecules, and assemble libraries, these catalysts make them switchable in the computer, so the computer would do the reactions, but it wouldn't just be doing organic, covalent chemistry, but it'd be doing mechanical bond chemistry if I use the term correctly. So I know Fraser likes the term being used precisely. The idea is to have a chemical state machine that makes a molecular machine and we have a feedback loop. And that's what we're doing at the moment. So if, if I haven't insulted you, or when you want me back one day, we will I can update you on that and maybe a year or so.

Okay, I'm one minute over, I just want to stop here by saying, you know, all these robots, you've seen use the same programming language, the same approach. So they should nickname we should work on any qualified robot, not just the things we built, that's very important, that abstraction has to work properly. And one of the things I wanted to kind of suggest is that we should go for a vision of computation that allows us to make molecules and complex materials that are universal with respect to the platform, but you do it with by hand, if you're a human, or it can speed robot, or one of my robots or some other robot, you've got its task universal in terms of looking for synthesis, optimization, and discovery. And there's a chemical universality because we're able to use blueprints for making specific chemistries. And as I said before, computation is the process of running our chemical code reliably or any compatible hardware. This is compared to computation, which I think I've explained. I'm going to stop right there. Sadly, my team could not get together for their group photo last November. So we did a lockdown selfie instead of here. We all are there absolutely amazing bunch of people. It's a pleasure to work with them. They've obviously had a very stressful year being locked down. But I think that they survived really well. And it helped keep me sane. And I'd like to also fund thank funders, particularly UK, dsrc, ERC, DARPA, and all these other people here. But last but not least, I'd like to thank you for attention. And I'm really looking forward to answer your questions and having a discussion. Thank you. All. Thanks very much. That was fascinating. So there's a there are a few I wouldn't call necessarily questions in the chat. Some of them are very long sort of statements are a few questions and that are quite long, though. If you want to engage with a few of those. And let's remind people, as this is off the record, if you do want to unmute yourselves to have a discussion, that's also fine. So two, we've referred to being able to take a property of a material that you might want to make and then you predict almost from first principles using your machine. So what do you envisage if you said that was a long term goal one of the things that this group is interested in is identifying even at a very coarse grain level once they have 515 and 30 year goal for this group. All research in a specific area will be an add on if you can offer some comment on that apologize, it's a slightly nebulous question. In five years, or even now, this year, we've got a system whereby we have a problem. So let's say we want to make an enzyme inhibitor, we then produce using a surrogate model candidates for that. But we screened those candidates for what we can make already in our machine. So we only consider those molecules are fit for purpose, that seems bad. But then when you think about the fraction of the universe that is really big in chemical space, you actually realize if you allow the system to go around, go, well, intelligent inverse design, no limitations, and you give it some limitations, that performance that you theoretically generate is really, really high. So it's kind of cool, because it's a bit like turning up in London, where they say, I'm going to drive around London, I need to go from A to B, what roads are open. And you're only allowed to use those roads, or I've just turned up in Tokyo, and those roads don't exist yet. I'm just going to make them you know. And so I think that so in five years, we will be routinely just rerouting within 10 years, we'll be inventing new reactions, and using our knowledge base to go back to basically say, Okay, I could take a route that maybe takes me 20 steps, but I get by just to have this one magic bond I can make here. The thing that's going to happen, I make a prediction that may be exciting for this group, is that right? Now, chemists are thinking about general reaction on small molecules. When the molecules get bigger, and bigger and bigger, there's no such thing as a methodology, the molecule, the bond you make will be locked, the molecule you're interested in, you will have to optimize for that particular bond. That is really exciting and interesting. And I think if we can start to crack that nut in the next say, decade, okay, the dream of making highly complex molecules that are highly functional, in ways that this group is interested in will becoming will become not quite routine, but the route to routine will be there.

He knows that to be annoying, Lee, did I just use a bond?

Yes. I don't mind using metaphysics to explain, you know, my, I was like crayons, referring to the fact on Twitter, I'm winding people up because chemists don't know what bonds are. Because I'll tell you a little secret in my robot. If you put in reagents, and you use Bayesian chemists, Bayesian mathematics, you can get some reactivity profile. But hey, if you start to put in solvents that can hydrogen bond that are not formal bonds, they change the mathematics in such a way, if you just treated them like bonds, you get better answers out. So when I said on Twitter bonds don't exist. What I'm saying is our classification of the energy level of bonds is misguided. Why don't we just do things random, basically, and then just work out if a molecule sticks together? Because a bond on Jupiter may be stable, more stable than a bond say on earth? Because of you know, the fact that the zillion pass bills that are there, but anyway, you know where I'm coming from? I tell you to play playful, where you could put the put the question.

So, are there any other questions that people want to ask us? The Alison's put a poll out there for people to comment on had led to the duration of the group to go Sorry, I

cut someone else has got a great, I have a question. Is anyone any one of these labs that currently has a computer planning on offering cloud computing?

that, yes, my I'm starting a company called chemic phi. And one of the things that came if I will do in the end is to basically have a kind of Amazon for chemistry, but it's ways away because there's regulatory issues. But the nice The nice thing is, let's say, I think it was 10. Right? You ask the question, let's just say that, you know, let's say your your secret name is Heisenberg. I'm like, Oh, my gosh, it's 10. There's gonna be submitting jobs to the, you know, or actually, you're working in a, you know, FDA regulated company, and you need to make compound x pretty urgently and that what you need is legitimate within the regime that the legal say regimes wrong word, you know, what would be the legal jurisdiction in which you live? There's a really interesting corruption technique, which is why having a Turing complete language is what you can do is you can submit the code to the robot in the warehouse, and it's encrypted. It's even encrypted the point of execution. And the robot is known as executing the person who doesn't know what it's executing his tags put out a new then and you get usage rights. So according to your certificate, you know whether you have the right to fly that robot at that level, you can make appropriate thcs Whatever you want. And then you can give people rights to do that. And then so there'll be a lot of cloud computing available. And we'll be able to not just regulate it from a safety point of view, which is vital, but also understand the type of molecules that people using in r&d environments and these supply chains, and

homomorphic encryption or what homomorphic encryption, yeah, yeah, cool. Yeah. How many uses that away? Because we practice and our biotech group of, you know, how you could use,

I've already done it. Theoretically, well, theoretically, we've done one example. But obviously, it has to be qualified. I'm more of a mathematician than I am a chemist. So when I was knocked down, I played around with my, my own robot at home. To I mean, I wasn't making compounds, I was actually changing pH, because I didn't think it was a, you know, on my lap, my home lab is only kind of certified for doing a brisk chemistry, you know, playing with lemon juice. But nevertheless, playing these four more ideas is pretty important. It's really important that we get people not to be scared of the technology, we have an open debate about where and how we want that technology to be used, right? Because I can imagine a time in a, you know, say, 50 years from now, let's say 75 years from now, where you've got continuous gene sequencing. So people and you can, you can basically detect people's illnesses, as they're just starting, in fact, a decade ahead, and they say, I've detected this problem, we're going to make the drug you need, the drug gets designed on site gets emphasized, you get given the drug, you don't even know anything about it. It should be a very, very simple future, if you can digitize chemistry and biology, of course, that's hell of a claim. And we don't know what's gonna happen, because we can't we don't even understand the genome. So,

yeah, there's some really exciting literature on that, from open mind the group at Oxford, but that's a conversation for another day, but yeah, very exciting. Okay, so

Alex has given me is that some NF t from chemical structures? That's brilliant. Alex, can you can you can you send me some for free? Just?

just just just just to get going? Yeah, no, I don't think I'm doing that. I you should

be if you are interested, watch the page for the ken coin is going to come but I need to figure it out properly. Because in the first place, really, what I want to do is imagine the following scenario, you go to you want to make a molecule simply, really urgently, you have two options. There's a really nice paper in nature published by your your, you know, your your one of your favorite chemists ever who is reliable and brilliant. And you can follow it, but it's going to take one month to follow the procedure. Or you can just get the code from Kemi file from a database. And you can see the code has been validated by a number of people like, you know, Facebook likes or whatever, where people gone through mine that code and version did. So then do you, let's say spend $1,000, either leasing a computer or going to a Chem cloud to make it? Or do you waste a month of someone's time, or potentially two months to make the molecule? Well, this is the the problem I had during my PhD is that I spend half my time drawing nice molecules, nice molecular machines, or these Mel organic frameworks that protect signs and all this crazy stuff. But I would always joke that 90% of my PhD was just in the lab doing organic chemistry, which I'm not that good at. So if my project was going badly, I'm an organic chemist, if it was going well. I was an inorganic chemist. And it was going really well. I was a materials chemist. So if I could skip that 90% of the work that I actually did that I didn't want to do. I'm all for it. Yeah, I'm sad. I didn't put in our mouth box. So we actually got an off button. We use image recognition to basically index the crystal faces and they put it on diffractometer. So we have this high throughput thing, also looking at polymorphs. But that's another talk. But But anyway, good to hear that. It's hard doing synthetic chemistry.

What's the big I have something really big. I have something really into ask, which is some since you reusing the same, I think because I understand that you keep reusing the same reactor apparatuses and crystal is that you have some kind of like, tell me a little about the cleaning step.

So yeah, there's a there's an automatic cleaning routine that happens all the time on the system. And not only that, within the computer system, we've got one where we have But not a fruit. But what we have six reactions going at once. And what you can do is we use scheduling software. And we're able to do reactions at different concentrations optimization. So what it does is it says, I'm going to start you first, why are t equals zero, then one hour later, I'll start the next reaction because the extraction site takes one hour and you have an amount of 60, you have an in sync. And then you can basically do the reaction like landing planes at airport one after the other. So to do that, we have to validate a cleaning procedure. When the cleaning procedure we finding, as long as you basically aspirate solvent through the backbone regularly, after you're done your, your unit operation, you very rarely have a problem unless you have a precipitate or you have a particularly sticky material or something like that. So we've been playing with that, unfortunately, I wanted my computer be like a normal silicon computers, like, you don't have to wash out the electrons, right? It's all you know, it's fine. But sadly, but there is a nuclear button quite literally, where if you really have to do an acid wash, you push the button and they will acid wash. And that's, that's really like the nuclear option. And the other thing I've been playing with is double routers, like I like the idea of having a RAID array. So raid arrays, a redundant array of independent disks, why have a rare array or run a redundant array of independent reactions, if I'm doing a 10 step synthesis, and I really need that material, and something messes up, I want to get a failover. And we're playing with failover. At the moment, it's really cool. Because people like work, it's like it just works. And cutting Glasgow, the temperature goes up and down. And things change here and there. You can't control everything. I am a control freak, but I'm not totally perfect. So there's lots of little games playing, we're only just getting started. So cleaning is important. failure, tolerance is important. And also scheduling is important.

For Lee, I had a question, I really enjoyed your presentation, particularly for automating directed evolution, not just in bacteria, which people use. And it's a very messy environment. But my question was, the reaction language that you described reminds me a lot of the development of the LSI. And the people eventually got to design rules. And that led to less skilled people develop more and more complex circuits. So an example for the reaction sequences that you've been talking about and you're in the automation. It does the user need to know enough to avoid making some kind of highly reactive intermediate that will then quickly react with the solution before the next step? Or is there some support tools that to help users get around?

We have our very, first of all, that's a really good question. And I'm a believer in giving the robot some not not ethics. But I mean, I was kind of inspired and depressed by why not actually inspired, I explained there in years ago, there's a tragedy where a gentleman leaves, pyre locked out, locked out, the copilot, locked out the pilot, and he programmed in the autopilot crashed plane into the Alps. Right? committed suicide took the entire it's a very sad. Now have we been thinking ahead? Oh, we don't want planes to fly into error to fly in the mountains, we'll just disallow that. It's not worth it. It's not outside of the width of our software person to do that. We just didn't think about it. Sadly, inspired by that tragedy, I thought well, are there ways we can put in design rules so we can stop people making another chalk by accident? Or basically making a bomb when they don't mean to? And also just messing up reactions got the timings wrong? And the answer is yes, is not that complex when you're doing certain design rules. So say for the RNA synthesis, it's really easy if you're doing the same thing again. And again. And again, when you're doing other reactions, because we have a controlled sequence of where you start and stop a reaction. And the compatibility, there is a compatibility flag. So basically, if the solvent that you're going to clean isn't compatible with the stuff you've just done, then it just stops. And I think there's a very good point going forward. And we want to put a bit more in there and also have an expert mode. But the RNA synthesizer we just made, first of all, when it did 5000 kdl. Like that is the longest synthetic sequence ever done by human being with he did it with a robot, right over two weeks. And rather than than right, he was a bit he was, like, I'll say is lazy. It wasn't lazy to genius, right? He's like, I'm not going to write out these codes. I'm going to make this recursively enumerate and expand out from a blueprint. And that's what he did. And so I was really inspired by that. So I think at the higher level, I one of the things I did a lot during lockdown as I'm pretty obsessed with hardware. So I'm playing around with FPGAs to solve graph problems. So I like instantiating the graph On the FPGA and pruning was the logic. But I mean, you're in very log. And all these higher level languages, right, I'm using a different language that was made by a little company called Alka tree where they make these nice FPGAs and a little compiler like wave for writing FPGAs, you can do a lot without having to know a lot at that level. So I think what is important is you then turn the expertise now, they don't become an expert in the actual logic design. They're all the way up. So I think we all should be inspired by, you know, the way micro processors and computer designers work, right? What we are right now is we're that interface between designing the transistors and the logic arrays, then we go above and you measure the logic arrays robust with respect to the next array up, those are solid, then you can fully express yourself. And there's some really weird stuff that you do in microprocessors like really weird things that shouldn't work. And I think I listened to Jim Keller, who did some stuff at AMD. There's some crazy stuff. They're doing binary decision trees, that just works. But how they know that to work. It's scary when you think it's like this deterministic thing, but it's non deterministic inside. But that's another side. So yeah, it's a great point. And it's something that we're trying to do. And it depends on how I commercialize this because I am starting a company that's going to try and quantify the world if you knows that the company problem is like, like Tesla. I'm trying to do what Elan Musk has done to test that. And I'm not Elon Musk, sadly, or maybe, because I have a software problem, a hardware problem and a culture problem. And a bunch of grumpy chemists. You know, that's, that's a very interesting kind of culture to solve. But we'll get there. And that can get if

you were describing as universality is when you're you're thinking also of expanding that concept, it sounded like you were talking about chemistry in solutions only.

In solids everywhere, it's universal everywhere. The only reason I'm using liquids is because I'm being pretty lazy in any engineering, but it's really important that we understand mathematically in terms of the state machine, we're building that it should be universal, I have built a distillation system of sublimator. One of my one of my postdocs today did a sublimation that is took a material paper basically sublimed it onto a cold finger. So with that, so we've worked out all the unit operations, we need to basically cover materials chemistry, because what I've been doing for chemists II, we are basically generate, you can use inverse design and develop pattern families. So we're going to discover new orleans, new solid state materials, new magnetic materials, but we need to get that underlying infrastructure, pick off the engineering problems. If someone says, okay, maybe a superconductor, we can then evolve a superconductor. But that's, that's a hard technical project. But if we have the language behind it, we know that in operations, it's not that hard to make a simple conductor for other stuff in the tube, evacuated the vacuum heater to 1000 degrees, and then pray. So you know, there's things we can we can make that better. So no, the universality will work. The only barrier we have is a little bit of engineering, and it's significant.

Yeah, that's so another. That reminds me of something I'm sure you've thought of because he's thought of this stuff for years. More like high energy reactions likes like, high pressure, high temperatures. sintering. You know, like you just talked about with the superconductor stuff, ceramic stuff. I mean, this is obviously extensible to that sort of,

yeah, we have to be making a ceramic robot right now. Basically, everyone says it can't work, you'll be too easy and cheap. So I'm like, okay, we'll just make a ceramic robot. It uses gravity is basically like dumpster ball at the moment. So we have a little vacuum chamber that just turns and stuff down. So we are making ceramics.

I spent some time at the ceramics lab at NASA Ames, which does a lot of refractory ceramics. And there's really beautiful things there for like mixing powders, I'm sure you know, all this technology. It's a lot of fun. Then the third question, Lee is um seems to me in sort of the spirit of flow cytometry and the LSI like, and and like this stuff can be typify cancer to some extent, and then maybe massively scaled, that massive you know, smaller quantities, but mass perils.

Yeah, I mean that there is once you know what you're making, then you can look at the scaling One of the nice things about the scale at which we do the chemistry right now there seems to be a sweet spot for discovery. The problem someone asked about comedy care another thing that comedy can if you're going too low down in value, and you don't know what you've made, you're not going to work it out comedy can only knows works when you're just confirming the the ad because you know the molecular weight or something. So you should be able to scale also made a microfluidic me qualified computer because we're putting more into low Earth orbit.

It's pretty weird, though that the if I read you correctly, that the sort of optimal discovery scale is like the scale of a flask good hold in your hand?

Well, there's a number of reasons that's cultural, isn't it? human beings build devices that human beings can interact with. Right? That's number one. And also, if you're looking at an NMR machine, you need a few miles of solvent. And there seems to be something really interesting on Earth, one atmosphere 298, that basically, if your surface area to volume ratio is too small, everything happens, everything happens at the boundary. If it's too much the other way, you've got too much solution and not enough interface. And, yeah, it's I mean, I'm trying to convince my group to do a, a project right now, which is basically called creative reaction volumes, which is, how creative is chemistry, the knomo, or nanoliter, are temporally and all the way up and look at where the optimal discovery is, right? I'm really interested in molecular transformations, single molecules, you know, the origin of life, there was just all the origin of information growing on a polymer, right, you want an isolated molecule, and you don't want polymer growth, you want basically that polymer growth. And that seems to happen in isolation, okay. And you want them to happen everywhere, but spaced apart. And one of the things I wanted to do is kind of program that and explore how much sequence space can I search, when I have nothing in the way, it's a bit like, if I'm going into a football stadium, and oh, my I can get to anything I want. But if I'm in if I mean, if there's loads of you, it's going to be difficult. So if I'm trying to work out how to generate peptide sequences, I need to be I go super dilute and super concentrated and oscillate from like, in my evolutionary system to search the state space, because actually, this is why directed evolution is hard, in some ways, because biology kind of restricts you, you know, to work in a certain state space, you can't get out of that. So I don't know. And there's one of the fascinating things we'll have to do. There's so much science to do here. You know, at the moment, I'm kind of confused by what you know, on what to do next, other than panic about the pandemic, and get funding and get back on an aeroplane or whatever is I need to do next

year, what, what were the will, because there's hundreds of cottage, special economic zones around the world. So what are the regulatory burdens that you're facing?

So at the moment, because of the way that we are going to work in terms of providing technology to people they most like most of those people will only provide technology, the company, to those that are qualified to handle it. For academics, it's easy. What will become harder is when the cloud stuff takes off. And I really, I talked to the FDA, and I'm talking to the NIH right now, because I'm doing some stuff with the NIH are starting to think about it. The real question is, how do we prove people's credentials? And then, and how do we then work at what they're allowed to do and not allowed to do and it all goes back to having a verifiable, auditable trail. And that's actually why blockchain works, right? People think blockchain is a way of hiding stuff from just buying, you know, buying yourself an aircraft carrier by accident. So I didn't mean to just, but actually, if you actually use a blockchain, you can actually make sure your transactions are transparent to the entire world. And there's an incentive for you making sure your track your transactions are transparent, because the regulatory authorities can check you're not doing the wrong thing by accident. In some chemistry lab, you could end up making the wrong molecule by accident, like I didn't mean to make it. And you're formally breaking the law, but you just made a, you know, grout or something that's not allowed in that jurisdiction. So so I don't know exactly how we'll overcome that yet. I think we need use cases, and we just need to talk to local authorities. But that's a long way away, you know, bit, although we will try and cloud it the amount of time. I mean, it depends what happens the next 18 months, if things go the way I want them to companies can expand ridiculously. But it needs a vast amount of money. And someone with a very real vision as big as mine, which I think we might have. But that's not that's not counting chickens. Couldn't you get rid of that problem by having a centralized laboratory space? So you do away with the traditional lab model? Instead, I log into a basically a lab class to collect cloud, it's just a lab that exists. Yeah, that is what we're doing it. I mean, I thought about it in the future, right? The way you can get away with it. People say oh, I want to have a 3d printer in my kitchen and print drugs. Well, no. Why don't you just have a local manufacturing facility pharmacy, and pharmacists will do what they used to do in the good old days, which is actually prepare The the medicine right? So if the if the pharmacy was regulated, then it would just it would just, you know, only dispose the things you had a prescription for. So I think, you know, I've talked to the FDA a bit about this in when we were 3d printing reactors, and using the 3d printed reactors to qualify reactions and so on. And that that caused quite a lot of interesting discussion, because I thought FDA would aid that this is no, no, we would definitely regulate this properly, because you have more control. And the other thing is you want when you're making some materials, you don't want the control system to go out a day, the problem a lot of drugs are being made, they're made on one manufacturing facility, that manufacturing facility is mothballed. After some time read on, you have all sorts of problems. So I think the idea of micro drug manufacturing and certified sites, where you post it from a cloud and gets validated is a will happen. James, I think that's a good point. If we if we reached the end, it's not a discussion, I'm going to read at some point, but I'm certainly interested in questions. I'm really pleased that yeah, it's gone. Really nice. I had one last question, which is how much does you've talked a little bit about how much these things cost to set up, depending on the application that you were going for. So how much is your organic synthetic setup to produce? Um, it's not. So I can say I want to clarify selling it for roughly, but to actually make that set up, if you're going to do it yourself, and by all the bits and bolts, and let's say you just magically have the software, probably 50 to 60,000 pounds maximum. And if you had your own ratio, evaporated, your stairs and so on, the cost goes down to about 20k because the main cost is the chiller, the rotary evaporator. And then some other some of the some of the pumps at the pumps and valve heads, right this is this I'm talking to many companies who are making fine chemicals at the moment as in finding chemical manufacturers, the robot that GSK have we bought we work with I lead them one to meet dietary needs. And I wrote the code dyes are in cost a million dollars of all reason why it costs a million dollars, a mole is only made in batch sizes of say 20 grams, and they decompose every three weeks, right because it's just UV sensitive. So basically,

the computer that GSK have lent them for free, has basically generated I think, a million pounds worth of products. So I missed the trick there. But I really needed them to kind of get. But what I'm saying is you what you could do is think about what compact let's say you're an expert in whatever area you tooled up. During the day you do research by night, you generate material that you then sell through the university legitimately that's not doing anything under. And then what you do is you then use it you work with university, I use my income from there, so augment my day job right. And then then labs can then become kind of on demand chemical manufacturing sites. Which I think is a cool idea. But I mean, maybe that's a little bit away, but I don't think it's that difficult.

Well, they can also be I think this is what you're alluding to. They can manufacture stuff that has short shelf life. Yeah.

Yeah, exactly. There's a lot of there's a lot of radioactive, I've invented the outcome meter. So we're gonna put a computer in in a nuclear reactor and just basically bombard it with neutrons, because we can. Because I would love to do alchemy, I'd like to be. But the other thing is you do radio chemistry in hospitals as well, right? One of the big things about labeling some of the pharmaceuticals that you've got, you want to be able to, you've got to be able to reliably and quickly get to the patient. So there's things like that. And also nuclear decommissioning, there's applications there as well. So they are all on the cards, just got to make sure the system doesn't keep breaking. It's not that bad. Actually, it's quite funny that we have a log of all the hardware failure. And I think about what crayon was talking about earlier, actually, about the what we're actually doing is keeping a log of how it all fails. It's like making it I imagined like when people making the first automobiles, like when the pistons fail, we know how often we have to change the syringes and the tubing and so on. And we and you know, with, you know, several 1000 hours a day a lot, so we just we can anticipate what will break. It's not magic, it's just wear and tear. Okay, I've almost it's almost half past eight my time. I think I should run but I'd love to continue the conversation. So people want to email me. I'm happy to continue it. Thank you very much, James and Alison for letting me come along. I'm going to come and attend these I realized the candidate so I should actually attend I didn't realize you had such an engaged bunch of audience committed people. So that's great.